name
stringlengths
5
6
title
stringlengths
8
144
abstract
stringlengths
0
2.68k
fulltext
stringlengths
1.78k
95k
keywords
stringlengths
22
532
588424
Convergence of a Difference Scheme for Conservation Laws with a Discontinuous Flux.
Convergence is established for a scalar finite difference scheme, based on the Godunov or Engquist--Osher (EO) flux, for scalar conservation laws having a flux that is spatially dependent through a possibly discontinuous coefficient. Other works in this direction have established convergence for methods employing the solution of 2 2 Riemann problems. The algorithm discussed here uses only scalar Riemann solvers. Satisfaction of a set of Kruzkov-type entropy inequalities is established for the limit solution, from which geometric entropy conditions follow. Assuming a piecewise constant coefficient, it is shown that these conditions imply L1-contractiveness for piecewise C1 solutions, thus extending a well-known theorem.
Introduction . The subject of this paper is a nite dierence algorithm for computing approximate solutions of the Cauchy problem for scalar conservation laws of the form 1). The ux k(x)f(u) has a possibly discontinuous spatial dependence through the coe-cient k, which is allowed to have jump discontinuities. A simple physical model corresponding to 1.1 is the Witham model of car tra-c ow on a highway [11]. The spatially varying coe-cient k corresponds to changing road conditions. Equations of this type have been addressed by several authors in recent years, often as the simplest examples of nonstrictly hyperbolic systems [6], [7], [20]. This approach is based on the observation that 1.1 can be written as a 2x2 system: The eigenvalues associated with this system are The system fails to be strictly hyperbolic if f 0 can vanish, in which case the system is described as resonant. Even when k and the initial data u 0 are smooth, solutions to 1.1 develop discon- tinuities, and so weak solutions are sought. A weak solution is a bounded measurable function Z RR Z R for all 2 C 1 (R [0; 1)). This paper considers the conservation law 1.1 within the framework of scalar nite dierence schemes. The spatial domain R is divided into cells I with centers at the points x Z. Similarly, the time domain [0; 1) is discretized via t nt for resulting in time strips I j;n be the characteristic function for the rectangle R n October 30, 1999 jtowers@cts.com J.D. Towers The nite dierence scheme then generates, for each mesh size constant approximation The initial data is discretized via the piecewise constant approximation x Z I j where j (x) is the characteristic function for the interval I j . The approximations u n are generated by the explicit, conservation-form, time marching algorithm Here, 1is based on a two point monotone (nondecreasing with respect to the second argument, nonincreasing with respect to the rst) numerical ux, h(v; u), consistent with the actual ux, i.e., h(u; In order to maintain the monotonicity of the scheme, the arguments of the numerical ux are transposed when the coe-cient k is negative, so that In particular, this paper focuses on the closely-related Godunov [3] and Engquist- Osher (EO) [2] uxes. The Godunov ux is derived by applying the exact solution operator for 1.1, with constant to piecewise constant initial data. The numerica ux that results is then h(v; Here u G (v; u) is the similarity solution of the resulting Riemann problem with right and left states v and u, evaluated anywhere along the vertical half line t > 0 in (x; t)-space where the jump in the initial data occurs. The following formula for the Godunov ux is derived in [14]: min [u;v] f(w); if u v The Godunov ux is Lipschitz continuous, and for f 2 C 1 it has piecewise continuous partial derivatives. With the notational convention that a = min(a; 0), it follows from 1.6 that The EO ux, which is dened by is also Lipschitz continuous, but is smoother than the Godunov ux. For f 2 C 1 , it has continuous partial derivatives satisfying (1. Dierence Scheme for a Discontinuous Flux 3 It follows from 1.7 and 1.9 that kf 0 k1 serves as a Lipschitz constant for both uxes. In the case where the ux f is convex or concave, with a single stagnation point, the EO and Godunov uxes are identical, except for the single case where the two arguments are joined by a sonic shock, i.e., f 0 (v) < 0 < f 0 (u). In the scheme 1.4, the coe-cient k is approximated at each cell boundary, resulting in a discretized version of k, where j+ 1(x) is the characteristic function for the interval I j+ 1= [x j ; x j+1 ). The discretized version k has jumps at cell centers, as opposed to cell boundaries, i.e., the discretization of k is staggered with respect to that of u. This results in a reduction in complexity, compared with the approach where the two discretizations are aligned. When the discretizations of the conserved quantity and ux coe-cient are aligned, more complicated 2x2 Riemann problems arise, and thus a 2x2 Riemann solver is required. The concept of reducing complexity via mesh staggering is well known. In [19], Temple used the Glimm scheme, solving 2x2 Riemann problems, to establish existence and uniqueness for a 2x2 resonant system of conservation laws. The singular function , used to prove compactness in x3 of this paper, originated in [19], in a slightly dierent form. (Unlike the system 1.2, the system studied was not reducible to a scalar conservation law, so the scalar scheme 1.4 discussed in this paper is not applicable to that problem.) The authors of [12] and [13] studied the 2x2 Godunov method as it applies to They adjoined the equation k treated 1.10 as a 2x2 system, for the purpose of modeling a resonant system. Using to establish compactness, they proved that the 2x2 Godunov method is convergent (modulo extraction of a subsequence) to a solution. They established a bound on the total variation of a function that corresponds to the function z of this paper, but their method of analysis was dierent, entailing a study of the various waves arising from the 2x2 Riemann problems. In [13] it was observed that time independent bounds on derivatives (measured via ) had been achieved for the 2x2 Glimm and Godunov schemes, but not for any scalar scheme that applies to 1.10. Section 3 of this paper provides a bound of this type for equations of the form 1.1, which constitute a nontrivial subset of 1.10, and are sometimes used to model the more general 1.10 [6], [7], [20] . Outside of the dierence scheme methods, the front tracking approach has been eective, computationally, and as an analytical tool for studying solutions to 1.1. Reference [7] establishes existence, uniqueness, and asymptotic behavior of the solution of 1.1, under essentially the same concavity assumptions about the ux f as this paper. The singular function , in the form appearing in x3 of this paper, is used in [7] also. A bound on the variation, measured via , is achieved by studying the wave interactions that arise. Additionally, satisfaction of a wave entropy condition is established for the limit of the front tracking scheme, and uniqueness is shown to follow from this entropy condition. In [6], it is shown that solutions of 1.1 and 1.10 depend continuously on k. Reference [20] discusses computational di-culties that can arise with schemes for 1.1 and 1.10 that are based on the solution of 2x2 Riemann problems. Specically, the 4 J.D. Towers random choice method and front tracking method can fail, due to variation blow-up, and non-existence of solutions to 2x2 Riemann problems. The 2x2 Lax-Friedrichs scheme does not experience these failures, as is shown in [20] by examples, since it avoids solving Riemann problems. The scheme 1.4 discussed in this paper, which is based on scalar Riemann solvers, is also unaected by these di-culties. In the example provided in [20] where a solution to the Riemann problem fails to exist, the source of this failure is a sign change in k. This provides the motivation for addressing the case of indenite k in x1 and x2 of this paper. This paper is organized as follows. Section 2 discusses the monotonicity and L 1 - contractiveness properties of the scheme 1.4, including the case of indenite k. Section 3 focuses on the case of positive k, and establishes the main result of the paper, convergence of a subsequence to a weak solution of the conservation law. Section 4 addresses the related issues of uniqueness and entropy satisfaction for the limit solution. 2. Monotonicity. For the case of constant k, the theory of monotone schemes has been essentially complete for many years [1], [5], [9], [17]. In that setting, approximations generated by monotone schemes are well-known to share many of the properties of the actual entropy weak solution to the conservation law 1.1, including monotonicity, L 1 -contractiveness, and satisfaction of a discrete entropy inequality. Their major drawback is that they are at best only rst order accurate even in regions where the solution is smooth [5]. Nevertheless, they provide the starting point for many of the modern higher order accurate schemes, some of which are constructed by modifying the two-point monotone ux function with higher order correction terms, and then applying ux limiters [4], [15]. The ux limiters damp out spurious oscillations that the correction terms often generate in regions of rapid transition. A nite dierence scheme such as 1.4 is monotone if When k is constant, monotonicity of the scheme follows from monotonicity of the numerical ux h under suitable CFL conditions. It follows that, in addition, the computed approximations u n remain within the convex hull of the initial data for all n. The following proposition provides an analog of these properties for the variable k situation. Proposition 2.1. With the CFL condition 2kkk1 kf 0 k1 1 both the Godunov and EO versions of the scheme are monotone. If k is nonnegative or nonpositive the CFL condition can be relaxed to kkk1kf 0 k1 1. If the initial data u 0 (x) lies within the interval j0 , for all j, and all n 0. Proof. Let . Expressing the three-point scheme as u n+1 the proof proceeds by showing that the partial derivatives @G=@u j+i are each non- negative, from which monotonicity of the scheme follows. That the partial derivative of G with respect to u j 1 is nonnegative is clear from similar formula for @G=@u j+1 shows that it is also nonnegative. For @G=@u j , there are four cases, depending on the signs of k j 1and k j+ 1. If k Dierence Scheme for a Discontinuous Flux 5 The case where k j 1 0 and k j+ 1 0 is similar, and the calculation is omitted. For the case where k j 1< 0 and k j+ 1> 0, The case where k j 1> 0 and k j+ 1< 0 is similar and is omitted. The stated invari- ance of with respect to the computed solution u j follows from the monotonicity of the scheme, along with the fact that f vanishes at u and u. Specically, For the remainder of this paper it will be assumed that u loc denotes the space of locally integrable functions w having bounded total variation, denoted TV (w). Also, it will be assumed that k is constant for large x, specically, that there are constants k(+1), k(1), and X such that . With the cell-average type discretizations discussed in x1 for u 0 and k, the discretized versions also satisfy these assumptions, and the various norms of the discretized quantities, u are bounded uniformly in by the corresponding norms for u 0 and k [1]. Proposition 2.2. With the CFL conditions in Proposition 2.1 and the assumption that the initial data satises u is constant for large x, the scheme 1.4 is L 1 -contractive, i.e., the inequality holds for a pair of approximate solutions u n generated by the scheme. The following inequality also holds: Proof. Both solutions u n remain in L 1 . The L 1 bound follows from Proposition 2.1. The L 1 bound follows from the nite range of in uence of the initial data, along with the assumption that k is constant for large x. These observations along with the fact that the scheme is monotone (due to the CFL condition) and conservative, add up to the hypotheses of the Crandall-Tartar Lemma [1], giving 2.1. The inequality 2.2 follows from L 1 -contractiveness, applied inductively, to two successive time steps, u n and u n+1 j , of a single computed solution. 6 J.D. Towers 3. Convergence. In the case where the coe-cent k is constant, monotonicity implies that the scheme is Total Variation Decreasing (TVD). However, even for smooth, but nonconstant, k, the total variation of the solution to the conservation law generally increases, at least initially, and so there is no hope that the numerical scheme for variable k will be TVD. As in [7], [12], and [19], the approach to convergence in this paper is to use the singular mapping rst used by Temple in his study of a resonant hyperbolic system [19]. Some assumptions about the ux function f(u) will be required before dening . The ux function f 2 C 2 [0; 1] is assumed to be strictly concave on [0; 1], i.e., the interval (0; 1), f(u) > 0, with a single maximum f at u 2 (0; 1). These are similar to the conditions imposed on f in [6] and [7]. Then the singular function is dened by f Z u Here In the remainder of this paper, k will be assumed to be bounded and strictly positive: 0 < k k(x) k. Then 1.5 becomes The convergence result of this paper remains valid (with a more restrictive CFL condition, and a larger bound on the variation of z ) if k is allowed to have nitely many points x where k(x )k(x In order to avoid obscuring the main idea of the paper, the analysis of this more general case is not presented. For each value of k in [k; k], (; is an increasing, 1 1 mapping. It is regular everywhere, except at the stagnation point u , where both f 0 and @ =@u vanish. The following Lipschitz continuity relationships in u and k follow directly from the denition of and the conditions imposed on the kf 0 k1 The function maps a function w(x; t) into a new function z(x; t) via z(x; (w(x; t); k(x)). The following elementary facts concerning z are readily veried. Proposition 3.1. Let w 1. For each t 0, z(; 2. For each t 0, z(; loc (R), and R +B It is now possible to state the main result of this paper. Theorem 3.2. Let k 2 L 1 loc taking constant values for large x, and assume that 0 < k k(x) k. Let f 2 C 2 [0; 1], f 00 < 0, in (0; 1), with a single maximum f at u 2 (0; 1). Let Let the mesh size ! 0 with xed and satisfying the CFL condition resulting in the sequence of approximations u . Then, there is a weak solution u of 1.1 and a subsequence u i such that u in L 1 loc (R [0; 1)). The rest of this section carries out the analysis required to prove Theorem 3.2. The basic approach is standard as in [17] and [18], with the exception that the quantity of immediate interest is z instead of the conserved variable u . Compactness for z follows in the usual way from bounds on the total variation and the along with a time-continuity estimate. Once the existence of a subsequential limit z has been established, the invertibility of then allows the corresponding Dierence Scheme for a Discontinuous Flux 7 solution u to be recovered from the limit z, with u ! u guaranteed by the continuity of . Finally, a version of the Lax-Wendro Theorem [10] demonstrates that the limit u is a weak solution to 1.1. Lemma 3.3. Let For both the Godunov and EO versions of the scheme, elsewhere. Proof. Like (u; k), is a strictly increasing function of u, so assume that u j > u j+1 . For concave f , and u j u j+1 , the Godunov and EO uxes are identical, so 1.9 holds for both uxes. Also, for either ux, Z h u Z u j+1 Using the following decomposition of j j+1 , Z Z Z it su-ces to show that the following two inequalities are satised: Z u j+1 Z u j+1 For 3.3, if are to the left of u , so the integral vanishes, and the inequality holds. So assume that R u j+1 R h u R h u In either case 3.3 holds, since then Z h u Z u j+1 Z u j+1 Z u j+1 For the integral vanishes. If Z u j+1 h u Z u j+2 Z u j+1 h u Z u j+1 8 J.D. Towers and denote forward and backward dierence operators. For example, With the assumptions stated in Theorem 3.2, uniformly, for all n 0, and all > 0. Specically, Proof. The superscript n is suppressed in the proof, except where two time levels are of interest. Taking into account the jumps in z at the cell boundaries (due to jumps in u ) and the jumps at the cell centers (due to jumps in k ), the total variation of z(x) is where u The second sum in 3.6, due to jumps in k , is bounded by TV (k), using 3.1. For the rst sum, let Summing over all of the jumps, and using the fact that z(x) ! k(1) as x ! 1, results in u It follows from 3.7 that Then, The following identity will be useful in estimating f By Lemma 3.3, Dierence Scheme for a Discontinuous Flux 9 The last inequality uses the fact that jh j+ 1j f , which follows directly from the conditions on f , along with 1.6 and 1.8. Substituting this into 3.8, shifting indices, and applying L 1 -contractiveness gives f f The nal step in the proof is to estimate the term j. For the jth term in this sum, Summing over j results in Substituting this into the last inequality in 3.9 yields When all of the terms are collected, the bound 3.5 stated in the theorem results. Lemma 3.5. With the assumptions stated in Theorem 3.2, there is a constant L, independent of the mesh size , such that for n > m 0, Z R Proof. Taking into account the constant values of z in each half-cell, the integral on the left side of 3.11 is An application of 3.2 and L 1 -contractiveness yields Z R f kf 0 k1 f (n m) J.D. Towers The proof is completed by bounding independently of , using 3.10. The following is essentially the Lax-Wendro theorem [10]. Theorem 3.6. Let be a sequence of approximations computed via the scheme 1.4 which converges in L 1 loc (R [0; 1)) and with uniformly bounded, to u 2 L 1 loc (R [0; 1)) (R [0; 1)). Then u is a weak solution of 1.1. Proof. Let (R [0; 1)) and t n . Jx such jxj B. As in [10], multiply the scheme 1.4 by n 0, then sum by parts to get where h j+ 1= h(u arguments [11], the sums involving j converge to their integral counterparts in 1.3, as It remains to show that xt Z RR kf(u) x dxdt: The proof of 3.12 reduces to verifying that Z I j tx The sum appearing in 3.13 does not exceed Z R xZ R which proves 3.13. The expression on the left side of 3.14 is bounded by Z The rst integral in 3.15 tends to zero by assumption, and the following estimate tx Z tx Dierence Scheme for a Discontinuous Flux 11 shows that the second integral also approaches zero. It is now possible to prove Theorem 3.2. Proof. The CFL condition guarantees that the computed solutions u remain within [0; 1]. An application of Proposition 3.1 gives uniform (with respect to both t and ) bounds on kz (; t)k 1 and kz (; t)k compact interval Theorem 3.4 provides a uniform bound on TV (z (; t)). Finally, from Lemma 3.5, it follows that where the constant C is independent of and t. By standard compactness arguments applied to the sequence z , there is a subsequence, z i , which converges in loc (R [0; 1)) to some function z 2 L 1 loc (R [0; 1)) (R [0; 1)). There is a further subsequence, also denoted z i , which converges a.e. to z. Let u(x; Due to the strict monotonicity of (; k), the function u is well- dened a.e., loc (R [0; 1)) (R [0; 1)). Dropping the subscript on , it remains to show that u ! u a.e. and in L 1 loc (R [0; 1)). Using the fact that u ju u jdxdt The rst integral tends to zero by the bounded convergence theorem, due to the continuity of 1 as a function of its rst argument. For the second integral, an estimate of j 1 (z ; dierentiation gives @k @k R u When the numerator and denominator are each expanded in a Taylor series about u , the result is @ 1 @k 2k kf 00 k1 ju u j: With this estimate, convergence of the second integral follows from the fact that Having established convergence in L 1 loc (R [0; 1)), there is yet a further subsequence u , which converges to u a.e., as well as in L 1 loc (R[0; 1)). Theorem 3.6 proves that u is a weak solution of the conservation law. 4. Entropy satisfaction. Even for constant k, solutions of 1.1 are not necessarily unique. Additional conditions, usually referred to as entropy conditions, are required to single out the physically relevant solution. When k 2 C 1 the Kruzkov entropy condition applies [8]. Specically, uniqueness is guaranteed if Z RR dxdt 0 holds for all c 2 R, and for all nonnegative 2 C 1 0 (R R condition 4.1 is not directly applicable if k is discontinuous, as was observed in [7]. 12 J.D. Towers For k discontinuous, the authors of [7] proved uniqueness within the class of solutions that satisfy a wave entropy condition. The wave entropy approach has the advantage of not requiring that the solution satisfy additional regularity conditions. The approach in this section is to concentrate on the Godunov version of the scheme, and assume, as in [6] and [7], that k has nitely many jumps. Proposition 4.1 ensures that the limit solution u satises the Kruzkov entropy inequalities 4.1 locally, away from the jumps in k. Proposition 4.3 provides Kruzkov-type entropy inequalities that apply when the test function has support which intersects one or more jumps in k. Theorem 4.5 shows that these entropy inequalities imply geometric entropy conditions for piecewise C 1 solutions, and Theorem 4.6 applies the geometric entropy conditions to show uniqueness with the additional assumption that k is piecewise constant. Proposition 4.1. In addition to the conditions stated in Theorem 3.2, let k be piecewise C 1 , with a bounded derivative, jk 0 (x)j for all x, and with nitely many jumps (in k and k 0 ) , located at 1 < . Let u be a convergent subsequence generated by the scheme 1.4 using the Godunov ux, converging to u, as in Theorem 3.2. Then the Kruzkov entropy inequalities 4.1 hold for every real number c, and every smooth test function 0 with compact support in t > 0, g. To avoid complications arising from the discontinuity in V 0 the following entropy inequality is established for smooth V before proceeding with the proof of Proposition 4.1. Lemma 4.2. In addition to the assumptions of Proposition 4.1, let (V; F ) be a convex entropy pair for 1.1, i.e., V is convex and F assume that 1]. For every smooth test function 0 with compact support in t > 0, g, and every c 2 R, the following inequality holds: Z RR Z RR Proof. Write the scheme 1.4 as Let u G the Godunov discrete entropy ux dened by H j+ 1= j+ 1), the discrete entropy inequality holds for w n . After rearranging terms, a discrete entropy inequality for the scheme results: Multiplying by a smooth, nonnegative test function with compact support in t > 0, g, and proceeding as in the proof of Theorem 3.6 gives xt xt Dierence Scheme for a Discontinuous Flux 13 As approaches zero, the rst sum in the top line of 4.2 converges to in the proof of the Lax-Wendro theorem. For the second sum in the rst line of 4.2, convergence to reduces to establishing tx where J , N , B, and T , are dened as in the proof of Theorem 3.6. Using the fact that kV 0 k1 kf 0 k1 is a Lipschitz constant for H j+ 1, the expression on the left side of 4.3 is bounded by Z u G The rst term in 4.4 tends to zero. Using jF and the fact that u G 1lies between and u n j+1 , the second term does not exceed Z The estimate 3.16 proves that this term tends to zero also. Consider the second line of 4.2. Since k 2 C 1 within the support of , xt Z RR The last term in 4.2 to be dealt with is xt ))=t. Expanding the divided dierence in a Taylor series yields Due to L 1 -contractiveness, and the time continuity estimate 3.10, the sums involving the last two terms on the right side of 4.5 approach zero as ! 0. The rst term on the right side of 4.5 approaches V 0 (u)f(u)k 0 in L 1 loc (R R + ), which gives xt Z RR It is now possible to prove Proposition 4.1. Proof. As in [8], approximate the convex function ju cj by a sequence of twice continuously dierentiable convex functions V i . Apply Lemma 4.2 for each V i , and 14 J.D. Towers Let -(x i ) denote the Dirac measure with support located at i . Proposition 4.3. With the same assumptions as in Proposition 4.1, the following inequality holds for the special case where c = u , for all nonnegative 2 0 (R R Z RR ju Z RR dxdt 0: Proof. Proceeding as in the proof of Lemma 4.2 gives an inequality of the form 4.2, with V and the corresponding versions of F (u) and H j 1. The sums in the rst line of 4.2 converge to their integral counterparts, as in the proof of Lemma 4.2. For the second line of 4.2, there are n such that Using the fact that h j 1= f(u G With this inequality, 4.2 becomes xt xt The term in the bottom line of 4.7 converges to the integral in the second line of 4.6. This can be veried by breaking the spatial portion of the sum in the bottom line of 4.7 into sums over intervals where k is dierentiable, and isolating the nite number of cells where the jumps in k are concentrated. Let u be a piecewise smooth solution to 1.1. It follows from a standard test function calculation that the Rankine-Hugoniot condition across a jump in k at one of the points where the subscripts L and R refer to limits from the left and right, respectively, at the jump in k. Lemma 4.4. Let F )). For a pair of states uL , uR satisfying the Rankine Hugoniot condition 4.8, kRF Proof. Take kL kR ; the other case is similar. By considering the various cases for a pair of states (u L ; uR ) which satisfy 4.8, the following relationships result: Dierence Scheme for a Discontinuous Flux 15 1. uR > u > uL ) kRF 2. 3. 4. uR < u < uL ) kRF The rst three cases cover the situation where the equation on the right side of 4.9 is satised. In each of those cases, the inequality on the left side of 4.9 also holds. Case 4 is the only case where the right side of 4.9 fails. In that situation, the left side of 4.9 also fails, since uL > u ) fL < f , and so The entropy inequalities 4.1 and 4.6 yield the following fact concerning the satisfaction of geometric entropy conditions. Theorem 4.5. In addition to the assumptions of Proposition 4.3, assume that the limit solution u is piecewise C 1 . Suppose that k jumps from kL to kR at and u is C 1 in some neighborhood of the point each side of the notation For a discontinuity located at away from the jumps in k, assume that u is some neighborhood of each side of a C 1 curve (t). With the following standard entropy condition holds: where the shock speed Proof. For 4.10, let N( a neighborhood of the type described in the statement of the theorem. Let 0 be a smooth test function with support in the rectangle centered at extending backward and forward in time from t 0 by an amount . Assume that and are small enough that standard test function calculation, applied to 4.6, gives Z The integral over can be made arbitrarily small by shrinking the width of the rectangle. It follows that kRF j. The entropy condition 4.10 then follows from Lemma 4.4. The geometric entropy condition 4.11 follows directly from Proposition 4.1. For k 2 C 1 , it is well known that if 4.1 holds for all c 2 R, then the geometric entropy condition 4.11 is satised [8]. The geometric entropy condition 4.11 is the usual geometric entropy condition satised by shocks in the smooth k case. It requires that characteristics on both sides of the shock extend toward the x-axis when followed backward in time from the shock. The geometric entropy condition 4.10 requires that the characteristics on at least one side extend toward the x-axis. In the constant-k setting, the author of [16] established L 1 -contractiveness of solutions which satisfy the geometric entropy condition 4.11. The following theorem is presented as evidence that limits of the scheme 1.4 are the physically J.D. Towers relevant solutions. It extends the theorem in [16] to the case of piecewise constant k, assuming that the additional geometric entropy condition 4.10 is satised. Theorem 4.6. In addition to the previous assumptions, let k be piecewise con- stant, with nitely many jumps, located at 1 < Let u and v be solutions of 1.1 satisfying the geometric entropy conditions 4.10 and 4.11. Assume that the initial data u 0 and v 0 are piecewise C 1 , and u Then, for t 0, Proof. Following [16], the approach is to show that the time derivative of the integral on the left side of 4.12 is nonpositive. That integral is broken up into integrals over segments where u v does not change sign. This decomposition yields which is constant within the interval Dierentiating 4.13 with respect to time results in terms of the form d dt _ In [16], it is noted that 4.14 holds even if there are shocks in the interior of For essentially the same reason, i.e., the Rankine-Hugoniot condition 4.8, the relationship 4.14 holds even if one or more of the points j lie within (or x i+1 ) away from the jumps in k, the contribution to 4.14 from x i (or x i+1 ) is non- positive, since the argument in [16] applies in this case, due to the geometric entropy condition 4.11. The only case remaining is where u v changes sign at a jump in k, say x assume that kL < kR , the other case being similar. There are contributions to 4.13 from two consecutive terms of the form 4.14. Taking into account that _ along the line the total contribution is The situation where can be eliminated, since then the Rankine-Hugoniot condition 4.8 implies that both i and i+1 vanish. Take the case where 1. The other case is similar. Then 4.15 becomes In order for there to be a sign change, at least one of uR < uL , v R > v L must hold. Take the case where uR < uL . With kL < kR and uR < uL , the geometric entropy condition 4.10 and Rankine-Hugoniot condition 4.8 require that uL < u . This is easily veried by viewing being determined by the intersection of a horizontal line with the graphs of kL f(u) and kR f(u). Since vL < uL u , and f is increasing in (0; u ), it is clear that f(vL ) < f(uL ). Then using 4.8, the expression 4.16 is equal to Dierence Scheme for a Discontinuous Flux 17 Next, take the case where v R > v L . With kL < kR and v R > v L , the Rankine-Hugoniot condition 4.8 requires that v R > u . If v L u , then uL > v L ) uL > u . The entropy condition 4.10 then requires that uR u . Then, with 4.8, uL > v L ) uR > v R , which is a contradiction, so it must be that v L < u . If also uL < u , then case 4.16 reduces to If uL > u , the entropy condition 4.10 requires that uR u . Then, uR < In this case 4.16 equals 2kR --R Monotone Di One sided di On a class of high resolution total-variation-stable nite dierence schemes Stability of conservation laws with discontinuous coe-cients Convex conservation laws with discontinuous coe-cients First order quasilinear equations in several independent variables Accuracy of some approximate methods for computing the weak solutions of a Systems of conservation laws Numerical methods for conservation laws A comparison of convergence rates for Godunov's method and Glimm's method in resonant nonlinear systems of conservation laws Suppression of oscillations in Godunov's method for a resonant non-strictly hyperbolic system the entropy condition High resolution schemes and the entropy condition An example of an L 1 On convergence of monotone di Shock waves and reaction-diusion equations Global solution of the cauchy problem for a class of 2x2 nonstrictly hyperbolic conservation laws The solution of nonstrictly hyperbolic conservation laws may be hard to compute --TR --CTR Michael Herty , Mohammed Sead , Anita K. Singh, A domain decomposition method for conservation laws with discontinuous flux function, Applied Numerical Mathematics, v.57 n.4, p.361-373, April, 2007 Sebastian Noelle , Normann Pankratz , Gabriella Puppo , Jostein R. Natvig, Well-balanced finite volume schemes of arbitrary order of accuracy for shallow water flows, Journal of Computational Physics, v.213 Adimurthi , Siddhartha Mishra , G. D. Veerappa Gowda, Conservation law with the flux function discontinuous in the space variable-II, Journal of Computational and Applied Mathematics, v.203 n.2, p.310-344, June, 2007 Raimund Brger , Anbal Coronel , Mauricio Seplveda, On an upwind difference scheme for strongly degenerate parabolic equations modelling the settling of suspensions in centrifuges and non-cylindrical vessels, Applied Numerical Mathematics, v.56 n.10, p.1397-1417, October 2006 S. Berres , R. Brger , K. H. Karlsen, Central schemes and systems of conservation laws with discontinuous coefficients modeling gravity separation of polydisperse suspensions, Journal of Computational and Applied Mathematics, v.164-165 n.1, p.53-80, 1 March 2004
conservation laws;difference approximations;discontinuous coefficients
588428
Long-Time Energy Conservation of Numerical Methods for Oscillatory Differential Equations.
We consider second-order differential systems where high-frequency oscillations are generated by a linear part. We present a frequency expansion of the solution, and we discuss two invariants of the system that determine the coefficients of the frequency expansion. These invariants are related to the total energy and the oscillatory harmonic energy of the original system.For the numerical solution we study a class of symmetric methods that discretize the linear part without error. We are interested in the case where the product of the step size with the highest frequency can be large. In the sense of backward error analysis we represent the numerical solution by a frequency expansion where the coefficients are the solution of a modified system. This allows us to prove the near-conservation of the total and the oscillatory energy over very long time intervals.
Introduction . Long-time near-conservation of the total energy and of adiabatic invariants in numerical solutions to Hamiltonian differential equations is important in a wide range of physical applications from molecular dynamics to nonlinear wave propagation. Backward error analysis [BG94, HaL97, Rei98] has shown that symplectic numerical integrators approximately conserve the total energy and adiabatic invariants over times that are exponentially long in the step size; more precisely, over times of length exp(c=h!) where ! is the highest frequency in the system. Such a result is meaningful only for h! ! 0, which is often not a practical assumption. For example, in spatially discretized wave equations h! is the CFL number, which is not chosen small in actual computations. Recently, in [GSS99, HoL99] new symplectic or symmetric time-stepping methods have been studied which admit second-order error bounds on finite time intervals independently of the frequencies of the dominant linear part of the system. In particular for such "long-time-step methods", the case is of no computational interest. The situation is reminiscent of stiff versus nonstiff differential equations, where stiff integrators are not appropriately analyzed by considering only the limit behavior h ! 0. In the stiff case, much insight has been gained by studying the behavior of numerical methods on well-chosen, rather simple linear and nonlinear stiff model problems. As a first step towards an understanding of the numerical energy behavior in Hamiltonian systems when the product of the step size and the highest frequency is not a small quantity, we consider in this article the nonlinear, highly oscillatory model problem x where Dept. de math'ematiques, Universit'e de Gen'eve, CH-1211 Gen'eve 24, Switzerland y Mathematisches Institut, Universit?t T-ubingen, Auf der Morgenstelle 10, D-72076 T-ubingen, Germany (E-mail: lubich@na.uni-tuebingen.de) (with blocks of arbitrary dimensions), and where the nonlinearity has a Lipschitz constant bounded independently of !. This situation arises in the celebrated Fermi-Pasta-Ulam model, for which we will present numerical experiments. Clearly, in the model problem (1.1) we take strong restrictions in that the high frequencies are confined to the linear part of the problem, and that the linear part has a single high frequency. The diagonal form of\Omega is not essential, since the numerical methods are invariant under a diagonalization of the matrix. We study the long-time energy behavior of a class of symmetric numerical methods which are used with step sizes h such that the product h! is bounded away from zero and can be arbitrarily large. The methods integrate the linear part of (1.1) exactly and reduce to the St-ormer/Verlet method for The class includes the methods of [GSS99, HoL99]. Classical symmetric methods such as the St-ormer/Verlet method, the trapezoidal rule, or Numerov's method are not considered in this article. However, using the results of the present paper, their energy behavior on (1.1) for h! in the range of linear stability is analyzed in [HaL99]. Our approach to the near-conservation of the energy is based on a frequency expansion of the solution x(t) of (1.1), which is an asymptotic series where the coefficient functions y(t) and z k (t) together with their derivatives are bounded independently of !. It turns out that the system determining the coefficient functions has two (formally exact) invariants. One of these is close to the total energy according to the partitioning of \Omega\Gamma The other invariant is close to which represents the oscillatory energy of the system. For the numerical solution we derive a similar frequency expansion which is valid on grid points Under a non-resonance assumption on h!, the equations determining the coefficient functions have a similar structure to those of the continuous problem. This allows us to obtain two almost-invariants close to H and I , and rigorous estimates for the near-conservation of the total and the oscillatory energy over time intervals of size CN h \GammaN . The only restriction on N comes from the above non-resonance condition. The analysis uses only the symmetry of the methods and does not require symplecticness. In Sect. 2 we describe the numerical methods and we present numerical experiments with the Fermi-Pasta-Ulam problem. These experiments illustrate the long-time conservation of the total and the oscillatory energy in non-resonance situations, which will later be completely explained by the theory. We also show the energy behavior of the methods near resonances. This behavior depends strongly on properties of the filter functions that determine the numerical method. We identify conditions that yield satisfactory energy conservation near resonances and the correct energy exchange between highly oscillatory components. Sect. 3 gives a complete analysis of the two-dimensional linear case of (1.1) over the whole range of non-resonant, near-resonant and exactly resonant cases. This gives already much insight into conditions determining the energy conservation in the general situation. The frequency expansion of the analytical solution of (1.1) is introduced in Sect. 4, that of the numerical solution in Sect. 5. The numerical invariants are derived in Sect. 6. The main result on the numerical long-time conservation of energy for (1.1) is formulated and proved in Sect. 7. 2. Numerical methods and numerical experiments. In this section we present the numerical methods and we illustrate the main results of this paper with the Fermi-Pasta-Ulam problem. 2.1. The discretization. We consider the differential equation (1.1), is a symmetric and positive semi-definite (not necessarily diagonal) real matrix, and we assume that initial values x 0 and - x 0 are given at t By the variation-of- constants formula, the exact solution of (1.1) satisfies t\Omega \Gamma\Omega sin t\Omega cos t\Omega s)\Omega ds (observe t\Omega is well-defined also for singular\Omega\Gamman It is therefore natural to consider, for a fixed step size h, the explicit discretization h\Omega xn \Gamma\Omega sin real functions OE(-), depending smoothly on - 2 . For method integrates the problem (1.1) without error. If the method is consistent of order 2. For fixed\Omega and for h ! 0, second-order convergence follows from classical results. In this article we are mainly interested in the situation where h\Omega can take large values. For long-time integrations, symmetric and/or symplectic methods are expected to have favorable properties. By exchanging n in (2.2)-(2.3), it is seen that the method is symmetric for all g(x) if and only if (where sinc sin -). It can be shown by direct verification that (2.2)-(2.3) is a symplectic discretization if, in addition to (2.4), also holds. This condition will not be required for the analysis of our paper. Since the nonlinearity g(x) in (1.1) does not depend on - x, we can eliminate - xn in (2.2) with the help of (2.3). In the case of a symmetric discretization we thus get the two-step recurrence The starting value x 1 is obtained from (2.2) with For the recognize the well-known St-ormer method. stiff harmonic soft nonlinear Fig. 1. Alternating soft and stiff springs Methods of the type (2.5) or (2.2)-(2.3) have been proposed and studied by several authors. Gautschi [Gau61] suggests to take (-=2). With this choice, (1.1) is integrated exactly even for const . Deuflhard [Deu79] discretizes the integral in (2.1) by the trapezoidal rule and thus arrives at (2.5) with recently, Garc'ia-Archilla, Sanz-Serna and Skeel [GSS99] introduce a function OE(-) in the argument of g, and they consider the case where the method is symplectic, so that Hochbruck and Lubich [HoL99] consider OE(-). The papers [GSS99] and [HoL99] derive error bounds on finite time intervals which are independent of ! and of the smoothness of the solution. In this article, we consider general functions OE; / with have no zeros except possibly at integral multiples of -. Since we are interested in the energy conservation of the numerical solution, we need also an approximation to the derivative if we use the two-term recurrence relation (2.5). This can be obtained by the relation (2.3) or, in the case of a symmetric method, also by the formula This is possible if h! is not a nonzero integral multiple of -. We obtain (2.6) by subtracting (2.5) from twice the formula of (2.2). For a symmetric method we obtain a formula for - by exchanging n in (2.3). Subtracting the resulting formula from (2.3), we obtain the two-step recurrence \Gamma2\Omega sin Formulas (2.5) and (2.7) give a symmetric two-step method even if (2.4) is not satis- fied. If -, then this method is exact for const . The choice / 1 been considered in [HoL99]. 2.2. Experiments with the Fermi-Pasta-Ulam problem. We consider a chain of springs, where soft nonlinear springs alternate with stiff harmonic springs (see [GGMV92] and Fig. 1). The variables x for the displacements of end-points of the springs. The movement is described by a Hamiltonian system with Using the symplectic change of variables u 2, we get a new Hamiltonian system with (a) H I I I (a) H I I I (a) H I I (c) H I 100 200 3001 (a) H I 100 200 300 I 100 200 300 (c) H I Fig. 2. Energy exchange of stiff springs where This is exactly of the form (1.1). For our numerical experiments we consider the case shown in Fig. 1) As initial values we take and zero for the remaining initial values. We apply the method (2.2)-(2.3) with the following data: (a) (c) with / 0 (-) and / 1 (-) given by (2.4). We study the total energy (2.8) and the oscillatory energy along the numerical solution on the interval 0 - t - 400. With the chosen initial values we have In Fig. 2 we have plotted, for three different step sizes and for all three methods, the numerical values for I 1 ; I 2 ; I 3 ; I and H \Gamma 0:8. We see that an exchange of energy takes place, going from the first stiff spring with energy I 1 to the second stiff spring and later to the third one. For the smallest step size we have also plotted in gray the numerical values for perturbed initial values obtained by adding 10 \Gamma8 to u 1 (0), - illustrates that the solution is very sensitive to perturbations. 6 E. HAIRER AND CH. LUBICH In all cases we see that H and I are well preserved over the whole interval, even for step sizes where the numerical solution is completely wrong. Further experiments have shown that such a preservation holds for much longer intervals (we tested up to An explanation of this phenomenon is the main objective of this paper. 2.3. Numerical experiments in near-resonant situations. When the product of the step size and the frequency h! is close to a multiple of -, then the different methods show widely different behavior. Energy is conserved only for some choices of / and OE. Satisfactory numerical behavior also in near-resonance situations is obtained if the numerical method satisfies the following additional conditions: These conditions yield long-time energy conservation for all values of h! with the exception of h! in intervals of width O(h) near integral multiples of 2-. The total energy appears to be conserved uniformly for arbitrary values of h! if The necessity of these conditions is seen from an analysis of the linear case, which is given in Sect. 3. When h! is close to 2m- with a positive integer m, the condition (2.9) requires a double zero of / at 2m-. Similarly, for h! close to an odd multiple of -, condition there requires a simple zero of /. For the choice HoL99] the condition (2.9) is obviously satisfied for all values of h!, but (2.10) is violated near odd multiples of -. For condition (2.10) is trivially satisfied for all h!, but condition (2.9) fails near even multiples of -. The choice satisfies the three conditions (2.9)- for all h!. Condition (2.12) is not satisfied by any of the methods previously proposed in the literature. Table Methods used for the numerical experiments of Sect. 2.3 A sinc (-) 1 2k- p Let us illustrate the effect of the conditions (2.9), (2.10) and (2.11) on the numerical solution when h! is close to a multiple of -. We consider the Fermi-Pasta-Ulam problem of Sect. 2.2 with the same initial values, and we apply six different methods. Their characteristics are given in Table 1. The sign p indicates that the corresponding condition on / and OE is satisfied. If a condition is not satisfied for all values of -22 (C) out of scale -22 (D) Fig. 3. Energy conservation of different methods for Fig. 4. Energy conservation of different methods for Fig. 5. Energy conservation of different methods for h!, we give the values close to which it is violated. For each of the methods (B), (C), (D) only one of the conditions (2.9)-(2.11) is not fulfilled. In Fig. 3 we show the errors of the Hamiltonian over the interval [0; 1000]. We have used the step size such that a maximal error of size 396497, because / 1 (-) given by (2.4) has a singularity at -. .2 .2 .2 I .2 I Fig. 6. Error in the total and oscillatory energies as a function of h! for the FPU problem Methods (A) and (D) show a clear drift from the constant value of the Hamiltonian. Only the methods (B), (E), and (F), for which all three conditions are satisfied close to -, conserve the Hamiltonian very well. For the pictures corresponding to these three methods, we have changed the scale so that the small oscillations become visible. Fig. 4 shows the same experiment, where this time ! is chosen such that 2:0000001 \Delta -. For both situations, we get the same qualitative behavior when we plot the oscillatory energy instead of the Hamiltonian. The results of these experiments confirm that the conditions (2.9), (2.10), (2.11) cannot be omitted if we are interested in long-time energy estimates that hold uniformly in h!. In Fig. 5 the numerical results of the nonresonant case are included. All methods give satisfactory results. The most accurate results are obtained by the method (C). In the upper pictures of Fig. 6 we plot the maximal errors in the Hamiltonian as a function of h!, and we take step sizes 0:025. The picture to the right corresponds to the method (E) of Table 1. The picture to the left is obtained with method (F) which satisfies (2.12). Uniform convergence of the error can be observed only in this case. The lower pictures of Fig. 6 show the analogue for the deviations in the oscillatory energy. Close to integral multiples of 2- this deviation is large for both methods. The same phenomenon can be observed already for linear problems (see Fig. 8), for which a complete analysis is given in Sect. 3. 2.4. Energy exchange. The energy exchange between stiff components takes place on time intervals of length O(!). In Fig. 2, this is reproduced qualitatively correctly for large h! only in the case where 1. The numerical frequency expansion of Sect. 5 shows that the condition .2 HH .2 HH .2 II .2 II Fig. 7. Same experiment as in Fig. 6 with method (2.5), with method (2.16) (right pictures) is needed for the approximation of the energy exchange between stiff components when h! is bounded away from zero (compare the equations for the z 2 component in (4.8) and (5.8)). This is a severe condition which excludes all methods considered so far with the exception of the above-mentioned method the other hand, we have seen that this method has rather poor energy conservation properties. We therefore extend the class of methods (2.5) to (h\Omega\Gamma2 For consistency, the functions / k , OE k must satisfy Conditions (2.9) and (2.10) are now needed for condition (2.13) is replaced with For example, the method with (2.6), or equivalently in one-step form h\Omega xn \Gamma\Omega sin shows the correct energy exchange for large h!, as for method (b) in Fig. 2. In Fig. 7, we compare the energy conservation for the method (2.5) with and the method for the same step sizes as in Fig. 6. For (2.16), the total and oscillatory energies are well conserved over long times except for h! in intervals of length O(h) around integral multiples of -. For ease of presentation, the following analysis will be done for methods of class (2.5), but the arguments extend in an obvious way to the class (2.14). 3. Long-time energy conservation for linear problems. We start our analysis with the case where with a two-dimensional symmetric matrix A satisfying a 11 ? 0, so that is linear. This gives already a lot of insight and illustrates the importance of the conditions (2.9)-(2.12). In this situation, the differential equation (1.1) becomes x The total energy H given by (1.5) is an invariant of the system. In the following we assume that H \Delta is bounded uniformly in !. This requires x 2 3.1. Analytical solution. The exact solution of (3.1) is given by O(! \Gamma2 )' i are the eigenvalues so that a 11 +O(! \Gamma2 For given initial values x(0), - Consequently, we have be This implies that the quantity remains O(! \Gamma1 )-close to the constant value I \Delta for all times t. 3.2. Numerical solution. We search for functions b v, such that xn := x(nh) satisfies the numerical scheme (2.5) with This implies so that has to be an eigenvalue of and v a corresponding eigenvector. In this section we use the short notation and OE(h!). Since the off-diagonal elements of (3.3) are small, the eigenvalues - are close to ff respectively. The corresponding eigenvectors are 'fl/ \Gammafl OE' the general solution can thus be written as where the complex coefficients are computed from the initial values b obtained from (2.2). Inserting To study the long-time near-conservation of H \Delta and I \Delta we distinguish two cases. Case I: well-separated eigenvalues. We assume that one of the conditions a 11 or is satisfied. This covers nearly all choices of h!. Only values in intervals of length O(h) are excluded. Theorem 3.1. Consider the numerical method (2.2)-(2.3) applied to (3.1) with a Under one of the restrictions (3.5) or (3.6) on the step size, and under the conditions (2.9), (2.10), (2.11) on the numerical method, the energies H and I of (1.5) and (1.6) along the numerical solution satisfy for all n - 0. The constants symbolized by O(\Delta) are independent of !, h and n. Proof. The characteristic polynomial of the matrix (3.3) is Its zeros can be computed explicitly. Each of the conditions (3.5) or (3.6) together with (2.11) implies that for both eigenvalues - i of (3.3). Hence, for sufficiently small h, - 1 and - 2 are real and in the interval [\Gamma1; 1]. The angles - i , defined by - are therefore also real and satisfy a We next estimate the coefficients a; a; b; b in (3.4). Since the - i and v i are real, the coefficients a; b are the complex conjugates of a; b. In both situations, (3.5) and (3.6), we have under the assumption (2.9), we further have Consequently, the condition b of using b more convenient to work with b x(h) \Gamma cos cos From the estimates sinc (h!=2)), which all follow from (2.9) and (2.10), we then get be The second relation is a consequence of the fact that (which follows from - and (2.11)). The statement (3.8) is now an immediate consequence of (3.9) and of the because the modulus of c = be i-2 t is independent of t. The near-conservation of the Hamiltonian can be seen similarly. If one of the conditions (3.5) and (3.6) is satisfied, we get flOE sin(- 2 so that This implies jbx 0 Const all t, this together with (3.8) proves the statement (3.7) for the total energy. Case II: nearly collapsing eigenvalues. We now consider the complementary case4 h 2 a so that the two eigenvalues of (3.3) are very close. Theorem 3.2. Consider the numerical method (2.2)-(2.3) applied to (3.1) with a Under the condition (2.12) on the numerical method, viz. for all n - 0 and uniformly in h! - c ? 0. The constants symbolized by O(\Delta) are independent of !, h and n. Proof. Under the condition (2.12) the numerical method satisfies (2.9), (2.10), so that Theorem 3.1 is applicable. It therefore remains to consider the situation where h! is restricted by (3.10). The condition (3.10) implies 1 2a 11 for sufficiently small h. We now assume that OE/ - 0, which is satisfied by the choice (2.12). This guarantees that the eigenvalues of (3.3) are real, that - OE/), and that a 11 +O(h 3 For the special choice (2.12) we further have jflj - 1= p O(!), so that the coefficients of (3.4) satisfy a = O(1) and (3. .2 .4 .2 .4 .2 .4 I .2 .4 I Fig. 8. Error in the total and oscillatory energies as a function of h! for a linear problem holds (recall that This is a consequence of the estimate OE/) and of the fact that jflj - 1= OE/. The relations (3.12) and (3.13) thus yield a 11 Together with (3.4) this implies To prove that the t-dependent term is small, we need the relation (2.12) between OE and /. Using also (3.12) and (3.13) we obtain a a sin This completes the proof of Theorem 3.2, because The proof above shows that in general I In the situation of Theorem 3.1 we have so that the first term in the right-hand expression of (3.14) becomes negligible, and long-time conservation of I can be concluded. If a 12 6= 0 and cos(h!) - 14 E. HAIRER AND CH. LUBICH OE/ and fl - 1= OE/, so that fl/ - /=OE and flOE - OE=/. The initial condition b so that none of the terms in (3.14) can be neglected. If - is different from an integral multiple of 2-, the expression (3.14) cannot remain close to a constant value. This result is intuitively clear, because in the situation (3.10) the two frequencies are indistinguishably close for the numerical method. The upper pictures of Fig. 8 show the maximal error in the Hamiltonian H(xn ; - in dependence of h! for the problem (3.1) with a initial values . The three curves correspond to the step sizes 0:05. The picture to the left it obtained with a method satisfying (2.12). Uniform convergence of the error can nicely be observed. The picture to the right corresponds to the method (E) of Table 1. The lower pictures of Fig. 8 plot the maximal deviation of the oscillatory energy I(xn ; - as a function of h!. It confirms the analysis above, which shows that for h! satisfying a 11 the oscillatory energy cannot be well conserved. 4. Frequency expansion of the analytical solution. The main tool of our analysis for nonlinear problems is a decomposition of the solution x(t) of (1.1) into a smooth part and into highly oscillatory terms with smoothly varying amplitudes. This decomposition is valid over finite time intervals. We show the existence of two almost-invariants for the coefficients of this decomposition, which are related to the total energy and the oscillatory energy of the system. A repeated use of these almost- invariants then allows us to prove the long-time near-conservation of the oscillatory energy. 4.1. The frequency expansion. We assume the nonlinearity g in (1.1) analytic on an open set D, and we consider solutions of (1.1) which satisfy where K is a compact subset of D. We assume further that the initial values have limited harmonic energy:2 k - where E is independent of !. Theorem 4.1. Under the assumptions (4.2) and (4.1) for 0 - t - T , the solution x(t) of Eq. (1.1) has for arbitrary N - 2 an expansion of the form e ik!t z k (t) +RN (t); where the remainder term and its derivative are bounded by The real functions and the complex functions z are bounded, together with all their derivatives, by and we have z . They are unique up to terms of size O(! \GammaN \Gamma2 ). The constants symbolized by the O-notation are independent of ! and t with on E, N , T , and on the order of the derivative). Proof. To determine the smooth functions we put e ik!t z k (t); insert this function into (1.1), expand the nonlinearity around y(t) and compare the coefficients of e ik!t . With the notation g (m) (y)z the following system of differential equations: z k z k Here the sums range over all m - 1 and all multi-indices integers ff j satisfying which have a given sum For large !, the dominating terms in these differential equations are given by the left-most expressions. However, since the central terms involve higher derivatives, we are confronted with singular perturbation problems. We are interested in smooth functions that satisfy the system up to a defect of size O(! \GammaN ). In the spirit of Euler's derivation of the Euler-Maclaurin summation formula (see e.g. [HaW96]) we remove the disturbing higher derivatives by using iteratively the differentiated equations (4.7)-(4.9). This leads to a system z k are formal series in powers of ! \Gamma1 . Since we get formal algebraic relations for , we can further eliminate these variables in the functions F . We finally obtain for y the algebraic relations z k z k and a system of real second-order differential equations for y 1 and complex first-order differential equations for z At this point we can forget the above derivation and we can take it as a motivation for the ansatz (4.10)-(4.11), which we truncate after the O(! \GammaN ) terms. Inserting this ansatz and its first and second derivatives into (4.7)-(4.9) and comparing like powers recurrence relations for the functions F k jl . This shows that these functions together with their derivatives are all bounded on compact sets. We determine initial values for (4.11) such that the function e x(t) of (4.6) satisfies e x(0). Because of the special structure of the ansatz (4.10)- (4.11), this gives a system which, by the implicit function theorem, yields (locally) unique initial values y 1 (0), (0). The assumption (4.2) implies that z 2 It further follows from the boundedness of F 2l that z 2 looking closer at the structure of the function G k jl it can be seen that it contains at least k times the factor z 2 . This implies the stated bounds for all other functions. We still have to estimate the remainder RN x(t). For this we consider the solution of (4.10)-(4.11) with initial values (4.12). By construction, these functions satisfy the system (4.7)-(4.9) up to a defect of O(! \GammaN ). This gives a defect of size O(! \GammaN ), when the function e x(t) of (4.6) is inserted into (1.1). Hence on a finite time To obtain the slightly sharper bounds (4.4), we apply the above proof with N replaced by N 2. 4.2. The Hamiltonian of the frequency expansion. Consider now the situation so that (1.1) is a Hamiltonian system x with Hamiltonian where U(x) is assumed to be analytic. Let v k that by (4.7)-(4.9) these functions satisfy Here, the sum is again over all m - 1 and all multi-indices integers ff which have a given sum and we write Further we denote From the above it follows that the vector (y; V ) satisfies the system y which, neglecting the O(! \GammaN ) terms, is Hamiltonian with Theorem 4.2. Under the assumptions (4.2) and (4.1) for The constants symbolized by O(\Delta) are independent of ! and t with depend on E, N and T . Proof. Multiplying (4.17) and (4.18) with - y T and ( - respectively, gives dt Integrating from 0 to t and using v By the bounds of Theorem 4.1, we have for On the other hand, we have from (4.14) and (4.3) that Using it follows from Inserted into (4.22) and (4.23) this yields the statement (4.21). 4.3. Another almost-invariant. Besides the Hamiltonian H(y; - V ), the coefficients of the frequency expansion have another almost-invariant. It only depends on the oscillating part and it is given by This almost-invariant turns out to be close to the energy of the harmonic oscillator, Theorem 4.3. Under the assumptions (4.2) and (4.1) for The constants symbolized by O(\Delta) are independent of ! and t with depend on E, N and T . Proof. With the vector holds that U(y; Differentiating the identity with respect to t yields z \Gammak because The proof of Theorem 4.3 is now very similar to that of Theorem 4.2. We multiply the relation (4.18) with \Gammai!k(v \Gammak ) T instead of ( - Summing up yields, with the use of (4.28), \Gammai! The derivative of I(V; - by (4.24), is d dt (v \Gammak In the sums the terms with k and \Gammak cancel. Hence, the statement (4.26) follows from (4.29) and (4.30). Using - from the bounds of Theorem 4.1 that On the other hand, using the arguments of the proof of Theorem 4.2, we have This proves the second statement of the theorem. Corollary 4.4. If x(t) 2 K for The constants symbolized by O(\Delta) are independent of ! and t with depend on E and N . Proof. With a fixed T ? 0, let V j denote the vector of frequency expansion terms that correspond to starting values (x(jT ); - x(jT )). For we have by (4.27) Vn ('T Vn ('T Vn We note that I(V j+1 (0); - by the uniqueness statement of Theorem 4.1, we have V j+1 we have the bound (4.26) of Theorem 4.3. The same argument applies to I(Vn ('T ); - Vn ('T Vn (0)). This yields the result. Remark. It is already known from the article [BGG87] that the oscillatory energy x(t)) is nearly preserved over long times. The proofs in [BGG87] are completely different. They use coordinate transforms from Hamiltonian perturbation theory and show that I is nearly preserved over time intervals which grow exponentially with !. By carefully tracing the N-dependence of the constants in the O(! \GammaN )-terms, it is possible to obtain near-conservation of I over exponentially long time intervals also within the present framework of frequency expansions. 5. Frequency expansion of the numerical solution. In this section we show that the numerical solution (2.2), (2.3) for nonlinear problems (1.1) has a frequency expansion similar to that of the analytical solution. Following the idea of backward analysis and motivated by the results of Sect. 4 we look for a function e ik!t z k (t) (with smooth y(t) and z k (t) depending on h) 1 such that, up to a small defect, We assume throughout this section that and that the numerical solution \Phix n remains in a compact subset of the region where g(x) is analytic, i.e., 5.1. Functional calculus. For the computation of the functions y(t) and z k (t) the following functional calculus is convenient. Let f be an entire complex function bounded by jf(i)j - C e fljij . Then, converges for every function x which is analytic in a disk of radius r ? flh around t. We note that (hD) k are two such entire functions, then whenever both sides exist. In particular, we have To avoid an overloaded notation with hats, we use the same letters y and z k as for the analytical solution. We hope that this does not cause confusion. We therefore introduce the operator h\Omega sin which, for h ! 0, is an approximation to h 2 (D 2 We next study the application of such an operator to functions of the form e i!t z(t). By Leibniz' rule of calculus we have (hD) k e i!t z(t). After a short calculation this also yields f(hD)e i!t 5.2. Modified equations for the coefficient functions of the frequency expansion. With the operator L(hD) of (5.5) the condition (5.2) becomes Inserting the ansatz (5.1), expanding the right-hand side of (5.7) into a Taylor series around \Phiy(t), and comparing the coefficients of e ik!t yields for the functions y(t) and z k (t) Here, multi-index as in the proof of Theorem 4.1, ff is an abbreviation for the m-tupel (\Phiz ff To get smooth functions y(t) and z k (t) which solve (5.8) up to a small defect, we look at the dominating terms in the Taylor expansions of L(hD) and L(hD ik!h). With the abbreviations 2 kh!) we have (ihD) The situation is now more complicated than in (4.7)-(4.9) for the frequency expansion of the analytical solution, because several of the coefficients in (5.9) may vanish due to numerical resonance. We here confine the discussion to the non-resonant case. We assume that h and ! \Gamma1 lie in a subregion of the small parameters for which there exists a positive constant c such that 2: LONG-TIME ENERGY CONSERVATION 21 The condition excludes that h! is o( close to integral multiples of -. For given h and !, the condition imposes a restriction on N . In the following, N is a fixed integer such that (5.10) holds. Theorem 5.1. Under the limited-energy condition (4.2), under the non-resonance condition (5.10), under the conditions (5.3), (5.4), and under the conditions (2.9) and (2.10) on the numerical method (2.2)-(2.4), the numerical solution is of the form uniformly for where the functions satisfy (5.8) up to a defect of O(h N+2 ) in their first components, and O(/(h!)h N+2 ) in their second components. Together with all their derivatives these functions are bounded by , and the constants symbolized by the O-notation are independent of ! and h, but depend on E, N , and T . Proof. Under assumption (5.10), the first non-vanishing coefficients in (5.9) are the dominant ones, and the derivation of the defining relations for y and z k is the same as for the analytical solution in Theorem 4.1. We insert (5.9) into (5.8) and we eliminate recursively the higher derivatives. This motivates the following ansatz for the computation of the functions y and z k : z k z k where the functions depend smoothly on the variables y 1 , - and on the bounded parameters /(h!). Inserting this ansatz and its derivatives into (5.8) and comparing like powers of h yields recurrence relations for the functions jl . The functions g k jl (for k - 1) contain at least k times the factor OE(h!)z 2 , and f 2l contains at least once this factor. Since the series in (5.12) need not converge, we truncate them after the ( We next determine the initial values y 1 (0), - x(h) of (5.1) coincide with the starting values x 0 and x 1 of the numerical scheme (x 1 is computed from x 0 and - x 0 via the formula (2.2) with Using the non-resonance 22 E. HAIRER AND CH. LUBICH assumption (5.10), the condition b The formula for the first component of (2.2), x with b implies that For the second component we have x x (2.2), and b x 2 \Delta , which after division by h sinc h! yields O The four equations (5.13), (5.14), (5.15) constitute a nonlinear system for the four quantities y 1 (0), - z . By the implicit function theorem and using the limited-energy assumption (4.2), we get a locally unique solution for sufficiently small h, if the conditions (2.9) and (2.10) are satisfied. The initial value for z 2 satisfies z 2 it follows from (2.10) that by (5.12). This implies z 2 for . The other estimates (5.11) are directly obtained from (5.12). Conse- quently, the values b x(nh) inserted into the numerical scheme (2.5) yield a defect of size O(h N+2 Standard convergence estimates then show that on bounded time intervals is of size O(t 2 h N ) in the first component and of size O(/(h!)t 2 h N ) in the second component. This completes the proof of Theorem 5.1. 5.3. Frequency expansion of the derivative approximation. Under the condition (5.10) we have h! 6= k- for integer k, so that the derivative approximation xn is given by (2.6). We now define b x 0 (t) by the continuous analogue Using condition (2.10), Theorem 5.1 implies that on bounded time intervals. We next write the function b x 0 (t) as e ik!t z 0k (t): Inserting the relation (5.1) into \Gammai which is equivalent to (5.17), and comparing the coefficients of e ik!t we obtain sinc (ihD) - In particular, we get for z 1 2 that z 01 cos !h sin !h Theorem 5.2. Under the assumptions of Theorem 5.1, the numerical solution xn , given by (2.6), satisfies uniformly for where the functions y together with all their derivatives are bounded by z 01 The constants symbolized by the O-notation are independent of ! and h, but depend on E, N and T . Proof. The estimates follow from (5.19) and from Theorem 5.1. For y 0 1 and z 01 2 we use the formulas (5.12) to get the sharper result. 5.4. Energy along the numerical solution. In the Hamiltonian case \GammarU (y), the total energy H and the oscillatory energy I are related to the frequency expansion coefficients as follows. Lemma 5.3. If the coefficients of the frequency expansions for xn and - xn satisfy (5.11) and (5.21), respectively, then Proof. By definition (4.25) we have I(bx; b e i!t z 1 e i!t z 1 the statement (5.23) follows from the fact that jv . The formula (5.22) can be proved in the same way. 6. Almost-invariants of the numerical frequency expansion. In this section we show that, in the Hamiltonian case the coefficients of the frequency expansion of the numerical solution have invariants that can be obtained as in Sect. 4. We denote z k are the coefficients of the frequency expansion (5.1). Similar to (4.16) we consider the function where the sum is taken over all m - 1 and all multi-indices non-vanishing integral components for which It then follows from Theorem 5.1 that the coefficients y and v k satisfy r y The factor OE(h!) in the defect of (6.3) is due to the presence of the factor OE(h!)z 2 in the relations (5.12) defining the z-components. The similarity of these relations to (4.17), (4.18) allows us to obtain invariants that are the analogues of H and I of Sect. 4. 6.1. First invariant. As in Sect. 4.2 we multiply (6.2) and (6.3) by - y T and respectively, and we thus obtain dt Since we know bounds on z k and on its derivatives (Theorem 5.1), we switch to the quantities z k and we get the equivalent relation dt We shall show that the left-hand side is the total derivative of an expression that only depends on y, z k and its derivatives. Indeed, the term - y T y (2l) can be written as dt Similarly, we get for z \Gammak that Re - z T z dt z T z Re z T z dt z T z z T z dt z T z Im z T z dt z T z z T Hence, there exists a function b which depends on the values at t of the functions y and of their first N derivatives, such that (6.4) reads d dt This yields immediately the first statement of the following result. Theorem 6.1. Under the assumptions of Theorem 5.1, the coefficient functions y and for Proof. The formula for b H 0 is obtained from the formulas (5.9) for L(hD together with the estimates of Theorem 5.1. Remark. Symplectic discretizations have 6.2. Second invariant. As in the proof of Theorem 4.3 we have for the function of (6.1) that Consequently, it follows from (6.3) that \Gammai! Written in the z variables, this becomes \Gammai! As in Sect. 6.1, the left-hand expression can be written as the total derivative of a function b I 0 [Z](t) which depends on the values at t of the function Z and its first N derivatives: d dt I Theorem 6.2. Under the assumptions of Theorem 5.1, the coefficient functions I I for with -(h!) as in Theorem 6.1. Proof. From (6.5) and the estimates of Theorem 5.1, we obtain Because of condition (2.10), this yields the stated formula for b I 0 . 7. Long-time energy conservation of the numerical discretization. We are now able to prove the main result of this paper. This shows that the total energy H and the oscillatory energy I are nearly conserved over time intervals of length CN h \GammaN , for any N for which the non-resonance condition (5.10) is satisfied. For the convenience of the reader we restate our assumptions: ffl the limited-energy condition 26 E. HAIRER AND CH. LUBICH ffl the boundedness condition (5.4) for the numerical solution sequence: \Phix n stays in a compact subset of the domain of analyticity of g; ffl the condition (5.3): h! - d ? 0; ffl the conditions (2.9) and (2.10) on the numerical method: ffl the non-resonance condition (5.10): for some N - 2, Theorem 7.1. Under the above conditions, the numerical solution of (1.1) obtained by the method (2.2)-(2.4) satisfies nh. The constants symbolized by O(\Delta) are independent of n and of h and ! satisfying the above conditions, but depend on N . Proof. (a) If we consider the linear combinations b I 0 and I 0 =-, it follows from Theorem 6.1 and Theorem 6.2 that Moreover, by Theorem 6.1 and Theorem 6.2 together with Lemma 5.3 we have I where again b x(t) is defined by the frequency expansion (5.1) with coefficients y(t) and defined by (5.17). The relations (7.1) and (7.2) hold only on finite time on which the frequency expansion is defined. (b) We now apply the above relations repeatedly on intervals of length h, for frequency expansions corresponding to different starting values. As long as satisfies the limited-energy condition (4.2) (possibly with a larger constant E), Theorem 5.1 gives us frequency expansion coefficients yn (t); Zn (t) corresponding to starting values Because of the uniqueness (up to O(h N+1 )) of the coefficients of the frequency expansion, the following diagram commutes up to terms of size O(h N+1 numerical method (up to O(h N+1 The construction of the coefficient functions via (5.12) shows that also higher derivatives of (y n ; Zn ) at h and (y n+1 ; Zn+1 ) at 0 differ by only O(h N+1 ). We thus have from (7.3) and (7.1) Using this relation repeatedly, we obtain Moreover, from (7.2) we have the following for the coefficient functions corresponding to the starting values xn ) and by construction, and b x 0 Theorem 5.2, we obtain which gives the desired bound for the deviation of the total energy along the numerical solution. The same argument applies to I(xn ; - Acknowledgment . We are grateful to Sebastian Reich for drawing our attention to the Fermi-Pasta-Ulam problem. --R Realization of holonomic constraints and freezing of high frequency degrees of freedom in the light of classical perturbation theory. On the Hamiltonian interpolation of near to the identity symplectic mappings with application to symplectic integration algorithms A study of extrapolation methods based on multistep schemes without parasitic solutions On the problem of energy equipartition for large systems of the Fermi-Pasta-Ulam type: analytical and numerical estimates Numerical integration of ordinary differential equations based on trigonometric polynomials The life-span of backward error analysis for numerical integrators Energy conservation by St-ormer-type numerical integrators Analysis by Its History A Gautschi-type method for oscillatory second-order differential equations Dynamical systems --TR --CTR M. Van Daele , G. Vanden Berghe, Geometric numerical integration by means of exponentially-fitted methods, Applied Numerical Mathematics, v.57 n.4, p.415-435, April, 2007 Ernst Hairer, Important Aspects of Geometric Numerical Integration, Journal of Scientific Computing, v.25 n.1, p.67-81, October 2005 J. M. Franco, New methods for oscillatory systems based on ARKN methods, Applied Numerical Mathematics, v.56 n.8, p.1040-1053, August 2006 Paul J. Atzberger , Peter R. Kramer , Charles S. Peskin, A stochastic immersed boundary method for fluid-structure dynamics at microscopic length scales, Journal of Computational Physics, v.224 n.2, p.1255-1292, June, 2007
Fermi-Pasta-Ulam problem;frequency expansion;backward error analysis;second-order symmetric methods;oscillatory differential equations;long-time energy conservation
588460
Iterative Substructuring Preconditioners for Mortar Element Methods in Two Dimensions.
The mortar methods are based on domain decomposition and they allow for the coupling of different variational approximations in different subdomains. The resulting methods are nonconforming but still yield optimal approximations. In this paper, we will discuss iterative substructuring algorithms for the algebraic systems arising from the discretization of symmetric, second-order, elliptic equations in two dimensions. Both spectral and finite element methods, for geometrically conforming as well as nonconforming domain decompositions, are studied. In each case, we obtain a polylogarithmic bound on the condition number of the preconditioned matrix.
Introduction . Since the late nineteen eighties, interest has developed in non-overlapping domain decomposition methods coupling different variational approximations in different subdomains. The mortar element methods, see [10], have been designed for this purpose and they allow us to combine different discretizations in an optimal way. Optimality means that the error is bounded by the sum of the subregion-by- subregion approximation errors without any constraints on the choice of the different discretizations. One can, for example, couple spectral methods of different polynomial degrees, or spectral methods with finite elements, or different finite element methods with different meshes. Also, the domain partitioning need not be geometrically con- forming, i.e. the intersection of the closures of two neighboring subdomains may only be parts of certain edges of these subdomains. The basic ideas of the mortar method can be outlined as follows: the skeleton of the decomposition (i.e. the union of the subdomains interfaces) is itself partitioned into mortars. Each mortar is an entire edge of one of the subdomains; the mortars are disjoint open sets. The chosen local discretizations may force the method to be nonconforming and we only impose a type of weak continuity. For each and for each nonmortar side \Gamma j k of @\Omega k , we introduce a carefully chosen discrete space ~ kh of functions supported on \Gamma j k . Weak continuity, in this context, then means that the k )\Gammaprojection of the jump across \Gamma j k into the space ~ kh vanishes. In the first version of the mortar method, strong continuity constraints were also imposed at the vertices of the subdomains but this turned out not to be necessary. A second version of the mortar method, developed and analyzed by Ben Belgacem and Maday [6],[7], does not require such constraints. In particular for problems in three dimensions, the second version offers important advantages over the first and in what follows, we shall exclusively work with this more recently developed method. We note that, in a finite element context, similar nonconforming methods have been studied by Le Tallec et al INSA Rennes, 20 Av des Buttes de Coesmes, 35043 Rennes, France and CMAP, Ecole Polytechnique 91128 Palaiseau, cedex France. Electronic mail address: achdou@cmapx.polytechnique.fr y Laboratoire ASCI, B"atiment 506, Universit'e Paris Sud, 91405 Orsay and Universit'e Paris 6, Paris, France. Electronic mail address: maday@ann.jussieu.fr z Courant Institute of Mathematical Sciences, 251 Mercer St, New York, NY 10012. Electronic mail address: widlund@cs.nyu.edu. URL: http://cs.nyu.edu/cs/faculty/widlund/index.html. This work was supported in part by the CNRS, while this author was visiting Universit'e Paris 6, in part by the National Science Foundationunder Grant NSF-CCR-9503408, and in part by the U. S. Department of Energy under contract DE-FG02-92ER25127. . Mortar element methods offer many advantages: ffl They increase the portability of spectral methods. ffl In the context of finite elements, they provide flexibility in the construction of the mesh. For example, they may be used in some cases to avoid updating the finite element mesh (sliding meshes [5]) or, on the contrary, to simplify the adaption of the meshes ([9]). ffl They are well suited for parallel computing. There has already been several implementations of the mortar methods, among them [5] with sliding meshes, [23] for spectral element methods, [18] for a nonconforming finite element method for elasticity problems, and [4] for the Navier Stokes equation. In the present paper, we propose algorithms for solving the algebraic linear systems arising from the mortar methods. After the elimination, in parallel, of the degrees of freedom internal to the subdomains, there remains to find the traces of the solution on the subdomain boundaries, i.e. to solve the Schur complement system. In our methods, we work only with the true unknowns of the Schur complement systems, i.e. the unknowns associated with the mortars and the vertices of the subdomains. The method presented here can be viewed as a generalization of an iterative substructuring algorithm first introduced by Bramble, Pasciak, and Schatz [12], for two-dimensional conforming discretizations and which was reinterpreted in terms of block-Jacobi methods in [14]. The algorithm consists essentially of decomposing suitably the discrete space into a direct sum of subspaces in such a way that the related block-Jacobi preconditioned conjugate gradient method has a satisfactory rate of convergence. Each mortar can be associated in a natural way with a subspace but, in addition, a global coarse space must be included to deal with the low frequency error. We obtain a polylogarithmic bound, in terms of the local number of unknowns, on the number of iterations required for a given accuracy. Therefore, the proposed algorithm can be considered as almost perfectly scalable. Other algorithms have also been proposed. A Neumann-Neumann preconditioner is studied and tested in [18]. In [2], a saddle point formulation of the system, as in [6], is considered, and iterative methods based on a certain class of preconditioners is suggested. A saddle point algorithm for which the internal degrees of freedom need not be eliminated is proposed in [16]. In [13], a method based on a hierarchical basis representation, cf. [24], is developed and tested for low order mortar finite elements and geometrically conforming decompositions of the regions. In three dimensions, the preconditioner of this paper is not satisfactory, and another iterative substructuring method has been proposed; see [22]. In addition, an extension of the theory for two-level Schwarz algorithms, using overlapping subregions, has been completed for the mortar finite element case; see [26]. The paper is organized as follows. In Section 2, a brief review is given of the mortar finite element method in the geometrically conforming case. An iterative substructuring preconditioner for that case is proposed and studied in Section 3. The geometrically nonconforming mortar finite element method is discussed in Section 4. Finally, a generalization to the spectral element method in the geometrically conforming and geometrically nonconforming cases is carried out in Section 5. 2. Mortar Element Methods in the Geometrically Conforming Case. Let\Omega be a bounded polygonal domain of IR 2 , and let k=1 be a partition of\Omega into K non-overlapping open quadrilaterals: We make this restriction to polygonal domains and subdomains only to simplify the presentation. The domain decomposition is called geometrically conforming if the intersection of the closure of two subdomains is either empty, a vertex, or an entire common edge of the two subdomains. For any 1 k 6= ' K, let \Gamma k' be the closed straight segment, possibly degenerate, given by \Gamma k' us also introduce V as the set of crosspoints of the domain decomposition which are not on @ \Omega\Gamma and the skeleton, defined by We assume that the subdomains have uniformly bounded aspect ratios but there is no need to assume that the subdomains form a quasiuniform coarse triangulation. All what follows concerns the Dirichlet problem for Poisson's equation (1) but our results hold for any self-adjoint, elliptic, second order operator. Families of finite element triangulations T k;h are associated with K, which we assume satisfy the classical shape regularity assumption on the elements. We denote by h k the maximum diameter of the elements of T k;h . To simplify our analysis, we also assume that the meshes are quasiuniform for each recall that quasiuniformity for a triangular mesh means that there exist two positive constants and oe such that for all triangles T of T k;h , h k hT oeae T . Here hT is the diameter of T , and ae T the diameter of the circle inscribed in T . Let X kh be the related space of piecewise linear continuous finite element functions which vanish on @ Denoting by T r k the trace operator The product spaces X h and X h are defined by: Y Y If j\Gamma k' j 6= 0, we also introduce X k;';h by Clearly, X k;';h is a subspace of the space X k;';h of the piecewise linear continuous functions on the corresponding mesh of \Gamma k' (X ;). The dimension of denoted by N k' the number of nodes of T k;h on \Gamma k' . We note that the meshes need not match at the interface between two subdomains. Thus, in order to discretize the space H 1 0(\Omega\Gamma8 we have to introduce, for each 1 k ! space ~ W k;';h of Lagrange multipliers used to impose a weak continuity constraint across \Gamma k' . A choice has to be made since this space of Lagrange multipliers can be associated with either X k;';h or One strategy is always to choose the one of largest dimension, but we emphasize that any other choice can also be supported by existing theory, and that the same asymptotical error bound results in all cases. In the case where the Lagrange multiplier space ~ W k;';h is based on be the shape functions of X k;';h associated with the nodes of T k;h on \Gamma k' , with OE 0 and OE Nk'+1 associated with the endpoints of \Gamma k' . Then, ~ W k;';h is Fig. 1. Shape functions of the spaces chosen as the space spanned by (OE it is a subspace of X k;';h of codimension two. Figure 1 illustrates the construction of ~ from It is now possible to define the subspace Y h of R oe and the subspace Y h of ae Z (v oe Consider an edge Assuming that the Lagrange multiplier space ~ is built from the mesh T kh , then the nodes of T are called slave and master nodes, respectively, because the value of v h 2 Y h at any slave node is completely determined by the values at the master nodes and crosspoints. Assuming that j\Gamma k' j ? 0, then the edge is said to be a slave, or nonmortar, and master, or mortar, edge respectively, if the space ~ W k;';h is based on the mesh T kh and T 'h , respectively. We denote by NV the number of degrees of freedom at the crosspoints of the domain decomposition, and by Nm and N s the number of master and slave nodes, respectively. Then, the dimension of X h and Y h are NV and NV +Nm , respectively. IR; be the bilinear form: Z\Omega The discretized problem corresponding to (1) is: Find u h 2 Y h such that Z\Omega It is natural to introduce two subspaces which are orthogonal in the sense of this energy inner product. The first, X ffi consists of functions which vanish on the interfaces, i.e. X ffi 0g. The other is the subspace of the discrete harmonic extensions ~ i.e. the unique solution ~ u h of a(~ ~ (2) In order to reduce the size of the problem, it is possible to solve, in parallel, a discrete Dirichlet problem for each subdomain, i.e. to find u h such that Z\Omega Defining the bilinear form s corresponding to a discrete Poincar'e- Steklov operator there remains to find u h 2 Y h such that The solution of (2) is then given by u The goal of the next section is to find a basis of Y h for which a block diagonal preconditioner for S h yields a condition number almost independent of the mesh parameters. 3. Preconditioners for the Geometrically Conforming Mortar Element Method. In the following c and C will denote positive constants uniformly bounded away from 0 and 1, respectively. They are, in particular, independent of the H k and the diameters of the subdomain\Omega k and its elements, and in the spectral case, of the degree of the polynomials. 3.1. Decomposition of the space Y h . The purpose of this section is to decompose the vector space Y h into the direct sum of a coarse space YH of dimension NV (the number of degrees of freedom associated with the crosspoints) and of a fine space Y H h of dimension Nm (the number of master nodes): Here A few notations will be needed in order to specify the coarse space YH . Let A be a crosspoint and let KA denote the set It is clear that cardinal (KA ). For each crosspoint A, and for each k 2 KA , we define a basis vector e A;k 2 Y h , such that, 1. e A;k 2. for all vertices B 6= A 3. for all ' 6= k, and for all vertices B 4. e A;k is linear on master edges. For a given crosspoint A 2 clearly vanishes on all edges except those which have A as an endpoint. Let A and B be the endpoints of In the case where \Gamma k' is a master side of @\Omega k , the restriction of e A;k k to \Gamma k' is the linear function ~ e k;';A with the value 1 at A and 0 at B. The restriction of e A;k l to \Gamma k' is the unique function ~ ';k;h such that R If conversely, \Gamma k' is a slave side of @\Omega k , the restriction of e A;k l to \Gamma k' is 0, while the restriction of e A;k k to \Gamma k' is the unique function ~ e k;';A in X k;';h such that, ~ Z ~ e k;';A The coarse space YH is defined by It is clear that the dimension of YH is NV . In what follows, it will be necessary to have accurate estimates of certain Sobolev norms of the basis vectors of the coarse space YH . Lemma 1. Let A be a crosspoint and k; ' 2 KA , with j\Gamma k' j ? 0. Assume that is a slave side of\Omega k and let ~ e k;';A be the unique function in X k;';h defined by (6). Then, Proof. Let ~ E be the vector of coordinates of ~ e k;';A in the previously described basis of shape functions of X k;';h associated to the nodes of T k;h which lie on . Using the boundary values for ~ e k;';A , we find that the vector ~ satisfies: ~ ~ l 13 l l N kl \Gamma23 (l N kl \Gamma2 l N kl \Gamma13 l Here l i is the length of the i-th mesh interval of \Gamma k' . Since the mesh T kh is quasiuniform, ~ B is spectrally equivalent to the diagonal matrix D j h k I and therefore the Euclidean norm of the vector ~ E is of order 1, since the Euclidean norm of F is of order h k . Therefore, (8) is proved. The next inequality, (9), now follows by using a well known inverse inequality for quasiuniform meshes. Finally, (10) is obtained from (8) and and the Gagliardo-Nirenberg interpolation inequality. Remark 1. In the same way, we can also prove the same estimates for the function ~ e ';k;A \Gamma ~ e k;';A when the side is a master side of\Omega k . Remark 2. From (10) and Remark 1, it follows immediately that 8A 2 V; 8k 2 KA , Denoting by j:j 1=2; the product semi-norm on useful to have bounds of je A;k j 1=2; for any A 2 V and k 2 KA . Three cases can be 1. Both sides adjacent to A are slave sides. 2. Both sides adjacent to A are master sides. 3. One side adjacent to A is a slave side, the other a master side. In the third case, we have the following result. Lemma 2. Let A be a crosspoint and let k; Assume that is a slave side of\Omega k and that master side of\Omega k . Then, je A;k Proof. From the quasiuniformity of the mesh T kh , there exists a constant C such that for all ffl 2 (0; 1=2), je A;k Ch \Gamma2ffl Let f k;m;A and f k;';A be the functions on @\Omega k , which coincide with e A;k respectively, and with 0 on @\Omega k n\Gamma km and and respectively. It is then clear that at all mesh points which are not vertices, e A;k The semi-norm jf k;m;A j 2 can be computed explicitly, because f k;m;A is piecewise linear, and the following bound is obtained: It now follows from (8) and an inverse inequality that Ch 2ffl Choosing combining (13) and (14), we obtain the desired result by using (12). The next lemma is proved in the same way as Lemma 2. Lemma 3. Let A be a crosspoint and let k; ' 2 KA , with j\Gamma k' j ? 0. Assume that is a master side of\Omega k . Then, je A;k It is also possible to prove the following result for the first and second cases: Lemma 4. Let A be a crosspoint and let k; Assume that the sides are either both master or both slave sides of\Omega k . Then, je A;k C: Proof. The result is very easy when both sides are master sides, because e A;k is then continuous and piecewise linear on @\Omega k . When both sides are slave sides, it follows from Lemma 1 that and the proof is completed by using an inverse inequality. Lemmas 2-4 can be summarized in the following corollary: Corollary 1. Let A be a crosspoint and let k 2 KA . Then, 3.2. A block-Jacobi preconditioner. Let ~ S be the matrix of S h in the new basis described above. The matrix ~ S can be written as ~ ShH ~ ~ In order to design a preconditioner for ~ S, we replace the block ~ ShH by 0, and the block ~ S hh by its block diagonal part with one block for each mortar. The resulting preconditioner is " S with In this section, we will develop bounds for the condition number of the preconditioned S. For that purpose, the following well known result will prove useful: There exist two constants c and C such that see, e.g., [11], in particular the discussion of an extension theorem for finite element spaces. Let " s h be the bilinear form corresponding to the matrix " S. The following lemma gives an upper bound for the eigenvalues of " Lemma 5. There exists a positive constant C such that Proof. Consider an element v h 2 Y h . There then exists a unique pair (v H h \Theta YH such that Obviously, Observing that for any x 2 \Gamma, there is a uniform bound on the number of subspaces with elements which do not all vanish at x, we deduce that there exists a constant C such that In addition, To find a lower bound for the eigenvalues of " S, the following lemma is needed: Lemma 6. There exists a constant C such that is the coarse space component of v h . Proof. Consider a vector v h 2 Y h , and let (v H h \Theta YH be given as in (17). Then, Consider specifically the subdomain\Omega ' and denote by fV i g 1iNV;' the vertices of and by 1iNV;' the subdomains adjacent to\Omega ' . We choose a numbering such that joins the crosspoints V i and V i+1 . It is clear that vH Denote by w'H the continuous piecewise linear function on @\Omega ' which interpolates v 'h at the V i . We can then write v'H as \Gamma k' is a slave side of (v kh Here ~ e ';k;V i is defined by (5). Proceeding exactly as in [12], we can prove that In addition, since v h denotes the mean value of v kh over the edge \Gamma k' . Therefore, k' is a slave side of But, see, e.g., [15], and In addition, from Lemma 3, which gives the desired result since\Omega ' has a uniformly bounded number of neighbors. We can now prove a lower bound for the eigenvalues of " Theorem 1. There exists a constant C such that Proof. Consider a vector v h 2 Y h , and let (v H h \Theta YH be given by (17). It is clear that C We now focus on the term jv . Using exactly the same arguments as in [12],[15], it is possible to bound this expression by which completes the proof of the theorem. To make our paper more self contained, we will outline a proof of this result. Assume that \Gamma k' is the segment (0; H). By the definition of the H 1=2 x dx: Clearly, and it is possible to use Lemma 6. Since the last two terms of (21) are very similar, we concentrate on the first. As in [15], this integral is split into two, over (0; h k ), respectively. It is easily seen that x and that Z hkjv kh x From (11), it follows that Thus, from (22),(23), and (24), We now use the following very important property of the projection into the coarse space: the component v kH depends only on v kh and v lh , and, for any c 2 IR, c is associated through this mapping to v kh Recalling that , and choosing c =! v kh , we find, ck 2 ck 2 as in (20). We can now obtain a bound on the condition number of " Theorem 2. There exists a constant C such that Remark 3. In order to design a convenient and inexpensive preconditioner, we should replace the blocks of " S hh in a suitable way. The preconditioners defined above can be simplified in two ways: first the fine space blocks can be replaced by more convenient matrices by using for instance hierarchical bases as described in [24] and [13]. Another possible simplification is crucial for parallelism: it makes sense to replace the block " S hh of the preconditioner, related to the fine space, by a matrix corresponding to a bilinear form s h \Theta Y H constructed as follows: Each h is mapped to v h 2 X h given by on the mortar sides, on the nonmortar sides, For the resulting preconditioner, it is easy to prove, by using the stability result of Ben Belgacem [6], Lemma 1, that the condition number estimate (25) remains valid in the geometrically conforming case. A full discussion will be given, in Subsection 4.3, of the geometrically nonconforming case. 4. Preconditioners for the Geometrically Nonconforming Mortar Element Methods. 4.1. The geometrically nonconforming mortar element method. In this section, we turn to the mortar element method in the case when the decomposition is no longer geometrically conforming. We will still assume that the aspect ratios of the subdomains are bounded by a positive constant, and we recall that H k is the diameter of the subdomain\Omega k . We also assume that there exists a constant c such that if Before formulating the discrete problem, we will adapt some of our previous notations and introduce some new ones. For denote the edges of @\Omega k . Among the set of all edges we select a family of mortars ffl m g 1mM , satisfying the following three conditions: 1. [ 2. 8(m; n) 3. 8m 2 there exists k(m); j(m) such that k(m) . Denoting by X j kh the vector space of the traces on \Gamma j k of the functions of X kh , we introduce the vector space W h Y As in the geometrically conforming case, let us introduce the space kh of the piecewise linear continuous functions on the corresponding mesh of \Gamma j k . Then ~ kh denotes the subspace of kh of the functions which are constant in the two end segments of T kh "\Gamma j k . The nonconforming approximation of H 1 0(\Omega\Gamma is given by the space if 9m such that (k; else R We can also introduce the trace space Y h As in Section 2, the edge k is called a mortar or master side of\Omega k if there exists Mg such that (k; and a nonmortar or slave side otherwise. Again, the unknowns interior to each subdomain can be eliminated by solving, in parallel, one Dirichlet problem for each subdomain, and we are led to the problem of solving (4). As in the previous section, the goal is to find a basis of Y h for which a block-Jacobi preconditioner leads to condition numbers which are almost independent of the size of the subdomains and elements. As in Section 3, the preconditioner will consist of a coarse space block and a block for each mortar. 4.2. Decomposition of the vector space Y h . As in Subsection 3.1, we decompose the vector space Y h into the direct sum of a coarse space YH of dimension NV (the number of degrees of freedom associated with the crosspoints) and a fine space h of dimension Nm (the number of master nodes). Thus, where A basis of YH is defined as follows. For each vertex A, and for each k 2 KA , the basis vector e A;k 2 YH is fully determined by the following four conditions: 1. e A;k 2. for all vertices B 6= A 3. for all ' 6= k, for all vertices B 4. e A;k is linear on the master edges. As in Subsection 3.1, the coarse space YH is defined by Consider first a vertex A k is a slave side of\Omega k and let B be the other end point of k . Then, e A;k e A;k R e A;k Exactly as in Lemma 1, we can prove that C Cp C: Assume now that \Gamma j k is a master side Let ~ e A;k be the trivial extension of e A;k k . Just as in Lemma 1, and Remark 1, we can prove that To give a flavor of the proof, let us consider the case depicted in Figure 2: Let C 0 and C 1 be the endpoints of \Gamma i ' and let D 0 and D 1 be the endpoints of the mesh segment containing the crosspoint A. We introduce the continuous function ~ e A;k;' defined on \Gamma i ' , which is piecewise linear on the mesh of \Gamma i ' , and satisfies ~ k on (D linear on (D l Fig. 2. In turn, ~ e A;k;' is split into the sum of two piecewise linear functions ~ e A;k;' 1 and ~ e A;k;'such that ~ e A;k;' e A;k;' where ~ e A;k;' 1 linear on (D Let h be the L 2 -projection onto X i 'h . It is clear that e A;k Therefore, It is clear that k~e A;k ' and from the L 2 stability of h , the first term of the right hand side of the inequality above is bounded by C h ' . Then, an argument as in Remark 1 yields the same bound for the second term k(I \Gamma h )~e A;k;' From the observation (30), it is possible to prove the following lemma in the same way as Lemma 3. Lemma 7. Let A be a crosspoint and let \Gamma j k be a master side of\Omega k with an end point A. Let ' 6= k, and let je A;k As in Subsection 3.2, let ~ S be the matrix of S h in the new basis described above. Again, the matrix ~ S can be described by formula (15) and it is possible to define a block diagonal preconditioner " S by (16). The bilinear form related to " S is called An upper bound for the eigenvalues of " S is given by a counterpart of Lemma 5, which is proved exactly as for the geometrically conforming case. To find a lower bound, we have to prove an analogue of Lemma Lemma 8. There exists a positive constant C such that where v H is the projection of v h on YH along Y H h . Proof. Consider a vector v h 2 Y h , and let (v H h \Theta YH be given by (17). One can check that We focus on one subdomain denoted wH be the continuous piecewise linear function on @\Omega 0 taking the same values as v 0h at the vertices of @\Omega 0 . v 0H can be rewritten as 0 is a slave side of ~ where the functions ~ H will be specified below. Let us focus on a single slave side \Gamma j 0 of which we call convenience. The related space of Lagrange multipliers ~ 0h is denoted ~ W flh . Let be the subdomains such that jfl " in such a way that and\Omega k+1 are adjacent. In the rest of the proof, we assume that k(fl) ? 1, but the results also hold for slight modification, which will not be discussed here. Also denote by the side of adjacent to fl, as shown in Figure 3: Fig. 3. The side fl of @\Omega 0 and adjacent subdomains The length of fl and fl k are called d fl and d fl k , respectively. A calculation shows that ~ where and where the function wH has been extended linearly outside fl. Let us further focus on the term fi k e Ck ;k 0 can be estimated in the same way. For can be rewritten as (v kh (C k For simplicity only, we assume that for 0 k k(fl), the intersection of fl and fl k contains the support of at least one basis function of the Lagrange multiplier space related to fl; this hypothesis can be eliminated by decomposing the quantities v kh (C k )\Gamma it is possible to choose a nonnegative h 2 ~ W flh supported in fl k , such that R R R R R R Therefore, as in the proof of Lemma 6, Additionally, because of (28), we can prove the following estimate: je Ck ;k which is slightly stronger than (31). For k(fl), the ratio d(A;Ck ) is smaller than one and the estimate follows from (32) and (33). For the situation is somewhat more difficult because d(A;Ck ) may be large; however, in this case je C k(fl) ;k(fl) is bounded by C h0 )). Therefore, we find that Exactly as for the geometrically conforming case, it is now possible to prove the following result: Theorem 3. There exists a constant C such that 4.3. The fine space block " S hh . To simplify the implementation, it makes sense to replace the block " S hh of the preconditioner, related to the fine space, by a matrix corresponding to a bilinear form s h \Theta Y H constructed as follows: Each h is mapped to v h 2 X h given by on the mortar sides, on the nonmortar sides, and The related preconditioner is called S, and it is easy to prove that To obtain a condition number estimate, we first state the following lemma. Lemma 9. Let\Omega 0 be a subdomain and let fl be a nonmortar side of @\Omega 0 . Let 1kk(fl) be the mortars adjacent to fl; here fl k is a whole side of @\Omega k . Then, there exists a constant C such that 8v h 2 Y H Proof. Let the function ~ v h be defined on [ 1kk(fl) fl k by the restrictions of v kh to Let be the mortar projection which maps ~ v h to v 0h j fl . As in the proof of Lemma 8, we set w be the piecewise linear continuous function which interpolates ~ v h at (C k ) 0kk(fl) , A and B. Of course, w vanishes at all the C k and thus depends linearly only on the two parameters v 1h (A) and v k(fl)h (B). When necessary, we shall use the notation w(\Delta; \Delta). It is clear that Let us first bound w. Using the same arguments as in Lemma 7, we find that the square of the H 1=2(fl)-norm of w(0; 1) and w(1; are bounded by C(1 Additionally, Therefore, There remains to estimate (~v h \Gamma w). From the stability of the operator , given in Lemma 1 of [6], we have C The right hand side of (36) can be bounded using the same argument as in the proof of Theorem 1, and we obtain The proof of the lemma follows by using (35), (36), and (37). The following corollary is an analogue of Lemma 5: Corollary 4.1. There exists a positive constant C such that The next result follows from Corollary 4.1, the analogue of Theorem 1 for the geometrically nonconforming case, and (34): Theorem 4. There exists a constant C such that 5. Preconditioners for the Mortar Spectral Method. 5.1. The geometrically conforming case. For simplicity only, we shall now assume that the subdomains\Omega k are rectangles with sides parallel to the coordinate axes. We first give a brief review of the mortar spectral method; see [7] for more details. For a review on the analysis of the Legendre spectral methods, see, e.g., [8]. Consider a family of integers fN k g k2f1:::Kg , all greater than two. Let X kN be the space of polynomials on\Omega k of degree N k in each space variables, which vanish on \Omega\Gamma and let X kN be the space of traces of functions of X kN on @\Omega k . A product space is defined by Y For each subdomain\Omega k , a quadrature formula is given, by the Gauss-Lobatto- Legendre formula on (\Gamma1; 1), an affine transformation, and tensorization. We denote the corresponding nodes and weights by i;j;k and ! i;j;k , and by GL;k the quadrature GL;k The nodes i;j;k define a grid T kN on\Omega k . A discrete bilinear form is given by GL;k Consider two adjacent and\Omega ' and let L k' be equal to N k or N ' . Let and denote by ~ W k;';N the space of polynomials of degree N k' . The subspace YN of XN is now introduced by R oe and the subspace YN of XN by R (v oe Slave and mortar sides are introduced exactly as in Section 2, and the Poincar'e-Steklov bilinear form s can be introduced in terms of the problem of finding YN such that GL;k f ~ Here N is obtained by solving the analogue of (3) and ~ vN is the discrete harmonic extension of vN . As before, we look for a decomposition of the space YN into the direct sum of a coarse space YH , of dimension NV , and a fine space Y H N . For each crosspoint A and for each k 2 KA , the basis vector e A;k of YH is fully determined by the following conditions: 1. e A;k 2. for all vertices B 6= A 3. for all ' 6= k, for all vertices B 4. e A;k is linear on the master edges. Suppose first that \Gamma k' is a slave edge of @\Omega k . We can then check, using coordinates such that \Gamma with the value H corresponding to A, that e A;k e A;' Here Ln is the Legendre polynomial of degree n. From (39) and (40), it is straight-forward to show that C: The next lemma provides estimates of the and H 1=2 \Gammanorms of e A;k . Lemma 10. log(N k' ) Proof. Without loss of generality, we can assume that \Gamma k' and to simplify the notations, we also set . Recalling that it is easy to show that In order to evaluate the H 1=2 -norm of ~ e k;';A , we compute the H 1=2 -norm of Ln . By the definition of the norm, we have to estimate dxdy: Integration by parts, in the case where n is even, leads to x\Gammay x\Gammay x\Gammay 1\Gammay \Gamma1\Gammay x\Gammay and similarly, in the case where n is odd, Hence, dxdy 1\Gammay 2 dy n even; 1\Gammay 2 dy n odd: In order to estimate the last terms in this formula, we use Gaussian quadrature based on the roots (i i ) 1in of Ln . It is well known that there exists positive quadrature weights (! i ) 1in such that We therefore obtain We now recall, see [25],(thm 6.21.3 & (15.3.14)) that and that there exist two positive constants c and C such that c We find that dy We recognize this last expression as a Riemann sum of ( 1 dy Clog(n): The same argument leads to dy Clog(n); which in turn leads to x\Gammay x\Gammay dxdyj Clog(n): Finally, we remark that Here (Ln (x)\GammaL n (y)\Gamma(x\Gammay)L 0 (x\Gammay) 2 is a polynomial of degree ! n in both x and y. The orthogonality of the Legendre polynomials then leads to By using (41), we find that dxdy Clog(n): Thus, the square of the H 1=2 -norm of Ln is bounded by Clog(n) and so is the square of the H 1=2 -norm of ~ The next result is the analogue of Lemma 2: Lemma 11. Let A be a crosspoint and let k; Assume that is a slave side of\Omega k and that master side of\Omega k . Then, je A;k Proof. Since we are interested in the H 1=2 is possible to rescale the problem and assume that the edge \Gamma k' is (\Gamma1; 1). We again set We first use the following inverse inequality, je A;k Cn 4ffl je A;k see [8]. Let us denote by f k;m;A and f k;';A the functions on @\Omega k which are equal to e A;k respectively, and vanish on @\Omega k n\Gamma k' , respectively. It is clear that, almost everywhere, e A;k The semi-norm jf k;m;A j 2 can be computed explicitly, because f k;m;A is piecewise linear, and the following estimate is obtained: To evaluate the contribution of f k;';A , it is enough to estimate the following quantity Here is the characteristic function of (\Gamma1; 1). It can be proved, as in Lemma 5.1, that the contribution of jx\Gammayj 2\Gamma2ffl is bounded by C(1+log(n)). There remains to estimate This is done as follows: Hence, je A;k Choosing log(n) yields je A;k Let ~ S be the matrix of s N in the new basis described above. The matrix ~ S can be written as ~ ~ ~ In order to design a preconditioner for ~ S, we replace the block ~ SNH by 0, and the block ~ SNN by its block diagonal with one block for each edge. The resulting preconditioner is s N be the bilinear form corresponding to the matrix " S. The bound, is obtained exactly as in Lemma 5. The following lemma will also be needed: Lemma 12. There exists a constant C such that 8v kN 2 X kN , Similarly, if x 0 is a point of Proof. To prove (43), we first rescale so that the diameter then prove that This formula is true for any function and can be derived by using a Fourier transform argument. From an inverse inequality for polynomials, we also have Choosing yields the desired result. Finally, (44) is obtained by a standard quotient space argument. Given Lemma 12, analogues of Lemma 6 and Theorem 1 can be obtained as in the finite element case. The only notable difference is when estimating the integral \GammaH cf. (21)-(24) in the proof of Theorem 1. The integral is again split into two, over respectively. Here the ( i ) i2f0;Nkg are the Gauss-Lobatto points in [\Gamma1; 1]. We note that 1 ) and that an inverse inequality for polynomials of degree n is given by Finally, since all necessary technical tools are at hand, the proof of the following main result can be obtained Theorem 5. There exists a constant C such that 5.2. The geometrically nonconforming case. The previous mortar spectral element method can be generalized straightforwardly to the case where the assumption of geometrical conformity is relaxed. The analysis of our preconditioner requires, in Lemma 5.5 below, a bound on the relative size of the intersection of two edges; in order to provide estimates which only depend polylogarithmically on the parameter N , we make the assumption that c; a condition less standard and more stringent than (28). The mortar method is based on the definition of the master and slave sides as described in Section 4 in the finite element context. The projection that gives slave values in terms of those on the master edges is defined as in Subsection 5.1. The iterative substructuring algorithm now considered is the same as that of Section 4. We now have the new tools of Subsection 5.1 at our disposal, and in order to prove a result equivalent to Theorem 3 for the spectral approximation, we additionally only have to derive an analogue of Lemma 7. We do so after some preliminary work. The definition of the mortar condition naturally leads to the introduction of a projection operator N (\Gamma1; 1); given by This projection operator has already been analyzed in [7] and certain stability and approximation properties are given in that paper. While we do not now know how to prove a uniform bound for the H 1=2 00 -norm of this operator, the following results will allow us to derive a strong result on our preconditioner. Lemma 13. Let a be any real number in a denote the piecewise linear, discontinuous function on that vanishes on linearly from 1 and 0 over the interval [a; 1]. There then exists a constant C such that log(N Proof. Let 'N be a polynomial of degree N in one variable that vanishes at \Sigma1. It is easy to see that it can be written as an Recalling that the Legendre polynomials Ln are L 2 (\Gamma1; 1)-orthogonal, with norm and that they satisfy the differential equation d dx we can first show that a 2 Using now the integral relation and the differential equation once more, we find that an Here we have used the L 2 \Gammaorthogonality of the Ln and the formula an (a Let us denote by [IP N (\Gamma1; 1)"L 2 (\Gamma1; 1); IP N (\Gamma1; 1)"H 1 the interpolation space, with index ', between the space of polynomials IP N (\Gamma1; 1) of degree N provided with the L 2 -norm and the H 1 respectively. By a main result of interpolation theory [20], it follows from our two estimates of the norms of 'N that there exists a constant C such that, for all 'N 2 IP N (\Gamma1; 1), C We recall now that, according to [21], [IP N (\Gamma1; 1)"L 2 (\Gamma1; 1); IP N (\Gamma1; 1)"H 1 coincides topologically with IP N (\Gamma1; 1) provided with the H ' -norm. In particular, we can conclude that C In order to complete the proof, we now evaluate the coefficients an of the decomposition of given by (46). By using the definition of a , we find that a (x)L 0 a a Ln (x) Combining the integral and the differential formulae for the Legendre polynomials, we deduce (a) (a) From [1], (22.14.9), we know that jLn (a)j C As another consequence of the integral formula, we find that which yields (a) Finally from (48),(49), and (50), we deduce that By using (46), we obtain Thus, an and hence By introducing this bound into formula (47), we can conclude that log(N It is now an easy matter to state an analogue of the Lemma 7. Lemma 14. Assume that (45) holds, let A be a crosspoint, and let \Gamma j k be a master side of\Omega k with A as an end point. Let ' 6= k, je A;k Proof. Let us denote by ' the function obtained from by a scaling which maps ' with the further property that ' B are the end points of \Gamma i k . By construction, e A;k continuous and it even belongs to H 3 2 \Gamma" for any positive ". We decompose e A;k ' into the following sum e A;k supplies the desired bound for the first term j ' (@\Omega l ) . The bound on the second, j ' (@\Omega l ) , is obtained as in the geometrically conforming case. Lemma 14, we are ready to prove the following result using the same arguments as in Section 4. Theorem 6. Under the assumption (45), there exists a constant C such that --R Handbook of mathematical functions Substructuring preconditioners for the Q 1 mortar element method M'ethodes it'eratives de sous-structuration pour les 'el'ements avec joints A fast solver for Navier-Stokes equations in the laminar regime using mortar finite element and boundary element method On the mortar element method: generalization and implementation The mortar finite element method with Lagrange multipliers A spectral methodology tuned to parallel implementations Approximations Spectrales de Probl'emes aux Limites Ellip Raffinement de maillage en 'el'ements finis par la m'ethode des joints A new nonconforming approach to domain de- composition: the mortar element method Iterative methods for the solution of elliptic problems on regions partitioned into substructures The construction of preconditioners for elliptic problems by substructuring A hierarchical preconditioner for the mortar finite element method Some domain decomposition algorithms for elliptic problems Domain decomposition algorithms with small overlap Efficient iterative solvers for elliptic finite element problems on nonmatching grids Domain decomposition methods in computational mechanics Domain decomposition with non matching grids: Augmented Lagrangian approach Probl'emes aux Limites non Homog'enes et Applications Rel'evement de traces polynomiales et interpolations hilbertiennes entre espaces de polyn In preparation. Nonconforming discretizations and a posteriori error estimators for adaptive spectral element techniques A domain decomposition algorithm using a hierarchical basis --TR --CTR C. Cinquini , P. Venini, A mortar approach for the analysis and optimization of composite laminated plates, Computational structures technology, Civil-Comp press, Edinburgh, UK, 02 Ilaria Perugia , Dominik Schtzau, On the Coupling of Local Discontinuous Galerkin and Conforming Finite Element Methods, Journal of Scientific Computing, v.16 n.4, p.411-433, December 2001 Dan Stefanica, FETI and FETI-DP Methods for Spectral and Mortar Spectral Elements: A Performance Comparison, Journal of Scientific Computing, v.17 n.1-4, p.629-638, December 2002 Micol Pennacchio, The Mortar Finite Element Method for the Cardiac Bidomain Model of Extracellular Potential, Journal of Scientific Computing, v.20 n.2, p.191-210, April 2004 Feng , Catherine Mavriplis , Rob Feng , Rupak Biswas, Parallel 3D Mortar Element Method for Adaptive Nonconforming Meshes, Journal of Scientific Computing, v.27 n.1-3, p.231-243, June 2006 T. P. Mathew , G. Russo, Maximum norm stability of difference schemes for parabolic equations on overset nonmatching space-time grids, Mathematics of Computation, v.72 n.242, p.619-656, 1 April
domain decomposition;mortar finite element method;iterative substructuring
588473
Analysis of Numerical Errors in Large Eddy Simulation.
We consider the question of "numerical errors" in large eddy simulation. It is often claimed that straightforward discretization and solution using centered methods of models for large eddy motion can simulate the motion of turbulent flows with complexity independent of the Reynolds number and dependence only on the resolution ``$\delta$'' of the eddies sought. This report considers this question analytically: Is it possible to prove error estimates for discretizations of actually used large eddy models whose error constants depend only on $\delta$ but not Re? We consider the most common, simplest, and most mathematically tractable model and the most mathematically clear discretization. In two cases, we prove such an error estimate and try to explain why our technique of proof fails in the most general case. Our analysis aims to assume as little time regularity on the true solution as possible.
Introduction The laminar or turbulent ow of an incompressible uid is modeled by solutions (u; p) of the incompressible Navier-Stokes equations: in r in in @ Z (1) Here is a bounded, simply connected domain with polygonal boundary d is the uid velocity, p R is the uid pressure, f(x; t) is the (known) body force, u 0 (x) the initial ow eld and Re the Reynolds number. Unfortunately, when Re is large the resulting turbulent ow is typically so complex that, so called, direct numerical simulation of (u; p) is not practically feasible. One conjecture of Leray is that \turbulence" in nature is associated with a breakdown of uniqueness of weak solutions to (1). It is known that, for example, weak solutions to (1) are unique for There numerous generalizations of this basic result, [GHR00], [Lad69]. With this in mind, solutions u to (1) with kruk are frequently described as \laminar". Thus, the L p {regularity in time which can be reasonably assumed is of critical importance. There are numerous approaches to the simulation of turbulent ows in practical settings. One of the most promising current approaches is large eddy simulation (LES) in which approximations to local spacial averages of u are calculated. A spacial length scale - is selected and the velocity scales of O(-) and larger are approximated directly while the eects of those smaller than O(-) on the O(-) and larger eddies are modeled. In computational turbulence studies using LES it is often reported that the resulting computational complexity is independent of the Reynolds number (but dependent on the resolution sought, -). There has been little or no analytical support for this observation however. The goal of this report is to begin numerical analysis in support of this claim. To be more specic, a smooth, nonnegative function g(x) with Z R d is selected and the mollifer g - (x) is dened in the usual way: One common example is a Gaussian, where the summation convention is used. The spacial averaging/ltering operation is now dened by convolution In LES, approximations to (u; p) are sought rather than to (u; p). The usual procedure is to rst lter the Navier-Stokes equations: in r in where the \Reynolds' stress tensor" T is Closure is addressed by a modeling step in which T is written in terms of u. The resulting (closed) space ltered Navier-Stokes equations are solved numerically. In this procedure, there are three essential issues: 1. The \modeling error" committed in approximating T. 2. The \numerical error" in solving the resulting system. 3. Correct boundary conditions for the ow averages. In this report, we study the numerical error analytically. Since there are many models in LES (see, e.g., [IL00], [GL00], [HMJ00], [FP99], [BFR80], [Lia99], [Sag98], [Par92] for examples) and few analytical studies, we take herein the simplest model commonly in use presented, for example, in Ferziger and Peric [FP99, Section 9.3]. To describe the model, let D (u) be the deformation tensor associated with the indicated velocity eld by: The Reynolds stresses are thought of as a turbulent diusion process based upon the Boussinesq assumption or eddy viscosity hypothesis that \turbulent uctuations are dissipative in the mean", [IL00], [Fri95], [MP94], [Par92]. We will accordingly consider a model of the where turb turbulent viscosity or eddy viscosity. This turbulent viscosity's determination can be very complex, involving even solutions of accompanying systems of nonlinear, partial dierential equations. In the simplest case, the turbulent viscosity depends on the mean ow u through the magnitude of the deformation of u; (D (u)), with a functional dependence. Under the Boussinesq assumption, r T should act like a physical viscosity. Following the reasoning of Ladyzhenskaya [Lad70], thermodynamic considerations imply that the Taylor series of turb (D ) should be dominated by odd degree terms. The simplest case is of linear dependence upon jD j turb (jD where jD j denotes the Frobenius norm of D . For specicity and for accord with the most commonly used Smagorinsky [Sma63] model, we take a 0 (-) 0 and a 1 Other scalings are possible, [Lay96], though less tested, as are many other subgridscale models typically either chosen to be around 0.1 or taken to be a function extrapolated as in the \dynamic subgridscale model" of Germano, [GPMC91]. With the model (2) the resulting system of equations for the approximations (w; q) to in in in Z Boundary conditions must be supplied for the large eddies. It is physically clear that large eddies do not adhere to solid walls. (For example, tornadoes and hurricanes move while touching the earth and lose energy as they move.) Therefore, in [GL00], [Sah00], (see also [Par92] for the use of similar boundary conditions in a conventional turbulence model), it was proposed that the large eddies w should satisfy a no-penetration condition and a slip with friction condition on @ @ where ~ t is the Cauchy stress vector on , for background information see Serrin [Ser59], is the friction coe-cient (calculated explicitly in [Sah00]), ^ n the outward unit normal j an orthonormal system of tangent vector's on . The friction coe-cient can be calculated once a specic lter is chosen, [Sah00]. It has the property ([Sah00]) that no slip conditions are recovered as - ! 0: A Dirichlet boundary condition ow on 0 , is appropriate if 0 is an in ow boundary upon which u can be calculated by extending the known, in ow velocity eld upstream. The Cauchy stress vector ~ t includes the action of both the viscous stresses and Reynolds stresses and is given by: Standard properties of convolution operators imply that the ow averages (u; p) are in space, have bounded kinetic energy Z Z have no solution scales smaller than O(-) and converge to u as On the other hand, it is not obvious, nor has it been proven yet, that solutions (w; q) to the large eddy model approximating (u; p) share any of these properties! Nevertheless, the spacial regularity of solutions (w; q) we shall consider to be a modeling issue (beyond the scope of this report studying numerical errors in LES). The time regularity of solutions (w; q) is still an important consideration. For example, we shall show that solutions of this model satisfy Z Tkrwk 3 uniformly in Re. One goal is to keep the assumed time regularity as close to L 3 (0; T ) as possible and below L 4 (0; T ). The fundamental error analysis of Heywood and Rannacher [HR82] for the Navier-Stokes equations is based, in part, on a laminar-type assumption 62 Weakening this to an assumption of the form ru 2 L 3 (0; T; L 3((as we seek to do herein) is nontrivial. Preliminaries This section sets the notation used in the report, describes the function spaces employed and collects several useful inequalities. The notation used is standard for the most part. The norms, for p 6= 2, are denoted explicitly as kfk L p . Sobolev spaces W k;p are dened in the usual way, [Ada75]. The associated norm is denoted k k k;p . If the domain in question is not (e.g., will be explicitly indicated. If norms will be for the W norm and k k k; for the W k;p () norm and k k and k k , respectively, for the L and L 2 () norms. Suppose the polygonal boundary = @ is composed of faces The spaces associated with the boundary conditions (4) are: d Z The boundary condition in X is dened to hold in the sense of the trace theorem on each is the outward unit normal to = @ The L and L 2 () inner products are denoted (; ) and (; ) respectively. denotes the usual deformation tensor, dened in the introduction. The denotes an orthonormal system of tangent vectors on . Whenever it will be understood that the term is to be summed over the two tangent vectors if For example: The following dual norms are dened in an equivalent but slightly nonstandard way: kfk := sup kfk W 1;q R t(f; v)dt 0 R tkD (v)k q Lemma 2.1 [Inf-Sup Condition]. Let ~ Jg: The velocity-pressure spaces ( ~ X;Q) satisfy the inf-sup condition: sup kk C > 0: (5) Proof: The trace theorem [Gri92] and Korn's inequality together show that (5) is implied by the usual inf-sup condition sup d kk krvk C > 0: Lemma 2.1 implies that the space of weakly divergence free functions is a well dened, nontrivial, closed subspace of ~ X. The inf{sup condition (5) is used to get a bound on kqk once a bound on kD (w)k is proved. The bound on kD (w)k follows from Lemma 3.1 below. An inf{sup condition with ~ replaced by X and kD (w)k by kD (w)k L 3 would also su-ce to get the bound on kqk. Remark 2.2 Since is not C 1 discontinuities in j and ^ have forced modications in the norms to piecewise denition. For example, but The conforming nite element method for this problem begins by selecting nite element spaces X h X and Q h Q, where \h" denotes as usual a representative meshwidth for satisfying the usual approximation theoretic conditions required of nite element spaces. The condition that X h X imposes the restriction that v h ^ . For intricate boundaries, this could possibly be onerous so it is interesting to consider imposing with penalty or Lagrange multiplier methods, following, e.g., the work in [Lia99]. Nevertheless, there is already considerable computational experience with imposing this condition in nite element methods, see, e.g., [GS00], [ESG82], so we shall not focus on the interesting detail of the treatment of corners. Without these additional regularizations in the numerical method, it is useful in the analysis to assume that (X h satises the discrete analogue of (5): sup kD (v h where C > 0 is independent of h. The next lemma shows, in essence, that if the computational mesh follows the boundary and if the velocity space restricted to no-slip boundary conditions and the pressure space satisfy the usual inf-sup condition, then (6) holds. Lemma 2.3 [Discrete Inf-Sup Condition]. If (X sup d then (6) holds. Proof: By trace theorem [Gri92] and the Poincare-Friedrichs inequality, for any h 6= 0, kD (v h Thus, (6) will be assumed throughout this report. Under (6), the space of discretely divergence free functions is a nontrivial closed subspace of X h , [GR86], [Gun89]. We shall frequently use Young's inequality in the form: ab a The generalization of Holder's inequality: Z r is also useful. We shall frequently use the Sobolev embedding theorem, often, but not always, in the form that in 3 dimensions W The nonlinear form in the subgridscale term: for v; w 2 W (jD (w)jD (w); D (v)) is of p-Laplacian type (with Thus, it is strongly monotone and locally Lipschitz continuous in the sense made precise in the following well-known lemma, see, e.g., [Lay96], [DG91]. Lemma 2.4 [Strong Monotonicity and Local Lipschitz-Continuity]. There are constants C and C such that for all 6 d and (jD (jD Korn's inequalities relate L p norms of the deformation tensor D (v) to those same norms of the gradient for 1 < p < 1, see Galdi [GHR00], Gobert [Gob62, Gob71], Temam [Tem83] or Fichera [Fic72], and fail if Theorem 2.5 [Korn's Inequalities]. There is a C > 0 such that for 1 < p < 1 for all v 2 (W 1 d . Further, if (v) is a semi-norm on L which is a norm on the constants, then holds for 1 < p < 1 and for all v 2 (W 2 d . As a consequence of Korn's inequality it follows that, taking for all v 2 fv 2 W 1;p We will often use Poincare's inequality, which holds since p. 56., kvk krvk; for all v 2 X: We shall use the Gagliardo-Nirenberg inequality in W 1;p \ X to reduce the time regularity required for w. This inequality [Ada75], [Nir59], [GHR00], [DiB93] states that provided satises a weak regularity condition (holding in particular for polygonal domains) and kvk L q Ckrvk a d where, for if s In particular, note that taking kvk L Ckrvk 2=3 The following combination of this and Korn's inequality will be useful in Section 4. Lemma 2.6 Let and Proof: This follows immediately from (7) and Korn's inequality. 3 Finite Element Formulation This section develops the nite element method for the LES model. The stability of the model is also studied. In particular, we show w and w h 2 L 1(uniformly in Re. Lastly, the error in an equilibrium projection is considered. The variational formulation is derived in the usual way by multiplication of (3) by (v; (X; Q) and applying the divergence theorem. The boundary integral terms require careful treatment (following, e.g., [Lia99]) on account of the slip with friction condition on . Let 0 be a constant. The formulation which results is to nd w (w and w(x; X. For compactness, dene the nonlinear and trilinear form: (w It is a simple index calculation to check that for v 2 X; w 2 V (since such functions have zero normal components on ) (w rw; Thus, the variational formulation can be rewritten as: nd (w; (w for all (v; Using Lemma 2.4, it is easy to prove that the LES model (3), (4) satises the analogue of Leray's inequality for the Navier-Stokes equations. Lemma 3.1 [Leray's inequality for the LES model]. A solution of Z t" J Proof: in (9) and use Lemma 2.4. Remark 3.2 1. Because of the slip with friction boundary conditions (4), it is important to choose the formulation of the viscous terms, as in (8), (9), involving the deformation tensor. 2. Leray's inequality immediately implies stability in various norms (which we will de- velop) and is the key, rst step in proving existence of weak solutions to (3), (4). This last question is fully investigated (under dierent boundary conditions) in remarkable papers by Ladyzhenskaya [Lad67], Pares [Par92] and Du and Gunzburger [DG91]. The continuous-in-time nite element method for (3) uses the variational formulation as follows. First, velocity-pressure nite element spaces X h X \ W satisfying (6) are selected. Next, the least squares parameter 0 is selected. The nite element approximations to (w; q) are maps (w h ; q h (w h for all (v h ; h is an approximation to w(x; It is straightforward to verify that Leray's inequality holds for w h as well as w. Lemma 3.3 [Leray's inequality for w h ]. Any solution of Using various inequalities in the right hand side, stability bounds for w h follow from Lemma 3.3. Proposition 3.4 [Stability of w h ]. The solution w h of (10) J Proof: Inequality (11) follows by applying Young's inequality to Lemma 3.3. The bound (12) follows from the denition of the dual norm and ab a 3 applied in the same manner. For in (10), use Lemma 2.4 and apply the Young's inequality on the right hand side. This gives: d dt Inequality now follows by using an integrating factor. In the analysis of the error in the approximation of the time dependent problem, it is useful to have a clear description of the error in the Stokes projection under slip with friction boundary conditions, [Lia99]. It is also necessary that any dependence on Re; - and be made explicit. Under the discrete inf-sup condition, the Stokes projection dened as follows. Let (w; q), where ( ~ (r (w ~ ((w ~ (r (w ~ This is equivalent to the following formulation provided Given (w; q), nd ~ (r (w ~ ((w ~ for all v h 2 V h and h 2 Q h . Under the discrete inf-sup condition, it is well-known that q) is a quasi-optimal approximation of (w; q). The dependence of the stability and error constants upon Re and is important to the error analysis. That dependence is described in the next lemma and proposition. Lemma 3.5 [Stability of the projection ~ w]. Let w 2 V be given. Then, ~ w satises if > 0: kr ~ in the second formulation of the Stokes projection. This gives immediately: kr ~ (w from which the rst result follows. If the term (q; r ~ w) is bounded by noting that r ~ trace (D ( ~ w)) so that (q; r ~ Proposition 3.6 Suppose the discrete inf-sup condition (6) holds. Then, ( ~ exists uniquely in kr (w ~ k(w ~ C inf Proof: The proof follows standard arguments, carefully tracking the dependence of the constants upon Re and . Note that the use of least squares penalization of the divergence allows an error estimate for the Stokes projection whose constants are essentially independent of the Reynolds number in a suitably weighted norm. 4 The Convergence Theorem Let us rst note that for standard piecewise polynomial nite element spaces it is known that, e.g., the L 2 -projection of a function in L p ; p 2, is in L p itself and the L 2 -projection operator is stable in L w denote an approximation of w in V h \ W 5 for example, the L 2 -projection under the conditions of [CT87]. We assume that each norm of ~ w can be bounded by the same norm of w times a constant which is independent of Re and h. This assumption is proven for many piecewise polynomial nite element spaces in [CT87]. The error is decomposed as w) (w h ~ w and An error equation is obtained by subtracting (9) from (10) and using the fact that w 2 V . This gives, for any v h 2 V h \ W and h 2 This is rewritten, adding and subtracting terms and setting v The monotonicity lemma (Lemma 2.4) implies that and with r := maxfkD (w)k L 3 ; kD ( ~ Remark 4.1 If ~ w is taken to be the Stokes projection of (w; q) into V h then, e.g., the term kD ()k" on this last right hand side does not occur. Inserting these two bounds in (15) and using the Cauchy-Schwarz and Young's inequalities d dt s r 3=2 kD ()k 3=2 Picking collecting terms gives2 d dt This is the basic dierential inequality for the error. Three cases will be considered, revolving around the treatment of the rst term on the RHS of (16). Remark 4.2 If completely analogous estimate holds with the pressure term modied to be either nonuniform in Re (e.g., \C(2Re 1 or nonuniform in -. Consider the convection terms The terms containing \" shall be bounded rst. Consider Using the inequalities in Section 2 appropriately gives 1 1h The term b(; w h ; h ) is bounded similarly as follows: 1 Korn's inequality and the stability bounds of Section 3, (12) and (13), immediately imply that D (w h uniformly in Re so that krw h k 2 uniformly in Re. The imbedding theorem and Korn's inequality also imply kwk 2 uniformly in Re. Thus, these bounds su-ce for a later application of Gronwall's inequality. The rst term containing only h ; b(w; h ; h ), is zero due to skew symmetry. Thus, there only remains the term Estimating the term is the essential, core di-culty in obtaining an error bound which is uniform in Re. There are only a few natural ways to bound this using Holder's inequality and the Sobolev embedding theorem. There are two cases in which the analysis is successful very regular, rw 2 L 1(There is one important case in which the analysis fails: (iii) a 0 3(To highlight subsequent analysis and, hopefully, spur further study, we shall rst present the case (iii) and explain the failure of the analysis. 4.1 The case rw 2 L 3 (0; 8 and a 0 If we assume only that rw 2 L 3 (0; T; L 84 there is no need to add and subtract terms since a priori bounds on krw h k L 3 (0;T;L 3 ) have been proven which are uniform in Re. Thus, we can use Holder's inequality to write: picking s using the embedding and Poincare's inequality give Remark 4.3 Using Lemma 2.6 instead of the embedding of W 1;3 ! L 6 changes the critical exponent on k h k \3=2" to 12=7 in the rst term of (20) but not the nal conclusion. Combining (18), (19), (20), with gives an initial bound on the convection term's dierence kk 3=2 Inserting (21) into (16), applying Korn's inequality and collecting terms gives2 d dt Thus, pick such that d dt Consider the three terms bracketed on this right hand side. The rst is approximation theoretic; the second is an L 1 function multiplying k h (t)k 3=2 the third is an L 1 function multiplying k h (t)k 2 . Let y(t) := k h (t)k 2 . This inequality may then be written as: d dt (nonnegative terms) C(t)h where The nal step would normally be to apply Gronwall's inequality to deduce y(t) =2 to be bounded by its initial values and approximation theoretic terms. Unfor- tunately, the term y 3=4 is not Lipschitz, so the argument fails at this last step. Tracing the inequalities backward, the problem term arises from the steps used to bound to obtain Re independence. The error analysis in the successful cases (i) and (ii) centers therefore on alternate bounds for this term. We shall rst consider case (i). Remark 4.4 If the estimate in (20) is improved as noted in Remark 4.3, the term y(t) 3=4 is changed to y(t) 6=7 but the nal conclusion still holds. 4.2 The case rw 2 L 3 (0; 8 and a 0 (-) > 0 Theorem 4.5 Assume > 0 and a 0 (-) > 0. Let a Then, there is a C independent of Re and h, such that Further, there is a C independent of Re and h, such that Then, the error w w h satises for T > 0 ~ F(w ~ with F(w ~ k(w ~ k(w ~ +kr (w ~ wk 3=2 Proof: This analysis follows closely the previous discussion except for the treatment of the and the nal application of Gronwall's inequality. Consider therefore Integration by parts and using the fact that h ^ on gives: Using the embedding H 1=2 ,! L 3 in d = 2; 3 and Young's inequality give 1 Consider now the last term on the above right hand side. By Holder's inequality, we obtain The Sobolev embedding theorem implies that for any s; 1 s < 1 in 2 or 3 dimensions C(s; kD (w h )k 2 This implies that for any r 0 > 2 Consider the last term on the above right hand side. The Sobolev embedding theorem also implies (The nal result is not improved by applying here instead the Gagliardo-Nirenberg inequal- ity.) As r inequality. Thus, picking r close enough to 2 implies, using an embedding inequality and Korn's inequality, for any t > 0. Thus, for these values of r 0 and s for any t > 0. For conjugate exponents inequality, we then have Picking these values of r 0 and s: Using this bound, (22) and (23) gives nally Remark 4.6 It appears on rst consideration that this last term (r h ; w h h ) can be agreeably bounded more directly and easily by: This bound, while certainly true, is not su-cient because of the condition that inevitably arises from using it that w h or w 2 L 1 . The extra work in the bound we use reduces the time regularity requirements arising from this term to w h 2 L 1;3((which is bounded uniformly in Re by problem data in Section 3). Substituting this bound for b( in the derivation of the upper estimate (21) for the dierence of the convection terms gives 3 To proceed further, (24) is inserted in the right hand side of (16). This yields the dierential inequality2 d dt Pick These choices simplify (25) to:2 d dt a Before applying Gronwall's inequality, let us rst verify that it will indeed give us an error bound that is uniform in the Reynolds number by considering the coe-cients on the right hand side of (26). By the stability estimates kwk 2 uniformly in Re. Thus, kwk 3=2 kk 3=2 Consider the (critical) bracketed coe-cient of the last term on the right hand side. We must show this coe-cient is in L 1 (0; T ) uniformly in Re. Indeed, by the stability estimates uniformly in Re. Since T < 1, L 3 (0; T ) thus the rst factor of the last term is in L 1 (0; T ) uniformly in Re. In addition, note that r CkD(w)k L 3. Hiding all constants in generic C's, Gronwall's lemma now implies for almost all t 2 [0; T Z tkD (w)k 3=2 Z t1 Note that by the Cauchy{Schwarz inequality in L 2 (0; T ) and the stability estimates Z tkD (w)k 3=2 C(-)kD ()k 3=2 Now, the essential supremum of t 2 [0; T ] is applied on both sides of the inequality. As the triangle inequality completes the proof of Theorem 4.5. 4.3 The case rw 2 L 2 (0; T ; L 8 and a 0 (-) 0 We now consider the case of smoother w, i.e., 9 uniformly in Re; allowing for the case a 0 (-) 0. This case is primarily of interest because many tests involve \academic" ow elds given in closed form (as in Section 5). These are typically smooth and bounded. In this case Theorem 4.7 gives an error estimate with constants independent of Re (but depending on - and ). It is noteworthy in this estimate that multiplicative constants depend on - but the rate constant in the (inevitable) exponential term takes the form with no explicit dependence on -. Theorem 4.7 Suppose a 0 (-) 0; > 0 and w 2 L 9 uniformly in Re. Let then there is a C (w) such that be such that Then, the error w w h satises: ~ F(w ~ with F(w ~ k(w ~ k(w ~ kr (w ~ kD (w ~ Proof: In this case, the dierence in the nonlinear terms is decomposed a bit dierently as: Consider the individual terms on the right hand side of (27): 1 (w h Combining these three estimates gives The term kw h k L 6 is bounded using the Gagliardo-Nirenberg inequality Since kw h k is bounded uniformly in and h by (12) or (13), it follows This bound, together with (28) is now inserted in the right hand side of (16) giving2 d dt To apply Gronwall's inequality we need4 in other words w 2 L The term on the right hand side of this inequality containing r 3=2 is treated as in the proof of Theorem 4.5. In the nal result of Gronwall's lemma, we must also verify that the resulting terms containing kD (w h )k L 3 are bounded uniformly in Re. To this end, apply Holder's inequality: Z TkD (w h )k 4=3 1. From the stability estimates, we clearly must take q such that 4q=3 3. Accordingly, take 9. This gives Z TkD (w h )k 4=3 Similarly, for q and q 0 conjugate exponents: Z TkD (w h )k 2 kD (w h )k 2 The stated error estimate now follows from Gronwall's inequality and the triangle inequality as in the proof of Theorem 4.5. 5 A Numerical Example To give a numerical illustration several decisions must be made, mainly to work on an 'aca- ow problem with a known exact solution or a more realistic ow problem containing the accompanying uncertainties. Since our aim is to illustrate a convergence theorem we have chosen the former. (To assess a model or study the limitations of an algorithm we would naturally have chosen the latter.) Accordingly, we have selected the vortex decay problem of Chorin [Cho68], used also by others, e.g., Tafti [Taf96]. The domain is choose For the relaxation time Re this is a solution of the Navier-Stokes equation consisting of an array of opposite signed vortices which decay as t ! 1. The right hand side f , initial condition and non-homogeneous Dirichlet boundary conditions are chosen so that this is the closed form solution of (3). Since we are studying convergence as h ! 0 for - xed and Re varying we have accordingly chosen Relaxation time Vortex conguration Final time Eddy Scale It is signicant that n so that the vortices are larger than O(-) and hence should be \visible" to the model. Tables 2 to 4 give the error in L 6 and show uniformity in Re and in L 5 and show weak dependence on Re (since this term in the error estimate is scaled by Re 1=2 ). The least squares constant is chosen to be zero. Time discretization is by the fractional step -method with the indicated time steps. The tables show a decrease of the error as h decreases until it reaches the error introduced by the time stepping procedure. The spacial discretization is done on a uniform square mesh with the Q 2 =P disc denoted in element with the associated meshwidths indicated. The viscous term is treated not as (rw h ; rv h ) but using the deformation tensor formulation, (D (w h ); D (v h )) as analyzed herein. Both the Smagorinsky subgridscale model and the convection term are treated implicitly. Using the above elements and meshes the calculations involved the numbers of degrees of freedom listed below in Table 1. mesh width velocity d.o.f. pressure d.o.f. total d.o.f. 1=8 578 192 770 Table 1: Mesh widths and degrees of freedom (d.o.f.) in space. These are certainly not extremely large numbers of degrees of freedom. However, their importance is only relative to the Reynolds numbers chosen the resolution sought Again, LES is focused on situations in which the number of degrees of freedom is small relative to Re. Thus, the chosen values of h and Re seem appropriate. The Tables { 4 present the L norms of the errors. Note that the trends are exactly as anticipated by the theory; there is none to minimal degradation in the error as Re increases from 10 2 to 10 5 and the error plateau's as h decreases at a value which seems to be the induced error in the time stepping procedure. Re n h 1/8 1/16 1/32 1/64 Table 2: 0:02. Note the uniformity in Re. Re n h 1/8 1/16 1/32 1/64 Table 3: observe the uniformity in Re. Re n h 1/8 1/16 1/32 1/64 Table 4: 0:005: Note the uniformity in Re and the reduction of the minimal error as t decreases. The Tables { 7 present the errors in L These are not predicted to be uniform in Re so some error degradation is expected as Re increases. Very mild degradation is indeed observed. The degradation is mild, possibly because are predicted to have the same uniform in Re convergence rates. Thus, the theory forecasts increase as Re increases until it reaches the (slower) uniform convergence rate predicted by the kD (e)k L 3 (0;T;L 3 ) bound. Re n h 1/8 1/16 1/32 1/64 Table 5: kD (e)k L 2 (0;T;L 2 Re n h 1/8 1/16 1/32 1/64 Table Re n h 1/8 1/16 1/32 1/64 Table 7: kD (e)k L 2 (0;T;L 2 --R spaces. Improved subgrid models for large eddy simulation. Numerical solution for the Navier-Stokes equations The stability in l p and w 1 Analysis of a ladyzhenskaya model for incompressible viscous ow. Degenerate parabolic equations. The implementation of normal and/or tangential boundary conditions in Existence theorems in elasticity. Computational Methods for Fluid Dynamics. Turbulence, the Legacy of A. An Introduction to the Mathematical Theory of the Navier-Stokes Equations Fundamental Directions in Mathematical Fluid Mechanics. Approximation of the larger eddies in uid motion ii: A model for space Une inequation fondametale de la th Sur une in A dynamic subgrid-scale eddy viscosity model Finite Element Methods for Navier-Stokes equa- tions Singularities in Boundary Value Problems Incompressible Flow and the Finite Element Method. Finite Element Methods for Viscous Incompressible Flows. Large eddy simulation and the variational multiscale method. Finite element approximation of the nonstationary navier-stokes problem i: Regularity of solutions and second order error estimates for spacial discretization An assessment of models in large eddy simulation. "Al.I.Cuza" New equations for the description of motion of viscous incompressible uids and solvability in the large of boundary value problems for them. The Mathemetical Theory of Viscous Incompressible Flow. A nonlinear Weak imposition of boundary conditions in the Stokes and Navier- Stokes equation Analysis of the K-Epsilon Turbulence Model On elliptic partial di New perspectives on boundary conditions for large eddy simulation. Mathematical principles of classical uid mechanics. General circulation experiments with the primitive equations. Comparison of some upwind-biased high-order formulations with a second-order central-dierence scheme for time integration of the incompressible Navier-Stokes equations Problemes Mathematiques en Plasticite. --TR --CTR Volker John, An assessment of two models for the subgrid scale tensor in the rational LES model, Journal of Computational and Applied Mathematics, v.173 n.1, p.57-80, 1 January 2005 Faisal A. Fairag, Analysis and finite element approximation of a Ladyzhenskaya model for viscous flow in streamfunction form, Journal of Computational and Applied Mathematics, v.206 n.1, p.374-391, September, 2007
navier-stokes equations;turbulence;large eddy simulation;finite element methods
588514
Superconvergence of the Local Discontinuous Galerkin Method for Elliptic Problems on Cartesian Grids.
In this paper, we present a superconvergence result for the local discontinuous Galerkin (LDG) method for a model elliptic problem on Cartesian grids. We identify a special numerical flux for which the L2-norm of the gradient and the L2-norm of the potential are of orders k+1/2 and k+1, respectively, when tensor product polynomials of degree at most k are used; for arbitrary meshes, this special LDG method gives only the orders of convergence of k and k+1/2, respectively. We present a series of numerical examples which establish the sharpness of our theoretical results.
Introduction . In this paper, we derive a priori error estimates of the Local Discontinuous Galerkin (LDG) method on Cartesian grids for the following classical model elliptic problem: @n where\Omega is a bounded domain of R d and n is the outward unit normal to its boundary we assume that the (d \Gamma 1)-measure of \Gamma D is non-zero. Recently, Castillo, Cockburn, Perugia and Sch-otzau [3] obtained the first a priori error analysis of the LDG method for purely elliptic problems. Meshes consisting of elements of various shapes and with hanging nodes were considered and general School of Mathematics, University of Minnesota, Vincent Hall, Minneapolis, MN 55455 (cockburn@math.umn.edu). Supported in part by the National Science Foundation (Grant DMS- 9807491) and by the University of Minnesota Supercomputing Institute. y Institut f?r Angewandte Mathematik, Universit?t Heidelberg, INF 293/294, 69120 Heidelberg, Germany (guido.kanschat@na-net.ornl.gov). This work was supported in part by the ARO DAAG55-98-1-0335 and by the University of Minnesota Supercomputing Institute. It was carried out when the author was a Visiting Professor at the School of Mathematics, University of Minnesota. z Dipartimento di Matematica, Universit'a di Pavia, Via Ferrata 1, 27100 Pavia, Italy (perugia@dimat.unipv.it). Supported in part by the Consiglio Nazionale delle Ricerche. This work was carried out when the author was a Visiting Professor at the School of Mathematics, University of Minnesota. x School of Mathematics, University of Minnesota, Vincent Hall, Minneapolis, MN 55455 (schoetza@math.umn.edu). Supported by the Swiss National Science Foundation (Schweizerischer Nationalfonds). 2 B. Cockburn, G. Kanschat, I. Perugia and D. Sch-otzau numerical fluxes were studied. It was shown that, for very smooth solutions, the orders of convergence of the L 2 -norms of the errors in ru and in u are k and k respectively when polynomials of degree at most k are used. On the other hand, Castillo [2] and Castillo, Cockburn, Sch-otzau and Schwab [4] proved that, for one- space dimension transient convection-diffusion problems, the order of convergence of the error in the energy norm is optimal, that is, k that the so-called numerical fluxes are suitably chosen. In this paper, we extend these results to the LDG method on Cartesian grids for the multi-dimensional elliptic model problem (1.1); we show that the orders of convergence in the L 2 -norm of the error in ru and respectively, when tensor product polynomials of degree at least k are used. Our proof of this super-convergence result is a modification of the analysis carried out in [3]; it takes advantage of the Cartesian structure of the grid and makes use of a key idea introduced by LeSaint and Raviart [10] in their study of the original DG method for steady-state linear transport. Since our analysis is a special modification of that of [3], in order to avoid unnecessary repetitions, we refer the reader to [3] for a more detailed description of the framework of our error analysis. The organization of this paper is as follows. In Section 2, we briefly display the LDG method in compact form, introduce the special numerical flux on Cartesian grids and present and discuss our main result. In Section 3, the detailed proofs are given and in Section 4, we present several numerical experiments showing the optimality of our theoretical results. We end in Section 5 with some concluding remarks. 2. The main results. In this section we recall the formulation of the LDG method and identify the special numerical flux we are going to investigate on Cartesian grids. Then we state and discuss our main results. As pointed out in the introduction, we refer to [3] for more details concerning the formulation of the LDG method. 2.1. The LDG method. We assume that the problem domain\Omega can be covered by a Cartesian grid. To define the LDG method, we rewrite our elliptic model problem (1.1) as the following system of first-order equations: \Gammar Next, we discretize the above problem on a Cartesian grid T . To obtain the weak formulation with which the LDG is defined, we multiply equations (2.1) and (2.2) by arbitrary, smooth test functions r and v, respectively, and integrate by parts over the d-dimensional rectangle K 2 T . Then we replace the exact solution (q; u) by its approximation (q N ; uN ) in the finite element space MN \Theta VN , where and of degree at most k in each variable on Kg: The LDG method on Cartesian grids for elliptic problems 3 The method consists in finding (q VN such that Z Z Z Z Z Z for all test functions (r; v) 2 S(K) d \Theta S(K), for all elements . The functions uN and b q N in (2.5) and (2.6) are the so-called numerical fluxes. These are nothing but discrete approximations to the traces of u and q on the boundary of the elements K and are defined as follows. Consider a face e of the d-dimensional rectangle K. If e lies inside the domain\Omega\Gamma we define uN - ffq N gg and, if e lies on the boundary of \Omega\Gamma g N on \Gamma N ; and b u := g D on \Gamma D ; Moreover, the stabilization parameter C 11 and the auxiliary parameter C 12 are defined as follows: where i is a positive real number and v is an arbitrary but fixed vector v with non zero components; see Fig. 3.1. 2.2. Error analysis on Cartesian grids. To state our main result, we need to recall some notation and to introduce new hypotheses. We restrict our analysis to domains\Omega such that, for smooth data, the solution u of problem (1.1) belongs to (\Omega\Gamma5 and such that when f is in L 2(\Omega\Gamma and the boundary data are zero, we have the elliptic regularity result k u Grisvard [8] or [9]. Since the domain\Omega will be triangulated by means of a Cartesian grid, the above requirements hold only if\Omega is a d-dimensional rectangle. We denote by hK the diameter of an element K, and set, as usual h := maxK2T hK . We denote by E I the set of all interior faces of the triangulation T , by ED the set of faces on \Gamma D , and by EN the set of faces on we assume that \Gamma e. The Cartesian triangulations we consider are regular, that is, if ae K denotes the radius of the biggest ball included in K, ae K Finally, we denote by EN ae\Omega a closed set containing the intersection between the and the set fx 0g. Moreover, we assume that the triangulation T is such that where K e denotes, from now on, an element containing the face e. 4 B. Cockburn, G. Kanschat, I. Perugia and D. Sch-otzau We are now ready to state our main result. Theorem 2.1. Assume that the solution (q; u) of (2.1)-(2.4) belongs to H k+1 H k+2 for k - 0; assume also that if the intersection between \Gamma N and fx non-empty, u belongs to W k+1;1 (EN ). Assume that the Cartesian grid T is shape-regular, (2.12), and that it satisfies the condition VN be the approximation of (q; u) given by the LDG method with k - 0 and numerical fluxes defined by (2.9), (2.11) and by (2.10). Then we have and where the constant C solely depends on i, k, d, oe and on the norms k u k k+2 and Several important remarks are in order before we prove this result in the next Section. Remark 2.2. This theorem is an extension to the bounded domain case of the corresponding result by Cockburn and Shu [7] for the LDG method for transient convection-diffusion problems. It is also an extension to the multi-dimensional case of the results obtained by Castillo, Cockburn, Sch-otzau and Schwab [4] in the one-space dimension case. The key ingredient of its proof is a super-convergence result of LeSaint and Raviart [10] used in their study of the original DG method for steady-state linear transport in Cartesian grids. Remark 2.3. Note that Theorem 2.1 holds true in the case approximate solution is piecewise constant. In [3], all the error estimates obtained for the corresponding LDG method on general grids are valid only for k - 1; moreover, no order of convergence is numerically observed for Remark 2.4. From an approximation point of view, the order of convergence in q, namely, k+1=2, is suboptimal by one half; however, it is confirmed to be sharp by our numerical experiments in Section 4. For general numerical fluxes and unstructured grids, an order of convergence in q of only k is obtained; see [3]. Remark 2.5. If we take the more general case e e are constants, we might conceive the possibility that a suitable tuning of the value of ff could improve the order of convergence in q. However, this is not true, as will be made clear in the proof of Theorem 2.1 displayed in the next section. See also [3] for other results about the influence of the value of ff on the orders of convergence of the general LDG method. Remark 2.6. In Theorem 2.1 an extra regularity condition on the exact solution u on the closed set EN containing part of the Neumann boundary is required. If this condition is dropped, and if 0g is not empty, only an order of convergence of k in the error in q can be proved by using our technique which represents a loss of 1=2. Note that whenever it is possible to choose v in such a way that regularity assumption on the exact solution is required. The LDG method on Cartesian grids for elliptic problems 5 3. Proofs. This section is devoted to the proof of Theorem 2.1. For simplicity, we consider only the case and\Omega rectangle; see Fig. 3.1. All the arguments we present in our analysis rely on tensor product structures and can be easily extended to the case d ? 2. \GammaE +Fig. 3.1. The Cartesian grid T and the auxiliary vector v used to define the numerical fluxes. To prove Theorem 2.1, we follow the approach used by [3]. Thus, we start, in Section 3.1, by briefly reviewing the setting of our error analysis. We proceed in Section 3.2, by introducing the projections \Pi and \Pi which generalize to several space dimensions the projections used by Castillo, Cockburn, Sch-otzau and Schwab [4] in their study of the LDG method for transient convection-diffusion problems in one- space dimension. Then, in Section 3.3, we derive the expressions of the functionals KA and KB needed in the setting of [3] to get error estimates. To do so, we make use of a super-convergence result essentially due to LeSaint and Raviart [10], and whose proof is presented in Section 3.4. The proof of Theorem 2.1 is completed in Section 3.5. 3.1. The framework of the error analysis. All the following results are collected from [3]. First, we start by reviewing that, by summation over all elements, the LDG method can be written in the compact form: Find (q such that for all (r; v) 2 MN \Theta VN , by setting 6 B. Cockburn, G. Kanschat, I. Perugia and D. Sch-otzau with a(q; r) := Z K2T Z ur Z e ds \Gamma Z e c(u; v) := Z e ds e2ED Z e C 11 uv ds: The linear forms F , G are defined by F (r) := e2ED Z e Z e2ED Z e ds Z e We also introduce the semi-norm j (q; u) j 2 A that appears in a natural way in the analysis of the LDG method and is defined as Z e ds e2ED Z e To prove error estimates for the LDG method, we follow [3] and introduce two func- tionals, KA and KB , which capture the approximation properties of the LDG method; the functionals are related to two suitably chosen projections \Pi and \Pi onto the FE spaces MN and VN , respectively. Namely, we require KA and KB to satisfy for any (q; u); (\Phi; 2(\Omega\Gamma3 and for any (r; v) 2 MN \Theta VN and (q; u) 2 H By Galerkin orthogonality, all the error estimates can then be solely expressed in terms of KA and KB as can be seen in the following result. Lemma 3.1 ([3]). We have A (q; u; q; u) +KB (q; u): Furthermore, KA (q; u; \Phi; ') with ' denoting the solution of the adjoint problem @n and The LDG method on Cartesian grids for elliptic problems 7 3.2. Projections. In this section we define the projections \Pi and \Pi we are going to use to prove Theorem 2.1 and list their properties. To this end, we start by introducing one-dimensional projections. Let I be an arbitrary interval, and let P k (I) be the space of the polynomials of degree at most k on I . We denote by - the L 2 (I)-projection onto P k (I), i.e., for a function w 2 L 2 (I) the projection -w is the unique polynomial in P k (I) satisfying Z I Furthermore, for w we define the projections - \Sigma w 2 P k (I) by the following I On a rectangle we define the following tensor product operators: with the subscripts indicating the application of the one-dimensional operators - or - \Sigma with respect to the corresponding variable. Finally, we define the projections \Pi and \Pi as In our error analysis, we use key properties of these projections displayed in the following result. Lemma 3.2. With the notation indicated in Figure 3.1, we have Z Z 2: We also need several approximation results which we gather in the lemma below. Lemma 3.3 (Cf. [5]). Let Furthermore, for any edge e i parallel to the x i -axis, we have Finally, if u 8 B. Cockburn, G. Kanschat, I. Perugia and D. Sch-otzau 3.3. The functionals KA and KB . In this subsection, we obtain the functionals KA and KB introduced in Section 3.1. We consider the stabilization parameter C 11 , defined by (2.14), in order to highlight the fact that any choice of ff 6= 0 in (2.14) deteriorates the rates of convergence of the estimates of Theorem 2.1. In [3, Corollary 3.4], KA has been investigated for general projection operators \Pi and \Pi satisfying the approximation results in Lemma 3.3 with 1. Thus, we we just report here the final result. Lemma 3.4 ([3]). Let u to be given by (2.14). Then, if we set the approximation property (3.2) holds true with KA (q; u; \Phi; \Theta Furthermore, in the particular case where (\Phi; there holds KA (q; u; q; \Theta In [3], the functional KB was only studied in the case where \Pi and \Pi are L 2 - projections. Next, we show that a better result for KB can be obtained on Cartesian grids for the projections defined by (3.4) and the numerical fluxes defined by (2.11). To obtain such a result, we use the following standard inverse inequality. Lemma 3.5 (Cf. [5]). There exists a positive constant C solely depending on k, d and oe such that for all s 2 MN we have for all K 2 T , e being any side of K. We set the sides of \Gamma. We are now ready to state our main lemma. Lemma 3.6. Let u to be given by and let \Pi and \Pi be the operators defined by (3.4). Then, for any (r; w) 2 MN \Theta VN , the approximation property (3.3) holds true, with KB given by where the constant C solely depends on k, d and oe. Proof. In order to be able to distinguish the many parts of \Gamma and facilitate the proof of the above result, we introduce the following notation: and define E these boundaries are indicated in Fig. 3.1. We set - q := The LDG method on Cartesian grids for elliptic problems 9 and estimate each of the forms separately. a. Estimate of T 1 . We have K2T Z K2T K2T K2T b. Estimate of T 2 . We can write K2T Z Z e ds e2ED Z e ds Taking into account the definition of the fluxes in (2.11) and the properties of the projection \Pi in Lemma 3.2, we conclude that Z Z e Z e Consequently, Z e ds Multiplying and dividing each term of the sum by C2 11 , and using the approximation properties of \Pi, we have Z e ds 0;e 0;e s+1;Ke Note that we have used the shape-regularity assumption (2.12) to bound C \Gamma1 11 by Ke . c. The estimate of T 4 . We have Z e ds e2ED Z e ds e2ED 0;e e2ED 0;e K2T 0;e K2T G. Kanschat, I. Perugia and D. Sch-otzau d. Estimate of T 3 . This estimate cannot be obtained as easily as the previous ones since it is here that the key idea introduced by LeSaint and Raviart [10] has to be suitably applied. We start by writing K2T Z Z e (ff- u gg +C 12 \Delta [[- u ds \Gamma Z e ds K2T Z K2T I Z e (ff- ds Z e ds Again with (2.11), we see that the contribution of an interior element K to this expression is Z Z ds \Gamma Z where the superscript 'out' denotes the traces taken from outside K. Since u out and [\Piu] out for the corresponding one-dimensional projection - i , this contribution can be written as Z Z ds \Gamma Z ds Z ds \Gamma Z For boundary elements, we add and subtract corresponding terms to obtain K2T Z e ds Z e ds ds K2T Z e ds Z e ds Z e with ZK (r; u) defined in (3.5). We start by bounding the contributions to T 3 stemming from a boundary edge e parallel to the x i -axis, 2. Since u 2 H s+2 implies see [9], by the property (3.3) and the inverse inequality in Lemma 3.5, we get Z ds - The LDG method on Cartesian grids for elliptic problems 11 Here, K e i denotes again the element containing the edge e i . Consequently, the global contribution to T 3 of the boundary edges belonging to E i n can be bounded by 0;Ke For the edges e in EN "E \Gamma , we have to use a different argument. Thus, by Lemma 3.5, we have Z e ds -k - u k L 1 Hence, by the Cauchy-Schwarz inequality, sup and so Finally, we estimate the contribution ZK (r; u), by using the following super-convergence result, essentially due to LeSaint and Raviart [10], whose proof is postponed to Section 3.4. Lemma 3.7. Let ZK (r; u) be defined by (3.5). Then we have for s - 0 By combining the result of Lemma 3.7 with the above estimates of the contribution of boundary edges, we obtain Conclusion. The result now follows by simply gathering the estimates for T i , obtained above. This completes the proof. 12 B. Cockburn, G. Kanschat, I. Perugia and D. Sch-otzau 3.4. Proof of Lemma 3.7. We can write where Z Z Z and Z Z Z The proof of the approximation results for Z K;1 and Z K;2 are analogous; therefore, we just present the one for Z K;1 , essentially following the same arguments as in [10]. First, we consider Z K;1 on the reference square (\Gamma1; 1) 2 . We claim that To prove (3.6), fix r are polynomial preserving operators, holds true for every u 2 Q k (K). Therefore, we just have to consider the cases . Let us start with 1 . On 2 we have 1, and on 2 we have is a polynomial of degree at most k \Gamma 1 in x 1 , we obtain Z Z Thus, Z K;1 (r 1 . In the case u(x 1 2 , we integrate by parts and obtain Z Z Z Z due to the special form of u, we conclude that Z K;1 (r 2 . This completes the proof of (3.6). For fixed r 1 2 Q k (K), the linear functional u 7! Z K;1 (r 1 ; u) is continuous on H s+2 (K) with norm bounded by Ckr 1 k 0;K . Due to (3.6), it vanishes over P s+1 (K) for k. Hence, by applying Bramble-Hilbert's Lemma (see [6, Lemma 6], for instance), we obtain for This proves the assertion on the reference element (\Gamma1; 1) 2 . The general case follows from a standard scaling argument. The LDG method on Cartesian grids for elliptic problems 13 3.5. Proof of Theorem 2.1. If the exact solution of our model problem, (q; u), belongs to H k+1 2 \Theta H k+2 KA (q; u; q; u) - h 2k+1+ff kuk 2 and with . The estimate of the error j follows now from Lemma 3.1. Notice that gives the best order of convergence in h equal to k . Our assumptions on the domain imply that the solution ' of the adjoint problem in Lemma 3.1 belongs to H 2 (\Omega\Gamma and that we have k'k 2 - Ck-k 0 ; see [8, 9]. Hence, we conclude that The estimate of ku \Gamma uN k 0 thus follows from Lemma 3.1. Notice that again the best order of convergence in h which is k + 1. 4. Numerical Experiments. In this section, we display a series of numerical experiments showing the computed orders of convergence of the LDG method; we show (i) that the orders given by our theoretical results are sharp, (ii) that they can deteriorate when the stabilization parameter C 11 is not of order one, (iii) that the exact capture of the boundary conditions induces an unexpected increase of 1in the order of convergence of the gradient, and (iv) that the orders of convergence are independent of the dimension. In all experiments, we estimate the orders of convergence of the LDG method as follows. We consider successively refined Cartesian grids T ' , ' - 0, consisting of 2 d ' uniform d-dimensional cubes with corresponding mesh size 2 \Gamma'+1 ; we present results in two and three space dimensions. If e(T ' ) denotes the error on the '-th mesh, then the numerical order of convergence is computed as follows: log The results have been obtained with the object-oriented C++ library deal.II developed by Bangerth and Kanschat [1]. 4.1. The sharpness of the orders of convergence of Theorem 2.1. We consider the two-dimensional model problem (1.1) on the f and boundary conditions chosen in such a way that the exact solution is given by We consider two cases: In the first, we impose inhomogeneous Dirichlet boundary conditions on the whole boundary, and in the second, we also impose inhomogeneous Neumann boundary conditions on the edge f\Gamma1g \Theta (\Gamma1; 1). The results are contained in Tables 4.1 and 4.2 where the numerical orders of convergence in the L 2 - and L 1 -norm in u, q 1 and q 2 of the LDG method with Q k elements for are shown. We take C and the coefficients C 12 as in (2.11) with 14 B. Cockburn, G. Kanschat, I. Perugia and D. Sch-otzau In Table 4.1, we report the results for Dirichlet boundary conditions imposed on the whole boundary. Note that, because of the symmetry of the problem, the orders of convergence are exactly the same for q 1 and q 2 . For we see the optimal order of convergence of 1 in the L 2 -norm of the error of both u and q; note that Theorem 2.1 predicts an order of convergence of 1 2 only for q. However, for k - 1 the L 2 -rates are of order k +1 in u and k 2 in q, in full agreement with Theorem 2.1. The orders on convergence in the L 1 -norm of the error in u and q appear to be k respectively. The results displayed in Table 4.2 are those for the case of inhomogeneous Neumann boundary conditions on part of the boundary. We see that the orders of convergence in this case are the same as the ones in the previous case. Thus, the above experiments show that the orders of convergence given by Theorem 2.1 are sharp. Table Orders of convergence for the LDG method with C 11 element 6 0.9683 0.9624 0.9724 0.3856 6 2.9661 2.8316 2.4678 1.9658 6 3.9661 3.8249 3.4676 2.9724 4.2. The effect of the choice of C 11 . Next, we test the effect of the choice of the coefficients C 11 on the orders of convergence of the LDG method. We consider the same problem as in the previous experiments, case use Q 1 and Q 2 elements. We only show the numerical orders of convergence for the finest grids. The LDG method on Cartesian grids for elliptic problems 15 Table Orders of convergence of the LDG method with C 11 element 6 0.9795 0.9555 1.0303 0.9283 0.9954 0.6270 5 2.9563 2.9428 2.5042 1.9392 2.4631 1.9392 6 2.9770 2.8316 2.5044 1.9658 2.4806 1.9658 6 3.9805 3.8264 3.5024 2.9722 3.4815 2.9721 The results are displayed in Tables 4.3 and 4.4. We must compare all these results with those with C obtained in the first set of experiments. We see that when C 11 is of order h \Gamma1 , the order of convergence in u remains but the order of convergence in q degrades from k 2 to only k, as predicted by our analysis; see section 3.5. We also see that taking C at the outflow boundary and C 11 of order one elsewhere only results in a slight reduction of the L 1 -orders of convergence. In the remaining cases, we take C 11 to be of order h in all the domain and then in all but the outflow boundary where it is taken to be of order h \Gamma1 . We observe a slight degradation of all the orders of convergence. These results indicate that the best choice of the stabilization parameter C 11 for the LDG method is to take it of order one, as predicted by our analysis. 4.3. Piecewise polynomial boundary conditions. The purpose of these numerical experiments is to show that if the boundary data are piecewise polynomials of degree k, the order of convergence of the L 2 -norm of the error in q is optimal, that 1, and not only k 2 as predicted by Theorem 2.1 and shown to be sharp in sub-section 4.1. B. Cockburn, G. Kanschat, I. Perugia and D. Sch-otzau Table Orders of convergence of the LDG method with Q 1 elements. 1=h 5 1.9607 1.9550 1.1409 0.8816 6 1.9792 1.9057 1.1019 0.9366 1=h on elsewhere 6 1.9646 1.7914 1.4605 0.9268 6 1.8603 1.7887 1.4564 0.9701 1=h on elsewhere 6 1.8563 1.7887 1.4556 0.9319 Table Orders of convergence of the LDG method with Q 2 elements. 1=h 5 2.9555 2.9541 2.2223 1.8475 6 2.9754 2.9584 2.1685 1.9228 1=h on elsewhere 6 2.9634 2.7424 2.4663 1.9427 6 2.8240 2.5482 2.4656 1.9742 1=h on elsewhere 6 2.8211 2.4554 2.4643 1.9365 We consider two test problems. In the first, we take homogeneous Dirichlet boundary conditions and f such that the exact solution is In the second, we take piecewise quadratic Dirichlet boundary conditions and f such that the exact solution is The results of the first problem are reported in Table 4.5 where we can see that the optimal order of convergence of k for the L 2 - and L 1 -norms of the errors in both u and q are obtained; the results for are displayed. The results of the second problem are reported in Table 4.6, where we see that the optimal order of convergence of k for the L 2 - and L 1 -norms of the errors in both u and q are obtained for k - 2, as claimed. For 2, the order of convergence in the -norm of the error in q is 2 only which nothing but the order of convergence predicted by Theorem 2.1. To better understand this phenomenon, we plot the errors in q 1 for Q 1 and Q 2 elements in Figs. 4.1 and 4.2, respectively; the triangulation has 16 \Theta 16 elements and corresponds to the index We immediately see the oscillatory behavior of the error typical of finite element methods. In Fig. 4.1, we see that the error obtained with elements is bigger at the boundary than at the interior. This, together with the fact that the order of convergence in L 2 is 3 2 whereas the order of convergence in L 1 is only 1, suggests that the error at the boundary is a factor of order bigger The LDG method on Cartesian grids for elliptic problems 17 than the error at the interior of the domain. On the other hand, the behavior of the error with Q 2 elements is rather different, as can be seen in Fig. 4.2. Indeed, the error behaves in the same way at the boundary and at the interior; this is further confirmed by the fact that both the the order of convergence in L 2 and the one in L 1 are equal to k + 1. These experiments justify our contention that the optimal order of convergence in q can be reached if the boundary conditions are piecewise polynomials of degree k. Our theoretical analysis does not explain this phenomenon. Table Orders of convergence for the LDG method with C element 6 0.9456 0.9658 0.9662 0.9483 6 2.0213 1.9878 2.0003 1.9858 6 2.9815 3.0150 2.9855 3.0161 6 4.0247 3.9918 4.0041 3.9748 Table Orders of convergence for the LDG method with C 11 element 6 2.0015 1.9775 1.4976 1.0091 6 2.9815 3.0150 2.9855 3.0162 6 B. Cockburn, G. Kanschat, I. Perugia and D. Sch-otzau -2e-022e-026e-02Fig. 4.1. The error in the first component of the gradient for Fig. 4.2. The error in the first component of the gradient for The LDG method on Cartesian grids for elliptic problems 19 4.4. A three-dimensional example. In this experiment, we consider the model problem (1.1) on the three-dimensional We take Dirichlet boundary conditions and f such that the exact solution is The results are displayed in Table 4.7; the computation on level 5 with Q 2 did not fit into the computers available to us. We can see that the orders of convergence are similar to those obtained in the corresponding two-dimensional test problem in the previous sub-section, cf. Table 4.6. This gives an indication that the orders of convergence of the LDG method in three space dimension behave in the same way they do in the two-dimensional case. Table Orders of convergence for the LDG method with C element 3 2.9204 2.8642 5. Concluding remarks. In this paper we have shown that the LDG method on Cartesian grids and with a special numerical flux super-converges; the proof of this result is based on suitable defined projections \Pi and \Pi exhibiting a tensor product structure. This work extends the corresponding result by LeSaint and Raviart [10] for the DG method for linear hyperbolic problems and that by Castillo [2] and Castillo, Cockburn, Sch-otzau and Schwab [4] for the LDG method applied to the one-dimensional transient convection-diffusion. Extensions of this work to more general elliptic and both steady and transient convection-diffusion problems can easily be made. --R Concepts for object-oriented finite element software - the deal An optimal error estimate for the local discontinuous Galerkin method An a priori error analysis of the local discontinuous Galerkin method for elliptic problems An optimal a priori error estimate for the hp-version of the local discontinuous Galerkin method for convection-diffusion prob- lems The finite element method for elliptic problems General Lagrange and Hermite interpolation in R n with applications to finite element methods The local discontinuous Galerkin finite element method for convection-diffusion systems Elliptic problems in nonsmooth domains On a finite element method for solving the neutron transport equa- tion --TR --CTR Paul Castillo, A review of the local discontinuous Galerkin (LDG) method applied to elliptic problems, Applied Numerical Mathematics, v.56 n.10, p.1307-1313, October 2006 Ilaria Perugia , Dominik Schtzau, On the Coupling of Local Discontinuous Galerkin and Conforming Finite Element Methods, Journal of Scientific Computing, v.16 n.4, p.411-433, December 2001 Ilaria Perugia , Dominik Schtzau, Thehp-local discontinuous Galerkin method for low-frequency time-harmonic Maxwell equations, Mathematics of Computation, v.72 n.243, p.1179-1214, July Bernardo Cockburn , Chi-Wang Shu, RungeKutta Discontinuous Galerkin Methods for Convection-Dominated Problems, Journal of Scientific Computing, v.16 n.3, p.173-261, September 2001
finite elements;elliptic problems;discontinuous Galerkin methods;cartesian grids;superconvergence
588515
Superlinear Convergence of Conjugate Gradients.
We give a theoretical explanation for superlinear convergence behavior observed while solving large symmetric systems of equations using the conjugate gradient method or other Krylov subspace methods. We present a new bound on the relative error after $n$ iterations. This bound is valid in an asymptotic sense when the size $N$ of the system grows together with the number of iterations. The bound depends on the asymptotic eigenvalue distribution and on the ratio $n/N$. Under appropriate conditions we show that the bound is asymptotically sharp.Our findings are related to some recent results concerning asymptotics of discrete orthogonal polynomials. An important tool in our investigations is a constrained energy problem in logarithmic potential theory.The new asymptotic bounds for the rate of convergence are illustrated by discussing Toeplitz systems as well as a model problem stemming from the discretization of the Poisson equation.
Introduction . The Conjugate Gradient (CG) method is widely used for solving systems of linear equations with a positive denite symmetric matrix A. The CG method is popular as an iterative method for large systems, stemming e.g. from the discretisation of boundary value problems for elliptic PDEs. The rate of convergence of CG depends on the distribution of the eigenvalues of A. A well-known upper bound for the error e n in the A-norm after n steps is where e 0 is the initial error and the condition number is the ratio of the two extreme eigenvalues of A. In practical situations, this bound is often too pessimistic, and one observes an increase in the convergence rate as n increases. This phenomenon is known as superlinear convergence of the CG method. It is the purpose of this paper to give an explanation for this behavior in an asymptotic sense. The error bounds are derived from the following polynomial minimization prob- lem. For any compact set S R, we dene p2Pn where Pn is the set of polynomials p of degree at most n with 1. The standard convergence analysis of the CG method leads to Laboratoire d'Analyse Numerique et d'Optimisation, UFR IEEA { M3, UST Lille, F-59655 Villeneuve d'Ascq CEDEX, France, e-mail: bbecker@ano.univ-lille1.fr y Department of Mathematics, Katholieke Universiteit Leuven, Celestijnenlaan 200 B, B-3001 Leuven, Belgium, e-mail: arno@wis.kuleuven.ac.be 2 B. BECKERMANN AND A. B. J. KUIJLAARS where (A) is the spectrum of A. The usual way to analyze (1.3) is to include the spectrum into a 'continuous' compact set S, so that The quantity En (S) can be estimated using notions from potential theory, since lim log En where gS (z) is the Green function for the complement of S with pole at 1. Thus one arrives at as an upper estimate for the error. For example, if one chooses the Green function evaluated at 0 is known to be min min which leads to the asymptotic estimate in terms of the condition number which is in agreement with (1.1). We refer the reader to the survey paper [DTT98] of Driscoll, Toh, and Trefethen for an excellent account on the interaction between iterative methods in Numerical Linear Algebra and logarithmic potential theory. The estimate (1.7) is typically accurate at early stages of the iteration. The reason for this is that for small n, a polynomial p 2 Pn that is small on (A) has to be uniformly small on the full interval [ min ; max ] as well. When n gets larger, however, a better strategy for p is to have some of its zeros very close to some of the eigenvalues of A, thereby annihilating the value of p at those eigenvalues, while being uniformly small on a subcontinuum of S only. Then the right-hand side of (1.7) may become a great overestimation of the error. This eect is the reason for the superlinear convergence behavior of the CG iteration, observed in practical situations. As an illustration we look at the case of a matrix A with 100 equally spaced 100g. The error curve computed for this example is the solid line in Figure 1. See also [DTT98, page 560]. The classical error bound given by (1.1) with is the straight line in Figure 1. For smaller values of n, the classical error bound gives an excellent approximation to the actual error. The other curve (the one with the dots) is the new asymptotic bound for the error that we nd in Corollary 3.2 below. This curve follows the actual error especially well in the region of superlinear convergence (for n 40). The phenomenon of superlinear convergence has been understood for compact operators, see [Win80, Mor97, Nev93]. Also, the above heuristic for the convergence behavior of CG for large matrices has been discussed and further analyzed by several authors [VSVV86, Gre89, SlVS96, DTT98]. To our knowledge, a formula for SUPERLINEAR CONVERGENCE OF CONJUGATE GRADIENTS 3 -5log(rel. iterations CG error energy norm Classical bound Our asymptotic bound Fig. 1. The CG error curve versus the two upper bounds for the system solution x, and initial residual r Our new asymptotic bound is given in formula (3.11). the relative error improving (1.7) and explaining the superlinear convergence is still lacking. Our goal in this paper is to provide a better understanding of the superlinear convergence of CG iteration, and in particular to explain the form of the error curve as seen in Figure 1. We will argue that for a large NN matrix A, the error En in the polynomial minimization problem (1.2) is approximatelyn log En ((A)) < t decreasing family of sets, depending on the distribution of the eigenvalues of A. The sets S() have the following inter- pretation: S() is the subcontinuum of [ where the optimal polynomial of degree uniformly small. From (1.3) and (1.8) we nd the improved estimate with Note that t depends on n, since . As the sets S() are decreasing as increases, their Green functions g S() (0), evaluated at 0, increase with . Hence the numbers t decrease with increasing n (see also Remark 2.3 below), and this explains the eect of superlinear convergence. 4 B. BECKERMANN AND A. B. J. KUIJLAARS The phenomenon of superlinear convergence may also occur for other Krylov subspace methods applied to a system A is no longer symmetric positive denite. For instance, for symmetric but not positive denite A one usually employs iterative methods like MINRES. Also, the method GMRES may be applied in case of a general matrix A. Supposing that A is diagonalizable, i.e., with D a diagonal matrix containing the (possibly complex) eigenvalues of A, the nth relative residual may be bounded for these methods by (see, e.g., [Saa96, Proposition 6.15]). In particular, for symmetric or more generally normal matrices, V is unitary, and thus again we may give bounds for the relative residual by describing the (asymptotic) behavior of En ((A)). Indeed, for the ease of presentation our results are stated for real spectra, but they remain equally valid for complex spectra (see also Remark 2.4 below). The paper is organized as follows: In x2, we describe the (sequence of) matrices AN under considerations. We explain the potential-theoretic origin of our sets S(t), and establish in Theorem 2.1 the estimate (1.8). Under some stronger assumption concerning the clustering of eigenvalues, we prove in Theorem 2.2 that estimate (1.8) is sharp. Section 3 contains a description of eigenvalue distributions where our sets are explicit intervals. Subsequently, we give an analysis of the plot of Figure 1. In x4 it is shown that our assumptions are valid for a large class of symmetric positive denite Toeplitz matrices. Our ndings are illustrated by considering some Toeplitz matrix occurring in time series analysis. The discretized two-dimensional Poisson equation on a uniform grid is analyzed in x5. Finally, a lemma used in the proof of Theorem 2.1 is proved in the appendix. We should mention that our results concerning the two applications above are more of theoretical nature since in the present paper neither preconditioning nor nite precision arithmetic are considered. The main aim of this paper is to illustrate that some recent results in logarithmic potential theory may help to understand better a classical phenomenon in Numerical Linear Algebra (see also [BeSa98, Kuij99]). 2. The main result. Properly speaking, the concept of superlinear convergence for the CG method applied to a single linear system does not make sense. Indeed, in the absence of roundo errors, the iteration will terminate after N steps if N is the size of the system. Also the notion that the eigenvalues are distributed according to some continuous distribution is problematic when considering a single matrix. Therefore we are not going to consider a single matrix A, but instead a sequence (AN ) N of symmetric positive denite (or more generally invertible symmetric) matri- ces. The matrix AN has size NN , and we are interested in asymptotics for large N . These matrices need to have an asymptotic eigenvalue distribution. By this we mean that there exists a positive Borel measure with compact support supp() such that the following condition is satised. Condition (i) The spectra (AN ) are all contained in a xed compact set S R, and for every function f continuous on S we have lim Z SUPERLINEAR CONVERGENCE OF CONJUGATE GRADIENTS 5 This condition is equivalent to the weak convergence of the normalized eigenvalue counting measures N dened by where is the unit point mass at , to . As all the AN have spectra contained in S, the measure is supported on S. The total mass kk is at most one, and it can be strictly less than one if the matrices AN have many coinciding eigenvalues. Note that in the sum in (2.1) each in (AN ) is taken only once, regardless of its multiplicity (see also Remark 2.5 below). For the use of the potential theory in what follows, we need to impose a condition on . The logarithmic potential of a Borel measure with compact support is the function U Z logj This is a superharmonic function on C taking values in (1;1]. In particular it is lower semi-continuous. We refer the reader to [Ran95, SaTo97] for detailed accounts of logarithmic potential theory. Our assumption is the following. Condition (ii) The logarithmic potential U of the measure from (i) is a continuous real-valued function on C . The condition (ii) is not very restrictive. For example, if is absolutely continuous with respect to Lebesgue measure with a bounded density then (ii) is satised. It is also satised if the density has only logarithmic-type or power-type singularities at a nite number of points. On the other hand, condition (ii) is not satised if has point masses. A consequence of (ii) is that for any measure satisfying , the potential U is also continuous. Indeed, U is lower semi-continuous, and since continuous and U lower semi-continuous, U is also upper semi-continuous; hence U is continuous. There is a third condition we impose on the sequence (AN ) N . Condition (iii) The limit (2.1) also holds for Notice that (iii) follows from (i) if 0 62 S, or even if the (in modulus) small eigenvalues of AN do not approach zero too fast. If (iii) would not hold, then the matrices AN are too ill-conditioned and the estimate (2.9) given below may very well fail. In many practical applications, the family (AN ) N of matrices appears as discretizations of a continuous operator, and then (i){(iii) are natural conditions, see for instance the discussion in x4 and x5 below. The sets S(t) that were announced in (1.8) depend only on the asymptotic eigenvalue distribution . They are determined by the solution of an energy minimization problem which we describe now. The logarithmic energy of a Borel measure with compact real support is the double integral Z U () ZZ logj For every t 2 (0; kk), we dene the class is a Borel probability measure on R : 6 B. BECKERMANN AND A. B. J. KUIJLAARS and we let t be the unique measure minimizing the logarithmic energy (2.2) in the class M(t; ) (compare [Rak96], [DrSa97, Theorem 2.1]). Thus The minimizer t depends on t and . The minimization problem (2.3) is a constrained problem, since measures in M(t; ) are dominated by the constraint =t. It is known that the minimizer t is characterized by the following variational conditions associated with (2.3). There exists a constant F t such that see [Rak96, Theorem 3] and [DrSa97, Theorem 2.1]. From these variational conditions one obtains Finally, the sets S(t) which are crucial in our ndings are dened by The extremal problem (2.3) has been studied before in connection with the asymptotic behavior of discrete orthogonal polynomials, see, e.g., [Rak96], [DrSa97], [KuVA98], [Beck98], [BeSa98], and [Joh99]. In particular, the monic analogue of (1.2) is covered by these results, that is, the study of where P n denotes the class of monic polynomials of degree n. Notice that if S [0; 1) (as for instance for symmetric positive denite matrices), then E n (S) and En (S) are realized (up to scaling) by the same polynomial, namely the generalized Chebyshev polynomial. Our main result is the following. Theorem 2.1. Let (AN ) N be a sequence of symmetric invertible matrices, AN of size N N , satisfying the conditions (i), (ii) and (iii) for some measure . Let the measures t , the constants F t , and the sets S(t) be dened by (2.3), (2.4){(2.5), and (2.7), respectively. Then for t 2 (0; kk), we have lim sup n=N!tn log En ((AN Proof. Let t 2 (0; kk), and let depend on N in such a way that Our goal is to construct for every large N a polynomial pN in Pn which is suciently small on (AN ), so as to obtain the estimate (2.9). We x > 0, and dene SUPERLINEAR CONVERGENCE OF CONJUGATE GRADIENTS 7 Since U t is a continuous function (cf. the discussion following condition (ii)), the set K is closed. It is disjoint from S(t) because of (2.4) and (2.7). Thus (K By choosing a smaller if necessary, we may assume that (@K possible to nd for every large N , a set ZN such that (3) for all continuous functions f , lim Z For the proof that this is indeed possible, we refer to Lemma A.1 in the Appendix. We write for N large, Y Then pN 2 Pn by property (1). We want to estimate max 2(AN ) jp N ()j. Let be such that Since pN vanishes on (AN ) \ K by property (2) and the denition (2.13), we have and the latter is a bounded set. Passing to a subsequence, if necessary, we may assume that the sequence (N ) converges as N !1 with limit We have by (2.13),n log jp N (N )j =n log log jj: (2.16) From property (3), we have that the normalized counting measures of ZN , i.e., converge in weak sense to t t . The principle of descent, see [SaTo97, Theorem I.6.8], and (2.15) then imply that U N (N t, this gives lim sup log 8 B. BECKERMANN AND A. B. J. KUIJLAARS and thus by (2.15) lim sup log The principle of descent also implies By property (2) we have N N , where N is the normalized counting measure of (AN ). Since N ! by condition (i), we nd that ( N N ) N is a sequence of positive measures that converges to t t in weak sense. Applying the principle of descent once more, we obtain U N N (0): (2.19) Also the condition (iii) gives U The relations (2.18){(2.20) easily imply that U N (0) and this is equivalent to lim log Combining (2.16) with (2.17) and (2.21), we obtain lim sup log jp N (N )j F t By (2.14) and the denition (1.2) of En , we then see lim sup En ((AN The number > 0 can be chosen arbitrarily close to 0. Hence (2.9) follows. To obtain (2.10) we need to show that To establish this and the inclusion property claimed in the introduction, we recall the connection of the constrained minimization problem (2.3) with the energy problem in the presence of an external eld. For a SUPERLINEAR CONVERGENCE OF CONJUGATE GRADIENTS 9 continuous function Q sucient growth at 1, the logarithmic energy of a measure in the presence of the external eld Q is I ZZ log 1 Z The minimizer Q s for the extremal problem Z exists and is unique if log jj (2.24) and it is characterized by the conditions U Q s U Q for some constant G s , see, e.g., [SaTo97, Theorem I.1.3]. Buyarov and Rakhmanov [BuRa99, Theorem 2] proved the following formula for Q s where !S is the equilibrium measure for the set S. In fact, the authors consider external elds where the limit on the right-hand side of (2.24) equals +1, and s 2 (0; +1). However, from their proof (see the last paragraph of [BuRa99, Section 2]) it becomes clear that (2.27) remains valid as long as (2.24) holds. Now, in our situation with the constraint , we take By comparing the conditions (2.4){(2.5) with (2.25){(2.26) we can easily check that for s; t > 0 with s In particular, ( Q converges in weak sense to for s ! kk. Then the Buyarov{ Z t! S() d: (2.29) From (2.29) we obtain the inequality t t for < t, and thus (2.23) holds. In order to show (2.22), notice that the Green function is connected with the potential of the equilibrium measure by the formula A. B. J. KUIJLAARS where cap denotes the logarithmic capacity. Combining this with (2.29), we obtain for Z tlog cap (S()) d: (2.30) For 2 S(t), the left-hand side of (2.30) vanishes according to (2.4). Also, by (2.23), each 2 S(t) belongs to S() for all < t, so that the integral in (2.30) involving the Green functions vanishes for 2 S(t). Consequently, Z tlog cap (S()) d; and the equation (2.22) follows from (2.30). This completes the proof of Theorem 2.1. Under additional conditions the inequality (2.9) can be improved to give equality lim n=N!tn log En ((AN These additional conditions are related to the separation of the eigenvalues. If many eigenvalues are very close to each other then the inequality (2.9) may be strict. For the extremal problem (2.8), various separation conditions were considered by Rakhmanov [Rak96], Dragnev{Sa [DrSa97], Kuijlaars{Van Assche [KuVA98], and Beckermann [Beck98], see also [KuRa98] for a survey. If one of these conditions holds in the present situation, the limit (2.31) can be proved. Indeed, according to Theorem 2.1, we only require a sharp lower bound for En ((AN )). For sets S with positive capacity, lower bounds for En (S) are usually obtained by applying the Bernstein-Walsh inequality. In our discrete setting, some analogue of the Bernstein-Walsh inequality in terms of the extremal measure t exists, see [KuVA98, Lemma 8.1 and Corollary 8.2] and [Beck98, Theorem 1.4(c)], implying (2.31). We will give here a proof using the separation condition of Beckermann [Beck98], which was rst conjectured by Rakhmanov [KuRa98]. For a nite subset Z C , we introduce I (Z) :=(#Z) 2 logj which may be thought of as the discrete energy of a system of #Z particles each having a charge 1=#Z. Beckermann's condition is: Condition (iv) With I() as in (2.2) we have lim I ((AN It can be shown that lim inf I ((AN already follows from condition (i). Notice also that the separation conditions of [Rak96, DrSa97] imply (iv). SUPERLINEAR CONVERGENCE OF CONJUGATE GRADIENTS 11 It is not dicult to prove using for instance [Beck98, Lemma 2.2(b)] that condition (iv) is equivalent to the fact that I (XN whenever (XN ) is a sequence of sets satisfying XN (AN ) for each N , lim(#XN )=N In this form the condition (iv) will be used. Theorem 2.2. Suppose that the assumptions of Theorem 2.1 are satised and that in addition the condition (iv) holds. Then for every t 2 (0; kk), the limit (2.31) holds. Proof. Let t 2 (0; kk). As in the proof of Theorem 2.1 we assume that depends on N in such a way that n=N ! t. For every N 2 N, let N be a set of n points in (AN ). That is, N has denoted by maximizes the product Y among all n 1-point subsets of (AN ). Equivalently, N minimizes the discrete energy I (N ). Since N (AN ), it is clear that En ((AN Our rst goal is to show that the normalized counting measures of the Fekete points tend to t , that isn as N ! 1. Since the sets N are all contained in the compact S, Helly's theorem asserts that from any subsequence of the sequence of normalized counting measures of the Fekete points, we may extract a further subsequence having a weak limit (which clearly is an element of M(t; )). Our claim (2.35) follows by showing that . According to (2.33), we nd that along an appropriate subsequence we then have I (N Let (ZN ) N be a sequence of sets satisfying It follows from Lemma A.1 in the Appendix that such a sequence exists. Again by (2.33), we nd that I (ZN 12 B. BECKERMANN AND A. B. J. KUIJLAARS by (2.3), and I (ZN ) I (N ) by the denition of Fekete points, we may conclude that I( by the uniqueness of the minimizer in (2.3). This proves the claim (2.35). Next, we dene for N 2 N and Y Then P k;N has degree n, and any polynomial p 2 Pn can be written in the form a k with coecients a k satisfying k=0 a and a k P k;N (0) P k;N (0) ng be such that it maximizes P k;N (0) among all k 2 f0; ng. Then it follows from (2.38) that Since this holds for every p 2 Pn , we nd We write shorter ~ with normalized counting measure SUPERLINEAR CONVERGENCE OF CONJUGATE GRADIENTS 13 Because of (2.35) we see that (N ) has the weak limit t for N ! 1. From the principle of descent [SaTo97, Theorem I.6.8] it follows that lim sup log jP kN ;N Also, we will show below that lim inf log jP kN ;N ( kN ;N )j F t (2.41) (compare with [Beck98, Lemma 2.6]). Combining (2.40) and (2.41) with (2.34) and (2.39), we may conclude that lim inf log En ((AN log which in view of Theorem 2.1 is the inequality required for the proof of Theorem 2.2. Finally, in order to establish (2.41), we note that by the denition of Fekete points we have for every 2 (AN ), Taking logarithms, and adding the inequalities for 2 (AN N , we obtain log jP kN ;N ()j and therefore log jP kN ;N ( kN ;N )j 1 log 1 One easily veries that the right-hand of (2.42) side equals2 which according to (2.33) converges to2t 2 Z where for the last equality we have used the variational condition (2.4). Since (#(AN assertion (2.41) follows from (2.42), and Theorem 2.2 is proved. Remark 2.3. We have shown in Theorem 2.1 that, for n; N !1, the quantity log En ((AN )) is asymptotically bounded above by log n 14 B. BECKERMANN AND A. B. J. KUIJLAARS and this bound is sharp (under some additional assumptions) according to Theorem 2.2. This conrms our claims (1.9) and (1.10) of the introduction. The graph of Nt log t for xed N and varying is drawn in the plots of Figures 1, 2, and 4. From (2.43) one sees that N t log t is dierentiable, up to at most a countable number of points, with derivative d Thus it follows that (2.43) is decreasing. Also because of (2.23) one sees that g S(n=N) (0) is increasing with n, and therefore the graph of (2.43) is concave. If S is a compact set containing all the spectra (AN ), then S(t) S, for every one easily checks that In other words, the bound (1.9) is sharper than (1.6). The equality holds if and only if which again is true if and only if the equilibrium distribution !S of S is less then or equal to =t. This may be translated by saying that, roughly, about tN out of the eigenvalues of AN are asymptotically distributed like the equilibrium distribution of S. Remark 2.4. Theorems 2.1 and 2.2 are equally valid for complex discrete sets (AN ), here supp() may be a subset of the complex plane. Indeed, the energy problems with constraint have been studied in a complex setting (see, e.g., [DrSa97]), and it is possible to show that the representation of F t U t in terms of Green functions remains equally true. Furthermore, all other arguments used in the proofs of Theorems 2.1 and 2.2 still apply for complex sets (AN ). As a consequence, our Theorems can also be used for bounding the relative residual while solving systems of linear equations with normal matrices AN via methods like MINRES or GMRES. Remark 2.5. In many applications (as for instance for symmetric Toeplitz ma- trices) it is dicult to know in advance the multiplicities of the eigenvalues, and one only obtains a measure ~ dened by a modication of condition (i) where multiple eigenvalues are counted according their multiplicities. We will refer to this modi- cation as condition (i'). Condition (ii) with this (possibly) new measure ~ will be called (ii'), and accordingly (iii) becomes (iii'), where again we count multiplicities. Notice that Theorem 2.1 remains valid if assumptions (i),(ii),(iii) are replaced by (i'),(ii'),(iii') (and is replaced by the new measure ~ ). In case of, e.g., real S, conditions (i') and (iii') have an interesting interpretation in terms of asymptotics of determinants: Indeed,N log jdet(I N AN log 1 (in this formula we count multiplicities), and from logarithmic potential theory we know that relation (2.1) holds for every function f continuous on S if and only if lim log jdet(I N AN SUPERLINEAR CONVERGENCE OF CONJUGATE GRADIENTS 15 for all 2 C n S. Furthermore, it is sucient that (2.44) holds for 2 where C has a nite accumulation point outside of S. Notice that condition (iii') may be rewritten as (2.44) with Finally, condition (i') is known to hold i lim trace for all 2 , C as above. Using (2.45), one can for instance easily show that condition (i') remains valid (with the same measure) if AN is perturbed by some matrix BN , with sup N kBN k < 1, and rank(BN )=N ! 0. 3. Equidistant eigenvalues. In order to apply Theorems 2.1 and 2.2 we have to calculate the sets S(t) from the eigenvalue distribution . This is a problem in itself. In general, the sets S(t) can have a complicated form. They may consist of several intervals, or even have a Cantor-like structure. The easiest case would be if all S(t) are single intervals. This would also be the most convenient case for the computation of the Green function at 0, since for an interval [a; b], we have a a Lemma 3.1. Suppose that is supported on the interval [a; b] and has a density w() with respect to Lebesgue measure. We write ~ w() := (a) Suppose ~ w is strictly increasing on (a; b). Then S(t) is an interval containing b for every t 2 (0; kk). More precisely, we have and is the unique solution in (a; b) of the equation Z r a r r w() d: (3.2) (b) Suppose ~ w is strictly decreasing on (a; b). Then S(t) is an interval containing a for every t 2 (0; kk). More precisely, we have w(b and w(b is the unique solution in (a; b) of the equation r r a r w() d: B. BECKERMANN AND A. B. J. KUIJLAARS (c) Suppose ~ w is symmetric with respect to the midpoint m := (a + b)=2 and strictly decreasing on (m; b). Let t 2 (0; kk). Then w(b and w(b is the unique solution in (0; (b a)=2) of the equation a w()d: Proof. (a) We consider as in the proof of Theorem 2.1 the external eld a log Let Q s be the extremal measure with external eld Q and normalization s 2 (0; kk) (cf. the paragraph preceding formula (2.26)). In [KuDr99, Theorem 2] it was proved that the support of Q s is an interval of the form [r; b] if Q and w are related as in (3.3) and if ~ w() increases on [a; b]. In [KuDr99] this is stated under the assumption that Q is dierentiable with a Holder continuous derivative. An inspection of the proof, however, shows that this assumption is not necessary. It was also assumed that s = 1. This is also not essential. Since s by (2.28), it thus follows that S(t) is an interval containing b for every t. We show that w(a+). For t ~ w(a+), we have from the fact that ~ w is strictly increasing, (b )( a) Thus the equilibrium measure ! [a;b] of [a; b], i.e., (b )( a) d belongs to the class M(; t). Since ! [a;b] minimizes the energy (2.2) among all probability measures on [a; b], it is then also the minimizer over M(; t). Thus and it follows from (3.4) that Conversely, if is a probability measure on [a; b] whose potential is constant on [a; b] by (2.4). This implies that Hence t! [a;b] and (3.4) holds. Thus t ~ w(a+). For the rest of the proof we assume that t 2 ( ~ w(a+); kk). Then with a < r(t) < b. From [DrSa97, Corollary 2.15] we know that t is the balayage (see [SaTo97, Section II.4]) of onto the interval [r; b]. Consequently, according to [SaTo97, Eqn. (II.4.47)], t t has the density w() if 2 (a; r), Z r a s SUPERLINEAR CONVERGENCE OF CONJUGATE GRADIENTS 17 For 2 (r; b), we rewrite the density as Z r a Since a < r < b and 0 v() w() < 1 for 2 (a; b), we must have lim In view of (3.5), the relation (3.2) follows. To show that there is only one r satisfying (3.2), we rewrite the right-hand side of (3.2) as Z r a r r Z r a ~ w() d Z =2~ where for the second equality, we used the change of variables Since ~ w is strictly increasing, it is then clear that (3.6) strictly increases for r 2 (a; b). This completes the proof of part (a). (b) The proof of part (b) is similar. (c) Part (c) follows using a quadratic transformation, compare with [Kuij99, Proof of Theorem 5.1]. Lemma 3.1 allows us to determine the sets S(t) in a number of situations. We consider here the case of equidistant eigenvalues. Suppose AN has N equidistant eigenvalues . Multiplying the matrix by a positive constant does not change the numbers En ((AN )). We multiply AN by 1=N and so we consider instead matrices with spectrum These matrices have an asymptotic eigenvalue distribution and the conditions (i){(iv) are satised. The explicit solution of the energy minimization problem (2.3) with given by (3.7) is due to Rakhmanov [Rak96]. We show how the sets S(t) can be determined from Lemma 3.1. The assumptions of Lemma 3.1(c) are clearly satised with a = 0, Therefore we have for 0 < t < 1, d =1=2 r B. BECKERMANN AND A. B. J. KUIJLAARS Thus and 11p Using (3.1) and (3.8) we nd after a little calculation log Hencet Z tlog d Theorem 2.2 and (3.10) then give the following result. Corollary 3.2. For every t 2 (0; 1) we have lim n=N!tn log En (f1; Corollary 3.2 gives the theoretical justication for our CG bound in the case of equidistant eigenvalues as given in Figure 1. Notice that, already for 100, the approximation for log En (f1; obtained by multiplying the right-hand side of (3.11) by n is quite accurate. To conclude this section, we note that Lemma 3.1(c) also applies to the case of ultraspherical eigenvalue distributions. The corresponding sets S(t) were determined in [Kuij99]. 4. Applications to Toeplitz matrices. Toeplitz matrices provide interesting examples for our results. Toeplitz systems arise in a variety of applications, such as signal processing and time series analysis, see [ChNg96] and the references cited therein. be an integrable function with Fourier coecients Z ()e ik d; We assume is bounded and not equal to a constant. The Nth Toeplitz matrix with symbol is given by TN SUPERLINEAR CONVERGENCE OF CONJUGATE GRADIENTS 19 Then TN () is a Hermitian matrix and it is well known that ( inf ; sup ) is the smallest interval containing the spectrum of TN () for every N , where inf and sup denote the essential inmum and essential supremum of , respectively. Thus, since is non-negative, all eigenvalues 1;N 2;N N;N of TN () are strictly positive, and the matrix TN () is positive denite. A classical result of Szeg}o, see e.g. [GrSz84, pp. 63-65], [BoSi99, Theorem 5.10 and Corollary 5.11], says that lim Z for every continuous function f on [ min ; max ]. It follows that the sequence (TN ()) satises the condition (i') (see Remark 2.5) with the measure given by Z Z Then is a probability measure and its support is equal to the essential range of . Another result of Szeg}o (see [Sz67, Eqn. (12.3.3)] or [GrSz84, p. 44 and p. 66]) is that lim Z log () d provided that satises the Szeg}o condition Z log Notice that this condition can be rewritten as U (0) < +1. It follows from (4.2), (4.3) that lim log det TN () Z log Z log d() 2 R; and the condition (iii') is satised. Consequently, for Toeplitz matrices TN () with non-negative, integrable, and bounded symbol and continuous real-valued potential U , the conditions (i')-(iii') are satised, and we may apply Theorem 2.1. We will discuss an example of Kac, Murdock and Szeg}o [KaMuSz53, p. 783] with 1). Toeplitz matrices with this symbol (or with a multiple of this arise as covariance matrices of rst-order autoregressive processes [ChNg96, Section 4.6.1]. The corresponding Fourier coecients are given by Suppose without loss of generality that > 0. Then the measure from (4.2) has support [a; b] where 20 B. BECKERMANN AND A. B. J. KUIJLAARS Since is even we have2 Z Making the substitution = (), we obtain after some calculations2 Z a d Thus the measure has density a < < b: (4.4) with respect to Lebesgue measure. From (4.4) it is easy to show that U is continuous, so that Theorem 2.1 applies in this case. Now we apply Lemma 3.1(b) in order to compute S(t). Notice that for r 2 [a; b), r r a r r d r a r Consequently, by Lemma 3.1(b), we have r a a if a < t < 1: In particular, we get from (1.10) and (3.1) the convergence rate log whereas for a < t < 1, we have t log a log a a d a log( a log d a log( a log It is quite interesting that, in the superlinear range, we obtain (up to some linear transformation) the same function as for equidistant nodes. Numerical experiments for the symmetric positive denite Toeplitz matrix T 200 of order 200 of Kac, Murdock and Szeg}o are given in Figure 2. The four dierent plots correspond to the choices 19=20g of the parameter. Notice that the CG error curve (solid line) of the last two plots is clearly aected by rounding errors leading to loss of orthogonality, whereas the GMRES relative residual curves (dotted line) behave essentially like predicted by our theory. In particular, the classical bound SUPERLINEAR CONVERGENCE OF CONJUGATE GRADIENTS 21 log( rel. residual log( rel. residual iterations iterations Fig. 2. The error curve of CG (solid line) and GMRES (dotted line) versus the classical upper bound (crosses) and our asymptotic upper bound (circles) for the system T 200 solution x, and initial residual r Here TN is the Kac, Murdock and Szeg}o matrix of x4, with parameter 19=20g. (1.1), (1.11) (crosses) does no longer describe correctly the size of the relative residual of GMRES for n 20 and 19=20g. Experimentally we observe that the range of superlinear convergence starts in the dierent examples approximately at the iteration indices 50, 30, 20, and 10, respectively. This has to compared with the predicted quantity N a which for the dierent choices of approximately takes the values 66, 40, 29, and 5, respectively. Though theses numbers dier slightly, we observe that the new bound (1.9), (1.10) re ects quite precisely the shape of the relative residual curve, and in particular allows to detect the ranges of linear and of superlinear convergence. Let us nally mention the Toeplitz matrices occuring in the context of the rst- order moving average process [ChNg96, Section 4.6.1], where the symbol is given by Here the eigenvalues are asymptotically distributed like the equilibrium distribution on and therefore there will be no superlinear convergence in 22 B. BECKERMANN AND A. B. J. KUIJLAARS this case. 5. The Model Problem. Consider the two dimensional Poisson equation for (x; y) in the unit square 0 < x; y < 1, with Dirichlet boundary conditions on the boundary of the square. The usual ve-point nite dierence approximation on the uniform grid leads to a linear system of size N N where After rescaling, the coecient matrix of the system may be written as a sum of Kronecker products Bm where . 1 mm and I m is the identity matrix of order m. It is well known and easy to verify that the eigenvalues of Bm are and that the eigenvalues j;k of AN are connected with the eigenvalues of Bm via most of the eigenvalues have multiplicity at least 2. Also, j;m+1 j and the eigenvalue 4 has multiplicity m. We suspect that which is conrmed by our numerical experiments presented below. To calculate the asymptotic distribution of the eigenvalues j;k as rst note that the eigenvalues k of Bm are in [0; 4] and have the asymptotic density (4 1. The asymptotic density of the is then given by the convolution of v with itself, i.e., SUPERLINEAR CONVERGENCE OF CONJUGATE GRADIENTS 23 Fig. 3. Bar chart of the eigenvalue distribution of the matrix A 1600 (without counting multi- plicities) resulting from discretizing the 2D Poisson equation on a uniform grid with points. The solid line corresponds to the asymptotic density function. where the factor 1=2 is added in accordance with the multiplicities of the eigenvalues of AN . The density w is symmetric around 4. For 2 (0; 4), we have from (5.4) and In (5.6) we put make the change of variables to obtain w(4 4x) =8 2 Z 11 By the Euler integral representation for hypergeometric functions, (5.7) is w(4 4x) =8 F (1=2; 1=2; is a Pochhammer symbol. It is interesting to observe that 4 2 w(4 4x) equals the complete elliptic integral of the rst kind, evaluated at Eqn. 7.3.2.(75)]. Since w is symmetric around 4 and the right-hand side of (5.8) is even in x, the formula (5.8) holds for 1 < x < 0 as well. B. BECKERMANN AND A. B. J. KUIJLAARS In the series in the right-hand side of (5.8) each term is clearly decreasing as increases. Thus w() is increasing for 2 (0; 4), which is also clear from Figure 3. Then also increases for 2 (0; 4), and therefore the assumptions of Lemma 3.1(c) are satised. From Lemma 3.1(c) we then conclude that, for every t 2 (0; 1=2), the set S(t) associated with w()d is with r 2 (0; w() d: (5.9) Putting in (5.9) we have x w(4 4x) dx: (5.10) Inserting the series (5.8) for w(4 4x) and interchanging integration and summation, we nd x For each k, the integral is easily transformed to a beta-integral, and it follows that x see also [Kuij99]. Inserting this in (5.11), we obtain This is a known series expansion for the arccos function see [PrBrMa90, Eqn. 7.3.2.(76)]. Inverting this we obtain the remarkably simple and so -5log(rel. iterations 2D Poisson discretized, 150 2 inner gridpoints CG error energy norm Classical bound Our asymptotic bound Fig. 4. The CG error curve versus the two upper bounds for the system AN resulting from discretizing the 2D Poisson equation on a uniform grid with points. We have chosen a random solution x, and initial residual As predicted by (5.15), we obtain superlinear convergence from the beginning. Notice that the classical upper bound for CG is far too pessimistic for larger iteration indices. Similar plots are obtained for other mesh sizes. Finally, after a small calculation using (1.10) and (3.1) we obtain the convergence factor log t =t Z tlog(tan( Notice that, for small t, the set S(t) of (5.14) approximately equals the set obtained for equidistant eigenvalues on [0; 8], compare with x3. This observation is in accordance with the behavior of the eigenvalues of AN at the endpoints of One might be curious about what CG error curve is obtained if other boundary conditions are imposed. In this case, we need to modify O(m) rows of AN , and such \small rank" perturbations have been covered in Remark 2.5. However, since multiplicities are in general not preserved by such modications, we need to have a closer look in order to obtain sharp error bounds. In our case we can be more precise since again the eigenvalues can be computed explicitly for a number of congurations (see, e.g., [ChEl89]). For instance, in case of periodic boundary conditions, most of the eigenvalues are of multiplicity 8. Thus, in accordance with [ChEl89], the convergence behavior for Dirichlet boundary conditions with mesh size h is similar to the one obtained for periodic boundary conditions with mesh size h=2. In case of \no- ow" Neumann boundary conditions on the vertical boundaries discretized by a rst order scheme, the corresponding eigen-values are given by (3.4) plus the m eigenvalues of Bm . Here we may expect the same convergence behavior as for Dirichlet boundary conditions. Appendix A. In the appendix we state and prove a lemma that is used in the proof of Theorem 2.1. 26 B. BECKERMANN AND A. B. J. KUIJLAARS Lemma A.1. Let be a nite Borel measure on R with compact support. Suppose (N ) N is a sequence of sets, all contained in a xed compact set, such that lim Z for all continuous functions f on R. let be a Borel probability measure such that t . Let #N such that n=N ! t. Then there exists a sequence of sets (ZN ) N such that (a) (b) ZN N , and (c) for all continuous functions f , lim Z Furthermore, if K is a closed set such that then the sets ZN can be chosen such that in addition to (a), (b) and (c), we also have for N large enough, (d) N \ K ZN . Proof. We have to prove that for some sets ZN satisfying (a) and (b) the normalized counting measures converge in weak sense to t. To show this, we proceed in three steps. Step 1 Suppose we have a nite partition of R consisting of measurable sets j. Since the normalized counting measures of the sets N tend to , we then have for every j, lim then possible to choose, for every j and N , a subset ZN;j N \ U j such that lim The sets ZN;j are disjoint and for their union Z we have Z Hence lim #(Z Then also lim lim #(Z so that #Z 1. The set Z N may not have exactly n elements. By adding or deleting o(N) elements, we obtain from Z N a set ZN with exactly n elements. If we add elements, we choose them from N . Then ZN N and the limits lim hold. Now assume we have a nite collection U j , of measurable sets such that (@U j j. The sets U j are not necessarily disjoint. For each I I =@ \ U jA \@ \ (R The sets V I with I ranging over all subsets of a partition of R. By Step 1, see (A.1), there exist sets ZN such that lim Since every U j is a nite disjoint union of some of the V I , it also follows that lim be a basis for the topology of R, chosen such that (@U j j. From Step 2 we get for each k, a sequence of sets (Z (k) such that #Z (k) N N and lim see (A.2). Then by a diagonal argument, it is possible to nd a sequence (k N ) tending to innity, such that the sets ZN dened by lim We also have 28 B. BECKERMANN AND A. B. J. KUIJLAARS so that (a) and (b) hold. if N is the normalized counting measure of ZN , then by (A.3) we have for every j, lim Since the U j form a basis for the open sets, it follows that the measures N tend in sense to t. Thus (c) holds. Next, assume that K is a closed set such that Then lim #(N \ K) and since also we have because of (c) lim #(ZN \ K) we then have Then we modify ZN by adding the elements of (N n ZN ) \ K to ZN and removing o(N) arbitrary elements from ZN n K. This is always possible for N large enough. Then clearly (d) is satised, while (a), (b) and (c) continue to hold. This completes the proof of the lemma. --R On a conjecture of The sensitivity of least squares polynomial approxima- tion Families of equilibrium measures with external Fourier analysis of iterative methods for elliptic problems Conjugate Gradient Methods for Toeplitz systems Constrained energy problems with applications to orthogonal polynomials of a discrete variable From potential theory to matrix iteration in six steps Comparisions of splittings used with the conjugate gradient algorithm Which eigenvalues are found by the Lanczos method? Equilibrium problems associated with fast decreasing polynomials Zero distributions for discrete orthogonal poly- nomials Extremal polynomials on discrete sets A note on the superlinear convergence of GMRES Convergence of Iterations for Linear Equations Integrals and Series Potential theory in the complex plane Equilibrium measure and the distribution of zeros of the extremal polynomials of a discrete variable Iterative methods for sparse linear systems Logarithmic potentials with external Further results on the convergence behavior of conjugate-gradients and Ritz values The rate of convergence of conjugate gradients Some superlinear convergence results for the conjugate gradient method --TR --CTR A. L. Levin , D. S. Lubinsky, Green equilibrium measures and representations of an external field, Journal of Approximation Theory, v.113 n.2, p.298-323, December 2001 S. Helsen , M. Van Barel, A numerical solution of the constrained energy problem, Journal of Computational and Applied Mathematics, v.189 n.1, p.442-452, 1 May 2006
superlinear convergence;conjugate gradients;toeplitz systems;krylov subspace methods;logarithmic potential theory
588536
Integral Operators on Sparse Grids.
In this paper we are concerned with the construction and use of wavelet approximation spaces for the fast evaluation of integral expressions. The spaces are based on biorthogonal anisotropic tensor product wavelets. We introduce sparse grid (hyperbolic cross) approximation spaces which are adapted not only to the smoothness of the kernel but also to the norm in which the error is measured. Furthermore, we introduce compression schemes for the corresponding discretizations. Numerical examples for the Laplace equation with Dirichlet boundary conditions and an additional integral term with a smooth kernel demonstrate the validity of our theoretical results.
Introduction . A naive Galerkin discretization of an integral operator Z with global kernel K leads to a dense stiness matrix. Hence, on a uniform full grid with O(2 nJ ) unknowns (n dimension, J maximal level in a multiscale discretization), the discrete operator has O(2 2nJ ) entries. This makes matrix vector multiplications, as they are needed in iterative methods, prohibitively expensive for large n or large J . A standard strategy to reduce this cost is to exploit the decay properties of the kernel that, for singular kernels, are typically of the form @ x @ y K(x; y) C x y f(;) Such decay properties do hold for pseudo-dierential operators [24]. Together with (isotropic) biorthogonal wavelets with a su-cient number of vanishing moments, most entries in the corresponding stiness matrices are then close to zero and can be replaced by zero without destroying the order of approximation [4, 9, 12]. Another approach to reduce the cost of the integral evaluation is to replace the full grid approximation space by the so-called sparse grid space [30]. The idea is to use an anisotropic tensor-product basis together with a specic subset of the full grid approximation space. It has been shown that under some assumptions on the approximation and smoothness properties of the underlying basis functions and provided that certain additional regularity assumptions are fullled, the resulting approximation spaces exhibit the same order of approximation as the full grid space, while having less dimension, see for example [32]. The sparse grid approach (that also appeared under the names hyperbolic cross approximation or boolean blending schemes) is well established in approximation and interpolation theory, see e.g. [1, 11, 13, 30, 33, 34]. First approaches to use sparse grids for integral operators can be found in [16, 18, 20, 27]. In [20] it has been observed that a discretization with adaptive sparse grid spaces leads to good approximation rates for the single layer potential equation on a square. In [18] theoretical results on the discretization of elliptic operators of arbitrary order with sparse grids and biorthogonal wavelets have been presented. In both papers sparse grids are used as test- and ansatz-spaces in a Galerkin discretization. Hence, without further com- pression, the resulting discrete systems are dense. This rises the question, whether additional compression schemes along the lines of [10, 29] are applicable for this type of tensor-product discretizations. For integral operators with smooth kernel, sparse grids may be used directly for the approximation of the kernel. The theoretical background to this approach is that the smoothness of the kernel translates into decay properties of its multiscale coe-cients. Then one can construct (modications of) sparse grid spaces such that the evaluation of integral expressions of the type (1.1) can be performed with much lower cost but the same accuracy as with the full grid approximation space. This approach has already been used for complexity estimates of Fredholm integral equations of the second kind with smooth kernels for the special case with isotropic Sobolev smoothness or dominating mixed smoothness for K und u [15, 16, 27]. In this paper we develop sparse grid type approximation spaces for the approximate evaluation of (1.1). We start with smoothness classes including isotropic Sobolev spaces as well as spaces of dominating mixed derivative, and construct sparse grid type approximation spaces for the kernel that are adapted to the smoothness of K and u from (1.1), as well as the norm in which the error of the integral evaluation is measured. Moreover we develop compression schemes for anisotropic tensor-product discretizations for kernels that obey (1.2). Note that although these two types of compression are based on dierent properties of the kernel, it is instructive to present both approaches in a somewhat unied framework. The remainder of the paper is as follows. Section 2 introduces the necessary notation and summarizes the basic facts about biorthogonal wavelet bases, tensor-product spaces, norm equivalences and the smoothness classes we consider. They are certain intersections of classes of functions with dominating mixed derivatives, see (2.1) below. In Section 3 we investigate strategies for the fast evaluation of (1.1). We introduce our new approximation spaces and give estimates on the order of approximation obtained with these spaces and discuss compression schemes for anisotropic tensor-product discretizations for kernels that obey (1.2). In Section 4 we present numerical examples for the integral evaluation and for the numerical solution of the Laplace equation with Dirichlet boundary conditions and an additional integral term with smooth kernel. 2. Preliminaries. Multi-indices (vectors) are written boldface, for example (j Inequalities like l t or l 0 are to be understood componentwise. We exist independent of any parameters x or y may depend on such that C 1 y x C 2 y: In the rest of the paper C denotes a generic constant which may depend on the smoothness assumptions and on the dimension n of the problem. Moreover, for By dist(x; y) we denote the euclidian distance between x and y. 2.1. Sobolev spaces. Let us denote by H t (I n ); t 2 IR, a scale of standard Sobolev spaces on I the space of L 2 -integrable functions on I n . Now, we dene the Sobolev spaces H t;l mix . They x the smoothness assumptions we consider. Definition 1. Let t 2 IR; l 2 the i-th unit-vector in IR n . Then we dene mix mix mix mix H kn (I) for k 2 IR n . For l < 0 we dene H t;l mix as dual of H t; l mix This denition includes the class of functions of dominating mixed derivative H t;0 mix as well as the standard isotropic Sobolev spaces H t mix additional information on these spaces, specically on the denition via Fourier transform, the treatment of boundary conditions and the connections between spaces of bounded mixed derivative and these anisotropic tensor-product spaces see [18, 22] and [21]. When the meaning is clear from the context, we will sometimes drop I n in H t;l mix and write H t;l mix . 2.2. Biorthogonal wavelet bases. The approximation spaces considered here stem from anisotropic tensor-products of univariate function spaces. We start from a one-dimensional multi-resolution analysis and we assume that the complement spaces W are spanned by some multiscale basis functions such that we have W set dened from the subdivision rate of successive renement levels. We will consider dyadic renement throughout the paper. Moreover, we assume that f jk forms a Riesz-basis of W j and that there exists a dual system f ~ such that holds. We assume k jk k L . In the following let jk and ~ jk have N and ~ vanishing moments, respectively. Moreover we write ~ for the complement spaces spanned by the dual wavelets. Under these conditions every u 2 L 2 has unique expansions To simplify things we assume in the following the validity of the norm equivalences where 0 and an analogous relation for the dual wavelet (with t 2 ( r)). Such two-sided estimates can be inferred from the validity of direct estimates (estimates of Jackson type) and inverse estimates (Bernstein inequalities) for the primal and the dual wavelets as a consequence of approximation theory in Sobolev spaces together with interpolation and duality arguments, see e.g. [8, 25]. For the higher-dimensional case n > 1, let j 2 IN n be given, and consider the tensor-product partition with uniform step size 2 j i into the i-th coordinate direction. By we denote the corresponding tensor-product function spaces W j := W W jn and ~ ~ basis of W j is then given by [k2 j f jk Under the assumption of the validity of (2.3) and an analogous relation for the dual wavelet, it is then straightforward to prove the following norm equivalences mix r) and mix ~ r), see [18]. For we regain the (standard) norm equivalences for the isotropic Sobolev space H t and the Sobolev space with dominating mixed derivative H t;0 mix . The dierent factors 2 2tjjj 1 and 2 2ljjj 1 in these equivalences re ect the dierent smoothness requirements. One of the merits of the validity of a norm equivalence for H s is the fact that it leads directly to optimal preconditioning for H s -elliptic problems and hence to fast iterative methods with convergence rates independent of the number of unknowns, see [8]. Moreover, (2.4) and (2.5) may be used to analyse the approximation of the embedding and to construct optimized sparse grid type Finite Element approximation spaces for elliptic variational problems including integral operators [18]. However, for integral operators the resulting stiness matrices are dense. In the next Section we develop schemes to avoid the additional cost due to the denseness of the matrices. 3. Tensor-product approximation and compression of integral opera- tors. The aim of this Section is to e-ciently approximate the coe-cients h jk of To this end, we represent K and u from (1.1) in the dual and primal basis, respectively, (j;k)2IN 2n (l;m)2 jk a jklm ~ jl (x) ~ where a jk := j k . Here we assume that eventual boundary conditions are implemented into the denition of the primal and dual wavelets, compare Section 4. Given an index set I IN 2n we consider the approximation space ~ I := (j;k)2I ~ ~ Typical examples are the full grid spaces ~ ~ ~ and the sparse grid (or hyperbolic cross) spaces ~ ~ ~ (3. Figure 3.1. Index sets corresponding to the full (left) and the sparse grid space (right) with maximal level in one dimension from [30]. These are associated with rectangular index sets and with the index sets respectively, compare Figure 3.1. The dimensions of the full grid and the sparse grid approximation space are O(2 2nJ ) and O(2 J J 2n 1 ), respectively, see [14, 30]. Note that the dimension of the sparse grid space compares favourably with the dimension of the full grid space. Given an additional index set comp jk jk we dene the following two approximations of the kernel K, K I (x; y) := (j;k)2I (l;m)2 jk a jklm ~ jl (x) ~ km (y) and I (x; y) := (j;k)2I (l;m)2 comp a jklm ~ jl (x) ~ In (3.5) the kernel K is approximated by tensor-product approximation and by setting all coe-cients a jklm with (j; I to zero. The second approximation step (3.6) consists of (additionally) setting some of the coe-cients a jklm to zero. The corresponding integral operators are the uncompressed operator A from (1.1) and the two operators A I u(x) := Z K I (x; y) u(y)dy and A comp I u(x) := Z I (x; y) u(y)dy: Now, to obtain schemes for an e-cient evaluation of the integral expression Au, we have to choose the index sets I and comp jk in such a way, that on the one hand a good approximation of the integral is obtained, and on the other hand, the cost of the integral evaluation remains as small as possible. The rest of this Section is devoted to this task. 3.1. Tensor product approximation. The following Theorem estimates the approximation error k(A A I )ukH s for a given index set I. Theorem 1. Let mix . We assume that the norm equivalences (2.4) and (2.5) hold. Moreover let the parameters ~ r from (2.4) and (2.5) be such that ~ holds. Then, (j;k)62I mix mix Proof: We use the biorthogonality between the dual and the primal wavelets and the validity of (2.4) and (2.5). It holds (j;k)62I (l;m)2 jk a jklm ~ jl (x) ~ km (y)A p2o uop op (y) (j;k)62I (l;m)2 jk a jklm ukm ~ k:(j;k)62I a jklm ukmA ~ k:(j;k)62I a jklm ukmA max (j;k)62I k:(j;k)62I (j;k)62I jklm km (j;k)62I mix mix :Now, for a given bound of the error k(A A I )ukH s we determine minimal index sets I. To simplify things, we assume in the rest of this Section l; t; we use in (3.7) for I the index set f(j; which corresponds to the standard full grid approximation space (3.3) for the kernel K, we obtain mix mix Note that the order of approximation in (3.8) does not explicitely depend on the smoothness of u (p and q only restrict the range of l and t). For example for a kernel with product structure K(x; the smoothness of Au is only dependent on the smoothness of k 1 . A closer look on (3.7) and (3.8) reveals that one could discard indices from the full grid index set without destroying the order of approximation from (3.8). This motivates the following denition of index sets. Figure 3.2. Index sets I100 (0; 0; from left to right. Definition 2. We introduce the parametrized index sets I J (s; t; l; p; q) := f(j; According to (3.2) the related approximation spaces are ~ (j;k)2IJ (s;t;l;p;q) ~ ~ When the meaning is clear from the context, we sometimes drop the parameters (s; t; l; p; q) and write I J instead of I J (s; t; l; p; q). For we regain the index sets corresponding to the sparse grid spaces (3.4) and the standard full grid spaces (3.3), respectively, compare Figure 3.2 (rst and second) and Figure 3.1. Additional smoothness in u, that is p > 0 or q > 0, further reduces the number of entries in the index sets I J . Figure 3.2 (third and fourth) gives two examples for the one-dimensional case. Due to the biorthogonality of the primal and dual wavelets, not all coe-cients of the wavelet representation of u are required for the evaluation of A I u or A comp I u. For example, the coe-cients of h IJ := A IJ u read a Therefore, we need the coe-cients ukm only if 9k 2 IN I J . In this way, I J also induces approximation spaces for u, VUI J := k2UI J where Similarly, for Au the denition of I J implies the index set Note that this index set is optimal for the approximation of functions from H t;l mix if the error is measured in the H s -norm, compare [18]. Moreover, it depends only on s l t . The following Lemma summarizes the approximation properties of our new approximation spaces. It is a direct consequence of (3.7) and the denition of I J in (3.9). Lemma 1. Let under the assumptions of Theorem 1 it holds mix Inequality (3.12) shows that the use of the index set I J (s; t; l; p; q) leads to the same rate of convergence as the use of the full grid index set, compare (3.8), although in most cases the number of elements in I J (s; t; l; p; q) is much smaller. Note that the validity of (3.7) and (3.12) may be extended to the maximal range because of the wider range of the -estimate in (2.3), see [29] (with eventual additional logarithmic terms in the extremal case and special cases of the index sets). If the optimal approximation order is bounded by the approximation order of the wavelets and not by the smoothness of K or u (i.e. for example for the case t > ~ then the optimal index set I J is obtained by setting Then, for example for the L 2 -norm, i.e. the optimal index set is I J (0; ~ Together with an estimate of the number of elements in I J (s; t; l; p; q) (which may be obtained from general results in [14]), it is straightforward to obtain estimates for the cost of approximating Au. For example in the case t > 0, i.e. for a kernel with dominating mixed smoothness, the number of elements N needed to obtain an approximation with bounded by O( 1 O( 1 O s l for s l < 0: In the case of isotropic smoothness, i.e. the number of elements needed to obtain an approximation A IJ u with k(A A IJ )ukH s is bounded by O( n O( n O 2n s l q for The above estimates have to be compared with the corresponding number of elements when using the full grid approximation spaces. Hence, for a kernel with dominating mixed smoothness and s > l, an -approximation may be obtained with a cost that is asymptotically independent of the dimension n, compare (3.13). For s l, there appears a slight n-dependence. This is similar to the approximation of functions from H t;l mix in H s . Indeed, in this case, the integral evaluation is asymptotically not more expensive than the approximation of a function from H t;l mix , compare [18]. For K and u with isotropic smoothness, the cost is dependent on n, compare (3.14). But in any case, the integral evaluation with the above scheme is less costly than the evaluation with the use of full grid approximation spaces. See also [26, 27], where special cases of this phenomenon have been described previously. Note however, that without further compression the above sparse grid approach still leads to dense discrete operators. 3.2. Additional compression. Additional compression may be obtained via a special choice of the index sets comp jk from (3.6). The idea is to drop small entries in the stiness matrix without destroying the order of approximation. Here, simple thresholding, i.e. dropping those entries in the stiness matrix that are under a certain threshold, is in general not feasible, as this approach requires the computation of all entries. Moreover, this approach does not take into consideration the smoothness of u and the norm in which the error is measured. Special compression schemes for isotropic discretizations that take these considerations into account can be found in [10, 29]. Then for isotropic tensor-product wavelets on a full grid the number of entries in the stiness matrix after compression is only of order O(J k 2 Jn ) with some k 2 IR + or even O(2 Jn ). Corresponding investigations concerning anisotropic tensor-product discretizations seem still to be missing, especially together with sparse grid type discretizations. Along the lines of [10] we obtain the following results. Decay properties of the kernel such as (1.2) introduce decay properties of the stiness matrix entries. Specically, if the decay property (1.2) holds for all ; N Taylor expansion of the kernel together with the cancellation properties of the primal wavelets shows for That is, the size of the entries depends on the distance of the supports of the basis functions lk , l 0 k 0 and on their scales. The rate of the decay of the entries is governed by the number of vanishing moments. The following Lemma estimates the error of setting entries in the stiness matrix to zero. Lemma 2. Let I mix . Moreover let the norm equivalences (2.4) and (2.5) hold. The parameters ~ r from (2.4) and (2.5) shall full ~ I A comp I )uk 2 (j;k)2I (l;m)2 jk n comp mix The proof of Lemma 2 is similar to the proof of Theorem 1. It uses again the biorthogonality of the primal and the dual wavelet and the stability of the dual wavelets in H s , see (2.5), and of the primal wavelets in H p;q mix , see (2.4). Now, we need to dene comp jk in such a way that the order of approximation does not deteriorate because of this compression. Assume for a moment that we have A I )ukH s ' O(2 RJ ) with some R 2 . Then we require that k(A I A comp I )ukH s is of the same order, i.e. k(A I A comp I )ukH s ' O(2 RJ ). According to Lemma 2 it is su-cient to require (j;k)2I (l;m)2 jk n comp jklm O 2 2RJ The following Theorem tells us how to dene the index sets comp jk . Note that the choice of the comp jk has to balance the overall complexity and the error after compression in an optimal way. Theorem 2. For (j; where J := maxfjj; Ig. We dene comp Under the above assumptions ((3.15) and Lemma 2) we then have I A comp I )ukH s C 2 RJ kuk H p;q mix Proof: Combination of (3.15), Lemma 2 and the Denition of comp jk shows (for shortness we use the abbreviation I A comp I )uk 2 (j;k)2I mix C (j;k)2I (l;m)2 jk mix Now we apply (l;m)2 jk I A comp I )uk 2 (j;k)2I mix (j;k)2I1 A kuk 2 mix mix :The resulting stiness matrix will in general be non-symmetric. A symmetric operator may be derived by taking the maximum of B jk and B kj in the denition of comp jk . At this point we need an estimate of the number of remaining non-zero elements in the compressed stiness matrix. A full analysis of this problem is presently missing. Note however, it must be possible to obtain a compression with anisotropic wavelets which is optimal up to a logarithmic factor: Each compactly supported univariate scaling function on level j has a representation by O(j) compactly supported univariate wavelets. Hence, each isotropic tensor product wavelet on level j has a representation by at most O(j product wavelets. Therefore, the compression schemes of [10, 29] translate into compression schemes for anisotropic wavelets which result in stiness matrices with O(J 2n 1 J k 2 Jn ) entries. 4. Numerical examples. In this Section we present two numerical examples in 2D which show the e-ciency of our scheme for the approximate integral evaluation and for the solution of integro-dierential equations. Here we use orthogonal Daubechies wavelets on the interval with e.g. [7, 19]. Now, there is a certain coarsest level such that the approximation space V j0 cannot be further decomposed into V j0 1 and W j0 . In the multivariate setting of our numerical experiments, this leads to some modications in the denition of the index sets I J , U IJ for the approximation of K and u. Here, we use I J := f(j; for the optimized scheme and I f for the classical full grid scheme. Except for the correction terms 6j 0 and 2j 0 , the index sets I J and U J correspond to the index sets I J (0; N; 0; N; 0); U I2J (0;N;0;N;0) which are optimized for smooth K and u (where the order of approximation is limited only by N) and for k:k L 2-norm estimates of the error. In our experiments K and u are arbitrarily smooth, see subsections 4.1 and 4.2. Table 4.1 shows the dimension of the approximation spaces ~ I for the dierent index sets. It holds dim( ~ and dim( ~ Table Dimension of approximation spaces for and the index sets from (4.1) and (4.2). In the previous Sections the exact projections K I and were used for the analysis. However, in numerical applications K I and u I are usually not explicitly known and have to be approximated by, say, K 0 I and u 0 I . This is an essential part in applications and needs to be considered also for accuracy and complexity estimates. For full grid approximation spaces one usually employs tensor products of one-dimensional quadrature schemes, see [31], to approximate the scalar products of K and u with the scaling functions on the nest level of renement. In all our experiments we used a quadrature scheme of order Then, by the wavelet transform one obtains the coe-cients of K 0 I and u 0 I . For the sparse grid spaces ~ IJ we use Smolyak's blending scheme [2, 11, 17, 30]. Then, for example, for mix and U I2J Jg, one can show I2J mix For mix and the index set I J (0; N; 0; N; obtain mix Note, that this error is of order O(J 2n 1 2 (M+1=2)J=2 ). However, since the overall error is dominated by k(A A IJ )ukH s and not by the approximation of K IJ and u I2J . Similar estimates (without the logarithmic terms) hold for the full grid case. For the calculation of errors the nodal values of the numerical solution h I and the exact solution h are computed on a grid with mesh size 2 12 for the x- and y-direction. Then, the exact (semi-) norms kh h I k L 2 and jh h I j H 1 are approximated using the discrete (semi-) norms These discrete norms yield accurate estimates of the true norms of the error. 4.1. Integral evaluation. We consider the integral evaluation (1.1) with K(x; y) := sin Y and The exact result is 0:577::: is Euler's constant. Furthermore, Si and Ci denote the Sinus and Cosinus Integralis. See Figure 4.5 for a plot of u and h. Figure 4.1 shows the (discrete) L 2 - and H 1 -norms of the error. The new scheme with index sets (4.1) yields approximately the same accuracy as the classical full grid scheme { with a much smaller work count, of course. Furthermore, the rate of error reduction agrees very well with the rate predicted by Theorem 1. Plots of error vs. number of degrees of freedom are given in Figure 4.2. The L 2 -optimized scheme leads to superior convergence rates in both norms. In all experiments the calculation of a 0 jklm was by far the most expensive part of the algorithm. Table 4.2 lists cpu-times for the numerical quadrature, the wavelet transforms involved in the blending scheme and the matrix-vector product (3.10) for the L 2 -optimized scheme. The measurements were carried out on a SGI O2000 with a 195 MHz R10k processor. Due to the blending scheme the numerical work count and, therefore, the cpu-times are only asymptotically proportional to dim( ~ This explains the growth of the normalized cpu-time shown in the fth column of Table 4.2. error full grid I J -optimized J error full grid I J -optimized J Figure 4.1. Discrete L 2 - and H 1 -errors against the maximal level J for the integral evaluation. degrees of freedom error full grid I J -optimized I J degrees of freedom error full grid I J -optimized I J Figure 4.2. Discrete L 2 - and H 1 -errors against dim( ~ I ) for the integral evaluation. Table -optimized scheme: cpu-times in [s] for the calculation of the coe-cients a 0 jklm (quadratures (Q), wavelet transforms (WT)), the multiplications (M) according to (3.10) and cpu-times for Q+ WT and M normalized with the dimensions of ~ I J . 4.2. Solution of integro-dierential equations. We consider the numerical solution of the integro-dierential equation Z Here, K is given by (4.3) and the right hand side f is chosen such that the exact solution is (4.4), see Figure 4.5. The trial and test functions for the Galerkin discretization are Daubechies wavelets with homogeneous Dirichlet boundary conditions and vanishing moments. A BiCGStab(2) iterative solver was used for the solution of the resulting linear system. The number of iterations required to reduce the residual to machine precision was between 40 and 50 (essentially independent of the maximal level J). As shown in Figure 4.3 the optimized scheme yields approximately the same accuracy as the classical scheme and the theoretical convergence rates are matched well. Again the new scheme leads to dramatically reduced errors when compared to the full grid scheme, see Figure 4.4. error full grid I J -optimized J error full grid I J -optimized J Figure 4.3. Discrete L 2 - and H 1 -errors against the maximal level J for the solution of (4.6). degrees of freedom error full grid f -optimized I J degrees of freedom error full grid f -optimized I J Figure 4.4. Discrete L 2 - and H 1 -errors against dim( ~ I ) for the solution of (4.6).0.5010.010.030.50126 Figure 4.5. Left: u, (4.4). Middle: h, (4.5). Right: f , (4.6) Acknowledgements : The authors would like to thank the referees for their useful comments. The second author has been supported by the Deutsche Forschungsge- meinschaft, GR 1144/7-2. --R Approximation by trigonometric polynomials in a certain class of periodic functions of several variables On the adaptive computation of integrals of wavelets Fast wavelet transforms and numerical algorithms I Wavelets on the interval and fast wavelet transforms Stability of Multiscale Transformations Wavelet and Multiscale Methods for Operator equations Theory 34 Compression of Wavelet decompositions Constructive Approximation 14 Number of integral points in a certain set and the approximation of functions of several variables Complexity of local solution of multivariate integral equations Information complexity of multivariate Fredholm equations in Sobolev classes Numerical Integration using Sparse Grids Optimized tensor-product approximation spaces Orthogonal Wavelets on the Interval Hyperbolic cross approximation of integral operators with smooth kernel Approximation und Kompression mit Tensorprodukt-Multiskalen-Raumen Wavelets: Calderon-Zymund and multilinear operators On discrete norm estimates related to multilevel preconditioners in the On the complexity of On the complexity of analysis of the combination technique Quadrature and interpolation formulas for tensor products of certain classes of functions Quadrature formulae and asymptotic error expansions for wavelet approximations of smooth functions Approximation of periodic functions Parallel Algorithms for Partial Di --TR
sparse grids;optimized approximation spaces;biorthogonal wavelets;hyperbolic cross approximation;compression;integral equations;boolean blending
588564
An A Priori Error Analysis of the Local Discontinuous Galerkin Method for Elliptic Problems.
In this paper, we present the first a priori error analysis for the local discontinuous Galerkin (LDG) method for a model elliptic problem. For arbitrary meshes with hanging nodes and elements of various shapes, we show that, for stabilization parameters of order one, the L2-norm of the gradient and the L2-norm of the potential are of order k and k+1/2, respectively, when polynomials of total degree at least k are used; if stabilization parameters of order h-1 are taken, the order of convergence of the potential increases to k+1. The optimality of these theoretical results is tested in a series of numerical experiments on two dimensional domains.
Introduction . In this paper, we present the first a priori error analysis of the Local Discontinuous Galerkin (LDG) method for the following classical model elliptic problem: @n where\Omega is a bounded domain of R d and n is the outward unit normal to its boundary for the sake of simplicity, we assume that the (d \Gamma 1)-dimensional measure of \Gamma D is non-zero. The LDG method was introduced by Cockburn and Shu in [25] as an extension to general convection-diffusion problems of the numerical scheme for the compressible Navier-Stokes equations proposed by Bassi and Rebay in [6]. This scheme was in turn an extension of the Runge-Kutta Discontinuous Galerkin (RKDG) method developed Scientific Computing Program, School of Mathematics, University of Minnesota, Vincent Hall, Minneapolis, MN 55455, email: castillo@math.umn.edu. y School of Mathematics, University of Minnesota, Vincent Hall, Minneapolis, MN 55455, email: cockburn@math.umn.edu. Supported in part by the National Science Foundation (Grant DMS- 9806956) and by the University of Minnesota Supercomputing Institute. z Dipartimento di Matematica, Universit'a di Pavia, Via Ferrata 1, 27100 Pavia, Italy, email: perugia@dimat.unipv.it. Supported in part by the Consiglio Nazionale delle Ricerche. This work was carried out when the author was a Visiting Professor at the School of Mathematics, University of Minnesota. x School of Mathematics, University of Minnesota, Vincent Hall, Minneapolis, MN 55455, email: schoetza@math.umn.edu. Supported by the Swiss National Science Foundation (Schweizerischer Na- tionalfonds). P. Castillo, B. Cockburn, I. Perugia and D. Sch-otzau by Cockburn and Shu [19, 21, 23, 24, 26] for nonlinear hyperbolic systems. The LDG method is one of several discontinuous Galerkin methods which are being vigorously studied, especially as applied to convection-diffusion problems, because of their applicability to a wide range of problems, their properties of local conservativity and high degree of locality, and their flexibility in handling adaptive hp-refinement. The state of the art of the development of discontinuous Galerkin methods can be found in the volume [20] edited by Cockburn, Karniadakis and Shu. Let us give the reader familiar with (classical and stabilized) mixed and mortar finite element methods for elliptic problems an idea of what kind of method is the LDG method. ffl The LDG is obtained by using a space discretization that was originally applied to first-order hyperbolic systems. Hence, to define the method, we rewrite our elliptic model problem as a system of first-order equations and then we discretize it. Thus, we introduce the auxiliary variable obtain the equations \Gammar Since these are nothing but the equations used to define classical mixed finite element methods, the LDG method can be considered as a mixed finite element method. However, the auxiliary variable q can be eliminated from the equations which is usually not the case for classical mixed methods. ffl In the LDG method, local conservativity holds because the conservation laws (1.2) and (1.3) are weakly enforced element by element. In order to do that, suitable discrete approximations of the traces of the fluxes on the boundary of the elements are needed which are provided by the so-called numerical fluxes; these are widely used in the numerical approximation of non-linear hyperbolic conservation laws and are nothing but the so-called approximate Riemann solvers; see Cockburn [17]. As in the case of non-linear hyperbolic conservation laws, these numerical fluxes enhance the stability of the method and hence the quality of its approximation. This is why the LDG method is strongly related to stabilized mixed finite element methods; indeed, the stabilization is associated with the jumps of the approximate solution across the element boundaries. ffl The LDG method allows general meshes with hanging nodes and elements of several shapes since no inter-element continuity is required. This is also a key property of the mortar finite element method. However, in the LDG method there are no Lagrange multipliers associated to the continuity constraint; instead, the Lagrange multiplier is replaced by fixed functions of the unknowns which are nothing but the above mentioned numerical fluxes. ffl In the LDG method, on each element, both the approximation to u as well as the approximations to each of the components of q belong to the same space, which is very convenient from an implementational point of view. Moreover, the lack of continuity constraints across element boundaries in the finite element spaces renders the coding of the hp-version of the LDG method much simpler than that of standard mixed methods. us briefly describe the recent work on error analysis of DG methods in order to put our results into perspective. Analyses of the LDG method in the context of An analysis of the LDG method for elliptic problems 3 transient convection-diffusion problems have been carried out by Cockburn and Shu [25], by Cockburn and Dawson [18], by Castillo [14] and more recently by Castillo, Cockburn, Sch-otzau and Schwab [15]. The DG method of Baumann and Oden [7, 8, 9, 32] has also been analyzed by several authors. Oden, Babu-ska and Baumann [31] studied this method for one dimensional elliptic problems and later Wihler and Schwab [41] proved robust exponential rates of convergence of the Oden and Baumann DG method for stationary convection-diffusion problems also in one space dimension. Rivi'ere, Wheeler and Girault [35] and Rivi'ere and Wheeler [34] analyzed several variations of the DG method of Baumann and Oden (involving interior penalty techniques) as applied to non-linear convection-diffusion problems and, finally, S-uli, Schwab and Houston [38] synthesized the self-adjoint elliptic, parabolic, and hyperbolic theory by extending the analysis of these DG methods to general second-order linear partial differential equations with non-negative characteristic form. As applied to purely elliptic problems, the LDG method and the method of Baumann and Oden are strongly related to the so-called interior interior penalty (IP) methods explored mainly by Babu-ska and Zl'amal [3], Douglas and Dupont [27], Baker [4], Wheeler [39], Arnold [2] and later by Baker, Jureidini and Karakashian [5], by Rusten, Vassilevski, and Winther [36] and by Becker and Hansbo [10]. All of these DG methods for elliptic problems can be recast within a single framework as shown by Arnold, Brezzi, Cockburn and Marini [1]; this framework should provide a basis for a better understanding of the connections among them and lead to a unified error analysis of these methods. As a contribution to this effort, we present in this paper an a priori error analysis of the LDG method for purely elliptic problems. We show that if polynomials of degree at least k are used in all the elements, the rate of convergence of the LDG method in the L 2 -norm of u and q are of order k and k, respectively, when the stabilization or penalization parameter C 11 is taken to be of order one. When the stabilization parameter C 11 is taken to be of order h \Gamma1 , the order of convergence of u is proven to be k + 1, as expected. Indeed, this is what happens for the interior penalty methods and for the modifications of the method of Bassi and Rebay [6] studied by Brezzi, Manzini, Marini, Pietra and Russo [13]; the penalization parameters of these methods are also of order h \Gamma1 . These results are summarized in the Table 1.1. Table Orders of convergence for k - 1. method penalization interior penalty O(h Brezzi et al. O(h Finally, let us point out that the order of convergence of u for the DG method for purely convective problems is k +1=2. This order of convergence was proven by Johnson and Pitk-aranta [30] and later confirmed by Peterson [33] to be sharp. Whether 4 P. Castillo, B. Cockburn, I. Perugia and D. Sch-otzau or not a similar phenomenon is actually taking place for the LDG method, with the stabilization parameter C 11 of order one, as applied to elliptic problems remains to be investigated. Our numerical experiments for the LDG method have all been performed on structured and unstructured triangulations without hanging nodes and give the optimal orders of convergence of k for u and q, respectively, with C 11 of order h \Gamma1 and, remarkably, with C 11 of order one. The organization of the paper is as follows. In section 2, we present the LDG methods and state and discuss our main a priori error estimates. We also give a brief sketch of the proofs in order to display the ideas of our analysis. The analysis is carried out in full detail in section 3 and several possible extensions are indicated in section 4. In section 5, we present several numerical experiments testing the sharpness of our theoretical results. We end in section 6 with some concluding remarks. 2. The main results. In this section, we formulate the LDG method and show that it possesses a well-defined solution. We then state and discuss our main result and, finally, we display the main ideas of our error analysis. We assume, to avoid unnecessary technicalities, that the exact solution u of our model problem (1.1) belongs to H 2 (\Omega\Gamma and that the solution of the so-called adjoint problem satisfies the standard ellipticity regularity property. Extensions to more general situations are discussed in section 4. 2.1. The LDG method. To introduce our LDG method, we consider a general discontinuous Galerkin (DG) method of which the LDG method is a particular but important case. We consider a general triangulation T with hanging nodes whose elements K are of various shapes. To obtain the weak formulation with which our DG method is defined, we multiply equations (1.2) and (1.3) by arbitrary, smooth test functions r and v, respectively, and integrate by parts over the element K 2 T to obtain Z Z ur Z Z Z Z Note that the above equations are well defined for any functions (q; u) and (r; v) in Next, we seek to approximate the exact solution (q; u) with functions (q in the finite element space MN \Theta VN ae M \Theta V , where and the local finite element space S(K) typically consists of polynomials. Note that for a given element K, the restrictions to K of uN and of each of the components of q N belong to the same local space; this renders the coding of these methods considerably simpler than that of the standard mixed methods, especially for high-order polynomial An analysis of the LDG method for elliptic problems 5 local spaces. In order to ensure the existence of the approximate solution of the DG method, we require the following local and quite mild condition: Z Other than these properties, there is complete freedom in the choice of the local spaces since no inter-element continuity is required at all. The approximate solution (q defined by using the above weak formula- tion, that is, by imposing that for all K 2 T , for all Z Z Z Z Z Z where the numerical fluxes b uN and b q N have to be suitably defined in order to ensure the stability of the method and in order to enhance its accuracy. As pointed out in the introduction, we can see that the numerical fluxes b uN and b q N are nothing but discrete approximations to the traces of u and q on the boundary of the elements. To define these numerical fluxes, let us first introduce some notation. be two adjacent elements of T ; let x be an arbitrary point of the set which is assumed to have a non-zero (d \Gamma 1)-dimensional measure, and let n + and n \Gamma be the corresponding outward unit normals at that point. Let (q; u) be a function smooth inside each element K \Sigma and let us denote by (q the traces of (q; u) on e from the interior of K \Sigma . Then, we define the mean values ff\Deltagg and jumps [[\Delta]] at x 2 e as follows: Note that the jump in u is a vector and the jump in q is a scalar which only involves the normal component of q. We are now ready to introduce the expressions that define the numerical fluxes in (2.2) and (2.3). If the set e is inside the domain\Omega\Gamma we take b u ffugg where the auxiliary parameters C 11 , C 12 and C 22 depend on x 2 e and are still at our disposal. The boundary conditions are imposed through a suitable definition of (b q; b u), namely, g N on \Gamma N ; and b u := where the superscript + denotes quantities related to the element the edge we are considering belongs to, and We remark that the definition of (b q; b u) on 6 P. Castillo, B. Cockburn, I. Perugia and D. Sch-otzau the boundary @\Omega is still of the form (2.4) if the exterior traces are taken to be and C 12 is chosen such that C 12 Let us stress once more that the numerical fluxes we just defined are nothing but a particular case of the so-called approximate Riemann solvers widely used in numerical schemes for non-linear hyperbolic conservation laws. This completes the definition of our DG method. The LDG method is obtained when C 22 j 0; in this case, the function q N can be locally solved in terms of uN and hence eliminated from the equations, as can be easily seen from (2.3). This local solvability gives its name to the LDG method. That this DG method actually defines a unique approximate solution depends in a crucial way on the coefficients C 11 and C 22 . Indeed, we have the following result. Proposition 2.1 (Well posedness of the DG method). Consider the DG method defined by the weak formulation (2.2) and (2.3), and by the numerical fluxes in (2.4) and (2.5). If the coefficients C 11 are positive and the coefficients C 22 are non-negative, the DG method defines a unique approximate solution (q Notice that the above result, which we prove in the next subsection, is independent of the auxiliary vector parameter C 12 . The choice C symmetry and stability of the DG method. Finally, let us point out that the role of the auxiliary parameters C 11 and C 22 is to enhance the stability and hence the accuracy of the method. 2.2. The classical mixed setting. The study of our DG method is greatly facilitated if we recast its formulation in a classical mixed finite element setting. To do that, we need to introduce some notation. We denote by E i the union of all interior faces of the triangulation T , by ED the union of faces on \Gamma D , and by EN the union of we assume that \Gamma Now, we sum equations (2.2) and (2.3) over all elements and obtain, after some simple manipulations, that the DG approximation (q N ; uN ) is the unique solution of the following variational problem: find (q VN such that \Gammab(v; for all (r; v) 2 MN \Theta VN . Here, the bilinear forms a, b and c are given by a(q; r) := Z\Omega Z C 22 ds Z K2T Z ur Z ds \Gamma Z c(u; v) := Z ds Z ED C 11 uv ds: The linear forms F , G are defined by F (r) := Z ED Z C 22 (g N \Delta n)(r \Delta n) ds; Z\Omega Z ED ds Z An analysis of the LDG method for elliptic problems 7 Note that these two linear forms contain all the data of the problem. In particular, they contain both the Dirichlet and Neumann data, which is not the case for the classical mixed finite element methods. Equations (2.6) and (2.7) can be rewritten in a more compact form as follows: by setting We end this section by proving Proposition 2.1. Proof of Proposition 2.1. Due to the linearity and finite dimensionality of the problem, it is enough to show that the only solution to the equations (2.6) and (2.7) with taking adding the two equations, we get which implies q As a consequence, equation (2.6) becomes after integration by parts, the form b(\Delta; \Delta) can be rewritten as K2T Z Z ds Z ED we get that K2T Z Hence, owing to (2.1), on every K 2 T and since uN is a continuous function equal to zero on the Dirichlet boundary, we get that uN j 0. This completes the proof of Proposition 2.1. 2 2.3. A priori error estimates. In this section we state and discuss our a priori error bounds for the DG method. As pointed out at the beginning of this section, we restrict our analysis to domains\Omega such that, for smooth data, the solution u of problem (1.1) belongs to H We also assume that when f is in L 2(\Omega\Gamma and the boundary data are zero, we have the elliptic regularity result k u Grisvard [28] or [29]. We assume that every element K of the triangulation T is affine equivalent, see [16, Section 2.3], to one of several reference elements in an arbitrary but fixed set; this allows us to use elements of various shapes with possibly curved boundaries. For each by hK the diameter of K and by ae K the diameter of the biggest ball included in K; we set, as usual, h := maxK2T hK . The triangulations we consider can 8 P. Castillo, B. Cockburn, I. Perugia and D. Sch-otzau have hanging nodes but have to be regular , that is, there exists a positive constant oe such that ae K see [16, Section 3.1]. Moreover, we let the maximum number of neighbors of a given element K be arbitrary but fixed. To formally state this property, we need to introduce the set hK; K 0 i defined as follows: interior of @K " @K 0 otherwise: Thus, we assume that there exists a positive constant 1 such that, for each element These three hypotheses allow for quite general triangulations and are not restrictive in practice. The only assumptions we use for the local space S(K) are that it contains the space of polynomials of degree at most k on K and satisfies (2.1). Next, we introduce a semi-norm that appears in a natural way in the analysis of these methods. We denote by H k (D), D being a domain in R d , the Sobolev spaces of integer orders, and by k \Delta k k;D and j \Delta j k;D the usual norms and semi-norms in H k (D) and H k (D) d ; we omit the dependence on the domain in the norms whenever We define j (q; u) j 2 A := A(q; u; q; u), that is, where \Theta 2 (q; u) := Z C 22 ds Z ds Z ds Z ED We assume that the stabilization coefficients C 11 and C 22 defining the numerical fluxes in (2.4) and (2.5) are defined as follows: ih ff -h fi independent of the mesh-size and jC 12 j of order one. Our main result will be written in terms of the parameters - ? and - ? defined as follows: An analysis of the LDG method for elliptic problems 9 We are now ready to state our main result. Theorem 2.2. Let (q; u) be the solution of (1.2)-(1.5) and let (q N ; uN ) be the approximate solution given by the DG method (2.2) and (2.3). We assume the hypotheses on the local spaces and on the form of the stabilization parameters described above. The triangulations are assumed to satisfy the hypothesis (2.11); if ff 6= 0 or fi 6= 0, we also assume that hypothesis (2.12) is satisfied. Then we have that, for where C solely depends on oe, ffi (not when Let us briefly discuss the above result: ffl We begin by noting that the orders of convergence depend on the size of the stabilization parameters C 11 and C 22 only through the quantities - ? and - ? . This fact has several important consequences: ffi The same orders of convergence are obtained with either C 22 = 0 or C 22 of order h. This means that there is no loss in the orders of convergence if instead of penalizing the jumps of the normal component of q N with a C 22 or order h, no penalization at all (the LDG method) is used. ffi The same orders of convergence are obtained with either the LDG method of order one or C 11 of order h \Gamma1 and C 22 of order one. In general, the same orders of convergence are obtained by taking (ff; or by taking (ff; ffi The most remarkable cases occur when \Gammaff; fi 2 f0; 1g since it is for those values that - ? and - ? achieve their maximum and minimum. The corresponding orders of convergence are displayed in Table 2.1 for k - 1. Table Orders of convergence for ffl In the case 1 - k - s, that is, when the degree of the polynomial approximation is less than needed to fit the smoothness of the exact solution, we see in Table 2.2 that the best orders of convergence for P. Castillo, B. Cockburn, I. Perugia and D. Sch-otzau respectively, are obtained for both C 11 and C 22 of order one. When C 22 is taken to be of order h or equal to zero, the stability of the method is weakened and, as a consequence, a loss in the orders of convergence of 1=2 takes place. If now C 11 is taken to be of order h \Gamma1 , the full order of convergence in the error of the potential is recovered. The numerical experiments of section 5 show that these orders of convergence are actually achieved. However, the expected loss in the orders of convergence when C 11 is taken of order one is not observed, which shows that in practice the LDG method is essentially insensitive to the size of the stabilization parameter C 11 . ffl The influence of the choice of the coefficients C 12 on the accuracy has not been explored in this paper; we only assume those to be of order one. In [22] it is shown that the LDG method, with a suitable choice of the coefficients C 12 , still gives the orders of convergence of k respectively, if Cartesian grids and tensor product polynomials of degree k in each variable are used. Table Orders of convergence for ffl For the case k - s + 1, that is, when the degree of the polynomial approximation is more than needed to fit the smoothness of the exact solution, we see in Table 2.3 that the LDG method performs at least as well as all the other methods; it performs better if C 11 is of order h \Gamma1 . Table Orders of convergence for An analysis of the LDG method for elliptic problems 11 ffl In the case the DG method converges provided C 22 6= 0; in particular, for constant coefficients C 11 and C 22 , we obtain estimates of order one for . This is one of the few finite element methods for second-order elliptic problems that actually converges for piecewise-constant approximations. When C 22 = 0, that is, for the LDG method, our numerical results, which we do not report in this paper, show that there is no positive order of convergence in this case, as predicted by Theorem 2.2. Finally, let us point out that the hypothesis (2.12) is not necessary when 2.4. The idea of the proof. The proof of Theorem 2.2 will be carried out in section 3. The purpose of this section is to display as clearly as possible the basic ingredients and the main steps of our error analysis. As usual, we express the error as the following sum: where \Pi and \Pi are projections from M and V onto the finite element spaces MN and VN , respectively. a. The basic ingredients. The basic ingredients of our error analysis are two. The first one is, as it is classical in finite element error analysis, the so-called Galerkin orthogonality property, namely, This property is a straightforward consequence of the consistency of the numerical fluxes. The second ingredient is a couple of inequalities that reflect the approximation properties of the projections \Pi and \Pi, namely, for any (q; u); (\Phi; for any (r; v) 2 MN \Theta VN and (q; u) 2 H As we show next, all the error estimates we are interested in can be obtained solely in terms of functionals KA and KB . b. The estimate of the error in the A-semi-norm. We have the following result. Lemma 2.3. We have A (q; u; q; u) +KB (q; u): Proof. j (\Delta; \Delta) j A is a semi-norm and, hence, Since by the definition of A, (2.9); by assumption (2.19); P. Castillo, B. Cockburn, I. Perugia and D. Sch-otzau we have that and so, The estimate now follows from a simple application of the assumption (2.18). This completes the proof. c. Estimate of the error in u in non-positive order norms. To obtain an estimate of k e u k \Gammat;D , where t is a natural number and D is a sub-domain of \Omega\Gamma we only have to obtain an estimate of the error in the approximation of the linear functional denotes the L 2 -inner product, by (u N ) since In this paper, we are only interested in the case but we give here the general argument to stress the fact that it is essentially the same for all natural numbers t. Error estimates in negative order norms are very important, as we point out in section 4 of this paper. To obtain our estimate, we need to introduce the solution ' of the so-called adjoint problem, namely, @n Lemma 2.4. Let t be a natural number. Then, we have KA (q; u; \Phi; ') with ' denoting the solution of (2.21)-(2.23) and Proof. Since ' is the solution of the adjoint equation, it is easy to verify that if we set for all (s; w) 2 M \Theta V ; indeed, note that problem (1.1) can be rewritten as in (2.8). Taking (s; by the definition of A, (2.9); , by the assumption (2.19) and the estimate (2.20), we obtain An analysis of the LDG method for elliptic problems 13 and hence, The estimate now follows from a simple application of assumption (2.18), and from the definition of a non-positive order norm. This completes the proof. d. Conclusion. Thus, in order to prove our a priori estimates, all we need to do is to obtain the functionals KA and KB ; this will be carried out in the next section. Then, Theorem 2.2 will immediately follow after a simple application of Lemmas 2.3 and 2.4. 3. Proofs. In this section, we prove our main results. We proceed as follows. First, we obtain the functional KA for general projection operators \Pi and \Pi. To obtain the functional KB , the projections \Pi and \Pi are taken to be the standard L 2 - projections, just as done by Cockburn and Shu [25] in their study of the LDG method for transient convection-diffusion problems. 3.1. Preliminaries. The following two lemmas contain all the information we actually use about our finite elements. The first one is a standard approximation result for any linear continuous operator \Pi from H r+1 (K) onto S(K) satisfying any w 2 P k (K); it can be easily obtained by using the techniques of [16]. The second one is a standard inverse inequality. Lemma 3.1. Let w \Pi be a linear continuous operator from H r+1 (K) onto S(K) such that for some constant C that solely depends on oe in inequality (2.11), k, d and r. Lemma 3.2. There exists a positive constant C inv that solely depends on oe in inequality (2.11), k and d, such that for all s 2 S(K) d we have for all K 2 T . We are now ready to prove our main result. 3.2. The functional KA . In this subsection we determine the functional KA in (2.18), up to a multiplicative constant independent of the mesh-size. We start by giving an expression for KA which is valid for coefficients C 11 and C 22 that vary from face to face, for any regularity of the solution. Then we write KA for the particular choice (2.15), (2.16) of C 11 and C 22 in Theorem 2.2. Let \Pi and \Pi be arbitrary projections onto VN and MN , respectively, satisfying (component-wisely) the assumptions in Lemma 3.1. Lemma 3.3. Assume (q; u) 2 H s+1 the approximation property (2.18) holds true with KA (q; u; \Phi; ') =X 14 P. Castillo, B. Cockburn, I. Perugia and D. Sch-otzau where K2T K2T K2T 22 h 2 minfs;kg+1 K2T 22 h 2 minft;kg+1 K2T K2T K2T K2T K2T K2T ii := supfC ii 2. The positive constant C is independent of the mesh-size but depends on the approximation constants in Lemma 3.1 and on the coefficients C 12 . Furthermore, in the case where (\Phi; KA (q; u; q; Proof. We set, for convenience, - q := We start by writing and then proceed by estimating each of the forms on the right-hand side separately. The form a(\Delta; \Delta) can be written as K2T Z C 22 (- q \Delta n)(- \Phi \Delta n) ds Z @Kn@\Omega C 22 (- q \Delta n)(- \Phi \Delta n) ds Z @Kn@\Omega C 22 (- out ds where the superscript 'out' denotes quantities taken on @K n @\Omega from outside K. By repeated applications of the Cauchy-Schwarz's inequality, we obtain that ja(- q ; - \Phi )j is bounded by K2T 22 - q \Delta nk 0;@K"\Gamma N kC2 22 - \Phi \Delta nk 0;@K"\Gamma N 22 - q \Delta nk 22 - out 22 - \Phi \Delta nk K2T K2T An analysis of the LDG method for elliptic problems 15 K2T 22 k- q \Delta nk 2 K2T 22 k- \Phi \Delta nk 2 Now, a straightforward application of Lemma 3.1 yields To deal with the second term, we first note that K2T r- Z ds Z ds and obtain, after repeated applications of the Cauchy-Schwarz's inequality with suitably chosen weights, that jb(- u ; - \Phi )j is bounded by K2T K2T k- K2T K2T k- K2T Once again, a straightforward application of Lemma 3.1 gives that For the third term, we use the same arguments to get Finally, proceeding as above, we get K2T K2T This proves the first assertion. The second one immediately follows by taking into account that and the proof of the lemma is complete. The following result is a straightforward consequence of the estimates in Lemma 3.3. Corollary 3.4. Let (q; u) 2 H s+1 2 \Theta H s+2 be the exact solution of 0, be the solution of the dual problem (2.21)-(2.23), P. Castillo, B. Cockburn, I. Perugia and D. Sch-otzau and \Gammar'. Assume that coefficients C 11 and C 22 satisfy (2.15), (2.16) . Then there exist a constant C that solely depends on oe, i, - , k and d such that KA (q; u; \Phi; for k - 1. Moreover, KA (q; u; q; Proof. From Lemma 3.3, we get KA (q; u; \Phi; ') =C \Theta h minfs;kg+1 and KA (q; u; q; \Theta Note that the above results hold for arbitrary ff and fi. If now we restrict ourselves to the case of Theorem 2.2, the result follows after simple algebraic manipulations. 3.3. The functional KB . In this subsection we determine the functional KB satisfying (2.19), up to a multiplicative constant independent of the mesh-size. Here, we take \Pi to be L 2 -projection and Again, we start by determining expressions which are valid for varying coefficients C 11 and C 22 , and we conclude by considering the particular case of Theorem 2.2. We proceed as follows. We show that there exists a form j (\Delta; \Delta) j B , which is a semi-norm in both variables, such that for any with C independent of the mesh-size. Then it is enough to determine KB such that for any (q; u) 2 M \Theta V . In the following lemma we prove that (3.1) is satisfied by defining the semi-norm j (\Delta; \Delta) j B as Z ED ds Z ds Z C 22 where for each internal or Neumann boundary face e we set C 22 (x) otherwise: Note that only the function values along faces enter the j (\Delta; \Delta) j B semi-norm. As can be inferred from the proof of Lemma 3.5 below, this is due to the particular choice of \Pi and \Pi as L 2 -projections. An analysis of the LDG method for elliptic problems 17 Lemma 3.5. Let \Pi and \Pi be the L -projection and L -projection onto VN and MN , respectively, and j (\Delta; \Delta) j B be defined by (3.3). Then (3.1) holds true, with a constant C that solely depends on oe, k and d. Proof. Setting by the definition of the form A in (2.9), Using Cauchy-Schwarz's inequality and the fact that \Pi is the L d -projection, we obtain 'Z C 22 ds Z ds 'Z ds Z ds Furthermore, Z ds Z ED ds Multiplying and dividing by C2 11 and then applying Cauchy-Schwarz's inequality, we obtain 'Z ds Z ED ds 'Z ds Z ds Analogously, Z (ff- ds Z ds 'Z ds Z ds 'Z (ff- ds Z ds The first factor can be estimated as follows: Z ds Z ds - Z C 22 ds Z ds Z ds Z ds Z ds Z P. Castillo, B. Cockburn, I. Perugia and D. Sch-otzau . By the inverse inequality in Lemma 3.2, Z ds Z ds - K2T Z e ds K2T K2T @Kg. Thus, combining the above estimates, we get Finally, Z ds Z ED ds 'Z ds Z ED ds 'Z ds Z ED ds To complete the proof, we simply have to gather the estimates of the terms T i , 4, and apply once again the Cauchy-Schwarz's inequality. The function KB can be easily defined by applying the estimates in Lemma 3.1 to defined in (3.3). Lemma 3.6. For any (q; u) 2 H s+1 2 \ThetaH s+2 the approximation property (3.2) holds true with K2T /e K2T e where e and C is a constant independent of the mesh-size and solely depending on the approximation and inverse inequality constants (cf. Lemmas 3.1 and 3.2). From this lemma, we immediately obtain the following result. Corollary 3.7. Let (q; u) 2 H s+1 2 \Theta H s+2 , s - 0. Assume that the coefficients C 11 and C 22 satisfy (2.15), (2.16). The triangulations are assumed to satisfy the hypothesis (2.11); if ff 6= 0 or fi 6= 0, we also assume that hypothesis (2.12) is Then there exists a constant C that solely depends on oe, ffi, i, - , k and d such that the constant C is independent of ffi. An analysis of the LDG method for elliptic problems 19 Proof. If we take the coefficients C 11 and C 22 as in Theorem 2.2, we get, after a simple computation, /e and where the parameter ffi is defined in (2.12), and b Note that the left-hand sides of the above inequalities are trivially uniformly bounded when otherwise, we must invoke the hypothesis (2.12) to ensure the boundedness of these quantities. We emphasize that this is the only instance in which this hypothesis is used. Hence we obtain \Theta where C is independent of the mesh-size but depends on ffi and on the approximation and inverse inequality constants, and -h fi otherwise: The result follows after simple algebraic manipulations. 3.4. The proof of Theorem 2.2. From Lemma 2.3 and Corollaries 3.4 and 3.7, we get and since minfPA ; the estimate follows. Next, consider the L 2 -norm of the error =\Omega in Lemma 2.4. From the elliptic regularity of the adjoint problem (2.21)-(2.23), we have . The estimates of ku \Gamma uN k 0 directly follow from substituting the expression of KA (q; u; \Phi; ') given by Corollary 3.4, and the expressions of KB (q; u), KB (\Phi; ') given by Corollary 3.7 in (2.24), and bounding k\Phik 1 and k'k 2 by k-k 0 . Indeed, we get and since minfQA j follows with This completes the proof of Theorem 2.2. 4. Extensions. In this section, we indicate how to the extend our main result in several possible directions. P. Castillo, B. Cockburn, I. Perugia and D. Sch-otzau 4.1. The case of polygonal domains. In the case of a non-convex polygonal domain in two dimensions, our assumptions on the smoothness of the solution u of our model problem (1.1) and on the elliptic regularity inequality are no longer true. Indeed, if for instance the Neumann boundary is empty, the Dirichlet data is smooth and f is in L 2 (\Omega\Gamma3 we have, see Grisvard [28], that u 2 H s+2 and ! is the maximum interior angle of @ Moreover, if the Dirichlet data is zero, we have see (1.7) in Schatz and Wahlbin [37] and the references therein. This is the elliptic regularity result that we must use. To prove our error estimates in this case, we proceed as follows. First, we note that our main result Theorem 2.2 can be easily extended to this case; indeed, a simple density argument shows that Lemmas 3.3 and 3.6 remain valid for s; t 2 (\Gamma1=2; 0). Now we proceed as in subsection 3.4 and obtain the desired estimates by using the above mentioned lemmas and the above described elliptic regularity inequality. The estimate of the error in the j(\Delta; \Delta)j A -seminorm remains the same but the estimate of the L 2 -norm of the potential has to be suitably modified. For turns out that only for non-zero orders of convergence for respectively. The results for k - 1 are displayed in Table 4.1 for smooth solutions and in Table 4.2 for non-smooth solutions simply write fl instead of Table Orders of convergence for 4.2. Estimates of the error in negative-order norms. It is very well known that the error in linear functionals can be estimated in terms of the error in negative- order norms. Moreover, Bramble and Schatz [11] showed how to exploit the oscillatory nature of finite element approximations, captured in estimates of the error in negative- order norms, to enhance the quality of the approximation by using a simple post-processing on regions in which the exact solution is very smooth and the mesh is locally translation invariant. An analysis of the LDG method for elliptic problems 21 Table Orders of convergence for estimates of negative-order norms can be easily obtained for our general DG method by following the argument described in subsection 2.4 and the technicalities displayed in section 3. 4.3. Curvilinear elements. The analysis in section 3 covers the case of triangulations of curvilinear elements affine-equivalent to fixed curvilinear reference elements. The aim of this subsection is to show how our main result can be extended to the more general case where such an affine equivalence can not be established anymore. This is, for instance, the case when the problem domain has a boundary with a generic curvature. There are two distinctive possibilities to do that. The first one is to keep the finite element spaces described in the introduction; in this case, the local space S(K) could be taken to be simply P k (K), for example. For our main result to hold in this case, only Lemmas 3.1 and 3.2 would have to be proven for these elements and for the case in which \Pi is the L 2 -projection. The other possibility is to consider elements obtained through the so-called Piola transformation [12, Section III.1.3]. This transformation associates the function (q; u) defined on K to the function (b q; b u) defined on b K by where FK denotes the mapping from b K to K. With the above notation, our finite element spaces are given by It is easy to verify that the following properties are satisfied on each element K of 22 P. Castillo, B. Cockburn, I. Perugia and D. Sch-otzau our triangulation: Z Z Z r Z Z Z This implies that with this choice of finite element spaces, our main result holds if Lemmas 3.1 and 3.2 hold for the reference element b K and for the standard L 2 -projection, provided the mappings FK are sufficiently smooth; see [12] and the references therein. Indeed, the proof of section 3 holds in this case if we use the projections \Pi and \Pi defined by d \Piq := b \Pi is the L 2 -projection into the space S( b K) and b \Pi). The only slight modification of the proof occurs in section 3.3 in the definition of j (\Delta; \Delta)j 2 to which we have to add the term k q k 2 This implies that an extra term in the upper bound of the term T 1 in the proof of Lemma 3.5 appears which is easily dealt with. modification of the proof is required at all. 4.4. General elliptic problems. The extension of our main result to more general elliptic problems which include lower order terms can be done in a straight-forward way by applying our techniques to the formulation used by Cockburn and Dawson [18]. 4.5. Exponential convergence of hp-approximations. In the analysis of the DG methods considered in this paper, we have only derived error estimates with respect to the mesh-size h and we have not exploited the dependence of our estimates on the approximation order k. However, this can be done by modifying Lemmas 3.3 and 3.6 correspondingly; see also the work of Houston, Schwab and S-uli [38] and the references there. In addition, by using the proper mesh design principles and by obtaining suitable approximation error estimates in the elements abutting at solution singularities, exponential convergence of the DG method can be proved. See, for example, the recent work of Wihler and Schwab [40] who showed exponential convergence for a model elliptic problem on a polygonal domain\Omega for the DG method of Baumann and Oden with interior penalties. 5. Numerical results for the LDG method. The purpose of this section is to validate our a priori error estimates for the LDG method (i.e., C 22 = 0) and to assess how the quality of its approximations depends on the size of the stabilization parameters C 11 . Since C 22 = 0, the function q N can be expressed locally in terms of uN and hence can be eliminated from the equations. In our examples we solve the resulting linear system for uN by using the standard Conjugate Gradient algorithm; in order to obtain as much precision as possible, the stopping criterion is such that the absolute residual norm is less than 10 \Gamma12 . The approximation q N is then recovered in a post-processing step by using the local expression of q N in terms of uN . We present numerical results using sequences of structured as well as unstructured triangular meshes fT i g, where the mesh-size parameter of T i+1 is half the one of T i . The numerical orders of convergence of the errors are computed for An analysis of the LDG method for elliptic problems 23 polynomials of degree 1 to 6 in the L 2 -norm and A-semi-norm. These orders are defined as follows. If e(T i ) denotes the error on mesh T i (in the corresponding norm), then the numerical order of convergence r i is In all our computations, we take C 12 normal to the edges and of modulus 1=2. The stabilization coefficient C 11 is chosen to be of order h \Gamma1 . We emphasize, however, that for all our experiments no significant difference has been observed in the errors of the approximations when C 11 is of order one. We also remark that results for are not included either, since no positive orders of convergence have been obtained, as predicted in Theorem 2.2. 5.1. Smooth solutions. In our first example, we investigate the order of convergence for smooth solutions. We solve the model problem (1.1) with homogeneous Dirichlet boundary conditions and empty Neumann boundary. The right hand side f is chosen such that the exact solution is given by cos The sequence of structured meshes used in this example is created from consecutive global refinement of an initial coarse structured mesh; at each refinement, every triangle is divided into 4 similar triangles. The number of triangles of the meshes are 16, 64, 256, 1024 and 4096. Since our analysis is valid for arbitrary meshes, we also perform some tests with a sequence of unstructured meshes. It consists of a set of meshes such that the maximum edge length is less than a certain value. This value is reduced by a factor of two, from one mesh to the next. In this way, if we take two consecutive meshes, one is not the global refinement of the other. The number of elements of the meshes are 22, 88, 312, 1368 and 5404. We show the orders of convergence in the L 2 -norm of the error in the gradient in the A-semi-norm of the error of (q; u) and in the L 2 -norm of the error in u in Tables 5.1, 5.2 and 5.3, respectively. For both types of meshes, we observe that the optimal order of convergence predicted by our theory, see Table 2.2, is achieved. Note that since machine precision is achieved for very fine grids and high polynomials, the corresponding orders of convergence are meaningless and are replaced by a horizontal line. To give the reader a better idea of this phenomenon, in Figure 5.1, we display the actual errors in the potential u whose orders of convergence appear in the left side of Table 5.3. Note how the very last part of the curve corresponding to polynomials of degree bends as a consequence of having reached machine accuracy. P. Castillo, B. Cockburn, I. Perugia and D. Sch-otzau Table Smooth solution; order of convergence of the L 2 error in the gradient q. k order of convergence Structured meshes Unstructured meshes 3 2.7216 2.9488 2.9924 3.0008 2.9303 2.4986 3.2825 2.9120 6 5.8878 5.9683 5.9820 - 6.4090 5.0744 6.4362 - Table Smooth solution; order of convergence of the A-semi-norm of the error in (q; u). k order of convergence Structured meshes Unstructured meshes 3 2.8380 2.9745 3.0018 3.0052 3.0120 2.5618 3.2984 2.9233 Table Smooth solution; order of convergence of the L 2 error in the potential u. k order of convergence Structured meshes Unstructured meshes 6 7.0129 6.9889 6.8763 - 7.3003 6. An analysis of the LDG method for elliptic problems 25 log|| Fig. 5.1. Smooth solution; the L 2 error in the potential u for the structured meshes. 26 P. Castillo, B. Cockburn, I. Perugia and D. Sch-otzau 5.2. An exact solution in H 5(\Omega\Gamma but not in H We solve the model problem (1.1) with exact Dirichlet boundary conditions in the convex domain 1). The right hand side is chosen such that the exact solution of the problem is the function u ff defined by ae cos in (\Gamma1; 0) \Theta (\Gamma1; 1); cos This function belongs to H ff+ 1 but does not belong to H ff+ 1+"(\Omega\Gamma6 for all " ? 0. In this test, ff = 4:5 and so u ff 2 H 5(\Omega\Gamma3 The predicted orders of convergence of the -norm of the error in the gradient and that of the A-semi-norm of the error are both 5, and the predicted order of convergence of the L 2 -norm of the error in the potential is 4; see Tables 2.2 and 2.3. These are precisely the orders observed in Tables 5.4, 5.5 and 5.6, respectively. We use the sequence of structured meshes from the previous test. Similar results not reported here are obtained using unstructured meshes. Table of convergence of the L 2 error in the gradient q. k order of convergence 3 2.1363 2.8375 2.9531 2.9844 6 3.8556 3.9387 3.9710 3.9860 Table of convergence of the A-semi-norm of the error in (q; u). k order of convergence 3 2.5591 2.8822 2.9641 2.9882 6 3.9801 3.9664 3.9776 3.9876 Table of convergence of the L 2 error in the potential u. An analysis of the LDG method for elliptic problems 27 5.3. Smooth solution on an L-shaped domain. We solve the model problem (1.1) in an L-shaped domain with Dirichlet boundary conditions. The exact solution is the function u ff , described above, with 4:5. For this test we use a sequence of unstructured meshes, created from a global refinement of an unstructured coarse mesh. The number of elements of the meshes are 22, 88, 352, 1408 and 5632. In Tables 5.7, 5.8 and 5.9 below, we can see that we obtain the same order of convergence as in the convex case even though the standard elliptic regularity result guarantees an order of convergence for the L 2 -error of the potential smaller by as indicated in Table 4.1. A similar phenomenon takes place with the very smooth solution from the first test. Table 5 -solution on L-shaped domain; order of convergence of the L 2 error in the gradient q. k order of convergence 3 2.6595 2.8369 2.9260 2.9644 6 3.0742 3.9120 4.0307 4.1347 Table 5 -solution on L-shaped domain; order of convergence of the A-semi-norm of the error in (q; u). k order of convergence 3 2.7984 2.8763 2.9379 2.9688 6 4.0916 3.9158 4.0313 4.1347 Table 5 -solution on L-shaped domain; order of convergence of the L 2 error in the potential u. 28 P. Castillo, B. Cockburn, I. Perugia and D. Sch-otzau 5.4. Non-smooth solution on an L-shaped domain. Finally, we present numerical results for the classical L-shaped domain test with a singularity at the reentrant corner. We consider the model problem (1.1) in an L-shaped domain with zero right hand side and Dirichlet boundary conditions such that the exact solution is given by For conforming finite element methods, it has been shown that the orders of convergence in the H 1 and L 2 norms are 2 respectively. The numerical results for the LDG method on the sequence of unstructured meshes described in the previous experiment are reported in Tables 5.10, 5.11 and 5.12. They show that the rates of convergence predicted by Table 4.2 are achieved by the LDG method. Observe that the same rates of convergence as in the conforming case are achieved. Table Non-smooth solution on L-shaped domain; L 2 error in the gradient q. k order of convergence Table Non-smooth solution on L-shaped domain; A-semi-norm of the error in (q; u). k order of convergence Table Non-smooth solution on L-shaped domain; L 2 error in the potential u. An analysis of the LDG method for elliptic problems 29 6. Concluding remarks. In this paper, we present the first a priori error analysis for a general DG method that includes the LDG method and allows for triangulations with hanging nodes and elements of several shapes. We have proven that the orders of convergence of the approximations given by the LDG method with the stabilization parameter C 11 of order h \Gamma1 are optimal; these results have been confirmed by our numerical experiments which also indicate that the quality of the approximation does not deteriorate when C 11 is taken to be of order one. Theoretically, a loss of 1=2 in the orders of convergence can take place but this phenomenon was not observed in the particular test problems we considered; as a consequence, the sharpness of our error estimates in this case remains to be studied. We have also theoretically shown that the effect of taking non-zero stabilization parameters C 22 does not significantly improve the orders of convergence of the LDG method. An exception is, of course, the piecewise constant case in which the LDG method has an order of convergence of 0 whereas the DG method with C 11 and C 22 of order one do converge with orders of convergence of at least 1=2 and 1 in the error of the gradient and potential, respectively. In this paper, nothing has been said about how to chose the parameters C 12 . In a forthcoming paper [22], it will be shown that, in the case of Cartesian grids and tensor product polynomials, the orders of convergence of the LDG method can actually increase if C 12 is suitably chosen. Let us end by pointing out tha the implementation of codes for hp-adaptive versions of the LDG method for general elliptic and transient convection-diffusion-reaction problems is the subject of ongoing work. --R Discontinuous Galerkin methods for elliptic problems An interior penalty finite element method with discontinuous elements Nonconforming elements in the finite element method with penalty Finite element methods for elliptic equations using nonconforming elements Piecewise solenoidal vector fields and the Stokes problem A discontinuous hp finite element method for the Navier-Stokes equations A discontinuous hp finite element method for the Euler and the Navier-Stokes equations A finite element method for domain decomposition with non-matching grids Higher order local accuracy by averaging in the finite element method Mixed and hybrid finite element methods Discontinuous finite elements for diffusion problems An optimal error estimate for the local discontinuous Galerkin method An optimal a priori error estimate for the hp-version of the local discontinuous Galerkin method for convection-diffusion prob- lems The finite element method for elliptic problems The devising of discontinuous Galerkin methods for non-linear hyperbolic con- servatin laws Some extensions of the local discontinuous Galerkin method for convection-diffusion equations in multidimensions TVB Runge-Kutta local projection discontinuous Galerkin finite element method for conservation laws IV: The multidimensional case First international symposium on discontinuous Galerkin methods TVB Runge-Kutta local projection discontinuous Galerkin finite element method for conservation laws III: One dimensional systems Superconvergence of the local discontinuous Galerkin method for elliptic problems on Cartesian grids TVB Runge-Kutta local projection discontinuous Galerkin finite element method for scalar conservation laws II: General framework Elliptic problems in nonsmooth domains An analysis of the discontinuous Galerkin method for a scalar hyperbolic equation A discontinuous hp finite element method for diffusion problems A conservative DGM for convection-diffusion and Navier-Stokes problems A note on the convergence of the discontinuous Galerkin method for a scalar hyperbolic equation A discontinuous Galerkin method applied to nonlinear parabolic equations Improved energy estimates for interior penalty Interior penalty preconditioners for mixed finite element approximations of elliptic problems Maximum norm estimates in the finite element method in plane polygonal domains. An elliptic collocation-finite element method with interior penalties Exponential convergence of the hp-DGFEM for diffusion problems in two space dimensions --TR --CTR Kanschat Guido, Block Preconditioners for LDG Discretizations of Linear Incompressible Flow Problems, Journal of Scientific Computing, v.22-23 n.1-3, p.371-384, January 2005 Rommel Bustinza, A unified analysis of the local discontinuous Galerkin method for a class of nonlinear problems, Applied Numerical Mathematics, v.56 n.10, p.1293-1306, October 2006 Guido Kanschat, Block preconditioners for LDG discretizations of linear incompressible flow problems, Journal of Scientific Computing, v.22-23 n.1, p.371-384, June 2005 Claes Eskilsson , Spencer J. Sherwin, Spectral/hp discontinuous Galerkin methods for modelling 2D Boussinesq equations, Journal of Computational Physics, v.212 n.2, p.566-589, 1 March 2006 Ilaria Perugia , Dominik Schtzau, An hp-Analysis of the Local Discontinuous Galerkin Method for Diffusion Problems, Journal of Scientific Computing, v.17 n.1-4, p.561-571, December 2002 Rommel Bustinza , Gabriel N. Gatica, A mixed local discontinuous Galerkin method for a class of nonlinear problems in fluid mechanics, Journal of Computational Physics, v.207 August 2005 Bernardo Cockburn , Guido Kanschat , Dominik Schtzau, A Note on Discontinuous Galerkin Divergence-free Solutions of the Navier-Stokes Equations, Journal of Scientific Computing, v.31 n.1-2, p.61-73, May 2007 Paul Castillo, A review of the local discontinuous Galerkin (LDG) method applied to elliptic problems, Applied Numerical Mathematics, v.56 n.10, p.1307-1313, October 2006 J. J. W. van der Vegt , S. K. Tomar, Discontinuous Galerkin method for linear free-surface gravity waves, Journal of Scientific Computing, v.22-23 n.1, p.531-567, June 2005 Rommel Bustinza , Gabriel N. Gatica , Bernardo Cockburn, An a posteriori error estimate for the local discontinuous Galerkin method applied to linear and nonlinear diffusion problems, Journal of Scientific Computing, v.22-23 n.1, p.147-185, June 2005 Rommel Bustinza , Gabriel N. Gatica , Bernardo Cockburn, An A Posteriori Error Estimate for the Local Discontinuous Galerkin Method Applied to Linear and Nonlinear Diffusion Problems, Journal of Scientific Computing, v.22-23 n.1-3, p.147-185, January 2005 J. J. W. Vegt , S. K. Tomar, Discontinuous Galerkin Method for Linear Free-Surface Gravity Waves, Journal of Scientific Computing, v.22-23 n.1-3, p.531-567, January 2005 Ilaria Perugia , Dominik Schtzau, On the Coupling of Local Discontinuous Galerkin and Conforming Finite Element Methods, Journal of Scientific Computing, v.16 n.4, p.411-433, December 2001 Slimane Adjerid , Andreas Klauser, Superconvergence of discontinuous finite element solutions for transient convection-diffusion problems, Journal of Scientific Computing, v.22-23 n.1, p.5-24, June 2005 Slimane Adjerid , Andreas Klauser, Superconvergence of Discontinuous Finite Element Solutions for Transient Convection--diffusion Problems, Journal of Scientific Computing, v.22-23 n.1-3, p.5-24, January 2005 Paul Castillo, An a posteriori error estimate for the local discontinuous Galerkin method, Journal of Scientific Computing, v.22-23 n.1, p.187-204, June 2005 Paul Castillo, An A Posteriori Error Estimate for the Local Discontinuous Galerkin Method, Journal of Scientific Computing, v.22-23 n.1-3, p.187-204, January 2005 Ilaria Perugia , Dominik Schtzau, Thehp-local discontinuous Galerkin method for low-frequency time-harmonic Maxwell equations, Mathematics of Computation, v.72 n.243, p.1179-1214, July Igor Tsukerman, A class of difference schemes with flexible local approximation, Journal of Computational Physics, v.211 n.2, p.659-699, 20 January 2006 Bernardo Cockburn , Chi-Wang Shu, RungeKutta Discontinuous Galerkin Methods for Convection-Dominated Problems, Journal of Scientific Computing, v.16 n.3, p.173-261, September 2001
finite elements;elliptic problems;discontinuous Galerkin methods
588569
Numerical Approximation of the Maximal Solutions for a Class of Degenerate Hamilton-Jacobi Equations.
In this paper we study an approximation scheme for a class of Hamilton--Jacobi problems for which uniqueness of the viscosity solution does not hold. This class includes the eikonal equation arising in the shape-from-shading problem. We show that, if an appropriate stability condition is satisfied, the scheme converges to the maximal viscosity solution of the problem. Furthermore we give an estimate for the discretization error.
Introduction Given a Hamilton-Jacobi equation, a general result due to Barles-Souganidis [3] says that any "reasonable" approximation scheme (based f.e. on finite differences, finite elements, finite volumes, discretization of characteristics, etc.) converges to the viscosity solution of the equation. Besides some simple properties that the approximation scheme has to satisfy, it is only requested that the equation satisfies a comparison theorem for discontinuous solutions, which in particular implies uniqueness of the viscosity solution. This result covers a wide class of first and second order Hamilton-Jacobi equations, yet there are interesting examples of equations coming from the applications for which uniqueness of the viscosity solution does not hold. A significant example is given by the Eikonal equation on some open and bounded domain\Omega ae R n coupled for example with a Dirichlet boundary condition on @ This equation arises in the Shape-from-Shading problem in image analysis and a large literature has been devoted to its study (see [4] for a description of the problem This paper was written while the second author was visiting the Dipartimento di Matematica, Universit'a di Roma "La Sapienza" supported by DFG-Grant GR1569/2-1. The research was partially supported by the TMR Network "Viscosity solutions and their applications". and [16] for a viscosity solution approach). It is well known that if f vanishes at some points, there are infinite many viscosity solutions to (1.1) (see [15]). Nevertheless, among these solutions, in general only one is the relevant solution (for example, from the physical point of view, from the control theoretic one, etc. In [6] (see also [14]), requiring a stronger condition for supersolution than that for the standard viscosity solution, a Comparison Principle, which characterizes the maximal viscosity solution of the problem, has been obtained for the following class of Hamilton-Jacobi problems Here\Omega is a bounded domain of R N , H and f are nonnegative continuous functions and f can have a very general zero set (the Eikonal equation (1.1) fits into this class of equation). It is worth noting that this maximal solution is the value function of a control problem associated in a suitable way to (1.2)-(1.3). There are, in general, two approaches to the discretization of problem (1.2)-(1.3). A first possibility is to discretize problem (1.2)-(1.3) directly, but imposing some additional condition which among the infinite many solutions singles out the one we want to approx- imate: for example, in [17], it is assumed that the solution is known on the zero set of f , which is now a part of the boundary of the domain where the problem is discretized. A second possible approach (see [4], [5] and references therein) is to discretize a regularized version of problem (1.2)-(1.3), obtained by cutting from below f at some positive level (note that for f ? 0 problem (1.2)-(1.3) has a unique viscosity solution). To prove the convergence of the scheme, both ffl and the discretization step h have to be send to 0. Since the limit problem does not have a unique viscosity solution, it is not possible to apply the Barles-Souganidis theorem and, to our knowledge, there is no convergence theorem for this class of schemes, at least for a general zero set of f . Furthermore, if ffl and h are not related by some condition, the approximation scheme shows numerical instability and it is not really known which solution is approximated (see [12] for some numerical tests in this sense). Aim of this paper is to describe an approximation scheme for which it is possible to prove the convergence to the maximal solution of problem (1.2)-(1.3), without requiring any additional assumptions. The scheme is based on a two step discretization of the control problem associated to the regularized problem: first in the time variable, discretization step h, and then in the space variable, discretization step k (see [2], [13] for related ideas). In the first part (Sections 3, 4), we study the approximation scheme obtained by discretization in time. We show that, if ffl and h are related in an appropriate way, the scheme converges to the maximal solution of (1.2)-(1.3) for ffl and h going to zero. This result is in the spirit of [3], in the sense that it is based on stability properties of the maximal viscosity solution and on its characterization given by the comparison theorem in [6]. Therefore, the proof of the convergence theorem can be easily modified to manage other boundary conditions instead of (1.3) or, also, different approximation schemes not necessarely based on the control theoretic interpretation of the problem. APPROXIMATION OF DEGENERATE HAMILTON-JACOBI EQUATIONS 3 In the second part (Section 5) we study the discretization error for the fully discrete scheme. We show that, if the zero set of f is not too "wild", it is possible to estimate in terms of ffl and of the discretization steps the L 1 -distance between the approximate solution and the maximal solution of the continuous problem. This part deeply employs the control theoretic interpretation both of the discrete problem and of the continuous one. Continuous problem: assumptions and results In this section we briefly recall the characterization of the maximal solution of problem obtained in [6]. Here and in the remainder of the paper by (sub, super)solutions we mean Crandall-Lions viscosity (sub, super)solutions (see [1] for a general treatment). We first set the assumptions on the data of the problem. The hamiltonian H :\Omega \Theta R N is assumed to be continuous in both variables and to verify lim uniformly for x strictly increasing for t 2 [0; 1] for any (x; p) 2\Omega \Theta R N , and is convex for any x Note that the hypothesis (2.2) replaces the stronger one of convexity of H in p. The function f R is nonnegative, continuous in\Omega . Moreover, defined K := fx 2 it is assumed that Finally we assume g : R N ! R to be a continuous and bounded function. We introduce the gauge function ae and the support function ffi of the convex set Z(x), namely for any (x; p) 2\Omega \Theta R N . Both these functions are convex and homogeneous in the variable p, and are l.s.c. and respectively continuous in\Omega (note that, if x 2 K, ae(x; are related by the following equality Example 2.1 Let / be a continuous function such that strictly increasing. Consider the equation In this case we have We now define a nonsymmetric semidistance on\Omega \Theta\Omega by R T and, for x 2\Omega and r ? 0, the open sets fy It can be shown that the family BL (x; r) induces a topology - L on \Omega\Gamma If K consists of isolated points this topology is equivalent to the Euclidean topology and the problem can be studied in the framework of viscosity solution theory (see [14]). In general, - L is weaker than the Euclidean topology and, for x 2 K, the set of points having zero L-distance from x is a subset of K. To obtain the characterization of the maximal solution, the definition of viscosity solution will be adapted to the topology - L . Definition 2.2 Given a l.s.c. function v continuous function OE is called L-subtangent to v at x 0 2\Omega if, for some ffl ? 0, The L-subtangent is called strict if OE(x) ! v(x) outside BL 0g. We remark that the convexity assumption (2.2) allows us to use Lipschitz continuous test functions instead of C 1 test functions as in the standard definition of viscosity solution. For a Lipschitz continuous function OE, we denote by @OE(x) the generalized gradient of OE at OE is differentiable at x n g: Definition 2.3 A l.s.c. function v is said to be a singular supersolution of (1.2) if for any x 0 2\Omega and for any OE, L-subtangent to v at x 0 such that there exists a sequence x n 2\Omega n K and a sequence p n 2 @OE(x n ) for which lim and lim APPROXIMATION OF DEGENERATE HAMILTON-JACOBI EQUATIONS 5 It is worth noting that the definition of singular supersolution reduces to the standard definition of viscosity supersolution if x 0 In fact, in this case, since the topology - L and the Euclidean topology are equivalent in neighborhood of x 0 , L-subtangents at x 0 coincide with standard subtangents. Moreover, ae(x; p) - 1) if and only if H(x; p) - f(x) (resp. H(x; p) - f(x)). In the following theorem, we compare viscosity subsolutions and singular supersolutions of equation (1.2). Theorem 2.4 Let u 2 USC(\Omega LSC(\Omega ) be a viscosity subsolution and a singular supersolution of equation (1.2), respectively, such that u - v on @ Then Hypothesis (2.2) allows us to give a control theoretic interpretation of problem (1.2)-(1.3). Let U be the value function of the control problem with dynamics where x 2\Omega and q is any bounded measurable function from [0; +1) to R n such that cost functional Z The dynamic programming equation associated to the control problem (2.10)-(2.11) is sup This equation turns out to be equivalent to equation (1.2), in the sense that any viscosity sub or supersolution of equation (2.12) is also a viscosity sub- or supersolution of equation (1.2) and vice versa. In the following we will assume that the boundary datum g verifies the compatibility condition It is standard to show that, under hypothesis (2.13), U is a viscosity solution of (1.2) and satisfies the boundary condition (1.3). Furthermore we have Proposition 2.5 The value function U is a singular supersolution of equation (1.2) in \Omega\Gamma Theorem 2.4 and Proposition 2.5 now allow us to characterize the maximal solution of denote the set of functions v 2 USC(\Omega ) which are viscosity subsolutions of (1.2) and which satisfy v - g on @ \Omega\Gamma From Theorem 2.4 and Proposition 2.5 it follows that the value function U of the control problem (2.10)-(2.11) is the maximal element of S, i.e. the maximal subsolution of problem (1.2)-(1.3). Moreover U is a singular supersolution of (1.2) satisfying on @ hence it is the maximal solution. 6 FABIO CAMILLI AND LARS GR - Remark 2.6 If H is convex in p, then U coincides with the value function of control problem with dynamics (2.10) and cost functional where H (x; \Delta) denotes the Legendre transform of H(x; \Delta), cp. [15]. Note, however, that ffi(x; q) and f(x) +H (x; q) in general do not coincide pointwise. We conclude this section stating a particular case of a general stability theorem proved in [6] needed for the construction of the approximation scheme. Proposition 2.7 Set f ffl be the sequence of viscosity solutions of Then lim ffl!0 uniformly in \Omega\Gamma where U is the maximal solutions of (1.2)-(1.3). Note that for any ffl ? 0 fixed, since f ffl ? 0 in \Omega\Gamma problem (2.14) admits a unique viscosity solution. Moreover this solution is given by the value function of the control problem with dynamics (2.10) and cost functional @\Omega and ffi ffl (x; q) is defined as ffi(x; q) with f ffl instead of f . We introduce some notations we will use in the following. We define Moreover, for ffl ? 0, we set Note that, for any ffl ? 0, (r) is bounded by ! rg. 3 The semidiscrete scheme Let us introduce the semidiscrete approximation scheme, obtained by discretizing in time the exit time control problem (2.10)-(2.15). For a fixed ffl ? 0, we choose a step in time define discrete dynamics by the recursive sequence APPROXIMATION OF DEGENERATE HAMILTON-JACOBI EQUATIONS 7 The cost is given by where (we assume the convention that 0). The value function for this control problem is such that N ! +1g: By a standard application of the discrete dynamic programming principle, the function u hffl is a solution of the problem The following result holds true Proposition 3.1 There is a constant C (independent of h and ffl) such that Moreover u hffl is the unique bounded solution of (3:1). Proof: We first observe that it is always possible to assume, by adding a constant, that - 0. It follows that u hffl - 0. Moreover where M is as in (2.16). be two bounded solution of (3.1) and set w i 2. Then 2\Omega where It follows that with in R N n \Omega\Gamma We conclude that for any ffl ? 0 and h ? 0 there exists at most one bounded solution of (3.3) and therefore of problem (3.1). This solution is given by u hffl . Remark 3.2 If we discretized the control problem (2.10)-(2.11) directly (which corresponds to setting in the previous approximation scheme), the resulting approximating equation does not have a unique bounded solution, similarly to what happens in problem (1.2)-(1.3). This causes the drawback that any algorithm designed to solve that approximating equation could not converge to the maximal viscosity solution and, in any case, displays high numerical instability (see [12]). 4 Convergence of the semidiscrete scheme In this section, we prove the convergence of the approximation scheme introduced in the previous section to the maximal solution of (1.2)-(1.3). Given a locally uniformly bounded sequence of functions v ffl lim inf ffl!0 ffl!0 lim sup ffl!0 ffl!0 for any x 2 \Omega\Gamma The functions lim inf ffl!0 ffl!0 are, respectively, l.s.c. and u.s.c. in\Omega . Lemma 4.1 Let u hffl be a sequence of solutions of (3.1) and assume that is such that Then ffl!0 2\Omega is a singular supersolution of (1.2). Proof: Because of (3.2), the function u is well defined in \Omega\Gamma Let OE R be L-subtangent to u at x 0 2 \Omega\Gamma It is possible to assume without loss of generality (see [6], Proposition 5.1) that OE is a strict L-subtangent to u at x 0 . Employing a standard argument in viscosity solution theory, we find a sequence x ffl of minimum points for u tends to 0 + . Then ae oe ae oe for some q ffl with jq ffl From the Mean Value Theorem for Lipschitz continuous functions (see Clarke [7]), there exist Substituting (4.3) into (4.2), we get Observe that x ffl 62 K, otherwise, since on K, we should have and from APPROXIMATION OF DEGENERATE HAMILTON-JACOBI EQUATIONS 9 which is impossible since ffi ffl is strictly positive in \Omega\Gamma Let By the homogeneity of ffi ffl (x; q) with respect to q, we have q ffl 2 fq 2 1g. Dividing (4.4) by recalling (2.6), we get Since the sequence x ffl belongs to hypothesis (4.1), that u is a singular supersolution of (1.2). Theorem 4.2 Assume that either or \Omega is convex: (4.6) If u hffl is a sequence of solutions of problem (3.1) and satisfies the assumption (4.1), then lim ffl!0 uniformly in\Omega , (4.7) where U is the maximal solution of problem (1.2)-(1.3). Proof: We set ffl!0 ffl!0 These function are well defined because of (3.2). From Proposition 4.1, it follows that u is a singular supersolution of equation (1.2). Moreover it is standard to show that u is a subsolution of (2.12) and therefore of (1.2) in\Omega (see, f.e., [1] or [2]). If we show that u - u on @ \Omega\Gamma then Theorem 2.4 and Proposition 2.5 imply that in\Omega and therefore (4.7). We will show that To show that u(x) - g(x) on @ \Omega\Gamma we need an estimate on the behavior of u hffl in a neighborhood of @ sufficiently small and jg. For x @\Omega be such that d(x; @ for the discrete control problem by and, denoted by x n the corresponding discrete trajectory, let 62\Omega g. Observing that Nh - jy \Gamma xj, we get where M is as in (2.16) and ! g is a modulus of continuity of g. If x @\Omega and x ffl 2\Omega is a sequence converging to x 0 , we have either u hffl @\Omega is such that d(x ffl ; @ converges to x 0 , we get u(x 0 To get the other inequality in (4.8), if g j 0, then u in\Omega and therefore u - 0 on @ If (4.6) holds, by adding a constant, we can always assume that g - 0. For x 2\Omega , let q n be an j-optimal control for u hffl (x), x n the corresponding discrete trajectory and N the exit time we have with C as in (3.2). Let q(t) be a control law for the continuous problem obtained by setting are respectively the trajectory and the time corresponding to q(t), we have R where the estimate j-(T holds because of the convexity of \Omega\Gamma Since u ffl for any x 2 @\Omega and the assumption (4.1) is satisfied, from (4.9) we easily get other inequality in (4.8). Remark 4.3 For the Eikonal equation (1.1) we have condition (4.1) reduces to f is the modulus of continuity of the function f on \Omega\Gamma 5 Discretization error for the fully discrete scheme In this section we will discuss a fully discrete scheme derived from the semidiscrete one as developed in the previous sections. In order to simplify the calculations we assume that APPROXIMATION OF DEGENERATE HAMILTON-JACOBI EQUATIONS 11 the function g defining the boundary condition is uniformly Lipschitz with constant L g , and that the domain\Omega is convex. We will introduce a space discretization which transforms (3.1) into a finite dimensional problem. For this purpose we choose a grid \Gamma covering\Omega consisting of simplices S j with nodes x i and look for the solution of (3.1) in the space const on S j g of piecewise linear functions on \Gamma. By the parameter k we denote the maximal diameter of the simplices S j . For simplicity we assume that the boundary of the gridded domain coincides with the boundary of \Omega\Gamma (In the general case we can always achieve an error scaling linearly with the distance between these two boundaries due to the fact that g is Lipschitz). Thus we end up with the fully discrete scheme ffl;h ffl;h for all nodes x i 2\Omega with the boundary condition u k ffl;h for the nodes x i 62\Omega and linear interpolation between the nodes. Note that there exists a unique bounded solution of (5.1). The boundedness of any solution of (5.1) follows from the fact that ffl;h ffl;h holds for any q 2 R n with 1. Thus we can always choose q such that u k ffl;h depends on nodes which are closer to the boundary @\Omega than x i and (if h ! but with a weight strictly less than one. Since the value in the boundary nodes is bounded we obtain boundedness for each node by induction. Due to the boundedness the existence of a unique solution u ffl;h is now easily proved by applying the Kruzkov transformation ffl;h as in the proof of Proposition 3.1. Note that the function ffi ffl appearing in the scheme is defined implicitely via H and f ffl . In order to solve the scheme we assume that we can compute this function analytically as e.g. in Example 2.1. (In the case of a convex Hamiltonian one may alternatively use a numerical approximation of the integrand from Remark 2.6 via the Legendre transform as given e.g. in [10]. Note, however, that this procedure yields a different cost function than in the following analysis.) We will now start by estimating the discretization error ju ffl ffl;h we allow nonconstant boundary conditions we introduce the following auxiliary functions which will be useful for the estimation of the error. Definition 5.1 For each point x 2\Omega we define where -(\Delta) is an optimal path for the initial value x and -(T For each node x i of the grid pick a control q i minimizing (5.1) and let w 2 2 W be the unique solution of with the boundary condition w 2 interpolation between the nodes. Finally we define Remark 5.2 The existence of optimal paths follows from the continuous dependence of the functional J(x; q) from the control function q using the weak -metric (as defined for control functions e.g. in [9]), using the Gronwall Lemma as in [8, Proof of Lemma 3.4(ii)] and the structure of ffi ffl . Note that the a-priori boundedness of the length of approximately optimal trajectories - following from the positivity of ffi ffl - is crucial for this continuous dependence. Thus in general the existence of optimal trajectories does not hold for the non-regularized problem since there for any sequence of approximately optimal trajectories the length of these trajectories may grow unbounded when we restrict Note that we do not require uniqueness of the optimal paths in Definition 5.1. In the case that there is no unique optimal path we may use one that minimizes w 1 . Definition 5.1 defines functions which are 0 at @\Omega and away from @\Omega essentially grow like ffl;h , respectively. More precisely we have that and ffl;h ffl;h for -(\Delta) and q i as used in the definition. Note that in particular if g(x) j c is constant we obtain Using this w we can give the following estimate for the discretization error. Proposition 5.3 Let u k ffl;h 2 W be the unique solution of (5.1). Then the estimate fi(ffl)h holds for each x 2\Omega and for all sufficiently small k ? 0 and h ? 0 with and ff as defined in (2.16)-(2.18), and some constant C independent from ffl; h and k. APPROXIMATION OF DEGENERATE HAMILTON-JACOBI EQUATIONS 13 The proof can be found in the appendix. Remark 5.4 (i) Note that estimate (5.3) is stronger than the usual L1 estimate since essentially the error scales with the function w(x) being 0 at @ \Omega\Gamma The reason for this behaviour origins in the fact that the error is estimated along the optimal trajectories whose length depends on the optimal value. (ii) The constant ff(ffl) essentially depends on the growth of H in jpj, e.g. in Example 2.1 we have (ffl). The constant fi(ffl) is determined by the difference between H(x; p) and H(x; q) for In particular if H(x; p) 2 [C 1 jpj we have that fi(ffl) - C 1 =C 2 independently from ffl. Finally, ! ffi (which gives a bound for ! for combines the continuity properties of H and f , i.e. in Example 2.1 we have that ! ffi (iii) Note that the requirement on h ensuring the convergence of the fully discrete scheme is thus it is consistent with condition (4.1) for the convergence of the semidiscrete scheme. (iv) The appearance of the value ff(ffl) in the denominator in (5.3) is due to the fact that here we implicitely included the worst case, i.e. that the length of the optimal trajectories may grow like 1=ff(ffl) for ff(ffl) ! 0. Since this is not necessarily the case in many practical examples one can expect better convergence behaviour for ff(ffl) ! 0. (v) A particular nice formulation of estimate (5.3) can be obtained if we consider the Eikonal equation (1.1) (implying assume that f is uniformly impose a homogeneous boundary condition, i.e. (implying L 0). In this case the estimate becomes for some constant C ? 0 independent from ffl; h and k. In particular this implies convergence of the scheme We will now turn to the discussion of the error obtained when equation (1.2) is replaced by equation (2.14), i.e. the error introduced by the regularization of the problem. Proposition 2.7 already implies that u " converges to U , where U is the maximal subsolution of (1.2). Unfortunately, in general this convergence can be arbitrary slow. In the optimal control interpretation this is due to the fact that the length of approximately optimal trajectories may grow unbounded as the approximation gets better and better. Since these long pieces of the trajectories can only appear in regions where f is sufficiently small (otherwise the cost would be large contradicting the approximate optimality), we can derive an estimate for the regularization error by defining a criterion for the sets where f is small which in turn gives a bound on the length of approximately optimal trajectories. The following definition is our main tool for this purpose. Definition 5.5 Let B ae R d be a compact set. For each connected component B i of B we define the inner diameter d(B i ) by 14 FABIO CAMILLI AND LARS GR - where and for B we define the inner diameter by where the sum is taken over all connected components of B. Using this definition we can state the following estimate for the regularization error. Proposition 5.6 Let U be the maximal subsolution of (1.2) and let u ffl be the unique viscosity solution of (2.14). Then the estimate holds where K ffl := fx The proof can be found in the appendix. Here the constant c(ffl) depends only on the sets Z ffl (x)g, i.e. on ffl and on the Hamiltonian H. In fact an easy calculation shows that Thus e.g. the estimate c(ffl) - C ffl fl for some constants C; fl ? 0 and all ffl ? 0 sufficiently small holds if H(x; p) - (jpj=C) 1=fl for all x 2 K ffl , all ffl ? 0 sufficently small and all p 2 R n with jpj sufficiently small. In particular for the Eikonal equation (1.1) this implies Observe that if f is piecewise polynomial then bounded for all ffl ? 0 and hence convergence with order c(ffl) follows for ffl ! 0. Piecewise polynomial maps are in particular interesting since they include the case where f is obtained from experimental data by some polynomial interpolation (e.g. using piecewise linear interpolations, multidimensional splines. The following theorem now gives the full a-priori estimates for the approximation error of the whole numerical approximation. Theorem 5.7 Let U be the maximal subsolution of (1.2) and let u k ffl;h be the unique solution of the numerical scheme (5.1). Then the estimate fi(ffl)h holds for each x 2\Omega and the constants from the Propositions 5.3 and 5.6. APPROXIMATION OF DEGENERATE HAMILTON-JACOBI EQUATIONS 15 Proof: Follows immediately from the Propositions 5.3 and 5.6. Remark 5.8 (i) A possible modification of the scheme can be made if we allow smaller time steps at the boundary @ i.e. for x i 2\Omega and x 62\Omega we use the restricted time step 2\Omega g: Although slightly more difficult to implement this modification usually gives better numerical results. The proof of Proposition 5.3 also applies to this modified scheme. (ii) Due to the structural similarity of the scheme described in this section with the scheme considered in [13], the adaptive grid scheme developed there can also be applied here. Similar convergence results as in [13] can be obtained for our scheme using the technique from the proof of Proposition 5.3. 6 Appendix Proof of the Propositions 5.3 and 5.6 In order to prove Proposition 5.3 we will first state a useful estimate for the local error along the functional. Lemma 6.1 For each measurable q(\Delta) with almost all t 2 [0; h] and the path -(\Delta) with - 2\Omega for all t 2 [0; h] there exists p 2 R n with that Z and Conversely, for each p 2 R n with each x 2\Omega with x 2\Omega there exists a measurable function q(\Delta) with Z and 2\Omega for all t 2 [0; h]. Proof: The convexity of ffi ffl in the second argument implies Z Hence by defining Z hq(t)dt the first assertion immediately follows from the continuity of ffi ffl which is measured by ! The second assertion follows directly from the continuity of ffi ffl setting q(t) j p and using the convexity of\Omega . Proof of Proposition 5.3 We start giving some preliminary estimates. First note that the error at the boundary can be estimated by which simply follows from the Lipschitz property of g. Furthermore it is easy to see that on each element S j of the grid we can estimate for each two points We show the estimate (5.3) by estimating seperately the quantities u k ffl;h (x). First, we consider u Observe that for any fl ? 0 there exists an j - 0 such that which easily follows from the fact that w 1 - 0 and u k ffl;h is bounded. Now we fix some arbitrary fl ? 0 and choose j - 0 to be minimal with (6.3). If the assertion immediately follows. Otherwise by the continuity of the functions and the compactness of\Omega we can conclude that there exists x 2\Omega such that ffl;h Now consider the element S j containing x . We can write x where the x i are the nodes of S j and the - i are nonnegative coefficients with Using estimate (6.2) we obtain ffl;h i2I ffl;h Now for each of the nodes we distinguish three cases. By (6.1) this implies ffl;h 2\Omega and for the optimal path - i (\Delta) with - i Definition 5.1 there exists a APPROXIMATION OF DEGENERATE HAMILTON-JACOBI EQUATIONS 17 @\Omega and - (t). In this case by the convexity of\Omega we can conclude that there exists p 2 R n with 62\Omega such that jx i we obtain ffl;h 2\Omega and for the optimal path - i (\Delta) with - i 5.1 the equality Z holds @\Omega and - In this case Lemma 6.1 and the definition of u k ffl;h imply ffl;h where by (6.8) we can estimate and by Definition 5.1 thus also Taking into account that the coefficients in (6.5) sum up to 1 we derive i2I ffl;h and combining (6.3), (6.4), (6.5), (6.10) and (6.11) we obtain (w 1 from which we conclude that Estimating (note that - j - ff(ffl)) this becomes h- Now we specify the assumption "h; k ? 0 sufficiently small" by choosing them such that h- for some constant C ? 0 and thus fi(")h h- j which implies the desired estimate for w values in this resulting inequality are independent from fl ? 0 this also implies the estimate for The inequality for u ffl ffl;h (x) follows with the same technique and the obvious modifications using note that here the convexity of\Omega is also needed in Lemma 6.1 used in case (iii). Proceeding in this way we end up with the analogous estimate to (6.12) which leads to the desired result here without using the assumptions on k and h. Proof of Proposition 5.6 For any measurable and bounded q and any x 2\Omega denote the solution of (2.10) by arbitrary and pick some x 2 \Omega\Gamma Then by the optimal control representation of U (2.10)-(2.11) there exists a solution - with @\Omega and We now divide the connected components K i I of K ffl into two classes by defining I 1 := and I Then by the continuity of H there exists a constant fl(ffl 2 ) with as Furthermore by the uniform continuity of f every set K i has a volume bounded from below by some uniform constant depending on ffl 2 and hence there are only finitely many of these sets; we may number them by Now we define for each of these K i which is hit by the trajectory - ffl 1 times by " g and t i where we omit those sets K i " for which [t i holds. This gives us a finite number r of pairwise disjoint intervals [t i which we assume to be numbered according to their order, i.e. t i For each trajectory piece - ffl 1 we have by (6.13) and by the fact that outside K ffl the functions ffi and ffi ffl coincide the estimate APPROXIMATION OF DEGENERATE HAMILTON-JACOBI EQUATIONS 19 For the points - ffl 1 yielding ffl for all t 2 [0; ffl is possible by the definition of d(\Delta) and the structure of the dynamics (2.10). We now define a sequence of times t i , and a measurable function ~ q(\Delta) by ~ This construction yields that (t) for all and thus in particular it follows that -(t r ; x; ~ obtain dt dt Now letting first ffl we obtain the assertion since u ffl - U is obvious --R Optimal Control and Viscosity Solutions of Hamilton-Jacobi-Bellman equations An approximation scheme for the minimum time function Convergence of approximation scheme for fully nonlinear second order equations The MIT Press Global "La Sapienza" Optimization and nonsmooth analysis Infinite fime optimal control and periodicity Some aspects of control systems as dynamical systems Fast Legendre-Fenchel transform and applications to Hamilton-Jacobi equations and conservation laws A numerical approach to the infinite horizon problem of deterministic control theory "Image analysis and processing" An adaptive grid scheme for the discrete Hamilton-Jacobi-Bellman equation Uniqueness results for a class of Hamilton-Jacobi equations with singular coefficients Generalized solutions of Hamilton-Jacobi equations A viscosity solution approach to Shape from shading --TR
maximal solution;singular Hamilton-Jacobi equations;regularization;discretization error;numerical approximation
588626
Multilevel Boundary Functionals for Least-Squares Mixed Finite Element Methods.
For least-squares mixed finite element methods for the first-order system formulation of second-order elliptic problems, a technique for the weak enforcement of boundary conditions is presented. This approach is based on least-squares boundary functionals, which are equivalent to the H-1/2 and H1/2 norms on the trace spaces of lowest-order Raviart--Thomas elements for the flux and standard continuous piecewise linear elements for the pressure, respectively. Continuity and coercivity of the resulting bilinear form is proved implying optimal order convergence of the resulting Galerkin approximation. The boundary least-squares functional is implemented using multilevel principles and the technique is tested numerically for a model problem.
Introduction . In the context of least-squares finite element methods for first-order systems, boundary conditions can be enforced as essential boundary conditions in the finite element spaces. This yields optimal order convergence of the Galerkin approximations under suitable assumptions on the regularity of the problem (see, for example, [18, 9, 10]). However, this approach constructs an approximation which is much more accurate on the boundary than in the interior of the domain. For least-squares finite element methods, a natural way of treating boundary conditions is to enforce them weakly by adding boundary functionals. The boundary functional approach is also a natural and simple way to handle the common situation that the boundary conditions cannot be satisfied exactly by trial functions in the finite element spaces (see, e.g., [16, 5, 13, 1], see also [4] for a different approach to handle boundary conditions within a least-squares method). Nonlinear boundary conditions in the modelling of flow in porous media (see, e.g., Bear [2, Chap. 7]) may also be handled effectively by boundary functionals. In this paper, we derive appropriate boundary functionals for the mixed formulation of second-order elliptic problems. These are equivalent to the H 1=2 (\Gamma D )-norm for the Dirichlet conditions and to the H \Gamma1=2 (\Gamma N )-norm on the Neumann boundary. From the view of trace and extension theorems of Sobolev norms this gives the proper balance between the interior and boundary least-squares functionals. These boundary functionals are in some sense the weakest possible without losing the optimal order of the Galerkin approximation. Related but different boundary functional methods in connection to least-squares finite elements were suggested before in the literature. For pure Dirichlet problems, a weighted L 2 (\Gamma) norm was used in [13]. This approach using weighted L 2 norms was generalized to elliptic boundary value problems of A-D-N type in [1]. Our aim in this paper is to justify our multilevel boundary functional approach theoretically and by computational experiments for the mixed formulation of the Poisson equation using lowest-order Raviart-Thomas elements. Clearly, the true potential of this methodology lies in its applicability to more complicated problems, e.g., in flow computations in porous media. The use of the least-squares mixed finite ele- Fachbereich 6 (Mathematik und Informatik), Universit?t-GH Essen, 45117 Essen, Germany ment approach for nonlinear boundary value problems arising in variably saturated subsurface flow is studied in [20]. The techniques presented in this paper can also be extended to least-squares formulations like, e.g., Maxwell, Stokes and Navier-Stokes equations. In the following section, we present the least-squares formulation including the boundary functionals and prove coercivity and continuity of the corresponding bilinear form. Section 3 reviews some results on the approximation properties of finite element spaces including the lowest-order Raviart-Thomas elements. In Section 4, we are concerned with an equivalent and computable boundary functional based on multilevel principles. Section 5 gives a study of the effect of the boundary least-squares functionals on the accuracy of the finite element approximation by computations for a two-dimensional model problem. 2. Least-Squares Formulation with Boundary Functionals. We consider the first-order system formulation of Poisson's equation, div oe (\Omega\Gamma in a bounded polygonal domain\Omega ae IR 2 . The boundary of\Omega is divided into are prescribed. Actually, for \Gamma ' @ may be defined as the space of traces from H 1(\Omega\Gamma and H \Gamma1=2 (\Gamma) as the corresponding dual norm with respect to L 2 (\Gamma) (see, for example, [11, Section I.1]). Despite the fact that we restrict ourselves to two-dimensional problems for the purpose of exposition, all the techniques presented below can be extended to higher dimensions. Note that these techniques can also be extended to more general diffusion problems like those arising in porous media flow. Clearly, each solution (u; 1(\Omega\Gamma of this boundary value problem also minimizes the least-squares functional 0;\Omega kn for any ff ? 0. With the corresponding bilinear form 0;\Omega this is equivalent to finding (u; p) 2 H(div ; \Omega\Gamma \Theta H 1(\Omega\Gamma such that for all (v; We have the trace inequalities kpk 1;\Omega for all (see [11, Theorem 1.5]) with a constant c T and kn \Gamma1=2;@\Omega - kuk div;\Omega for all u MULTILEVEL BOUNDARY FUNCTIONALS FOR LEAST-SQUARES METHODS 3 (see [11, Theorem 2.5]). Moreover, if \Gamma D is a curve of positive measure, the generalized Poincar'e-Friedrichs inequality in [15, Theorem 1.9] gives kpk for all (\Omega\Gamma with a constant c F . If \Gamma restrict ourselves to H normalizing p to, for example, (p; 1) This construction is necessary in order to ensure uniqueness of p and we have a Poincar'e-Friedrichs inequality of the form kpk 0;\Omega for all which satisfy (p; 1) (cf. [3, Section II.3]). These tools allow us to prove coercivity and continuity of the bilinear form for any ff ? 0. Theorem 2.1. Under the assumptions above, for any ff ? 0, the bilinear form B(\Delta; \Delta; \Delta; \Delta) is coercive and continuous with respect to H(div; \Omega\Gamma \Theta H and with positive constants c S and c E (which depend on ff, c T and c F ). Proof. The above trace inequalities and repeated use of the Cauchy-Schwarz inequality lead to 0;\Omega kdiv vk div;\Omega kvk 1;\Omega kqk which proves (2.9). For the coercivity proof we consider two separate cases: (i) \Gamma @\Omega a curve of positive measure. Case Cauchy-Schwarz inequality and ff kn ff 1=2;@\Omega lead to 0;\Omega 0;\Omega ff 1=2;@\Omega ff 1;\Omega for any ffi 2 (0; 1). Combined with (2.8), this implies \Gamma1=2;@\Omega F 1;\Omega F ff 0;\Omega F ff Choosing F )g. Case positive measure. The first step is as in case (i) and using (2.6) and (2.5) we obtain kpk kpk kn ff ff 1;\Omega which holds for any ffi 2 (0; ff). This leads to 0;\Omega Combining this with (2.7) gives 0;\Omega kuk ff F Choosing proves (2.10). We remark that in [18], coercivity and continuity of the bilinear form 0;\Omega is shown for MULTILEVEL BOUNDARY FUNCTIONALS FOR LEAST-SQUARES METHODS 5 Since more emphasis is put on enforcing the boundary conditions as ff increases, this result can be regarded as the limiting case for ff !1. Standard finite element theory implies that the Galerkin approximation in these spaces is of optimal order. There are many ways to achieve this optimal order convergence by boundary functionals which are stronger than the k \Delta k \Gamma1=2;\Gamma N and k \Delta k 1=2;\Gamma D norms, respectively. The importance of using the norms k \Delta k \Gamma1=2;\Gamma N and k \Delta k 1=2;\Gamma D in (2.2) is that the least-squares functionals on the boundary and in the interior are properly balanced. 3. Galerkin Approximation. For the numerical approximation we consider finite element spaces based on a quasiuniform sequence of triangulations fT l g l=0;1;::: of \Omega\Gamma Let h l denote the mesh-size of fT l g given, for example, by the maximal diameter of the triangles. We compute approximate solutions u l 2 V l and p l 2 W l for u and p, respectively, such that F(u l ; p l ; f) is minimized among all u l 2 V l and p l 2 W l . This is equivalent to the variational problem of finding l \Theta W l such that for all (v l ; q l l \Theta W l . The simplest choice is to use the lowest-order Raviart- Thomas space for V l and standard continuous piecewise linear functions for W l on the triangulation T l . Continuity and coercivity proved in Theorem 2.1 give us the usual quasi-optimality of the Galerkin approximation l 2V l q l 2W l Let us assume that f 2 H ff for some ff 2 (0; 1] and that g; h and the boundary are such that p 2 H 1+ff and, consequently, u 2 (H details on the conditions for such regularity results). For example, holds if domain\Omega with the additional property that the interior angles at boundary points separating \Gamma N from \Gamma D are at most -=2. In order to simplify our notation, we write - l . j l to indicate that - l - cj l holds with a constant c which is independent of l. We also write - l h j l to indicate that both - l . j l and j l . - l are satisfied. Thus, [8, Proposition III.3.9] implies l 2V l div;\Omega . h ff l [kfk for the approximation by lowest-order Raviart-Thomas elements and standard finite element interpolation results (cf., e.g., [7, Chapter 4]) give q l 2W l 1;\Omega . h ff l In order to compute the solution of (3.1), we use the bases f\Phi (-) l g M l l and l g N l -=1 for W l . The variational problem (3.1) may then be formulated as a linear system of equations - A uu A up A pu A pp - u 6 GERHARD STARKE l )] -=1;:::;N l ;-=1;::: ;M l l )] -=1;::: ;N l l l ) \Gamma1=2;\Gamma N l and l are the basis representations of u l 2 Setting up the matrices A uu and A pp involves the computation of (n \Delta \Phi (-) l ) l Since we cannot compute these inner products directly, we need to replace k \Delta k \Gamma1=2;\Gamma N by equivalent norms such that the corresponding inner products are computable. Alternatively, one could replace these norms by stronger norms, e.g., norms do not give the optimal balance between accuracy on the boundary and in the interior. 4. Multilevel Implementation of the Boundary Functionals. In this sec- tion, we derive computable norms which are equivalent to the norms k \Delta k \Gamma1=2;\Gamma N and using additive multilevel decompositions in the spirit of [6]. To this end, we define the L 2 (\Gamma)-orthogonal projection Q @\Omega onto (the trace space of) l . We will abuse notation and denote this trace space W l wherever it is clear from the context. With the operator l we have (cf. [17, Theorem 15] or [14, Corollary 3.2.4]) l for all q l 2 W l . For a computable expression that replaces k \Delta k \Gamma1=2;\Gamma N we use the following result. Theorem 4.1. Let V l and W l be the lowest-order Raviart-Thomas spaces and standard piecewise linear continuous finite element spaces, respectively, based on a quasiuniform sequence of triangulations. Then, for u l 2 V l , kn \Delta u l k \Gamma1=2;\Gamma N h sup z l 2W l ;z l 6=0 Proof. By definition (see [11, Section I.1]), kzk MULTILEVEL BOUNDARY FUNCTIONALS FOR LEAST-SQUARES METHODS 7 Clearly, since W l ae H 1=2 (\Gamma N ), this implies kn \Delta u l k \Gamma1=2;\Gamma N z l 2W l ; z l 6=0 For the upper bound, we use the L 2 (\Gamma N )-orthogonal projection Q l;\Gamma N to (the trace space of) W l . It is well-known (cf. [21, Section 4]) that for all z together with the trivial inequalities implies l kzk 1=2;\Gamma N and kQ l;\Gamma N zk 1=2;\Gamma N . kzk We may write sup kzk kzk kzk Using the fact that n \Delta u l is piecewise constant along the boundary, we may estimate the second term on the right hand side using zk . h l kzk kzk 1=2;\Gamma N . h 1=2 l kn \Delta u l k \Gamma1=2;\Gamma N kzk where the last step follows from the inverse inequality k- l k 0;\Gamma N . h \Gamma1=2 l k- l k \Gamma1=2;\Gamma N for piecewise constant functions - l on the subdivision of \Gamma N corresponding to T l . This implies kn \Delta u l k \Gamma1=2;\Gamma N kzk l kn \Delta u l k Choosing l 0 such that 2Ch 1=2 l 0 - 1, we are led to kzk . sup z l 2W l ; z l 6=0 From (4.2) and (4.3) we are led to a computable expression by kn \Delta u l k \Gamma1=2;\Gamma N h sup z l 2W l ; z l 6=0 z l ) 1=2 z l 2W l ; z l 6=0 (n \Delta u l ; C \Gamma1=4 where C is the adjoint of C l;\Gamma N such that (n \Delta u l ; C \Gamma1=4 z l ) holds for all u l 2 V l ; z l 2 W l . This gives rise to the modified bilinear form 0;\Omega The variational problem of finding l \Theta W l such that 0;\Omega for all (v l ; q l l \Theta W l replaces (3.1). The computation of the contributions of the boundary functionals to the Galerkin matrix is based on the mass matrices M j;\Gamma in W j , on the mass matrices ~ M j;\Gamma for coupling V j and W j , ~ and on the restriction matrices I j l which, for j - l, compute moments -=1 from s l = [(s; \Psi (-) l (note that I l l is just the identity matrix in IR N l ). It is easy to see that holds. Of course, the inverse of M j;\Gamma can only be formed on the subspace of nodal basis functions which do not vanish on \Gamma (we keep using this notation for simplicity). For the Dirichlet part, we have l l ;h 0 l l =h l l l MULTILEVEL BOUNDARY FUNCTIONALS FOR LEAST-SQUARES METHODS 9 which implies l h l l I j l M In order to set up the matrix entries associated with the Neumann boundary conditions, we start from l ) l I j l ]M l;\Gamma N l I l l I j l M l;\Gamma N This implies l ) ~ with B l;\Gamma N l I l l I j l which leads to l ); (C \Gamma1=4 l ~ Clearly, from the quasiuniformity of the sequence of triangulations we obtain l ); (C \Gamma1=4 l ~ ~ which is the form we actually use in our implementation. The expressions for the computation of the boundary contributions in (4.5) and (4.6) involve the inverses of mass matrices M j;\Gamma D and M j;\Gamma N These inverses are dense matrices on the subspace of unknowns associated with \Gamma D respectively. The actual computation of M \Gamma1 should be avoided, especially for three-dimensional problems, and replaced by an inner conjugate gradient iteration for solving linear systems with M j;\Gamma D and M j;\Gamma N . It is well-known that the number of iterations required to achieve a certain accuracy is bounded independently of l for these systems. Alternatively, one can use biorthogonal wavelets for the construction of an H \Gamma1=2 (\Gamma N )-equivalent boundary functional for the Neumann conditions. In the two-dimensional case, i.e., \Gamma N consists of a finite number of line segments, this is described in [19, Section 5.6]. Fig. 5.1. Pressure (left) and flux (right) approximation for Example 10.5 1 ComputationalExperiments. For our computational experiments, we consider on the rectangular domain shown in Figure 5.1. For the first example, the boundary conditions are on the right two thirds of the upper and on the right boundary segment, n \Delta on the left third of the upper boundary segment and n \Delta segments. This can be viewed as a simple model problem for flow in porous media where water infiltrates at a prescribed rate on the left part of the upper boundary segement. The left and lower boundary are impermeable and on the remaining part of the boundary pressure is set to zero (hydraulic potential equals gravitational potential). The solution for the flux u is indicated by the arrows in Figure 5.1. For this example, we have 3=2\Gamma"(\Omega\Gamma for all due to the jump from Neumann to Dirichlet boundary condition on the upper boundary segment. The coarsest triangulation T 0 consists of 74 triangles which are then uniformly refined. Figure 5.1 shows the first refinement T 1 . Table 5.1 shows the error with respect to the "exact solution", measured in the H(div ; \Omega\Gamma \Theta H As an "exact solution" we use the Galerkin approximation (with exact enforcement of the boundary conditions on an adaptively refined triangulation T ? 4 which consists of 24042 degrees of freedom for the edges and 8133 for the nodes. The triangulation 4 was constructed from T 3 using four steps of the refinement algorithm presented in [20] based on the least-squares functional as an a posteriori error esti- mator. The least-squares functional is reduced by a factor of more than 10 on T compared to T 3 . This leads us to the conclusion that it is legitimate to use it as reference solution in our experiments. Table Approximation error for the multilevel boundary functional for Example 1 MULTILEVEL BOUNDARY FUNCTIONALS FOR LEAST-SQUARES METHODS 11 Obviously, Table 5.1 shows that the approximation is slightly improved if the boundary functional is weighted by instead of enforcing the boundary conditions exactly 1). We need to compare the multilevel boundary functional with the weighted L 2 (\Gamma) approach of [13] and [1]. In our situation, this means replacing the least-squares functional (2.2) on V l \Theta W l by kn h l This leads to the numbers listed in Table 5.2 which are slightly larger in most cases. However, the difference between the multilevel and the weighted L 2 (\Gamma) functional is rather marginal in this case. Table Approximation error for the weighted L 2 (\Gamma) boundary functional for Example 1 In our second example, the boundary conditions are chosen such that it is not possible to satisfy them exactly with our finite element spaces. On the upper bound- ary, we set linearly interpolated for x 1 chosen to be smaller than the edge length of our finest triangulation. The boundary conditions on the other boundary segments are the same as in the first example. The numbers obtained with the multilevel and the weighted L 2 boundary functionals are shown in Tables 5.3 and 5.4, respectively. Table Approximation error for the multilevel boundary functional for Example 2 Table Approximation error for the weighted L 2 (\Gamma) boundary functional for Example 2 In conclusion, our computational results show the feasibility of using the multi-level boundary functional approach. In particular, this method seems promising in cases where it is not possible to enforce the boundary conditions exactly by the finite element spaces. The construction and use of similar multilevel boundary functionals for more complicated situations including nonlinear boundary conditions in porous media applications is the focus of current research. Acknowledgement . I would like to thank Tom Manteuffel for many helpful discussions related to the subject of this paper. The detailed comments of the anonymous referees are also very much appreciated. In particular, I am thankful to one of them for pointing out references [13] and [1] to me. --R Least squares methods for elliptic systems Dynamics of Fluids in Porous Media Cambridge University Press A generalized Ritz-least-squares method for Dirichlet problems Parallel multilevel preconditioners The Mathematical Theory of Finite Element Methods Mixed and Hybrid Finite Element Methods Finite Element Methods for Navier-Stokes Equations Elliptic Problems in Nonsmooth Domains A least squares decomposition method for solving elliptic equations Multilevel preconditioning - appending boundary conditions by Lagrange multi- pliers Multilevel Finite Element Approximation Two preconditioners based on the multi-level splitting of finite element spaces --TR --CTR Huo-Yuan Duan , Guo-Ping Liang, Nonconforming elements in least-squares mixed finite element methods, Mathematics of Computation, v.73 n.245, p.1-18, January 2004
multilevel boundary functionals;least-squares mixed finite element method;raviart-thomas spaces
588632
Detection of Edges in Spectral Data II. Nonlinear Enhancement.
We discuss a general framework for recovering edges in piecewise smooth functions with finitely many jump discontinuities, where $[f](x):=f(x+)-f(x-) \neq 0$. Our approach is based on two main aspects--- localization using appropriate concentration kernels and separation of scales by nonlinear enhancement. To detect such edges, one employs concentration kernels, $K_\epsilon(\cdot)$, depending on the small scale $\epsilon$. It is shown that odd kernels, properly scaled, and admissible (in the sense of having small $W^{-1,\infty}$-moments of order ${\cal O}(\epsilon)$) satisfy recovering both the location and amplitudes of all edges. As an example we consider general concentration kernels of the form $K^\sigma_N(t)=\sum\sigma(k/N)\sin kt$ to detect edges from the first $1/\epsilon=N$ spectral modes of piecewise smooth f's. Here we improve in generality and simplicity over our previous study in [A. Gelb and E. Tadmor, Appl. Comput. Harmon. Anal., 7 (1999), pp. 101--135]. Both periodic and nonperiodic spectral projections are considered. We identify, in particular, a new family of exponential factors, $\sigma^{exp}(\cdot)$, with superior localization properties. The other aspect of our edge detection involves a nonlinear enhancement procedure which is based on separation of scales between the edges, where $K_\epsilon*f(x)\sim [f](x) \neq 0$, and the smooth regions where examples demonstrate that by coupling concentration kernels with nonlinear enhancement one arrives at effective edge detectors.
Introduction We discuss a general framework for recovering edges from the spectral projections of piecewise smooth functions. Our approach for edge detection is based on two fundamental aspects - localization to the neighborhood of the edges using appropriate concentration kernels and separation of scales by nonlinear enhancement. Both the location and amplitudes of all edges are recovered. Let SN f(x) denote the spectral projection of a piecewise smooth f . Given SN f , one can accurately reconstruct f away from its discontinuous jumps, e.g., [10],[14, x2.1], as well as up to the discontinuities, [11]. In either case, an a priori knowledge on the location of the edges and their amplitudes is required. This issue was treated in recent literature, consult [1], [5], [13], [15]. In [7], we unified the previous treatments as special cases of appropriate concentration kernels. Here we improve on these results in both generality and simplicity. To this end, let [f ](x) := denote the local jump function and let us consider a concentration kernel K ffl (\Delta), depending on a small scale ffl. It is shown that odd kernels, properly scaled, and admissible (- in the sense of having small W \Gamma1;1 -moments of order O(ffl), (2.6)), recover both the locations and the amplitudes of the jumps so that Thus, K ffl tends to "concentrate" near the singular support of f . Differentiation of ffl-supported mollifiers is one example for local concentration kernels outlined in x2.2.1. In x2 we also address the issue of detecting edges in global Fourier projections. Given the first modes, we seek concentration kernels of the form K oe sin kt: It is shown that if the concentration factors oe(-) j -) are normalized so that K oe N (t) is an admissible concentration kernel, K oe and the following error estimate holds -( log N The non-periodic case is studied in x3. The analogous results for the Chebyshev case reads, consult Corollary 3.2 log N The special cases of Fourier concentration factors oe ff (- sin ff- and oe p were considered earlier in [1],[7],[9],[13], and [15]. Our general framework motivates a new set of C 1 -exponential concentration factors which yield superior localization properties away from the detected edges. While (1.1) refers to the asymptotic behavior of the concentration kernel as a function of the small parameter ffl # 0, it is essential to recover the exact locations of the edges of f for the accurate reconstruction of f . In x4 we discuss another essential aspect of edge detection, namely nonlinear enhancement. To this end, one introduces a critical threshold, J crit , for the amplitude of admissible edges, and an enhancement exponent, p, to amplify the separation of scales in (1.1) between the edges, where K ffl f(x) - [f ](x) 6= 0, and the smooth regions where K ffl the enhanced kernel Enhanced detection of edges in spectral data 3 ae Clearly, with p large enough, one ends up with a sharp edge detector where K ffl;J [f at all but O(ffl)-neighborhoods of the jump discontinuities. In this sense, the enhancement procedure actually "pinpoints" the location jump discontinuities, allowing an accurate reconstruction of f . The particular case corresponds to the quadratic filter studied in [12],[22], in the special context of concentration kernels based on localized mollifiers. Acknowledgment . Research was supported in part by the Sloan Foundation (AG) and by NSF Grant No. DMS97-06827 and ONR Grant No. N00014-1-J-1076 (ET). Edge Detection by Concentration Kernels 2.1 Concentration Kernels We want to detect the edges in piecewise smooth functions. Assume that f(\Delta) has jump discontinuities of the first kind with well defined one-sided limits, f(x\Gamma) denote the local jump function. By piecewise smoothness we mean 1 In practice one encounters functions f(x) with finitely many jump discontinuities, and (2.2) requires the differential of f(x) on each side of the discontinuity to have bounded variation. For example, if f 0 (x\Sigma) are well defined (for finitely many jumps), then (2.2) holds. We will detect the edges in such piecewise smooth f 's using smooth concentration kernels, depending on a small parameter ffl. Such kernels are characterized by Thus the support of K ffl f(x) tends to "concentrate" near the edges of f(x). One recovers both the location of the jump discontinuities as well as their amplitudes. To guarantee the concentration property of K ffl , we seek odd kernels, which are normalized so that Z and which satisfy the main admissibility requirement Z Const Remarks. 1 Here and below we use BV [a; b] to denote the space of functions with bounded variation, endowed with the usual semi-norm kOEk BV [a;b] := R b a jOE 0 jdx A. Gelb and E. Tadmor 1. For example, if K ffl (t) concentrates near the origin so that its first moment does not exceed Z Const \Delta ffl; (2.7) then it is clearly admissible in the sense that (2.6) holds. We note that our admissibility condition also allows for more general oscillatory kernels, K ffl (t), where (2.7) might fail, yet (2.6) is satisfied due to the cancelation effect of the oscillations, consult (2.18 below. 2. Observe that the admissibility requirement (2.6) generalizes both properties P 3 and P 4 in the definition of admissible kernel [7, definition 2.1]. Our main result states that Theorem 2.1 Consider an odd kernel K ffl (t), (2.4), normalized so that (2.5) holds, and satisfying the admissibility requirement (2.6). Then the kernel K ffl (t) satisfies the concentration property (2.3) for all piecewise smooth f 's, and the following error estimate holds Const \Delta ffl: (2.8) Proof. Using the fact that K ffl (t) is odd, we have Z Z Z Applying (2.5) yields Z By our assumption in (2.2), F x (t) is BV and it is therefore bounded. Consequently, in the particular case that the moment bound (2.7) holds, the first term on the right of (2.9) is of order O(ffl), yielding Const \Delta Z In the general case, F x (t) has bounded variation, and the admissibility requirement (2.6) implies that the first term on the right of (2.9) is of order O(ffl), and we conclude 2.2 Examples of Concentration Kernels 2.2.1 Compactly supported kernels Our first example consists of concentration kernels which 'concentrate' near the origin, so that (2.7) holds. We consider a standard mollifier, OE ffl (t) := 1 ), based on an even, compactly supported bump function, OE 2 C 1 We then set Enhanced detection of edges in spectral data 5 Clearly, K ffl is an odd kernel satisfying the required normalization (2.5) Z Z In addition, its first moment is of order Z Z and hence (2.7) holds. Theorem 2.1 then implies Corollary 2.1 Consider the odd kernel K ffl ffl (t), based on even OE 2 C 1 Then K ffl (t) satisfies the concentration property (2.3), and the following error estimate holds 2.2.2 The conjugate Dirichlet kernel The conjugate Dirichlet kernel, log N ~ sin kt; is an example of an oscillatory concentration kernel. Clearly, KN (t) is an odd kernel. Moreover, the normalization (2.5) holds with ffl - 1 log N , log N log N Finally, summing ~ sin cos sin t; we find that the first moment of DN (t)= log N does not exceed Const \Delta ffl; log so that the requirement (2.7) is fulfilled. Theorem (2.1) then yields the classical result regarding the concentration of conjugate partial sums, [2, x42],[23, xII Theorem 8.13], log N ~ log N We note in passing that in the case of Dirichlet conjugate kernel, KN (t) does not concentrate near the origin, but instead (2.7) is fulfilled thanks to its uniformly small amplitude of order O(1= log N ). The error, however, is only of logarithmic order, consult [7, x2]. 6 A. Gelb and E. Tadmor 2.2.3 Oscillatory kernels. general concentration factors To accelerate the unacceptable logarithmically slow rate of Dirichlet conjugate kernel in (2.12), we consider general form of odd concentration kernels K oe based on concentration factors, oe( k N ) which are yet to be determined. Clearly K oe N (t) is odd. Next, for the normalization (2.5) we note that K oe Z 1oe(x) x dx: In fact, the above Riemann's sum amounts to the midpoint quadrature, so that for oe(-) one has Z - K oe Z 1oe(-) and thus (2.5) holds for normalized concentration factors oe(-), Z 1oe(-) Consult [7] for further refinement concerning the assumed regularity of oe(\Delta) (We note that oe(\Delta) is rescaled here with an additional factor of \Gamma- compared to [7]). Finally, we address the admissibility requirement (2.6) (and in particular (2.7)). To this end, we proceed along the lines of [7, Assertion 3.3], utilizing the identity (abbreviating - This leads to the corresponding decomposition of K oe K oe sin t: Here, R oe N (t) consists of the first four terms on the right hand side of (2.16), and it is easily verified that each one of these terms has a small first moment satisfying (2.7) (and consequently, (2.6) holds), i.e. jtR oe Const log N Enhanced detection of edges in spectral data 7 For example, using the standard bound j sin(kt)=2 sin(t=2)j - minfk; 1=tg, the contribution corresponding to the first term, I 1 (t), does not exceed I 1 (t) Similar estimates hold for the remaining contributions of I 2 ; I 3 and I 4 . In particular, since oe(- is Finally, the admissibility of the fifth term on the right of (2.16) is due to standard cancelation which guarantees that (2.6) holds, sin tOE(t)dt Const \Delta oe(1) It is in this context of spectral concentration kernels that admissibility requires the more intricate property of cancellation of oscillations. Summarizing (2.14), (2.17) and (2.18), we obtain as a corollary an improved version of the main result in [7, Theorem 3.1] regarding spectral edge detection using concentration kernels, K oe N (t). In particular, since K oe N (t) are N \Gammadegree trigonometric polynomials, one detects the edges of the piecewise smooth function f(x) directly from its spectral projection SN (f) := K oe Corollary 2.2 Consider the odd concentration kernel (2.13) K oe sin kt; oe(-) Assume that oe(\Delta) is normalized so that (2.15) holds Z 1oe(-) Then K oe admits the concentration property (2.3), and the following estimate holds Const \Delta log N Remark. One can relax the regularity on the concentration factor oe(\Delta), [7]. Corollary 2.2 is a generalization of [7, Theorem 3.1] 2 ; in particular, the error estimate (2.19) is valid throughout the interval, including at the location of the jump discontinuities. Let us introduce few prototypical examples of concentration factors oe(\Delta) for the detection of edges from spectral data. In this context we note that other detection methods of discontinuities in periodic spectral data can be found in the works of Eckhoff [5], [6] and of Mhaskar & Prestin, e.g., [15] and the references therein. We note that our results apply to the non-periodic expansions as discussed in x3.2.2 below. We note the different rescaling here of oe(\Delta) by an additional factor of \Gamma-, compared with the formulation in [7, Theorem 3.1] 8 A. Gelb and E. Tadmor 1. Trigonometric factors. We consider concentration factors of the form with the proper normalization Si(ff) := R ff(sin j=j)dj. The edge detector introduced originally by Banerjee & Geer, [1] corresponds to oe -); the general case is found in [7, x3.2]. 2. Polynomial factors. As a first example consider oe(-. In this case, K x corollary 2.2 recovers Fej'er's result, [23, xIII Theorem 9.3], with the following error estimate Const \Delta log N This is the first member of a whole family of polynomial concentration factors, e.g., [7, x3.4], which correlate to concentration kernels satisfying (2.4), (2.5), and (2.6). For odd p's, K oe p N (f ); for even p's, K oe p These edge detectors were introduced in [9] and were recently analyzed by Kvernadze in [13]. Corollary 2.2 yields i- Const \Delta log N The last error estimate is (essentially) first order. It is sharp. It was noted in [7, x3.4], however, that oe p 's with higher p's lead to faster convergence rate at selected interior points, bounded away from the singularities of f . This leads us to the next example of 3. Exponential factors. Polynomial concentration factors (of odd degree) correspond to differentiation in physical space; trigonometric factors correspond to divided differences in the physical space - consult the original derivation in [1]. Our main result stated in Corollary 2.2 provides us with the framework of general concentration kernels which are not necessarily limited to a realization in the physical space. In particular, we seek concentration factors, oe(\Delta), which vanish at to any prescribed order, The higher p is, the more localized the corresponding concentration kernel, K oe becomes. Here is why. Evaluating K oe N (t) at the equidistant points t K oe sin 2-k' we observe that K oe coincide with the '-discrete Fourier coefficient of oe(\Delta); since oe(-) and its first p-derivatives vanish with at both ends, there is a rapid decay of its (discrete) Fourier coefficients, j-oe ' j - Const:' \Gammap , Enhanced detection of edges in spectral data 9 Thus, for t away from the origin, K oe N (t) is rapidly decaying for large enough N 's. Moreover, we claim an increasing number of moments of K oe vanish. To this end we consider the odd moments of K oe N (\Delta) (- its even moments vanish, of course). With - sin kt \Gamma2 Z -N Z -N Integrate by parts - respectively, sum by parts the summation on the right of (2.23). Thanks to (2.22) the boundary terms vanish and we have sin(-N-) d p\Gammaj d- p\Gammaj As an example, we consider the exponential concentration factors oe exp Const Z exp normalized so that 1. Here, the C 1 concentration factor oe exp (-) vanishes exponentially at both ends, so that (2.22) holds for all p's. Figure 2.3, 2.4 confirms the improved localization of these exponential concentration factors. 4. Band pass filter. Bauer [3] have considered a family of what he termed as 'band pass filter', supported in the range of middle frequencies, say suppj ae [1=4; 3=4]. We note in passing that these are special cases of p-order admissible concentration factors, (2.22), although the normalization used in ([3, eq. (1.35)]), R prevented the recovery of the amplitude of the jumps. To demonstrate the detection of edges by the concentration factors outlined above, consider the following two examples of discontinuous f 's (defined on [\Gamma-]): f a (x) := \Gammasgnx \Delta cos( x In both cases, f a (x) and f b (x) are recovered from their Fourier coefficients using the Fourier partial sums SN [f ](x), and we wish to recover their jump discontinuities [f a 0; else. 0; else: Figures 2.1 and 2.2 demonstrate the use of trigonometric and polynomial concentration factors for the detection of edges from Fourier spectral data. A. Gelb and E. Tadmor -3.1 -2.1 -1.1 -0.1 0.9 1.9 2.9 x -2.2 -1.2 -3.1 -2.1 -1.1 -0.1 0.9 1.9 2.9 x Figure 2.1: Trigonometric concentration factor for (left) f a (x) where the exact jump value is where the exact jump values are [f ](\Sigma - 2. -3.1 -2.1 -1.1 -0.1 0.9 1.9 2.9 x -2.2 -1.2 -3.1 -2.1 -1.1 -0.1 0.9 1.9 2.9 x Figure 2.2: Jump value obtained by the polynomial concentration factor oe p=1 (- for (left) f a (x) and (right) f b (x). Enhanced detection of edges in spectral data 11 Figure 2.3: Edge detection using the exponential concentration factor oe exp vs. oe p=1 for S 40 f a (x) and (right) oe exp for SN f a (x) with modes. -0.50.5s p1 Figure 2.4: Edge detection using the exponential concentration factor oe exp vs. oe p=1 for S 40 f b (x) and (right) oe exp for SN f b (x) with modes. As noted in [10, x3.4], polynomial factors of higher degree yields improved results away from the jump discontinuities. Indeed, the corresponding concentration kernel, K oe p N (\Delta) have additional vanishing moments. In the limit, one arrives at exponential factors, K oe exp Figures 2.3,2.4 demonstrate the edge detection of these factors in Fourier expansions SN f a (x) and SN f b (x). The improved localization is evident, due to the faster convergence rate in the smooth parts of f 's. In particular, the superiority of the exponential factors is illustrated in the figures on the left, when compared with the first-order accurate polynomial concentration factor, oe p=1 (-. At the same time, Gibbs A. Gelb and E. Tadmor oscillations can be noticed in the vicinity of the jump discontinuities. 3 Edge Detection in Non-periodic Projections Consider a piecewise smooth f(\Delta). To simplify our presentation, we assume f experiences a single jump discontinuity at The localization property of the appropriate concentration kernel in the presence of a single jump applies to the case with finitely many jump discontinuities. We begin with an alternative derivation of our results for the periodic case. 3.1 Revisiting the periodic case If a 2-periodic f(\Delta) experiences a single jump, [f ](c), then it dictates the Fourier coefficients decay, To extract information about the location of the jump from the phase of the leading term, we examine the special concentration kernel, K oe N with oe(-, where K - ik [f ](c) e ik(x\Gammac) Here we used the concentration property of the Dirichlet kernel localized at O The same property applies to the class of concentration factors, oe(-), such that (2.15) holds, -( O It then follows that the corresponding K oe N in (2.13) is an admissible concentration kernel, so that K oe 3.2 Non-periodic expansions 3.2.1 General Jacobi expansions We begin with the Jacobi expansion of a piecewise smooth f(\Delta), Enhanced detection of edges in spectral data 13 Here are the Jacobi polynomials - the eigenfunctions of the singular Sturm Liouville problem with corresponding eigenvalues - Different families of Jacobi polynomials are associated with different weight functions To simplify the computations, we assume that the P k 's are normalized so that kP k As in the periodic case, integration by parts (against (3.28)) shows that a single jump disconti- nuity, [f ](c), dictates the decay of the Jacobi coefficients, To extract information about the location of jump, we consider the conjugate sum of the following -( \Theta P 0 corresponding to concentration factors oe(-). We shall focus our attention on the particular case \Theta P 0 This is the non-periodic analogue of the Fourier concentration kernel K - with the additional pre-factor weight of We want to quantify the localization property of the last summation. To this end we note that if fP (ff) k (x)g are the Jacobi polynomials with respect to the weight function ! ff (x), then fP 0 k (x)g are the Jacobi polynomials w.r.t. the modified weight function ! fi -orthogonality follows from integration by parts of (3.28) against P (ff) k . Thus, The coefficients C k;fi are determined by normalization where by using (3.28) once more we find and hence we set C so that fP (fi) is the orthonormal family w.r.t. ! fi weight. Inserted into the leading term of (3.31), we end up with a Jacobi kernel associated with weight function 14 A. Gelb and E. Tadmor \Theta (c)P (fi) \Theta We rewrite this as \Theta KN (c; x): (3.32) By virtue of Christoffel-Darboux formula, e.g., [19, Theorem 3.2.2], the kernel KN (c; x) is given by KN and it remains to quantify the concentration property of KN (c; x). To this end we use the asymptotic behavior of P (fi) N which is stated as 3 denotes the separation between the interior and boundary regions. Using this to upper bound KN (c; x) in (3.33), we find \Theta 1 The upper bound on the right is in fact the leading term in the asymptotics of KN (c; x) for large N 's as long as Similarly, the behavior at The desired concentration property now follows, similar to the localization of the periodic Dirichlet kernel DN (3.27). We restrict our attention to interior jumps, so that for 3 The first term on the right pf (3.34) follows from the classical asymptotic formula, e.g., [19, Theorem 12.1.4], which tells us the behavior of the L 2 -normalized P (fi) N (x) at the interior The second term on the right of (3.34) reflects the fact that as x approaches the \Sigma1-boundaries, the L 2 -normalized approaches to its maximal value e.g., [19, 4.7.3, 4.7.15] r Enhanced detection of edges in spectral data 15 large enough, c (3.35), (3.36) and (3.32) yield \Theta KN (c; x) - O \Theta 1 We summarize by stating Corollary 3.1 Let SN (f) denote the truncated Jacobi expansion (3.27) of a piecewise smooth f , associated with a weight function the concentration property Const \Delta log N It is instructive to examine the above discussion for the special case of Chebyshev expansion corresponding to dx: (Observe that except for Chebyshev expansion, the concentration bound (3.37) deteriorates as we approach the boundaries, depending whether jxj - 1 for ff 1.) The conjugate sum corresponding to (3.32) reads k (c) In this case, we can sum the corresponding Chebyshev kernel: setting O( 1 [f 3.2.2 Chebyshev expansion Our discussion above on edge detection in the non-periodic expansions is based on expansion of the Jacobi coefficients to their leading order in (3.29). More precise information is obtained using the general framework introduced in the main Theorem 2.1. Corollary 3.2 Let f(\Delta) be a piecewise smooth function with Chebyshev expansion SN f(x) - Consider the concentration factors, oe(-), with -(\Delta) normalized so that A. Gelb and E. Tadmor Then K oe admits the concentration property (2.3), and the following estimate holds -( Const \Delta log N Proof. With a piecewise smooth f(x) defined over the interval [\Gamma1; 1] we utilize the usual Chebyshev transformation We consider the even extension f(cos '); \Gamma-. Using Theorem 2.1 along the lines of Corollary 2.2, we find that the odd concentration kernel, K oe recovers the jumps of f(cos '), i.e., log N computation shows the sum on the left equals -( and the result follows. We turn to numerical examples. The following tables summarize our results for the edge detection in Legendre expansion , corresponding to ff = 0, and in Chebyshev expansion, corresponding to \Gamma1=2. Scaled to the unit interval [\Gamma1; 1], we consider f a ( x ). The results confirm the linear convergence rate stated in Corollary 3.1, both away from the jumps - consult Tables 3.1 and 3.2, as well as at the jump itself, Table 3.3. N Legendre expansion Chebyshev expansion Table 3.1: Pointwise error estimate j- away from the jump discontinuity at We note that the critical threshold must be very high for to eliminate the artificial jumps. This indicates that 40 nodes are not enough to resolve the jumps of f b (x) in either the Chebyshev or Legendre case. N Legendre expansion Chebyshev expansion Table 3.2: Pointwise error estimate j- away from the jump discontinuities at Enhanced detection of edges in spectral data 17 Legendre Chebyshev Table 3.3: Pointwise error estimate j- [f ](c)j at the point(s) of discontinuity, It is clear from tables 3.1 and 3.2 that convergence is nonuniform at the boundaries. We have observed in our numerical experiments, that the edge detector, - experiences larger oscillations near the boundaries which do affect the linear convergence rate there. In this context we note the dependence of the error bounds on the smoothness of f( The first-order convergence is re-confirmed, in table 3.4 below, when measuring the L 1 -error away from the jumps discontinuities (and up to the boundaries) . N Legendre Chebyshev Table away from discontinuities. 4 Nonlinear Enhancement The detection of edges in Theorem 2.1 is based on separation of scales. Thus, consider for example a piecewise smooth f with finitely many jump discontinuities at . If K ffl is an admissible concentration kernel, then jK ffl f(x)j !! 1 for x away from these jumps, where as at ae O(ffl); x The last statement refers to the asymptotic behavior of the concentration kernel as a function of the small parameter ffl # 0. In this section we outline a new, nonlinear enhancement procedure, which is easily implemented to 'pinpoint' finitely many edges in piecewise smooth f 0 s. To this end we enhance the separated scales in (4.41) by considering Const By increasing the exponent p ? 1, we enhance the separation between the vanishing scale at the points of smoothness (- of order O(ffl p )), and the growing scale at the jumps (- of order Next one must introduce a critical threshold which will eliminate all the unacceptable jumps. Only those edges with amplitudes larger than the critical threshold, [f ](x) ? J 1=p crit ffl, will be detected. A. Gelb and E. Tadmor Thus crit is a measure which defines the small scale in our computation of edge detection. We note that data dependent and is typically related to the variation of the smooth part of f . Given this critical threshold, we form our enhanced concentration kernel K ffl;J [f ae Clearly, with p large enough, one ends up with a sharp edge detector where K ffl;J [f at all but O(ffl) neighborhoods of the jumps In practical applications, a moderate enhancement exponent, p - 5 will suffice. We consider two examples. 1. The quadratic filter. Consider the peaked concentration kernel (2.10) K ffl ffl (t). Then, with one finds the so called quadratic filter [12],[22], where 2. Enhanced spectral concentration kernels. We apply the procedure of nonlinear enhancement in conjunction with spectral concentration kernels K sin kt by considering the corresponding enhanced spectral concentration kernel K oe ae K oe The enhanced spectral concentration kernel depends on four ingredients which are at our disposal ffl The number of modes, N ffl The enhancement exponent, p ffl The critical threshold, J ffl The concentration factor, oe(-). Figures 4.5 and 4.6 demonstrate the enhancement procedure to the spectral detection of edges depicted earlier in the corresponding Figures 2.2 and 2.1. Enhanced detection of edges in spectral data 19 -3.1 -2.1 -1.1 -0.1 0.9 1.9 2.9 x -2.2 -1.2 -3.1 -2.1 -1.1 -0.1 0.9 1.9 2.9 x Figure 4.5: Jump value obtained by applying the polynomial concentration factor oe(- with where the exact jump value is and (b) f b (x) where the exact jump values are [f ](\Sigma - 2. -3.1 -2.1 -1.1 -0.1 0.9 1.9 2.9 x -2.2 -1.2 -3.1 -2.1 -1.1 -0.1 0.9 1.9 2.9 x Figure obtained by applying the trigonometric concentration factor oe 1 with modes and enhancement exponent where the exact jump value is where the exact jump values are [f ](\Sigma - 2. We conclude with non-periodic examples. In Figures 4.7 we show the detection of a single edge in f a (x=-) from its Legendre expansion, f a (x). The detection in Chebyshev expansion is A. Gelb and E. Tadmor shown in Figure 4.8 for f b (x=-). In both cases we used an enhancement factor critical threshold -0.5x -0.5x Figure 4.7: Detection of edges in Legendre expansion of f a (x=-) with exact jump value is [f a (left) before and (right) after enhancement with x x Figure 4.8: Detection of edges in Chebyshev expansion of f b (x=-) with exact jump value is [f b ](\Sigma before and (right) after enhancement with Concluding Remarks Accurate reconstruction of piecewise smooth functions from their spectral projections is only plausible when the location (and amplitude) of the underlying jump discontinuities are known, consult Enhanced detection of edges in spectral data 21 [1],[7],[5],[6],[16],[11] and references therein. Theorem 2.1, and its corollaries 2.2, and 3.1 provide the general framework for the detection of edges from spectral data, in both periodic and non-periodic cases. The detection is based on admissible concentration kernels which include as particular cases classical examples of Fej'er as well as additional examples in recent literature, [1],[9],[13]. In particular, we introduce here a new family of exponential concentration kernel, (2.24), with a superior convergence rate away from the edges. A linear convergence rate is observed near the detected edges. We also introduce a nonlinear enhancement (4.43) procedure which enables one to "pinpoint" edges with amplitude larger than a critical threshold. Recently the edge detection and enhancement method was applied to non-linear conservation laws, [8], as a post-processing tool to improve the overall convergence rate of the spectral viscosity solution. Since the edge detection occurs only at the post-processing stage, very little cost is added to the procedure yet the results are dramatically improved. Future applications, in both one- and several space dimensions, will also include image processing, where edge detection is needed to de-noise the contamination by the O(1)-Gibbs' oscillations in the neighborhoods of the undetected edges. --R Exponential approximations using Fourier series partial sums Treatise of Band filters for determining shock locations Introduction to the Theory of Fourier's Series and Integrals Accurate reconstructions of functions of finite regularity from truncated series expansions On a high order numerical method for functions with singularities Detection of edges in spectral data Enhanced spectral viscosity method for nonlinear conservation laws Determination of the jump of a function of bounded p-variation by its Fourier series "Progress and Supercomputing in Computational Fluid Dynamics" On the Gibbs phenomenon and its resolution Determination of the jump of a bounded function by its Fourier series On the detection of singularities of a periodic function The Fourier method for nonsmooth initial data Multiresolution approximations and wavelets orthonormal bases of L 2 (R) Convergence of spectral methods for nonlinear conservation laws Family of spectral filters for discontinuous problems Asymptotic behavior of quadratic edge filters Cambridge University Press --TR --CTR R. Pasquetti, On inverse methods for the resolution of the Gibbs phenomenon, Journal of Computational and Applied Mathematics, v.170 n.2, p.303-315, 15 September 2004 Anne Gelb, Parameter Optimization and Reduction of Round Off Error for the Gegenbauer Reconstruction Method, Journal of Scientific Computing, v.20 n.3, p.433-459, June 2004 Rick Archibald , A. Gelb, Reducing the Effects of Noise in Image Reconstruction, Journal of Scientific Computing, v.17 n.1-4, p.167-180, December 2002 Anne Gelb, A Hybrid Approach to Spectral Reconstruction of Piecewise Smooth Functions, Journal of Scientific Computing, v.15 n.3, p.293-322, Sept. 2000 Bernie D. Shizgal , Jae-Hun Jung, Towards the resolution of the Gibbs phenomena, Journal of Computational and Applied Mathematics, v.161 n.1, p.41-65, 1 December Scott A. Sarra, The spectral signal processing suite, ACM Transactions on Mathematical Software (TOMS), v.29 n.2, p.195-217, June Scott A. Sarra, Chebyshev super spectral viscosity method for a fluidized bed model, Journal of Computational Physics, v.186
piecewise smoothness;concentration kernels;spectral expansions
588637
Multiplicative Schwarz Algorithms for the Galerkin Boundary Element Method.
We study the multiplicative Schwarz method for the h- and the p-version Galerkin boundary element method for a hypersingular and a weakly singular integral equation of the first kind. For both integral equations we prove that the contraction rate of the multiplicative Schwarz operator is strictly less than 1 for the h-version for the two level and the multilevel methods, whereas for the p-version we show that the contraction rate approaches one only logarithmically in p for the 2-level method. Computational results are presented for both the h-version and the p-version which support our theory.
Introduction . We study multiplicative Schwarz methods for the h and p versions of the Galerkin boundary element method applied to hypersingular and weakly singular integral equations on open or closed curves. These equations are integral reformulations for boundary value problems with the Laplace equation and the Dirichlet or Neumann boundary conditions (see [4]). Application of the Galerkin method to solve these integral equations yields linear systems with dense, symmetric and positive-definite stiffness matrices. Since the condition numbers of these matrices grow like h \Gamma1 in the h version and like p 2 in the p version (see [10]), the convergence rate of the conjugate gradient algorithm approaches 1 (the non-convergent status of the iterative method) like ch 1=2 as positive constant c. We shall propose for both versions multiplicative Schwarz algorithms which significantly improve the rate of convergence of the conjugate gradient method. We shall analyze for the h-version a 2-level method which corresponds to a multigrid algorithm using the Gau-Seidel algorithm for smoothing, and a multilevel method which corresponds to a multigrid algorithm using the Jacobi smoother. We shall prove that both the Institut f?r Angewandte Mathematik, University of Hannover, Germany y Institut f?r Angewandte Mathematik, University of Hannover, Germany z School of Mathematics, University of New South Wales, Sydney, 2052, Australia 2-level and multilevel methods yield an error reduction which is independent of the mesh sizes and number of levels. This is an improved result compared with [8]. For the p-version we propose a 2-level method which has an error reduction factor approaching one like The analysis is based on the abstract framework of [3] (see also [13]) which requires two main ingredients. The first ingredient is estimates for the extremum eigenvalues of the additive Schwarz operator which is correspondingly defined from the multiplicative operator. For the boundary integral operators considered in this paper, these estimates were recently obtained for both versions [10, 11]. The second ingredient is a strengthened Cauchy-Schwarz inequality. We will in this paper prove inequalities of this type for both versions for the hypersingular integral equation, and for the p-version for the weakly singular equation. A strengthened Cauchy-Schwarz inequality for the h-version of the Galerkin method applied to the weakly singular integral equation is still an open question to us. It is noted that since the equations considered in this paper yield dense stiffness matrices, the proofs of these inequalities are much more complicated than those appear in the finite element method where differential operators are considered which yield sparse matrices. In Section 2 we introduce the gereral setting of multiplicative Schwarz methods and recall the abstract analysis from [3, 13]. We prove in Section 3 strengthened Cauchy-Schwarz inequalities for the hypersingular equation (Lemmas 3.2 and 3.12) and appropriately apply the result of Section 2 to obtain estimates for the rates of convergence (Theorems 3.3, 3.9, and 3.13). Similar treatment for the weakly singular equation is proceeeded in Section 4 (Lemma 4.2 and Theorem 4.3). In Section 5 we present our numerical results which clearly underline the the- ory. Some useful lemmas which are used in the proofs of the strengthened Cauchy-Schwarz inequalities are proved in the Appendix. For simplicity of notation, the integral equations considered in this paper are defined on the interval (\Gamma1; 1). A generalization to a polygonal curve is straight forward. General setting of multiplicative Schwarz methods. Multiplicative (and additive) Schwarz methods are in general defined via a subspace decomposition of the space of test and trial functions together with projections onto these subspaces. More precisely, let and let projections defined by Here a(\Delta; \Delta) is a symmetric and positive-definite bilinear form on V . The multiplicative Schwarz operator is then defined as where I is the identity map and is the error propagation operator. We note that is the corresponding additive Schwarz operator. By defining we obtain which in turn yields We now present in this section some results which were mainly proved in [3] and [13]. We include the proofs here for completeness. These results will be used in the analysis of the following sections. Lemma 2.1 For any v 2 V there holds Proof. We have, for or equivalently Summing up we obtain the desired result. 2 The analyses in [3] and [13] suggest that we would have to prove a strengthend Cauchy-Schwarz inequality corresponding to the decomposition (1). The following lemma shows that it is possible to avoid the coarse grid subspace V 0 in this inequality and use instead a bound for the maximum eigenvalue of the additive Schwarz operator. Lemma 2.2 Let \Theta be an whose elements are defined by If there exist positive constants C 1 and C 2 such that a(P AS v; v) - C 2 a(v; v) 8v 2 V; (3) and that then there holds for any v 2 V a(P AS v; v) - C 1 Proof. Using (2) we obtain implying Cauchy-Schwarz's inequality implies !1=2 On the other hand it follows from the definitions of ' ij and the k \Delta k 2 -norm of matrices that !1=2 Inequalities (4), (5), and (6) yield This inequality and (3) imply The lemma is proved. 2 Theorem 2.3 If there exist positive constants C and a constant C 1 satisfying then there holds a - a is the norm given by the bilinear form a(\Delta; \Delta). Proof. For any v 2 V we have using Lemma 2.2 Using Lemma 2.1 we deduce which implies a - The theorem is proved. 2 Remark 2.4 If we use the multiplicative Schwarz method as a preconditioner for the CG- scheme, we have to use the symmetrized version The condition number is bounded by a a Remark 2.5 C 0 is a lower bound for the minimum eigenvalue and C 2 is an upper bound for the maximum eigenvalue of the additive Schwarz operator P AS . Multiplicative Schwarz methods for the hypersingular integral equation. We consider the hypersingular integral equation f:p: Z ds where f.p. denotes a finite part integral in the sense of Hadamard. Let e \Gamma be an arbitrary closed curve containing \Gamma. We define, as in [7], the Sobolev spaces e loc and H \Gamma1=2 (\Gamma) being the dual space of e H 1=2 (\Gamma) with respect to the L 2 inner product on \Gamma. As was shown in [4], D is continuous and invertible from e H 1=2 (\Gamma) to H \Gamma1=2 (\Gamma). Moreover, D is strongly elliptic, i.e., there exists a constant fl ? 0 such that where h\Delta; \Deltai denotes the L 2 duality on \Gamma. Hence D defines a continuous, symmetric and positive-definite bilinear form a(v; 3.1 The h-version. We consider a uniform mesh of size h on \Gamma and define on this mesh the space V h of continuous piecewise-linear functions on \Gamma which vanish at the endpoints of \Gamma. We note that V h is a subset of e (\Gamma). The h-version boundary element method for Equation (7) reads as: Find h such that The stability and convergence of the scheme (9) was proved in [12]. It is known that the condition number of the matrix system derived from (9) is N 2 . We show in this paper that the multiplicative Schwarz method yields a preconditioned system which has convergence rate strictly less than 1. 3.1.1 2-level method. Let OE h;j , the hat functions forming a basis for V h . We then decompose where VH is defined as V h with mesh size g. For notational convenience we identify respectively. The projections and the operators PMS and P AS are then well defined. Our task is to verify the assumptions of Theorem 2.3. The following result was proved in [11]. Lemma 3.1 There exist constants C independent of h such that for any v 2 V h Let \Theta be an whose elements are defined by It follows from the Cauchy-Schwarz inequality that 0 1. Therefore, in general there holds In view of Theorem 2.3 we will prove in the next lemma that the bound is indeed independent of N . During the course of the proof, a strengthened Cauchy-Schwarz inequality is proved. Lemma 3.2 There exists a positive constant C 3 such that Proof. First we note that u denotes the derivative of u with respect to the arc-length, and Z log Let z 2 log jzj for z 6= 0; Then by using the formula Z d c a log we can prove where This implies log 2X and hence it suffices to prove that To do so, we define to obtain Since f is a concave function on (1; 1), which results in G(m) ! 0 for m - 2. By the mean value theorem there exist ' and ' 0 2 (0; 1) such that if (For the mean value theorem argument used above, the reader is referred to [1, p. 101].) We note that Since we deduce from (12) and (13) for any m - 6; which in turn yieldsX x x dx +X This completes the proof of the lemma. 2 As a consequence we obtain, due to the abstract Theorem 2.3, Theorem 3.3 Let C 0 , C 2 , and C 3 be given by Lemmas 3.1 and 3.2, and let C g. Then for any v 2 V h there holds Remark 3.4 Due to the subspace decomposition (10) we see that this 2-level multiplicative Schwarz algorithm corresponds to the unsymmetric 2-level multigrid algorithm using the Gau- Seidel smoother. 3.1.2 Multilevel method. We shall in this subsection design a multilevel method for the hypersingular equation. The analysis for this method is slightly different from that of the general framework in Section 2. The main reason for this difference is the non-availability of a strengthened Cauchy-Schwarz inequality which can yield an estimate like Lemma 3.2. Starting with a coarse mesh we divide each subinterval into two equal intervals. Hence, if h l is the meshstep of N l , . For l be the spline space associated with N l which contains continuous and piecewise-linear functions vanishing at the endpoints \Sigma1. Let OE l l be the nodal basis for V l , where N is the dimension of V l . We then decompose V l as l l with V l l . Eventually, V L is decomposed as l l Let l for i is defined for any v 2 V L by and let Y The multilevel multiplicative Schwarz operator is now defined as Analogously, we define Letting we deduce which in turn yields l The following lemma follows easily in the same manner as Lemma 2.1. Lemma 3.5 For any v 2 V L The following two lemmas were proved in [11]: Lemma 3.6 There exists a positive constant C 0 independent of h and L such that for any there holds Lemma 3.7 There exist constants such that for any v 2 V k where there holds The above lemma plays the role of Lemma 3.2 in the analysis of the multilevel method. We also need the following technical lemma: Lemma 3.8 For any there holds where C 1 is the constant given by Lemma 3.7. Proof. By using the Cauchy-Schwarz inequality for the symmetric and positive-definite bilinear form a(T k \Delta; \Delta), and Lemma 3.7 we obtain which implies The lemma is proved. 2 Theorem 3.9 Let C 0 , C 1 , and fl be given by Lemmas 3.6 and 3.7. Then for any v 2 V L where Proof. The theorem is proved if we can prove that By Lemma 3.5 it suffices to prove We have from Lemma 3.6, the Cauchy-Schwarz inequality, the inequality 2ab - a noting that a(T l v; v) It follows successively from (16), the Cauchy-Schwarz inequality, Lemmas 3.7 and 3.8 that Inequalities (18) and (19) yield (17) and theorem is proved. 2 Remark 3.10 Due to the subspace decomposition (14) this multilevel multiplicative Schwarz algorithm corresponds to the unsymmetric multigrid algorithm using the Jacobi-smoother. 3.2 The p-version. We shall in this subsection design a multiplicative method for the p-version of the Galerkin boundary element method applied to the hypersingular integral equation. We define on the mesh (8) the space V p of continuous functions on \Gamma whose restrictions are polynomials of degree at most p, p - 1. In order to guarantee that these functions belong to e H 1=2 (\Gamma), we also require that the functions vanish at the endpoints \Sigma1 of \Gamma. For the p-version of the Galerkin scheme, we approximate the solution of (7) by functions in V p and increase the accuracy of the approximation not by reducing h (which is fixed) but by increasing p. More explicitly, the p-version boundary element method for Equation (7) reads as: Find u such that The stability and convergence of the scheme (20) was proved in [9]. Note that the dimension of Choosing a basis for V p , we derive from (20) a system of equations to be solved for u p . In practice, we use the following basis. Let OE k , defined as hat functions satisfying For we also define L q;j as the affine image onto \Gamma j of where L q\Gamma1 is the Legendre polynomial of degree q \Gamma 1. We extend L q;j by 0 outside \Gamma j . It is clear that is a basis for V p . Solving the equation (20) amounts to solving where the matrix AN has entries a(v; w) with v; w 2 B. The condition number of (21) grows at least like p 2 and at most like p 3 . We will define a multiplicative Schwarz method to solve instead of (21) a preconditioned system which has condition number growing significantly slower than We decompose V p as a direct sum where the space of continuous piecewise-linear functions on \Gamma vanishing at \Sigma1, and The space V 0 serves the same purpose as the coarse grid space in the h-version. We note that functions in V p are supported in - With the projections P j appropriately defined as in Section 2, we can define the multiplicative and additive Schwarz operators as where Analogously to Lemma 3.1, we have Lemma 3.11 [10] There exist constants C independent of p and N such that for any there holds Our next task is to show a strengthened Cauchy-Schwarz inequality for this version. Lemma 3.12 Let V p i be defined as in (23) and let Then there exists a constant C independent of i; j; p such that If 1-i;j-N , then there holds where Proof. Let l=2 v l L l;j . Then we have Due to Lemma A.6 (compare with [5, Lemma 1]) we have with From (24) and (25) we obtain \Theta \Theta oe For oe oe Due to Lemma A.1 we have (note the scaling)@ p Using (26), (27) and (28) we obtain for Due to the Cauchy-Schwarz inequality we have independent of p and N . Therefore, for any there holds Finally we have This completes the proof of the lemma. 2 Using the abstract Theorem 2.3 we can prove Theorem 3.13 Let C 0 , C 2 , and C 3 be given by Lemmas 3.11 and 3.12. If C 1 then for any v 2 V p there holds Amultiplicative method for the weakly singular integral equation We shall now in this section design a multiplicative algorithm for the p-version Galerkin method applied to the weakly integral equation of the form Z log ds As was shown in [4], V is continuous and invertible from e H \Gamma1=2 (\Gamma) to H 1=2 (\Gamma). Here the Sobolev space H 1=2 (\Gamma) is defined as the space of functions which are traces of functions in H 1 loc e H \Gamma1=2 (\Gamma) is its dual. It is known that there exists a constant fl ? 0 such that e where h\Delta; \Deltai denotes the L 2 duality on \Gamma. Hence V defines a continuous, symmetric and positive-definite bilinear form a(v; H \Gamma1=2 (\Gamma). We define on the mesh (8) the space - of piecewise continuous functions on \Gamma whose restrictions on \Gamma j := are polynomials of degree at most p, p - 1. For the p-version of the Galerkin scheme, we approximate the solution of (29) by functions in - increase the accuracy of the approximation not by reducing h (which is fixed) but by increasing p. More explicitly, the p-version boundary element method for Equation (29) reads as: Find u p such that The stability and convergence of the scheme (30) was proved in [9]. Note that the dimension of Choosing a basis for - we derive from (30) a system of equations to be solved for u p . In practice, we use the following basis. Let OE k , be defined as piecewise constant functions satisfying For we also define L q;j as the affine image onto \Gamma j of the Legendre polynomial L q of degree q. We extend L q;j by 0 outside \Gamma j . It is clear that is a basis for - . Solving the equation (30) amounts to solving where the matrix AN has entries a(v; w) with v; w 2 B. The condition number of (31) grows at least like p 2 and at most like p 3 . We will define a multiplicative Schwarz method to solve instead of (31) a preconditioned system which has condition number growing significantly slower than We decompose - as a direct sum where the space of piecewise constant functions on \Gamma, and The space - serves the same purpose as the coarse grid space in the h-version. We note that functions in - are supported in - With the projections P j defined appropriately as in Section 2 we can define the multiplicative and additive Schwarz operators as where Similarly to Lemma 3.11, we have Lemma 4.1 [10] There exist constants C 0 and C 2 independent of p and N such that for any there holds Again our next task is to prove a strengthened Cauchy-Schwarz inequality. Lemma 4.2 Let - i be defined as in (33), and let there exists a constant C independent of i; j; p such that Moreover, if where C 3 := C- 2 =3. Proof. Let Due to Lemma A.6 (compare with [5, Lemma 1]) we have with From (34) and (35) we obtain l l \Theta \Theta oe For oe oe Due to Lemma A.3 we have for (note the scaling) Using (36), (37) and (38) we obtain for Due to the Cauchy-Schwarz inequality we have independent of p and N . Therefore, for any there holds Finally we have This completes the proof of the lemma. 2 Using the abstract Theorem 2.3 we can prove Theorem 4.3 Let C 0 , C 2 , and C 3 be given by Lemmas 4.1 and 4.2. If C 1 then for any v 2 - there holds 5 Numerical results. We consider the hypersingular integral equation (7) with the right hand side f(x) j 1. We solve the Galerkin equations (20) for the p-version by the multiplicative Schwarz algorithm, for different subspace decompositions and observe the expected behavior of the convergence rate with respect to N and p (see Tab. 1). For the h-version we note the considerably lower contraction rates of the 2-level method compared with the multilevel method in Tab. 3. But if we take into account the high costs of solving a system of mesh width in each iteration step we see that the multilevel method is the far superior method. Analogously we consider the weakly singular integral equation (29) with the right hand side We solve the Galerkin equations (30) for the p-version by the multiplicative Schwarz algorithm. The numerical results in Tab. 2 show the expected behavior of the convergence rate. In the case of the h-version we observe that the elements of the Galerkin matrix x f:p: x ds y ds x =: a i\Gammaj depend only on the difference for an uniform mesh. Therefore we can reduce the memory used to store the Galerkin matrix from O(N 2 ) to O(N ). In a more general case we have to use a clustering or multipole technique to reduce the amount of memory needed. This also reduces the amount of time for computing the Galerkin matrix. On a vectorcomputer SNI VPP 300/4 we have achieved a performance of 1000 MFlops/s. In the case of the p-version we calculate the elements of the Galerkin matrix analytically. Due to the smaller size of the Galerkin matrix we can store the full matrix in the main memory. We have to note that our subspace decomposition in this case is actually a reordering of the basis functions. Therefore the projections involved are simplyfied considerably. Contraction rate 9 0.6245 0.6383 0.6524 Table 1: Hypersingular integral equation, p-version Contraction rate 9 0.6401 0.6576 0.6745 Table 2: Weakly singular integral equation, p-version A Appendix p. Then there exists a constant C Contraction rate 2-level multilevel Table 3: Contraction rates for the hypersingular integral equation, h-version independent of u and p such that ~ Proof. From the definition of the interpolation norm k \Delta k ~ by the real K-method [2] we have ~ Z 1' u=v+w We have Therefore it is sufficient to take the infimum in H 1 ~ Z 1t \Gamma2 dt: (40) Due to v 2 H 1 0 (I) we can expand v also in antiderivatives of Legendre polynomials For the norms in (40) there hold and, due to Lemma A.2, Let l and l 0 Using the the monotonicity of the square function and the inequality a when we obtain from (40), (41), and (42) ~ Z 1t \Gamma2 dt Z 1t \Gamma2 inf dt Z 1t \Gamma2 inf dt Z 1t \Gamma2 inf (v k )2l / dt Z 1t \Gamma2 inf (v k )2l 0/ dt Z 1t \Gamma2 dt: (43) Note that if a+b . Therefore we obtain from equation (43) ~ Z 1t \Gamma2 dt Z 1t \Gamma2 dt Z 1Ck \Gamma5 2 dt -s This completes the proof. 2 . Then there holds Proof. With the normalized antiderivatives of Legendre polynomials defined by L there holds due to the proof of [6, Lemma 3.1] Due to there holds This completes the proof. 2 Lemma A.3 Let (x=h). Then there exists a constant C independent of u, h and p such that ~ Proof. Due to ~ kuk ~ For there holds kuk ~ Due to Lemma A.4 we have This completes the the proof. 2 (x=h). Then there exists a constant C independent of v, h and p such that Proof. Let 1). From the definition of the interpolation norm k \Delta k H 1=2 (\Gammah;h) by the real K-method [2] we have Z 1' v=w+a Z 1i Z 1t \Gamma2 inf w Z 1t \Gamma2 inf w Z 1t \Gamma2 inf w Z 1t \Gamma2 inf w By the orthogonality property of the Legendre polynomials we have On the other hand It follows from the definition of matrix norms that Lemma A.5 and a straightforward calculation give 2: (50) Inequalities (48), (49), and (50) yield This inequality and (47) imply Z 1t \Gamma2 inf w k dt Z 1t \Gamma2 inf w k dt Z 1t \Gamma2 Lemma A.5 Let k; m - 1. Then we have Proof. From the recurrence formula of the Legendre polynomials we have which results in and be the Legendre polynomial of degree p i linearly transformed onto the open straight line I ae IR 2 . L p i ;I is supposed to be continued by 0 outside I on the entire line containing I where necessary. Let J ae IR 2 be another open straight line with - there exists a constant C such that Proof. Let x 0 be the midpoint of I and y 0 be the midpoint of J . Then there holds using the Taylor expansion of log jx \Gamma yj log y log jx and y log with 1). Due to the orthogonal properties of the Legendre polynomials the first sum vanishes and we have Z I Z ds y ds x Z I Z ds y ds x : (52) Applying the Cauchy-Schwarz inequality two times we obtain 'Z I ds x ds y I Z ds y ds x 'Z I Z ds y ds x sup Since y log there holds sup (jI (jI Inequalities (53) and (54) imply This completes the proof. 2 --R Mathematical Analysis: A Modern Approach to Advanced Calculus Convergence Estimates for Product Iterative Methods with Applications to Domain Decomposition Boundary Integral Operators on Lipschitz Domains: Elementary Results Efficient Algorithms for the p-Version of the Boundary Element Method A Multilevel Additive Schwarz Method for the h-p Version of the Galerkin Boundary Element Method Multigrid Solvers and Preconditioners for First Kind Integral Equations On the Convergence of the p-Version of the Boundary Element Galerkin Method Additive Schwarz Algorithms for the p-Version Boundary Element Method Additive Schwarz Methods for the h-Version Boundary Element Method A Hypersingular Boundary Integral Method for Two-Dimensional Screen and Crack Problems Iterative Methods by Space Decomposition and Subspace Correction --TR --CTR Matthias Maischak, Multiplicative Schwarz algorithms for the p-version Galerkin boundary element method in 3D, Applied Numerical Mathematics, v.56 n.10, p.1370-1382, October 2006 T. Tran , E. P. Stephan, Two-level additive Schwarz preconditioners for the h-p version of the Galerkin boundary element method for 2-d problems, Computing, v.67 n.1, p.57-82, July 2001
p-version Galerkin boundary element method;h-version Galerkin boundary element method;multiplicative Schwarz;multigrid algorithm
588642
Smoothing Methods and Semismooth Methods for Nondifferentiable Operator Equations.
We consider superlinearly convergent analogues of Newton methods for nondifferentiable operator equations in function spaces. The superlinear convergence analysis of semismooth methods for nondifferentiable equations described by a locally Lipschitzian operator in Rn is based on Rademacher's theorem which does not hold in function spaces. We introduce a concept of slant differentiability and use it to study superlinear convergence of smoothing methods and semismooth methods in a unified framework. We show that a function is slantly differentiable at a point if and only if it is Lipschitz continuous at that point. An application to the Dirichlet problems for a simple class of nonsmooth elliptic partial differential equations is discussed.
Introduction . This paper considers the nonlinear operator equation Y is a continuous mapping, X and Y are Banach spaces, and D is an open domain in X . In a number of problems, the operator F is nondifferentiable. For example, a class of such problems arising in optimal control problems for parabolic partial differential equations with bound constraints on the control [16, 17]: where K is a completely continuous map from L 1 to C for some bounded is the map on C given by for given l and u in C : A paradigm for such problems is the Urysohn integral equation of the second kind. Another class of nonsmooth equations related to MHD (magnetohydrodynamics) equilibria [18, 37] will be discussed in Section 4. A paradigm for this class is the Dirichlet problem for nonsmooth elliptic partial differential equations as discussed in Section 4. The nonsmoothness poses serious difficulties and challenges for devising for non-smooth problems analogues of existing iterative methods, which use smoothness. For Department of Mathematics and Computer Science, Shimane University, Matsue 690-8504, Japan (chen@math.shimane.ac.jp). The work of this author was supported in part by the Japan Society of the Promotion of Science Grant C11640119. y Department of Mathematical Sciences, University of Delaware, Newark, DE 19716, U.S.A. (nashed@math.udel.edu). The research of this author was partially supported by a grant from the National Science Foundation. z Department of Applied Mathematics, Hong Kong Polytechnic University, Kowloon, Hong Kong. (maqilq@polyu.edu.hk). School of Mathematics, University of New South Wales, Sydney 2052. (L.Qi@unsw.edu.au). The work of this author was partially supported by the Australian Research Council. Nashed and L.Qi example, the Newton method assumes that F is Fr'echet differentiable and is defined by where F 0 is the Fr'echet derivative of F . What are suitable analogues of Newton's method when F is not smooth ? Iterative methods for nondifferentiable equations have been studied for decades [7, 10, 15, 16, 31, 34, 38, 40]. Among these methods, smoothing methods and semi- smooth methods for nondifferentiable equations arising from variational inequalities and complementarity problems in R n have been studied extensively in the last few years [2, 4, 5, 9, 11, 14, 34]. Superlinear convergence analysis of semismooth Newton methods for equations defining a locally Lipschitian operator in R n uses the notions of generalized Jacobian [12] and semismoothness [32, 35], which are based on the Rademacher theorem. The Rademacher theorem states that if F : R n locally Lipschitzian function, then F is differentiable almost everywhere. For a locally Lipschitzian function F : defined the generalized Jacobian [12] by where DF is the set of points where F is differentiable. For nonsmooth equations described by a locally Lipschitzian operator F from R n into R n , the generalized Newton method is defined by Qi and Sun [32, 35] established superlinear convergence of (1.3) by using a concept of semismoothness. The concept of semismoothness was introduced by Mifflin for real-valued functions [20]. In [35], F is said to be semismooth at x if the limit lim exists for any h 2 R n . Local behavior of the generalized Newton method is analyzed in [13, 31, 32, 35]. The Rademacher theorem does not hold in function spaces. Hence the above definitions of generalized Jacobian and semismoothness cannot be used in function spaces. In this paper, we introduce notions of slanting functions and slant differentiability of operators in Banach spaces, and use these notions to formulate a concept of semi- smoothness in infinite dimensional spaces, which coincides with the above notion of semismoothness in the case of a locally Lipschitzian mapping on R n . These notions will play a pivotal role in the formulation and convergence analysis of analogues of Newton's method (smoothing and semismooth methods) for nondifferentiable operator equations in function spaces. The main feature of smoothing Newton methods in R n is to approximate F by a parametric function f(x; ffl) : R n \Theta R++ ! R n which is continuously differentiable with respect to x, and then to use the partial derivative f x at each step of the Newton-like iteration. The error of f(\Delta; ffl k ) to F is bounded by Smoothing Methods and Semismooth Methods 3 1. For complementarity problems, many smoothing functions have the following properties [9] exists for every x 2 R n and lim khk 0: The properties (1.4) and (1.5) suggest a superlinearly convergent Newton method [6, 11]: Note that f n\Thetan is a single valued function, and the superlinear convergence of (1.6) is not based on the Rademacher theorem. This opens a way to study Newton methods for nonsmooth problems in function spaces. The organization of the paper is as follows. In Section 2 we introduce the notion of a slanting function f o and slant differentiability for a general nonsmooth function F in Banach spaces, and study some of their interesting properties. Using slanting functions, we extend in Section 3 the semismooth Newton method and the smoothing Newton method to Banach spaces. An application to a class of nonsmooth Dirichlet problems is studied in Section 4. We use ff; to denote scalars. The set of all positive real numbers is denoted by R++ . Let L(X; Y ) denote the set of all bounded linear operators on X into Y . 2. Slant differentiability. A function F : D ae X ! Y is said to be (one- sided) directionally differentiable at x if the limit exists, in which case is called the (one-sided) directional derivative of F at x with respect to the direction h. For brevity we will drop "one-sided" in what follows since this is the only notion of directional derivative that occurs in this paper. A function F : D ae X ! Y is said to be B-differentiable at a point x if it is directionally differentiable at x, and lim khk 0: In this case, we call \Delta) the B-derivative of F at x. See [39] for the B- differentiability (B for Bouligand). In finite dimensional Euclidean spaces, Shapiro [41] showed that a locally Lipschitzian function F is B-differentiable at x if and only if it is directionally differentiable at x. Moreover, Qi and Sun [35] showed that F is semismooth at x if and only if F is B-differentiable (hence directionally differentiable) at x and for each V 2 @F 4 X.Chen, Z. Nashed and L.Qi However, these results do not hold in function spaces since the generalized Jacobian is defined only in finite dimensional spaces. To circumvent this difficulty in infinite dimensional spaces we introduce the following notion of slant differentiability. Definition 2.1. A function F : D ae X ! Y is said to be slantly differentiable at x 2 D if there exists a mapping f such that the family of bounded linear operators is uniformly bounded in the operator norm for sufficiently small and lim The function f o is called a slanting function for F at x. Definition 2.2. A function F : D ae X ! Y is said to be slantly differentiable in an open domain D 0 ae D if there exists a mapping f f o is a slanting function for F at every point x 2 D 0 . In this case, f o is called a slanting function for F in D 0 . Definition 2.3. Suppose that f slanting function for F at We call the set xk!x the slant derivative of F associated with f o at x 2 D. Here, lim xk!x f the limit of f sequence fx k g ae D such that x k ! x and lim xk!x f exists, and @S F (x) is the set of all such limits. (Note that f is always nonempty.) Slant differentiability captures a property that appears implicitly in some convergence proofs of Newton-type methods for solving nonsmooth equations as well as ill-posed smooth equations. For example, consider the parameterized Newton method for solving ill-posed smooth equations. To overcome ill-posedness and singularity, we use is chosen such that F 0 I is nonsingular. Let f and assume slanting function for F at x if F 0 (x) is uniformly bounded in a neighborhood of x . We now make a few comments on some unusual properties of slanting functions which also explain the choice of the terms "slanting function" and "slant derivative". Remarks 1. Unlike other notions of derivatives, the term "f does not appear in Definition 2.1, so for a slanting function f for F at x, f itself is not characterized in general by a limit of a quotient or a sequence. 2. A function F may be slantly differentiable at every point of D, but there is no common slanting function of F at all points of D. For example, if F is Fr'echet differentiable at x, we take f slanting function for F at x. But f o in general is not a slanting function of F at other points of D. If F is continuously differentiable in D and we take slanting function for F at every point of D. Smoothing Methods and Semismooth Methods 5 3. A slanting function f for F at x is a single valued function. A slantly differentiable function F at x can have infinitely many slanting functions at x. Even if F is continuously differentiable in D, F still can have infinitely many slanting functions for all points of D. For example, we may let f o take the same values of F 0 except at a finite number of points of D, and take arbitrary values at these finite number of points. Then such f o is still a slanting function of F for all points of D. One may conjecture that if F is continuously differentiable in D and f o is a slanting function for F in D, then f with F 0 except possibly on a set of measure zero. 4. If f are both slanting functions for F at x (in D), then is also a slanting function for F at x (in D), where - 2 [0; 1]: Moreover, lim kf On the other hand, if f are slanting functions for F and G at x (in D), respectively, then h slanting function for ffF at x (in D) where ff and fi are constants. Note that such a result for linear combination does not hold for the generalized Jacobian [12]. 5. f o is not continuous in general. For example, let be a real number. Then the function is a slanting function for F in X . The slant derivative of F for x 2 X is In fact it is easy to see that if f o is continuous at x, then F is differentiable at x and F 0 (x). The slant derivative of F associated with f o at x reduces to a singleton @S F 6. For a locally Lipschitzian function F : R n is semismooth at x, then any single-valued selection of the Clarke Jacobian or the B-subdifferential is a slanting function of F at x. This may not be true if F is not semismooth at x. For example, let ae The derivative F 0 (x) is discontinuous at 0. The function F is slantly differentiable at Indeed let f o be any function for which lim h!0 f Then lim 0: Hence F is slantly differentiable at 0 with infinitely many slanting functions for F at 0. Note that such f o is not a slanting function for F at every point f0g. If we let f slanting function for F at every point but not a slanting function for F at 0. 6 X.Chen, Z. Nashed and L.Qi 7. For a slantly differentiable function F at x, the set @S F (x) is dependent on the choice of a slanting function for F at x. Associated with any slanting function, the set @S F (x) is bounded, since f uniformly bounded for h sufficiently small. For example, let ae Let ae sin 1 slanting function for F at 0 and @S F that F is neither directionally differentiable at 0 nor Lipschitzian in any neighbourhood of 0. Note that the function f o in this example is not slantly differentiable at 0. 8. A continuous function is not necessarily slantly differentiable. For example, let ae p jhj, and 1= there is no uniformly bounded function f o such that F Definition 2.4. A function F is said to be Lipschitz continuous at x if there is a positive constant L such that for all sufficiently small h, We now present a necessary and sufficient condition for the slant differentiability, for the proof we need the following lemma, which is a corollary of the Hahn-Banach theorem. Lemma 2.5. Let X be a normed space and h be a fixed element of X, h 6= 0. Then there exists an element g 2 X , where X is the (norm) dual of X, such that (Note by definition of X , g is a continuous linear functional on X, so it is bounded.) Theorem 2.6. An operator slantly differentiable at x if and only if F is Lipschitz continuous at x. Proof. Suppose that F is slantly differentiable at x. By the definition of slant differentiability, there are C ? 0 and ffi ? 0 such that for all khk - ffi, kf and khk Hence, for all khk - ffi, Smoothing Methods and Semismooth Methods 7 Conversely, suppose that F is Lipschitz continuous at x. We shall show that F is slantly differentiable at x by constructing a slanting function for F at x. For each fixed h 6= 0, by Lemma 2.5, there exists a continuous linear functional g h 2 X such that g h fixed as above, define the following function on an open domain containing x khk for h 6= 0, and define f to be any bounded linear operator on X into Y . Then f maps D into L(X; Y ) since each g h is in X . For any z 2 X , khk Thus kf khk khk kzk: Therefore sup z 6=0 kf kzk khk that is kf Thus for sufficiently small h, kf Now using (2.2), and g h khk If X is a Hilbert space, then by the Riesz representation theorem every continuous linear functional on X can be represented by an inner product. Thus the formula (2.2) can be written in the form: khk Corollary 2.7. (Mean Value Theorem for Slantly Differentiable Functions) Suppose that F : D ae X ! Y is slantly differentiable at x. Then for any h 6= 0 such that x + h is in D, there exists a slanting function for F at x such that 8 X.Chen, Z. Nashed and L.Qi Proof. This follows from the first part of Theorem 2.6 and the proof of the second part of the same theorem. Note that the above form of the mean value theorem is in equality form. It is a selection theorem from the set of slanting functions for F at x. Mean value theorems for smooth operators whose range is an infinite dimensional space are usually given in the form of inequalities involving norms or majorants, or an inclusion form involving the closed convex hull of the set of values of the derivative. For a comprehensive overview of various types of mean value theorems for smooth operators, see pp. 171- 186 of [23]. Proposition 2.8. Suppose that F is slantly differentiable at x, and let f o be a slanting function for F at x. (a) F is directionally differentiable at x if and only if lim exists. If F is directionally differentiable at x, then (b) F has a B-derivative at x if and only if lim exists uniformly with respect to h on each bounded set (say on Proof. (a) Let h 2 X with lim is equivalent to lim Hence if F is directionally differentiable, then The converse is also true. (b) This follows from part (a) and the known (and easy to prove) fact that F has a B-derivative at x if and only if lim exists uniformly with respect to h on each bounded set (see, for example, Nashed [23], where a hierarchy of notions of differentiability are characterized by convergence of "remainder" quotients R(th)=t as t approaches zero, uniformly with respect to h in various classes of subsets). Theorem 2.9. Suppose F is slantly differentiable at x and let f o be a slanting function for F at x. Then the following statements are equivalent. Smoothing Methods and Semismooth Methods 9 (a) For some function which is o(khk), f positively homogeneous of degree 1 in h. (b) lim t!0 exists for every h 2 X and lim khk 0: (c) F is B-differentiable at x, and Proof. (a)) (b): If f positively homogeneous of degree 1 in h, then for any fixed t ? 0, so Note that only if for each fixed h, uniformly in h on each bounded set. Hence for any h 2 X lim uniformly with respect to h on each bounded set. Moreover, lim khk khk 0: (b) ) (c): By part (b) of Proposition 2.1, statement (b) implies that F is B- differentiable and lim t!0 statement (c) holds. (c) positively homogeneous of degree 1 in h, we have (a). Proposition 2.10. Suppose that F is slantly differentiable in a neighborhood of x and let f o be a slanting function for F in the neighborhood of x. Then the following two statements are equivalent. (a) There are a neighborhood N x of x and a positive constant C such that for any (b) There are a neighborhood - N x and a positive constant - C such that for any u 2 - every V 2 @S F (u) is nonsingular and kV C: Proof. Part (a) =) part (b): It is straightforward from the definition of @S F (u). Part (b) =) part (a): It is due to the fact that f Proposition 2.11. Suppose that F is slantly differentiable at x and let f o be a slanting function for F at x. If there are a neighborhood N x of x and a positive constant C such that for any u 2 N x , f o (u) is nonsingular and kf then there is a positive constant - C such that every V 2 @S F (x) is nonsingular and C: Moreover, if Y is a finite dimensional space, the converse holds. Proof. The first part follows the definition of @S F (x). The second part is due to the fact that every bounded sequence has convergent subsequence in finite dimensional spaces. Indeed, if the second part is not true, then there is a sequence fu k g such that are singular or kf 1. By the definition of @S F (x), there is a subsequence fu k j g ae fu k g such that f is singular. This contradicts the assumption. Nashed and L.Qi 3. Smoothing functions and semismooth functions. We generalize the definition of smoothing functions for a nonsmooth function and the concept of semism- mothness of a nonsmooth function in finite dimensional spaces to infinite dimensional spaces. Definition 3.1. We say that Y is a smoothing function of F if f is continuously differentiable with respect to x and for any x 2 D and any ffl ? 0, where - is a positive constant. The smoothing function f is said to satisfy the slant derivative consistency property at - x (in D) if lim exists for x in a neighborhood of - x (in D) and f serves as a slanting function for F at - x (in D). Note that the limit in (3.2) is in the topology of the operator norm, so the pointwise convergence of f x (x; ffl)h to f o (x)h for each fixed h is uniform on the set Definition 3.2. We say that F is semismooth at x if there is a slanting function f in a neighborhood N x of x, such that f o and the associated slant derivative satisfy the following two conditions. (a) lim t!0 exists for every h 2 X and lim (b) Theorem 3.3. Suppose that F is slantly differentiable in a neighborhood N x of x, and let f o be a slanting function for F in N x : Then F is semismooth at x if and only if F is B-differentiable at x and where @S F is the slant derivative associated with f o in N x . Proof. Suppose that F is semismooth at x. Then from Theorem 2.9, part (a) of Definition 3.2 implies that F is B-differentiable and Thus part (b) of Definition 3.2 implies (3.3). Now, we suppose that F is B-differentiable at x and (3.3) holds. Then for all Smoothing Methods and Semismooth Methods 11 Hence part (b) of Definition 3.2 holds. Moreover, we have By Theorem 2.9, part (a) of Definition 3.2 holds and so F is semismooth at x. Theorem 3.3 implies that the definition of semismoothness used here coincides with the definition of Qi-Sun [35] in finite dimensional spaces if we take a single-valued selection of the Clarke Jacobian or the B-subdifferential as the slanting function. To illustrate Theorem 3.3, we consider the system of "min" nonsmooth equations in R n where p and q are continuously differentiable functions from R n into itself. This system is equivalent to the complementarity problem Chen, Qi and Sun showed [9] that every smoothing function f(x; ffl) in the Chen- Mangasarian smoothing function family [5] for the nonsmooth function F satisfies (3.2). In particular, for where ff 2 [0; 1] is dependent on the choice of a smoothing function. Such f belongs to the set at every point x 2 R n . (See [33] for @C F (x):) Hence, every smoothing function in the Chen-Mangasarian smoothing function family satisfies the slant derivative consistency property in R n . Moreover, the associated slant derivative is, for which is bounded, nonempty and satisfies Furthermore, the following fact is known [35] On the other hand, by Proposition 2.8, we know Hence, the nonsmooth function F is semismooth in the sense of Definition 3.2. Now we consider superlinearly convergent Newton-type methods for nonsmooth equations with slanting differentiable operators. Nashed and L.Qi Theorem 3.4. Suppose that F is slantly differentiable at a solution x of (1.1). Let f o be a slanting function for F at x and kf in a neighborhood N of x , where M is a positive constant. Then the iterative sequence fx k g generated by the Newton-type method superlinearly converges to x in a neighborhood N 0 of x . Here A(x) 2 L(X; Y ) and Proof. By Definition 2.1 and the Banach Lemma [29], there is a neighborhood N 0 of x , N 0 ae N , and positive constants M for any x 2 N 0 , A(x) is nonsingular and kA(x) and Therefore, starting from any x 0 2 N 0 the Newton method (3.4) is well defined and the successive iterates satisfy the following inequalities: Hence the sequence fx k g converges to x : Moreover, using Definition 2.1 and (3.5), the inequalities above imply Using Theorem 3.4 and Proposition 2.10, we can immediately obtain the following theorem. Theorem 3.5. Suppose that F is slantly differentiable at a solution x of (1.1). Let f o be a slanting function for F at x and kf in a neighborhood N of x , where M is a positive constant. Then the following statements hold. (a) The Newton-type method (1.6) superlinearly converges to x in a neighborhood N 0 of x . (b) If f : D \Theta R++ ! Y is a smoothing function of F which satisfies the slant derivative consistency property (3.2) in N , then the smoothing Newton method superlinearly converges to x in a neighborhood N 0 of x . (c) If F is semismooth at x , then the semismooth Newton method superlinearly converges to x in a neighborhood N 0 of x . Smoothing Methods and Semismooth Methods 13 4. An application to a class of nonsmooth elliptic partial differential equations. Let\Omega ae R 2 be a bounded region with piecewise smooth boundary @\Omega and let W be the class of functions from\Omega to R satisfying Z R) be the space of functions in W endowed with the norm Z \Omega ju(x)jdx: We consider the following nonsmooth Dirichlet problem ae in\Omega R is a continuous function. Let Z \Omega G(x; y)P (u(y))dy; where G is the Green function for the boundary value problem (e.g. see [42]) ae in\Omega is the Dirac "generalized function" at y in \Omega\Gamma The nonsmooth integral equation is equivalent to the nonsmooth Dirichlet problem (4.1). Theorem 4.1. Suppose that smoothing function of P satisfying where - is a positive constant. Then Z\Omega is a smoothing function of F and where x2\Omega Z\Omega 14 X.Chen, Z. Nashed and L.Qi Proof. It is easy to see that f is continuously differentiable with respect to u and Z Moreover Z Z Z \Omega kG(x; y)kdy -ffl: Theorem 4.2. Suppose that P is slantly differentiable at u, and let p o be a slanting function for P at u. Then F is slantly differentiable at u, and Z \Omega G(x; y)p is a slanting function for F at u Proof. Using the definition of f o as given in the statement of the theorem, we have 'Z Z\Omega Z\Omega 'Z\Omega slanting function for P at u, the above equality implies that f o is a slanting function for F at u. By properties of integrals, we have the following proposition. Proposition 4.3. (1) If smoothing function of P which satisfies the slant derivative consistency property (3.2) at u, then is a smoothing function of F which satisfies the slant derivative consistency property (3.2) at u (2) If P is semismooth at u, then F is semismooth at u. The above results demonstrate that superlinearly convergent smooth methods or semismooth methods can be developed for the nonsmooth Dirichlet problem (4.1). For example, we consider that is ae where - and ff are constants. Let Smoothing Methods and Semismooth Methods 15 Then p is a smoothing function of P in X , and satisfies the slant derivative consistency property in X since lim ffl!0 ffl!0 otherwise: We next show that p o is a slant function for P in X . In fact, for any u; h(6= we have Z -f Z Z u(x)+h(x)=ff Note that Letting khk ! 0, we have khk f Z Z u(x)+h(x)=ff dx 0: Moreover, we can show that P is semismooth in X . Hence, by using Z\Omega we can obtain superlinearly convergent smoothing methods and semismooth methods for the nonsmooth Dirichlet problem ae in\Omega This nonsmooth Dirichlet problem is related to MHD (magnetohydrodynamics) equilibria [18]. We report numerical results for the following example. Example 4.1. ae where OE(0; problem has an exact solution ae Z. Nashed and L.Qi Table Numerical result of Example 4.1: kFn We use method (1.6) with the five-point finite difference method. The stopping criterion is kFn Here Fn is the finite difference approximation function with grids n. We report the value of kFn (x)k 1 at the last five iterations. Nonsmooth optimization and operator equations involving nonsmooth operators are becoming crucial in various areas of computational and applied mathematics, for example in nonsmooth mechanics [21, 22, 30], optimal design of electromagnetic devices [21], ill-posed problems involving nonsmooth operators and variational inequalities [19, 24], bounded variation regularization and nondifferentiable optimization problems in nonreflexive spaces [26, 27], inverse source problems [25], free boundary problems [28], multi-body system identification [1], and nonlinear complementary problems (see [8] and references cited therein). Various classes of these problems can be reformulated as nonsmooth equations with locally Lipschitzian operators. Hence the smoothing methods and semismooth methods studied in this paper can be applied to these problems. Acknowledgement . This research was initiated while the first author worked in the School of Applied Mathematics at the University of New South Wales where the second author visited in 1993 and 1998. We would like to express our appreciation for the support. --R Outer inverses and multi-body system identification The global linear convergence of a non-interior path-following algorithm for linear complementarity problem Smooth approximations to nonlinear complementarity problems A class of smoothing functions for nonlinear and mixed complementarity problems Superlinear convergence of smoothing quasi-Newton methods for nonsmooth equa- tions Global and superlinear convergence of inexact Uzawa methods for saddle point problems with nondifferentiable mappings Convergence of Newton's method for singular smooth and nonsmooth equations using adaptive outer inverses Global and superlinear convergence of the smoothing Newton method and its application to general box constrained variational inequalities Convergence domains of certain iterative methods for solving nonlinear equations On homotopy-smoothing methods for variational inequalities Optimization and Nonsmooth Analysis Solution of monotone complementarity problems with locally Lipschitzian functions Identification of the support of nonsmoothness Multilevel algorithms for constrained compact fixed point prob- lems Finite element analysis of a nondifferentiable nonlinear problem related to MHD equilibria Regularization of nonlinear ill-posed variational inequalities and convergence rates Semismooth and semiconvex functions in constrained optimization Topics in Nonsmooth Mechanics Mathematical Theory of Hemivariational Inequalities and Applications Differentiability and related properties of nonlinear operators: some aspects of the role of differentials in nonlinear function analysis On nonlinear ill-posed problems II: Monotone operator equations and monotone variational inequalities Stable approximation of nondifferentiable optimization problems with variational inequalities Stable approximation of a minimal surface problem with variational inequalities Least squares and bounded variation regularization with non-differentiable functionals Iterative Solution of Nonlinear Equations in Several Variables Inequality Problems in Mechanics and Applications motivation and algorithms Convergence analysis of some algorithms for solving nonsmooth equations A globally convergent successive approximation method for severely non-smooth equations A nonsmooth version of Newton's method Computational Solution of Nonlinear Operator Equations Approximation of a nondifferentiable nonlinear problem related to MHD equilibria A unified convergence theory for a class of iterative process Local structure of feasible sets in nonlinear programming. Newton's method for a class of nonsmooth functions On concepts of directional differentiability --TR --CTR B. Mermri , X. Chen, On characterizations and regularity of the solution of bilateral obstacle problems, Journal of Computational and Applied Mathematics, v.152 n.1-2, p.333-345, 1 March Xiaojun Chen, Applications of smoothing methods in numerical analysis and optimization, Focus on computational neurobiology, Nova Science Publishers, Inc., Commack, NY, 2004
nonsmooth elliptic partial differential equations;superlinear convergence;semismooth methods;nondifferentiable operator equation;smoothing methods
588815
Perfect Packing Theorems and the Average-Case Behavior of Optimal and Online Bin Packing.
We consider the one-dimensional bin packing problem under the discrete uniform distributions $U\{j,k\}$, $1 \leq j \leq k-1$, in which the bin capacity is $k$ and item sizes are chosen uniformly from the set $\{1,2,\ldots,j\}$. Note that for $0 < this is a discrete version of the previously studied continuous uniform distribution $U(0,u]$, where the bin capacity is 1 and item sizes are chosen uniformly from the interval $(0,u]$. We show that the average-case performance of heuristics can differ substantially between the two types of distributions. In particular, there is an online algorithm that has constant expected wasted space under $U\{j,k\}$ for any $j,k$ with $1 \leq j < k-1$, whereas no online algorithm can have $o(n^{1/2})$ expected waste under $U(0,u]$ for any $0 < u \leq 1$. Our $U\{j,k\}$ result is an application of a general theorem of Courcoubetis and Weber that covers all discrete distributions. Under each such distribution, the optimal expected waste for a random list of $n$ items must be either $\Theta (n)$, $\Theta (n^{1/2} )$, or $O(1)$, depending on whether certain ``perfect'' packings exist. The perfect packing theorem needed for the $U\{j,k\}$ distributions is an intriguing result of independent combinatorial interest, and its proof is a cornerstone of the paper. We also survey other recent results comparing the behavior of heuristics under discrete and continuous uniform distributions.
Introduction . Suppose one is given items of sizes 1, 2, 3, . , j, one of each size, and is asked to pack them into bins of capacity k with as little wasted space as possible, i.e., one is asked to find a least cardinality partition (packing) of the set of items such that the sizes of the items in each block (bin) sum to at most k. For what Published electronically February 1, 2002. This paper originally appeared in SIAM Journal on Discrete Mathematics, Volume 13, Number 3, 2000, pages 384-402. http://www.siam.org/journals/sirev/44-1/39542.html # Department of Computer Science, Athens University of Economics and Business, Athens, Greece (courcou@csi.forth.gr). - Bell Labs, Murray Hill, NJ 07974. # Statistical Laboratory, University of Cambridge, Cambridge CB3 0WB, UK (R.R.Weber@ statslab.cam.ac.uk). # Avaya Labs Research, Basking Ridge, NJ 07920 (mihalis@research.avayalabs.com). values of j and k can the set be packed perfectly (i.e., so that the sizes of the items in each block sum to exactly k)? Clearly the sum of the item sizes must be divisible by k, but what other conditions must be satisfied? Surprisingly, the divisibility constraint is not only necessary but su#cient. Readers might want to try their hand at proving this. Relatively short proofs exist, as illustrated in the next section, but a certain ingenuity is required to find one. The exercise serves as a warm-up for the following more general and more di#cult theorem, also proved in the next section, in which there are r copies of each size for some r # 1. Theorem 1 (perfect packing theorem). For positive integers k, j, and r, with perfectly pack a list L consisting of rj items, r each of sizes 1 through into bins of size k if and only if the sum of the rj item sizes is a multiple of k. In set-theoretic terms, the question answered by Theorem 1 is an intriguing puzzle in pure combinatorics. But our motivation to work on it came from its relevance to certain fundamental questions about the average-case analysis of algorithms. In particular, consider the standard bin packing problem, in which one is given a bin capacity b and a list of items each a i has a positive size s i # b, and is asked to find a packing of these items into a minimum number of bins. In most real-world applications of bin packing, as in Theorem 1, the item sizes are drawn from some finite set. However, the usual average-case analysis of bin packing heuristics has assumed that item sizes are chosen according to continuous probability distributions, which by their nature allow an uncountable number of possible item sizes (see [3, 10], for example). The assumption of a continuous distribution has the advantage of sometimes simplifying the analysis and has been justified on the grounds that continuous distributions should serve as reasonable approximations for discrete ones. But there are reasons to ask whether this is actually true. For example, consider the continuous uniform distributions U(0, u], 0 < u # 1, where the bin capacity is 1 and item sizes are chosen uniformly from the interval (0, u], and the discrete uniform distributions U{j, k}, 1 # j # k - 1, where the bin capacity is k and item sizes are chosen uniformly from the set {1, 2, . , j}. The limit of the distributions U{mj,mk}, as m #, is equivalent to U(0, j/k] (after scaling by dividing the item sizes and bin capacities by mk). However, in the limit, combinatorial questions such as those addressed by Theorem 1 evaporate. This suggests that something important (and may in fact be lost in the transition from discrete to continuous models. The results in this paper show that this is indeed the case. To describe the results, we need the following notation. If A is a bin packing algorithm and L is a list of items, then A(L) is the number of bins used when A is applied to L, and s(L) is the sum of the item sizes in L divided by the bin capacity. Note that s(L) is a lower bound on the number of bins needed to pack L. The waste in the packing of L by A is denoted by W A an algorithm that always produces an optimal packing. In what follows, L n denote a list of n items whose sizes are independent samples from a given distribution F . The expected waste rate EW A for an algorithm A and distribution F is defined to be the expected value of W A (L n as a function of n. In what follows we typically abbreviate this as simply the "expected waste." We say a distribution F is a bounded waste distribution if EW OPT As a consequence of Theorem 1 and a classification theorem of Courcoubetis and Weber [11], we can prove the following. Theorem 2. For any j, k with 1 < j < k - 1, EW OPT This in itself does not represent a departure from the continuous model, since U(0, u] is also a bounded waste distribution for all u, 0 < u < 1 [3, 16]. The distinction comes when we consider "online" algorithms. In an online algorithm, items are assigned to bins in the order in which they occur in the input list L. Each assignment must be made without knowledge of the sizes or number of items later in the list, and once an item is packed it cannot subsequently be moved. This mirrors many practical situations but clearly is a substantial restriction on the power of an algorithm. In particular, we can prove the following. Theorem 3. If A is an online algorithm and u # (0, 1], then it cannot be the case that E[W A In contrast, in the discrete uniform case there is a single online algorithm that has bounded expected waste for all the distributions U{j, k}, 1 # j < k - 1. This is the recently discovered sum-of-squares algorithm (SS) of [12], defined as follows. Suppose we are packing integer-sized items into bins of capacity b. When an online algorithm packs an item x from such a list, the only thing relevant about the current packing is the number N h of bins whose current contents total h, 1 # h # b - 1. SS chooses the bin into which x is to be placed (either a new, previously empty bin or one that is already partially filled but has enough room for x) so as to minimize the resulting sum h . Note that SS can be implemented to take O(b) time per item [12], and so runs in linear time for any fixed bin size. As shown in [12], algorithm SS performs well on average in a surprisingly general sense. Let us say that a discrete distribution F is any triple (b, S, # p), where b is an integral bin size, is a finite set of integral item sizes in the range from 1 to b - 1, and # rational probability vector, where p i > 0 is the probability of item size s i and ignore the possibility of items of size b since such items always must start a new bin and completely fill it, leaving the rest of the packing una#ected.) The following specialized version of the result of [12] su#ces for our needs. Theorem (see [12]). For any discrete distribution Hence by Theorem 2, EW SS The current paper is organized as follows. The proof of the perfect packing theorem (Theorem 1) appears in section 2. Section 3 then presents the proof of Theorem 2. We begin by describing the classification result of Courcoubetis and Weber [11] upon which the proof depends. This result says that for any discrete distribution F , EW OPT must be one of #(n), #(n 1/2 ), or O(1). Which case applies depends on the existence of certain perfect packings and is in general NP- hard. However, Theorem 1 allows us avoid this complexity in the case of the discrete uniform distributions U{j, k}. Theorem 3, this paper's contribution to the theory of continuous distributions, is proved in section 4. We conclude in section 5 with a survey of the results that have been proved about the average-case behavior of bin packing algorithms under discrete and continuous distributions. As we shall see, there are other significant di#erences between the discrete and continuous cases. 2. The Perfect Packing Theorem. We begin our proof of Theorem 1 with three lemmas that list a number of special instances that lead to perfect packing. The first lemma takes care of the special case, r = 1. Lemma 4. Suppose m, j, and k are positive integers such that j # k and 1)/2. Then the set of j items, one each of sizes 1, . , j, perfectly packs into m bins of size k. Proof. The proof is by induction. Pick j and k and assume the theorem is true for all pairs that are smaller in lexicographic order than (k, j). The theorem is clearly true for k # 2 or j # 2, so assume k, j > 2. then we can start by perfectly packing bins with pairs of items (j after which the remaining items are those of sizes plus the item of size k/2 if k is even. Since the sum of the sizes of the items that have been packed at this point is a multiple of k, the sum of the sizes of the remaining items is also a multiple of k. If k is odd, the unpacked items are an instance of (k, and the induction hypothesis applies. If k is even, then k/2 divides j(j + 1)/2 and all remaining items are no larger than k/2. Thus the items 1, . , k - j - 1 form an instance of (k/2, k - j - 1) and by the induction hypothesis can be perfectly packed into bins of size k/2. These half-bins and the item of size k/2 can then be combined into bins of size k. Now suppose j # k/2. If k is even, then we have an instance of (k/2, j) and the induction hypothesis applies. If k is odd, first note that k/2 # j and j > 2 imply which together with 2m. Thus we can construct m pairs of items each of total size k 1. If we place one pair in each of our m bins, we now have bins with gaps of size k - k items of sizes 1, . , j - 2m. Because 1)/2, the sum of these item sizes must be m(k - k # ), and so an application of the induction hypothesis to the instance completes the proof. Lemma 5. Consider r > 1 sets, the ith of which consists of j items of consecutive (a) r is even or (b) j is odd. Then these rj items perfectly pack into j bins of size equal to the sum of the average item sizes in the r groups, i.e., r(j Proof. The lemma will follow if we can show that for # bins of size r(j + 1)/2 it is possible to pack perfectly the items into the j bins in such a manner that each bin contains exactly one item from each of the r sets. If r is even, then we simply take two of the sets and pack the ith largest item in one set with the ith smallest item in the other set, i.e., as the pair (i, bins to level j + 1. By repeating this r/2 times we fill j bins of size r(j + 1)/2. If r and j are both odd, then an extra step is required. The idea is first to pack items in triples, one item from each of three sets, such that the sum of each triple is the same. It is easiest to appreciate the construction by considering an example, say 9. The triples, which each sum to 15, are given in the columns below. In general, the triples are (i, i j. The result of packing these one per bin is to fill all j bins to level 3(j 1)/2. The number of remaining sets is even and the remaining spaces in the j bins are equal. Thus, the procedure for case (a) can be applied to complete the packing in each bin. The following lemma provides part of the induction step used in the proof of Theorem 1. Lemma 6. Consider a quadruple (k, j, r, m) of positive integers such that k # j and 1)/2. Then there exists a perfect packing of r copies of 1, . , j into m bins of size k if there exists a perfect packing for each lexicographically smaller quadruple of this form, and if any one of the following holds: (b) r does not divide k. (c) k or r is even. Proof. First, using the arguments of Lemma 4, we demonstrate how to reduce the problem to a smaller instance if (a) holds. If j # k/2 and k is odd, then we can pack bins with pairs (j - 1)/2. The remaining items, which are of sizes 1, . , k - j - 1, define the smaller instance (k, k - and k is even, then we can pack bins in the same way, k/2. The remaining items, which are of sizes 1, . , and k/2, can be packed into bins of size k/2 by the induction hypothesis that there exists a perfect packing for If (b) holds, then r and m must have a common factor p > 1 and the problem reduces to the instance (k, j, r/p, m/p). Now suppose neither (a) nor (b) holds but (c) does. If k is even, then k/2 divides k/2. Thus the problem reduces to a smaller instance in which the bin size is k/2. If r is even, then since (b) does not hold, k is divisible by r and so must be even too. Thus the same argument applies. Finally, for case (d), assume that (a), (b), and (c) do not hold, i.e., j < k/2, r divides k, and k and r are both odd. Let r k. The fact that r is odd implies that r 1 and r 2 are integers. The fact that r divides k implies that k 1 and k 2 are integers, with 1)/2. Since by assumption j < k/2, we have hence by hypothesis for the instance (k 1 , j, r 1 , m), we can pack r 1 copies of 1, . , bins of size k 1 . Similarly, if also then we can pack copies of 1, . , bins of size k 2 . Since we can combine pairs of bins of sizes k 1 and k 2 into bins of size k. Thus there is a reduction to smaller instances holds. Proof of the perfect packing theorem. Instances for which the theorem is to be proved are described by the quadruples of Lemma 6. Notice that it would be enough to specify the triple (k, j, r); however, it is helpful to mention m explicitly. The proof of the theorem is by induction on (k, j, r) under lexicographical ordering. By Lemma 4 it is true for 1. Assume all quadruples that are smaller than (k, j, r, m) can be perfectly packed and r > 1. We show there exists a perfect packing of r copies of 1, . , bins of size k. By Lemma 6, we need only consider the case when and r are odd, r divides k, and (r - 1)k/2r < j < k/2. Note that in this case (r - 1)k/2r is an integer and k/2r is 0.5 more than an integer. We show below that we can perfectly pack all the items of sizes from bins of size k. (Note that the lower bound on this range is greater than 1 because of the above lower bound on j.) The theorem then follows because the remaining items form a smaller quadruple, so by the induction hypothesis they can be perfectly packed into bins of size k. To follow the construction below, the reader may find it helpful to consider a specific example. Consider the quadruple (k, j, r, m) = (165, 77, 5, 91). Note that k and r are odd, r divides k, and j lies between (r - show below how to perfectly pack all items of sizes 12, . , 77. The remaining items form the smaller quadruple (165, 11, 5, 2). To pack all items of sizes from j +1-(r-1)k/2r through j, we divide the range of item sizes into intervals, i.e., sets of consecutive integers. Each interval is symmetric about a multiple of k/2r and has one of two lengths depending on whether the interval 100 COFFMAN ET AL. is symmetric about an odd or even multiple of k/2r. To form the intervals, we first take the largest interval that is symmetric about (r - 1)k/2r; this is the interval [(r -1)k/r-j, j]. Note that this interval does not include (r -2)k/2r since j < rk/2r. Next we take the largest interval that can be formed from the remaining items that is symmetric about (r-2)k/2r, obtaining the interval [j-k/r+1, (r-1)k/r-j-1]. Continuing in this fashion and taking intervals symmetric about further multiples of k/2r, we end up with intervals of two kinds. First, there are (r - 1)/2 intervals centered on even multiples of k/2r, with the interval centered on (r - 1 - 2i)k/2r being ranges from 0 to (r - 3)/2. Second, there is an equal number of intervals centered on odd multiples of k/2r, with the interval centered on (r - 2i)k/2r being [j - ik/r ranges from 1 to (r - 1)/2. Note that the smallest endpoint is claimed above. For the numerical example above, there are two intervals of each type. Intervals of the first type are [22, 44] and [55, 77]; they are of length 23 and symmetric about 33 and 66. Intervals of the second type are [12, 21] and [45, 54]; they are of length 10 and symmetric about 16.5 and 49.5. In general, intervals of the first type have odd length are symmetric about an even multiple of k/2r. Intervals of the second type have even length k -2j -1 and are symmetric about an odd multiple of k/2r. Our plan is to use Lemma 5 to perfectly pack into bins of size k those items whose sizes lie in intervals of the same type. We begin by considering all those intervals of the first type. These have odd lengths and they are symmetric about points ik/r, 1)/2. There are r items of each size in each of these intervals. Our strategy is to partition these intervals into groups that satisfy the hypotheses of Lemma 5(b). That is, we arrange for the midpoints of the intervals within each group to sum to k. Since the midpoints correspond to the average item sizes for the corresponding intervals, and the number of items in the intervals is odd, Lemma 5(b) implies that we can perfectly pack the items in the intervals of each group. Constructing these groups is a bin packing problem in which the midpoints of the intervals take on the role of item sizes. In what follows we write "items" in quotes when speaking of the midpoints of intervals, possibly normalized, and viewing them as items to be perfectly packed in bins of some required size. In considering intervals of the first type, it is as though we had r "items" of each of the sizes ik/r, and wished to pack them into bins of size k. After a normalization that multiplies each item size by r/k, this is equivalent to the problem of packing r "items" of each of the sizes 1, . , (r -1)/2 into bins of size r. That is, we have a smaller version our packing problem, with But by the induction hypothesis this means that the desired packing can be achieved. In the example, it is as though we had 5 "items" of sizes 33 and 66 that are to be packed in bins of size 165. Normalizing by a factor of 1/33, this is equivalent to the problem instance (5, 2, 5, 3). We must now pack items whose sizes lie in intervals of the second type. These intervals are of even length, symmetric about the points ik/2r, for i odd and 1, . , r - 2. Again, there are r items of each size in these intervals. As above, we exhibit a reduction to a smaller perfect packing problem. After we multiply item sizes by 2r/k the problem is equivalent to perfectly packing r copies of "items" of sizes 1, 3, 5, . , r - 2 into bins of size 2r. For the example, this is 5 copies of "items" of sizes 1 and 3, to be perfectly packed into 2 bins of size 10. Unfortunately, if the sum of the "item" sizes is an odd multiple of r the "items" cannot be perfectly packed into bins of size 2r. For this reason, and also because it is convenient to do so even when the sum of the "item" sizes is a multiple of 2r, we consider perfect packings into bins of sizes r and 2r. Assume for the moment that r copies of "items" of sizes can be perfectly packed into bins of sizes r and 2r. If they are packed entirely into bins of size 2r, then the number of "items" in each bin must be even (as all "item" sizes are odd), and so Lemma 5 applies and implies that the original items can be perfectly packed into bins of size k. On the other hand, suppose a bin of size r is required. The set of "items" that are packed into a bin of size r corresponds to a set of intervals whose midpoints sum to k/2. Recall that the intervals are of even length. We divide each such interval into its first half and its second half, obtaining twice as many intervals, whose midpoints now sum to k. Now we can again use Lemma 5 to construct the perfect packing. The final step in the proof is to show that we can indeed perfectly pack r copies of each of the item sizes 1, 3, 5, . , r - 2 into bins of sizes r and 2r. We shall use a di#erent packing depending upon whether For the case the "items" are perfectly packed by the following simple procedure. We begin by packing one bin with (1, 1, r - 2), one bin with (2i 1, r-2i-2, r-2i) for each one bin with (2i-1, 2i+1, r-2i, r-2i) for each (noting that in the final case, when #, we get three "items" of size 2i This packs four "items" of each size larger than 1, and three "items" of size 1. We can apply this packing # times, leaving us with one "item" of each size larger than 1 and # "items" of size 1. Then we pack one bin with This uses up all the remaining "items" (where #+1 items of size 1 are used because there are two "items" of size 1 when 1). For the numerical example, in which construction says that we should pack 5 copies of 1 and 3 into bins of size 5 and 10 by first packing one bin with (1, 1, then one bin with (1, 3, 3, 3). This leaves one "item" of size 3 and two of size 1. These perfectly pack into a bin of size 5. When the procedure is very similar to that above. We begin by packing one bin with (1, 1, r - 2), one bin with (2i +1, 2i+1, r -2i-2, r -2i) for each As before, this packs four "items" of each size larger than 1, and three "items" of size 1. We apply this packing # times, leaving us with three "items" of each size larger than 1 and #+ 3 "items" of size 1. Then we pack one bin with (2i - 1, 2i for each This leaves us with # "items" of size 1, two "items" of size one "item" of each other size. Finally, as before, we pack one bin with uses up all remaining items. 3. Proof of Theorem 2. Recall the theorem statement: For any distribution U{j, k}, with j < k - 1, EW OPT We rely on a general result of Courcoubetis and Weber [11]. Suppose is a discrete distribution as defined in section 1. Note that a packing of items with sizes from into a bin of size b can be viewed as a nonnegative integer vector interest are those vectors that give rise to a sum of exactly b, which we shall call perfect packing configurations. For instance, if one such configuration would be (1, 0, 2). Let P S,b denote the set of all perfect packing configurations for a given S and b. Let # S,b be the convex cone in R d spanned by all nonnegative linear combinations of configurations in P S,b . Theorem (Courcoubetis and Weber [11]). For any discrete distribution (b, S, # p), the following hold. (a) If # p lies in the interior of # S,b , then EW OPT (b) If # p lies on the boundary of # S,b , then EW OPT (c) If # p lies outside of # S,b , then EW OPT In general it is NP-hard to determine which of the three cases applies to a given distribution (as can be proved by a straightforward transformation from the PARTITION problem [13]). However, for the distributions U{j, k}, j < k - 1, we can use the following lemma, which we shall prove using the perfect packing theorem, to show that (a) applies. Lemma 7. For each i, j, k with exist positive integers such that the set of r items consisting of r items of size i together with r i items of each of the other j - 1 sizes can be packed perfectly into m i bins of size k. Note that this lemma implies that the j-dimensional vector - is strictly inside the appropriate cone when 1. This is because - e is in the interior of the cone spanned by vectors of the form (r i , . , r those vectors are sums of perfect packing configurations by Lemma 7. The proof of Theorem 2 thus follows from case (a) of the above theorem. Proof of Lemma 7. We make use of the perfect packing theorem. There are two cases. If k we simply set r that the total size of r i items each of the sizes 1, . , so by the perfect packing theorem, we can perfectly pack them into j(j bins of size i. The remaining s items of size i can then go one per bin to fill these bins up to size precisely k. On the other hand, suppose Now things are a bit more complicated. We have r 1)/2. By the perfect packing theorem r i items each of the sizes 1, . , perfectly pack in s i bins of size k - i. (Such items exist because by assumption j < k - 1.) We then add the additional s i items of size i to these bins, one per bin, to bring each bin up to size k. There remain r i items each of sizes k - j through j, for a total of items. These can be used to completely fill the remaining bins with pairs of items of sizes (j, k - j), (j - 1, k Note that if k is even, the last bin type contains two items of size k/2, but we have an even number of such items by our choice of r presents no di#culty. It is easy to verify that in both cases r i , s i , and m i are all less than 2k 2 . 4. Proof of Theorem 3. Recall the theorem statement: If L n has item sizes generated according to U(0, u] for 0 < u # 1, and A is any online algorithm, then there exists a constant c > 0 such that E[W A (L n )] > cn 1/2 for infinitely many n. Proof. Let w(t) denote the amount of empty space in partially filled bins after t items have been packed. We show that for any n > 0 the expected value of the average of w(1), . , w(n) is # n 1/2 u 3 ). This implies that E[w(n)] must be# (n 1/2 u 3 ), i.e., not o(n 1/2 ). Consider packing item a t+1 . Let #(t) denote the number of nonempty bins that have a gap of at least u 2 /8 after the first t items have been packed. There are at most bins into which one can put an item larger than u 2 /8. Therefore, if a t+1 is to leave a gap of less than # in its bin, either it must have size less than u 2 /8 or its size must be within # of the empty space in one of these #(t) bins with gaps larger than /8. The probability of this is at most [u 2 /8+#(t)]/u. By choosing conditioning on whether #(t) is greater or less than n 1/2 , and noting that the size of a t+1 is distributed as U(0, u] independent of #(t), we have Now (a s+1 is last in a bin and leaves gap #) [P (a s+1 is last in a bin) - P (a s+1 is last in a bin and leaves gap < #)] [P (a s+1 is last in a bin) - P (a s+1 leaves gap < #)] (a s+1 is last in a bin) - u/4. t be the sum of the first t item sizes, and note that S t is a lower bound on the number of bins and hence on the number of items that are the last item in a bin. We thus have (a s+1 is last in a bin) # E[S t Using the fact that we then have If we have for all t # n/2, This implies On the other hand, if w(t) #n These imply that E[w(n)] is# (n 1/2 ). It should be noted that the above proof relies heavily on the fact that the distribution is continuous, since this is the reason why the union of n 1/2 intervals of size cannot cover the full probability space. Our discrete distributions U{j, k} do not have this failing, and for this reason we can obtain significantly better average-case behavior for them. 104 COFFMAN ET AL. Table Expected waste in the symmetric case. BF #(n 1/2 log 3/4 n) [26] #(n 1/2 log 3/4 Best online #(n 1/2 log 1/2 n) [26, 27] #(n 1/2 log 1/2 The upper bound is proved in the reference; the lower bound is conjectured based on experiments. # The upper and lower bounds here appear to follow from the corresponding results for the continuous case, but the details of the upper bound in particular still need to be worked out. 5. Concluding Remarks. The results in this paper were among the first to be obtained about the average-case behavior of bin packing algorithms under discrete distributions. Since they were announced in [4], many additional results have been proved, illustrating further contrasts with (and similarities to) the case of continuous distributions. In this concluding section we survey the literature and point out some of the remaining open problems. Let us begin by considering symmetric uniform distributions, as represented by U(0, 1] and U{k - 1, k}, k # 1. (In general, a symmetric distribution is one that satisfies p(s #) = p(s # b - #) for all #, 0 # b.) Table 1 summarizes what is known about average-case behavior under these distributions. A horizontal line separates the o#ine algorithms from the online ones. Except where noted, all results in this table are theorems. Four famous classical algorithms have been extensively studied. First fit (FF ) is an online algorithm in which each item is placed in the first bin that has room for it, where bins are sequenced according to the order in which they received their first item. If no bin has room, a new bin is started. Best fit (BF ) is similar, except now the item is placed in the bin with the smallest gap large enough to contain it (ties broken in favor of the earlier bin). First fit decreasing (FFD) and best fit decreasing (BFD) are the corresponding o#ine algorithms in which the list is first sorted so that the items are in nonincreasing order by size, and then FF (BF ) is applied. From a worst-case point of view, FF and BF are equivalent: in an asymptotic sense each can produce packings that use 70% more bins than optimal, but neither can do any worse [15]. The corresponding o#ine versions FFD and BFD each can use more bins than optimal but can do no worse [14, 15]. The results in Table 1 show that these algorithms perform much better on average than in the worst case, since they now have sublinear expected waste, a surprise when it was first observed empirically in [2]. The o#ine versions continue to have an advantage over their online counterparts, but it is of reduced practical significance. And now there is a distinction in the behavior of FF and BF , with BF being the better of the two. The above remarks apply equally well to the discrete and continuous cases. As to the comparison between these cases, we once again have a significant di#erence for online algorithms. For any fixed value of k, the online algorithms in the table all have expected waste, in contrast to the expected wastes in the continuous case of #(n 2/3 ) for FF , #(n 1/2 log 3/4 n) for BF and #(n 1/2 log 1/2 n) for the best possible online algorithm. (Here the notation #(f(n)) means that the lower bound is taken in the Hardy and Littlewood sense of "not o(f(n))," i.e., "greater than cf(n) for some Table Possibilities for expected waste in the nonsymmetric case. OPT #u (1) [3] # k (1) [.] online# u (n 1/2 ), Ou (n 1/2 log 3/4 n) [.] [25] # k (1) [.] . Results proved in this paper. ruled out by theorems, but no occurrences are known either. For BFD and FFD, it does not occur for any k # 10,000 [5]. # This is conjectured to hold for all u # (0, 1), based on experimental studies. To date it has been proved only for u # [0.66, 2/3) and BF [18]. c > 0 and infinitely many n," rather than in the standard Knuthian sense of "greater than cf(n) for some c > 0 and all su#ciently large n.") In a sense, however, the online results for the discrete case are consistent with those for the continuous one. Although technically an online algorithm is not allowed to know the magnitude of n, if one formally sets in the formulas for expected waste for the discrete case, one gets EW A BF , and the best possible online algorithm. SS is not applicable to continuous distributions, but note that in this asymptotic discrete sense it appears to be worse than FF and BF . Indeed, experiments suggest that EW SS Let us now turn to the nonsymmetric distributions U(0, u], u < 1, and U{j, k}, 1. The known results for these distributions are summarized in Table 2. Here for the first time we see di#erences between the continuous and discrete cases for o#ine algorithms. In particular, for A # {FFD,BFD}, EW A for all u # 1/2 [3, 16], but for many of the distributions U{j, k} with j # k/2 (the corresponding discrete uniform distributions), we have EW A Moreover, for u # (1/2, 1), EW A growth rate never occurs for U{j, k}. This follows from a theorem in [5] that says that for all discrete distributions F , EW FFD must be either O(1), #(n 1/2 ), or #(n). The theorem also provides algorithms that determine the answers for a given distribution (b, S, # p), find the constants of proportionality when the expected waste is linear, and run in time polynomial in b and |S|. Unfortunately, although the answers for the distributions U{j, k} with k # 10,000 have all been computed [5], these do not suggest any simple rule as to how the choice among O(1), #(n 1/2 ), and #(n) might depend on j and k. (As an example of the type of behavior that can occur, for the U{j, 151} distributions the choice between linear and bounded expected waste switches back and forth 10 times as j increases from 1 to 149.) The results for k # 10,000 do, however, exhibit several suggestive patterns. First (and this can be proved to hold for arbitrarily large k), the expected waste is O(1) whenever j < # k or expected waste #(n 1/2 ) does not occur for any U{j, k} with suggesting that it may never occur. Third, for each U{j, k} with (U{j, k}) are either both linear or both bounded. If linear, the constant of proportionality for BFD is never larger than that for FFD (but is occasionally smaller). Here again there is a sense in which the discrete case is asymptotically consistent with the continuous case, even though the expected waste for FFD and BFD is always sublinear in the latter. As k increases, the maximum constants of proportionality for the linear expected waste under U{j, k} appear to decrease. Indeed, it can be 106 COFFMAN ET AL. shown that the constants for FFD are bounded by a function that declines at least as fast as (log k)/k [5]. (This presumably holds for BFD as well.) The worst case is the distribution U{6, 13}, for which the expected waste for both FFD and BFD is n/624, which is less than 0.6% of the expected optimal number of bins. Moreover, this is easily avoided, since not only does SS have bounded expected waste for this distribution, but so do FF and BF (although this is the only case we have identified where the online FF and BF algorithms outperform their o#ine cousins). Turning now to the online algorithms FF and BF , we observe that their behavior under discrete uniform distributions appears empirically to be similar to their behavior under continuous ones. Based on extensive experiments, it is conjectured that FF and BF both have linear expected waste under U(0, u] for 0 < u < 1, although to date this has been proved only for u # [0.66, 2/3) and BF [18]. In the discrete case (U{j, k} with experiments suggest that for su#ciently large k, the expected waste for FF and BF is bounded when linear. Some of this has been proved. In [19] it was shown that EW BF 0, and this result was extended to FF in [1]. In [4] it was shown that EW FF In practice, bounded expected waste is more common, at least for small k. The growth rates for BF under U{j, k} with k # 11 were completely characterized using multidimensional Markov chain arguments in [8], and linear expected waste only occurs for U{8, 11}. The only general result proving linear expected waste mirrors the result for the continuous case: EW BF k is su#ciently large [18]. At this point we do not know if expected wastes other than O(1) and #(n) are possible for FF or BF under any distributions U{j, k} with classification theorem such as those for FFD, BFD, or OPT has been proven, so the range of possibilities is not known to be limited to O(1), #(n 1/2 ), and #(n), as it was for FFD and BFD. There is also a gap between the lower bound proved in this paper on the best possible online expected waste for continuous distributions U(0, u] and the best rate known to be achievable. The former is# (n 1/2 ) and the latter is O(n 1/2 log 3/4 n), as proved in [25]. The algorithm of [25] works for any distribution, discrete or continuous, but has drawbacks from a pragmatic point of view: the best current bound we have on its running time is O(n 8 log 3 n) [12]. If one is willing to consider more specialized algorithms, better running times are possible, at least theoretically. For any fixed distribution F , there is an algorithm AF that runs in time O(n log n) and again has expected waste of O(n 1/2 log 3/4 n) [24]. These algorithms have drawbacks too, however, since the proof that they exist is nonconstructive. The question of whether practical algorithms exist that attain these bounds, or indeed whether lower bound is achievable, remains open. Finally, in addition to the open problems mentioned above for the discrete and continuous uniform distributions U(0, u] and U{j, k}, there is the question of what happens for arbitrary discrete and continuous distributions. In the discrete case, the above-mentioned classification theorems apply for BFD, FFD, and OPT , and say that the corresponding expected waste must be O(1), #(n 1/2 ), or #(n). As also mentioned above, the applicable cases for BFD and FFD and any specific distribution can be determined in time polynomial in b and |S|. For OPT there is also an algorithm for determining which case applies, as noted in [12]. This involves solving up to |S| +1 linear programs with |S|b variables and |S| constraints. None of these algorithms technically runs in polynomial time since b may be exponentially larger than its contribution to instance size (log b). However, all are feasible for b in PERFECT PACKING THEOREMS 107 excess of 1,000, which makes it possible to characterize the behavior of FFD, BFD, and OPT for many interesting distributions on a case-by-case basis. The theorem about SS presented in section 1 can be generalized to arbitrary discrete distributions if one replaces SS by a simple variant SS # : As in SS, items are packed so as to minimize h , but now the choice must be made subject to the following additional constraint: No item may be placed in a partially filled bin if the resulting gap cannot be exactly filled with items whose sizes have already been encountered in the list L. The resulting algorithm still runs in time O(nb) and satisfies distributions F [12]. In addition, there is a more complicated randomized variant that runs in time O(nb log b), satisfies the above property, and also has the same constant of proportionality as OPT when the expected waste is linear [12]. As to the case of arbitrary continuous distributions, we as yet have no general classification theorems, although some partial results have been proved. Rhee [22] provided a complicated measure-theoretic characterization of those F for which is sublinear, but this does not appear to be computationally useful. A result of Rhee and Talagrand [23] implies that if EW OPT must be O(n 1/2 ) or better. Rates strictly between O(1) and #(n 1/2 ) have not yet been ruled out, however. Moreover, there is as yet no algorithm with the general e#ectiveness of SS and its variants. The results of [24, 25] imply that there are online algorithms whose expected waste is at most O(n 1/2 log 3/4 n) worse than the optimal expected waste. For o#ine algorithms, Karmarkar and Karp have devised an algorithm which in the worst case never uses more than the optimal number of bins plus O log 2 (OPT O(log 2 n) [17]. This means that its expected waste is never more than the maximum of O(log 2 n) and EW OPT the algorithms of [24, 25], however, it is impractical, having a running time for which our best current bound is O(n 8 log 2 n). To conclude with an open problem that hearkens back to the main result of this paper, note that our ability to determine the expected waste for FFD, BFD, and OPT on a case-by-case basis can only take us so far, and more general results would be desirable. Results for U(0, 1] and U{k - 1, k} typically continue to hold for arbitrary continuous and discrete symmetric distributions, respectively, but the real world is not dominated by symmetric distributions. It would be nice if we could identify additional interesting classes of nonsymmetric distributions F for which general results about EW OPT can be proved, as we did in this paper for the discrete uniform distributions. Are there interesting classes for which new perfect packing theorems can provide us with similar general answers? --R An experimental study of bin packing Some unexpected expected behavior results for bin packing Stability of on-line bin packing with random arrivals and long-run average constraints On the sum-of-squares algorithm for bin packing Computers and Intractability: A Guide to the Theory of NP-completeness Average Case Behavior of First Fit Decreasing and Optimal Packings for Continuous Uniform Distributions U(0 Linear waste of best fit bin packing on skewed distribu- tions stochastic analysis of best fit bin packing An Average-Case Analysis of Bin Packing with Uniformly Distributed Item Sizes Optimal bin packing with items of random sizes Optimal bin packing with items of random sizes. The average case analysis of some on-line algorithms for bin packing How to pack better than Best Fit: Tight bounds for average-case on-line bin pack- ing --TR --CTR Janos Csirik , David S. Johnson , Claire Kenyon , James B. Orlin , Peter W. Shor , Richard R. Weber, On the Sum-of-Squares algorithm for bin packing, Journal of the ACM (JACM), v.53 n.1, p.1-65, January 2006
approximation algorithms;bin packing;average-case analysis;online
588918
Convergence Properties of an Augmented Lagrangian Algorithm for Optimization with a Combination of General Equality and Linear Constraints.
We consider the global and local convergence properties of a class of augmented Lagrangian methods for solving nonlinear programming problems. In these methods, linear and more general constraints are handled in different ways. The general constraints are combined with the objective function in an augmented Lagrangian. The iteration consists of solving a sequence of subproblems; in each subproblem the augmented Lagrangian is approximately minimized in the region defined by the linear constraints. A subproblem is terminated as soon as a stopping condition is satisfied. The stopping rules that we consider here encompass practical tests used in several existing packages for linearly constrained optimization. Our algorithm also allows different penalty parameters to be associated with disjoint subsets of the general constraints. In this paper, we analyze the convergence of the sequence of iterates generated by such an algorithm and prove global and fast linear convergence as well as show that potentially troublesome penalty parameters remain bounded away from zero.
Introduction . Introduction In this paper, we consider the problem of calculating a local minimizer of the smooth function where x is required to satisfy the general equality constraints and the linear inequality constraints Here f and c i map ! n into !, A is a p-by-n matrix and b 2 ! p . A classical technique for solving problem (1.1)-(1.3) is to minimize a suitable sequence of augmented Lagrangian functions. If we only consider the problem (1.1)- (1.2), these functions are defined by where the components - i of the vector - are known as Lagrange multiplier estimates and - is known as the penalty parameter (see, for instance, Hestenes [18], Powell [23] and Bertsekas [3]). The question then arises how to deal with the additional linear inequality constraints (1.3). The case where A is the identity matrix (that is when specifies bounds on the variables) has been considered by Conn et al. in [5] This research was supported in part by the Advanced Research Projects Agency of the Department of Defense and was monitored by the Air Force Office of Scientific Research under Contract No F49620-91-C-0079. The United States Government is authorized to reproduce and distribute reprints for governmental purposes notwithstanding any copyright notation hereon. This work was also supported by the Belgian national Fund for Scientific Research. and [7]. They propose keeping these constraints explicitly outside the augmented Lagrangian formulation, handling them directly at the level of the augmented Lagrangian minimization. That is, a sequence of optimization problems, in which (1.4) is approximately minimized within the region defined by the simple bounds, is attempted. In this approach all linear inequalities other than bound constraints are converted to equations by introducing slack variables and incorporated in the augmented Lagrangian function. This strategy has been implemented and successfully applied within the LANCELOT package for large-scale nonlinear optimization (see Conn et al. [6]). How- ever, such a method may be inefficient when linear constraints are present as there are a number of effective techniques specifically designed to handle such constraints directly (see Arioli et al. [1], Forsgren and Murray [14], Toint and Tuyttens [24], or and Carpenter [25], for instance). This is especially important for large-scale problems. The purpose of the present paper is therefore to define and analyze an algorithm where the constraints (1.3) are kept outside the augmented Lagrangian and handled at the level of the subproblem minimization, thus allowing the use of specialized packages to solve the subproblem. Our proposal extends the method of Conn et al. [5] in that not only bounds but general linear inequalities are treated separately. Fletcher [13, page 295] remarks on the potential advantages of this strategy. Furthermore, it is often worthwhile from the practical point of view to associate different penalty parameters to subsets of the general constraints (1.2) to reflect different degrees of nonlinearity. This possibility has been considered by many authors, including Fletcher [13, page 292], Powell [23] and Bertsekas [3, page 124]. In this case, the formulation of the augmented Lagrangian (1.4) can be refined: we partition the set of constraints (1.2) into q disjoint subsets fQ j g q redefine the augmented Lagrangian as where - is now a q-dimensional vector, whose j-th component is - j ? 0, the penalty parameter associated with subset Q j . Because of its potential usefulness, and because its analysis is difficult to directly infer from the single penalty parameter case, this refined formulation will be adopted in the present paper. The theory presented below handles the linear inequality constraints in a purely geometric way. Hence the same theory applies without modifications if linear equality constraints are also imposed and all the iterates are assumed to stay feasible with respect to these new constraints. It is indeed enough to apply the theory in the affine subspace corresponding to this feasible set. As a consequence, linear constraints need not be included in the augmented Lagrangian and thus have the desirable property that they have no impact on the structure of its Hessian matrix. The paper is organized as follows. In Section 2, we introduce our basic assumptions on the problem and the necessary terminology. Section 3 presents the proposed algorithm and the definition of a suitable stopping criterion for the subproblem. The global convergence analysis is developed in Section 4 while the rate of convergence is analyzed in Section 5. Second order conditions are investigated in Section 6. Section 7 considers some possible extensions of the theory. Finally, some conclusions and perspectives are outlined in Section 8. 2. The problem and related terminology. We consider the problem stated in (1.1)-(1.3) and make the following assumptions. AS1: The region nonempty. AS2: The functions f(x) and c i (x), are twice continuously differentiable for all x 2 B. Assumption AS1 is clearly necessary for the problem to make sense. We note that it does not prevent B from being unbounded. We now introduce the notation that will be used throughout the paper. Let g(x) denote the gradient r x f(x) of f(x) and H(x) denote its Hessian matrix r xx f(x). We also define J(x) to be the m-by-n Jacobian of c(x), where Hence denote the Hessian matrix r xx c i (x) of c i (x). Finally, let g ' (x; -) and denote the gradient, r x '(x; -), and the Hessian matrix, r xx '(x; -), of the Lagrangian function We note that '(x; -) is the Lagrangian solely with respect to the c i constraints. If we define first-order Lagrange multiplier estimates componentwise as where w [S] denotes the jSj-dimensional subvector of w whose entries are indexed by the set S, we shall use the identity r x \Phi(x; Now suppose that fx k 2 Bg, f- k g and f- k g are infinite sequences of n-vectors, m-vectors and positive q-vectors, respectively. For any function F , we shall use the notation that F k denotes F evaluated with arguments x So, for instance, using the identity (2.2), we have that where we have written (2.1) in the compact form We denote the vector w at iteration k by w k and its i-th component by w k;i . We also use w k;[S] to denote the jSj-dimensional subvector of w k whose entries are indexed by S. Now let fx k subset K of the natural numbers N, be a convergent subsequence with limit point x . Then we denote the matrix whose rows are those of A corresponding to active constraints at x - that is the constraints which are satisfied as equalities at x - by A . Furthermore, we choose Z to be a matrix whose columns form an orthonormal basis of the null space of A , that is A Z We define the least-squares Lagrange multiplier estimates (corresponding to A ) at all points where the right generalized inverse of J(x)Z is well defined. We note that, whenever J(x)Z has full rank, -(x) is differentiable and its derivative is given in the following lemma Lemma 2.1. Suppose that AS2 holds. If J(x)Z Z T J(x) T is nonsingular, -(x) is differentiable and its derivative is given by where the i-th row of R(x) is (Z T Proof. The result follows by observing that (2.5) may be rewritten as g(x) and J(x)Z for some vector r(x). Differentiating (2.7) and eliminating the derivative of r(x) from the resulting equations gives the required result. 2 We stress that, as stated, the Lagrange multiplier estimate (2.5) is not directly calculable as it requires a priori knowledge of x . It is merely introduced as an analytical device. Finally, the symbol k \Delta k will denote the ' 2 -norm or the induced matrix norm. We are now in position to describe more precisely the algorithm that we propose to use. 3. Statement of the algorithm. We consider the algorithmic model we wish to use in order to solve the problem (1.1)-(1.3). This model proceeds at iteration k by computing an iterate x k which satisfies (1.3) and approximately solves the subproblem min x2B where the values of the Lagrange multipliers - k and penalty parameters - k are fixed for the subproblem. Subsequently we update the Lagrange multipliers and/or decrease the penalty parameters, depending on how much the constraint violation for (1.2) has been reduced within each subset of the constraints. The motivation is simply to ensure global convergence by driving, in the worst case, the penalty parameters to zero, in which case the algorithms essentially reduce to the quadratic penalty function method (see, for example, Gould [15]). The tests on the size of the general constraint violation are designed to allow the multiplier updates to take over in the neighbourhood of a stationary point. The approximate minimization for problem (3.1) is performed in an inner iteration which is stopped as soon as its current iterate is "sufficiently critical". We propose to base this decision on the identification of the linear constraints that are "dominant" at x (even though they might not be active) and on a measure of criticality for the part of the problem where those constraints are irrelevant. Given ! - 0, a criticality tolerance for the subproblem, we define, for a vector x 2 B, the set of dominant constraints at x as the constraints whose indices are in the set for some - 0 ? 0. Here a T is the i-th row of the matrix A and b i is the corresponding component of the right-hand side vector b. Denoting by A D(x;!) the submatrix of A consisting of the row(s) whose index is in D(x; !), we also define the cone spanned by the outward normals of the dominant constraints. The associated polar cone is then where cl(V ) denotes the closure of the set V . The cone T (x; !) is the tangent cone with respect to the dominant constraints at x for the tolerance !. Note that D(x; !) might be empty, in which case A D(x;!) is assumed to be zero, N(x; !) reduces to the origin and T (x; !) is the full space. We then formulate our "sufficient criticality" criterion for the subproblem as fol- lows: we require that is the projection onto the convex set V and ! k is a suitable tolerance at iteration k. Once x k satisfying (3.3) has been determined by the inner iteration, we denote For future reference, we define Z k to be a matrix whose columns form an orthonormal basis of V k , the null space of AD k , and Y k to be a matrix whose columns form an orthonormal basis of W k . As above, we have that T k is the full space and N k reduces to the origin when D k is empty. We note that, in this case, Z I , the identity operator, and Y We also note that V k ' T k , and hence that since Z k Z T k is the orthogonal projection onto V k . It is important to note that the stopping rule (3.3) covers a number of more specific choices, including the rule used in much existing software for linearly constrained optimization (such as MINOS [21], LSNNO [24], or VE14 and VE19 from the Harwell Subroutine Library [17]). The reader is referred to Section 7.2 for further details. We are now in position to describe our algorithmic model more precisely. In this model, we define ff k to be the maximum penalty parameter at iteration k (see (3.10)). At this iteration, the parameters ! k and j k represent criticality and feasibility levels, respectively. partition of the set disjoint subsets is given, as well as initial vectors of Lagrange multiplier estimates - 0 and positive penalty parameters - 0 such that The strictly positive constants - are specified. Set ff approximately solves (3.1), i.e. such that (3.3) holds. [Test for convergence]. If kP T k Step 3 [Disaggregated updates]. For or Step 3b otherwise. Step 3a [Update Lagrange multiplier estimates]. Set Step 3b [Reduce the penalty parameter]. Set where Step 4 [Aggregated updates]. Define If then set otherwise set Increment k by one and go to Step 1. Algorithm 3.1 is specifically designed for the first-order estimate (2.1), a formula with potential advantages for large-scale computations. We refer the reader to Section 7.1 for a further discussion of a more flexible choice of the multipliers, covering, among others, the choice of the least-squares estimates -(x) as defined in (2.5). We immediately verify that our algorithm is coherent, in that lim Indeed, we obtain from (3.6) that ff k ! 1 for all k, and (3.14) then follows from (3.12) and (3.13) if ff k tends to zero, or from (3.13) alone if ff k is bounded away from zero. The restriction (3.6) is imposed in order to simplify the exposition. In a more practical setting, it may be ignored provided the definition of ff 0 and (3.10) are replaced by and ff respectively, for some constant fl s 2 (0; 1), and that (3.11) is replaced by Algorithm 3.1 may be extended in other ways. For instance, one may replace the definition of ! 0 , the first equation in (3.12) and the first equation of (3.13) by . The definition of j 0 and the second equation in (3.12) may then be replaced by for some j s ? 0. None of these extensions alter the results of the convergence theory developed below. The values used in the LANCELOT package in a similar context are (relation (3.15) is also used with ensuring that 0:01). The values also seem suitable. The parameters ! and j specify the final accuracy requested by the user. Finally, the purpose of the update (3.9) is to put more emphasis on the feasibility of the constraints whose violation is proportionally higher, in order to achieve a "balance" amongst all constraint violations. This balance then allows the true asymptotic regime of the algorithm to be reached. The advantage of (3.9) is that this balancing effect is obtained gradually, and not enforced at every major iteration, as is the case in Powell [23]. Furthermore Powell's approach increases the penalties corresponding to the constraints that are becoming too slowly feasible, based on the ' 1 -norm. Thus it is only when they have changed sufficiently so that they are all within the constraint violation tolerance that the Lagrange multiplier update is performed. By contrast, we update the multipliers of the well-behaved constraints (assuming they correspond to a particular partition - which is likely since that is, partly at least, why the partitions exist) independently of more badly behaved ones. In addition, by virtue of using the -norm, we do not give quite the same emphasis to the most violated constraint. 4. Global convergence analysis. We now proceed to show that Algorithm 3.1 is globally convergent under the following assumptions. AS3: The iterates fx k g considered lie within a closed, bounded domain\Omega\Gamma AS4: The matrix J(x )Z has column rank no smaller than m at any limit point, x , of the sequence fx k g considered in this paper. We notice that AS3 implies that there exists at least a convergent subsequence of iterates, but does not, of course, guarantee that this subsequence converges to a stationary point, i.e. that "the algorithm works". We also note that it is always satisfied in practice because the linear constraints (1.3) includes lower and upper bounds on the variables, either actual or implied by the finite precision of computer arithmetic. Assumption AS4 guarantees that the dimension of the null space of A is large enough to provide the number of degrees of freedom that are necessary to satisfy the nonlinear constraints and we require that the gradients of these constraints (projected onto this null space) are linearly independent at every limit point of the sequence of iterates. This assumption is the direct generalization of AS3 used by Conn et al. [5]. We shall analyse the convergence of our algorithm in the case where the convergence tolerances ! and j are both zero. We first need the following lemma, proving that (3.3) prevents both the reduced gradient of the augmented Lagrangian and its orthogonal complement from being arbitrarily large when ! k is small. Lemma 4.1. Let fx k g ae B; k 2 K, be a sequence which converges to the point x and suppose that where the ! k are positive scalar parameters which converge to zero as k 2 K increases. Then for some - 1 ? 0 and for all k 2 K sufficiently large. Proof. Observe that, for k 2 K sufficiently large, ! k is sufficiently small and sufficiently close to x to ensure that all the constraints in D k are active at x . This implies that the subspace orthogonal to the normals of the dominant constraints at x k , V k , contains the subspace orthogonal to the normals of the constraints active at x . Hence, we deduce that where we have used (3.5) to obtain the second inequality and (3.3) to deduce the third. This proves the first part of (4.1). We now turn to the second. If D k is empty, then Y k is the zero matrix and the second part of (4.1) immediately follows. Assume therefore that D k 6= ;. We first select a submatrix - of AD k that is of maximal full row-rank and note that the orthogonal projection onto the subspace spanned by the fa i g i2D k is nothing but A T A T Hence we obtain from the orthogonality of Y k , the bound jD k j - p, (3.2) and (3.4) and the fact that all constraints in D k are active at x for k sufficiently large, that A T A T A T A T But there are only a finite number of nonempty sets D k for all possible choices of x k and we may thus deduce the second part of (4.1) from (4.2) by defining A T A T where the minimum is taken on all possible choices of D k and - . 2 We now examine the behaviour of the sequence fr x \Phi k g. We first recall a result extracted from the classical perturbation theory of convex optimization problems. This result is well known and can be found, for instance, in [12, pp. 14-17]. Lemma 4.2. Assume that U is a continuous point-to-set mapping from S ' ! ' into the power set of ! n such that the set U(') is convex and non-empty for each S. Assume that the real-valued function F (y; ') is defined and continuous on the space and convex in y for each fixed '. Then, the real-valued function F defined by is continuous on S. We now show that, if it converges, the sequence fr x \Phi k g tends to a vector which is a linear combination of the rows of A with non-negative coefficients. Lemma 4.3. Let fx k g ae B,k 2 K, be a sequence which converges to the point x and suppose that the gradients r x \Phi k , k 2 K, converge to some limit r x \Phi . Assume furthermore that (3.3) holds for k 2 K and that ! k tends to zero as k 2 K increases. Then, r x \Phi for some vector - - 0, where A is the matrix whose rows are those of A corresponding to active constraints at x . Proof. We first define with the aim to show that this quantity tends to zero when k 2 K increases. We obtain from (4.3), the Moreau decomposition [20] of r x \Phi k and the Cauchy-Schwarz inequality, that 1g. As, for sufficiently close to x and ! k sufficiently small, all the constraints in D k must be active at x , we have that N k is included in the normal cone N(x ; 0) and therefore the vector PN k belongs to this normal cone. Moreover, since the maximization problem of the last right-hand side of (4.4) is a concave program, since x is feasible for (1.3), and since kx large enough, we thus deduce that is a global solution of this problem. Observing that we obtain that where we used the Cauchy-Schwarz inequality to deduce the last inequality. We may now apply Lemma 4.1 and deduce from the second part of (4.1), (4.5) and the contractive character of the projection onto a convex set containing the origin that and thus, from (4.4) and our assumptions, that Our assumption on the ! k sequence then implies that oe k converges to zero as k increases in K. Consider now the minimization problem d; subject to A(x Since the sequences fr x \Phi k g and fx k g converge to r x \Phi and x respectively, we deduce from Lemma 4.2 applied to the optimization problem (4.3) (with the choices and the convergence of the sequence oe k to zero that the optimal value for problem (4.6) is zero. The vector thus a solution for problem (4.6) and satisfies r x \Phi for some vector - - 0, which ends the proof. 2 The important part of our convergence analysis is the next lemma. Lemma 4.4. Suppose that AS1 and AS2 hold. Let fx k g ae B; k 2 K, be a sequence satisfying AS3 which converges to the point x for which AS4 holds and let - where - satisfies (2.5). Assume that f- k g, k 2 K, is any sequence of vectors and that nonincreasing sequence of q-dimensional vectors. Suppose further that (3.3) holds where the ! k are positive scalar parameters which converge to zero as increases. Then (i) There are positive constants - 2 and - 3 such that and for all sufficiently large. Suppose, in addition, that c(x (ii) x is a Kuhn-Tucker point (first-order stationary point) for the problem (1.1)- is the corresponding vector of Lagrange multipliers, and the sequences converge to - for k 2 K; (iii) The gradients r x \Phi k converge to g ' Proof. As a consequence of AS2-AS4, we have that for k 2 K sufficiently large, exists, is bounded and converges to (J(x )Z . Thus, we may write for some constant - 2 ? 0. Equations (2.3) and (2.4), the inner iteration termination criterion (3.3) and Lemma 4.1 give that for all k 2 K large enough. By assumptions AS2, AS3, AS4 and (2.5), -(x) is bounded for all x in a neighbourhood of x . Thus we may deduce from (2.5), (4.10) and (4.11) that Moreover, from the integral mean value theorem and Lemma 2.1 we have that Z 1r x -(x(s))ds \Delta where r x -(x) is given by equation (2.6), and where Now the terms within the integral sign are bounded for all x sufficiently close to x and hence for all k 2 K sufficiently large and for some constant - 3 ? 0, which implies the inequality (4.8). We then have that -(x k ) converges to - . Combining (4.12) and (4.14) we obtain which gives the required inequality (4.7). Then, since by assumption ! k tends to zero as k increases, (4.15) implies that - k converges to - and therefore, from the identity (2.3), r x \Phi k converges to g ' (x ; - ). Furthermore, multiplying (2.1) by - k;j , we obtain Taking norms of (4.16) and using (4.15), we derive (4.9). Now suppose that c(x Lemma 4.3 and the convergence of r x \Phi k to g ' (x ; - ) give that for some vector - - 0. This last equation and (4.17) show that x is a Kuhn-Tucker point and - is the corresponding set of Lagrange multipliers. Moreover (4.7) and (4.8) ensure the convergence of the sequences f -(x K. Hence the lemma is proved. 2 We finally require the following lemma in the proof of global convergence, which shows that the Lagrange multiplier estimates cannot behave too badly. Lemma 4.5. Suppose that, for some j (1 - j - q), - k;j converges to zero as k increases when Algorithm 3.1 is executed. Then the product - k;j k- converges to zero. Proof. As - k;j converges to zero, Step 3b must be executed infinitely often for the j-th subset. Let K be the set of indices of the iterations in which Step 3b is executed. We consider how the j-th subset of Lagrange multiplier estimates changes between two successive iterations indexed in the set K j . Firstly note that - kv+1;[Q j At iteration where the summation is null if Substituting (4.19) into (4.18), multiplying both sides by - kv+t;j , taking norms and using (3.9), yields and hence Using the fact that (3.7) holds for we deduce that Now defining we obtain that for all t such that k Thus, from (4.22) and the inequality - ! 1, if ae v converges to zero, then ffi v and hence, from (4.21), - kv+t;j k- both converge to zero. To complete the proof it therefore suffices to show that ae v converges to zero as v tends to infinity. Suppose first that ff k is bounded away from zero. Then we must have that (3.13) is used for all k sufficiently large, with ff 1). This and the definition of ae v in (4.20) imply that min for sufficiently large v. As (3.13) also guarantees that j k tends to zero, we deduce that ae v converges to zero. This completes the proof for the first case. Suppose now that ff k converges to zero. This implies that each of the q independent penalty parameters is reduced an infinite number of times. Consider the progress of ff k over the course of q successive decreases (3.11). As (3.11) only happens when the currently largest penalty parameter, - k;j say, is reduced, as (3.9) requires that this penalty parameter is reduced by - , and because there can only possibly be at most parameters in the interval (- k;j ; - k;j ], it follows that ff k must be reduced by at least - over q successive decreases (3.11). Thus, considering the possible outcomes (3.12) and (3.13), each j kv+l must be bounded by a quantity of the form t. Furthermore, at most q such terms can involve any particular i and t. Therefore, since - ff kv ! 1, we obtain that Thus we see that, as ff kv converges to zero, so does ae v , completing the proof for the second case. 2 We can now derive the desired global convergence property of Algorithm 3.1, which is analogous to Theorem 4.4 in Conn et al. [5]. Theorem 4.6. Assume that AS1 and AS2 hold. Let x be any limit point of the sequence fx k g generated by Algorithm 3.1 of Section 3 for which AS3 and AS4 hold and let K be the set of indices of an infinite subsequence of the x k whose limit is x . Finally, let - conclusions (i), (ii) and (iii) of Lemma 4.4 hold. Proof. Our assumptions are sufficient to reach the conclusions of part (i) of Lemma 4.4. We now show that c(x and therefore that c(x To see this, we consider a analyze two separate cases. The first case is when - k;j is bounded away from zero. Hence Step 3a must be executed every iteration for k sufficiently large, implying that (3.7) is always satisfied for k large enough. We then deduce from (3.14) that c(x k ) converge to zero. The second case is when - k;j converges to zero. Then Lemma 4.5 shows that tends to zero. Using this limit and (3.14) in (4.9), we obtain that tends to zero, as desired. As a consequence, conclusions (ii) and (iii) of Lemma 4.4 hold. 2 We finally note that global convergence of Algorithm 3.1 can be proved under much weaker assumptions on - k;j and ! k . The reader is again referred to Conn et al. [9] for further details. 5. Asymptotic convergence analysis. The distinction between dominant and non-dominant (floating) linear inequality constraints has some implications in terms of the identification of those constraints that are active at a limit point of the sequence of iterates generated by the algorithm. Given such a point x we know from Theorem 4.6 that it is critical, i.e. that \Gammag ' for the corresponding Lagrange multipliers - . If we now consider a linear constraint with index is active at x , we may define the normal cone N [i] to be the cone spanned by the outwards normals to all linear inequality constraints active at x , except the i-th one. We then say that the i-th linear inequality constraint is strongly active at x if \Gammag ' . In other words, the i-th constraint is strongly active at a critical point if this point ceases to be critical when this constraint is ignored. Let us denote by S(x ) the set of strongly active constraints at x . All non-strongly active constraints at x are called weakly active at x . We now prove the reasonable result that all strongly active constraints at a limit point x are dominant for k large enough. Theorem 5.1. Assume that AS1-AS3 hold. Let fx k g, k 2 K, be a convergent subsequence of iterates produced by Algorithm 3.1, whose limit point is x with corresponding Lagrange multipliers - . Assume furthermore that AS4 holds at x . Then for all k sufficiently large. Proof. Consider a linear inequality constraint i 2 S(x ). Then, by definition of this latter set, we have that \Gammag ' . Since Theorem 4.6 guarantees that r x \Phi k converges to g ' (x ; - ) and as N [i] is closed, we have that \Gammar x \Phi k 62 N [i] for large enough. Therefore, one obtains from the Moreau decomposition [20] of \Gammar x \Phi k that for some ffl ? 0 and for all sufficiently large k 2 K, where T [i] . We have also from (3.3) that kP T k (\Gammar x \Phi k )k is arbitrarily small, because ! k tends to zero (see (3.14)). Assume now that, for some arbitrarily large k 2 K, we have that i 62 D k . This implies that T [i] hence that (5.1) is impossible. We therefore deduce that i must belong to D k , which proves the theorem. 2 This result is important and is the generalization of Theorem 5.4 by Conn et al. [5]. It can also be interpreted as a means of active constraint identification, as is clear from the following easy corollary. Corollary 5.2. Suppose that the conditions of Theorem 5.1 hold. Assume furthermore that all linear inequality constraints active at x have linearly independent normals and are non-degenerate, in the sense that where ri[V ] denotes the relative interior of a convex set V . Then D k is identical to the set of active linear inequality constraints at x for all k 2 K sufficiently large. Proof. The non-degeneracy assumption and the linear independence of the active constraints normals imply that - is unique and only has strictly negative components. Therefore each of the active linear inequality constraints at x is strongly active at x , and the desired conclusion follows from Theorem 5.1. 2 We note here that the non-degeneracy assumption corresponds to strict complementarity slackness in our context (see, for instance, Dunn [11], or Burke et al. [4]). We now make some additional assumptions before pursuing our local convergence analysis. We intend to show that all penalty parameters are bounded away from zero. AS5: The second derivatives of the functions f(x) and c i (x) (1 - i - m) are Lipschitz continuous at any limit point x of the sequence of iterates fx k g. Suppose that (x ; - ) is a Kuhn-Tucker point for problem (1.1)-(1.3) and let I be any subset of the linear inequality constraints which are active at x that contains all strongly active constraints (S(x plus an arbitrary subset of weakly active constraints at x . Then, if the columns of the matrix Z form an orthonormal basis of the subspace orthogonal to the normals of the constraints in I, we assume that the matrix is nonsingular for all possible choices of the weakly active constraints in the set I. We note that AS6 implies AS4 and seems reasonable in that the definition of strongly and weakly active constraints may vary with small perturbations in the prob- lem, for instance when lies in one of the extreme faces of the cone N . Our assumption might be seen as a safeguard against the possible effect of all such perturbations. We now make the distinction between the subsets for which the penalty parameter converges to zero and those for which it stays bounded away from zero. We define We also denote and ae k We now prove an analog to Lemma 5.1 by Conn et al. [5] which is suitable for our more general framework. Lemma 5.3. Assume that AS1-AS3 hold. Let fx k g, k 2 K, be a convergent subsequence of iterates produced by Algorithm 3.1, whose limit point is x with corresponding Lagrange multipliers - . Assume that AS5 and AS6 hold at x . Assume furthermore that Z 6= ;. there are positive constants - and an integer k 1 such that, if ff k 1 - ff, then and (ii) If, on the other hand, P 6= ;, there are positive constants - and an integer k 1 such that, if - k 1 ;Z - ff, then and Proof. We will denote the gradient and Hessian of the Lagrangian function, taken with respect to x, at the limit point and H ' , respectively. Similarly, J will denote J(x ). We also define We observe that the assumptions of the lemma guarantee that Theorem 4.6 can be used. We first note that there is only a finite number of possible D k , and we may thus consider subsequences of K such that D k is constant in each subsequence. We also note that each k 2 K belongs to a unique such subsequence. In order to prove our result, it is thus sufficient to consider an arbitrary infinite subsequence - K such that, is independent of k. This "constant" index set will be denoted by D. As a consequence, the cones N k and T k , the subspaces V k and W k and the orthogonal matrices Z k and Y k are also independent of k; they are denoted by N , T , V , W , Z and Y , respectively. Using (2.3) and Taylor's expansion around x , we obtain that where ds and The boundedness and Lipschitz continuity of the Hessian matrices of f and c i in a neighbourhood of x , together with the convergence of - - k to - then imply that and for some positive constants - 8 and - 9 . Moreover, using Taylor's expansion again, along with the fact that Theorem 4.6 ensures the equality c(x we obtain that where Z 1s ds (see Gruver and Sachs [16, page 11]). The boundedness of the Hessian matrices of the c i in a neighbourhood of x then gives that for some positive constant - 10 . Combining (5.9) and (5.12), we obtain !/ where we have suppressed the arguments of the residuals r 1 , r 2 and r 3 for brevity. Using the orthogonal decomposition of ! n into V \Phi W and defining we may rewrite (5.14) as r 4 where r 4 . Expanding this last equation gives thatB @ (5.15)B @ r We now observe that (3.3), the inclusion V ' T and the fact that ! k tends to zero imply that 0: Substituting (5.16) in (5.15), removing the middle horizontal block and rearranging the terms of this latter equation then yields that !/ Roughly speaking, we now proceed by showing that the right-hand side of this relation is of the order of ' k We will then ensure that the vector on the left-hand side is of the same size, which is essentially the result we aim to prove. We first observe that from (4.1). We then obtain from (4.7) and (5.19) that . Furthermore, from (5.10), (5.11), (5.13), (5.19) and (5.20), . We now bound c(x k ) by distinguishing components from Z and P . We first note that, since the penalty parameters for each subset in P are bounded away from zero, the test (3.7) is satisfied for all k sufficiently large. Moreover, the remaining components of c(x k ) satisfy the bound for all j 2 Z and all k sufficiently large, using (4.9). Hence, using (5.3), (3.7) and (5.22), we deduce that Note that the first term of the last right-hand side only appears if P is not empty. Since the algorithm ensures that because may obtain from (4.1), (5.23) and (5.19) that assumption AS6, the coefficient matrix on the left-hand side of (5.17) is nonsingular. Let M be the norm of its Multiplying both sides of the equation by this inverse and taking norms, we obtain from (5.18), (5.21), (5.24) and (5.25) that Suppose now that k is sufficiently large to ensure that and let Recall that ff 0 and hence - ff, the relations (5.26)-(5.28) give As converge to zero, we have that for k large enough. Hence inequalities (5.29) and (5.30) yield that If P is empty, we use (5.19), (5.31) and (5.18), the fact that - and the inequality to deduce (5.4), where - 4 deduce (5.5) from (4.7) and (5.4). Now, using (2.1), and (5.6) then follows from (5.32) and (5.5). If, on the other hand, P is not empty, (5.7) results from (4.7), (5.19), (5.31) with Finally, (5.8) results from (2.1) and (5.7). 2 For the remaining of this section, we will restrict our attention to the case where the sequence of iterates converges to a single limit point. Obviously, this makes AS3 unnecessary. We briefly comment at the end of the section on why this additional assumption cannot be relaxed. We now show that, if the maximum penalty parameter ff k converges to zero, then the Lagrange multiplier estimates - k converge to their true values - . Lemma 5.4. Assume AS1 and AS2 hold. Assume that fx k g, the sequence of iterates generated by Algorithm 3.1, converges to the single limit point x at which AS6 holds, and with corresponding Lagrange multipliers - . Then, if ff k tends to zero, the sequence - k converges to - . Proof. Recall that AS6 implies AS4 and therefore that our assumptions are sufficient to apply Theorem 4.6. We observe that the desired convergence holds if - k;[Q j ] converges to - It is thus sufficient to show this latter result for an arbitrary j between 1 and q. The result is obvious if Step 3a is executed infinitely often for the j-th subset. Indeed, each time this step is executed, - and the inequality (4.7) guarantees that - converges to - ;[Q j ] . Suppose therefore that Step 3a is not executed infinitely often for this subset. Then k(- remain fixed for all executed for each remaining iteration. But then implies that kc(x k ) As ff k tends to zero and ff k sufficiently large for which ff k strictly decreases. But then inequality (3.7) must be satisfied for some k - k 3 , which is impossible, as this would imply that Step 3a is again executed for the j-th subset. Hence Step 3a must be executed infinitely often.We now consider the behaviour of the maximum penalty parameter ff k and show the important result that, under stated assumptions, it is bounded away from zero. The proof of this result is inspired by the technique developed by Conn et al. [5]. When the single penalty parameter definition of the augmented Lagrangian (1.4) is used (or, equivalently, when one then avoids a steadily increasing ill-conditioning of the Hessian of the augmented Lagrangian. Note that this ill-conditioning is also avoided when q ? 1, as we show below in Theorem 5.6. Theorem 5.5. Assume AS1 and AS2 hold and suppose that the sequence of iterates fx k g of Algorithm 3.1 converges to a single limit point x with corresponding Lagrange multipliers - , at which AS5 and AS6 hold. Then there is a constant ff min 2 (0; 1) such that ff Proof. Suppose otherwise that ff k tends to zero (that is that - k;j tends to zero for each j between 1 and q. Then Step 3b must be executed infinitely often for each subset. We aim to obtain a contradiction to this statement by showing that Step 3a is always executed for each subset for sufficiently large k. We note that our assumptions are sufficient to apply Theorem 4.6. Furthermore, we may apply Lemma 5.3 to the complete sequence of iterates. First observe that for all k - k 1 , where - ff and k 1 are those of Lemma 5.3. Note that for all k - k 1 . This follows by definition if (3.12) is executed. Otherwise it is a consequence of the fact that ff k is unchanged while ! k is reduced, when (3.13) occurs. Let k 4 be the smallest integer k such that and that (5.33) and (5.35) imply that 5 be such that for all k - k 5 , which is possible because of Lemma 5.4. Now define k let \Gamma be the set fk j (3.12) is executed at iteration k \Gamma 1 and k - k 6 g and let k 0 be the smallest element of \Gamma. By the assumption that ff k tends to zero, \Gamma has an infinite number of elements. By definition of \Gamma, for iteration k 0 , . Then inequality (5.6) gives that, for each j, (from (5.36)) (from (5.33)) (from (5.34)). As a consequence of this inequality, Step 3a will be executed for each j with - k 0 Inequality (5.5) together with (5.37) guarantee that We shall now make use of an inductive proof. Assume that, for each j, Step 3a is executed for iterations and that Inequalities (5.38) and (5.39) show that this is true for We aim to show that the same is true for Our assumption that Step 3a is executed gives that, for iteration . Then, inequality (5.6) yields that, for each j, (from (5.40)) (from (5.36)) Hence Step 3a will again be executed for each j with Inequality (5.5) then implies that (from (5.40)) (from (5.35)) which establishes (5.40) for executed for each all iterations k - k 0 . But this implies that \Gamma is finite, which contradicts the assumption that Step 3b is executed infinitely often for each subset. Hence the theorem is proved.This theorem was all that was needed in Conn et al. [5]. However, the situation is more complex here because q may be larger than one. If the ill-conditioning of the Hessian is to be avoided, we must now prove the stronger result that all penalty parameters stay bounded away from zero. Theorem 5.6. Assume AS1 and AS2 hold and suppose that the sequence of iterates fx k g of Algorithm 3.1 converges to a single limit point x with corresponding Lagrange multipliers - , at which AS5 and AS6 hold. Then there is a constant - ? 0 such that - k;j - for all k and all Proof. Assume otherwise that Z is not empty, and hence that - k;Z converges to zero. Then Step 3b must be executed infinitely often for j 2 Z . We aim to obtain a contradiction to this statement by showing that, for any j 2 Z , Step 3a is always executed for sufficiently large k. We may deduce from Theorem 5.5 that ff k attains its minimum value ff min 2 (0; 1) at iteration k max , say. Hence, P 6= ;. Furthermore, we may apply Lemma 5.3 to the complete sequence of iterates. Let k 7 - k max be the smallest integer for which ff and k 1 are those of Lemma 5.3, and where Note that ff fi j +ffl Consider the j-th subset, for some j 2 Z . At iteration k - k 7 , the algorithm ensures that if Step 3b is executed for the j-th subset, while (5.7) ensures that if Step 3a is executed for the same subset. Summing on all j 2 Z , and defining executed for the j-th subset at iteration kg executed for the j-th subset at iteration kg; we obtain that ae For the purpose of obtaining a contradiction, assume now that ae k -2 for all k - k 7 . Then (5.42) gives that, for all k - k 7 , ae k+1 ae k because of (5.41). Hence we obtain from (5.44) that ae Therefore, since ae k 7 ff (k\Gammak 7 +1)ffl min tends to zero, we obtain that ae k+1 !2 ff ff j+(k 7 \Gammak min ff (k\Gammak 7 +1)fij for all sufficiently large k, where the last equality results from the definition of k max and (3.13). But this contradicts (5.43), which implies that (5.43) does not hold for all sufficiently large. As a consequence, there exists a subsequence K such that ae k !2 for all k 2 K. Consider such a k. Then, using (5.42) and (5.45), we deduce that ae where we have used (5.41) to obtain the second inequality. As a consequence, k+1 2 K and (5.45) holds for all k sufficiently large. Returning to subset j 2 Z , we now obtain from (5.8) and (5.45) that for all k sufficiently large, because of (5.41). Hence Step 3a is executed for the subset j and for all sufficiently large k, which implies that j does not belong to Z . Therefore Z is empty and the proof of the theorem is completed. 2 As in Conn et al. [5], we examine the rate of convergence of our algorithms. Theorem 5.7. Under the assumptions of Theorem 5.6, the iterates x k and the Lagrange multipliers - - k of Algorithm 3.1 are at least R-linearly convergent with R-factor at most ff fi j min , where ff min is the smallest value of the maximum penalty parameter generated by the algorithm. Proof. The proof parallels that of Lemma 5.3. First, Theorem 5.5 shows that the maximum penalty parameter ff k stays bounded away from zero, and thus remains fixed at some value ff min ? 0, for k - k max . For all subsequent iterations, (5. hold. Moreover, Theorem 5.6 implies that, for all hold for all sufficiently large. Hence and because of (4.1), the bound on the right-hand side of (5.25) may be replaced by - Therefore, if k is sufficiently large that and inequalities (5.47)-(5.49) can be rearranged to yield But then (5.19) gives that (5.50) show that x k converges to x at least R-linearly, with R-factor ff fi j min . Inequalities (4.7) and (5.50) then guarantee the same property for - To conclude this section, we note that the conclusions of Theorems 5.5, 5.6 and 5.7 require that the complete sequence of iterates converges to a unique limit point. As indicated above, this assumption cannot be relaxed. The counterexample presented by Conn et al. [5] (where the linear inequality constraints are simple bound constraints on the problem's variables) shows that the sequence of penalty parameters may indeed converge to zero, if there is more than a single limit point. 6. Second order conditions. If we further strengthen the stopping test for the inner iteration beyond (3.3) to include second-order conditions, we can then guarantee that our algorithms converge to an isolated local solution. More specifically, we require the following additional assumption. AS7: Suppose that x k satisfies (3.3), converges to x for k 2 K, such that Z has a rank strictly greater than m. Then, if Z is defined as in AS6, we assume that Z T r xx \Phi k Z is uniformly positive definite (that is, its smallest eigenvalue is uniformly bounded away from zero) for all k 2 K sufficiently large. We can then prove the following result. Theorem 6.1. Under assumptions AS1-AS3, AS5-AS7, the iterates x k , k 2 K, generated by Algorithm 3.1 converge to an isolated local solution of (1.1)-(1.3). Proof. By definition of \Phi, r xx \Phi (x) is the Jacobian of c(x) [Q j ] . Note that the rank of Z is at least that of Z . AS7 then implies that there exists a nonzero vector s such that and hence for each j. For any such vector, AS7 further implies that for some - 21 ? 0, which in turn gives that because of (6.1) and (6.2). By continuity of H ' as x k and - approach their limits, this ensures that for all nonzero s satisfying which implies that x is an isolated local solution of (1.1)-(1.3) (see, for instance, Avriel [2, Thm. 3.11]). 2 If we assume that the inner iteration stopping test is tightened so that r xx \Phi k is required to be uniformly positive definite in the null space of the dominant constraints, and if we assume that the non-degeneracy condition (5.2) holds, then Corollary 5.2 ensures that Z sufficiently large k and Theorem 6.1 holds. A weaker version of this result also holds, where only positive semi-definiteness of the augmented Lagrangian's Hessian is required, yielding then that x is a (possibly not isolated) minimizer of the problem. 7. Extensions. 7.1. Flexible Lagrange multiplier updates. The formula (2.1) has definite advantages for large-scale computations, but may otherwise appear unduly restrictive. The purpose of the first extension we consider is to introduce more freedom in our algorithmic framework, by replacing this formula by a more general condition, allowing a much larger class of Lagrange multiplier updates to be used. More specifically, we consider modifying Algorithm 3.1 as follows. Algorithm 7.1 This algorithm is identical to Algorithm 3.1, except that Step 3 is replaced by the following, where fl is a constant in (0; 1). Step 3 [Disaggregated updates]. Compute a new vector of Lagrange multiplier estimates - k+1 . For or Step 3b otherwise. Step 3a [Update Lagrange multiplier estimates]. Set Step 3b [Reduce the penalty parameter]. Set where - k;j is defined by (3.9) in Algorithm 3.1. Algorithm 7.1 allows a more flexible choice of the multipliers than Algorithm 3.1, but requires that some control is enforced to prevent their growth at an unacceptably fast rate. It covers, among others, the choice of the least-squares estimates -(x) as defined in (2.5). The global convergence theory presented in Section 4 for Algorithm 3.1 can be extended to cover Algorithm 7.1. This extension is detailed in Conn et al. [9]. Conn et al. [10] extend the local convergence analysis of Section 5 to Algorithm 7.1, under the additional condition that holds for some positive constants - 22 and - 23 and all k 2 K sufficiently large, where K is the index set of a subsequence of iterates (generated by Algorithm 7.1) converging to the critical point x with corresponding Lagrange multipliers - . Both (2.1) and (2.5) satisfy this condition because of Theorem 4.6. We also note that Corollary 5.2 ensures that the least-squares multiplier estimates are implementable when the non-degeneracy condition (5.2) holds. By this we mean that the estimates are identical to those defined in (2.5) for all k sufficiently large, and, unlike (2.5), are well defined when x is unknown. 7.2. Alternative criticality measures. In Algorithms 3.1 and 7.1 we used the criticality measure kP T k in order to define the stopping criterion of the inner iteration (see (3.3)), because it is general. However, this quantity might not be easily computed in the course of the numerical method used to calculate x k , especially when the dimension of the problem is high. It is therefore of interest to examine other criticality measures that might be easier to calculate. It is the purpose of this section to analyze such alternative proposals. Given D k , N k , and AD k as above, we first claim that (3.3) can be replaced by the requirement that there exists a set of non-positive "dominant multipliers" f- ik g i2M k is the jD k j-dimensional vector whose i-th component is - ik if or zero otherwise. We prove this claim. Lemma 7.1. Assume that there exists a non-positive - k such that (7.1) holds at x k . Then (3.3) also holds at x k . Proof. Since the vector A T belongs, by construction, to the cone N k defined in (3.4), we can immediately deduce from the definition of the othogonal projection and (7.1) that which is the desired inequality. 2 Condition (7.1) is appealing for two reasons. Firstly, a set of (possibly approx- imate) multipliers is available in many numerical procedures that might be used to perform the inner iteration and to compute a suitable x k ; one can then select those multipliers which correspond to the dominant constraints, further restrict this choice to the non-positive ones and finally check (7.1). Such a scheme is implicitly used by both the Harwell [17] barrier-function quadratic programming codes VE14 and VE19 and the IMSL [19] general linearly constrained minimization package LCONG. Alternatively, suitable multipliers can be computed, for instance by (approxi- mately) solving the least-squares problem min -k and selecting the non-positive components of the resulting vector -, or by (approxi- mately) solving the constrained least-squares problem min -k: Condition (7.1) is also appealing as it provides, in a single condition, both a stopping condition on the inner iteration and a measure of the tolerated "inexactness" in solving the associated least-squares problem, if this is the procedure chosen to obtain the dominant multipliers. We may therefore deduce from Lemma 7.1 that the convergence theory holds for Algorithms 3.1 and 7.1 whenever (7.1) is used instead of (3.3). Condition (7.1) can be further specialized. For instance, one might choose to impose the familiar "reduced gradient" criterion is an orthogonal matrix whose columns span the null space of the constraints active at x k , provided that the multipliers associated with these linear constraints are all non-positive. In this case, we have that because T the tangent cone to the set determined by the linear inequality constraints active at x k , contains T k . As a consequence, the convergence theory still holds when this criterion, which has been implemented by several subroutines for minimizing a general objective function subject to linear constraints (for example, the NAG [22], quadratic programming code E04NFF and the more general package E04UCF), is used as an inner-iteration stopping rule within Algorithms 3.1 and 7.1. This is also true for reduced gradient methods (e.g. MINOS [21], or LSNNO [24]) which compute a full column rank matrix - whose columns are generally non- orthonormal but depend upon a subset of the (finite number) of coefficients for the linear constraints. Indeed, the norm of - bounded above and away from zero, and a relationship that is a weighted form of (7.3) thus also holds in these cases. In order to preserve coherence with the framework presented in Conn et al. [8], we finally note that oe k as defined in (4.3) may also be viewed as a criticality measure. Hence we might decide to stop the inner iteration when The reader is referred to Conn et al. [9] for a proof that global convergence is still obtained for this modification of Algorithms 3.1 and 7.1. However, the authors have not been able to prove the desired local convergence properties with only (7.4). Instead, the local convergence theory is covered for Algorithms 3.1 and 7.1 for the stronger condition (see Conn et al. [10] for details). This condition is theoretically interesting, but might be practically too strong. Note, as we now show, that it implies a variant of (3.3). Theorem 7.2. Assume that fx k g, k 2 K, is a convergent subsequence of vectors of B such that (7.5) holds for each k 2 K, where the ! k converge to zero as k increases in K. Then the inequality also holds for each k 2 K sufficiently large and for some - 24 - 1. Proof. We first consider the simple case where that is when no linear inequality is present. In this case, it is easy to check from (4.3) that oe But we must have that D (\Gammar x \Phi k )k. We therefore obtain that holds with - large enough to ensure that ! k - 1. Assume now that p ? 0. The Moreau decomposition of \Gammar x \Phi k [20] is given by obviously holds for any choice of - 24 . Assume therefore that P T k nonzero. We now show that x k B, where we define kAk1 Assume first that i 2 D k . Then \Gammaa i 2 N k and a T because of the polarity of N k and T k . Since x k 2 B, we obtain that a T On the other hand, if i 62 D k , we have that a T (a T Gathering (7.8) and (7.9), we obtain that x k B, as desired. Furthermore, since kd k k - 1 by definition, we have verified that d k is feasible for the minimization problem (4.3) associated with the definition of oe k . Hence, where we have used successively the Moreau decomposition of \Gammar x \Phi k , the definition of d k and the orthogonality of the terms in the Moreau decomposition. If ffl (7.5) and (7.10) imply that sufficiently large. Otherwise, we deduce from (7.10), (7.5) and (7.7) that As a consequence of (7.11) and (7.12), we therefore obtain that (7.6) holds with Combining all cases, we conclude that (7.6) holds with this last value of - 24 . 2 We finally note that Lemma 7.1 and Theorem 7.2 do not depend on the actual form of the augmented Lagrangian (1.5), but are valid independently of the function minimized in the inner iteration. This observation could be useful if alternative techniques for augmenting the Lagrangian are considered for a merit function. 8. Conclusion. We have considered a class of augmented Lagrangian algorithms for constrained nonlinear optimization, where the linear constraints present in the problem are handled directly and where multiple penalty parameters are allowed. The algorithms in this class have the advantage that efficient techniques for handling linear constraints may be used at the inner iteration level, and also that the sparsity pattern of the Hessian of the augmented Lagrangian is independent of that of the linear constraints. The global and local convergence results available for the specific case where linear constraints reduce to simple bounds have been extended to the more general and useful context where any form of linear constraint is permitted. We finally note that the theory presented is directly relevant to practical compu- tation, as the inner iteration stopping rule (3.3) covers the type of optimality tests used in available packages for linearly constrained problems. This means that these packages can be applied to obtain an (approximate) solution of the subproblem, and constitutes a realistic and attractive algorithmic development. It is now the authors' intention to perform extensive numerical experiments on large-scale problems. This development requires considerable care and sophistication if an efficient solver for the subproblem is to be integrated with the class of algorithms described here. Acknowledgements . The authors wish to acknowledge funding provided by a NATO travel grant. They are also grateful to J. Nocedal and the anonymous referees for their constructive comments. --R Computing a search direction for large-scale linearly constrained nonlinear optimization calculations Nonlinear Programming: Analysis and Methods. Constrained Optimization and Lagrange Multiplier Methods. Convergence properties of trust region methods for linear and convex constraints. A globally convergent augmented Lagrangian algorithm for optimization with general constraints and simple bounds. LANCELOT: a Fortran package for large-scale nonlinear optimization (Release On the number of inner iterations per outer iteration of a globally convergent algorithm for optimization with general nonlinear equality constraints and simple bounds. Global convergence of a class of trust region algorithms for optimization using inexact projections on convex constraints. Global convergence of two augmented Lagrangian algorithms for optimization with a combination of general equality and linear constraints. Local convergence properties of two augmented Lagrangian algorithms for optimization with a combination of general equality and linear constraints. On the convergence of projected gradient processes to singular critical points. Introduction to sensitivity and stability analysis in nonlinear programming. Practical Methods of Optimization. Newton methods for large-scale linear equality-constrained minimization On the convergence of a sequential penalty function method for constrained minimization. Algorithmic Methods in Optimal Control. A catalogue of subroutines (release 11). Multiplier and gradient methods. D'ecomposition orthogonale d'un espace hilbertien selon deux c-ones mutuellement polaires Mark 16. A method for nonlinear constraints in minimization problems. LSNNO: a Fortran subroutine for solving large scale nonlinear network optimization problems. Indefinite systems for interior point methods. --TR
constrained optimization;augmented Lagrangian methods;convergence theory;linear constraints
588924
The Effective Energy Transformation Scheme as a Special Continuation Approach to Global Optimization with Application to Molecular Conformation.
This paper discusses a generalization of the function transformation scheme used in Coleman, Shalloway, and Wu [Comput. Optim. Appl., 2 (1993), pp. 145--170; J. Global Optim., 4 (1994), pp. 171--185] and Shalloway [Global Optimization, C. Floudas and P. Pardalos, eds., Princeton University Press, 1992, pp. 433--477; Global Optim., 2 (1992), pp. 281--311] for global energy minimization applied to the molecular conformation problem. A mathematical theory for the method as a special continuation approach to global optimization is established. We show that the method can transform a nonlinear objective function into a class of gradually deformed, but ``smoother'' or ``easier'' functions. An optimization procedure can then be successively applied to the new functions to trace their solutions back to the original function. Two types of transformation are defined: isotropic and anisotropic. We show that both transformations can be applied to a large class of nonlinear partially separable functions, including energy functions for molecular conformation. Methods to compute the transformation for these functions are given.
Introduction We are interested in solving the global minimization problem for molecular conformation, especially protein folding. How protein folds is one of the key biophysical problems of the decade. Protein folding is fundamental for almost all theoretical studies of proteins and protein-related life processes. It has many applications in the biotechnology industry, notably, structure-based drug design for the treatment of important diseases such as cancer and AIDS. Optimization approaches to the protein folding problem are based on the hypothesis that the protein native structure corresponds to the global minimum of the protein energy. The problem can be attacked computationally by minimizing the protein energy over all possible protein structures. The structure with the lowest energy is presumed to be the most stable protein structure. Mathematically, for a protein molecule of n atoms, let ng represent the molecular structure with each x i specifying the spatial position of atom i. Then the energy minimization problem for protein folding is to globally minimize a nonlinear function f(x) for all x 2 S, namely, min x2S f(x); (1) where S is the set of all possible molecular structures. The objective function f(x) is the energy function for the protein. The usual form of f(x) is is the pairwise energy function determined by the distance between atoms i and j. A widely used pairwise energy function is the Van der Waals energy function, are all physical constants (see [2]). Problem (1) is very difficult to solve in general. The reasons are as fol- lows: First, in theory even simple versions of the problem have been proved to be NP-complete [9]. Second, in practice the objective function often contains exponentially many local minimizers; therefore, search for the global minimizer can be computationally intractable. Third, the protein molecules tend to be very large, typically containing O(10,000) atoms. For such large problems, the required computation is unaffordable using general global optimization methods. However, because of its great practical importance, Problem (1) has been studied intensively in many areas of computational science and optimiza- tion. New algorithms on both sequential and parallel machines have been developed; a variety of small to medium sizes of problems have been studied [3, 4, 5, 6, 11, 12, 13, 14, 15, 17, 18, 19, 20]. In recent efforts smoothing techniques are specifically designed for molecular conformation via global minimization. Examples include the diffusion equation method [11, 14], the packet annealing method [17, 18], as well as the effective energy simulated annealing method [4, 5]. The basic idea behind these methods is to use special techniques to smooth a given energy function so that search for a global minimizer becomes more tractable. The methods usually use function transformation schemes to transform a given energy function into a class of new functions. A solution tracing procedure is then applied to the new functions to locate a solution for the original function. In this paper, we discuss an important generalization of the effective energy transformation scheme introduced in [4, 5, 17, 18]. Instead of applying the transformation to the probability distribution function, we now transform the functions directly, generalizing the method to a broader class of functions. More important, with this generalization, a mathematical theory for the transformation as a special continuation approach to global optimization is established. We show that the method can transform a nonlinear objective function into a class of gradually deformed, but "smoother" or "easier" functions. An optimization procedure can then be applied to the new functions successively, to trace their solutions back to the original func- tion. Two types of transformation are defined: isotropic and anisotropic. We show that both transformation types can be applied to a large class of nonlinear partially separable functions which includes typical energy functions for molecular conformation. Methods to compute the transformation for these functions are given. The paper is organized as follows. Section 2 introduces the basic approach and describes the function transformation method. Section 3 studies the mathematical properties of the transformation as a special continuation process. Section 4 characterizes the "smoothness" property and shows that the transformed function becomes "smoother" in the sense that the small high-frequency variations in the original function are averaged out after the transformation. The numerical applicability of the transformation is discussed in Section 5. The transformation is extended to the anisotropic type in Section 6. The formulas to compute the transformation for molecular conformation energy functions are derived. Finally, Section 7 contains concluding remarks. 2 The Approach In this section, we describe our function transformation idea which, in turn, defines our basic approach to global optimization. Suppose that we have a "poorly-behaved" nonlinear function with many local minimizers. Because of "nonsmoothness," this type of function can be very hard to minimize either locally or globally. To overcome this difficulty, we suggest using a special technique to transform the objective function into a class of gradually deformed, but "smoother" or "easier" functions. An optimization procedure can then be applied to these new functions successively, to trace their solutions back to the original function. To deform the function, we define a parametrized integral transformation as follows: Given a nonlinear function f , the transformation f is defined such that for all x, Z or equivalently; Z where - is a positive number and C - is a normalization constant such that Z Note that in contrast to the approaches in [4, 5, 17, 18], the transformation here applies directly to the given function instead of its probability distribution. This approach simplifies the transformation, and also makes it much easier to compute and analyze. To understand this transformation, consider that, given a random function distribution function p(x 0 ) for the random variable the expectation of the function g with respect to p is Z In light of (7), the transformation (4) yields a function value for any x equal to the expectation for f sampled by a Gaussian distribution function centered at x. For example, consider the following nonlinear function: which is a quadratic function augmented with a "noise" function. The transformation for this function can be computed: The function value fixed x is equal to the integration with respect to the product of two functions, the original function f(x 0 ) and the Gaussian distribution function p(x 0 Figure (a)). The parameter - determines the size of the dominant region of the Gaussian. Since the most significant part of the integration is that within the dominant region of the Gaussian, !f? - (x) can be viewed as the average value for the original function f within a small -neighborhood around x. If - is equal to zero, the transformed function is exactly the original function. For positive -, the original function variations in small regions are averaged out, and the transformed function will become "smoother" (Figure 1 (b)). Figure shows how the function behaves with increasing -. Observe that when 0:0, the function is the original function; when we increase - to 0.1, the function becomes "smoother;" when - is increased further to 0.2, the function becomes entirely "smooth." Figure 3 illustrates what the transformation implies for optimization. A standard optimization procedure, the quasi-Newton method, is applied to (a) Figure 1: A one dimensional transformation example Figure 2: A class of gradually deformed functions the three functions in Figure 2. Figure 3 (a), (b), and (c) contain the corresponding solutions x obtained with different choices of initial guesses x ffi . Although globally convergent, the method may not find the right solution if the "noise" is large. So for the function in Figure 2 (c), the method converged to the right solution only when the initial guess was close enough to the solution. When the initial guess was far from the solution, the method failed to find the right solution (Figure 3 (c)). For the function in Figure 2 (b), although it is "smoother," the behavior of the method is essentially unchanged. However, for the function in Figure 2 (a), the method always converged to the right solution (Figure 3 (a)). If we apply the procedure to the functions in Figure 2 (a) to (c) successively, and at each step take the solution for the previous function as the starting point, the solutions for all these functions can then be obtained. The experiment above suggests a general global optimization method: to optimize a difficult function, use the transformation technique to deform the function into a class of "smoother" or "easier" functions, and then apply an optimization procedure to the functions successively, to trace their solutions back to the original function. Continuation What is the difference between the suggested approach and general homotopy methods? The answer is that this approach is indeed a special type of homotopy method. But the transformation is different from conventional ho- motopies, and has the following three special features: First, the transformed functions are not arbitrarily deformed functions. They all are approximations to the original function in the sense that they are coarse estimates. Second, the transformation is defined by a special parametrized integral transforma- tion. If we increase the value of the parameter, the transformed function will become "smoother" with small variations gradually removed, but maintaining the overall function structure. Finally, if we apply an optimization procedure to a transformed function, the obtained solution usually is close to the solution for the original function. All these features are good for global optimization (also for robust local optimization), but are not necessarily the properties of conventional homotopies. We show in the following that the proposed transformation is indeed x* (a) x* (b) x* (c) Figure 3: The solutions for the functions in Figure 2 obtained by the quasi-Newton method with different initial guesses a well-defined homotopy and determines for any initial solution a unique solution curve containing the stationary points for the transformed functions. Assumption 1 The objective function f is twice continuously differentiable, and the transformation (4) is well defined for the function as well as all its derivatives. Assumption 2 Let g be the gradient of f , and \Delta the Laplace operator Then the operation \Delta can be applied to g, and the transformation (4) is well defined for all derivatives involved. Also, \Deltag(x) is uniformly bounded and satisfies a Lipschitz condition: Assumption 3 The transformation !r 2 f ? - (x) satisfies a Lipschitz condition and its inverse is uniformly bounded. Note that to guarantee Assumptions 1 to 3, a sufficient condition on f is that f and its derivatives are all integrable in terms of parametrized integration (4). We first state two sets of standard results for the proposed transformation in the following lemmas without proof. Assumption 1, 8-; x, Assumption 1, 8x, lim lim lim For convenience, we define a function h(-; x), - 2 \Gamma, and x 2 S such that is a vector space. With this definition, the condition for x to be a stationary point of x Theorem 1 and h be defined as in (18). Then under Assumptions 1 and 2, h 00 x- exists and is uniformly bounded for all and x 2 S, and also satisfies a Lipschitz condition in x: In addition, x- Proof: Let p(-; x) be the Gaussian distribution function defined as follows: -). Then by the definition of !f? - , Z By Lemma 1, x Z After differentiating (24) with respect to -, it follows that Z \Gamma- Z where Z Z It is easy to verify that Also note that Z Z Therefore, Replacing by (31), we see that By Assumption 2, ! f 000 ? - (x) is well defined and uniformly bounded. x- exists and is uniformly bounded for all f 000 (x) satisfies a Lipschitz condition by Assumption 2, satisfies a Lipschitz condition: immediately. So h 00 x- (-; x) satisfies a Lipschitz condition in x. 2 Theorem 2 Let f : R n ! R and h be defined as in (18). Then under Assumptions 1 and 2, h 00 x- exists and is uniformly bounded for all and x 2 S, and also satisfies a Lipschitz condition in x: In addition, x- \Deltag ? - Proof: Let p n (-; x) be the Gaussian distribution function where c n -) n . Then by the definition of !f ? - , Z By Lemma 1, x Z Differentiate (38) with respect to - to obtain Z where Z Z Z Z Z where Z Z From the proof of Theorem 1, Substitute (44) back into (41) to obtain Then n- 2!g? - (x) (46) and Similar to the proof of Theorem 1, it follows immediately that h 00 x- exists and is uniformly bounded for all - 2 \Gamma and x 2 S, and also satisfies a Lipschitz condition in x. 2 Finally, we state and prove the main theorem in this section as follows: Theorem 3 Let f be a function for which Assumptions 1, 2, and 3 all hold. Then for any stationary point x 0 of there is a continuous and differentiable curve x(- 2 \Gamma, such that x stationary point of !f? - . The curve x(-) is also the unique solution of the initial value problem Proof: Since x 0 is a stationary point of !f? - 0 By Assumptions 1, 2, 3, Lemmas 1, 2, and Theorem 2, function h 0 x is continuously differentiable at all (-; S. So by the Implicit Function Theorem, there is a continuously differentiable function x(-) at a neighborhood of - 0 , such that x x for all - in the neighborhood. We now show that x(-) also is defined uniquely in \Gamma. By differentiating (51), we see that x(-) is a solution to the initial value problem: xx x- which, by Lemma 1 and Theorem 2, is equivalent to the problem (48)-(49). Then it suffices to show that the right-hand side of (52) satisfies a Lipschitz condition in x on \Gamma \Theta S, which guarantees a unique solution x(-) in \Gamma by standard ordinary differential equation theory [10]. Under Assumption 3, for h 00 xx xx By Theorem 2, for h 00 x- x- Let xx x- Then it is easy to verify that G(-; x) satisfies a Lipschitz condition in x on with which completes the proof. 2 4 Smoothness In Section 2 we illustrated that the transformed functions are "smoother" than the original function in the sense that they vary slower and may even have fewer local minimizers. In the following, we characterize more rigorously the "smoothness" of the transformation. f be the Fourier transformation for function f , and d !f? - for function Recall that the transformation !f ? - for f is just a convolution of f and p, where p is the Gaussian distribution function Therefore, the Fourier transformation of ! f ? - is equal to the product of the Fourier transformations of f and p. The Fourier transformation of the Gaussian distribution function is where ! is the frequency. So, we have d We see from (62) that if - ! 0, then d converges to - f , and converges to f . This is exactly the fact we stated in Lemma 2. Also by (62), d will be very small if ! is large and - is fixed. This implies that high-frequency components of the original function become very small after the transformation. This is why the transformed function is "smoother." In addition, for larger - values, wider ranges of high-frequency components of the original function practically vanish after the transforma- tion, and therefore the transformed function becomes increasingly smooth as increases. We state these properties formally in the following theorem. Theorem 4 Let f , - all be given and well defined. fixed -, such that 8! with k!k ? ffi, f (!)j fixed -, let (1=")=-. Then 8! with ": (64)From this theorem we learn that the relative size of d can be made arbitrarily small for all ! with k!k greater than a small value ffi. Since ffi is inversely proportional to -, high-frequency components are removed when - is large. 5 Numerical Applicability The definition of the transformation (4) involves high-dimensional integration which cannot be computed in general (except perhaps by the Monte Carlo method, which is not appropriate for our purposes because it is too expen- sive). So the transformation may not be applicable to arbitrary functions, at least numerically. However, this transformation is computationally feasible for a large class of nonlinear partially separable functions, and especially to typical molecular conformation and protein-folding energy functions. We state several useful properties of the transformation in the following: First, for the sum of functions the transformation of f is equal to the sum of the transformations of the f i 's: Second, for the product of functions Y where the g i 's do not share common variables, the transformation of g is equal to the product of the transformations of the g i 's: Y Finally, for a large subclass of nonlinear partially separable functions, called the generalized multilinear functions, Y where the g i j 's are one-dimensional nonlinear functions, we have Y involves only one-dimensional integration, the transformation for a generalized multilinear function can be computed using a standard quadrature rule. In particular, let us consider a typical n-atom molecular conformation energy function, ng and h ij is the pairwise energy function determined by the distance between atoms i and j. By (66), the transformation of this energy function is equal to the sum of the transformations of the pairwise energy functions. However, the computation for the transformation still cannot be carried out directly, because there is still more than one variable in each term. Nevertheless, the following theorem provides a feasible way to compute the molecular energy transformation: Theorem 5 Let f be defined as in (71). Then the transformation of f can be computed using the formula Z Proof: We show the case when x 8i. The general case can be proved similarly. By the definition of !f? - , in form (5), for any x, Z where c - is such that Z Make the following variable transformation: Then it is easy to verify that Z -Z Z (77)The integral for involves only variable r ij and can be computed with a standard numerical integration technique; therefore, the transformation !f ? - (x) can be computed in this fashion. Note that the integral for !f ? - (x) must exist, for otherwise the transformation will have no definition. In practice, if the integral for a given f does not exist, a modified function may need to be considered instead. For example, the energy function given in (3) cannot be integrated directly because the function goes to infinity when r ij becomes very small. Usually, this can be cured by replacing the function for small r ij with finite interpolation (see [4, 11, 17]). Note also that the result in Theorem 5 applies only to energy functions that can be formulated in form (71). Most popularly used energy functions for molecular conformation and protein folding can be expressed as pairwise forms, for example, the Lennard-Jones potential, the electrostatic potential, the interaction potential for bonded atoms, etc. [2, 16]. However, some energy functions do contain terms that are not pairwise distance functions; for instance, the torsional potential usually is given as a function of the dihedral angle. Special approximation techniques may be needed to transform this type of function, We will not address this issue in this work. 6 Anisotropic vs. Isotropic The transformation we have discussed so far is of the isotropic type in the sense that it averages function variations equally along all directions in the variable space. In practice, we might wish to average different sizes of function variations along different directions (i.e., use different - values for different variables) in order to obtain a more accurate overall structure of the function. For this purpose, we can define a more general transformation, called the anisotropic transformation. Given a nonlinear function f , the anisotropic transformation !f? for f is defined such that for all x, Z or equivalently; Z where is a diagonal matrix with positive diagonal elements: and C with c - i determined such that Z Note also that in this definition, From this definition, we see that the anisotropic transformation will be reduced to the isotropic transformation when the diagonal elements of are all identical. Many of the important properties of the isotropic transformation carry over to the anisotropic case. We state these properties in the following: First, for the sum of the functions we have Second, for the product of the functions Y where the g i 's do not share common variables, we have Y where i 's are small diagonal matrices. If g i is a function of j variables positive numbers - i . Third, for the generalized multilinear functions, Y where the g i j 's are one dimensional nonlinear functions, we have Y We can also derive a simple formula to compute the anisotropic transformation for the molecular conformation energy function: Theorem 6 Let f be defined as in (71). Then the anisotropic transformation of f can be computed using the formula Z k)e Proof: We show only the case when x 8i. The general case can be proved similarly. By the definition of !f? , in form (79), for any x, Z is such that Z Make the following variable transformation: Then we have Using these relations we can verify that Z Z which completes the proof. 2 The anisotropic transform determines for any initial solution a unique solution function x(-) for the transformed functions, and therefore can also be used as a continuation process for optimization, more general and powerful than the isotropic transform. We state these results in Theorem 7 and 8. The details for the proof are quite similar to those for Theorem 2 and 3, so we will not present them. Parallel to Assumptions 1, 2 and 3 for Theorem 2 and 3, we make the following assumptions: Assumption 4 The objective function f is twice continuously differentiable, and transformation (78) is well defined for the function as well as its derivatives Assumption 5 Let g be the gradient of f , and \Psi an operator, Then the operation \Psi can be applied to g, and \Psig is a matrix with Transformation (78) is well defined for all derivatives involved in \Psig. Also, \Psig(x) is uniformly bounded and satisfies a Lipschitz condition: Assumption 6 The transformation !r 2 f ? (x) satisfies a Lipschitz condition and its inverse is uniformly bounded. Let S be a vector space, and for a positive vector - we define function h(-; x) such that 8(-; where - is the diagonal vector of , that is, Theorem 7 Let f be a given function and h be defined as in (101). Then under Assumptions 4 and 5, h 00 x- exists and is uniformly bounded for all also satisfies a Lipschitz condition in x: x- x- In addition, x- Theorem 8 Let f be a function for which Assumptions 4, 5, and 6 all hold. Then for any stationary point x 0 of there is a continuous and differentiable function x(- 2 \Gamma, such that x is a stationary point of !f ? . The function x(-) is also the unique solution of the initial value problem 7 Concluding Remarks In this paper, we have discussed a generalization of the effective energy transformation scheme used in [4, 5, 17, 18] for the global energy minimization applied to molecular conformation. Instead of applying the transformation to the probability distribution, here we transform the functions directly, generalizing the scheme in [4, 5, 17, 18] to a broader class of functions. A mathematical theory for the transformation as a special continuation approach to global optimization is established. We have established that the proposed method transforms a given nonlinear objective function into a class of gradually deformed, but "smoother" or "easier" functions. A continuation procedure can then be applied to these "smoother" or "easier" functions, to trace their solutions back to the original function. Two types of transformation are defined: isotropic and anisotropic. We have demonstrated that both transformation types can be applied to a large sub-class of nonlinear partially separable functions, and in particular, the energy functions for molecular con- formation. Methods to compute the transformation for these functions are given. We believe that the proposed method provides a powerful and effective tool for global or robust local optimization. We can see this partially from the work in [4, 5], which can be viewed as a special application of the method. In [4, 5], the transformation method, combined with simulated annealing, was applied to the global energy minimization problem for molecular confor- mation. Promising results were observed even if only simple algorithms and approximated transformation were implemented. More numerical work will be done in our future research. We will implement a group of algorithms based on the theory presented in this paper. While the transformation can now be computed with provided formulas, tracing the solution curve can be carried out using advanced numerical methods. There are at least three choices for the implementation of the tracing procedures 1. Use a general random search procedure to trace the changes of the global solution when the transformed function is gradually changed back to the original function. 2. Apply only local optimization procedures to each transformed function to trace a set of solution curves, and choose the best among all solutions obtained. 3. Solve the initial value problems for a set of solution curves, and choose the best solution. The first method is similar to the approach in [4, 5] where a simulated annealing procedure was applied to the transformed functions. This method converges to the global solution with certain probability, but a large number of random trials usually are required to obtain the convergence. The second approach is the most simple and efficient method, but the solution curves to be traced must be selected cleverly, for otherwise the global solution will not be guaranteed. The third method provides a more accurate and reliable way to trace the solution curves. As we have shown in this paper, the curves are solutions to well defined initial value problems. So standard numerical IVP-methods can be used (e.g., predictor-corrector methods) [1]. The implementation of all three tracing procedures and the numerical comparison among them will be of great interest for the further development of the algorithms. We are especially interested in applying these methods to the global energy minimization problems for molecular conformation, especially protein folding. A set of test problems will be considered including the Lennard-Jones microcluster conformation problem, the distance geometry problem, and several protein conformation problems. While searching for native structures of protein molecules is certainly very important, the proposed methods can also provide information about the paths that solutions follow. Such information may contain insights about how protein molecules change from arbitrary configurations to their native structures. Acknowledgments This research was supported partially by the Cornell Theory Center, which receives funding from members of its Corporate Research Institute, the National Science Foundation (NSF), the Advanced Research Projects Agency (ARPA), the National Institutes of Health (NIH), New York State, and IBM Corporation. The author thanks Lizhi Liao, Michael Todd, Lloyd Trefethen, and Wei Yuan for constructive suggestions. He especially thanks Thomas Coleman for many discussions relating to this work and for his helpful comments and suggestions on the manuscript, and David Shalloway for many discussions on the protein-folding problem as well as the original effective energy transformation ideas. --R Allgower and Kurt Georg Brooks III David Shalloway and Zhijun Wu David Shalloway and Zhijun Wu David Kincaid and Ward Cheney David Shalloway David Shalloway --TR --CTR Olivier Chapelle , Mingmin Chi , Alexander Zien, A continuation method for semi-supervised SVMs, Proceedings of the 23rd international conference on Machine learning, p.185-192, June 25-29, 2006, Pittsburgh, Pennsylvania Jorge J. Mor , Zhijun Wu, Distance Geometry Optimization for Protein Structures, Journal of Global Optimization, v.15 n.3, p.219-234, October 1999 Mark S. Lau , C. P. Kwong, A Smoothing Method of Global Optimization that Preserves Global Minima, Journal of Global Optimization, v.34 n.3, p.369-398, March 2006 Jack Dongarra , Ian Foster , Geoffrey Fox , William Gropp , Ken Kennedy , Linda Torczon , Andy White, References, Sourcebook of parallel computing, Morgan Kaufmann Publishers Inc., San Francisco, CA,
integral transformation;global/local minimization;continuation methods;molecular conformation
588931
An Unconstrained Convex Programming Approach to Linear Semi-Infinite Programming.
In this paper, an unconstrained convex programming dual approach for solving a class of linear semi-infinite programming problems is proposed. Both primal and dual convergence results are established under some basic assumptions. Numerical examples are also included to illustrate this approach.
Introduction . Many linear semi-infinite programming problems including the L1 and Chebychev approximation problems [14, 15] appear in the following "dual Program (D) is a compact set in R n , a(t) a are continuous functions defined on T . A corresponding "primal form" linear semi-infinite programming problem can be represented as follows. Min Z c(t)x(t)d-(t) s.t. Z a where -(t) is the Lebesgue measure, a particular regular Borel measure, on T and is measurable on T; x(t) - 0 a.e. with respect to -; and R To simplify our expressions, Eq. (1.2) will be denoted by R The duality theory relating Programs (P ) and (D) can be found in [17]. Under certain conditions, a strong duality theorem holds. According to [17], there exist three "basic" solution approaches, namely, the exchange methods, discretization methods, and methods based on local reduction. All these methods usually replace Program (D) by a sequence of finite linear programming problems for approximation and require a Department of Industrial and Operations Engineering, The University of Michigan, Ann Arbor, Michigan 48109, U.S.A.(cjlin@engin.umich.edu). y Operations Research and Industrial Engineering, North Carolina State University, Raleigh, NC 27695-7913, U.S.A. The work of this author was supported in part by the 1994 NCSC-Cray Research Grant (fang@eos.ncsu.edu). z Institute of Applied Mathematics, National Cheng-Kung University, Tainan, Taiwan, R.O.C. (soonyi@mail.ncku.edu.tw). search procedure for global optimization on T . This could be very costly, especially when T has higher dimensionality [13, 15, 16, 18]. In this paper, we propose a solution approach which identifies an optimal solution of Program (D) as a limit point of the solutions of a sequence of finite dimensional unconstrained convex programming problems. No global optimization is required in the proposed approach but instead, a multidimensional integration is required. The basic ideas are introduced in Section 2 and main results are given in Section 3. Primal convergence results are derived in Section 4 and numerical examples are given in Section 5. The last section includes some concluding remarks. 2. Basic Ideas. Following [11], given - ? 0, we perturb Program (P ) by adding an entropic term - R log x(t)d-(t) to its objective to form a perturbed problem: Min Z Z log x(t)d-(t) s.t. Z The conjugate dual [4, 21] of Program (P - ) is given as follows. Program (D - ) Z e When its optimal solution exists, we denote x - (t) as an optimal solution of Program with an optimal objective value v(P - ). Similarly, w - denotes an optimal solution of Program (D - ) with an optimal objective value v(D - ). Note that (D - ) is an unconstrained convex program which can be treated by various numerical techniques [8]. This provides an alternative approach even though the bottleneck becomes the numerical integration of a multidimensional integration over a compact set T in R n . Compared to Program (D), solving (D - ) is like a penalty function method [5] with an exponential penalty term - R penalty function methods used for semi-infinite programming problems can be referred to [6, 20]. In this paper we focus on using the exponential penalty function for linear semi-infinite programming as an extension of the methods developed for solving linear programming problems [10, 12]. In order to show that the proposed approach works, we need to address three basic issues [1]: (i) The existence of an optimal solution w - to Program (D - (ii) The existence of a compact set and a positive constant \Theta such that w lies in the compact set, 8 (iii) If a sequence fw g converges to w as - i ! 0, then w is an optimal solution of (D). In general, issues (i) and (ii) are much more difficult to deal with than the third one. In this paper, we refer to Borwein and Lewis [2] for issue (i) and resolve the other two in Section 3. Related results for the finite dimensional case can be found in [9]. Throughout this paper, two commonly seen assumptions are made: CONVEX PROGRAMMING APPROACH TO LSIP 3 Program (D) has an interior feasible solution, i.e., fw (A2) The primal constraint qualification (PCQ) holds, i.e., there exists - x in L 1 (T; is measurable on T and R with respect to the Lebesgue measure, R R and R With the assumption (A2), Theorem 4.2 of [2] resolves issue (i) and assures that both Programs (P - ) and (D - ) attain optimality with v(P - w - is optimal to Program (D - ), then x is the unique optimal solution of Program (P - ). This provides a dual-to-primal conversion formula. Also note that since the objective function of Program (D - ) is strictly concave, under the assumption (A2), its optimal solution w - can be obtained by solving the following first order condition: Z a(t)e 3. Main Results. The objective of this section is to address issues (ii) and (iii). Since issue (ii) is more difficult to handle, we start with issue (iii). Theorem 3.1. If fw is an optimal solution of Program (D). Proof. Under the assumption (A1), Program (D) has a feasible solution - with Z e Z e If w is not feasible for Program (D), then there exists - t 2 T such that a( - Since T is a compact set in R n and a j (t); c(t) are continuous on T , there exists a neighborhood N and an ffl ? 0 such that a(t) T w \Gamma c(t) - ffl; for t 2 N By using L'Hopital's rule, we have Z e Hence the right-hand-side of Eq. since Z e Z Consequently, the left-hand-side of Eq. (3.1) approaches b T - causes a contradiction. Therefore, w must be feasible for Program (D). 4 C.-J. LIN, S.-C. FANG AND S.-W. WU (Optimality) If w is not an optimal solution of Program (D), then we can find a feasible solution - w with b T - is optimal for Program (D - i ), we know Z e Z e \Gammac(t) With the same reason as we had for Eq. (3.3), as This again causes a contradiction and hence completes the proof. As a direct consequence, we have the next result. Corollary 3.2. If fw lim Z e Proof. Since Z e Z e as Z e \Gammac(t) Therefore, lim Z e To handle issue (ii), we make an additional assumption: (Bounded level set assumption) There exists a constant L such that fw j is nonempty and bounded. Note that the assumption (A3) implies that Program (D) is solvable. We shall keep using this constant L throughout the rest of the paper. Given that ffl ? 0; l ? 0, - is the Lebesgue measure on T , and - the following notations: 1. - 2. 3. 4. - Consequently, the feasible region of the original problem F 4 becomes S(0) and the assumption (A3) becomes that nonempty and bounded. Now we prove that, under some conditions, there exist \Theta ? 0 and - l ? 0 such that lies in a compact set S( - l) " B, 8 In this way, we have a convergent subsequence which goes to an optimal solution w . CONVEX PROGRAMMING APPROACH TO LSIP 5 The basic idea is to use -(w - ; ffl) as a measure of those t 0 s at which the constraint lies outside S( - l) " B, we can prove that -(w bounded above by zero, which causes a violation of the first order necessary Z a(t)e as Theorem 3.3. Under the assumption (A3), there exists - l ? 0 such that S( - is compact. Proof. It is easy to see that S(l) " B is closed, for l ? 0. Hence we only have to prove that there exists - l ? 0 such that S( - If our claim is not true, there must exist a sequence fl i g, with lim i!1 l that S(l i unbounded. Hence, we can select an unbounded sequence fw l i with for each l i . It is obvious that f w l i is a bounded sequence, thus we can find a subsequence fk i g of fl i g such that f wk i converges to a point, say - as nonempty, we can find at least one w being sufficiently small. Equivalently, for any t 2 T , As Therefore, for any fi ? 0, This contradicts the assumption (A3), under which By using the constant - l ? 0 obtained in Theorem 3.3, we can derive the following results: Lemma 3.4. Given ffl ? 0 and - l Proof. Since - is a closed subset of S( - l)"B, it is compact. With - we have there exists a sequence f - is compact, there exists a subsequence which converges to a point - with - Hence a.e. in T: Remembering that a j (t) and c(t) are continuous on T , we know a(t) T - . However, since - B, from its definition, there exists - t 2 T such 6 C.-J. LIN, S.-C. FANG AND S.-W. WU that a( - contradicts Eq. (3.8). Thus we have and the proof follows. Lemma 3.5. Given any w 0 2 the line segment w intersects - at exactly one point. Proof. Noting that w there exists at least one - t 2 T such that Note that for t 2 T with Together with (3.11), we know such - ff exists and - Note that b T (w We claim that w Otherwise there are two possible cases: (Case 1) There exists - In this case, we can find a small number fl ? 0 such that a( - This contradicts the definition of - ff. (Case 2) For all l. In this case, we can find a small number fi ? 0 such that This again contradicts the definition of - ff. Hence We now prove that w is the only point where w we know that there exists - t 2 T such that By Eq. (3.9), we have a( - Combining with Eq. (3.14), we know there exists another intersection point, say w then by the definition of - ff, we see ff ? - ff and a( - t) T (w Hence this causes a contradiction. Lemma 3.6. Given ffl ? 0 and - l Proof. Take w be the intersection point of w We can find ff ? 1 such that (3. CONVEX PROGRAMMING APPROACH TO LSIP 7 Note that Therefore for each t with By Lemma 3.4, we further have The next lemma shows that b T w sufficiently small. Lemma 3.7. There exists \Theta 1 ? 0 such that b T w - L, for Proof. By the assumption (A3), we can find a point - w that is feasible for Program (D) and b T - Z e Z e Z e Z e w is feasible for (D), lim -!0 - R there exists a sufficiently small \Theta 1 such that, for Z e This completes the proof. Combining the lemmas derived before, we now ready to prove the main result. Theorem 3.8. Under the assumptions (A1)-(A3), if there exists - that a(t) T - , then there exists \Theta ? 0 such that w - lies in a compact set Proof. Suppose that the claim is not true. Then, Lemma 3.7 implies that there exists an infinite sequence fw whose elements are not in S( - for each - i . Since Z a j (t)e \Gammac(t) we have Z w)e By Lemma 3.6, for each i, we have -(w - i 8 C.-J. LIN, S.-C. FANG AND S.-W. WU Z w)e Z \Gammac(t)-fflg Z \Gammac(t)-fflg R \Gammac(t)-fflg (a(t) T - d-(t) is bounded away from zero. Hence the right hand side of Eq. (3.19) approaches 1, as i !1. This contradicts Eq. (3.18) and completes the proof. Consequently, the issues (ii) and (iii) have been taken care by the following corollary Corollary 3.9. Under the assumptions (A1)-(A3), if there exists - that a(t) T - , then there exists \Theta ? 0 such that w - lies in a compact set sequence fw with has at least one convergent subsequence in S( - l) " B, whose limiting point is an optimal solution of Program (D). 4. Dual Unboundedness and Primal Convergence in Optimal Value. Note that (A3) is an assumption on bounded level sets. To detect the unboundedness of the dual program (D), the following lemma may help. Lemma 4.1. With the assumptions (A1) and (A2), suppose there exists - such that a(t) T - goes to infinity, as - i decreases to 0, then either the program (D) is unbounded above or its optimal solution set is unbounded. Proof. Assume our conclusion is false, then the program (D) is bounded above and its optimal solution set is nonempty and bounded. In this case, we let L be the optimal objective value which is sufficient for the assumption (A3) to hold. With the assumptions (A1) and (A2) and the existence of - Theorem 3.8 shows that the sequence fw lies in a compact set. This clearly causes a contradiction and the proof follows. We now turn our attention to investigate the primal convergence of Program in terms of the optimal objective value. It is important to point out that in our formulation of program (P ), since the primal variable where -(t) is the Lebesgue measure, a particular regular Borel measure, the program may not achieve P -attainment [2, 3, 4]. In other words, program (P ) may not have an optimal solution x (t) such that x x Z e and x (t) However, this does not affect the dual convergence results and we shall show the convergence of program (P - i ) in terms of optimal objective value. We start with the following two lemmas: Lemma 4.2. If fw such that, for each i ? CONVEX PROGRAMMING APPROACH TO LSIP 9 Proof. If the claim is not true, then there exists a particular K such that, for any lies in a compact set T , there is a subsequence which converges to - t with This contradicts the fact that Hence the proof is complete. Note that when w is optimal to Program (D - i ), Eq. (2.4) holds as Z a j (t)e m. Now, if there exists - then Eq. (4.1) implies that Z w)e \Gammac(t) and Z e This leads to the following result: Lemma 4.3. If there exists - R Combining Lemmas 4.2 and 4.3, we have the following convergence result: Theorem 4.4. If fw 0, and there exists - such that a(t) T - lim Z x (t) log x and lim Z c(t)x Proof. Given ffl ? 0, let w . Lemma 4.2 provides an N ? 0 such that, for Then from Eqs. (2.3), (4.2) and (4.3), for i ? N , we have Z x (t) log x Z Z e \Gammac(t) Therefore, lim Z x (t) log x Since Z c(t)x Z x (t) log x Z e as lim Z c(t)x This clearly shows the primal convergence of program (P - i ) in terms of the optimal objective value. 5. Numerical Examples. In this section, three numerical examples are re- ported. Our purpose is not to claim any computational superiority of the proposed method. Instead, we simply intend to illustrate the computational behavior the proposed approach. Note that the proposed approach is flexible to use any commercial or public unconstrained nonlinear optimizer, instead of developing special codes for our own implementation. In our examples, the L-BFGS-B software [22] was applied. 5.1. L1 Problems. In this subsection, the following two commonly seen L1 problems [14, 15] were tested: Problem 1 Problem 2 CONVEX PROGRAMMING APPROACH TO LSIP 11 Table Problem 1 const. comp. slackness It was reported in [14, 15] that 0.6931 and -1.78688 are approximate optimal solutions to Problem 1 and Problem2, respectively. We applied the L-BFGS-B subroutines [22] to solve Program (D - ) with the standard default setting in the "driver1.f" of the L-BFGS-B solver. The stopping criteria of L-BFGS-B was given by where f is the function to be optimized and f k , f k+1 are the values of f in the k-th and (k+1)-th iteration, respectively. Moreover, epsmch = 2:22 \Theta 10 \Gamma16 , which reflects the machine accuracy for our Silicon Graphics workstation, and factr was set to be A trivial initial solution with chosen for Problem 1. The numerical results of the proposed approach is shown by Table 5.1. In the table, "argmin \Gammac(t)g, "min \Gammac(t)g, and "comp. Z c(t)e Z c(t)x Z x (t) log x Z x The final w we obtained for problem 1 is Similarly, we chose a trivial initial solution with 2. The numerical results is shown in Table 5.2. The final w we obtained for problem 2 is The first three columns of both tables clearly show that b T w converges to the reported solution up to the 4-th digit after the decimal point, as - i decreases. The Table Problem 2 const. comp. slackness Table Two dimensional problem argmins argmin t min const. comp. slackness last two columns of both tables also show that the dual feasibility and complementary slackness conditions are satisfied up to the 4-th digit after the decimal point, when - i is small enough. The computational bottleneck of the proposed approach lies in the following in- tegration Z e Numerically, we need to discretize T for integration. When - i is getting smaller, a finer discretization of T is needed. This could be very time-consuming, especially when T has a high dimensionality. In our examples, we partition 400,000 intervals and use the Simpson's method for integration. 5.2. Two Dimensional Problem. In this subsection, the following two-dimensional problem on page 112 of [15] was tested: It was reported in [7] that 2.4356 is an approximate optimal objective value. We followed all settings as described in the previous subsection with an initial solution The numerical results are shown in Table 5.3. CONVEX PROGRAMMING APPROACH TO LSIP 13 The final w we obtained is For this two dimensional problem, we partitioned [0; 1] \Theta [0; 1] into 1500 2 intervals for integration. The computational behavior of this case is quite similar to those reported in the previous subsection. The proposed approach indeed generates a convergent solution which satisfies the dual feasibility and complementary slackness conditions. Because the numerical integration of R could be difficult for a general problem, we do not claim any computational superiority of the proposed approach. However, the proposed approach does provide a new angle to solve the linear semi-infinite programming problems. 6. Concluding Remarks. 1. We have proposed an unconstrained convex programming approach to solving linear semi-infinite programming problems. In this approach, solving Program (D - ) with a sufficiently small - ? 0 provides an approximate solution to Program (D). 2. A dual convergence result shows that, under the assumptions (A1)-(A3), (D in terms of optimal solutions. 3. A primal convergence result shows that (P - in terms of optimal objective values. 4. Compared to most known methods, the proposed approach does not require any search procedures for finding a global optimizer over T . Instead, an integration over T is required. We do not claim any computational superiority of the proposed approach. However, it does provide an alternative. For some problems, it might be easier to do integration than to perform a global optimization over T . 5. The proposed approach is flexible to use any unconstrained nonlinear opti- mizer. Three examples were included to illustrate the proposed approach. Acknowledgment :. The authors would like to thank Dr. A. R. Conn and referees for their constructive comments. --R Analysis and Methods Convergence of best entropy estimates Survey of penalty An exact penalty function for semi-infinite programming A projected Lagrangian algorithm for semi-infinite program- ming Numerical Methods for Unconstrained Optimization and Nonlinear Equations An existence theorem for penalty function theory Linear Optimization and Extensions: Theory and Algo- rithms Linear programming with entropic perturbation An inexact approach to solving linear semi-infinite programming problems An interior point algorithm for semi-infinite linear pro- gramming A dual affine scaling based algorithm for solving linear semi-infinite programming Problems The Math Works Inc. The potential method for conditional maxima in the locally compact metric spaces Convex Analysis --TR --CTR A. Ismael F. Vaz , Edite M. G. P. Fernandes , M. Paula S. F. Gomes, SIPAMPL: Semi-infinite programming with AMPL, ACM Transactions on Mathematical Software (TOMS), v.30 n.1, p.47-61, March 2004
convex programming;entropy optimization;linear programming;semi-infinite programming
588960
A Potential Reduction Newton Method for Constrained Equations.
Extending our previous work [T. Wang, R. D. C. Monteiro, and J.-S. Pang, Math. Programming, 74 (1996), pp. 159--195], this paper presents a general potential reduction Newton method for solving a constrained system of nonlinear equations. A major convergence result for the method is established. Specializations of the method to a convex semidefinite program and a monotone complementarity problem in symmetric matrices are discussed. Strengthened convergence results are established in the context of these specializations.
Introduction In the paper [11], we have introduced the problem of solving a system of nonlinear equations subject to additional constraints on the variables, i.e., a constrained system of equations. We have demonstrated that constrained equations (CEs) provide a unifying framework for the study of complementarity problems of various types, including the standard nonlinear complementarity problem and the Karush-Kuhn-Tucker system of a variational inequality. Postulating a partitioning property of the CE, we have introduced an interior point potential reduction algorithm for solving the CE and applied this method to convex programs of different kinds. The goal of this paper is to present a potential reduction Newton method for solving a CE, without assuming the existence of the partitioning property that is key to the previous work. The central problem studied in this paper is stated as follows. Let be a given mapping from the real Euclidean space ! n into itself and let\Omega be a given closed subset of ! n . The constrained equation defined by the find a vector x We refer the reader to [11] for the initial motivation to study the CE. Later in this paper, we will consider applications of our results to a semidefinite convex program and a monotone complementarity problem on the cone of positive semidefinite matrices. These applications yield new interior This work was based on research supported by the National Science Foundation under grants INT-9600343 and CCR-970048 and the Office of Naval Research under grant N00014-94-1-0340. y This work was based on research supported by the National Science Foundation under grant CCR-9213739 and by the Office of Naval Research under grant N00014-93-1-0228. point methods for solving these problems whose convergence can be established under some mild assumptions. The method proposed in this paper for solving the combines ideas from the classical damped Newton method for solving the unconstrained system of equations and the family of interior point methods for solving constrained optimization and complementarity problems. A general convergence theory for the proposed method will be presented and specializations of the results to the aforementioned applications will be described. Unlike the previous study [11] where we assume that the function H(x) has a certain partition conformal to the set\Omega\Gamma we make no such assumption herein. Instead, the present work is based on a set of broad hypotheses on the We explain some terminology and fix the notation used throughout the paper. For a given subset S of ! n , we let int S, cl S, and bd S denote, respectively, the interior, closure, and boundary of S. If the mapping H is (Fr'echet) differentiable at a point x in its domain, the Jacobian matrix of H at x is denoted H 0 (x); thus the (i; j)-entry of H 0 (x) is equal to @H i (x)=@x j , for is the Fr'echet derivative of H at x along the direction v. If H(x; y) is a function of two arguments (x; x denote the partial Jacobian matrix of H with respect to the variable x. For a real-valued function we write rOE(x) for the gradient vector of OE at the vector x . The p-norm of a vector x is denoted by kxk p ; in particular, its 2-norm or Euclidean norm is denoted by kxk. For a nonnegative vector a 2 ! n , we let [0; a] denote the line segment joining the origin and a. The set of real matrices of order n is denoted M n ; the subset of symmetric matrices in M n is denoted S n . The set M n forms a finite-dimensional inner-product vector space with the inner product given by where "tr" denotes the trace of a matrix. This inner product induces the Frobenius norm for matrices given by The subsets of S n consisting of the positive semidefinite and positive definite matrices are denoted by S n ++ respectively. For two matrices A and B in S n , we write A similarly, A OE B means ++ . For any matrix A 2 S n denotes the square root of A; i.e., A 1=2 is the unique matrix in S n such that A. Description and Analysis of the Algorithm In this section, we describe the potential reduction Newton algorithm for solving the where\Omega is a closed subset of ! n and H is a continuous mapping This section is divided into four subsections as follows: in the first subsection, we lay down the basic assumptions satisfied by the in the second subsection, we give some results which guarantee the existence of a solution for the in the third subsection, we present the detailed statement of the algorithm; in the fourth subsection, we establish a convergence theorem for the algorithm. 2.1 Basic assumptions We introduce several key assumptions on the Subsequently, these assumptions will be verified in the context of several applications of the CE. (A1) The closed set\Omega has a nonempty interior. There exists a closed convex set S ' ! n such that (a) (b) the (open) int\Omega is nonempty; (c) the set H \Gamma1 (int bd\Omega is empty. H is continuously differentiable on\Omega I , and H 0 (x) is nonsingular for all x 2\Omega I . Assumption (A1) is needed for the applicability of an interior point method. The sets S and \Omega I in assumption (A2) contain the key elements of the proposed algorithm. Notice that S pertains to the range of H and\Omega I to the domain. Initiated at a vector x 0 in\Omega I , the algorithm generates a sequence of iterates fx k g ae\Omega I so that the sequence fH(x k )g ae int S will eventually converge to zero, thus accomplishing the goal of solving the at least approximately. Assumption facilitates the application of a Newton scheme for the generation of fx k g; this scheme relies on a potential function for the set\Omega I that is induced by such a function for int S. Specifically, we postulate the existence of a potential function satisfying the following properties: (A4) for every sequence fu k g ae int S such that either lim we have lim continuously differentiable on its domain and u T rp(u) ? 0 for all nonzero u 2 int S. A condition equivalent to (A4) is stated in the following straightforward result. Condition (A4) holds if and only if for all 0, the set is compact. The notion of the central path has played a fundamental role in all interior-point methods for solving optimization and complementarity problems [2, 4, 5]. Inspired by this notion, we introduce an important assumption on the potential function p that postulates the existence of a vector satisfying a certain property; this vector will be used to define a modified Newton direction that is key to the generation of the iterates for solving the Although the vector a is inspired by the central vector of all ones in the case where S is the nonnegative orthant, since our present setting is very broad, the vector a should not be thought of as just a "central vector" for int instead, a is closely linked with the potential function p which itself is fairly loosely restricted. There exists a nonzero vector a 2 ! n and a scalar oe (a T u) (a T rp(u)) The basic role of the potential function p is to keep the sequence fH(x k )g away from the set bd S n f0g while forcing it towards the zero vector. Hence, its role is slightly different from that of a standard barrier function used in nonlinear programming, which in contrast penalizes an iterate when it gets close to any boundary point of S. Later, we will identify this function for various sets S in the applications to semidefinite problems. For now, we will consider the simple case where S is the nonnegative orthant ! n and establish the validity of conditions (A4)-(A6) for the function log is an arbitrary scalar. (Note: the ' 1 norm of u, instead of u T u, could also be used in the first logarithmic term. The analysis remains the same with the constant i properly adjusted.) Clearly, p is norm-coercive on ! n lim because for u ? 0, log log n log where the first and second inequalities follow from the fact that kuk 1 p nkuk and n log( log u i n log n, respectively. Moreover, for any positive sequence fu k g converging to a nonzero nonnegative vector with at least one zero component, the limit (1) clearly holds. Thus holds. Moreover, with a taken to be the vector of all ones, (A6) also holds. Indeed, we have for a thus (a T rp(u)) (a T u) where the last inequality follows from the fact that kuk 1 p n kuk and the arithmetic-geometric mean inequality. Other choices for the function p exist for . The above choice will be generalized to the case where S involves the cone of symmetric positive semidefinite matrices. 2.2 Existence of solutions In this subsection, we study conditions that guarantee the existence of solutions of the We start by giving a few definitions. Assume that M and N are two metric spaces and that N is a map between these two spaces. For Eg. The map G is said to be proper with respect to a set is compact for every compact set K ' E. If G is proper with respect to N , we will simply say that G is proper. For D ' M , and E ' N such that G(D) ' E, the restricted defined by ~ denoted by Gj (D;E) ; if we write this ~ G simply as GjD . We will also refer to Gj (D;E) as "G restricted to the pair (D; E)", and to GjD as "G restricted to D". We say that partition of the set V if space M is said to be connected if there exists no partition which both O 1 and O 2 are non-empty and open. A metric space M is said to be path-connected if for any two points there exists a continuous that The following result and its proof can be found in Monteiro and Pang [7] (see Corollary 1 of this reference). Proposition 1 Let M and N be two metric spaces and N be a continuous map. Let M 0 ' M and N 0 ' N be given sets satisfying the following conditions: is a local homeomorphism, Assume that F is proper with respect to some set E such that N 0 ' E ' N . Then F restricted to the pair (M local homeomorphism. If, in addition, N 0 is connected, then F cl N 0 . Using Proposition 1, we now derive two existence results for the Theorem 1 Assume that conditions (A1)-(A3) hold and that there exists a convex set E ae S such that I ) is nonempty and H proper with respect to E. Then, . In particular, solution. Proof. To apply Proposition 1, let j\Omega I , N Using (A2), we easily see that F (M by (A3) and the inverse function theorem, it follows that F j M 0 is a local homeomorphism. Since with respect to E by assumption, it follows from Proposition 1 that cl cl where the last equality follows from the fact that cl E) " cl (int by elementary properties of convex sets (see subsection 2.1 in Chapter 3 of [1]). Moreover, it also follows from Proposition 1 that restricted to the int S) is a proper local homeomorphism. Theorem 2 Assume that conditions (A1)-(A3) hold and that F is proper with respect to S. Then restricted to\Omega I maps each path connected component of\Omega I homeomorphically onto int S. In particular, solution. Proof. Conclusion (i) follows immediately from Theorem 1 with S. Using the last conclusion obtained in the proof of Theorem 1 and setting we conclude that F restricted to the pair (\Omega I ; int S) is a proper local homeomorphism. If T '\Omega I is a path connected component of\Omega I then F restricted to the pair (T ; int S) is a proper local homeomorphism since T is both open and closed with respect to\Omega I . Since every proper local homeomorphism from a path connected set into a convex set is a homeomorphism (see for example Theorem 1 of [7]), (ii) follows. 2.3 The algorithm The algorithm for solving the damped Newton method applied to the Referring the reader to [8] for the basic family of Newton methods for solving this unconstrained equation, we highlight the modifications to deal with the presence of the constraint In essence, there are two major modifications. One, the Newton equation to compute the search directions is modified using the (central) vector a in assumption (A6). Two, the merit function for the line searches is based on the merit function: This is different from the norm functions of H that are the common merit functions used in a classical damped Newton method. Note that by (A3) and (A5) the function / is continuously differentiable on\Omega I . With the above explanation, we now give the full details of the promised Newton method for solving the under the setting given in the last subsection. Step 0. (Initialization) Let a vector x 0 2\Omega I and scalars ae 2 (0; 1) and ff 2 (0; 1) be given. Let a sequence of scalars foe k g ae [0; oe) be also given. (The scalar oe is as given in assumption (A6).) Set the iteration counter Step 1. (Computing the modified Newton direction) Solve the system of linear equations a a (3) to obtain the search direction d k . Step 2. (Armijo line search) Let m k be the smallest nonnegative integer m such that x k +ae m d k 2\Omega I and Step 3. (Termination test) If prescribed tolerance; stop; accept x k+1 as an approximate solution of the Otherwise, return to Step 1 with k replaced by k + 1. By (A3) and the fact that x k 2\Omega I , the Newton equation (3) has a unique solution which we have denoted by d k . The following lemma guarantees that d k is a descent direction for the function / at x k . This property, along with the openness of\Omega I , ensures that the integer m k can be determined in a finite number of trials (starting with increasing it by one at each thus guaranteeing the well-definedness of the next iterate x k+1 . Lemma 2 Suppose that conditions (A5) and (A6) hold. Assume also that x 2\Omega I , d are such that a T H(x) where a are as in condition (A6). Then, r/(x) T d ! 0. Proof. Let due to (4) and the assumption that x 2\Omega I . This together with (2), (5), (4), (A5) and (A6) imply a T u a oe as claimed. 2.4 A convergence result In what follows, we state and prove a limiting property of an infinite sequence of iterates fx k g produced by the algorithm. Before stating the theorem, we observe that such a sequence necessarily belongs to the set\Omega I ; thus fH(x k )g ae int S. Since the sequence fx k g is infinite, we have for all k. Theorem 3 Assume conditions (A1)-(A6) hold and that lim sup k oe k ! oe. Let fx k g be any infinite sequence produced by the potential reduction Newton algorithm. Then, the following statements hold: (a) the sequence fH(x k )g is bounded; (b) any accumulation point of fx k g, if it exists, solves the in particular, if fx k g is bounded then the solution. Moreover, for any closed subset E of S containing the sequence fH(x k )g, (c) if H is proper with respect to (d) if H is proper with respect to E, then fx k g is bounded. Proof. Let all k. Hence, for any " ? 0 we have fu k g ae "g. Since by Lemma 1 the set ("; fl) is compact, and hence bounded, we conclude that fu k g is bounded. Hence, (a) follows. To show (b), let x 1 be an accumulation point of fx k g. Clearly x 1 2\Omega because\Omega is a closed set. Assume for contradiction that be a subsequence converging to x 1 and assume without loss of generality that foe converges to some scalar oe 1 . Since oe k 0 for all k and lim sup k oe k ! oe, we must have oe 1 2 [0; oe). Since p(u k ) and lim there exists " ? 0 such that the subsequence fu ae ("; fl). Since by Lemma 1 the set ("; fl) is compact, we conclude that u int S, and hence that x 1 2 assumption (A2), it follows that x 1 2\Omega I . Hence, by assumption (A3), H 0 exists. This implies that the sequence fd converges to a vector d 1 satisfying a a: Hence, it follows from Lemma 2 that r/(x 1 converges to x 1 2\Omega I where / is continuous, it follows that f/(x k converges. This implies that the whole sequence f/(x k )g converges due to the fact that it is monotonically decreasing. Using the relation for all k, we conclude that lim and hence that lim because lim Thus lim which implies that m k 2 for all k 2 sufficiently large. Consequently, by the definition of m k , we deduce that for all k 2 sufficiently large. Letting k 2 tend to infinity in the above expression, we obtain which contradicts the fact that ff ! 1 and r/(x 1 Consequently, we must have 0, and hence (b) follows. Assume now that E is a closed subset of S containing the sequence fH(x k )g. To prove (c), assume for contradiction that for an infinite subset ae f0; lim inf By an argument similar to that employed above, we conclude that for some " ? 0 we have and the fact that E is closed, we conclude that ("; a compact subset of int S " E. Since H is proper with respect to int S " E, the inverse image of is compact, and hence bounded. This implies that fx is bounded. By (b), every accumulation point of the latter subsequence is a zero of H . This contradiction establishes (c). Finally, using (a) and the fact that E is closed, we conclude that fu k g is contained in a compact subset E 1 of E. Since H is proper with respect to E, it follows that the set H is bounded. Hence, (d) follows. Statements (a), (b) and (c) of Theorem 3 do not claim the boundedness of the sequence fx k g. In particular, existence of a solution to the established only under the properness condition of statement (d). A consequence of statement (c) is consequently, in the sense that for any such ", there exists a vector x " can be computed by the potential reduction Newton method starting at the given vector x 0 . The framework of the that we have set forth so far is very broad. In addition to not assuming any sign restriction on the components of H (like we did in [11]; see Assumption 1 therein), part of the generality of the present framework stems from the freedom in the choice of the set S and the associated potential function p. Indeed, as we shall see in the special cases below, the set S and the function p can often be constructed under very mild assumptions. 3 Monotone Complementarity Problems in Symmetric Matrices We consider a mixed complementarity problem defined on the cone of symmetric positive semidefinite matrices. The linear version of this problem was introduced by Kojima, Shindoh, and Hara [3] and has received a great deal of research attention recently. In what follows, we consider a non-linear version of this problem defined in [6]. This reference contains a fairly extensive bibliography on interior point methods for solving optimization and complementarity problems defined on the cone of semidefinite matrices; it will be the source for several results that will be used freely in the subsequent development. 3.1 Implicit mixed complementarity problems We recall the framework considered in [6]. Let F : S n be a given mapping. The mixed complementarity problem in symmetric matrices is to find a triple (X; Y; z) 2 S n \Theta S n \Theta ! m satisfying F (X; Y; \Theta S n As explained in [6] and the references therein, there are several equivalent ways of stating the complementarity condition each leading to a different interior point method for solving the above problem. In what follows, we consider the equivalent formulation of this problem as the CE defined by the where the set\Omega and the \Theta S n are defined by F (X; Y; z) Similar treatment can be applied to other equivalent formulations and to generalizations of the basic problem (6). Throughout the following discussion, F is assumed to be continuous on its domain and continuously differentiable on S n ++ \Theta S n Associated with the above mapping H , define the set \Theta S n g: It has been shown in Lemma 1 of [6] that \Theta S n g: The fundamental role of the set U in the study of the problem (6) is well explained in this reference. This set continues to have an important role in the present algorithmic setting for solving the cone complementarity problem. We introduce an important assumption on the mapping F that will be used to verify the nonsingularity of the Jacobian matrix H 0 (X; Y; z). (B1) The mapping F is (X; Y )-differentiably-monotone at every triple (X; Y; z) 2 U \Theta ! m ; i.e., for any such triple, (B2) The mapping F is z-differentiably-injective at every triple (X; Y; z) 2 U \Theta ! m ; i.e., for any such triple, The following lemma asserts that the basic assumptions (A1)-(A3) in Subsection 2.1 are valid under the above hypotheses. Lemma 3 Consider the with\Omega and H defined by (7) and (8), and let S j S n If conditions (B1) and (B2) hold, then \Omega I j moreover, the and the set S satisfy conditions (A1), (A2) and (A3). Proof. Only the second assertion requires a proof. Conditions (A1) and (A2)(a) obviously hold. Clearly U is an open set; since (I ; I) 2 U , (A2)(b) holds. Moreover, it is easy to see that the alternative representation (10) implies (A2)(c). Next we establish that (A3) holds under (B1) and (B2). This amounts to showing that for every (X; Y; z) the following implication holds: Assume the left-hand condition holds. Then, Condition (B1) and (14) imply that dX ffl dY 0. This together with (13) and the fact that (see the proof of Theorem 3.1(iii) of [10]). In turn, this together with imply which yields due to (B2). Next we deal with conditions (A4)-(A6). For this purpose, consider the potential function ++ \Theta S n \Theta defined by ++ \Theta S n is an arbitrary constant. Lemma 4 The potential function (15), the vector a j and the scalar conditions (A4), (A5) and (A6). Proof. Since for a matrix Z 2 S n , kZk 2 F is equal to the sum of the squares of the n eigenvalues of Z, and det Z is equal to the product of these eigenvalues, the verification of (A4) for the function p(M; N; v) is the same as in the previous case of a nonnegatively constrained equation (discussed at the end of Subsection 2.1). Noting that we have and thus (A5) holds. We now show that (A6) is satisfied with the given a and oe. Indeed we have which implies Noting that (i) tr(M) equals the sum of the eigenvalues of M , (ii) tr(M the sum of the inverses of the same eigenvalues, and (iii) kMk 2 the sum of these eigenvalues squared, it follows from the same derivation as in the end of Subsection 2.1 that condition (A6) holds. According to (2), the potential function (15) induces the following merit function on the set log det for any triple (X; Y; z) 2 U \Theta ! m . Here, k \Delta k F;2 denotes the norm on S n \Theta ! m defined by We now give a detailed description of a specialized algorithm for solving the mixed complementarity problem in symmetric matrices (6), based on the potential reduction Newton method for solving the oe defined as in (7), (8), Lemma 3, (15) and Lemma 4, respectively. Step 0. (Initialization) Let a pair of matrices (X and ff 2 (0; 1) be given. Let a sequence of scalars foe k g be also given, where oe k 2 [0; 1) for all k. Set the iteration counter Step 1. (Computing the modified Newton direction) Solve the system of linear A to obtain the search triple (dX Step 2. (Armijo line search) Let m k be the smallest nonnegative integer m such and Step 3. (Termination test) If prescribed tolerance; stop; accept the triple (X as an approximate solution of the problem (6). Otherwise, return to Step 1 with k replaced by k + 1. As an immediate consequence of Lemma 3, Lemma 4 and Theorem 3, we have the following convergence result for the above algorithm. Theorem 4 Assume that conditions (B1) and (B2) hold and lim sup k oe k ! 1. Let f(X be any infinite sequence produced by the above algorithm for solving problem (6). Then, the following statements hold: (a) the sequence fH(X (b) any accumulation point of exists, solves the problem (6); in particular, if bounded then problem (6) has a solution. We now make a few remarks. The above theorem guarantees neither that f(X k bounded nor that it has an accumulation point. The conclusion that f(X k ; Y k ; z k )g is bounded would follow from Theorem 3(d) with could prove that the map H is proper with respect to the set S j S n . Unfortunately, this requirement is rather strong. For monotone mixed complementarity problems, we state in Proposition 2 below a result (from Monteiro and Pang [6, Lemma 2]) asserting that the map H is proper with respect to S n \Theta F (U \Theta ! m ). Hence, if the latter set contains the set equivalently if the equality F (U \Theta ! m then the sequence generated by the above algorithm f(X k ; Y k ; z k )g is bounded. Intuitively, the equality F (U \Theta ! m hold for maps F satisfying some kind of strong monotonicity condition. But since this type of condition is fairly restrictive, we do not pursue this issue any further. Another possible approach which would guarantee the boundedness of f(X is to reduce the set S so as to have S ' S n \Theta F (U \Theta ! m ). This approach requires some knowledge of the set F (U \Theta ! m ). We will see that for the complementarity problems studied in Subsection 3.2 and Section 4, enough information about the set F (U \Theta ! m ) is available which allows us to choose a set S together with a potential function satisfying the inclusion S ' S n \Theta F (U \Theta ! m ) and the conditions (A1)-(A6) of Subsection 2.1. Before stating the properness result mentioned above, we give a few basic definitions. mapping J(X; Y; z) defined on a subset dom(J) of M n \Theta M n \Theta ! m is said to be Y )-equilevel-monotone on a subset V ' dom(J) if for any (X; Y; z) 2 V and (X such that F (X; Y; will simply say that J is (X; Y )-equilevel-monotone. In the following two definitions, we assume that W , Z and N are three normed spaces and that OE(w; z) is a function defined on a subset of W \Theta Z with values in N . 2 The function OE(w; z) is said to be z-bounded on a subset V ' dom(OE) if for every sequence f(w k ; z k )g ae V such that fw k g and fOE(w k ; z k )g are bounded, the sequence fz k g is also bounded. When dom(OE), we will simply say that OE is z-bounded. Definition 3 The function OE(w; z) is said to be z-injective on a subset V ' dom(OE) if the following implication holds: (w; we will simply say that OE is z-injective. The following is the promised result from Lemma 2 of Monteiro and Pang [6]. m be a continuous map and let H m be the map defined by (8). Assume that the map F is (X; Y )-equilevel-monotone and z-bounded on its domain. If the map H restricted to U \Theta ! m is a local homeomorphism, then H is proper with respect to S n \Theta F (U \Theta ! m ). 3.2 Standard complementarity problem In this section, we consider the standard nonlinear complementarity problem (NCP) in symmetric matrices: is a given continuous mapping which is continuously differentiable on S n ++ . This problem is a special case of the implicit mixed complementarity problem of Subsection 3.1 z is not present) and F : S n \Theta S n \Theta S n We make the following assumption on the mapping f . is monotone on S n Lemma 5 If condition (C1) holds then the \Theta S n defined by (17) satisfies condition (B1) of Subsection 3.1. Proof. By (C1), it follows that for every X 2 S n , the linear map f 0 (X) is monotone in the sense that To verify (B1), assume that (dX; dY equivalently by (18), we have This shows that implication (11) holds for since implication (12) holds vacuously for It is possible to solve the NCP (16) with the use of the potential reduction algorithm described in Subsection 3.1. However, the sequence of iterates generated by this algorithm might not be bounded. We now develop a different potential reduction algorithm in which the set S is reduced so as to have S ' S n \Theta F (U ), thus ensuring the boundedness of the sequence f(X (see the discussion at the end of the previous subsection). To describe the alternative algorithm, it is sufficient to identify the the set S, the potential function and the vector a and scalar oe in condition (A6). We let \Theta S n \Theta S n where F is given by (17). Moreover, we let S j S n \Theta S n be defined by F \Theta S n is an arbitrary constant. Finally, we let a j (I ; I) and oe j 1. Clearly, the set\Omega I and the merit function / and log det Lemma 6 The the set S, the potential function !, the vector a and the scalar oe defined above satisfy conditions (A1)-(A6) of Subsection 2.1. Proof. Condition (A2)(b) follows from the fact that 2\Omega I for all sufficiently large. The other conditions are either straightforward or are shown using Lemma 5 and the same arguments used in the proofs of Lemma 3 and Lemma 4. Before giving the convergence result for the potential reduction Newton method in the above framework, we state the following result which will be used to establish boundedness of the iterates generated by this method. Lemma 7 Suppose that f : S n is a continuous map which is continuously differentiable on S n ++ and satisfies condition (C1). Then, for the maps F and H defined by (17) and (19) respectively, we have: (a) F \Theta S n proper with respect to S n \Theta S n (c) if 0 2 F (S n \Theta S n proper with respect to S n \Theta S n Proof. By Proposition 4(a) and Corollary 3 of [6] with \Theta S n \Theta S n ). Using this inclusion, we easily see that statement (a) holds. We next show (b). By Lemma 6, H 0 (X; Y ) is invertible for all (X; Y restricted to U is a local homeomorphism. Thus it follows from Lemma 2 that H is proper with respect to once we prove that S n ++ \Theta S n ++ be arbitrary. \Theta S n \Theta S n such that ~ X). For ffl ? 0, let X). Clearly, X ffl 0 for every ffl ? 0. By the continuity of f and the fact that U Y 0, we have Y ffl 0 for ffl ? 0 sufficiently small. Since that U belongs to F (S n \Theta S n We omit the proof of (c) which is similar to that of (b). We will skip the straightforward formulation of the potential reduction Newton method specialized to the above choices of the potential function and scalar oe; instead we directly give the convergence properties of the method. Theorem 5 be a continuous function which is continuously differentiable on ++ and satisfies condition (C1). Suppose that f(X k ; Y k )g is a sequence generated by the potential reduction Newton method with the potential function and scalar oe as specified above. Then, the following statements hold: (a) every accumulation point of f(X k ; Y k )g is a solution of the NCP (16); (b) if there exists ~ such that f( ~ (c) if there exists " ++ such that f( " , then the sequence f(X k ; Y k )g is bounded. Proof. Statement (a) follows from Theorem 3(b). To prove statement (b), note first that the assumption implies that 0 2 F (S n \Theta S n Hence, by Lemma 7(b), we conclude that H is proper with respect to S n \Theta S n ++ . It follows from Theorem 3(c) with to zero. The proof of (c) follows similarly from Lemma 7(c) and Theorem 3(d) with Statement (a) is within expectation; statement (b) is interesting because its assumption is the feasibility of the NCP in symmetric matrices (16). A consequence of of statement (b) is that feasibility of this problem (which is also monotone by assumption (C1)) is sufficient for the sequence to converge to zero although no boundedness of the sequence f(X k ; Y k )g is asserted. The latter assertion is established under the strict feasibility of the problem (16); this is statement (c). 4 Convex Semidefinite Programs In this section we consider the convex semidefinite program studied in [6, 9], namely: minimize '(x) subject to G(x) 0 are given smooth mappings. Under a suitable constraint qualification, if x is a locally optimal solution of the semidefinite program, then there must exist (j ; U such that is the Lagrangian function defined by Clearly, the first-order optimality condition (21) and the feasibility of x is equivalent to the implicitly mixed complementarity problem (6) in which the map F : S n \Theta S n is defined by \Theta S n and the following correspondence of variables are made: (U; V ) Hence, as in Subsection 3.1, the feasibility of x and the first-order optimality condition (21) can be formulated as the set\Omega and the \Theta S n are defined by \Theta S n \Theta S n Our goal is to solve the by the potential reduction Newton method. For this purpose, we make the following blanket assumptions on problem (20): (D1) the objective function continuously differentiable and convex; (D2) the map G continuously differentiable and positive semidefinite convex (psd-convex), that is (D3) the map affine, and the (constant) gradients frh j (x)g p are linearly independent; (D4) for every (x; U; the following implication holds: (D5) the feasible set is nonempty and bounded. We propose below a new interior point method for solving the convex semidefinite program (20) based on the potential reduction Newton algorithm of Subsection 2.3. This method not only generalizes the algorithm developed in Section 4.2 of [11] to the context of the nonlinear semidefinite programming problem but it also allows for more general choice of starting points. The new algorithm uses a novel potential function / which depends on the starting point. A key advantage of the new algorithm is that strong convergence properties can be established for arbitrary starting points. This differs from the results in [11] which either require the starting point to satisfy the linear equality constraint (Theorem 5 in the reference) or do not guarantee the boundedness of the sequence of multipliers (Theorem 4 in the reference). p+m denote an arbitrary starting point and let c 0 j h(x 0 ) and any matrix such that \Theta S n Note that S depends on the starting point when The following technical lemma is a partial restatement of Lemma 6 of [6] and is used in the subsequent Lemma 9 to establish that the and the set S defined above satisfy conditions (A1)-(A3) of Subsection 2.1. Lemma 8 Assume that G is an affine function. Then the following statements hold: (a) for every U 2 S n , the function x (b) if condition (D5) holds then, for every !, the set is bounded. Lemma 9 Assume that problem (20) satisfies conditions (D1)-(D4). The following three statements hold: (a) the map F defined by (23) satisfies (B1) and (B2) of Subsection 3.1; (b) the with\Omega and H defined by (25) and (24), respectively, and the set S defined by (26), satisfy conditions (A1), (A2) and (A3) of Subsection 2.1; and (c) the map H restricted to the set U \Theta ! p+m is a local homeomorphism. Proof. Since the case where c is easy to deal with, the proof below focuses on the case where Conditions (A1) and (A2)(a) are obvious. Clearly, we have which is nonempty because it contains the tuple (U using (10) we easily see that the set H \Gamma1 (int bd\Omega is empty. We have thus proved that condition (A2) holds. Using the same arguments as in the proof of Lemma 3, we can show that if statement (a) holds, then is nonsingular for every (U; V; j; x) 2 U \Theta ! p+m ; in particular, we can conclude that holds due to (27), and that H restricted to the set U \Theta ! p+m is a local homeomorphism by the inverse function theorem. Thus the remaining proof is to show that F satisfies (B1) and (B2). For this purpose, assume that (U; V; x; for some (dU; dV; dx; dj) 2 S n \Theta S n \Theta ! p+m , or equivalently Lemma 8(a) together with conditions (D1), (D2) and (D3) and the fact that U 0 imply that L(x; U; j) is a convex function of x. Hence, we have that dx T L 00 xx (x; U; j)dx 0. Multiplying (29) on the left by dx T and using this last observation together with (28) and (30), we obtain xx (x; U; j)dx 0: (31) Thus F satisfies (B1). Assume now that Then all the relations above hold with (dU; dV In particular, (28), (30) and (31) imply that h 0 (x; Hence, we conclude that to (D4). Using this and the fact that relation (29) hold with which in turn implies that due to (D3). We have thus shown that F satisfies (B2). Associated with the set S, we now introduce the following potential function defined for any tuple (A; B; c; d) 2 int S by det where i is a suitable constant. We establish in the next result that if i 3n=2, then the above potential function satisfies conditions (A4), (A5) and (A6) of Subsection 2.1. then the potential function (32), the tuple a and the constant conditions (A4), (A5) and (A6) of Subsection 2.1. Proof. The verification of (A4) is similar to the one of Lemma 4. Define ~ It is easy to see that ~ ~ ~ The definition of and ~ together with a simple algebraic manipulation reveals that and hence that (A5) holds. Moreover, using the fact that F and (trP for every P 2 S n and i 3n=2, we obtain for every (A; B; c; d) 2 int S, 2i F Hence (A6) holds with The next two results will be used in Theorem 3 to establish the boundedness of the sequence of iterates generated by the potential reduction Newton method under the framework of this section. Lemma 11 Assume that problem (20) satisfies conditions (D1)-(D5). Then the \Theta defined in (25) is proper with respect to the set S n \Theta F (U \Theta ! p+m ). Proof. Using Proposition 4(a) and Lemma 7 of [6], we conclude that the map F defined in (23) is )-equilevel monotone on S n \Theta S n \Theta ! p+m . Moreover, by Proposition 4(c) and Lemma 9 of [6], it follows that F is (j; x)-bounded on S n p+m . Since by Lemma 9 the map H restricted to U \Theta ! m+p is a local homeomorphism, we conclude from Proposition 2 that H is proper with respect to S n \Theta F (U \Theta ! p+m ). In the next result we describe in more detail the set F (U \Theta ! p+m ) for the map F given by (23). Lemma 12 Assume that problem (20) satisfies conditions (D1)-(D5). Then F (U \Theta ! p+m is the map given by (23) and Moreover, F is a convex set. Proof. The inclusion F (U \Theta ! p+m straightforwardly from the definition of the map F and the set U . Assume now that (B; c; d) 2 F \Theta ! m . We have proved in Lemma 10 of [6] that if conditions (D1)-(D5) holds and (0; Consider now the problem minimize ~ subject to ~ where ~ . It is easy to see that the functions ~ G and ~ h also satisfies conditions (D1)-(D5). Hence, applying Lemma of [6] to this new problem, we conclude that (0; 0; F is defined like the function F in (23) with ', G and h replaced by ~ G and ~ h, respectively. A simple verification shows that (0; 0; F (U \Theta ! p+m ) is equivalent to (B; c; d) 2 F (U \Theta ! p+m ). We have thus shown that F (U \Theta ! p+m Using conditions (D2) and (D3), and some standard arguments, we can easily show that F is a convex set. We establish one technical lemma which will be used to prove an important conclusion of the main result of this section, Theorem 6. Lemma 13 Let fU k g and fV k g be two sequences in S n ++ such that lim Then lim Proof. Since its eigenvalues are all real. Since it follows that all the eigenvalues of U k V k are real too. This implies that the eigenvalues of (U k are all positive. Therefore, Since the right-hand norm converges to zero as k ! 1, the same holds for the left-hand norm. Thus the spectrum of U k V k converges to the single element f0g. Since this spectrum is the same as that of (U k the desired limit (33) follows. The following is the main convergence result of the potential reduction Newton method specialized to the convex semidefinite program (20). Theorem 6 Suppose that problem (20) satisfies conditions (D1)-(D5), and that f(U is a sequence generated by the potential reduction Newton method of Subsection 2.3 initialized at an arbitrary tuple (U (25), (26) and (32), respectively, a j Assume also that 1=2. Then, the following statements hold: (a) every accumulation point of f(U is a solution of the (b) the sequence f(V k ; x k )g is bounded; thus fx k g has at least one accumulation point; (c) lim k!1 H(U (d) every accumulation point of the sequence fx k g is an optimal solution of problem (20); (e) if there exists that is problem (20) has a Slater point, then the whole sequence f(U Proof. By Lemmas 9 and 10, the assumptions of the theorem guarantee conditions (A1)-(A6) of Subsection 2.1. Hence, by Theorem 3, we conclude that statement (a) holds and that the sequence fH(U bounded. By the definition of H , this implies that fV k are bounded, and hence fx k g ae fx 8(b) the latter set is bounded, we conclude that fx k g is bounded. Clearly, this and the fact that bounded imply that fV k g is also bounded. Hence, statement (b) follows. The proofs of statements (c) and (e) are based on statements (c) and (d) of Theorem 3. For simplicity, we assume in the remaining proof that c 0 j h(x 0 ) 6= 0; the proof when c analogous. Define \Theta Note that E is a closed subset of S. Moreover, using (D3) and the fact that the third component of a is zero, we easily see that fh(x k )g ae [0; c 0 ]. Clearly, this implies that fH(U In view of (c) and (d) Theorem 3, statement (c) and (e) follow once we establish that the map H is proper with respect to \Theta and also proper with respect to E under the assumption that (0; We prove first the properness assertion with respect to int S " E. By Lemmas 11 and 12, we know that H is proper with respect to S n \Theta F (U \Theta ! p+m Hence, it suffices to show that int S " E is contained in S n \Theta F \Theta ! m , or equivalently that Using the definition of F and Lemma 8(b), it is easy to see that cl Moreover, it follows immediately from the definition of F and (35) that cl F Let (B; c) be an arbitrary element of the left-hand set in (34). Since c 2 [0; c 0 ], we have some t 2 [0; 1]. Hence, Since (0; cl F by (D5), by (26), and cl F is a convex set due to Lemma 12 and Proposition III.1.2.7 of [1], we conclude that (tG cl F . Hence, by (37) and (38), we have (B; c) = (B; holds. Assume now that (0; To prove the properness assertion with respect to E, it suffices to show that E ' S n \Theta F \Theta ! m , or equivalently If (B; c) is in the left-hand set then we have by (26) and F is convex by Lemma 12, we conclude that Hence, by (36) and the fact that B tG 0 , we have (B; c) = (B; holds. Finally, we prove statement (d). For each k, let It follows that x k is an optimal solution of the convex program due to the fact that x k together with the multiplier pair (U the optimality condition for this problem. Now let x 1 be an arbitrary accumulation point of fx k g. Clearly, x 1 is a feasible solution of (20) due to Theorem 6(c). To show the global optimality of x 1 , assume that ~ x is an arbitrary feasible solution of (20). Let t k 2 [0; 1] be such that h(x k ~ is feasible to (40). Since ft k g converges to zero, it follows that converges to ~ x. Moreover, since H(U by the definition of S (26), we have for each k (cf. (38)), Hence, it follows that log det for all k. Rearranging this inequality, we obtain log det log det I log det log det where the last inequality follows from the fact that Hence, as k goes to 1, we may invoke Lemma 13 to conclude that We have thus proved that x 1 is an optimal solution of (20). Remark. The significance of part (d) of Theorem 6 is that it does not require the sequence of multipliers to be unbounded. Assuming that G 0 0, it is possible to show that the potential function (32), a j and the inequality in condition (A6) for every (A; B; c; d) in the set E is defined in the proof of Theorem 6. Using this fact, it is possible to establish a convergence result similar to Theorem 6 for a j 1. The crucial point to note is that Theorem 3 still holds if we assume the inequality in condition (A6) to be valid only for points in the sequence )g. Details are omitted. --R Convex Analysis and Minimization Algorithms I A unified approach to interior point algorithms for linear complementarity problems The complementarity problem for maximal monotone multifunctions Pathways to the optimal set in linear programming On two interior-point mappings for nonlinear semidefinite complementarity problems Iterative Solution of Nonlinear Equations in Several Variables First and second order analysis of nonlinear semidefinite programs Existence of search directions in interior-point algorithms for the SDP and monotone SDLCP programs An interior point potential reduction method for constrained equations --TR --CTR Samuel Burer , Renato D. C. Monteiro , Yin Zhang, Interior-Point Algorithms for Semidefinite Programming Based on a Nonlinear Formulation, Computational Optimization and Applications, v.22 n.1, p.49-79, April 2002 Tong , Liqun Qi , Yu-Fei Yang, The Lagrangian Globalization Method for Nonsmooth Constrained Equations, Computational Optimization and Applications, v.33 n.1, p.89-109, January 2006 Christian Kanzow , Nobuo Yamashita , Masao Fukushima, Levenberg-Marquardt methods with strong local convergence properties for solving nonlinear equations with convex constraints, Journal of Computational and Applied Mathematics, v.172 n.2, p.375-397, 1 December 2004 Christian Kanzow , Nobuo Yamashita , Masao Fukushima, Levenberg-Marquardt methods with strong local convergence properties for solving nonlinear equations with convex constraints, Journal of Computational and Applied Mathematics, v.173 n.2, p.321-343, 15 January 2005 Stefania Bellavia , Maria Macconi , Benedetta Morini, STRSCNE: A Scaled Trust-Region Solver for Constrained Nonlinear Equations, Computational Optimization and Applications, v.28 n.1, p.31-50, April 2004 Jong-Shi Pang , Defeng Sun , Jie Sun, Semismooth homeomorphisms and strong stability of semidefinite and Lorentz complementarity problems, Mathematics of Operations Research, v.28 n.1, p.39-63, February
potential function;global convergence;interior point methods;primal-dual methods;complementarity problems;semidefinite programming;potential reduction algorithm;newton method;constrained equation;variational inequality
588961
A Practical Algorithm for General Large Scale Nonlinear Optimization Problems.
We provide an effective and efficient implementation of a sequential quadratic programming (SQP) algorithm for the general large scale nonlinear programming problem. In this algorithm the quadratic programming subproblems are solved by an interior point method that can be prematurely halted by a trust region constraint. Numerous computational enhancements to improve the numerical performance are presented. These include a dynamic procedure for adjusting the merit function parameter and procedures for adjusting the trust region radius. Numerical results and comparisons are presented.
Introduction . In a series of recent papers, [3], [6], and [8], the authors have developed a new algorithmic approach for solving large, nonlinear, constrained optimization problems. This proposed procedure is, in essence, a sequential quadratic programming (SQP) method that uses an interior point algorithm for solving the quadratic subproblems and achieves global convergence through the application of a special merit function and a trust region strategy. Over the past several years the theory supporting this approach has been analyzed and strengthened. This theory is presented in a companion paper [4]. In addition, implementations of the algorithm have been extensively tested on a variety of large problems, including standard test problems and problems of engineering and scientific origin, ranging in size from several hundred to several thousand variables with up to several thousand constraints. Specific strategies have been developed for handling the parameters utilized by the algorithm and for dealing with nontrivial pathologies (e. g. , linearly dependent active constraint gradients or inconsistent linearized constraints in the quadratic subprob- lem) that often occur in large scale problems. In this paper we present the results of these efforts. Based on its theoretical foundation and on our numerical experience we are confident that this algorithm provides an efficient means for attacking a large, sparse, nonlinear program with equality and/or inequality constraints. Rigorous comparisons of algorithms for large nonlinear problems is notoriously difficult, especially given the extensive set of options typically available in codes for such problems. Nevertheless, our algorithm, with the (conservative) default parameter settings, has been successful on problems that have caused difficulties for other algorithms and, consequently, we are encouraged to believe that it is competitive at the current stage in the development of methods for solving these large problems. Below we give an outline of our basic procedure and in the succeeding sections we provide more specific detail on the component parts of the implemented algorithm, including the strategies and safeguards that we have used. We also exhibit and comment on the results of some of our numerical tests. This paper relies heavily on the results from the paper on the theory for motivation of the basic ideas. Applied and Computational Mathematics Division, National Institute of Standards and Tech- nology, Gaithersburg, MD 20899 y Department of Mathematical Sciences, Carnegie Mellon University, Pittsburgh, PA 15213-3890 z Mathematics Department, University of North Carolina, Chapel Hill, NC 27599 We assume the general nonlinear programming problem to be of the form subject to: g(x) 0 are smooth functions. Nonlinear equality constraints are not included in our description here in order to avoid distracting technicalities. The modifications necessary for their insertion can be inferred from [6]. Nonlinear equality constraints are included in our code and in some of the problems we tested. The sequential quadratic programming method is the backbone of our algorithm. (See [7] for a review of these techniques.) At the kth step we have an iterate, x k , denoting the current approximation to the solution of (NLP ). In addition to the x-iterate we also maintain a non-negative iterate, z k 2 R m , which measures the infeasibility at x k . At this stage (NLP ) is modeled by a quadratic program of the form min subject to: rg(x k is taken to be an appropriate approximation to the Hessian of the Lagrangian for (NLP ), i.e., where and H xx represents the Hessian with respect x of the function to which it is applied. (See Section 4.5 for a discussion of the choice of B k used in our numerical experiments.) In this form (QP ) generates a step that provides a search direction for improving the current iterate. There are two significant points to be made concerning this phase of our algorithm. First, we apply an interior point quadratic program solver to (QP ); more specifically, we use the method found in [1] where solutions are calculated by solving a sequence of low dimensional quadratic programs. Pertinent details of this solver and its properties relative to its use in our SQP method can be found in Section 2. Second, we do not try to solve (QP ) with complete accuracy at each iteration; rather, we often terminate the interior point method prematurely. In particular, we halt the quadratic program solver when the steplength exceeds a "trust region radius" that is modified at each iteration according to how well the improvement in our merit function is predicted. Thus our algorithm can be said to be a "truncated Newton method" in the sense of [18] (see also [15]). This particular merit function and a more useful "working version" are discussed in Section 3 and our strategy for updating the trust region radius is given in Section 4.2. The output of the (QP ) solver is a vector that determines the direction of the step in the x-variable, which in turn yields a step direction for the "slack" variable z as explained in Section 3. The combined step direction of these two variables is a descent direction for the working version of the merit function and also for constraint thus we can choose the steplength in this direction to decrease the merit function and/or the infeasibility of the iterate. The choice of steplength determines the new iterate x k+1 and also the new value z k+1 . The strategy for choosing the steplength and other algorithmic details, including the modifications and safeguards necessary to make an implementation robust, are given in Section 4. The results of our numerical tests are contained in Section 5. These results demonstrate the overall effectiveness of the procedure and highlight the beneficial effect of our trust region strategy and other procedures. Finally, in Section 6 we briefly consider weaknesses in the current version of the algorithm and suggest possible avenues of research to improve its efficiency. For a discussion of the theoretical and practical questions related to large scale nonlinear programming see the recent surveys [12], [14] and [21]. 2. An Interior Point QP Solver. Interior point methods for linear programming have been demonstrated to be very successful, especially on large problems, and recent research has lead to their extension to quadratic programs. A particular method, the method of optimizing over low-dimensional subspaces, has performed well on linear programs and has been extended to the quadratic programming case (see [1], and [2] and the references contained therein). This method, for which good numerical results for quadratic programs have been reported, has properties that make it particularly compatible with the SQP algorithm we are describing in this paper. A brief description of the essential features of this method and their importance for our purposes follow. The many details of the actual algorithm that are not reported here may be found in the above references. The quadratic program that we solve, (QP ), has the form subject to: A T s n\Thetan , A 2 R n\Thetam , and b 2 R m . The assumptions on (2.1) that are necessary to apply the interior point algorithm are that the problem be bounded, that A have full column rank, and that there exist feasible points (i.e., that the constraints be consistent). Note that Q can be indefinite and that no assumption of a full-dimensional interior is required. If equality constraints are present, they are handled by writing them as two inequalities. An important prerequisite for solving (2.1) by an interior point method is a feasible initial point. Our algorithm uses a "Big M " method to construct the Phase I problem min subject to: A T s where e is a vector of ones and ' is the "artificial" variable. Clearly for ' large enough the point (s; is feasible for (2.2) and if M is sufficiently large the algorithm applied to (2.2) will reduce ' until the artificial variable is nonpositive, at which point the current value of s is feasible and the M' and e' terms are dropped. If no such value of the artificial variable can be found, then (2.2) is not consistent and the algorithm stops. As discussed below, we make use of the step obtained from (2.2) even if it is not feasible for (QP ). Note that when equality constraints are present, the entire solution procedure takes place in Phase I and ' will always be present. The defining characteristic of the algorithm is that it proceeds by solving a sequence of low-dimensional subspace approximations to (2.1). In our application we follow the reported results in which the dimension of the subspace is taken as three. The following is an outline of the O3D (for Optimizing over 3-Dimensional subspaces) version of the algorithm. As the variable ' is treated essentially the same as the components of s in the O3D algorithm (see, however, Step 6 below) the dependence on ' is incorporated into the formulation given in (2.1). O3D Algorithm for Quadratic Programming 1. Given a feasible point, s 0 2. Generate 3 independent search directions be the matrix whose columns are p i . 3. Form and solve the restricted quadratic program subject to: A T ~ s where ~ Call the solution i . 4. Set s j+1 := s for an appropriate value of the steplength ae 2 (0; 1): 5. If stopping criteria are met, exit. 6. Go to 2. (At this step, if the component of the vector s corresponding to the artificial variable ' has become nonpositive, it is eliminated from the problem.) The search directions in Step 2 are solutions to \Theta AD 2 A T where fi is a scalar depending on the current iterate, with , and the t i are particular values chosen such that one of these directions is always a descent direction with respect to the objective function. The steplength ae is set to the lesser of 99% of the distance to the boundary and the distance to the minimum of the objective function. The form of the matrix in (2.3) allows for efficient exploitation of the sparsity. Note that if Q is positive semi-definite, then the matrix in (2.3) is positive definite for all interior points; otherwise, it may not be. In the latter case, a modification similar to that in [20] is used. In our application of this algorithm, using this procedure obviates the need for the matrix B k to be positive definite, which in turn allows us to use the Hessian of the Lagrangian or a finite difference approximation thereof. The standard stopping criterion for the algorithm is that at least one of the following holds: (a) the relative change in two successive values of the objective function is small; (b) the relative difference between the primal and the dual objective function values is small; or (c) the difference between two successive iterates is small. For use in our SQP algorithm we have added: (d) the length of the solution vector exceeds a specified value. This additional condition has been implemented to allow for trust region strategies; in particular, this criterion will cause the algorithm to halt if (QP ) is unbounded. In any case, the terminal vector will be a useful direction in the context of our purposes; this point will be discussed in the next section. The most recent version of O3D described in [1] contains an option to perform a special "recentering step" after each subspace optimizing step (Step 4) that has generally improved the efficiency. This option is not used in the results reported here. (See Section 6 for a further comment.) 3. Updating the Iterates: the Merit Functions. In this section we review the definitions and properties of our merit functions and provide formulas for updating the iterates. The reader is referred to the companion paper for proofs and motivations of these concepts. As stated in Section 1, at each iteration our algorithm yields a pair x k is an approximation to the solution of (NLP ) and z k is the corresponding approximate slack vector. The step directions for the updated values of these approximations are based on the (approximate) solution, to the quadratic program min ffi;' subject to: rg(x k obtained as described in the preceding section. The vector ffi k gives the step direction for x k and we determine the step direction, q k , for the slack vector z k by the formula \Theta rg(x k Note that if ffi k is feasible for (QP ) then \Theta rg(x k In this case z k is the slack vector for (QP ) corresponding to ffi k and thus is the slack variable for the linear approximation of g(x k+1 ). Given the step direction we then update the iterate by means of the formulas z for some value of the steplength parameter ff: Observe that if z k 0 then the fact that means that z k+1 will be non-negative if ff 2 [0; 1]. In our algorithm the non-negativity of the slack vector iterates is preserved and, in fact, it sometimes turns out to be useful to maintain the z k at a positive level (see Section 4.8). It is important to emphasize that the ffi k are determined by (QP ), the quadratic approximation to (NLP ), and are not dependent on the choice of z k . The z k are generated solely for use with the merit function described below. That is, we do not solve the slack variable problem. A comment on the notation is also in order at this point: We denote the iterate by and the step by (ffi k ; q k ), whereas conventional notation would be to use ' x k z k and It should be clear from the context what is meant. In optimization algorithms the value of a steplength parameter is generally chosen so as to reduce the value of a suitably chosen merit function. Typically, a merit function for (NLP ) is a scalar-valued function that has an unconstrained minimum at x , a solution to (NLP ). Because a reduction in this function implies that progress is being made towards the solution, it can be used to determine an appropriate steplength in a given search direction. In [5] and [6] a merit function for equality-constrained problems was derived that has important properties vis-a-vis the steps generated by the SQP algorithm. Using a slack-variable formulation of (NLP ) a merit function for the inequality constrained problem can be constructed having the form d c(x; z) T where z is nonnegative, d is a scalar, and We use this merit function (and its approximations defined below) for choosing the value of the steplength parameter ff. As noted above, the approximate slack vectors generated by our algorithm, z k , always remain non-negative; thus the non-negativity constraint on the z for / d imposes no theoretical difficulty. The function c(x; z) defined above plays an important role in our algorithm as it is used to measure the feasibility of the pair (x; z). That is, if we define the function where k\Deltak denotes the standard Euclidean norm and set then C 0 corresponds to the feasible set of (NLP ) and hence close to feasible if it is in C j for small j. For d sufficiently small the merit function / d has the desirable property that a solution of (NLP ) corresponds to a (constrained) minimum of / d . In addition, if d is small and ffi k is the exact solution to (QP ) (which implies that ' then the step descent direction for / d when sufficiently close to feasibility. Despite these useful properties, / d has two deficiencies that limit its use in an efficient algorithm. First, (ffi k ; q k ) is a descent direction of / d only near feasibility, and, second, the evaluation of rf and rg and additional nontrivial computational algebra are required to assess a prospective point. In order to overcome these difficulties, the approximate merit function d d where is developed as a "working" version of / d at As the values of k and are fixed, / k d can be more easily evaluated than / d in a line search algorithm for choosing an appropriate value of ff. This approximate merit function, / k d , not only has essentially the same properties as / d with respect to the step has the stronger property that the step is a descent direction for / k d everywhere. Moreover, for sufficiently small and outside of a ball around the solution a "sufficient" reduction in / k d implies a "sufficient" reduction in / d . (We mean by "sufficient" reduction that a Wolfe condition is satisfied.) Thus we are able to use / k d as a surrogate for / d for testing the progress of our iterates towards a minimum. A further important property of the step ffi k , under the assumption that it is the exact solution to (QP ), is that it is a descent direction for the function r defined by (3.4). Thus a basic algorithm for the case where the (QP ) can be solved exactly is as follows: Given an initial value of j use the steps (ffi k ; q k ) to reduce r until the iterates are in C j . Once the iterates are contained in C j if a sufficient reduction in / k d does not yield a sufficient reduction in / d then reduce j. If, in the course of the algorithm, remains bounded away from zero, then convergence follows from the fact that the Wolfe condition is satisfied for / d . If j goes to zero, then convergence follows from the observation that the radius of the ball in which the Wolfe condition is not satisfied also goes to zero. This is essentially the algorithm for which global convergence is proved in the paper on the theory. In this paper we are primarily interested in enhancements that convert the theoretical algorithm into one that is practical and efficient. This requires that we make provisions for situations when the assumptions under which we performed the convergence analysis are not valid and that we adopt numerical procedures to reduce the computational effort. As we note below, not all of these modifications have been (or even can be) theoretically justified, but we believe that the firm foundation of the underlying algorithm and the evidence accumulated in extensive numerical testing validate their use. In the implementation of our algorithm a trust region constraint is used that possibly truncates the quadratic programming algorithm before an exact solution is achieved. In this case the theory described above does not apply for the step obtained from the approximate solution, Although a general convergence theory based on this step is not yet available, it is shown in the theory paper that if the approximate solution is obtained from the O3D algorithm and if ' k is not too large then the resulting step has the appropriate descent properties for the functions r, / d , and / k d at In particular, convergence can be achieved if ' k goes to zero in a suitable manner. These properties justify our use of the truncation procedure to speed up the algorithm. It is important to note that this approximation procedure also allows us to handle the difficulty that arises in sequential quadratic programming methods when the quadratic subproblem is inconsistent. 4. The Truncated SQP Algorithm. In this section we give a somewhat detailed description of our algorithm. Initially we assume that the Hessian approxima- are positive definite, the matrices A k , are nonsingular, and the linearized constraints in (QP ) are consistent. In real-world applications these assumptions are not always valid so we have tried to make our algorithm flexible enough to perform well in situations where these assumptions fail to hold. We describe some of these adaptations at the end of this section. The implementation of the algorithm depends upon four important parameters that need to be either computed or modified throughout the course of the algorithm. The globalization parameter, j, was introduced in (3.5). It is a measure of the size of the domain about the feasible region in which the direction (ffi k ; q k ) is a descent direction for the true merit function / d . A current estimate of j is maintained in the algorithm. The trust region parameter, , is an upper bound on the (weighted) norm of our approximate solution to (QP ), where D is a positive definite diagonal matrix. The trust region radius is updated at every iteration. The parameter, ff, is the steplength parameter. It determines the length of the step in the variables (x; z) in the direction (ffi k ; q k ). It is chosen to guarantee progress towards the solution in decreasing either the merit function or infeasibility. Finally, d, the merit function parameter, must be small enough to guarantee that the theoretical properties described in the preceding section are valid. Although the theory allows arbitrarily small values of d, the algorithm becomes very slow if d is too small, thus it is monitored throughout the algorithm and either increased or decreased as appropriate. The outline of the algorithm is followed by specific comments on the procedures and their justifications. This version contains some of the practical modifications described above. To simplify the notation we define Recall that r is given by (3.4). Basic Truncated SQP Algorithm 1. Initialization: Given x a. Initialize the slack variable z 0 0; b. Set k := 0: 2. Calculation of the basic trust region step: a. While kffik ! , iterate (using O3D) on min subject to: rg(x k to obtain ffi k and ' k . b. Set \Theta rg(x k \Theta rg(x k otherwise c. Decrease d if necessary. 3. Computation of the steplength parameter: a. Choose ff 2 (0; 1] such that / k d is sufficiently reduced. b. If reduce ff if necessary until r is sufficiently reduced. c. If reduce ff if necessary so that 4. Update of the estimate of the globalization parameter: a. If set 5. Update of the variables and check for termination: a. Set z b. If convergence criteria are met, quit. c. Update B k to B k+1 . 6. Adjustment of the merit function and trust region parameters: a. Update d if necessary. b. Adjust the trust region radius . 7. Return: a. b. Go to Step 2. 4.1. The Globalization Parameter. The globalization step is based on work in [6] and [4]. In Step 3 we require that the approximate merit function be reduced and, in addition, if the current iterate lies outside the set C j we require that the constraint infeasibilities also be reduced. This is possible as a result of the descent properties described in Section 3. If we have a good estimate of j and then the true merit function can also be reduced; if this is not the case, then our estimate of j is too large and we reduce its value in Step 4. This procedure will eventually lead to a sufficiently small value of j. Note that this arrangement allows steps that may increase the merit function, but only in a controlled way. It also allows steps that may increase the constraint infeasibilities, but only when inside of C j . 4.2. Updating . Our procedure for updating , the trust region radius, in Step 6b is similar to the standard strategy used in trust region algorithms (see [17] or [31]) in that we base the decision on how to change on a comparison of a predicted relative reduction, pred k ; and an actual relative reduction, ared k , in a function used to measure the progress toward the solution. (Various formulas for the predicted relative reduction, pred k , have been suggested for different merit functions, especially for equality constrained programming problems; see, for example, [19]). What is distinctive about our procedure is that we use different functions for computing pred k and ared k depending on the current status of the algorithm. When the linearized constraints are satisfied we use the approximate merit function to compute the predicted and actual reductions. When the trust region constraint causes O3D to terminate in Phase I, i.e., when the linearized constraints are not satisfied, predicted and actual reductions in infeasibility are used. In the case when a feasible solution to (QP ) is obtained then / k d is used to compute the predicted and actual reductions. Our method for defining pred k differs from the standard methods used in unconstrained optimization because the step- finding subproblem is not based solely on the merit function and, moreover, the trust region constraint does not appear explicitly in the subproblem. Nevertheless in updating we want to assess how well an approximation to / k d agrees with / k d in the direction uses a quadratic approximation of the Lagrangian for the objective function with linearized constraints, we form our approximation to d based on a quadratic approximation to the function / k 1 given by and a linear approximation to Note that / k d z). Based on these considerations and the results of [16] we define the predicted relative reduction by pred ae d oe where the derivatives are with respect to x and z and the steplength parameter ff k is the size of the most recently accepted step. The value of the actual relative reduction, ared k , is taken to be the difference in the values of / k d at the points and divided by the value of / k valid criticism of the formula for pred k is its dependence on higher order derivatives. Therefore we use the available approximation of the Hessian of the Lagrangian for r 2 / k 1 . For example, cell-centered finite difference approximations to the Hessian of the Lagrangian function were used in the numerical results presented here, unless analytic second derivative formulas were readily available. The above choice for pred k is not used when the step returned by O3D is not feasible. In these situations the resulting step is dominated by a feasibility improving component and it makes little sense for the adjustment to to be determined by / k rather, a comparison of the predicted and actual improvement in constraint infeasibility seems more appropriate. Therefore, in this case the function r(x; z) is used for comparison purposes. The values of pred k and ared k are given as follows for the case when the O3D algorithm terminates in Phase I: pred and ared These heuristics for choosing pred k and ared k appear to work well. Specifically, they allow the trust region radius, , to be increased even in the event that the step returned by O3D does not satisfy linearized constraints or it results in an increase in the true merit function. In our experience, the alternative formulas based solely on constraint violations never are employed close to the solution. Indeed, the iterates preceding convergence have always been observed to be well inside C j where satisfying the linearized constraints and decreasing the merit functions usually pose no problem. 4.3. The steplength ff. The steplength ff is determined in Step 3 of the al- gorithm. The "sufficient decrease" referred to in 3a and 3b requires that the Wolfe condition be satisfied. For a given function OE and potential step w from point v this condition requires that ff satisfy for some fixed oe 2 (0; 1). In the numerical experiments reported in Section 5 we employed a simple backtracking procedure (with factor one-half) to find ff to satisfy this condition for both / k d and for r. We have also experimented with more sophisticated line search methods motivated by unconstrained optimization techniques as in [18], but the observations to date suggest that the more complicated line searches result in very little improvement of our algorithm, except when the iterates are quite far from the solution. 4.4. Adjusting d. Choosing an effective value for the merit function parameter d is essential in our algorithm. While it is clear that (in a compact set) a sufficiently small value of d will assure that the results given in [4] are valid, there are three very important practical reasons why the parameter must be adjusted rather than fixed. First, if the angle between the direction generated by O3D and the gradient of the approximate merit function becomes nearly orthogonal the steps might become too small. We adjust d to avoid this possibility. Second, the approximate merit d , is changing at each iteration and it is possible a previous iterate might be acceptable to the current / k d , i.e., cycling might occur. This worry can also be alleviated by adjusting d. A third reason for changing d is to allow for larger steps. It is seen from the theory and has been verified by numerical experience that if d is too small then the form of the merit function forces the path of the iterates to follow the "nearly active" constraints closely. This causes the algorithm to take very small steps and, in particular, to be slow in moving away from a nonoptimal active set. By making it possible to increase d we can significantly improve the algorithm's performance. In the implementation of our algorithm there are two opportunities to adjust d: in Step 2, after solving the quadratic subproblem, and in Step 6, after the step has been taken. In the first of these adjustments d can only be decreased; in the second, the parameter may be increased or decreased. In Step 2, the angle between the gradient of the approximate merit function d and the step direction (ffi k ; q k ) is computed. If these two vectors become nearly orthogonal, we conclude that d is not small enough to ensure a good decrease in / k d , and we decrease the parameter. To be more specific, we compute If w(d) \Gamma:1 we calculate a value " d so that w( " d) \Gamma:5: We safeguard the procedure by not allowing more than a certain percentage decrease in d. In the current version we use 50%. If d was not decreased in Step 2 we consider modifying it after a step has been taken (Step 6). Here the primary concern is to avoid cycling. To do so we compute an interval for the penalty parameter as follows. For a fixed integer we seek a value of the parameter, d, such that d Inequality (4.2) implies that none of the past iterates will be acceptable to the approximate merit function with the new value of d. (Thus if cycling would be possible). To accomplish this, we use the decomposition 1 and / k are defined in Section 4.2. We then compute the values of / k and / k consider the inequalities d d We define d u i and d l i to be the upper and lower values of d that ensure that inequality (4.4) is satisfied. Then letting d and we obtain an interval (d l ; d u ). Assuming that this interval exists it is the case that if the value of d for the next step is chosen in this interval, the next iterate will not return to one of the previous iterates. In practice a value of 5 is usually more than sufficient to prevent cycling. If the interval doesn't exist, then we make no change. Given that we can choose d to avoid cycling, our second objective at this juncture is to increase d to allow bigger steps. If the d u is larger than the current d then we can safely increase d without worrying about possible cycling. However, we safeguard this increase in two ways. First, we require that the predicted reduction based on the approximate merit function must be greater than the predicted reduction of infeasibility in the linearized constraints. This restriction prevents d from being increased prematurely due primarily to a large decrease in constraint infeasibilities. Specifically, writing the predicted reduction in / k d (see (4.1)) as we insist that for a new value of d Second, we use a maximum allowable change (currently a factor of 2) to limit the growth of d. Computationally, these simple procedures for updating d appear to be effective, especially in the presence of highly nonlinear constraints and poorly scaled problems. 4.5. The Hessian Approximation. In the numerical experimentation reported here, we have used a finite difference approximation to the Hessian of the Lagrangian . Although the Hessian of the Lagrangian at a strong solution is positive definite on the appropriate subspace, it may be indefinite in general. Even if it is positive definite the finite difference approximation may not be. We experimented with two approaches for handling this possibility. First, we simply modified the approximate Hessian matrix by adding non-negative elements to the diagonal ensuring that the Cholesky factorization of the matrix had positive elements along its diagonal (see [20]). This modification was easy to implement, but it was observed to slow convergence on some problems. While this modification guarantees that a positive definite matrix will be delivered to the (QP ) solver, if it takes place when the iterates get close to the solution, it generally precludes local q-superlinear convergence. An alternative to modifying the approximate Hessian of the Lagrangian is simply to allow O3D to iterate on the indefinite QP subproblem, halting the iterations when the solution exceeds the trust region radius. We implemented this approach and it seemed to yield superior results to those obtained by making the approximate Hessian positive definite (especially when the iterates were close to a solution) even though, theoretically, we can only prove that we obtain a descent direction when the approximate Hessian is positive definite. 4.6. Convergence Criteria. The convergence criteria used are standard, and similar to those in [3]. We first insist that the constraints be satisfied to a close tolerance; specifically we require We also require that either or The criterion (4.9) is a stronger indication that a KKT point has been reached. The weaker criterion (4.10) suggests that progress slowed drastically and that iterates may or may not have drawn close to a solution. For this reason criterion (4.9) is usually preferable to criterion (4.10). The Lagrange multipliers returned by the quadratic program are used in (4.9) unless the trust region constraint determines the approximate solution of the (QP ). In that case, we use the least squares approximation to the multipliers, replacing all negative multipliers with machine zeros. In all of the problems solved to date, the trust region never comes into play when the iterates get close to the solution; therefore the (QP ) multipliers are used for the convergence test at the solution. 4.7. Inconsistent Quadratic Subproblems. One difficulty that can occur when making linear approximations to nonlinear constraints is that (QP ) may be inconsistent. In this case O3D will, even if it runs to completion, not exit Phase I and will return a positive value of the artificial variable. (Note that this always occurs if equality constraints are present.) For small ' the resulting direction is a descent direction for / k d and for r. As a result, the step taken in this direction will generally decrease infeasibility, making it less likely that an inconsistent set of linearized constraints will be encountered during subsequent iterations. More recent versions of our algorithm include a constraint relaxation procedure that appears to yield an acceptable step, even in the event that inconsistent linearizations of constraints are encountered. Because this situation did not surface during the numerical experiments presented in this paper, we do not include a description of our perturbation procedure. We do note, however, that we have encountered important application problems where this procedure was crucial to the performance of our algorithm (see for example [24]). 4.8. Updating slack variables. One difficulty in our algorithm is the updating of slacks in the event that the SQP step does not satisfy the linearized constraints well enough, i.e., ' k is not small enough. This can occur when (QP ) is inconsistent or when a trust region bound is encountered during the solution of (QP ). In this case our slack variable updating scheme would ensure that non-negative slacks remain non- negative, but the direction may not be one of descent. We resolve this dilemma by opting for descent, i.e., computing q k with ' replacing any negative slacks using the following rule: If z k+1 z k+1 ae ffl Mach g i \Gammag i where ffl Mach is machine epsilon. This is sometimes referred to as 'closing' the constraints (see for example [33]). 4.9. Linearly Dependent Constraint Gradients. Linearly dependent constraint gradients cause many theoretical and computational difficulties in constrained optimization. In our theoretical algorithm we obtain convergence even when there are linearly dependent constraint gradients provided the approximate multipliers do not become unbounded. In practice, even though O3D has no difficulty in dealing with this problem, evaluating the merit function and computing the least squares approximation to the Lagrange multipliers become problematical. Computational experience shows we solve many problems with degeneracy in the constraints. Simply maintaining slacks to to be positive as described above allows us to factor the crucial matrices and continue with the algorithm. However, the algorithm failed to solve some problems that had a large amount of degeneracy in the linearized constraint matrix. This was, of course, problem dependent but it was observed that the current implementation can usually solve problems where up to 25 percent of the constraint gradients are linearly dependent. This degeneracy causes the performance of the merit functions to deteriorate. In particular, the least squares approximation to Lagrange multipliers seems to be especially poor, resulting in only very small steps being allowed, even close to the solution. 5. Numerical Results. The modified algorithm was coded in Fortran and is installed on a SPARCstation 10 using IEEE floating point arithmetic (64 bit). The current implementation is being used to solve a wide variety of medium to large scale problems. In this section we report the results of a set of performance tests designed specifically to answer questions about the trust region strategy and the procedure to update the penalty parameter, d. We conclude the section with the results of our algorithm applied to some test problems that are publicly available. We emphasize that all of the problems were solved with the same default settings of the parameters, (see Table 1), i.e., no attempt was made to pick parameter settings to optimize performance on individual problems. Although in many of the applications some analytic derivatives were available, no use of analytic derivative information was used in these numerical experiments. When possible, first and second derivatives were computed using forward and central finite differences respectively. A costly one-time calculation provided a zero/non-zero stencil of the Hessian of the Lagrangian and the Jacobian matrix of the constraint function. These stencils were then used for the duration of the solution process. For some problems, these finite difference approximations are not convenient to use. This can be the case with control problems governed by partial differential equations (see [29] or [30]). If the partial differential equation is solved using a finite element method, with piecewise linear elements, then evaluating the derivative of the objective Parameter Value z Mach ) Table Numerical values of default parameters function with respect to the control variables can be quite cumbersome. In such cases, which occurred in the control problems in our test suite, one can approximate the first derivatives of the objective function by solving an adjoint problem with a computational cost comparable to one function evaluation. (For examples, see [22].) The objective function portion of the Hessian of the Lagrangian can then be approximated with forward finite differences. A set of eight problems was chosen as the first test suite. These problems ranged in size from 500-1000 variables and from 1000-2000 constraints. The first four are relatively straightforward nonlinear programming test examples, while the last four are from actual applications: two discretized control problems, a density estimation problem from statistics, and a "molecular distance" problem. A more complete description of these problems is found in the Appendix. The problems all have nonlinear inequality constraints and exploitable sparsity. Problem 4 (NLP4) was designed to have a controllable percentage of linear dependency in the constraint gradients to demonstrate any weaknesses in the algorithm associated with this difficulty. We ran three versions of our algorithm on each problem; using a positive definite modification of the Hessian matrix, as discussed in Section 4, with and without the trust region strategy and using the unmodified Hessian with the trust region. (Using the unmodified Hessian results in failure in most cases if no trust region strategy is employed.) In addition, each problem was run from two starting points; one, labeled "c", which was close to the solution in the sense that each of the variables was of the same order of magnitude as in the solution and a distant start, labeled "f". The results of the numerical tests on these problems are summarized in Tables 1-3. The first two columns of each table gives the number of SQP iterations ("nl-i") and the total number of O3D iterations ("qp-i"). The next two columns contain the stopping criterion that was met and the value of the gradient of the Lagrangian at the solution. Unless the algorithm failed, (which is denoted by "Failure" in the tables) feasibility condition (4.8) was satisfied for all solutions. The stopping criterion is denoted by either a "1" or a "2" depending on whether (4.9) or (4.10) was satisfied. If both were conditions were satisfied, a "3" appears in the column. The remaining columns give information about the values of the parameter d for each run; columns five through eight giving the initial, maximum, minimum, and final values of this parameter and the final column giving the last iteration at which d was changed. The results of the tests illustrate that using the unmodified Hessian with the trust region was most effective in reducing the number of O3D iterations and the number of SQP iterations. The trust region strategy prevented long, unprofitable steps from being generated when far from the solution and the use of the unmodified Hessian allowed the trust region to become inactive near the solution thus allowing rapid local convergence. Requiring the Hessian to be positive definite often precluded rapid local (q-superlinear) convergence and, when used in conjunction with the trust region strategy, resulted in the trust region's being active close to the solution. The results also show that the value of the parameter d varied over several orders of magnitude. The procedures discussed in Section 4 that allowed the value of d to increase or decrease greatly enhanced the algorithm; earlier tests using either a fixed value of d or only allowing a reduction in d yielded inferior results. Another modification in our algorithm, not reflected in the table or included in the description in the preceding section, was made to force the O3D algorithm to take a minimum number of steps. We found that when the trust region radius became small the algorithm would sometimes exit O3D after only one iteration, resulting in a poor step direction. This poor step would result in a further decrease in , and eventually the algorithm would fail. When we required a minimum number of steps to be taken in O3D (our choice was 7) this problem disappeared. Recently a collection of test problems has become available for the testing and comparing of optimization algorithms, (see [13]). The Constrained and Unconstrained Testing Enviroment (CUTE), are quickly becoming standards with which researchers can establish the viability and effectiveness of their numerical algorithms. These problems are replacing the smaller and well scaled test problems of Hock and Schittkowski [25] and Schittkowski [32] which were not intended to be used to test large scale al- gorithms. Our results on the CUTE test problems are summarized in Tables 6, 7 and 8. These problems were solved to the same stopping conditions as the problems above. Likewise, the same table format was used to present these numerical results. For detailed description of these problems, structure, motivation, and sources see [9]. While it appears that the CUTE test problem set is rich in both large and small scale unconstrained and equality constrained test problems, at present there are not many large scale problems that include inequality constraints (and particularly non-linear inequality constraints). We chose problems that reflected the class of problems our algorithm was designed to solve. At least one inequality constraint was present in each problem. The number of variables and/or constraints was large enough so that the exploitation of special sparsity structure was important. The problems we selected from CUTE to report on were "CORKSCREW, MANNE, SVANBERG" and "ZIGZAG". The associated problem sizes are recorded in Table 5. It is worth commenting that much of the machinery developed in this paper deals with effectively handling nonlinear inequality constraints. The performance of our algorithm on the CUTE test problem set is, therefore, slightly deceiving since many of the constraints in these problems are simple bounds on the primal variables or purely linear. (For instance, approximately 83% of the constraints in CORKSCREW, 50% of the constraints in MANNE, and 66% of the constraints in ZIGZAG were linear and many of them were equality constraints). Although these caused no problem for our algorithm, the structure of these constraints was not completely exploited and the extra machinery of our code resulted in an overhead with no performance benefit. Clearly an algorithm designed specifically to deal with linear equality constraints should outperform our algorithm on these problems. The problem on which our algorithm appeared to perform best was SVANBERG, a problem with only inequality constraints (and a substantial number of them are nonlinear). We succeeded in solving all four problems with a reasonable number of inner and outer iterations. However, many of our algorithmic enhancements contributed little to the solution process. The measure of distance to feasibility (the j-tube strategy), the nonmonotone updating of penalty parameter d, and the trust region strategy were essentially dormant during the solution process regardless of the iterates' proximity to the solution or to feasibility. In fact, the only evidence of our enhancements on the small number of CUTE test problems that we solved occurred when d was decreased slightly while solving the problem MANNE employing modified Hessians with a trust region strategy (see the third and fourth rows of Table 7). It is noteworthy that the iterates that resulted from solving this problem with the penalty parameter artificially held fixed at were identical to iterates that resulted for the adjusted d solution. This appears to illustrate that in this case the adjustment of d was purely superficial. Truss - c 103 3561 1 4.4e-8 9.87e-1 1.01e00 9.57e-2 9.57e-1 100 Molec-c 37 1376 1 9.8e-9 9.88e-1 1.21e1 1.17e-2 9.81e-1 34 Molec-f Table Modified Hessians with no trust region BCHeat -c 257 2898 1 8.1e-8 9.18e-1 4.15e00 1.38e-1 4.10e00 249 BCHeat -f 289 3071 1 9.9e-8 1.00e00 3.74e00 2.44e-1 3.89e00 281 Molec-c Molec-f 44 621 1 6.6e-8 1.00e00 1.48e00 7.39e-2 9.52e-2 38 Table Modified Hessians with trust region Truss-c 94 2242 1 3.9e-8 9.87e-1 1.03e00 1.26e-1 9.11e-1 87 Molec-c Molec-f Table Un-modified Hessians with trust region Problem Variables Constraints CORKSCREW 96 159 ZIGZAG 304 1206 Table Minimization parameters Problem nl-i qp-i conv krx lk1 d0 maxd mind final d last d-cha Table Modified Hessians with no trust region Problem nl-i qp-i conv krx lk1 d0 maxd mind final d last d-cha Table Modified Hessians with trust region Problem nl-i qp-i conv krx lk1 d0 maxd mind final d last d-cha Table Un-modified Hessians with trust region 6. Future Directions. In this paper we have discussed in some detail an SQP algorithm for solving large scale nonlinear problems. The numerical results with default parameter settings indicate that the procedures that we have implemented are robust, effective, and efficient; the convergence theory in [4] provides a sound theoretical basis for the procedure. Nevertheless, there are several areas in which the techniques used here can be improved to allow the solution of larger and more difficult problems. Algorithmically, we observe that the current implementation requires the factorization of both (rg T rg+Z) and (rgrg T ), the latter in O3D. While the sparse matrix package makes this reasonable for the problems that we have currently considered, it is clearly expensive to maintain both. The results reported here use analytic or finite difference Hessian approximations. An examination of the details of O3D reveals that a limited memory BFGS or limited memory SR1 could be readily incorporated into the code. We have done some experimentation with such techniques; the results will be reported elsewhere [26]. Many of the problems that we have seen have been degenerate and this significantly slows the convergence of the method. The primary culprit is the extremely poor multiplier estimates provided by the least squares procedure. Improvements in this area are certainly required. In some problems (not reported here) that have nonlinear equality constraints, we have occasionally observed significant difficulty in trying to satisfy the linearized equality constraints, i.e., in completing Phase I. In these cases we have had some success in relaxing the constraints [26]. In the context of O3D, this can be accomplished by simply fixing the artificial variable at some positive value and continuing the O3D iterations. In this approach, we often find that O3D converges and the "recentering" procedure mentioned in Section 2 has led to further improvements. The theory in [4] supports these ideas. The details will, again, be reported elsewhere. --R An interior-point method for general large scale quadratic programming problems An interior point method for linear and quadratic programming problems A merit function for inequality constrained nonlinear programming problems A family of descent functions for constrained optimization A truncated SQP algorithm for large scale nonlinear programming problems Cute: Constrained and unconstrained testing environment Functional and numerical solution of a control problem originating from heat transfer On exact and approximate boundary controlla- bilities for the heat equation Large scale numerical optimization: Introduction and overview Lancelot: A Fortran Package for Large-Scale Nonlinear Optimization SIAM Journal on Numerical Analysis Globally convergent inexact Newton methods A robust trust region algorithm with nonmonotonic penalty parameter scheme for constrained optimization New York Constrained nonlinear pro- gramming Numerical Methods for Nonlinear Variational Problems Molecular conformations from distance matrices Optimal signal sets for non-gaussian detectors Test Examples for Nonlinear Programming Codes The Use of Optimization Techniques in the Solution of Partial Differential Equations from Science and Engineering The solution of the metric stress and sstress problems in multidimensional scaling using Newtons method Numerical solution of a nonlinear parabolic control problem by a reduced SQP method Optimal Control of Systems Governed by Partial Differential Equations Computing a trust region step More Test Examples for Nonlinear Programming Codes On the role of slack variables in quasi-Newton methods for constrained optimiza- tion Nonparametric Probability Density Estimation Nonparametric Function Estimation where it was used to illustrate separability in nonlinear programming. "distance matrices" --TR --CTR Thomas F. Coleman , Jianguo Liu , Wei Yuan, A New Trust-Region Algorithm for Equality Constrained Optimization, Computational Optimization and Applications, v.21 n.2, p.177-199, February 2002 E. Bradley , A. OGallagher , J. Rogers, Global solutions for nonlinear systems using qualitative reasoning, Annals of Mathematics and Artificial Intelligence, v.23 n.3-4, p.211-228, 1998 Rubin Gong , Gang Xu, Quadratic surface reconstruction from multiple views using SQP, Integrated image and graphics technologies, Kluwer Academic Publishers, Norwell, MA, 2004
interior point;large scale;trust region;nonlinear programming;SQP;merit function
589001
A Multiple-Cut Analytic Center Cutting Plane Method for Semidefinite Feasibility Problems.
We consider the problem of finding a point in a nonempty bounded convex body $\Gamma$ in the cone of symmetric positive semidefinite matrices ${\cal S}^m_+$. Assume that $\Gamma$ is defined by a separating oracle, which, for any given $m\ti m$ symmetric matrix $\hat{Y}$, either confirms that $\hat Y \in \Gamma$ or returns several selected cuts, i.e., a number of symmetric matrices Ai, i=1,. . .,p, p\le p_{\max}$, such that $\Gamma$ is in the polyhedron $ \{ Y \in {\cal S}^m_+ \mid A_i \bullet Y \le A_i \bullet \hat{Y}, i=1,\ldots,p \}.$ We present a multiple-cut analytic center cutting plane algorithm. Starting from a trivial initial point, the algorithm generates a sequence of positive definite matrices which are approximate analytic centers of a shrinking polytope in ${\cal S}^m_+$. The algorithm terminates with a point in $\Gamma$ within $O(m^3p_{\max}/\epsilon^2)$ Newton steps (to leading order), where $\epsilon$ is the maximum radius of a ball contained in $\Gamma$.
Introduction be the set of mm symmetric matrices and let S m be its subset of symmetric positive semidenite matrices. We consider the problem of nding a point in a convex subset of We assume that contains a full-dimensional closed ball with radius > 0: The set is implicitly dened by a separating oracle, which, for any given mm symmetric matrix ^ Y , either conrms that ^ Y 2 or returns several cuts, i.e., a number of symmetric matrices A such that is in the polyhedron fY Here p max is the maximum number of cuts admitted in each iteration. In a recent paper [8], we presented an analytic center cutting plane method for the case p 1, in which a single cut is added in each iteration. The method was shown to have a worst-case complexity of O(m leading order). However, to make a cutting plane algorithm practically e-cient, adding multiple cuts is often necessary. The purpose of this paper is to propose and analyze an analytic cutting plane method that uses multiple cuts for solving the convex semidenite feasibility problem mentioned above. In admitting multiple cuts in an analytic center cutting plane method, we face some new theoretical problems that are dierent from the single-cut situation, these include (a) the problem of nding a feasible starting point for the Newton iteration after several new cuts have been added; (b) the estimation of the number of Newton steps needed to obtain a new approximate center through estimating the changes in the primal-dual potential function. Our paper extends the multiple-cut schemes of Go-n and Vial, Luo, and Ye [2, 5, 10] from Such extensions not only broaden the applications of cutting plane methods, but also extend several classical theoretical results for non-negative vectors to positive semidenite matrices. We note that for our multiple-cut analytic center cutting plane algorithm, the complexity analysis on the number of Newton iterations per oracle call follows the approach in [3].For the complexity analysis on the number of oracle calls, we follow the approach in [10], but we simplify the proofs of some results analogous to those in [10] by considering all the added cuts simultaneously instead of inductively. In this paper we will show that, starting from a trivial initial point, the multiple-cut algorithm generates a sequence of positive denite matrices which are approximate analytic centers of a shrinking polytope in S m . The algorithm will stop with a solution in at most O(m 3 (to leading order) Newton steps. Our analysis show that when the the problem is specialized to the space of positive semidenite diagonal matrices (which is equivalent to the non-negative the complexity bound is reduced to O(m 2 p This complexity bound is lower than the existing bound of O(m obtained in [2] and [10], where the same cuts are considered. Our bound appears to be better than that obtained in [5]. (Note that the proof for the bound appeared in [5] is incomplete, and to our best knowledge, a provable bound should be O(m Furthermore, the analysis in [5] is carried out only for the so-called shallow cuts that are placed at some distances away from the current testing point and hence may not be as e-cient as our proposed algorithm where the cuts pass through the testing point. We are able to obtain better complexity results than existing ones even when the problem is specialized to IR m because in each oracle call, we only admit cuts that are su-ciently good. We shall not give the precise denition of \goodness" here but refer the reader to section 4. Roughly speaking, base on our criteria, the admitted cuts A in each oracle call are eective in reducing the size of polytope in the sense that each should be able to delete a sizable portion of the current polytope that can not be otherwise deleted by the other admitted cuts. One obvious advantage of having such a selection criterion is that the number of cuts added in each iteration will be reduced since only eective cuts are admitted, and this translates into saving in the computational cost in each Newton step. We will now introduce some notations. For matrices A Y := tr(AY We write Y 0 and Y 0 if Y is positive denite and positive semidenite, respectively. For Y 0, we denote its symmetric square root by Y 1=2 . The 2-norm of a vector x is denoted by kxk, and the matrix 2-norm of a matrix A is denoted by kAk. For A are the eigenvalues of A. Note that k(A)k1 . We will use these facts in the paper without explicitly mentioning them. For a positive vector x 2 IR n , we write Generally, we use capital letters for matrices, lower case ones for vectors, and Greek letters for scalars. For convenience, we let Let svec be an isometry identifying S m with IR m so that K smat be the inverse of svec. Given any G 2 S m , we let G m to be the unique symmetric matrix such that It is easy to see that if G is positive denite, then G G is positive denite, and (G G 1=2 G 1=2 . If G is nonsingular, then (G Throughout, we make the following assumptions: A1. is a convex subset of S m . A2. where A3. contains a full dimensional ball of radius > 0. That is, there exists Y c that fY Note that Assumption A2 is made for convenience. It can be satised by scaling if the original set ^ is bounded. That is, suppose there exists a constant > 0 such that for all . Then the scaled set = fY= 2 A multiple-cut analytic center cutting plane method We rst dene the analytic center and then propose a multiple-cut analytic center cutting plane method at the end of this section. , be all the cuts dening the kth working set k . Dene Then the set k can be represented as We dene the following potential function on the set and denote The unique minimizer of k (Y ) over k is known as the analytic center of k . It is easy to see that the analytic center of the initial working set 0 is I=2; where I is the identity matrix. As a matter of fact, Y Y The minimum of 0 (Y ) must satisfy 1 (Y It is known [7, Proposition 5.4.5] that k is a strongly 1-self-concordant function on and diag (s), and should be the mm matrix within the round brackets. However, we have identied that mm matrix with a vector in IR m through the linear isometry svec. The optimality conditions for minimizing k are: denotes the vector of ones) I Y 0; Z; V 0; s; x > 0: With a slight abuse of language, we also call the solution ( V ) of (2.1) the analytic center of k . Denition 2.1 Given a point (Y; s; We call (Y; s; x; Z; V ) an -approximate (analytic) center of all the linear equalities in (2.1) are satised, and x; s > 0, Z; V 0. Obviously, a 0-approximate center is exactly the analytic center of . Denition 2.2 Given Y It was shown [8] that the following lemma holds. Lemma 2.3 Given Y We have Remark. Given Y the minimizer For such a Y , we will call Y an -approximate center of k in the sense that the point (Y; s; x Y is an -approximate center. We will now describe our algorithm. A multiple-cut analytic center cutting plane algorithm. 3=2), and pick - 2 (; 1). Set Let 0 be the initial working set and let Y be the initial point. Step 1. At the k-th iteration, call the oracle to either conrm that Y k is a feasible point of or return p k matrices A n k Otherwise, construct the new working set Step 2. Find a point ~ Y in the interior of (discussed in section 3). Step 3. (Recentering) Starting with the point Y in Step 2, perform the dual Newton method: 3.1. If - k+1 (Y ) < , set Y to Step 1. 3.2. Otherwise, Set smat where is determined as follows: if - k+1 (Y ) -, . Go to Step 3.1. 3 Restoration of centrality In our cutting plane algorithm, approximate analytic centers are found by using the dual Newton method. Our aim in this section is to estimate the number of Newton steps required to nd an approximate analytic center for a newly constructed working set. We do so by estimating the amount of potential value we should reduce for the new set. The mechanics are as follows. Since the potential function is 1-self-concordant, each Newton step can reduce the potential function by a constant amount. Thus to estimate the number of Newton steps needed to nd an approximate analytic center for a new working set, all we need is to estimate the amount of potential value we should reduce for the new set. To nd an approximate analytic center for a new working set, ideally, we would want the Newton method to start with the preceding approximate analytic center Y k . However, Y k is not in the interior of the new working set k+1 since the new cuts pass through this point. Thus our immediate task is to nd an interior point in k+1 , and then use this point as the starting point for the Newton method. Let n k be the number of cuts dening the set k . Suppose that p k new cuts are added to form the new set . Recall that Then the sets and k+1 can be written as Suppose is an -approximate center with < 1 3=2. (Note that by lemma We will now construct a point ( ~ s; ~ that is in the interior of . To this end, consider the following convex minimization problem: Evidently, the above problem has a unique solution that is also the unique solution to the KKT-conditions: Let (~!; ~ ) be an approximate solution of the above KKT conditions where (3.1a) is satised exactly and maxfj2p k ~ ~ 1=2. Note that in this case, ~ Note that to nd such a pair (~!; ~ ), we can apply Newton method to (3.1a) and (3.1b), where the computational work for each Newton iteration is O(p 3 In general, this constitutes only a very small fraction of the total computational work involved in nding an approximate analytic center for k+1 . In order not to lengthen the paper unnecessarily, we shall not establish the complexity of the Newton method for nding (~!; ~ ) in this paper. Interested reader can refer to [3] for such results. ~ ~ ~ ~ We refer the reader to [3] for an illuminating discussion on the motivation for considering the optimization problem (3.1a){(3.1b) in constructing the strictly interior point of above. It is readily shown that the following result holds: F Lemma 3.1 For any vector the following inequality holds: Proof. Refer to [11]. Lemma 3.2 Suppose (Y is an -approximate center with < 1. Then the following inequalities hold: Proof. We shall omit the proof of the rst equality as it is easy. Now we proceed with the proof of the second one. We have where we have used a theorem of Ostrowski [4, p. 225] in the second equality above, and i 's are scalars such that min (Z 1=2 Noting that max (Z 1=2 proved the required inequality. The last inequality in the lemma can be proven similarly. Theorem 3.3 The point ( ~ s; ~ constructed in (3.3){(3.4) satises the last three conditions in (2.1). Proof. First, we show that ~ Y I. We have since k(S k 3=2 < 1 from (3.7). On the other hand, we also have ~ since kY 1=2 3=2 < 1. The fact that ~ Y I can be shown similarly. Furthermore where we used the fact that from (3.1a), B T . Next we show that ~ x > 0 and ~ We have ~ since by lemma 3.2, Furthermore, Up to this point, we have succeeded in nding an interior point of k+1 that is derived from Y k . Our next task is to estimate the potential value of the new point in . Lemma 3.4 Suppose - k (Y k ) . Then the potential value k+1 ( ~ Y ) satises the following inequality Proof. Let ~ Y and U We have Note that we used the fact that d B T . Now e (U 1=2 where Note that e T By applying lemma 3.1 to (3.10), we have ~ Note that in the last second inequality above, we used the Cauchy inequality to derive the ~ !. Substituting the result in (3.12) into (3.9), we prove the lemma. >From lemma 3.4, we see that the upper bound for the dual potential value k+1 ( ~ the term ln ~ . If we were to consider the dual potential value alone, then nding an upper bound for ln ~ is necessary. But we have found that nding a tight upper bound for this term is di-cult. As a result, we have decided to consider the primal-dual potential value for which nding an upper bound for ln ~ is not necessary. To this end, let us dene the primal potential function associated with k . For any k (x; Z; V ++ that satises the primal potential of (x; Z; V ) is dened by The primal-dual potential function associated with k is We should emphasis that the primal-dual potential function is introduced solely for the purpose of estimating the potential value of ( ~ V ). It is not needed in our cutting plane algorithm described in section 2. Now we shall proceed to establish an analog of lemma 3.4 for the primal potential function. Before doing that, we need the following lemma. Lemma 3.5 For the directions (x; Z;V ) given in (3.4), the following inequality holds: Proof. Noting that !, we have d T ~ Thus (Y 1=2 (I U 1=2 F F Note that in the last inequality above, we used (3.7) and the fact that Lemma 3.6 For the point (~x; ~ constructed in (3.6), the following inequality holds: Proof. We have where Note that e T by lemma 3.2, F F By lemma 3.1 and (3.17), we get from (3.16), By applying lemma 3.5 and (3.7), we prove the lemma. The next lemma is an analog of lemma 3.4 for the primal-dual potential function. Lemma is an -approximate center with < 1 3=2. Then where Proof. Combining the results in lemmas 3.4 and 3.6, we have Note that ~ ~ ~ ~ By substituting (3.20) into (3.19), the lemma is proven. With lemma 3.7, we can nally establish an explicitly known upper bounded for the primal-dual potential value k+1 ( ~ Theorem 3.8 Suppose (Y is an -approximate center with < 1 3=2. Then where () is the constant given in (3.18). Proof. We have It is readily shown that Next we need to get an upper bound for the term k (Y k (3.22). By following the proof of lemma 2.1 in [1] and using the quadratic convergence result in [8], it is readily shown that Similarly, it can be shown that Combining (3.24) and (3.25), we get By putting the results in lemma 3.7, (3.23) and (3.26) into (3.22), the theorem is proven. With the estimate of k+1 ( ~ theorem 3.8, we are now ready to estimate the number of dual Newton steps required to nd an approximate analytic center for k+1 by using the point ~ Y as the initial point. Theorem 3.9 Given an -approximate center Y k of k with < 1 3=2. The total number of dual Newton steps required to nd an -approximate center Y k+1 of k+1 is O (p k where the constant O(1) is independent of k. Proof. By theorem 2.2.3 in [7], each dual Newton step reduces k+1 by a positive constant long as a point ^ Y with - not yet found, while keeping the primal iterate xed. Now, starting at ( ~ V ), the total value of k+1 needed to be reduced is not more than k+1 ( ~ theorem 3.8 implies that at most" Newton steps are required to reach a point ^ Y with - Y onwards, by Lemma 4.3 in [8], quadratic convergence can be achieved, so it needs at most ln(ln( additional full Newton steps to nd a point Y k+1 satisfying - k+1 (Y k+1 ) . (We can choose for example, 4 Potential changes and Complexity Recall that is an -approximate analytic center of k with < 1 3=2. Let Then Let Y k and Y k+1 be the analytic centers of and where In this section, we estimate the amount that the dual potential will increase when the working set change from to . To this end, we rst establish a lemma that is an extension of a result in [10]. Lemma 4.1 Suppose n; p are positive integers, and is a positive n-vector with e T Then for any positive constant , the following inequality holds: Y where is a positive constant no greater than 1:3 Proof. We need only to consider the case where n 2 as the inequality holds trivially when 1. Consider the maximization problem: Y It is shown in [10] that the maximizer has the form and p=2 Thus Y Y 1=p Y 1=p Y 1=p where 1+1=p e 1=(p+1) e Lemma 4.2 Suppose Y k is an approximate analytic center of 3=2. Then where is a constant depending only on . Proof. For simplicity, we will drop the subscripts k and k our notations in this proof, and denote for example, and by and Let Y , Y+ , and Y Let U 1=2 U 1=2 ]; Note that G G T . First, we establish an upper bound for ln We have G Thus By part (iii) of theorem 2.2.2 in [7], we have [1 3-(Y )] 1=3 3: Thus Hence Y pln and the desired upper bound is established. Now observe that Y Y det det Y det det U Using the bound in (4.3), we have Y det det Y det det U The inequality (4.2) follows once we have shown that Y det det Y det det U Note that A U U e U 1=2 U 1=2 )C A and by using (2.1), we have U 1=2 Z Z U Z U By Lemma 4.1, (4.5) is proved. The complexity analysis is based on the following idea. For the sequence of working set k , we can establish upper and lower bounds on ). The upper bound is approximately n k which is a consequence of the assumption that contains a ball of radius and the fact that k is dened by n k cuts. The lower bound is obtained by estimating which is a consequence of Lemma 4.2. A sophisticated estimation of r k gives rise to a lower bound that is proportional to n k ln(n k =m 3 ). The algorithm must terminate before the lower and upper bounds con ict each other. We rst establish an upper bound for Lemma 4.3 Let k be dened by n k linear inequalities and the positive semidenite constraint. Suppose Assumptions A1-A3 hold. Then Proof. Assumptions A1-A3 imply that there exists a point Y c 2 , such that (i) All eigenvalues of Y c and I Y c are greater than or equal to ; (ii) For any A 2 S m with We will brie y describe how to prove (Y c ) e before continuing with the proof of the lemma. Suppose j is an eigenvalue of Y c and v j is a corresponding unit eigenvector. Consider the . Since this matrix has a zero eigenvalue, it lies on the boundary of 0 and by Assumption A3, we have The fact that (Y c ) (1 )e can be proven similarly. Now we continue with the proof of the lemma. Since Noting that Y Y we have the desired inequality. Now we turn to nding a lower bound for r i2 Obviously, we need to estimate r i for each i. We rst seek to bound i by D 1 dened as follows. Let I is the identity matrix of order m. For let Lemma 4.4 Let A n i +j (with be the cuts generated from the approximate analytic center Y ii , k. For any point Yk , let Then In particular, Proof. We rst estimate s n i +j . We have The last inequality holds because by Assumption A2, I implying that e (Y Next, let Note that in deriving (4.8), we used the fact that S i for each i, and that In our complexity analysis, we will make the following assumptions. Assumption A4. p max m, where p Assumption A5. Let There exists a xed constant 1 such that for each Assumption A4 is made for technical reason. It is used in proof of lemma 4.5. Such an assumption also appeared in the papers [3] and [10]. Note that Assumption A4 can be relaxed to p max O(m). But for simplicity, we xed the constant at 1. Note that Assumption A5 holds trivially with . For the special case where a single cut is used in each iteration, it holds with 1. Thus by xing at an intermediate value between 1 and p max , we admit only cuts that are su-ciently good in the sense that the matrix have too many small eigenvalues. Of course, one may not want to x at the extreme value 1 since then the criterion is likely to reject most of the cuts unless there are many mutually orthogonal (with respect to The main advantage of having Assumption A5 is that in each oracle call, we have an objective criterion to select only cuts that are useful among possibly a large number of ineective cuts. In this way, the number of cuts added in each iteration will not be unnecessarily large, and hence the computational time in each iteration will not grow as rapidly compared to the case where the cuts are admitted unchecked. The choice of in practice would depend on the problem at hand. It should dynamically be adjusted as information on the quality of the cuts are obtained as the cutting algorithm progresses. If the choice of is too stringent and many good cuts are rejected, then we can progressively increase its value so that more good cuts are selected. However, without a priori information on the quality of the cuts, we propose to choose to be a small constant, say 5, based on the following empirical observation we made from numerical experiments. We conducted numerical experiments on random matrices of the form V T V mp , for 260. The elements of V are drawn independently from the standard normal distribution. We computed the ratio between the largest eigenvalue of V T V and Tr(V T V )=p for each V , and found that these ratios are less than 2 for all the 3510 cases we tested. Now let us continue with our complexity analysis. Let Since we have Next, we establish an upper bound for the right hand side of the above inequality. Its proof is modeled after that of [10, lemma 3.5]. However, we simplied the proof by considering all the cuts simultaneously instead of handling them one by one as in [10]. Lemma 4.5 9m 8m Proof. From the equation Y we have 9m where we used the fact that max (B T and the inequality ln(1 +x) 8x=9 for 0 x 1=8. We also made use of Assumption A4 that p i m. >From (4.10), it follows immediately that But m+m implying that 8m Combining (4.11) and (4.12), the lemma is proved. With the above lemma, we can now formally state a lower bound for Lemma 4.6 Suppose Assumptions A1{A5 hold. Then 8m where is the constant appeared in (4.2). Proof. The proof is similar to that of theorem 10 in [10] by making use of (4.9) and Lemma 4.5. We will next estimate the number oracle calls required to nd a feasible point of . Lemma 4.7 Suppose the Assumptions A1{A4 hold, and p max m. Then the analytic center cutting plane method stops with a feasible before k violates the following inequality 8m exp Proof. From Lemmas 4.3 and 4.6, we have 8m Thus, the algorithm must terminate before k violates the above inequality, i.e., the algorithm must terminate before k violates the following inequality: 8m Since the algorithm must terminate before k violates the inequality in the lemma. Theorem 4.8 Suppose the Assumptions A1-A4 hold, and p max m. Then the analytic center cutting plane method terminates in at most O (m 3 p max steps, where the notation O means that lower order terms are ignored. The total number of cuts added is not more than O (m 3 Proof. Ignoring lower order terms (assuming k m) and by the assumption that is a constant independent of p max , the above lemma implies that the algorithm stops as soon as k For large k, ln n k is negligible compared to n k , hence the algorithm requires at most cuts. By Theorem 3.9, the total number of Newton steps is O The theorem is proved. For feasibility problems in IR m m should be replaced by m in Lemma 4.7. Thus the complexity bound is O(m 2 p max for the number of required Newton steps. This bound is better than the bounds obtained in [2], [5], and [10]. Acknowledgement We thank the referees for their constructive comments that greatly help to improve the paper. --R Complexity analysis of an interior cutting plane method for convex feasibility problems Multiple cuts in the analytic center cutting plane method Convex nondi Matrix Analysis Analysis of a cutting plane method that uses weighted analytic center and multiple cuts Cutting plane algorithms from analytic centers: e-ciency estimates An analytic center cutting plane method for semide A potential reduction algorithm allowing column generation Complexity analysis of the analytic center cutting plane method that uses multiple cuts Interior Point Algorithms: Theory and Analysis --TR
multiple cuts;analytic center;cutting plane methods;semidefinite programming
589002
A Superlinearly Convergent Sequential Quadratically Constrained Quadratic Programming Algorithm for Degenerate Nonlinear Programming.
We present an algorithm that achieves superlinear convergence for nonlinear programs satisfying the Mangasarian--Fromovitz constraint qualification and the quadratic growth condition. This convergence result is obtained despite the potential lack of a locally convex augmented Lagrangian. The algorithm solves a succession of subproblems that have quadratic objectives and quadratic constraints, both possibly nonconvex. By the use of a trust-region constraint we guarantee that any stationary point of the subproblem induces superlinear convergence, which avoids the problem of computing a global minimum. We compare this algorithm with sequential quadratic programming algorithms on several degenerate nonlinear programs.
Introduction . Recently, there has been renewed interest in analyzing and modifying the algorithms for constrained nonlinear optimization for cases where the traditional regularity conditions do not hold [5, 12, 11, 20, 24, 23]. This research has been motivated by the fact that large-scale nonlinear programming problems tend to be almost degenerate (have large condition numbers for the Jacobian of the active constraints). It is therefore important to define algorithms that are as little dependent as possible of the ill-conditioning of the constraints. In this work, we term as degenerate those nonlinear programs (NLPs) for which the gradients of the active constraints are linearly dependent. In this case there may be several feasible Lagrange multipliers. Many of the previous analysis and rate of convergence results for degenerate NLP [5, 12, 11, 20, 24, 23] are based on the validity of some second-order conditions. These are essentially equivalent to the condition in unconstrained optimization that, for a critical point of a function f(x) to be a local minimum, f xx - 0 is a necessary condition and f xx - 0 is a sufficient condition. Here - is the positive semidefinite ordering. The place of f xx in constrained optimization is taken for these conditions by L xx , the Hessian of the Lagrangian, which is now required to be positive definite on the critical cone for one or all of the Lagrange multipliers [7, 21]. This work differs from previous approaches in that we assume only that 1. At a local solution x of the constrained nonlinear program, the first-order Mangasarian-Fromovitz [18, 17] constraint qualification holds. 2. The quadratic growth condition (QG) [6, 15] is satisfied: for some oe ? 0 and all x feasible in a neighborhood of x . 3. The data of the problem are twice continuously differentiable. These assumptions are equivalent to a weaker form of the second-order sufficient conditions [14, 6] which does not require the positive semidefiniteness of the Hessian of the Lagrangian on the entire critical cone. In a recent a paper [2] it has been shown Thackeray 301, Department of Mathematics, University of Pittsburgh, Pittsburgh, PA 15213 (anitescu@math.pitt.edu). Part of this work was completed while the author was the Wilkinson Fellow at the Mathematics and Computer Science Division, Argonne National Laboratory. This work was supported by the Mathematical, Information, and Computational Sciences Division sub-program of the Office of Advanced Scientific Computing, U.S. Department of Energy, under Contract W-31-109-Eng-38. This work was also supported by award DMS-9973071 of the National Science Foundation. that these conditions guarantee that x is an isolated stationary point and that a steepest-descent like algorithm induces linear convergence to x . The framework used here accommodates even problems for which no locally convex augmented Lagrangian exists [2], which do not satisfy the assumptions of most other convergence results [5, 12, 11, 20, 24]. In this paper we define an algorithm that is superlinearly convergent even in the very general conditions outlined above. The trade-off is that the subproblems to be solved are more complex than a quadratic program. The algorithm can be justified by a particular perspective on Newton's method for unconstrained optimization. If f(x) is the function to be minimized without constraints then sufficiently close to a solution x Newton's direction, d, is a solution of the quadratic minimization problem. min xx f(x)d: The term f(x) is constant for this minimization problem, but we include it to emphasize that we can regard d as a solution of the second-order approximation to the problem. If we have an inequality constrained nonlinear program, subject to g i its second-order approximation at x is the following problem min xx f(x)d subject to g(x) +r x We call such a problem a quadratically constrained quadratic problem (QCQP). To ensure that the problem is bounded even for x far from the solution x , we add to the problem a trust-region constraint, which is also quadratic: The problem is generally not convex and thus finding the global optimum may be a difficult problem. Also, the trust-region constraint may interfere with the order of convergence. However, we show that for x close to x and for sufficiently small but fixed: 1. The trust region constraint is inactive at any stationary point of the QCQP. 2. Any stationary point d of the QCQP used as a progress direction induces superlinear convergence. Therefore, finding a local solution to the QCQP is sufficient to induce superlinear convergence of the iterates, which considerably reduces the conceptual complexity of a sequential QCQP (SQCQP) algorithm. Note that the QCQP subproblem is identical to the one used in [16], although the analysis conditions in this work are more general. The paper is structured as follows. In Subsection 1.1 we discuss the different conditions defining a stationary point of a nonlinear program and the quadratic growth condition. Section 2 characterizes stationary points of the second-order approximation (QCQP) of the nonlinear program at x . We show that, if the trust-region constraint defines a sufficiently small region then the Mangasarian-Fromovitz constraint qualification is satisfied at any feasible point and is the unique stationary point of the QCQP. As a result, in Section 3 we prove that, for x sufficiently close to x , the trust-region constraint is inactive at any stationary point of QCQP and we prove the superlinear convergence of the SQCQP algorithm. We conclude with Section 4, where we briefly discuss possible approaches to solving the QCQP subproblem. 1.1. Previous Work, Framework, and Notations. We deal with the NLP problem f(x) subject to g(x) - 0; are twice continuously differentiable. We call x a stationary point if the Fritz John conditions conditions hold: There exist Here L is the Lagrangian function A local solution x of (1.2) is a stationary point [19]. If certain regularity conditions hold at x (discussed below), then there exists - 0 such that x with - In that case (1.3) are referred to as the KKT (Karush- Kuhn-Tucker) conditions [3, 4, 8] and - are referred to as the Lagrange multipliers. For that case, which is the one that most oftenly appears in this work, we define the Lagrangian as and the Karush-Kuhn-Tucker conditions become r x L(x; Since our analysis is limited to a neighborhood of a point x that is a strict local minimum, we assume that all constraints are active at x , or g(x Such a situation can be obtained by choosing a sufficiently small trust-region and simply dropping the constraints i for which g i (x since this relationship holds in an entire neighborhood of x . This does not reduce the generality of our results, but it simplifies the notation because now we do not have to refer separately to the active set. The regularity condition, or constraint qualification, ensures that a linear approximation of the feasible set in the neighborhood of x captures the geometry of the feasible set. Often in local convergence analysis of constrained optimization algo- rithms, it is assumed that the constraint gradients r x are linearly independent, so that the Lagrange multiplier in (1.6) is unique. We assume instead the Mangasarian-Fromovitz constraint qualification (MFCQ) [18, 17]: It is well known [9] that MFCQ is equivalent to boundedness of the set M(x ) of Lagrange multipliers that satisfy (1.6), that is, Note that M(x ) is certainly polyhedral in any case. Another condition equivalent to MFCQ (1.7) is [10] such that The critical cone at x is [7, 22] \Phi We briefly review some of the second-order conditions in the literature. In the framework of [7], the second-order sufficient conditions for x to be an isolated local solution of (1.2) are [7, 8]: If these conditions hold at x for some - , then the quadratic growth condition is satisfied, irrespective of the validity of the first-order constraint qualification [7, 8]. An important consequence of the condition (1.11) is that x is a local minimum of the augmented Lagrangian for a sufficiently large constant c. A refinement of the second-order conditions was introduced in [14]. In the presence of MFCQ, those conditions require that Further analysis shows that, in presence of MFCQ, these conditions are necessary and sufficient for the quadratic growth condition to hold [6, 14, 15, 22]. If the condition (1.12) holds, but (1.11) does not, then there may be no augmented Lagrangian with a positive semidefinite Hessian, as it is shown with an example in [2]. This is an interesting aspect since it invalidates the usual working assumption of Lagrange multiplier methods [4]. It also shows that the analysis in this paper is done without assuming the existence of an augmented Lagrangian that has x as an unconstrained minimum. In our analysis we use the L1 nondifferentiable exact penalty function: If the MFCQ (1.7) conditions hold at x , then the quadratic growth condition (1.1) and the second order conditions (1.12) are each equivalent to the following condition [6] for some oe ? 0 and all x in a neighborhood of x . For some function h : we denote by c 1h , c 2h bounds depending on the first and second derivatives of h. The positive and negative parts of h(x) are h With this notation Also, in our notation, r x g i (x), - and r x g(x)- are column vectors. In this work we need to estimate distances to sets described by linear constraints: where M eq and M in are n eq \Theta n and, respectively, n in \Theta n matrices and q eq and q in are respectively, n in dimensional vectors. By Hoffman's Lemma [13], if P 6= ;, there exists c P ? 0 such that where by D( ~ d; P) we denoted the distance from ~ d to the set P. This result allows us to relate the distance from a point ~ d to a polyhedral set in terms of the infeasibility of ~ d in the representation (1.15). 2. Stationary Points of Quadratically Constrained Quadratic Pro- grams. In this section we investigate the stationary points of the quadratically constrained quadratic program min d2IR n a T d subject to b T are n \Theta n symmetric matrices and a 2 m. We denote this program by TRQCQP(fl). Our assumptions concerning (2.1) are: 1. At 2. The quadratic growth condition is satisfied near 0: There exists fl 0 and oe 1 ? 0 such that a A local solution of (2.1) is clearly The aim of this section is to show that under assumptions (2.3) and (2.2), there exists is the only stationary point of TRQCQP(fl) (2.1), for any 0 - As a consequence any algorithm that reaches a stationary point of finds its global optimum. The results from [2] ensure that an isolated stationary point of TRQCQP(fl) (2.1). However, the developments of this section are necessary to ensure that additional stationary points are not introduced by the trust region constraint. The proof has the following steps, each stated for sufficiently small fl. Lemma 2.4 proves that MFCQ (1.7) is satisfied for all stationary points ~ d of (2.1). Therefore, at any stationary point there exist Lagrange multipliers that satisfy (1.6); ultimately implies that for any Lagrange multiplier - at a stationary point ~ d of (2.1) there exists a sufficiently close Lagrange multiplier at active subset is included in the active subset of -. This leads to the identity (- ~ ~ which helps bound above the variations in the objective function of (2.1) in the proof of Theorem 2.7. proves that the multiplier of the trust-region constraint is bounded above. This in turn implies Lemma 2.6: the Lagrange multipliers of all potential stationary points are uniformly bounded. ffl Theorem 2.7, the main result of this section, proves that ~ is the unique stationary point of (2.1). Subsection 2.1 contains additional results implied by Hoffman's Lemma (1.16), which are used in Section 3. 2.1. Sensitivity results for Lagrange Multipliers. An immediate consequence of MFCQ (2.2) is that the set of Lagrange Multipliers of TRQCQP(fl) (2.1) at is nonempty and bounded. Lemma 2.1. There exists c M ? 0 such that, for any w 2 IR n , and for any - 2 IR satisfying a there exists a - 2 M such that Proof Follows by direct application of Hoffman's Lemma (1.16), after using that jjwjj 1 - jjwjj. \Pi Lemma 2.2. There exists j ? 0 such that for all w 2 IR n with jjwjj - j and any satisfying a there exists - 2 M such that - Proof Assume the contrary: For any k 2 IN , there exists w k 2 IR n , such that k and there exists - k satisfying a and an index set I k ae f1; mg, such that - k I From Lemma 2.1, D(- is a compact set, and the set of subsets of f1; is finite, there exists a subsequence such that I From our assumptions - I 6= 0, 8- 2 M . On the other hand, since - kq I I and - kq ! - we must have - I which is a contradiction. The proof is complete. Lemma 2.3. There exists c M ? 0 and j ? 0 such that for any w 2 IR n with any - satisfying a there exists - 2 M with and such that - Proof Let j be the quantity defined by Lemma 2.2. Let I ae f1; that there exists a - satisfying (2.5) and - I = 0. Lemma 2.2 implies that there exists such that - I = 0. Let M I be the set of - 2 IR m such that From Lemma 2.2, M I is not empty. From Hoffman's Lemma (1.16), there exists c M I ? 0 such that, for all - 2 IR m , we have From Lemma 2.1 choose - 2 M such that From the definition of M (2.4) we have that Thus, from (2.7) we must have I I We also have from our choice of - (2.8) that we thus have conjunction with the preceding inequality and (2.9) implies that Hence from (2.8) and the preceding inequality we have that The conclusion now follows after taking I 2.2. Stationary Points of Quadratically Constrained Quadratic Pro- grams. In this section we analyze the stationary points of TRQCQP(fl) (2.1) for sufficiently small values of the parameter fl. We choose fl 00 1 such that are the quantities appearing in MFCQ (2.2) with choose which guarantees that whenever jjdjj - fl 1 , both (2.10) and the quadratic growth condition Lemma 2.4. There exists (1.7) at all its stationary points d with fl such that The important consequence of this lemma is that Lagrange multipliers exist at any stationary point of TRQCQP(fl) (2.1). Proof Take the quadratically constrained quadratic program min d2IR n d T d subject to b T with global solution as well as the quadratic growth condition (1.1). From [2], is an isolated stationary point of (2.12). Therefore there exists a fl 0 such that the only stationary point d of (2.12) that satisfies d T d - (fl 0 Take now g. Assume that there exists fl, MFCQ (1.7) is not satisfied at some stationary point - d of TRQCQP(fl) (2.1). From and (1.3) it follows that there exists - not both equal to 0, such that If - would imply or, after multiplying with p from (2.10) we get d) which implies - contradiction with the assumption that not both - 0 and - are and from (2.13) we get - dividing with - (b (b T But this means that - d 6= 0 is a stationary point of (2.12) with a Lagrange multiplier , which contradicts the properties of our choice of fl 2 . The proof is complete. \Pi Lemma 2.5. Consider the following quadratically constrained quadratic program subject to \Gamma i Then there exists the only stationary point of (2.15) that satisfies jjdjj - fl 3 is Proof Choose fl 0 g. From (2.11) this implies that for all d with 3 the quadratic growth condition (2.3) and (2.10) holds. Also, from Lemma 2.4, MFCQ (1.7) holds at any stationary point of (2.15). Take ~ d 6= 0 a feasible point of (2.15). We now estimate the variation of the constraints and objective function in a specific direction from ~ d, in order to decide under what conditions ~ can be a stationary point of (2.15). Let the active set at ~ d be We estimate the first-order behavior of \Gamma i (d) in the direction \Gamma ~ is the vector from (2.10) and fi - 0. For d we get ~ d) T (\Gamma ~ \Gammab T ~ ~ ~ d) \Gammab T ~ d) d where we used (2.10) and that, from (2.16), if d then \Gammab T ~ ~ For the objective function we have that (r d \Psi( ~ ~ d) T (\Gamma ~ \Gammaa T ~ d T A ~ ~ d T ~ ~ ~ d T ~ ~ d T ~ d T A ~ ~ where we used the quadratic growth condition (2.3). Choose now d 2: Assume that ~ d 6= 0, d . Using that d T A ~ ~ d d d d d d d d d where we used that from our choice of fl 00 3 (2.21) and since d 3 we have c fi d and c fi d We also used the definition of c ffi (2.22) and that c 1 - c ffi . Using (2.23) in (2.18) we get r d \Psi( ~ d) T (\Gamma ~ d d d Using (2.19) and (2.20) in (2.17) we get for all d d) T (\Gamma ~ d d d d 0: From (2.25) and (2.24) we get that if ~ and ~ 3 (2.21) then there exists a direction ~ that produces strict decreases in the objective function and the active constraints. Therefore ~ d cannot be a stationary point of (2.15). Otherwise (1.3) implies that there exist the multipliers - m , not all of them of 0 such that d) d From (2.25) and (2.24) we get, after multiplying with ~ that d) T ~ d d) T ~ which is a contradiction that proves the lemma with c ffi defined in (2.22) and Lemma 2.6. There exists d with d stationary point of TRQCQP(fl) (2.1) with Lagrange multipliers - 2 IR Proof We take defined in (2.11), fl 2 is the quantity from Lemma 2.4 and fl 3 is the quantity from Lemma 2.5. Lemma 2.4 ensures that the Lagrange multipliers exist at any stationary point of TRQCQP(fl) (2.1). Assume the contrary of the conclusion of the Lemma: For any k 2 IN , there exists ~ d k a stationary point of TRQCQP(fl) (2.1) with multipliers and (1.6). In particular, a +A ~ ~ By Lemma 2.5 since must have c k can choose - such that for a subsequence k q , q !1, we have lim q!1 with jj- jj d , where d We can now divide through (2.27) with take the limit as and d ~ multiply with p and use (2.10) and the fact that jj- jj to get ~ d which is a contradiction. This proves the lemma. \Pi Theorem 2.7. There exists fl 5 ? 0, such that, for any fl such that TRQCQP(fl) (2.1) has the unique stationary point Proof Choose ae oe where j is the quantity from Lemma 2.3, c ffi is the quantity from Lemma 2.5, -1 is the quantity from Lemma 2.6 and 4, are the bounds on the trust regions that ensure that all preceding results hold. Let ~ d 6= 0 be a stationary point of TRQCQP(fl) (2.1) with Lemma 2.4 TRQCQP(fl) (2.1) satisfies MFCQ (1.7) at ~ d. Therefore there exist the Lagrange multipliers - 0, c 1 - 0 which, together with ~ d satisfy (1.6), or a d) ~ d) ~ d T ~ ~ ~ d T ~ Since d applies to give that c 1 - c ffi . Since d 4 we have that jj-jj 1 -1 from Lemma 2.6. We define ~ ~ d: After applying the triangle inequality and using (2.28) we have that d ~ d ~ d d d For the last inequality, c- and d results in jjwjj - j. From (2.30) and (2.31) we have that a This implies, from Lemma 2.3, that there exists - 2 M ( a Lagrange multiplier for TRQCQP(fl) (2.1) at (2. a Adding the last equality to the first equation in (2.30), dividing by 2 and multiplying with ~ d T we obtain a T ~ dA ~ ~ ~ d) d 0: We now use the identity u well as the fact that (- ~ ~ from (2.33) and (2.30) to obtain d +2 ~ dA ~ d +2 ~ ~ ~ d) d a T ~ dA ~ (- ~ d) d which results in a T ~ dA ~ d (- ~ Since ~ d is feasible for TRQCQP(fl) (2.1) and since d the quadratic growth condition (2.3) holds to give that a T ~ dA ~ d . Define From (2.33), (2.31) and (2.32) we have jj- d Using all these bounds in (2.34), together with the arithmetic-quadratic mean inequality we get d - a T ~ dA ~ d (- d) -4 d d d Since d our assumption, we obtain, after dividing through the previous inequality with d that d Choose now ae oe From (2.35) it follows that the unique stationary point of TRQCQP(fl) (2.1) with The proof is complete. \Pi 3. Sequential Quadratically Constrained Quadratic Programming. In this section, we introduce the sequential quadratically constrained quadratic programming algorithm. We prove that under the conditions set forth in the introduction, the algorithm induces superlinear convergence. Since our main interest is the rate of convergence of the algorithm, we do not address global convergence issues. We consider the following form of the algorithm: 1. Choose a starting point x k , 2. Let stationary point of xx f(x)d subject to g i (x) +r x 3. Take x restart. At every step, the algorithm solves a problem with quadratic constraints and a quadratic objective, none of which are assumed to be convex. We name the above algorithm sequential quadratically constrained quadratic programming or SQCQP. As outlined in Subsection 1.1, we assume without loss of generality that g i (x eventually considering a sufficiently small trust-region, and that the quadratic growth condition (1.1) and MFCQ (1.7) hold at a local solution x of the nonlinear program (1.2). From [14, 6] these conditions are equivalent to MFCQ (1.7) and (1.12), which are expressed only in terms of the derivatives of the data up to the second order. We show that (3.1) is feasible for fl fixed and x in some neighborhood of x . Since it is also bounded, a stationary point must exist. Due to the fact that it captures the entire information up to second order for (1.2) at x , the quadratically constrained quadratic program xx f(x )d subject to r x g(x ) satisfies MFCQ (1.7) and (1.12) at As a result of [14, 6] it follows that (3.2) satisfies MFCQ (2.2) and the quadratic growth condition (2.3). Therefore, all the results from Section 2 apply for (3.2). We follow a line of proof similar to the one in Section 2. ffl Theorem 3.1 proves that MFCQ (1.7) is satisfied by (3.1) in a neighborhood of x and that the trust-region constraint is inactive at any stationary point d of (3.1). Corollary 3.2 further insures that in a neighborhood of x , the Lagrange multipliers of (3.1) are uniformly bounded. ultimately implies that for any Lagrange multiplier - at a stationary point d of (3.1) at there exists a sufficiently close Lagrange multiplier - at whose active subset is included in the active subset of -. This in turn leads to the conclusions of Lemma 3.4 that (- i +- is a stationary point of (3.1). This helps bound above the variations in the objective function of (3.1) in the proof of Theorem 3.5. Theorems 3.5 and 3.6 prove the superlinear convergence of a sequence x initiated sufficiently close to x , where d k is any stationary point of (3. Theorem 3.1. There exists fl 6 ? 0 and a neighborhood N fl 6 any fl with there exists a neighborhood N fl (x ) of x such that (i) The QCQP (3.1) is feasible for any x 2 N fl (x ). (ii) For any x 2 N fl 6 any d with jjdjj - fl 6 we have are the quantities entering the definition of MFCQ (1.7). (iii) For any sequence x k 2 N fl (x and with ~ d k a stationary point of (3.1) at must have ~ (iv) The constraint d T d - fl 2 is inactive for any x 2 N fl (x ) and d stationary point of (3.1). Proof Since (3.2) satisfies MFCQ (2.2) and the quadratic growth condition (2.3) at 0, from Theorem 2.7 there exists 6 , such that, for any 0 ! 6 , ~ is the only stationary point of (3.2). Choose now fl such that 6 . Since (3.2) satisfies MFCQ (2.2), then, from [21], for any sufficiently small perturbation of (3.2) we still obtain a feasible nonlinear program. We regard (3.1) as a perturbation of (3.2) and we therefore have, from the fact that f; g are twice continuously differentiable, that there exist a neighborhood N 2 such that (3.1) is feasible for any x 2 N 2 which proves part (i) as long as N fl which will be established later. We also have that, for all since from MFCQ (1.7) is a bound on the second derivatives of m. If we chose fl 00 4c2g , d with jjdjj - fl 00 6 and N fl 6 4c2g we get from the previous bound that, since now c 2g jjx \Gamma x jj - i0, c 2g jjdjj - i0, 0which shows part (ii), after defining 00g. We now choose N 3 both the conclusions of (i) and (ii) hold. In particular, for any fl 2 (0; fl 6 must have a stationary point since it is feasible and bounded. Assume now that the conclusion (iii) does not hold: There exists fl ? 0, with and a sequence x k ! x , x k 2 N 3 and the corresponding stationary points d k of (3.1) are bounded bellow sufficiently large. Since d k is a stationary point of (3.1) at must satisfy the first-order necessary conditions (1.3) for some multipliers - k Since the multipliers - 1, and the direction we can extract a subsequence k q such that x kq ! x , 1. Taking the limit as q ! 1 in (3.3) we obtain from the continuity of all data involved in terms of (x; d), that d is a stationary point of (3.2). Since d 6= 0 this contradicts the outcome of Theorem 2.7 that is valid due to our choice of fl 6 . This proves (iii). Assume now that (iv) does not hold. It then follows that there exists a sequence stationary point and such that fl. But this contradicts the conclusion of (iii) and thus there exists a neighborhood N fl that for x 2 N fl (x ) any stationary point of (3.1) satisfies d T d ! fl 2 and for which the conclusions of parts (i),(ii) and (iii) hold. The proof is complete. \Pi Corollary 3.2. Any stationary point of (3.1) satisfies the Kuhn-Tucker conditions 1.6, for any 0 - There exists -1 such that, for any x 2 N fl (x ), any stationary point d of (3.1) and any Lagrange multipliers - satisfying the Kuhn-Tucker conditions we have jj-jj 1 -1 . Proof Theorem 3.1(iv), we have that for any stationary point d, we must have jjdjj ! fl. Therefore only the can be active at a stationary point d. Then by Theorem 3.1(ii), MFCQ (1.7) is satisfied at d and thus there exist multipliers - 0 satisfying the Kuhn-Tucker conditions and in particular: Multiplying through with p we get after using the usual norm inequalities we get Since on N fl (x ) the expression from the right hand side is bounded above, there exists -1 for which the conclusion of this corollary holds. \Pi Lemma 3.3. There exists fl 7 ? 0 and a constant c ? 0 such that for any fl with there exists a neighborhood N 1 for any d a stationary point of (3.1) with the Lagrange multipliers - there exist the Lagrange multipliers at and Proof Take fl such that d be a stationary point of (3.1) with the Lagrange multipliers - 0 (which exist from Corollary (3.2)). From the Kuhn-Tucker conditions we obtain and thus Using that jjr x and that jjr x g i c 2g jjx \Gamma x jj, where c 2f and c 2g are bounds on the second derivatives of f and g, we get from (3.5) and Corollary 3.2 that We choose , where j is the quantity from Lemma 2.2. From (3.6) it follows that, for any fl - that, since jjdjj - We can therefore apply Lemma 2.2 and (3.6) to get that there exists - 2 M(x ) with the properties required, after taking is the constant from Lemma 2.3. \Pi Lemma 3.4. Let x 2 N 1 is the neighborhood obtained in Lemma 3.3. Let - be a Lagrange multiplier associated with a stationary point d at x of (3.1). Let - 2 M(x ) such that - and such that where \Theta P (d) is a continuous function that satisfies \Theta P Proof Using the first-order Taylor remainder formula [1] for g i (y) around for the fact that g i (x) is twice continuously differentiable for we obtain that is a continuous function, it follows that is a continuous function on jjwjj - fl 7 with the property that \Theta i We have that d is a stationary point of (3.1) and as a result satisfies g i Replacing w with d in (3.7) we obtain now \Theta P \Theta i (d): From the definition of \Theta i (d) we have that \Theta P is continuous and that \Theta P From the definition of P (x) (1.13), we get that, This proves point (i). Since - is such that - our hypothesis, and since d is a stationary point of (3.1) and thus satisfies the complementarity condition xx g(x)d this implies that, for xx g(x)d or, by using (3.7), and thus which completes the proof of (ii) and of the Lemma. \Pi From here on we use extensively that, for h twice continuously differentiable, we have is a continuous function with / 3h Indeed by Taylor's theorem we have that there exist continuous functions / 1 3h 3h 3h which satisfy / 1 3h 3h 3h and which in turn implies, after choosing / 2 3h 3h and using the Cauchy-Schwarz inequality, that The relation (3.9) now follows by comparing (3.9), (3.10) and (3.12) and taking 3h (z). If h were three times continuously differentiable, then would be related to the third derivative of h, from the error formula of trapezoidal integration [1], which is the origin of our subscript notation. Theorem 3.5. Let be a sequence such that x k ! x , x k 6= x . Let d k be a stationary point of (3.1) for is the quantity from Lemma 3.3. Then lim Proof Since x k ! x , the sequence x k eventually reaches N 1 means that Lemmas 3.4 and 3.3, as well as all preceding results apply for sufficiently large k. Using (3.9) we get that is a bound obtained by using (3.11) for f(x) between x k and x k . / 0 3f is a continuous function satisfying / 0 From Corollary 3.2, there exist the Lagrange multiplier - k , which, together with d k satisfies the Kuhn-Tucker conditions (1.6) for (3.1) at there exists a - \Lambdak 2 M(x ) such that Using the Kuhn-Tucker conditions (1.6) to replace r x r x f(x ) in terms of g and the Lagrange multipliers, and using the bounds that follow from Corollary 3.2, we get from (3.13) 3g is a bound obtained from applying (3.11) to g i (x), between the points x k taking the maximum among the resulting bounds. / 0 3g a continuous function satisfying / 0 We now make use of the identity ab for the terms Continuing the bounding in (3.15) we get 3f (3. We now bound all terms involving - and - . Using that and that g is twice continuously differentiable and thus we get mc c 2g ( Using that jj-jj 1 -1 from Corollary 3.2, (3.9) for g i (x) and that g i (x as well as Lemma 3.4 (ii) we obtain that Putting together the bounds from (3.16), (3.17) and (3.18) we obtain Since the bound on the right hand side is nonnegative, we can use Lemma 3.4 (i) and the quadratic growth condition (1.14) to get that oe where are continuous functions of their arguments that satisfy \Phi 1 (0; We now use that ab - to get from (3.20) that oe 0, from Theorem 3.1, there exists K 1 such that 8k - K 1 we have Taking the corresponding term to the right-hand side, we get that, 8k - K 1 , Now using the continuity of \Phi 2 and \Theta P , and that, from Theorem 3.1 (iii), d k ! 0, we get that lim oe lim 0: or that lim Using now the consequence of the triangle inequality and dividing the relation with and taking the limit, this implies that lim and thus lim Dividing (3.22) by the last limit we get that lim which proves the claim of the Theorem. \Pi Theorem 3.6. Let fl be such that is the quantity from Lemma 3.3. There exists a radius r such that for any x 2 B(x a stationary point of (3.1), then Whenever started inside B(x ; r ), the SQCQP algorithm produces a sequence x k ! x that is superlinearly convergent, lim Proof Assume the contrary: For any q 2 IN , there exists x q 6= x such that q and d q a stationary point of (3.1) such that Therefore x q ! x , and by Theorem 3.5 lim which contradicts (3.23). As a result there exists r with the properties required by the Theorem. When started with x 0 2 B(x ; r ), the SQCQP algorithm produces a sequence x , such that We can now use Theorem 3.5 to claim that lim which proves the superlinear convergence of x k to x . The proof is complete. \Pi Note If the data of the problems are three times continuously differentiable, then the functions / are Lipschitzian in their respective arguments, which considerably simplifies the notation for the proof of Theorem 3.5. For instance, case. Using essentially the same proof, the conclusion can then be strengthened to show that the order of convergence is at least 34. Conclusions. We present an algorithm that achieves superlinear convergence of the iterates to a local minimum of the nonlinear program (1.2) at which MFCQ (1.7) and the quadratic growth condition (1.1) are satisfied. The conditions we impose allow even situations for which no locally convex augmented Lagrangian exists, a case not accommodated by most previous results in the literature. At each step we solve a subproblem generated by approximating the function and the constraints by the second-order Taylor series at the current iterate. We also add a trust-region constraint, which insures that the problem is bounded. The algorithm therefore solves at each step a quadratically constrained quadratic program (QCQP) and we thus call it sequential quadratically constrained quadratic program (SQCQP). The subproblem to be solved is not necessarily convex. However we prove that for a suitable, fixed size of the trust region, the associated constraint is inactive at any stationary point of QCQP. As a result, any stationary point of the QCQP induces superlinear convergence of the iterates, which obviates the need for finding the global optimum of the subproblem. A subproblem that has quadratic constraints is more difficult to solve than a subproblem with linear constraints, the latter being the case of Sequential Quadratic Programming algorithms [19]. One could of course solve the QCQP with a nonlinear programming technique. The algorithm in [2] achieves at least linear convergence on the subproblem under the conditions considered here. Since in this work a more accurate model of the constraints is considered, compared to SQP, it would be expected that a smaller number of exterior iterations and thus of function evaluations is needed before completion. However, given the complexity of the subproblem, this will not necessarily results in superior runtime. Nevertheless, algorithms can be derived to deal directly with quadratically constrained problem via semidefinite relaxation [16]. Devising methods that specifically accommodate quadratic constraints will be the subject of future research. --R An Introduction to Numerical Analysis Degenerate New York Local Analysis of Newton-Type Methods for Variational Inequalities and Non-linear Programming Introduction to Sensitivity and Stability Analysis in Practical Methods of Optimization A necessary and sufficient regularity condition to have bounded multipliers in nonconvex programming Differential stability in nonlinear programming Stabilized sequential quadratic programming Stability in the presence of degeneracy and error estimation On approximate solutions of systems of linear inequalities Necessary and sufficient conditions for a local minimum. On Sensitivity analysis of nonlinear programs in Banach spaces: the approach via composite unconstrained optimization The Fritz John necessary optimality conditions in the presence of equality constraints Superlinear convergence of an interior-point method despite dependent constraints Applications to nonlinear programming Mathematical Programming Study 19 Sensitivity analysis of nonlinear programs and differentiability properties of metric projections Superlinear convergence of a stabilized SQP method to a degenerate solution Modifying SQP for degenerate problems --TR
quadratic constraints;sequential quadratic programming;degenerate constraints;superlinear convergence
589013
A Computationally Efficient Feasible Sequential Quadratic Programming Algorithm.
A sequential quadratic programming (SQP) algorithm generating feasible iterates is described and analyzed. What distinguishes this algorithm from previous feasible SQP algorithms proposed by various authors is a reduction in the amount of computation required to generate a new iterate while the proposed scheme still enjoys the same global and fast local convergence properties. A preliminary implementation has been tested and some promising numerical results are reported.
Introduction Consider the inequality-constrained nonlinear programming problem min f(x) s.t. are continuously differentiable. Sequential Quadratic Programming (SQP) algorithms are widely acknowledged to be among the most successful algorithms available for solving (P ). For an excellent recent survey of SQP algorithms, and the theory behind them, see [2]. Denote the feasible set for (P ) by In [17, 8, 14, 15, 1], variations on the standard SQP iteration for solving are proposed which generate iterates lying within X. Such methods are sometimes referred to as "Feasible SQP" (or FSQP) algorithms. It was observed that requiring feasible iterates has both algorithmic and application-oriented advantages. Algorithmically, feasible iterates are desirable because ffl The QP subproblems are always consistent, i.e. a feasible solution always exists, and ffl The objective function may be used directly as a merit function in the line search. In an engineering context, feasible iterates are important because ffl Often f(x) is undefined outside of the feasible region X, ffl Trade-offs between design alternatives (all requiring that "hard" constraints be satisfied) may be meaningfully explored, and ffl The optimization process may be stopped after a few iterations, yielding a feasible point. The last feature is critical for real-time applications, where a feasible point may be required before the algorithm has had time to "converge" to a solution An important function associated with the problem (P ) is the Lagrangian which is defined by Given a feasible estimate x of the solution of (P ) and a symmetric matrix H that approximates the Hessian of the Lagrangian L(x; -), where - is a vector of non-negative Lagrange multiplier estimates, the standard SQP search direction, denoted d 0 (x; H) or d 0 for short, solves of the Quadratic Program (QP) s.t. Positive definiteness of H is often assumes as it ensures existence and uniqueness of such solution. With an appropriate merit function, line search pro- cedure, Hessian approximation rule, and (if necessary) Maratos effect [13] avoidance scheme, the SQP iteration is known to be globally and locally superlinearly convergent (see, e.g., [2]). A feasible direction at a point x 2 X is defined as any vector d in R n such that x+ td belongs to X for all t in [0; - t ], for some positive - t. Note that the SQP direction d 0 , a direction of descent for f , may not be a feasible direction at x, though it is at worst tangent to the active constraint surface. Thus, in order to generate feasible iterates in the SQP framework, it is necessary to "tilt" d 0 into the feasible set. A number of approaches have been considered in the literature for generating feasible directions and, specifically, tilting the SQP direction. Early feasible direction algorithms (see, e.g., [27, 17]) were first-order methods, i.e. only first derivatives were used and no attempt was made to accumulate and use second-order information. Furthermore, search directions were often computed via linear programs instead of QPs. As a conse- quence, such algorithms converged linearly at best. Polak proposed several extensions to these algorithms (see [17], Section 4.4) which took second-order information into account when computing the search direction. A few of the search directions proposed by Polak could be viewed as tilted SQP directions (with proper choice of the matrices encapsulating the second-order information in the defining equations). Even with second-order information, though, it is not possible to guarantee superlinear convergence of these algorithms because no mechanism was included for controlling the amount of tilting. A straightforward way to tilt the SQP direction is, of course, to perturb the right-hand side of the constraints in QP 0 (x; H). Building on this obser- vation, Herskovits and Carvalho [8] and Panier and Tits [14] independently developed similar feasible SQP algorithms in which the size of the perturbation was a function of the norm of d 0 (x; H) at the current feasible point x. Thus, their algorithms required the solution of QP 0 (x; H) in order to define the perturbed QP. Both algorithms were shown to be superlinearly convergent. On the other hand, as a by-product of the tilting scheme, global convergence proved to be more elusive. In fact, the algorithm in [8] is not globally convergent, while the algorithm in [14] has to resort to a first-order search direction far from a solution in order to guarantee global convergence. Such a hybrid scheme could give slow convergence if a poor initial point is chosen. The algorithm developed by Panier and Tits in [15], and analyzed under weaker assumptions by Qi and Wei in [20], has enjoyed a great deal of success in practice as implemented in the FFSQP/CFSQP [26, 12] software packages. We will refer to their algorithm throughout this paper as FSQP. In [15], instead of directly perturbing QP 0 (x; H), tilting is accomplished by replacing d 0 with the convex combination an (essentially) arbitrary feasible descent direction. To preserve the local convergence properties of the SQP iteration, ae is selected as a function ae(d 0 ) of d 0 in such a way that d approaches d 0 fast enough (in particular, as the solution is approached. Finally, in order to avoid the Maratos effect and guarantee a superlinear rate of convergence, a second order correction d C (denoted ~ d in [15]) is used to "bend" the search direction. That is, an Armijo-type search is performed along the arc x+td+t 2 d C , where d is the tilted direction. In [15], the directions d 1 and d C are both computed via QPs but is is pointed out that d C could instead be taken as the solution of a linear least squares problem without affecting the asymptotic convergence properties. From the point of view of computational cost, the main drawback of algorithm FSQP is the need to solve three QPs (or two QPs and a linear least squares problem) at each iteration. Clearly, for many problems it would be desirable to reduce the number of QPs at each iteration while preserving the generation of feasible iterates as well as the global and local convergence properties. This is especially critical in the context of those large-scale nonlinear programs for which the time spent solving the QPs dominates that used to evaluate the functions. With that goal in mind, consider the following perturbation of QP 0 (x; H). Given a point x in X, a symmetric positive definite matrix H, and a non-negative scalar j, let (d(x; H; j); fl(x; H; j)) solve the QP s.t. where fl is an additional, scalar variable. The idea is that, away from KKT points of (P), fl(x; H; j) will be negative and thus d(x; H; j) will be a descent direction for f (due to the first constraint) as well as, if j is strictly positive, a feasible direction (due to the m other constraints). Note that when j is set to one the search direction is a special case of those computed in Polak's second-order feasible direction algorithms (again, see Section 4.4 in the book [17]). Further, it is not difficult to show that when j is set to zero, we recover the SQP direc- tion, i.e. d(x; H; values of the parameter j, which we will call the tilting parameter, emphasize feasibility, while small values of j emphasize descent. In [1], Birge, Qi, and Wei propose a feasible SQP algorithm based on QP (x; H; j). Their motivation for introducing the right-hand-side constraint perturbation and the tilting parameters (they use a vector of parameters, one for each constraint) is, like ours, to obtain a feasible search direction. Specifically, motivated by the high cost of function evaluations in the application problems they are targeting, their goal is to ensure that a full step of one is accepted in the line search as early on as is possible (so that costly line searches are avoided for most iterations). To this end, their tilting parameters start out positive and, if anything, increase when a step of one is not accepted. A side-effect of such an updating scheme is that the algorithm cannot achieve a superlinear rate of convergence, as the authors point out in Remark 5.1 of [1]. In the present paper, our goal is to compute a feasible descent direction which approaches the true SQP direction fast enough so as to ensure superlinear convergence. Furthermore, we would like to do this with as little computation per iteration as possible. While computationally rather expen- sive, algorithm FSQP of [15] has the convergence properties and practical performance we seek. We thus start with reviewing its key features. For x in X, define the index set of active constraints at x. In FSQP, in order for the line-search (with the objective function f used directly as the merit function) to be well-defined, and in order to preserve global and fast local convergence, the sequence of search directions fd k g generated by algorithm FSQP is constructed so that the following properties hold: is a KKT point for (P ), is not a KKT point, is not a KKT point, and We will show in Section 3 that given any symmetric positive definite matrix H k and non-negative scalar automatically satisfies P1 and P2. Furthermore, it satisfies P3 if j k is strictly positive. Ensuring that P4 holds requires a bit more care. In the algorithm proposed in this paper, at iteration k, the search direction is computed via solving QP and the tilting parameter j k is iteratively adjusted to ensure that the four properties are satisfied. The resultant algorithm will be shown to be locally superlinearly convergent and globally convergent without resorting to a first-order direction far from the solution. Further, the generation of a new iterate only requires the solution of one QP and two closely related linear least squares problems. In contrast with the algorithm presented in [1], our tilting parameter starts out positive and asymptotically approaches zero. Recently there has been a great deal of interest in interior point algorithms for nonconvex nonlinear programming (see, e.g., [5, 6, 24, 4, 16, 23]). Such algorithms generate feasible iterates and typically only require the solution of linear systems of equations in order to generate new iterates. SQP-type algorithms, however, are often at an advantage over such methods in the context of applications where the number of variables is not too large but evaluations of objectives/constraint functions and of their gradients are highly time-consuming. Indeed, because these algorithms use quadratic programs as successive models, away from a solution, progress between (expensive) function evaluations is often significantly better than that achieved by algorithms making use of mere linear systems of equations as models. In Section 2, we present the details of our new FSQP algorithm. In Section 3, we show that under mild assumptions our iteration is globally convergent, as well as locally superlinearly convergent. The algorithm has been implemented and tested and we show in Section 4 that the numerical results are quite promising. Some related issues are discussed in Section 5. Finally, in Section 6, we offer some concluding remarks and discuss some extensions to the algorithm which are currently being explored. Algorithm We begin by making a few assumptions that will be in force throughout. Assumption 1: The set X is non-empty. Assumption 2: The functions f are continuously differentiable. Assumption 3: For all x 2 X with I(x) 6= ;, the set frg j is linearly independent. A point x 2 R n is said to be a Karush-Kuhn-Tucker (KKT) point for the problem (P ) if there exist scalars (KKT multipliers) - ;j , such that (1) It is well known that, under our assumptions, a necessary condition for optimality of a point x 2 X is that it be a KKT point. Note that, with x 2 X, QP (x; H; j) is always consistent: (0; the constraints. Indeed, QP (x; H; j) always has a unique solution (d; fl) (see by convexity, is its unique KKT point; i.e. there exist multipliers - and - j , together with (d; fl), satisfy \Gammaj (2) A simple consequence of the first equation in (2), which will be used through-out our analysis, is an affine relationship amongst the multipliers, namely Parameter j will be assigned a new value at each iteration, j k at iteration k, to ensure that d(x k has the necessary properties. Strict positivity of j k is sufficient to guarantee that Properties P1 to P3 are satisfied. As it turns out however, this is not enough to ensure that, away from a solution, there is adequate tilting into the feasible set. For this, we will force j k to be bounded away from zero away from KKT points of (P ). Finally, P4 requires that j k tend to zero sufficiently fast as d 0 tends to zero, i.e., as a solution is approached. In [14], a similar effect is achieved by first computing of course, we want to avoid that here. Given an estimate I E k of the active set I(x k ), we can compute an estimate k ) of d 0 by solving the equality constrained QP s.t. which is equivalent (after a change of variables) to solving a linear least squares problem. Let I k be the set of active constraints, not including the "objective descent" constraint hrf(x k I k We will show in Section 3 that d E sufficiently large. Furthermore, we will prove that, when d k is small, choosing is sufficient to guarantee global and local superlinear convergence. Proper choice of the proportionality constant (C k in the algorithm statement below), while not important in the convergence analysis, is critical for satisfactory numerical performance. This will be discussed in Section 4. In [15], given x, H, and a feasible descent direction d, the Maratos correction d C (denoted ~ d in [15]) is taken as the solution of the QP s.t. if it exists and has norm less than minfkdk; Cg, where - is a given scalar satisfying 2 - 3 and C a given large scalar. Otherwise, d C is set to zero. (Indeed, a large d C is meangingless and may jeopardize global convergence.) In Section 1, it was mentioned that a linear least squares problem could be used instead of a QP to compute a version of the Maratos correction d C with the same asymptotic convergence properties. Given that our goal is to reduce the computational cost per iteration, it makes sense to use such an approach here. Thus, at iteration k, we take the correction d C k to be the solution d C exists and is not too large (specifically, if its norm is no larger than that of d k ), of the equality-constrained QP (equivalent to a least squares problem after a change of variables) s.t. direct extension of an alternative considered in [14]. In making use of the best available metric, such an objective, as compared to the pure least squares objective kd C k 2 , should yield a somewhat better iterate without significantly increasing computational requirements (or affecting the convergence analysis). Another advantage of using metric H k is that, asymptotically, the matrix underlying LS C will be the same as that underlying LS E resulting in computational sav- ings. In the case that LS C inconsistent, or the computed solution d C k is too large, we will simply set d C k to zero. The proposed algorithm is as follows. Parameters ff, fi are used in the Armijo-like search, - is the "bending" exponent in LS C , and ffl ' , C, C, and D are used in the update rule for j k . Algorithm FSQP 0 Parameters: positive definite, Computation of search arc. (i). compute (d k ; the active set I k , and associated multipliers (ii). compute d C exists and satisfies kd C k. Otherwise, set d C the first value of t in the sequence that satisfies Updates. (i). set x k+1 / x k . (ii). compute H k+1 , a new symmetric positive definite estimate to the Hessian of the Lagrangian. (iii). select C k+1 2 [C; C]. has a unique solution and unique associated multipiers, compute d E and the associated multipliers In such case, D and - E else set j k+1 / C k+1 ' . (iv). set k 3 Convergence Analysis Much of our analysis, especially the local analysis, will be devoted to establishing the relationship between d(x; H; j) and the SQP direction d 0 (x; H). Given x in X and H symmetric positive definite, d 0 is a KKT point for solution d 0 (x; H)) if and only if there exists a multiplier vector - 0 such that Further, given I ae mg, an estimate d E is a KKT point for LS E (x; H; I) (thus its unique solution d E (x; H; I)) if and only if there exists a multiplier vector - E such that Note that the components of - E for j 62 I play no role in the optimality conditions. 3.1 Global Convergence In this section we establish that, under mild assumptions, FSQP 0 generates a sequence of iterates fx k g with the property that all accumulation points are KKT points for (P ). We begin by establishing some properties of the tilted SQP search direction d(x; H; j). Lemma 1. Suppose Assumptions 1 through 3 hold. Then, given H symmetric positive definite, x 2 X, and j - 0, d(x; H; j) is well-defined and is the unique KKT point of QP (x; H; j). Further- more, d(x; H; j) is bounded over compact subsets of X \Theta P \Theta R + , where P is the set of symmetric positive definite n \Theta n matrices and R + the set of nonnegative real numbers. Proof. First note that the feasible set for QP (x; H; j) is non-empty, since Now consider the cases separately. From (2) and (4), it is clear that, if is a solution to only if d is a solution of QP 0 It is well known that, under our assumptions, d 0 (x; H) is well-defined, unique, and continuous. The claims follow. Suppose now that j ? 0. In that case, (d; fl) is a solution of QP (x; H; j) if and only if d solves the unconstrained problem ae oe and ae oe Since the function being minimized in (6) is strictly convex and radially unbounded, it follows that (d(x; H; j); fl(x; H; j)) is well-defined and unique as a global minimizer for the convex problem QP (x; H; j), and thus unique as a KKT point for that problem. Boundedness of d(x; H; subsets of X \Theta P \Theta R + follows from the first equation in (2), our regularity assumptions, and (3), which shows (since j ? 0) that the multipliers are bounded. Lemma 2. Suppose Assumptions 1 through 3 hold. Then, given H symmetric positive definite and j - 0 (i). fl(x; H; only if (ii). d(x; H; only if x is a KKT point for (P ). Moreover, if either (thus both) of these conditions holds, then the multipliers - and - for QP (x; H; are related by and - . Proof. To prove (i), note that since (d; QP (x; H; j), the optimal value of the QP is non-positive. Further, since H ? 0, the quadratic term in the objective is non-negative, which implies Now suppose that d(x; H; feasibility of the first QP constraint implies that fl(x; H; Finally, suppose that fl(x; H; it is clear that and achieves the minimum value of the objective. Thus, uniqueness gives Suppose now that d(x; H; by (2) there exist a multiplier vector - and a scalar multiplier - 0 such that We begin by showing that - ? 0. Proceeding by contradiction, suppose by (3) we have Note that, I Thus, by the complementary slackness condition of (2) and the optimality conditions (7), By Assumption 3, this sum vanishes only if - contradicting (8). Thus - ? 0. It is now immediate that x is a KKT point for (P ) with multipliers - Finally, to prove the necessity portion of part (ii) note that if x is a KKT point for (P ), then (1) shows that (d; is a KKT point for Uniqueness of such points (Lemma 1) yields the result. The next two lemmas establish that the line search in Step 2 of Algorithm FSQP 0 is well defined. Lemma 3. Suppose Assumptions 1 through 3 hold. Suppose x 2 X is not a KKT point for (P ), H is symmetric positive definite and j ? 0. Then (i). (ii). Proof. Both follow immediately from Lemma 2 and the fact that d(x; H; and fl(x; H; must satisfy the constraints in QP (x; H; j). Lemma 4. Suppose Assumptions 1 through 3 hold. Then, if a KKT point for (P ) and the algorithm will stop in Step 1(i) at iteration k. On the other hand, whenever the algorithm does not stop in Step 1(i), the line search is well defined, i.e. Step 2 yields a step t k equal to fi j k for some Proof. Suppose that j with The latter case cannot hold, as the stopping criterion in Step 1(i) would have stopped the algorithm at iteration k \Gamma 1. On the other hand, if then in view of the optimality conditions (5), and the fact that x k is always feasible for (P ), we see that x k is a KKT point for (P ) with multipliers 0; otherwise: Thus, by Lemma 2, d and the algorithm will stop in Step 1(i). The first claim is thus proved. Also, we have established that Step 2 is reached. The second claim now follows immediately from Lemma 3 and Assumption 2. The previous lemmas imply that the algorithm is well-defined. In addi- shows that if Algorithm FSQP 0 generates a finite sequence terminating at the point xN , then xN is a KKT point for the problem (P ). We now concentrate on the case in which an infinite sequence fx k g is gen- erated, i.e. the algorithm never satisfies the termination condition in Step 1(i). Note that, in view of Lemma 4, we may assume throughout that Before proceeding, we make an assumption concerning the estimates H k of the Hessian of the Lagrangian. Assumption 4: There exist positive constants oe 1 and oe 2 such that, for all k, Lemma 5. Suppose Assumptions 1 through 4 hold. Then the sequence fj k g generated by Algorithm FSQP 0 is bounded. Further, the sequence fd k g is bounded on subsequences on which fx k g is bounded. Proof. The first claim follows from the update rule in Step 3(iii) of Algorithm . The second claim then follows from Lemma 1 and Assumption 4. Given an infinite index set K, we will use the notation \Gamma! x to mean Lemma 6. Suppose Assumptions 1 through 3 hold. Suppose K is an infinite index set such that x k is bounded on K, and d k \Gamma! 0: sufficiently large and the QP multiplier sequences are bounded on K. Further, given any accumulation point is the unique solution of QP Proof. In view of Assumption 2 frf(x k )g k2K must be bounded. Lemma 2(i) and the first constraint in QP Thus, \Gamma! 0. To prove the first claim, let j 0 62 I(x ). There exists such that g j 0 sufficiently large. In view of Assumption 2, and since d k \Gamma! 0, fl k \Gamma! 0, and fj k g is bounded on K, it is clear that sufficiently large, proving the first claim. Boundedness of f- k g k2K follows from non-negativity and (3). To prove that of f- k g k2K , using complementary slackness and the first equation in (2), write Proceeding by contradiction, suppose that f- k g k2K is unbounded. Without loss of generality, assume that k- k k 1 ? 0, for all k 2 K and define for all Note that, for all k 2 K, k- k k Dividing (10) by k- k k 1 and taking limits on an appropriate subsequence of K, it follows from Assumptions 2 and 4 and boundedness of f- k g that for some - ;j , 1. As this contradicts Assumption 3, it is established that f- k g k2K is bounded. To complete the proof, let K 0 ' K be an infinite index set such that \Gamma! j and assume without loss of generality that H k and - k \Gamma! - . Taking limits in the optimality conditions (2) shows that, indeed, (d; and - . Finally, uniqueness of such points (Lemma 1) proves the result. Lemma 7. Suppose Assumptions 1 through 4 hold. Then, if K is an infinite index set such that d k \Gamma! 0, all accumulation points of fx k g k2K are KKT points for (P ). Proof. Suppose K 0 ' K is an infinite index set on which x k \Gamma! x 2 X. In view of Assumption 4 and Lemma 5, assume, without loss of generality that H k \Gamma! H , a positive definite matrix, and j k In view of Lemma 6, (0; 0) is the unique solution of QP It follows from Lemma 2 that x is a KKT point for (P ). We now state and prove the main result of this subsection. Theorem 1. Under Assumptions 1 through 4, Algorithm FSQP 0 generates a sequence fx k g for which all accumulation points are KKT points for (P ). Proof. Suppose K is an infinite index set such that x k \Gamma! x . In view of Lemma 5 and Assumption 4, we may assume without loss of generality that \Gamma! d , j k are considered separately. Suppose first that j there exists an infinite index set K 0 ' K such that either d E d \Gamma! 0. If the latter case holds, it is then clear that x \Gamma! x , since \Gamma! 0. Thus, by Lemma 7, x is a KKT point for suppose instead that d E From the second set of equations in (5), one can easily see that I sufficiently large, and using an argument very similar to that used in Lemma 6, one can show that f- E k g k2K 0 is a bounded sequence. Thus, taking limits in (5) on an appropriate subsequence of K 0 shows that x is a KKT point for (P ). Now consider the case j ? 0. We show that d k \Gamma! 0. Proceeding by contradiction, without loss of generality suppose there exists d ? 0 such that kd k k - d for all k 2 K. From non-positivity of the optimal value of the objective function in QP Assumption 4, we see that Further, in view of (9) and since j ? 0, there exists j ? 0 such that From the constraints of QP and using Assumption 2, it is easily shown that there exists such that for all k 2 K, k large enough, The rest of the contradiction argument establishing d k exactly the proof of Proposition 3.2 in [14]. Finally, it then follows from Lemma 7 that x is a KKT point for (P ). 3.2 Local Convergence While the details are often quite different, overall the analysis in this section is inspired by and occasionally follows that of Panier and Tits in [14, 15]. The key result is Proposition 1 which states that, under appropriate assumptions, the arc search eventually accepts the full step of one. With this and the fact, to be established along the way, that titled direction d k approaches the standard SQP direction sufficiently fast, superlinear convergence follows from a classical analysis of M.J.D. Powell's. As a first step, we strengthen the regularity assumptions. Assumption are three times continuously differentiable. A point x is said to satisfy the second order sufficiency conditions with strict complementary slackness for (P ) if there exists a multiplier vector ffl The pair (x ; - ) satisfies (1), i.e. x is a KKT point for (P ), positive definite on the subspace ffl and - ;j ? 0 for all j 2 I(x ) (strict complementary slackness). In order to guarantee that the entire sequence fx k g converges to a KKT point x , we make the following assumption. (Recall that we have already established, under weaker assumptions, that every accumulation point of is a KKT point for (P ).) Assumption 5: The sequence fx k g has an accumulation point x which satisfies the second order sufficiency conditions with strict complementary slackness. It is well known that Assumption 5 guarantees that the entire sequence converges. For a proof see, e.g., Proposition 4.1 in [14]. Lemma 8. Suppose Assumptions 1, 2', and 3 through 5 hold. Then the sequence generated by Algorithm FSQP 0 converges to a point x satisfying the second order sufficiency conditions with strict complementary slackness. From this point forward, - will denote the (unique) multiplier vector associated with KKT point x for (P ). It is readily checked that, for any symmetric positive definite H, (0; - ) is the KKT pair for QP 0 As announced, as a first main step, we show that our sequence of tilted SQP directions approaches the true SQP direction sufficiently fast. (This is achieved in Lemmas 9 through 18.) In order to do so, define d 0 k to be equal to d 0 are as computed by Algorithm FSQP 0 . Further, for each k, define - 0 k as a multiplier vector such that (d 0 (4) and let I 0 g: The following lemma is proved in [15] (with reference to [14]) under identical assumptions. Lemma 9. Suppose Assumptions 1, 2', and 3 through 5 hold. Then (iii) For all k sufficiently large, the following two equalities hold I 0 We next establish that the entire tilted SQP direction sequence converges to 0. In order to do so, we establish that d(x; H; j) is continuous in a neighborhood of positive definite. Complicating the analysis is the fact that we have yet to establish that the sequence fj k g does, in fact, converge. Given j - 0, define the set ae' rf(x ) \Gammaj oe Lemma 10. Suppose Assumptions 1, 2', and 3 through 5 hold. Then, given any j - 0, the set N (j ) is linearly independent. Proof. Let H be symmetric positive definite. Note that, in view of Lemma 2, Now suppose the claim does not hold, i.e. suppose there exists scalars - j , j 2 f0g [ I(x ), not all zero, such that \Gammaj 0: (11) In view of Assumption 3, - 0 6= 0 and the scalars - j are unique modulo a scaling factor. This uniqueness, the fact that d(x and the first scalar equations in the optimality conditions (2) imply that - are KKT multipliers for QP Thus, in view of (3), 0: But this contradicts (11), which gives hence N (j ) is linearly independent. Lemma 11. Suppose Assumptions 1, 2', and 3 through 5 hold. Let j - 0 be an accumulation point of fj k g. Then, given any symmetric positive definite is the unique solution of QP (x ; H; j ) and the second order sufficiency conditions hold, with strict complementary slackness Proof. In view of Lemma 2, QP its unique solution. Define the Lagrangian function L : R n \Theta R \Theta R \Theta R m ! R for Suppose - are KKT multipliers such that (2) holds with - and -. Let be the index for the first constraint in QP (x ; H; j ), i.e. hrf(x ); di - fl. Note that since (d ; fl the active constraint index set I for QP (Note that we define I as including 0, while I k was defined as a subset of Thus the set of active constraint gradients for QP is N (j ). Now consider the Hessian of the Lagrangian for QP (x ; H; j ), i.e. the second derivative with respect to the first two variables (d; fl), and given an arbitrary h 2 R n+1 , decompose it as clearly, and for h 6= 0, h T r 2 L(0; 0; - Hy is zero if and only if ff 6= 0. Since for such h ff it then follows that r 2 L(0; 0; - -) is positive definite on N (j ) ? , the tangent space to the active constraints for QP (x ; H; j ) at (0; 0). Thus, it is established that the second order sufficiency conditions hold. Finally it follows from Lemma 2(ii) that - - which, together with Assumption 5, implies strict complementarity for QP at (0; 0). Lemma 12. Suppose Assumptions 1, 2', and 3 through 5 hold. Then, if K is a subsequence on which fj k g converges, say to and - k Finally, Proof. First, proceed by contradiction to show that the first two claims hold and that, in addition, \Gamma! (0; 0); (12) i.e., suppose that on some infinite index set K 0 ' K either - k is bounded away from - -, or - k is bounded away from - from zero. In view of Assumption 4, there is no loss of generality is assuming that H k \Gamma! H for some symmetric positive definite H . In view of Lemmas 10 and 11, we may thus invoke a result due to Robinson (Theorem 2.1 in [21]) to conclude that, in view of Lemma 2(ii), \Gamma! (0; 0); - k a contradiction. Hence the first two claims hold, as does (12). Next, proceeding again by contradiction, suppose that d k 6! 0. Then, since fH k g and fj k g are bounded, there exists an infinite index set K on which fH k g and fj k g converge and d k is bounded away from zero. This contradicts (12). Thus It immediately follows from the first constraint in QP that Lemma 13. Suppose Assumptions 1, 2', and 3 through 5 hold. Then, for all k sufficiently large, I Proof. Since is bounded and, in view of Lemma 12, (d k ; Lemma 6 implies that I k ' I(x ), for all k sufficiently large. Now suppose it does not hold that I sufficiently large. Thus, there exists and an infinite index set K such that j 0 62 I k , for all k 2 K. Now, in view of Lemma 5, there exists an infinite index set K 0 ' K and j - 0 such that j k Further, Lemma 12 shows that - j 0 all k sufficiently large, k 2 K 0 , which, by complementary slackness, implies this is a contradiction, and the claim is proved. Now define and, given a vector - define the notation Note that, in view of Lemma 9(iii), for k large enough, the optimality conditions (4), yield R T The following well-known result will be used. Lemma 14. Suppose Assumptions 1, 2', and 3 through 5 hold. Then the R T is invertible for all k large enough and its inverse remains bounded as k ! 1. Lemma 15. Suppose Assumptions 1, 2', and 3 through 5 hold. For all k sufficiently large, d E are uniquely defined, and d E k . Proof. In view of Lemma 13, the optimality conditions (5), and Lemma 14, for all k large enough, the estimate d E k and its corresponding multiplier vector are well defined as the unique solution of R T The claim then follows from (13). Lemma 16. Suppose Assumptions 1, 2', and 3 through 5 hold. Then (iii) For all k sufficiently large, I Proof. Claim (i) follows from Step 3(iii) of Algorithm FSQP 0 , since in view of Lemma 12, Lemma 15, and Lemma 9, fd k g and fd E both converge to 0. In view of (i), Lemma 12 establishes that (ii) is proved. Finally, claim (iii) follows from claim (ii), Lemma 13, and Assumption 5. We now focus our attention on establishing relationships between d k , d C and the true SQP direction d 0 k . Lemma 17. Suppose Assumptions 1, 2', and 3 through 5 hold. Then Proof. In view of Lemma 15, for all k sufficiently large, d E k exist and are uniquely defined, and d E k . Lemmas 12 and 9 ensure that Step 3(iii) of Algorithm FSQP 0 chooses sufficiently large, thus (i) follows. It is clear from Lemma 13 and the optimality conditions (2) that d k and - k satisfy R T for all k sufficiently large, where 1 jI(x )j is a vector of jI(x )j ones. It thus follows from (13), Assumption 2, and Lemmas 12, 14 and 16 that and in view of claim (i), claim (ii) follows. Finally, since (from the QP constraint and Lemma is clear that O(kd k Lemma 18. Suppose Assumptions 1, 2', and 3 through 5 hold. Then d C O(kd 0 Proof. Let Expanding we see that, for some - j 2 (0; 1), z - \Gammag Assumption 2' we conclude that c O(kd 0 sufficiently large, in view of Lemma 13, d C k is well-defined and satisfies thus R T Now, the first order KKT conditions for LS C us there exists a multiplier - C k 2 R jI(x )j such that R T Also, from the optimality conditions (15) we have where In view of Lemma 17, q k and - C R T d C or equivalently, with - 0 R T The result then follows from Lemma 14. In order to prove the key result that the full step of one is eventually accepted by the line search, we now assume that the matrices fH k g suitably approximate the Hessian of the Lagrangian at the solution. Define the projection (R T Assumption lim 0: The following technical lemma will be used. Lemma 19. Suppose Assumptions 1, 2', and 3 through 5 hold. Then there exist constants (ii) for all k sufficiently large sufficiently large, Proof. To show part (i), note that in view of the first QP constraint, negativity of the optimal value of the QP objective, and Assumption 4, The proof of part (ii) is identical to that of Lemma 4.4 in [14]. To show (iii), note that from (15) for all k sufficiently large, d k satisfies R T Thus, we can write d (R T The result follows from Assumption 3 and Lemma 17(i,iii). Proposition 1. Suppose Assumptions 1, 2', and 3 through 6 hold. Then, sufficiently large. Proof. Following [14], consider an expansion of g j (\Delta) about x k I(x ), for all k sufficiently large, where we have used Assumption 2', Lemmas 17 and 18, boundedness of all sequences, and (16). As - ! 3, it follows that g j for all k sufficiently large. The same result trivially holds for j 62 I(x for k large enough, the full step of one satisfies the feasibility condition in the arc search test. It remains to show that the "sufficient decrease" condition is satisfied as well. First, in view of Assumption 2 0 and Lemmas 17 and 18, From the top equation in optimality conditions (2), equation (3), Lemma 17(i), and boundedness of all sequences, we obtain The last line in (2) and Lemma 17(i,iii) yield Taking the inner product of (19) with d k , then adding and subtracting the quantity using (20), and finally multiplying the result by 1gives2 hrf(x k ); d k Further, Lemmas 17 and Combining (18), (21), and (22), and using the fact that, for k large enough, With this in hand, arguments identical to those used following equation (4.9) in [14] show that for all k sufficiently large. Thus the "sufficient decrease" condition is satisfied A consequence of Lemmas 17, 18, and Proposition 1 is that the algorithm generates a convergent sequence of iterates satisfying Two-step superlinear convergence follows. Theorem 2. Suppose Assumptions 1, 2', and 3 through 6 hold. Then Algorithm generates a sequence fx k g which converges 2-step superlinearly to x , i.e. lim 0: The proof is not given as it follows step by step, with minor modifications, that of [18, Sections 2-3]. Finally, note that Q-superlinear convergence would follows if Assumption 6 were replaced with the stronger assumption lim 0: (See, e.g., [2].) 4 Implementation and Numerical Results Our implementation of FSQP 0 (in C) differs in a number of ways from the algorithm stated in Section 2. (It is readily checked that none of the differences significantly affect the convergence analysis of Section 3.) Just like in existing C implementation of FSQP (CFSQP: see [12]) the distinctive character of linear (affine) constraints and of simple bounds is exploited (provided the nature of these constraints is made explicit). Thus the general form of the problem description tackled by our implementation is min f(x) s.t. where a (componentwise). The details of the implementation are spelled out below. Many of them, including the update rule for H k , are exactly as in CFSQP. In the implementation of QP no "tilting" is effected in connection with the linear constraints and simple bound, since clearly the un- tilted SQP direction is feasible for these constraints. In addition, each non-linear constraint is assigned its own tilting parameter j j Thus QP replaced with The k 's are updated independently, based on independently adjusted C j 's. In the algorithm description and in the analysis all that was required of was that it remain bounded and bounded away from zero. In practice, though, performance of the algorithm is critically dependent upon the choice of C k . In the implementation, an adaptive scheme was chosen in which the new values C j are selected in Step 3 based on their previous values C j on the outcome of the arc search in Step 2, and on a preselected parameter if the full step of one was accepted are left unchanged; (ii) if the step of one was not accepted even though all trial points were feasible, then, for all j, C j k is decreased to minfffi c C j (iii) if some infeasibility was encountered in the arc search, then, for all j such that g j caused a step reduction at some trial point, C j k is increased to k is kept constant. Here, g j is said to cause a step reduction if, for some trial point x, g j is violated (i.e., g j (x) ? but all constraints checked at x before g j were found to be satisfied at that point. (See below for the order in which constraints are checked in the arc search.) It was stressed in Section 2 that the Maratos correction can be computed using an inequality-constrained QP such as QP C , instead of a LS C . This was done in our numerical experiments, in order to more meaningfully compare the new algorithm with CFSQP, in which an in an inequality-constrained QP is indeed used. The implementation of QP C and LS E involves index sets of "almost active" constraints and of binding constraints. First we define I n I a is the machine precision. Next, the binding sets are defined as I b;n I b;l mn is now the QP multiplier corresponding to the nonlinear constraints and where - a are the QP multipliers corresponding to the affine constraints, the upper bounds, and the lower bounds, respectively. Of course, no bending is required from d C k in connection with affine constraints and simple bounds, hence if I n simply set d C Otherwise the following modification of QP C is used: s.t. I a Since not all simple bounds are included in the computation of d C k , it is possible that x k k will not satisfy all bounds. To take care of this, we simply "clip" d C k so that the bounds are satisfied. Specifically, for the upper bounds, we perform the following: for j 62 I b;u do if (d C;j d C;j The same procedure, mutatis mutandis, is executed for the lower bounds. We note that such a procedure has no effect on the convergence analysis of Section 3 since, locally, the active set is correctly identified and a full step along k is always accepted. The least squares problem LS E used to compute d E k is modified similarly. Specifically, in the implementation, d E k is only computed if m n ? 0, in which case we use s.t. The implementation of the arc search (Step 2) is as in CFSQP. Specif- ically, feasibility is checked before sufficient decrease, and testing at a trial point is aborted as soon as infeasibility is detected. Like in CFSQP, all linear and bound constraints are checked first, then nonlinear constraints in an order maintained as follows: (i) at the start of the arc search from a given iterate x k the order is reset to be the natural numerical order; (ii) within an arc search, as a constraint is found to be violated at a trial point, its index is moved to the beginning of the list, with the order of the others left unchanged. An aspect of the algorithm which was intentionally left vague in Sections 2 and 3 was the updating scheme for the Hessian estimates H k . In the implementation, we use the BFGS update with Powell's modification [19]. Specifically, define where, in an attempt to better approximate the true multipliers, if - k ? we normalize as follows A scalar ' k+1 2 (0; 1] is then defined by 0:8 the rank two Hessian update is Note that while it is not clear whether the resultant sequence fH k g will, in fact, satisfy Assumption 6, this update scheme is known to perform very well in practice. All QPs and linear least squares subproblems were solved using QPOPT [7]. For comparison sake, QPOPT was also used to solve the QP subproblems in CFSQP. While the default QP solver for CFSQP is the public domain code QLD (see [22]), we opted for QPOPT because it allows "warm starts" and thus is fairer to CFSQP in the comparison with the implementation of more QPs are solved with the former). For alls QPs in both codes, the active set in the solution at a given iteration was used as initial guess for the active set for the same QP at the next iteration. In order to guarantee that the algorithm terminates after a finite number of iterations with an approximate solution, the stopping criterion of Step 1 is changed to small. Finally, the following parameter values were selected: Further, we always set H experiments were run on a Sun Microsystems Ultra 5 workstation For the first set of numerical tests, we selected a number of problems from [9] which provided feasible initial points and contained no equality con- straints. The results are reported in Table 1, where the performance of our implementation of FSQP 0 is compared with that of CFSQP (with QPOPT as QP solver). The column labeled # lists the problem number as given in [9], the column labeled ALGO is self-explanatory. The next three columns give the size of the problem following the conventions of this section. The columns labeled NF, NG, and IT give the number of objective function eval- uations, nonlinear constraint function evaluations, and iterations required to solve the problem, respectively. Finally, f(x ) is the objective function value at the final iterate and ffl is as above. The value of ffl was chosen in order to obtain approximately the same precision as reported in [9] for each problem. The results reported in Table 1 are encouraging. The performance of our implementation of Algorithm FSQP 0 in terms of number of iterations and function evaluations is essentially identical to that of CFSQP (Algorithm FSQP). The expected payoff of using FSQP 0 instead of FSQP however is that, on large problems the CPU time expended in linear algebra, specifically in solving the QP and linear least squares subproblems, should be much less. To assess this, we carried out comparative tests on the COPS suite of problems [3]. The first five problems from the COPS set [3] were considered, as these problems either do not involve nonlinear equality constraints or are readily reformulated without such constraints. (Specifically, in problem "Sphere" the equality constraint was changed to a "-" constraint; and in "Chain" the equality constraint (with replaced with two inequalities, with the left-hand side constrained to be between the values the solution was always at 5.) "Sawpath" was discarded because it involves few variables and many constraints, which is not the situation at which the new algorithm is targeted. The results obtained with various instances of the other four problems are presented in Table 2. The format of that table is identical to that of Table 1 except for the additional column labeled NQP. In that column we list the total number of QP iterations in the solution of the two major QPs, as reported by QPOPT. (Note that QPOPT reports CFSQP CFSQP CFSQP 9 19 7 6.0000000E+00 43 CFSQP Table 1: Numerical results on Hock-Schittkowski problems. zero iteration when the result of the first step onto the working set of linear constraints happens to be optimal. To be "fair" to FSQP 0 , we thus do not count the work involved in solving LS E either. We also do not count the QP iterations in solving QP C , the "correction" QP, because it is invoked identically in both algorithms.) The reason for the smaller ffl on Cam is this allowed CFSQP to reach the globally optimal objective function values (as per [3]). The results show a typical significantly lower number of QP iterations with the new algorithm and, as in the case of the Hock-Schittkowski prob- lems, a roughly comparable behavior of the two algorithms in terms of number of function evaluations. Note that in the two instances where the NQP count is less for CFSQP than for FSQP 0 , different local minima are reached, which makes the comparison meaningless. Finally, the abnormal terminations on Sphere-50 and Sphere-100 are both due to QPOPT's failure to solve a QP-the "tilting" QP in the case of CFSQP. One issue of interest is whether our convergence results still hold under weaker assumptions. To wit, Qi and Wei showed in [20] that the algorithm of [15] still enjoys global convergence (all limit points are KKT) and local (two-step) superlinear convergence when Assumption 3 (LICQ) is replaced with the Mangasarian-Fromovitz constraint qualification (MFCQ)-or even with a condition slightly weaker than MFCQ. They further showed that, if that algorithm is slightly modified, local superlinear convergence is preserved without strict complementarity assumption, provided the strong second-order sufficiency condition (SSOSC) is assumed. For algorithm FSQP 0 as stated however, LICQ and strict complementarity are essential, in connection with Step 3 (iii). First, Assumption 3 is needed in order for LS E to be well-defined close to a solution x . Second, strict positivity of the multipliers associated with active constraints at the solution is needed in order for the components of - E k to be nonnegative when a solution is approached. Barring this, the condition - E may never hold and the update rule j k+1 / C k+1 \Delta kd k k 2 may not be used close to the solution, in which case superlinear convergence would not take place. Careful modifications of Algorithm FSQP 0 might (at least in theory) accommodate weaker assumptions though. First, in the absence of Assumption 3, LS E could possibly be replaced with a QP with jI E inequality constraints (with essentially no penalty in CPU cost). Second, by replacing the nonnegativity condition by a requirement of the type - E;j - \Gammaffl k , where - E;j is now a "reg- ular" multiplier (see [20]) and where ffl k ? 0 would be made to go to zero at an appropriate (slow) rate, it may be possible to preserve convergence to KKT points while insuring that, even in the absence of strict comple- mentarity, the test would be satisfied close to the solution, thus allowing superlinear convergence to take place. This would require detailed analysis though, and may be as likely to hurt as to help in practice. A second issue worth discussing is that of possible low cost solution for the QP and two linear least-squares problems. The indexes of constraints appearing in LS E and LS C at iteration k are generally different, I E 28 142 .776859 1.E-4 CFSQP 42 8177 44 350 .776859 CFSQP 591 345458 154 2771 .783873 CFSQP 795 28328 246 587 660.675 CFSQP failure CFSQP 977 3784 575 1259 4.81189 CFSQP Table 2: Numerical results on COPS problems. for the former, I k for the latter. However, it was proved that, for k large enough, both of these sets are equal to I(x ). When that is the case, it is readily checked that LS E involve the same matrix, and thus that the latter can be solved at low cost once the former has been solved. Unfortunately, the linear systems arising in the solution of QP are different. In particular, they involve j k . 6 Conclusions We have presented here a new SQP-type algorithm generating feasible it- erates. The main advantage of the algorithm presented here is a reduction in the amount of computation required in order to generate a new iterate. While this may not be very important for applications where function evaluations dominate the actual amount of work to compute a new iterate, it is very useful in many contexts. In any case, we saw in the previous section that preliminary results seem to indicate that decreasing the amount of computation per iteration did not come at the cost of increasing the number of function evaluations required to find a solution. A number of significant extensions of Algorithm FSQP 0 are being ex- amined. It is not too difficult to extend the algorithm to handle mini-max problems. The only real issue that arises is how to handle the mini-max objectives in the least squares sub-problems. Several possibilities, each with the desired global and local convergence properties, are being examined. Another extension that is important for engineering design is the incorporation of a scheme to efficiently handle very large sets of constraints and/or objectives. We will examine schemes along the lines of those developed in [11, 25]. Further, work remains to be done to exploit the close relationship between the two least squares problems and the quadratic program. A careful implementation should be able to use these relationships to great advantage computationally. For starters, updating the Cholesky factors of H k instead of H k itself at each iteration would save a factorization in each of the sub-problems. Finally, it is possible to extend the class of problems (P ) which are handled by the algorithm to include nonlinear equality con- straints. Of course, we will not be able to generate feasible iterates for such constraints, but a scheme such as that studied in [10] could be used in order to guarantee asymptotic feasibility while maintaining feasibility for all inequality constraints. --R A variant of the Topkis-Veinott method for solving inequality constrained optimization problems Sequential quadratic programming. An interior point algorithm for large scale nonlinear programming. On the formulation and theory of the Newton interior-point method for nonlinear programming A primal-dual interior method for nonconvex nonlinear programming User's guide for qpopt 1.0: A fortran package for quadratic programming. A successive quadratic programming based feasible directions algorithm. Test Examples For Nonlinear Programming Codes Nonlinear equality constraints in feasible sequential quadratic programming. Feasible sequential quadratic programming for finely discretized problems from SIP. User's Guide for CFSQP Version 2.5: A C Code for Solving (Large Scale) Constrained Nonlinear (Minimax) Optimization Problems Exact Penalty Functions for Finite Dimensional and Control Optimization Problems. A superlinearly convergent feasible method for the solution of inequality constrained optimization problems. On combining feasibility Computational Methods in Optimization. Convergence of variable metric methods for nonlinearly constrained optimization calculations. A fast algorithm for nonlinearly constrained optimization calculations. On the constant positive linear dependence condition and its application to SQP methods. Perturbed Kuhn-Tucker points and rates of convergence for a class of nonlinear-programming algorithms QLD: A Fortran Code for Quadratic Programming A primal-dual interior-point method for nonconvex optimization with multiple logarithmic barrier parameters and with strong convergence properties An interior point algorithm for non-convex nonlinear programming An SQP algorithm for finely discretized continuous minimax problems and other minimax problems with many objective functions. User's Guide for FSQP Version 3.7: A FORTRAN Code for Solving Nonlinear (Minimax) Optimization Problems Methods of Feasible Directions. --TR --CTR Dudy Lim , Yew-Soon Ong , Bu-Sung Lee, Inverse multi-objective robust evolutionary design optimization in the presence of uncertainty, Proceedings of the 2005 workshops on Genetic and evolutionary computation, June 25-26, 2005, Washington, D.C. Zhibin Zhu, An efficient sequential quadratic programming algorithm for nonlinear programming, Journal of Computational and Applied Mathematics, v.175 n.2, p.447-464, 15 March 2005 L. Bauwens , C. M. Hafner , J. V. K. Rombouts, Multivariate mixed normal conditional heteroskedasticity, Computational Statistics & Data Analysis, v.51 n.7, p.3551-3566, April, 2007 David Cardoze , Alexandre Cunha , Gary L. Miller , Todd Phillips , Noel Walkington, A bzier-based approach to unstructured moving meshes, Proceedings of the twentieth annual symposium on Computational geometry, June 08-11, 2004, Brooklyn, New York, USA Matthew J. Tenny , Stephen J. Wright , James B. Rawlings, Nonlinear Model Predictive Control via Feasibility-Perturbed Sequential Quadratic Programming, Computational Optimization and Applications, v.28 n.1, p.87-121, April 2004 Daniel Mueller , Helmut Graeb , Ulf Schlichtmann, Trade-off design of analog circuits using goal attainment and "Wave Front" sequential quadratic programming, Proceedings of the conference on Design, automation and test in Europe, April 16-20, 2007, Nice, France Y. S. Ong , K. Y. Lum , P. B. Nair, Hybrid evolutionary algorithm with Hermite radial basis function interpolants for computationally expensive adjoint solvers, Computational Optimization and Applications, v.39 n.1, p.97-119, January 2008 Borys Shchokin , Farrokh Janabi-Sharifi, Design and kinematic analysis of a rotary positioner, Robotica, v.25 n.1, p.75-85, January 2007
feasible SQP;feasible iterates;sequential quadratic programming;FSQP;SQP
589022
Monotonicity of Fixed Point and Normal Mappings Associated with Variational Inequality and Its Application.
We prove sufficient conditions for the monotonicity and the strong monotonicity of fixed point and normal maps associated with variational inequality problems over a general closed convex set. Sufficient conditions for the strong monotonicity of their perturbed versions are also shown. These results include some well known in the literature as particular instances. Inspired by these results, we propose a modified Solodov and Svaiter iterative algorithm for the variational inequality problem whose fixed point map or normal map is monotone.
Introduction . Given a continuous function f : R n ! R n and a closed convex set K in R n ; the well-known finite-dimensional variational inequality, denoted by f ), is to find an element x 2 K such that It is well-known that the above problem can be reformulated as nonsmooth equations such as the fixed point and normal equations (see e.g. [9, 18]). The fixed point equation is defined by and the normal equation is defined by positive scalar, and \Pi K (\Delta) denotes the projection operator on the convex set K; i.e., Throughout the paper, k \Delta k denotes the 2-norm (Euclidean norm) of the vector in R It turns out that x solves VI(K; f) if and only if - ff (x and that if x solves ff f(x ) is a solution to \Phi ff conversely, if \Phi ff (u \Pi K (u ) is a solution to VI(K; f): Recently, several authors studied the P 0 property of fixed point and normal maps when K is a rectangular box in R n , i.e., the Cartesian product of n one-dimensional This work was partially supported by Research Grants Council of Hong Kong under grant CUHK4392/99E. y Department of Systems Engineering and Engineering Management, The Chinese University of Hong Kong, Shatin, New Territory, Hong Kong, and Institute of Computational Mathematics and Scientific/Engineering Computing, Chinese Academy of Sciences, Beijing, China (Email: ybzhao@se.cuhk.edu.hk). z Corresponding author. Department of Systems Engineering and Engineering Management, The Chinese University of Hong Kong, Shatin, New Territory, Hong Kong (Fax: (852) 26035505; Tel: Y. B. ZHAO AND D. LI intervals. For such a K, Ravindran and Gowda [17] (respectively, Gowda and Tawhid [8]) showed that - ff (x) (respectively, \Phi ff (x)) is a P 0 -function if f is. Notice that the monotone maps are very important special cases of the class of P 0 -functions. It is worth considering the problem: (P) When are the mappings - ff (x) and \Phi ff (x) monotone if K is a general closed convex set? Intuitively, we may conjecture that the fixed point map and the normal map are monotone if f is. However, this conjecture is not true. The following example shows that for a given ff ? 0 the monotonicity of f , in general, does not imply the monotonicity of the fixed point map - ff (x) and the normal map \Phi ff (x): Example 1.1. Let K be a closed convex set given by and For any x; y we have that 0: Hence the function f is monotone on R 2 . We now show that for an arbitrary scalar ff ? 0 the fixed point mapping - ff not monotone in R 2 . Indeed, let It is easy to verify that - ff Thus, we have which implies that - ff (\Delta) is not monotone on R Example 1.2. Let K be a closed convex set given by given as in Example 1.1. We now show that for an arbitrary the normal mapping \Phi ff not monotone in R 2 . Indeed, let We have that \Phi ff Thus, we have which implies that \Phi ff (\Delta) is not monotone on R From the above examples, we conclude that certain condition stronger than the monotonicity of f is required to guarantee the monotonicity of - ff (x) and \Phi ff (x). One such condition is so-called co-coercivity condition. We recall that f is said to be co- coercive with modulus fi ? 0 on a set S ae R n if there exists a constant fi ? 0 such that The co-coercivity condition was used in several works, such as Bruck [1], Gabay [7] (in which this condition is used implicitly), Tseng [25], Marcotte and Wu [15], Magnanti and Perakis [13, 14], and Zhu and Marcotte [29, 30]. It is also used to study the strict feasibility of complementarity problems [27]. It is interesting to note that in an affine MONOTONICITY OF FIXED POINT AND NORMAL MAPPINGS 3 case the co-coercivity has a close relation to the property of psd-plus matrices [12, 30]. A special case of the co-coercive map is the strongly monotone and Lipschitzian map. We recall that a mapping f is said to be strongly monotone with modulus c ? 0 on the set S if there is a scalar c ? 0 such that It is evident that any co-coercive map on the set S must be monotone and Lipschitz continuous (with constant not necessarily strongly monotone (for instance, the constant mapping) on the same set. In fact, the aforementioned problem (P) is not completely unknown. By using the co-coercivity condition implicitly and using properties of nonexpansive maps, Gabay actually showed (but did not explicitly state) that - ff (x) and \Phi 1=ff (x) are monotone if the scalar ff is chosen such that the map I \Gamma fff is nonexpansive. Furthermore, for strongly monotone and Lipschitzian map f , Gabay [7] and Sibony [20] actually showed that - ff (x) and \Phi 1=ff (x) are strongly monotone if the scalar ff is chosen such that the contractive. Throughout this paper, we use the standard concept "nonexpansive" map and "contractive" map in the literature to mean a Lipschitzian map with constant However, it is easy to give an example to show that - ff (x) and \Phi ff (x) are still monotone (strongly monotone) even when ff is chosen such that I \Gamma fff is not nonexpansive (contractive). For instance, let We see that the function f is co-coercive with modulus fff is not nonexpansive for remains monotone. As a result, the main purpose of this paper is to expand the results of Sibony [20] and Gabay [7]. We show that if f is co-coercive (strongly monotone and Lipschitz continuous, respectively), the monotonicity (strong monotonicity, respectively) of the maps - ff (x) and \Phi ff (x) can be ensured when ff lies in a larger interval in which the map I \Gamma fff may not be nonexpansive (contractive, respectively). The results derived in this paper are not obtainable by the proof based on the nonexpansiveness and contractiveness of maps. The other purpose of the paper is to introduce an application of the monotonicity of - ff (x) and \Phi ff (x): This application (see Section 3) is motivated by the globally convergent inexact Newton method for the system of monotone equations proposed by Solodov and Svaiter [21]. See also [22, 23, 24]). We propose a modified Solodov and Svaiter method to solve the monotone equations - ff 0: This modified algorithm requires no projection operations in the line-search step. 2. Monotonicity of - ff (x) and \Phi ff (x). It is known (see Sibony [20] and Gabay [7]) that if f is strongly monotone with modulus c ? 0 and Lipschitz continuous with constant L ? 0, then I \Gamma fff is contractive when nonexpansive, this in turn implies that - ff (x) and \Phi 1=ff (x) are both strongly monotone Similarly, it follows from Gabay [7] (see Theorem 6.1 therein) that if f is co-coercive with modulus fi ? 0; then I \Gamma fff is nonexpansive for and thus we can easily verify that - ff (x) and \Phi 1=ff (x) are monotone for In this section, we prove an improved version of the above-mentioned results. We prove that i) when ff lies outside of the interval (0; 2c=L 2 ), for instance, 2c=L 2 - ff - are still strongly monotone although I \Gamma fff , in this case, is not contractive, and ii) when ff lies outside of the interval (0; 2fi], for instance, 2fi ! remain monotone although I \Gamma fff is not nonexpansive. This new result on monotonicity (strongly monotonicity) of - ff (x) and \Phi 1=ff (x) for ff ? not obtainable by using the nonexpansive (contractive) property 4 Y. B. ZHAO AND D. LI of I \Gamma fff: The reason goes as follows: Let f be co-coercive with modulus fi ? 0 on the set S ' R n ; where Clearly, such a scalar fi is unique and is not a constant mapping. We now verify that I \Gamma fff is nonexpansive on S if and only if It is sufficient to show that if ff ? 0 is chosen such that I \Gamma fff is nonexpansive on S, then we must have ff - 2fi: In fact, if I \Gamma fff is nonexpansive, then for any x; y in S we have which implies that By the definition of fi, we deduce that ff=2 - fi; the desired consequence. Similarly, let f be strongly monotone with modulus c ? 0 and Lipschitz continuous with constant on the set S, where and We can easily see that is not a single point set. It is also easy to show that I \Gamma fff is contractive if and only if Since the map I \Gammafff is not contractive (nonexpansive, respectively) for ff - 2c=L 2 our result established in this section cannot follow directly from the proof of Sibony [20] and Gabay [7]. We also study the strong monotonicity of the perturbed fixed point and normal maps defined by and respectively. This is motivated by the well-known Tikhonov regularization method for complementarity problems and variational inequalities. See for example, Isac [10, 11], Venkateswaran [26], Facchinei [3], Facchinei and Kanzow [4], Facchinei and Pang [5], Gowda and Tawhid [8], Qi [16], Ravindran and Gowda [17], Zhao and Li [28], etc. It is worth mentioning that Gowda and Tawhid [8] showed that when perturbed mapping \Phi 1;" (x) is a P-function if f is a P 0 -function and K is a rectangular set. We show in this paper a sufficient condition for the strong monotonicity of - ff;" (x) and \Phi ff;" (x): The following lemma is helpful. Lemma 2.1. (i) Denote (2. MONOTONICITY OF FIXED POINT AND NORMAL MAPPINGS 5 Then (ii) For any ff ? 0 and vector b 2 R n ; the following inequality holds for all v 2 R n ; Proof. By the property of projection operator, we have Adding the above two inequalities leads to i.e., This proves the result (i). Given ff ? 0 and b 2 R n ; it is easy to check that the minimum value of ffkvk 2 +v T b is \Gammakbk 2 =(4ff): This proves the result (ii). We are ready to prove the main result in this section. Theorem 2.1. Let K be an arbitrary closed convex set in R n and K ' S ' R (i) If f is co-coercive with modulus fi ? 0 on the set S; then for any fixed scalar the fixed point map - ff (x) defined by (1.1) is monotone on the set S. (ii) If f is strongly monotone with modulus c ? 0 on the set S; and f is Lipschitz continuous with constant L ? 0 on S; then for any fixed scalar ff satisfying the fixed point map - ff (x) is strongly monotone on the set S: (iii) If f is co-coercive with modulus fi ? 0 on the set S; then for any 0 ! ff ! 4fi the perturbed map - ff;" (x) is strongly monotone in x on the set S. Proof. Let ff ? 0 and 0 - 2=ff be two scalars. For any vector x; y in By using the notation of (2.1) and Lemma 2.1, we have 6 Y. B. ZHAO AND D. LI If f is co-coercive with modulus fi ? 0, using " - 2=ff we see from the above that 'ff Setting in the above inequality, we see that for 0 ! ff - 4fi the right-hand side is nonnegative, showing that - ff is monotone on the set S: This proves the result (i). Also, if ff ! 4fi and 4fi ); the right-hand side of the above inequality is greater than or equal to rkx \Gamma showing that - ff;" is strongly monotone on the set S: The proof of the result (iii) is complete. Assume that f is strongly monotone with modulus c ? 0 and Lipschitz continuous with constant L ? 0: We now prove the result (ii). For this case, setting (2.2), we have that For it is evident that the scalar 4c 0: Result (ii) is proved. Similarly, we have the following result for \Phi ff (x): Theorem 2.2. Let f be a function from R n into itself and K be a closed convex set and K ' S ' R (i) If f is co-coercive with modulus fi ? 0 on the set S; then for any constant ff such that ff ? 1=(4fi); the normal map \Phi ff (x) given by (1.2) is monotone on the set S. (ii) If f is strongly monotone with modulus c ? 0 and Lipschitz continuous with constant L ? 0 on the set S; then for any ff satisfying ff ? L 2 =(4c); the normal map \Phi ff (x) given by (1.2) is strongly monotone on the set S. (iii) If f is co-coercive with modulus fi ? 0 on the set S; then for any constant ff ? 1=(4fi); the perturbed normal map \Phi ff;" (x), where strongly monotone in x on the set S. MONOTONICITY OF FIXED POINT AND NORMAL MAPPINGS 7 Proof. Let ff; "; r be given such that ff 0: For any vector x; y in u x and u y be defined by (2.1) with 2.1 we have kf (\Pi K and that further implies By using the above three inequalities, we have \Gammaf (\Pi K \Gammaf (\Pi K kf (\Pi K Let f be co-coercive with modulus fi ? 0 on the set S. Setting in the above inequality, and using the co-coercivity of f , we have 4ff kf (\Pi K 4ff kf (\Pi K For ff ? 1=(4fi); the right-hand side is nonnegative, and hence the map \Phi ff is monotone on the set S: This proves the result (i). By the co-coercivity of f , the inequality (2.6) can be further written as kf (\Pi K 8 Y. B. ZHAO AND D. LI the right-hand side of the above is nonnegative, and thus the map \Phi ff;" is strongly monotone on the set S: Result (iii) is proved. Finally, we prove the result (ii). Assume that f is strongly monotone with modulus continuous with constant L ? 0. For any vector x; y in we note that the equation (2.5) holds for any ff ? the equation (2.5) reduces to Given ff ? L 2 =(4c): Let r be a scalar such that Notice that Substituting the above into (2.7) and using inequalities (2.3) and (2.4), we have where the last inequality follows from the Lipschitz continuity and strong monotonicity of f: The right-hand side of the above is nonnegative. Thus, the map \Phi ff is strongly monotone on the set S: This proves the result (ii). The following result is an immediate consequence of Theorems 2.1 and 2.2. Corollary 2.1. Assume that f is monotone and Lipschitz continuous with constant L ? 0 on a set S ' K: ; then the perturbed map - ff;" (x) is strongly monotone in x on the set S. then the perturbed normal map \Phi ff;" (x) is strongly monotone in x on the set S. Proof. Let " 2 (0; 1) be a fixed scalar. It is evident that under the condition of the corollary, the function F strongly monotone with modulus continuous with constant L+ ": Therefore, from Theorem 2.1(ii) we deduce that if 0 the map - ff;" (x) is strongly monotone on S: Similarly, the strong monotonicity of \Phi ff;" (x) follows from Theorem 2.2(ii). The Item (iii) of both Theorem 2.1 and Theorem 2.2 shows that for any sufficiently small parameter ", the perturbed fixed point and normal maps are strongly monotone. This result is quite different from Corollary 2.1. When ff is a fixed constant, Corollary MONOTONICITY OF FIXED POINT AND NORMAL MAPPINGS 9 2.1 does not cover the case where " can be sufficiently small. Indeed, for a fixed ff ? 0, the inequalities 4" fail to hold when " ! 0: Up to now, we have shown that the fixed point map - ff (x) (respectively, the normal map \Phi ff (x)) is monotone if f is co-coercive with modulus fi ? 0 and ff 2 (0; 4fi] (respectively, ff 2 (1=(4fi); 1)): This result includes the known ones from Sibony [20] and Gabay [7] as special cases. Under the same assumption on f and ff; we deduce from the Item (iii) of Theorems 2.1 and 2.2 that the perturbed forms - ff;" and \Phi ff;" are strongly monotone provided that the scalar " is sufficiently small. In the succeeding sections, we will introduce an application of the above results on globally convergent iterative algorithms for VI(K; f) whose fixed point map or normal map is monotone. 3. Application: Iterative algorithm for VI(K; f) . Since - ff (x) and \Phi ff (x) are monotone if the function f is co-coercive and ff lies in certain interval, we can solve the co-coercive variational inequity problems via solving the system of monotone equation - ff Recently, Solodov and Svaiter [21] (see also, [22, 23, 24]) proposed a class of inexact Newton methods for monotone equations. Let F(x) be a monotone mapping from R n into R n : The Solodov and Svaiter's algorithm for the equation proceeds as follows: Algorithm SS. [21] Choose any x 0 Inexact Newton step. Choose a positive semidefinite matrix G k . Choose and where Line-search step. Find y being the smallest nonnegative integer m such that \GammaF Projection step. Compute repeat. As pointed out in [21], the above inexact Newton step is motivated by the idea of proximal point algorithm [2, 6, 19]. Algorithm SS has an advantage over other Newton methods that the whole iteration sequence is globally convergent to a solution of the system of equations, provided a solution exists, under no assumption on F other than continuity and monotonicity. Setting or \Phi ff (x), from Theorems 2.1 and 2.2 in this paper and Theorem 2.1 of [21], we have the following result. Theorem 3.1. Let f be a co-coercive map with constant fi ? 0: Substitute F(x) in Algorithm SS by - ff (x) (respectively, \Phi ff (x)) where ff ? 1=4fi). If - k is chosen such that C 2 - k - C 1 kF(x k )k, where C 1 and C 2 are two constants, then Algorithm SS converges to a solution of variational inequality provided that a solution exists. While Algorithm SS can be used to solve the monotone equations - ff and \Phi ff each line-search step needs to compute the values of - ff and \Phi ff that represents a major cost of the algorithm in calculating projection operations. Hence, in general cases, Algorithm SS has high computational Y. B. ZHAO AND D. LI cost per iteration when applied to solve \Phi ff To reduce this major computational burden, we propose the following algorithm which needs no projection operations other than the evaluation of the function f in line-search steps. Algorithm 3.1. Choose x 0 Inexact Newton Step: Choose a positive semidefinite matrix G k . Choose Compute where Line-search step. Find y being the smallest nonnegative integer m such that Projection step. Compute The above algorithm has the following property. Lemma 3.1. Let - ff (x) be given as (1.1). At kth iteration, if m k is the smallest nonnegative integer such that (3.2) holds, then y satisfies the following estimation: Proof. By the definition of - ff (x), the nonexpansiveness of projection operator and (3.2), we have \Gamma\Pi K Also, \Gamma- ff MONOTONICITY OF FIXED POINT AND NORMAL MAPPINGS 11 By (3.1) and positive semi-definiteness of G k , we have \Gamma- ff Combining (3.3), (3.4) and (3.5) yields \Gamma- ff The proof is complete. Using Lemma 3.1 and following the line of the proof of Theorem 2.1 in [21], it is not difficult to prove the following convergence result. Theorem 3.2. be a continuous function such that there exists a constant ff ? 0 such that - ff (x) defined by (1.1) is monotone. Choose G k and - k such that kG k k - C 0 and - are three fixed positive numbers and p 2 (0; 1]: Then the sequence fx k g generated by Algorithm 3.1 converges to a solution of the variational inequality provided that a solution exists. Algorithm 3.1 can solve the variational inequality whose fixed point mapping - ff (x) is monotone for some ff ? 0. Since the co-coercivity of f implies the monotonicity of the functions - ff (x) and \Phi ff (x) for suitable choices of the value of ff; Algorithm 3.1 can locate a solution of any solvable co-coercive variational inequality problem. This algorithm has an advantage over Algorithm SS in that it does not carry out any projection operation in the line-search step, and hence the computational cost is significantly reduced. 4. Conclusions . In this paper, we show some sufficient conditions for the monotonicity (strong monotonicity) of the fixed point and normal maps associated with the variational inequality problem. The results proved in the paper encompass some known results as particular cases. Based on these results, an iterative algorithm for a class of variational inequalities is proposed. This algorithm can be viewed as a modified Solodov and Svaiter's method but has lower computational cost than the latter. Acknowlegements. The authors would like to thank two anonymous referees for their incisive comments and helpful suggestions which help us to improve many aspects of the paper. They also thank Professor O. L. Mangasarian for encouragement and one referee for pointing out refs. [7, 20]. --R On the Douglas-Rachford splitting method and the Beyond monotonicity in regularization methods for Total Stability of Variational Inequalities Finite termination of the Applications of the method of multipliers to variational inequalities Existence and limiting behavior of trajectories associated with P0 equations A survey of theory Tikhonov's regularization and the complementarity problem in Hilbert spaces A generalization of Karamardian's condition in complementarity theory A decomposition property for a class of square matrices A unifying geometric solution framework and complexity analysis for variational inequalities The orthogonality theorem and the strong-f-monotonicity condition for variational inequality algorithms On the convergence of projection methods: application to the decomposition of affine variational inequalities regularization methods for variational inequality problems Regularization of P 0 Normal maps induced by linear transformations Control Optim. M'ethods it'eratives pour les 'equations et in'equations aus d'eriv'ees partielles non- lin'earies de type monotone A globally convergence inexact Newton method for systems of monotone equations A truly globally convergent Newton-type method for the monotone nonlinear complementarity problem A new projection method for variational inequality problems Further applications of a matrix splitting algorithm to decomposition in variational inequalities and convex programming An algorithm for the linear complementarity problem with a P 0 On condition for strictly feasible condition of On a new homotopy continuation trajectory for New classes of generalized monotonicity --TR --CTR Zhang , Weijun Zhou, Spectral gradient projection method for solving nonlinear monotone equations, Journal of Computational and Applied Mathematics, v.196 n.2, p.478-484, 15 November 2006
iterative algorithm;fixed point and normal maps;strongly monotone maps;variational inequalities;cocoercive maps
589026
Second-Order Algorithms for Generalized Finite and Semi-Infinite Min-Max Problems.
We present two second-order algorithms, one for solving a class of finite generalized min-max problems and one for solving semi-infinite generalized min-max problems. Our algorithms make use of optimality functions based on second-order approximations to the cost function and of corresponding search direction functions. Under reasonable assumptions we prove that both of these algorithms converge Q-superlinearly, with rate at least 3/2.This paper is a continuation of [E. Polak, L. Qi, and D. Sun, Comput. Optim. Appl., 13 (1999), pp. 137--161].
Introduction As is also the case with ordinary min-max problems, generalized min-max problems can be either finite or semi-infinite. Both are of the form where # is a smooth function and # n # m is a nonsmooth, vector-valued function. In the case of finite min-max problems, the components of #(-) are of the form 2 where the functions f are continuously di#erentiable and the sets q j := {1, 2, ., q j } are of finite cardinality 3 . In semi-infinite generalized min-max problems the components of #(-) are of the form where the functions Finite generalized min-max problems are obviously a special case of semi-infinite generalized min-max problems, since when the sets we can define the functions f j,k (x) by The best known generalized minimax problem occurs when an optimization problem with a max function cost and equality and inequality constraints is set up for solution using exact penalty functions, which results in an unconstrained optimization problem with f 0 (x) in (1.1) of the form: r where # e and # i are two positive penalty parameters. Another simple example occurs in a least squares problem involving max functions, in which case We denote components of a vector by superscripts and elements of a sequence or a set by subscripts. 3 Given any positive integer q, we use the notation q := {1, 2, ., q}. where each # j (x) is as in (1.3). As a last example, in trying to approximate a structural optimization problem the aim of which was to minimize the sum of the probability of failure 4 plus the cost of the steel in the structure, using linearizations of a state-limit function, we obtained a cost function of the form u#B# # is a ball of radius #, centered at the origin in the space of the random variables u, and g(x, u) is a smooth state-limit function which defined the boundary between outcomes that result in structural failure from those that do not. Functions of the form f 0 with #(-) as in (1.4), are the best known example of quasidi#erentiable functions and are treated in depth in [4]. Hence generalized min-max problems can be solved using algorithms developed for quasi-di#erentiable functions, see, e.g., [2, 3, 6, 7, 8, 9, 11, 21]. Under the additional assumption that #F (y)/#y j > 0 for all y # m and generalized min-max problems can be solved using transformations 5 into a smooth, constrained nonlinear programming problem (see e.g., [1, 5, 12]). Direct methods that depend on the assumption that #F (y)/#y j > 0 for all can be found, for example, in [6, 9] and in [20]. We will consider semi-infinite generalized min-max problems under the following hypotheses Assumption 1.1 We will assume that (a) The functions F (-) and # j (-, y), j # m, y # Y j , are at least once continuously di#er- (b) There exists a positive number c F > 0 such that #F (y)/#y j (c) The sets Y j are either compact sets of infinite cardinality, or sets of finite cardinality, of the form given in (1.5). 2 Parts (a) and (b) of Assumption 1.1 ensure that when both the F (-) and the # j (-) are convex, the function f 0 (-) is also convex. In addition, as we will see, when all parts of Assumption 1.1 hold, the function f 0 (-) has a subgradient. In [20], this fact was used in defining an optimality function and an associated descent direction for the problem 4 The probability of failure was given by # g(x,u)#0 #(u)du, with #(-) the normal probability density function. 5 These transformations result in a smooth problem with more variables than in the nonsmooth prob- lem. There is a fair bit of anecdotal evidence that they can induce considerable ill-conditioning in the smooth problem because they introduce arbitrary scaling. In general, solving nonsmooth problems using transformation techniques appears to be less e#cient than using algorithms that exploit problem structure. P, and in extending the Pshenichnyi-Pironneau-Polak (PPP) Algorithm 4.1 in [17] (see also [22, 13, 14]) to finite generalized min-max problems, and the Polak-He PPP Rate- Preserving Algorithm 3.4.9 in [17] (see also [15]) to semi-infinite generalized min-max problems. In this paper we show that techniques used in [18] and [19] for constructing Q- superlinearly converging algorithms for solving finite and semi-infinite min-max problems, of the form (1.1) and (1.2), can be extended to construct Q-superlinearly converging algorithms for the solution of both finite and semi-infinite generalized min-max problems. In Section 2, we present a continuous optimality function and its associated search direction function which, together with a backstepping rule, constitute the backbone of our algorithms. In Section 3, we extend the Polak-Mayne-Higgins Newton's method [18], for solving finite min-max problems, to generalized finite min-max problems. We prove the Q-superlinear convergence of this extention in Section 4. In Section 5, we make use of the theory of consistent approximations developed in [17] and the algorithm presented in Section 3 to develop an algorithm for solving generalized semi-infinite min-max problems and prove its convergence and Q-superlinear convergence. Section 6 is devoted to some numerical results to demonstrate the behavior of the proposed algorithms. We sum up in the concluding Section 7. 2. Optimality Conditions We will now present optimality conditions for the semi-infinite generalized min-max problem problem, defined in (1.1), (1.2), (1.4), both in "classical" form and in terms of an optimality function which leads to a superlinearly converging second-order algorithm. Lemma 2.1 [20] Suppose that F # is continuously di#erentiable and that # : # m is a locally Lipschitz continuous function that has directional derivatives at every x # n . Let f # be defined by Then, given any x # n , and any direction vector h # n , the function f 0 (-) has a directional derivative df 0 (x; h) which is given by Suppose that Assumption 1.1 is satisfied. Then it follows from Lemma 2.1 that the directional derivative of f 0 (-), at a point x # n in the direction h, is given by #F #F where When all the sets Y j are as in (1.5), (2.3) assumes the form #F where the functions f j,k (-) are defined by and the sets - q j (x) by Hence the following result is obvious. Theorem 2.2 Suppose that - x is a local minimizer for the problem (1.1), (1.2), (1.4). Then for all h # n , #F #F Furthermore, (2.8) holds if and only if 0 #f 0 (-x), where the subgradient #f 0 (-x) is given by #F Since (2.8) is a necessary condition of optimality, any point - x # n that satisfies (2.8) will be called stationary. When all the sets Y j are of the form (1.5), the expressions (2.8) and (2.9) assume the following #F #F (#(-x))#f j,k (-x) # . (2.11) Definition 2.3 We will say that # is an optimality function for problem (1.1), (1.2), (a) #(-) is upper semi-continuous, (c) for any - holds if and only if Assumption 2.4 We will assume that (a) the functions # in (1.1), (1.2), (1.4), are twice Lipschitz continuously di#erentiable on bounded sets, (b) the functions # are locally Lipschitz continuous, (c) there exist constants 0 < c # C < #, such that for all j # m, y and . (2.13)For the sake of convenience, for any x, h # n and w # m , we define u(x, h, w) := #F (#(x)), - and v(x, h, w) := 1 The reason for the introduction of the artificial variable w is as follows. The function is a perfectly good second-order approximation to F (#(x + h)), but unfortunately, it is not always convex and hence leads to problems in developing an algorithm for solving semi-infinite generalized min-max problems. By introducing the artificial variable w, we can define the function which, as we will later see, is a convex second-order approximation to F (#(x hence much more useful in algorithm construction. We define the function # n # and the associated search direction function # n by { min and { min h, w))} . (2.20) Note that We will shortly see that the function #(-) is an optimality function for the problem (1.1), (1.2), (1.4). For any y, #y # m , let Lemma 2.5 Suppose that Assumptions 1.1 and 2.4 are satisfied. For any y, #y # m , let (y, #y) be the solution set of (2.22). Then# (y, #y) is non-empty and compact and for any w # (y, #y), we have Proof. Since #F (y) > 0 and # 2 F (y) is positive semi-definite, for any w # 0 and #w# we have (y, #y) is nonempty and compact. Suppose that w # # (y, #y). Then w # satisfies the following first-order optimality conditions which follow directly from (and are equivalent to) the KKT conditions: i.e., #w #F (y) +# 2 F (y)(#y Clearly, (2.26) implies that for any w # (y, #y), we have #F (y) +# 2 F (y)#y +# 2 F (y)w # 0. (2.27)Lemma 2.6 Suppose that Assumptions 1.1 and 2.4 are satisfied. Then for any z # n there exists an # > 0 such that for all h # n with #h# and for all x # n with #x - z# we have i.e., defined by (2.17). Proof. Since h, -) is a convex quadratic function, any w # m satisfying the following first-order conditions #w, #F (#(x)) +# 2 F (#(x))( - is a solution of (2.18). Then, because #F (y)/#y j # c F , for every #(-) is uniformly continuous on any compact set and - we see that for any z # n there exists an # > 0 such that for all h # n with #h# and for all x # n with #x - z#, This implies that for all those h and x, we have Hence our proof is complete. 2 The above lemma shows that - is identical to - su#ciently small. This fact will be used in proving our superlinear convergence results. In general, - not convex in h. We will now show that - is convex in h. Lemma 2.7 Suppose that Assumptions 1.1 and 2.4 are satisfied. Then for any fixed is a convex function. Moreover, - f 0 (-) is continuous. Proof. First we will show that - is a convex function. For any y # m and #y # m , we have where It is easy to verify that S(#y) is a concave function and that its subgradient is given by # (y, #y)}, (2.34) (y, #y) is the solution set of (2.33). It now follows from (2.32) that - locally Lipschitz continuous in #y and that its subgradient gradient at #y is given by # (y, #y)}. (2.35) Since, by Lemma 2.5, for any w # # (y, #y) we conclude that for any s # - every is convex in (because it is the composition of a convex function with positive elements in the subgradient and a vector function whose components are convex). Next, we will prove that - continuous. First, since #F (y)/#y j # c F > 0 and positive semi-definite for all j # {1, - , m} and y # m , it follows from (2.22) (y, #y) is uniformly bounded in a neighborhood of given point (z, It now follows from Corollary 5.4.2 in Polak [17] that - F (-) is continuous. Hence which implies that - is continuous on # n with y := #(x) and #y := - The following theorem shows that #(-) is indeed an optimality function for the problem (1.1), (1.2), (1.4) and that the set-valued function H(-) is a descent direction function for Theorem 2.8 Suppose that Assumptions 1.1 and 2.4 are satisfied. Consider the functions #(-) and H(-) defined by (2.19) and (2.20), respectively. Then (i) For all x # n , (ii) For all x # n , where df 0 (x; h) is the directional derivative of f 0 at x in the direction h and # =2 (iii) For any x # n , 0 #f 0 (x) if and only if is the subgradient of f 0 (-) at x, defined in (2.9). Moreover, for any x # n such that have (iv) The set valued map H(-) is (a) bounded on bounded sets, (b) compact valued, and (c) outer-semicontinuous, i.e., for any x # n , H(x) is closed and, for every compact set S such that H(x) # there exists a # > 0 such that H(z) # z # B(x, #) := {y # n |#y - x#}. (v) The function #(-) is continuous. Proof. (i) admissible in (2.19) that #(x) # 0 for all x # n . directly from the definition of #(x) in (2.19) that for any h # H(x), #F Thus we have shown that (2.40) holds. (iii) For any x # n , let min u(x, h, u(x, h, 0) . (2.42) We will first prove that It is easy to see that Hence we only need to show that Suppose that there exists an h # n such that For any j # {1, - , m}, we have and Thus, there exists a constant C 0 such that which further implies that there exists a constant C 1 such that Since u(x, -, 0) is a convex function and u(x, 0, su#ciently small we have which contradicts that Next, with #f 0 (x) the subgradient of f 0 (-) at x, defined in (2.9), by emulating the proof of Lemma 2.5.5 in [17], we can prove that for any x # n , 0 #f 0 (x) if and only if therefore if and only if Finally we will show that for any x # n such that For the sake of contradiction, suppose that there exists an x # n such that {0}. Then there exist 0 such that u(x, h, w) which, together with the fact that v(x, h, w) # 0 implies that u(x, h, w) # 0. Hence we conclude that both u(x, h, u(x, h, w) < 0, which contradicts (2.43). However, implies that h, 0) is strongly convex in h and u(x, 0, (iv) According to our definition, for each h # n there exists a w(h) # m such that which, together with the facts that #F (y) > 0, y # m and v(x, h, w(h)) # 0, implies that Since for each j # {1, - , m} and h # n , it follows from (2.51) that for all y in any bounded neighborhood of x, Consequently, for any x # n , H(x) is nonempty and bounded and H(-) is bounded on bounded sets. Since - continuous (Lemma 2.7), it follows that H(x) is closed. Next we will prove that for every x # n and every compact set S such that there exists a # > 0 such that H(z) # not, then there exists an x # n and a compact set S such that H(x) # and a sequence {x i } converging to x such that H(x i ) # S #. Hence there exists a sequence {h i } such that is a compact set, without loss of generality, we can assume that By definition of H(x i ), f 0 (-) is continuous, it follows from (2.55), that which implies that - h # H(x). This contradicts that H(x) # Thus, we have shown that H(-) is outer-semicontinuous. (v) Finally, it follows from Corollary 5.4.2 in Polak [17] that # is continuous. 2 By introducing an additional variable, we can rewrite the expression for #(x), defined in (2.19), as follows Problem (2.57) is a quadratic problem with quadratic constraints. Under suitable assump- tions, (2.57) is actually a convex quadratic problem with convex quadratic constraints, and hence provides a convenient means for computing the optimality function value #(x) and an associated search direction h # H(x). Theorem 2.9 Suppose that Assumptions 1.1 and 2.4 are satisfied and the sets Y j are as in (1.5). For any x # n , let #(x) be the solution set of (2.57), i.e., any (p, h) #(x) solves (2.57). Then (i) Problem (2.57) is a convex quadratic problem with convex quadratic constraints. (ii) For x # n , #(x) is nonempty and compact and #(-) is outer-semicontinuous and bounded on bounded sets. (iii) If z # n is such that there exist a neighborhood N(z) of z and an # > 0 such that for any (p, h) #(x), x # N(z), we have Proof. (i) Under the conditions of Assumptions 1.1 and 2.4, # 2 F (#(x)) is positive semidefinite and for each j # {1, 2, - , m}, - strongly convex. Hence (2.57) is a convex quadratic problem with convex quadratic constraints. (ii) Since for all y in a bounded neighborhood N(x) of x and j # {1, 2, - , m}, it follows that for all y # N(x) and (p, #(y, we have # as #(p, h)#. (2.61) Hence, for all x # n , #(x) is nonempty and compact, and #(-) is bounded on bounded sets. The outer-semicontinuity of #(-) follows from the fact that #(-) is continuous and the constraint set in (2.57) is outer-semicontinuous. (iii) Since z # n is such that #(z). For any x # n , the KKT conditions for (2.57) are is the subgradient of - with respect to h. Suppose that (p, h) #(z). By (iii) of Theorem 2.8, we have Hence it follows from (2.62) and the fact that - which implies that positive semidefinite. Thus, we have proved that since #(-) is outer- semicontinuous, it follows that if x # z and (p, h) #(x), then It now follows from (2.62), (2.64), and the fact that for any y # m , #F (y)/#y j m}, that there exists a neighborhood N(z) of z such that for all x # N(z), the multiplier # in the KKT (2.62) must have all components positive and hence for all x # N(z), the KKT conditions for (2.57) become Thus, for any x # N(z) and j # {1, 2, - , m}, there exist nonnegative numbers - j,k satisfying # k#q j such that for any (p, h) #(x), where and for any k # q j such that we have We conclude from (2.65), (2.66) and (2.67) that for all x # N(z) and (p, h) #(x), #, p# where the last inequality follows from the fact that f j,k (x) # j (x) for all k # q j and m. By shrinking N(z) if necessary, we conclude from (2.67), (2.70) and Assumptions 1.1 and 2.4 that there exists a positive number # > 0 such that for all x # N(z) and 3. An Algorithm for Solving Generalized Finite Min-Max Problems An algorithm for solving generalized finite min-max problems is obviously of interest in its own right. However, we will also need it as a subroutine for our algorithms for solving generalized semi-infinite min-max problems. Hence, for the time being, we will assume that the sets Y j are of the form (1.5) and that the functions f j,k (-) are as in (2.6). As a result, our generalized finite min-max problem assumes the form (1.1), (1.2), (1.4), with min where, in view of Assumption 1.1, the functions F (-) and f j,k (-), j # m, k # q j are all continuously di#erentiable, where f j,k (-) are defined by (2.6). We are now ready to state an algorithm for solving generalized finite min-max prob- lems. This algorithm is a generalization of the Polak-Mayne-Higgins Newton's algorithm for solving finite min-max problems [18]. Algorithm 3.1 (Solves Problem (3.1)) Parameters. # (0, 1), # (0, 1), and # > 0. Step Step 1. Compute the optimality function value # i := #(x i ) and a search direction h i # according to the formulae (2.19) and (2.20). Step 2. If # Else, compute the step-size where N := {0, 1, 2, . Step 3. Set replace i by to Step 1. 2 Lemma 3.2 [20] Suppose that Assumption 1.1 holds. Then for any y, y # m such that y, Lemma 3.3 [20] Suppose that Assumptions 1.1 and 2.4 are satisfied. Then there exists a constant # > 0 such that for all x, x # n and # [0, 1], Theorem 3.4 Suppose that Assumptions 1.1 and 2.4 are satisfied and that all the Y j , are of the form (1.5), so that problem (1.1), (1.2), (1.4) reduces to problem (3.1). If {x i } # i=0 is an infinite sequence generated by Algorithm 3.1 and x # is the unique solution of (3.1), then {x i } # i=0 converges to x # . Proof. Suppose that {x i } # i=0 is an infinite sequence generated by Algorithm 3.1. Since f(-) is strongly convex by Lemma 3.3, the sequence {x i } # i=0 is bounded. Suppose that - x is an accumulation point of this sequence. Since the cost function f 0 (-) is continuous, f 0 (-x) is an accumulation point of the cost sequence. Hence, since, by construction, the cost sequence {f 0 Now, for the sake of contradiction, suppose that #(-x) < 0. Since for any x # n , H(x) is compact, and H(-) is bounded on bounded sets and is outer-semicontinuous ((iv) of Theorem 2.8), it follows from Theorem 5.3.7 (b) in Polak [17] that there exists a subsequence {j i } # i=0 of the integers such that x x and h j i # - h # H(-x), as i #. It follows from (ii) and (iii) of Theorem 2.8 that - be such that 0 < # < 1. Then it follows from the definition of the directional derivative of f 0 (-) that there exists a k # N such that Hence, for all # > 0 such that . (3.10) # := 1# . Then, since f 0 (-) and #(-) are continuous and h j i # - h, as i #, there exists a # > 0 such that for all x which shows that for all x . Next, since #(-) is continuous, there exists - # (0, #) such that for all x It therefore follows from the step-size rule (3.2) that for all x #), #(-x) . (3.12) implies that f 0 contradicting the fact that f 0 Hence we conclude that and therefore that - strongly convex, the whole sequence {x i } converges to x # . 2 4. Rate of Convergence of Algorithm 3.1 Proposition 4.1 Suppose that Assumptions 1.1 and 2.4 are satisfied and that - x is the unique solution of f 0 (-). Then for all x # n , Proof. By Lemma 3.3, f 0 (-) is a strongly convex function. Hence, for any x # n we have #F #F c #F c where - q j (x) is defined by (2.7). It now follows from (2.5) and (4.2) that Proposition 4.2 Suppose that Assumptions 1.1 and 2.4 are satisfied. Then for any compact convex set S there exists a # > 0 such that for any x, z # S, defined in (2.17). Proof. First, it follows from Polak [17, Lemma 2.5.4] or [18], that there exists a constant such that for any x, z # n , be a compact set, and let L 2 < # be a Lipschitz constant for # 2 F (-) on S, such that for any z # S, #F (#(z))# L 2 . (4.6) Then for all x, z # S, where L 3 := mL 2 L 1 /6. Since, by assumption, S is compact, it follows that there exists a positive number L 4 such that for all x, z # S, and Hence for all x, z # S, with 4 . (4.12) Similarly, we can prove that for all x, z # S, Thus we have shown that (4.4) holds. 2 Theorem 4.3 Suppose that Assumptions 1.1 and 2.4 are satisfied, that all the Y j , are of the form(1.5), so that problem (1.1), (1.2), (1.4) reduces to problem (3.1). If is a sequence constructed by Algorithm 3.1, in solving problem (3.1), then, {x i } # i=0 converges superlinearly with Q-order at least 3/2. Proof. First we will prove that after a finite number of iterations, the step-size # i stabilizes to 1, so that eventually x holds for the sequence {x i } # i=0 . We will then complete our proof by making use of results in [17, Corollary 2.5.8]. It follows from Theorem 3.4 that the sequence {x i } # i=0 converges to the unique minimizer x of f 0 (-). Hence we conclude from Theorem 2.8 that In view of this, we conclude from Lemma 2.6 that there exist a positive number # > 0 and a nonnegative integer i 0 such that for all Suppose that i 0 is su#ciently large to ensure that for all x# . (4.16) Then, making use of (4.1), we find that, for because - by (4.15). It now follows from Proposition 4.2 that there exists a # > 0 such that for all Now, by Theorem 2.9, there exist a positive integer and an # 1 > 0 such that for all Next, Proposition 4.2 and (4.15), imply that for all Hence, from (4.20) and (4.19), we have It now follows from (4.21) and the fact that h i # 0 as i # that for all i su#ciently large, We therefore conclude from [17, Corollary 2.5.8] or [18], (4.18) and (4.19) that {x i } # i=0 converges to - x superlinearly with Q-order at least 3/2. 2 5. An Algorithm for Solving Generalized Semi-Infinite Min-Max Problems We are now ready to tackle the generalized semi-infinite min-max problems defined in (1.1), (1.2), (1.4). Such problems can be solved only by discretization techniques. We will use discretizations that result in consistent approximations (as defined in Section 3.3 of [17]) and use them in conjunction with a master algorithm that calls Algorithm 3.1 as a subroutine. We will see that under a reasonable assumption, the resulting algorithm retains the rate of convergence of Algorithm 3.1. 5.1. Consistent Approximations Let N 0 be a strictly positive integer, and, for N # N 0 := {N 0 , be finite cardinality subsets of Y j , j # m, such that Y j,N # Y j,N+1 for all N and the closure of the set lim Y j,N is equal to Y j , j # m. Then we define the family of approximating problems PN , N # N 0 , as follows: PN min where N (x)), and for j # m, It should be clear that the approximating problems PN are of the form (3.1) and that one can define optimality functions # N (-) for them of the form (1.5). We will refer to the original problem (1.1), (1.2), (1.4) as P. Definition 5.1 [17] We will say that the pairs (PN , # N ), in the sequence {(PN , # N )} N#N0 are consistent approximations to the pair (P, #), if the problems PN epi-converge to P, (i.e., the epigraphs of the f 0 N (-) converge to the epigraph of f 0 (-) in the sense defined in Definition 5.3.6 in [17]), and for any infinite sequence {xN } N#K , K # N 0 , such that Assumption 5.2 We will assume as follows: (a) For every N # N 0 , the problem (5.1) has a solution. (b) There exists a strictly positive valued, strictly monotone decreasing function # : N #, such that #(N) # 0, as N #, and a K < #, such that for every there exists a y # Y j,N such that #y - y # K#(N). (5.4)For example, if for all j # m, Y j is the unit cube in # m j , i.e., Y then we can define Y I with a(N) := 2 N-N 0 . In this case it is easy to see that constructions can be obtained for other polyhedral sets. For any x, h # n and w # m , we define uN (x, h, w) := #F (#N (x)), - and where and We infer from (2.19) that the optimality functions # N (-), for the problems PN have the following form: { min h, w))}. (5.9) Since the cardinality of the sets Y j,N is finite, it is obvious that the # N (x) can be evaluated. As was also done in the Polak-Mayne-Higgins Rate-Preserving method [19] (see also [20]), we use an alternative optimality function for the problems PN for precision adjustment in our algorithm. This optimality function is defined by #F where #F with # > 0, a constant. Similarly (as in [20]), we define an alternative optimality function for the problem P by #F where #F with # > 0 the same constant as in (5.11). Proposition 5.3 [20] Suppose that Assumptions 1.1 and 5.2 are satisfied, and that for N (-) is defined by (5.2) and - # N (-) by (5.10). Let S # n be a bounded subset and let L < # be a Lipschitz constant valid for the functions # j (-) and # x # j (-) on q. Then there exists a constant C S < # such that for all x # S, N # N 0 , and | - 5.2. The Superlinear Rate Preserving Algorithm Algorithm 5.4 (Solves Problem (1.1), (1.2), (1.4)) Parameters. # (0, 1), # > 0, D > 0, # > 3. Step Step 1. Compute the optimality function value - according to (5.10) and (5.11), i.e., #F where #F (#N Step 2. If go to Step 3. Else, replace N by N + 1, and go to Step 1. Step 3. Compute the second optimality function value # N according to (5.9), i.e., { min and the corresponding search direction h i according to { min Step 4. Compute the step-size and go to Step 5. Step 5. Set to Step 1. 2 Remark. (a) It follows from Proposition 5.3 that - the loop consisting of Step 1 and Step 2 of Algorithm 5.4 yields a finite discretization parameter N i . For simplicity, we will assume that Algorithm 5.4 does not produce an iterate x i such that - (b) Note that the work needed to compute x i by Algorithm 5.4 increases with the iteration number i. 2 Lemma 5.5 Suppose that Assumptions 1.1, 2.4, and 5.2 are satisfied, and that Algorithm 5.4 has constructed a sequence {x i } # i=0 together with the corresponding sequence of discretization parameters {N i } # i=0 . If the sequence {x i } # i=0 has at least one accumulation point, then N i #, as i #. Proof. For the sake of contradiction, suppose that the sequence {x i } # i=0 has an accumulation point - x and that the sequence {N i } # i=0 is bounded. Then, because {N i } # i=0 is a monotonically increasing sequence of integers, there exists an i 0 # N , such that for all . Hence for the construction of the sequence {x i } # i=0 is carried out by Algorithm 3.1 applied to problem (5.1) with Furthermore, it follows from (5.18) that there exists an # > 0, such that - However, it follows from Theorem 3.4 that # N # and from the continuity of - that - # N #(x i where the infinite subsequence {x i } i#K , converges to - which contradicts the previous finding, and hence completes our proof. 2 Theorem 5.6 Suppose that Assumptions 1.1, 2.4, and 5.2 are satisfied, and that Algorithm 5.4 has constructed a bounded sequence {x i } # i=0 . Then every accumulation point - x of {x i } # i=0 satisfies - Proof. By applying Theorem 3.3.23 of [17] or Theorems in Section 5 of [16] and Lemma 5.5 to Algorithm 5.4, we obtain the desired result. 2 Theorem 5.7 Suppose that Assumptions 1.1, 2.4, and 5.2 are satisfied, and that Algorithm 5.4 has constructed a bounded sequence {x i } # i=0 . Then {x i } converges to the unique x of f 0 (-) with Q-order 3/2. Proof. First, by Theorem 5.6 and the fact that f 0 (-) has a unique minimizer - x, the whole sequence {x i } converges to - x. Hence, one can deduce from Theorem 4.3 and the proof of [17, Theorem 3.4.20], that {x i } converges to - x with Q-order 3/2. Since the derivation is straightforward, we omit the details here. 2 6. Some Numerical Results We now present some numerical results that illustrate the behavior of the algorithm proposed in Section 5 for generalized semi-infinite programming problems. The algorithm was implemented in Matlab. Throughout the computational experiments, the parameters used in the algorithm were 3.1. For both examples, we used the starting point (1, 1). The iteration of the algorithm is stopped at x i if for some N the meshsize #(N) < 0.005 and |# N developed in [24], which was based on a smoothing Newton method [23] for variational inequalities, was used to solve our search direction finding subproblem (2.57). Example 1. In this case, f 0 with and Example 2. In this case, the functions f are also defined as in Example 1, but F (-) is defined by The numerical results are summarized in Table 1 and Table 2. In these two tables the first column gives the residue ||x i - x|| (we used the last iterate as a substitute for x) and the discretization level (the meshsize at the present level is decreased to half of the previous one) refined by the master algorithm at the i-th step. It is clear from the numerical results that the rate of convergence is superlinear. Iteration Discretization level Table 1: Numerical results for Example 1 Iteration Discretization level Table 2: Numerical results for Example 2 7. Conclusion We have presented two superlinearly converging algorithms, one for solving finite generalized min-max problems of the form (1.1), (1.2), (1.3) and one for solving generalized semi-infinite min-max problems of the form (1.1), (1.2), (1.4). These algorithms were obtained by making use of the concepts underlying the construction of the Polak-Mayne- Higgins Newton's method [18] and the Polak-Mayne-Higgins Rate-Preserving method [19], respectively. The construction of the algorithms depends on the cost unction having a subgradient and their rate of convergence depends on convexity and second order smooth- ness, and hence Assumption 2.4 is essential. Our numerical results are consistent with our theoretical prediction that the algorithms converge Q-superlineary. Acknowledgement . The authors wish to thank Prof. R. T. Rockafellar for suggesting the function - as a way to get around the possible non-convexity of the function in h, as well as for the formula (2.57) which shows that our optimality function is defined by a quadratically constrained quadratic programming problem. --R "Nondi#erentiable optimization via approximation," "Quasidi#erentiable functions: necessary conditions and descent directions," "An algorithm for minimizing a certain class of quasidi#erentiable functions," "A smooth transformation of the generalized minimax problem," "A quadratic approximation method for minimizing a class of qua- sidi#erentiable functions," "A linearization method for minimizing certain quasidi#erentiable functions," "Randomized search directions in descent methods for minimizing certain quasidi#erentiable functions," "Descent methods for quasidi#erentiable minimization," "Proximal control in bundle methods for convex nondi#erentiable minimization," "The method of common descent for a certain class of quasidi#erentiable functions," "Algorithms for a class of nondi#erentiable problems," "On the rate of convergence of certain methods of centers," "Basics of minimax algorithms," "Rate-preserving discretization strategies for semi-infinite programming and optimal control," "On the use of consistent approximations in the solution of semi-infinite optimization and optimal control problems," Optimization: Algorithms and Consistent Approximations "A superlinearly convergent algorithm for min-max problems," "On the extension of Newton's method to semi-infinite for minimax problems," "First-Order algorithms for generalized finite and semi-infinite min-max problems," "On the minimization of a quasidi#erentiable function subject to equality-type quasidi#erentiable constraints," Numerical Methods in Extremal Problems (Chislennye Metody v Ekstremal'nykh Zadachakh) "A new look at smoothing Newton methods for non-linear complementarity problems and box constrained variational inequalities," "Numerical experiments for a class of squared smoothing Newton methods for box constrained variational inequality problems" --TR --CTR Huang , Defeng Sun , Gongyun Zhao, A Smoothing Newton-Type Algorithm of Stronger Convergence for the Quadratically Constrained Convex Quadratic Programming, Computational Optimization and Applications, v.35 n.2, p.199-237, October 2006
optimality functions;superlinear convergence;consistent approximations;generalized min-max problems;second-order methods
589032
A Primal-Dual Method for Large-Scale Image Reconstruction in Emission Tomography.
In emission tomography, images can be reconstructed from a set of measured projections using a maximum likelihood (ML) criterion. In this paper, we present a primal-dual algorithm for large-scale three-dimensional image reconstruction. The primal-dual method is specialized to the ML reconstruction problem. The reconstruction problem is extremely large; in several of our data sets the Hessian of the objective function is the product of a 1.4 million by 63 million matrix and its scaled transpose. As such, we consider only approaches that are suitable for large-scale parallel computation. We apply a stabilization technique to the system of equations for computing the primal direction and demonstrate the need for stabilization when approximately solving the system using an early-terminated conjugate gradient iteration.We demonstrate that the primal-dual method for this problem converges faster than the logarithmic barrier method and considerably faster than the expectation maximization algorithm. The use of extrapolation in conjunction with the primal-dual method further reduces the overall computation required to achieve convergence.
Introduction . In this paper we consider the image reconstruction problem in emission tomography. This problem is encountered in the eld of nuclear medicine, which is concerned with the study of organ function through radioactively labeled \tracer" compounds. The quantity of interest in this problem is the spatial concentration of radioactive emissions within the object under study. The quality of the reconstructed image can depend upon a number of factors including the number of emission events (i.e., counts) collected by the scanner and the method used to reconstruct the image. In studies that are characterized by poor counting statistics (that is, few counts), statistical reconstruction methods that model the Poisson nature of the emission process have been shown to improve image quality over traditional, non-statistical reconstruction methods [35, 57]. The low-count problem has generated considerable interest in the medical imaging community because low radiotracer doses and short scanning durations are highly desirable. The estimation of emission density in an organ is an inherently three-dimensional (3-D) process. Volume, or 3-D acquisition improves the counting statistics compared with 2-D acquisition (in which axially oblique coincidences are either physically or electronically blocked from detection) but increases the problem size considerably. Since the 3-D problem may involve image and measurement vectors with millions of elements, the amount of computation required to perform 3-D statistical reconstructions can be quite substantial. In our computational studies for example, the larger reconstructions consist of 1.4 million image variables which are reconstructed from a measurement vector with 63 million elements. As such, it is important to use reconstruction methods that converge rapidly. The statistical image reconstruction Ariela Sofer is partly supported by National Science Foundation grants DMI-9414355 and DMI 9800544. y Center for Information Technology, National Institutes of Health, Bethesda, Maryland, 20892- 5624 (johnson@mail.nih.gov). z Department of Systems Engineering and Operations Research, George Mason University, Fairfax, Virginia 22030-4444 (asofer@gmu.edu). C.A. JOHNSON AND A. SOFER problem can be posed as a constrained nonlinear optimization problem. In this paper we present a primal-dual method for performing statistical 3-D reconstructions in emission tomography that has been specialized to the intricacies of the application. We demonstrate the rapid convergence of our primal-dual method in computational studies on low-count, 3-D positron emission tomography (PET) data. This paper is organized as follows. In Section 2 we present the statistical model and develop the objective function. Section 3 reviews the EM method for ML recon- struction. In Section 4 we develop a primal-dual method for ML reconstruction and discuss initialization, stabilization, and extrapolation enhancements. Computational tests comparing the primal-dual results to a logarithmic barrier approach and the EM method on small animal data are presented in Section 5. Some concluding remarks are made in Section 6. 2. Statistical model and objective function. We begin our discussion by forming a nite parameter space for the image estimates, as is customary [20]. Consider the situation depicted in Figure 2.1 where a grid of boxes or voxels has been imposed over the emitting object (for simplicity, the Figure is depicted in 2-D; the concept is readily extended to 3-D). Given a set of measurements along lines of coin- cidence, we seek to estimate x expected number of counts emitted from voxel i. Let X i be the number of radioactive events emitted from voxel are assumed to be independent Poisson-distributed random variables with mean system matrix C 2 < nN is used to model a number of physical eects including spatially dependent resolution and attenuation. The elements C i;j of the system matrix represent the probability that an event emitted from voxel i will be detected by detector pair (coincidence line). The number of events emitted from voxel and detected at coincidence line j is therefore are also independent Poisson variables. The measurements y j are thus realizations of sums of independent Poisson variables y The above is a considerably simplied model of the actual measurement process; for further discussion on its validity to the present situation, see [24]. Given our simplied Poisson model, the likelihood may be written as Y Y e The ML objective function is formed by taking the log likelihood log Ignoring the constant term, we dene our objective function fML (x) as is a vector of 1's, so that q is the sum of the columns of C (which need not necessarily be 1). Dening Fig. 2.1. Relationship between estimate x i and measurement y j . Shown here is the case of PET, where emission-count measurements are taken along coincidence lines from pairs of detectors. A nite parameter space is formed by imposing a grid of voxels over the emitting region. The estimate of the expected emission intensity within voxel i is x i . to be a forward transformation, we can write the gradient and Hessian of the objective function, respectively, as The Hessian is negative semidenite (since so the objective function (2.1) is concave. Thus, any local maximum will also be a global maximum. Equation (2.2) sheds some insight into the computational costs associated with maximizing the objective function. Given a current solution estimate x k ; computing the gradient requires rst computing a forward transformation ^ then computing a backward transformation k y from the forward transformation. The costs of performing the forward transformation and backward transformation are similar and together dominate the computation associated with iterative reconstruction methods, especially in large scale. We shall revisit this computational structure, which is common to all iterative reconstruction methods. Since the underlying activity distribution is non-negative, the ML reconstruction problem is a constrained optimization problem with lower-bound constraints: maximize fML (x) subject to x 0: The ML objective function has a nite maximum and compact level sets on x 0 [36]. 2.1. Maximum a posteriori reconstruction. Without regularity conditions on x, estimating the spatial emission distribution is a statistically ill-posed problem [7, 33]. The fully converged ML reconstruction, being dominated by noise and edge artifact, is not generally of biomedical interest [55]. Regularization can be included 4 C.A. JOHNSON AND A. SOFER in the objective function by introducing a Bayesian formulation [20, 37]. Given prior probabilities P fxg and P fyg for the image and measurements, respectively, we dene the posterior probability The estimate of x is then obtained by maximizing the posterior probability P fxjyg. A common choice for the image prior is the Gibbs distribution P although other priors (e.g., Gaussian, Gamma) have been investigated [37, 39]. The popularity of Gibbs priors stems in part from their ability to capture the local correlation property of images [19]. The energy function R is dened as a sum of "potential" functions designed to discourage non-smoothness in a neighborhood denotes the neighborhood of voxel i. In order to maintain concavity and twice continuous dierentiability in the objective function, the potential function V i;l is chosen to be convex with continuous rst and second derivatives. In our studies we have used the potential function V i;l z log z and - is a shaping constant that we typically set to 1 [38]. For maximum a posteriori (MAP) reconstructions, the objective function is the log-posterior likelihood log P fxjyg. Ignoring a constant, our objective function become The MAP reconstruction problem can also be posed as a constrained optimization problem maximize fMAP (x) subject to x 0: We note for future reference the following: Although the function R (with the potential function (2.5)) is concave it is not strictly concave. only for vectors v that are a scalar multiple of the unit vector e N , and since e T negative denite and that fMAP is strictly concave [38]. In addition, fMAP has a nite maximum and bounded level sets on x 0 [37]. 2.2. The optimization problem. For convenience of notation, let us pose the reconstruction problem as a constrained minimization problem: subject to x 0; (2. 0. The case corresponds to the unregularized ML; in general we shall be more interested in the fully converged MAP where The Karush-Kuhn-Tucker (KKT) rst-order necessary conditions for optimality of (2.8) at a point x are existence of Lagrange multipliers so that x is the Lagrangian function. Due to the strict convexity of f , the second-order su-ciency conditions are satised, and x is the unique minimizer of f . 3. The EM algorithm. The expectation maximization (EM) method, as presented by Dempster, Laird, and Rubin [8] for ML estimation, is an iterative algorithm for computing ML estimates when the measurements are viewed as incomplete data. Shepp and Vardi [53] and Lange and Carson [36] applied the EM method to emission and transmission tomography problems, respectively. The EM algorithm has been proven to converge to an optimal solution of (2.4) [36, 56]. The EM algorithm for emission tomography can be derived [56, 27] from the optimality conditions for the reconstruction problem. For the unregularized problem can be written as diag (x), and Premultiplication by X , and utilizing the complementary slackness condition yields or since Applying a xed-point algorithm x to the above equation yields the EM update equation where x k is the current image estimate, diag . Given a positive initial solution x the algorithm maintains non-negativity at every iteration and converges to a xed point x which is an optimal solution of (2.4). The asymptotic rate of convergence is governed by the spectral radius of rM which is typically very close to unity. In one example using reasonable assumptions about the scanner geometry, the lower bound of the spectral radius was calculated to be .99938 [17]. Indeed, EM has been observed to converge very slowly, especially close to the optimal solution. The slow convergence of the EM algorithm has limited its clinical applicability. The cost of one EM iteration is equivalent to the cost of one gradient calculation. In MAP-EM, the presence of the regularizing term in (2.6) precludes a closed-form update equation such as (3.1) for ML-EM. We mention two algorithms that are commonly used for MAP-EM reconstructions: the \one step late" (OSL) algorithm and 6 C.A. JOHNSON AND A. SOFER DePierro's algorithm. Green's OSL algorithm approximates R (x) with the constant , thereby permitting a closed-form approximated update [16, 17] . OSL converges to the MAP solution provided that , where is an upper threshold for the prior strength. DePierro's algorithm is a \true" MAP-EM implementation that substitutes the convex function R (x) with a separable, convex, and twice continuously dierentiable function R x; x k R (x), so that separable maximizations can be performed on the variables [9, 10]. Regularization improves the convergence rate of EM, with larger prior strengths resulting in lower spectral radii. However, for reasonable prior strengths (mild to moderate smoothing), the convergence rates of OSL and DePierro's algorithm are still quite close to unity. The EM update formula on the right-hand side of (3.1) follows Kaufman [27], who was the rst to pose the EM algorithm as an optimization algorithm (namely, a scaled steepest-ascent method). This representation allows for the inclusion of a line search [27, 28] to accelerate the method's performance. Likewise, the MAP update (3.2) can be enhanced by a line search. Several other approaches for solving the maximum likelihood estimation problem have been proposed. These include preconditioned conjugate gradient techniques [27, 28, 34, 42] or truncated-Newton methods [27, 28]. The nonnegativity constraints are maintained either be limiting the step length or by using a bending line search. The paper [44] explores active set methods, while [43] enforces nonnegativity via a quadratic penalty in the objective. In other work [29, 30, 31] a penalized least-squares objective is used instead of the maximum likelihood. These problems are solved by a preconditioned conjugate gradient and use specialized techniques to drive the complementary slackness to zero. There is considerable debate within the PET community regarding the appropriate model for reconstruction. It has long been observed that the unregularized maximum likelihood estimator gives grainy images. However if the EM algorithm is stopped early, the resulting solution often produces images of acceptable quality. For this reason some researchers argue that early termination is a form of smoothing, and that no regularization is needed. Proponents of MAP argue that the approach allows the user to control the amount of regularization through the parameter, and that the regularized objective function is better conditioned. In either case, it has been observed that EM-type algorithms may lead to nonuniform convergence. In particular, the algorithms may converge slowly in \cold spots" (regions of low activity within regions of activity) and in areas of isolated activity within cold spots. The use of an interior-point algorithm oers the hope of more uniform convergence. 4. A primal-dual approach. The drawbacks of the EM algorithm motivate our investigation into interior-point approaches for the ML and MAP reconstruction problems. As is clear from (2.1), the objective function can be undened outside the feasible region x 0. Thus the ML and MAP reconstruction problems would appear to be \natural" candidates for interior-point algorithms. The reconstruction problem is especially suited to an interior-point approach, because its output is a gray-scale image. Whether a particular value is exactly \zero" or just very close to zero is immaterial. Slight inaccuracies below the gray scale threshold are inconsequential; obtaining an image rapidly is a neccessity. Primal-dual methods have enjoyed considerable success in linear programming [18, 32, 40], and have recently been proposed for nonlinear programming [5, 13, 41]. Although they are closely related to the logarithmic barrier method, primal-dual methods may pose some advantages. In the logarithmic barrier method, the Lagrange multiplier estimates may be inaccurate when the primal variables are not close to the barrier trajectory [11]. Primal-dual methods oer the potential of improved \center- ing" over barrier methods. Given the size of the current problem, the developments presented here must be suitable for large-scale parallel computation. In a manner similar to classical barrier methods, primal-dual methods attempt to follow the \barrier trajectory," a smooth trajectory characterized by a barrier parameter [12]. The points along the trajectory satisfy a perturbed version of the KKT conditions: Dening ng and our method maintains (4.3) while attempting to solve (4.1), (4.2), that is Xen en 0: Given the point x and the barrier parameter , the search direction prescribed by Newton's method satises the \unsymmetric" primal-dual equations [41]: I Elimination of the (1,2) block of the matrix in (4.5) yields the reduced system where the \condensed" primal-dual matrix is given by We have implemented an algorithm in which the primal and dual variables are permitted to take separate steplengths: The primal steplength x is chosen to ensure su-cient decrease in the merit function log Observe that F (x; ) is simply the logarithmic barrier function and that 8 C.A. JOHNSON AND A. SOFER is identical to the right-hand side of (4.6) for . The unconstrained minimizer x () of F (x; ) satises the perturbed KKT conditions (4.1)-(4.3) with corresponding multiplier i Furthermore, the solution of the condensed primal-dual Newton equation (4.6) is guaranteed to be a descent direction of the merit function for > 0, since and M is positive denite. We shall discuss in further detail the computation of the primal search direction and step length. The formula for the dual step length follows a suggestion by Conn, Gould, and Toint (CGT) [5]. If lies component-wise in the interval (where is a constant parameter that we have set to 100) then otherwise nd 0 < < 1 such that subject to k+1 being in the interval (4.9). These conditions on the dual step might appear at rst glance to be overly restrictive but are actually designed to give maxi- mum exibility in the choice of k+1 . CGT use these bounds on and nonsingularity of M to prove that, for any xed parameter value , the minimization of F (x; must be successful, that is, eventually a solution is found that satises the perturbed KKT conditions (4.1)-(4.3). In general it is neither necessary nor desirable to reach full subproblem conver- gence. Rather, we have implemented a \short-step" algorithm in which only one primal-dual step is usually needed before adjusting . Setting the barrier parameter is an important consideration in primal-dual algorithms, and has a strong in uence on the convergence rate. A reduction in k is performed whenever the \-criticality" conditions [5, 54] are satised: k+1 are constant parameters. If the above conditions are satised, the barrier parameter is reduced according to where is a constant parameter such that A consequence of (4.14) is that k cannot increase. Furthermore, since the minimization of F (x; ) must be successful, a -critical solution (a weaker requirement) must eventually be found. Thus it is impossible for k to be non-decreasing. Using this argument, CGT prove that the algorithm must converge to a KKT solution [5]. In practice we nd that both the primal and dual direction vectors are well scaled, and that x and are both typically close to 1. By far the most costly operations are computing the primal direction p x and updating the gradient rF (x), as we shall explore. In contrast, the costs of the line search for the primal steplength, the computation of the dual search direction (4.7), and the dual line search (4.9) are relatively insignicant. From empirical evidence in our computational studies, we have found that a \short-step" algorithm with gradual reduction in achieves the fastest convergence to the KKT conditions. Specically, we dene # These parameter values enable the -critical conditions to be met after only one primal-dual step for most subproblems. 4.1. Computing the primal direction. For large problems, factoring the condensed primal-dual matrix M or even forming the Hessian r 2 f (x) would be prohibitive due to the size of the matrix (376,000376,000 for even the smaller reconstructions being considered in this paper) and the enormous amount of computation that would be required. Thus we must consider methods for approximating the Newton direction in (4.6). The approach we have successfully applied to this problem is motivated by the truncated-Newton [6] method of unconstrained optimization. The search direction is an approximate or truncated solution to the Newton equations [47, 49] An early-terminated conjugate gradient (CG) iteration [21] is used to obtain an approximate solution to (4.15). An equivalent statement of (4.15) is we seek to nd the direction p x that approximately minimizes the quadratic Q (p x x reasonable and eective truncation point for (4.15), based on the monotonicity of Q (p x ), is proposed in [48]; the CG is terminated at subiteration l if x x The CG termination rule (4.16) has been an important component of the reconstruction software in that it consistently yields a well-scaled primal direction vector as long as s , where s is a threshold value below which stabilization is required (we shall discuss the < s case in Section 4.2). The CG method does not require storage of the Hessian or condensed primal-dual matrix, but rather only application of matrix-vector products. From (2.3) we can write the rst term of the matrix-vector product for an arbitrary vector v 2 < n . Computationally, (4.17) consists of a forward transformation (C T v) followed by a diagonal scaling (^y is already available from the computation of rf (x)), followed by a backward transformation (premultiplication by C). To be explicit, recalling (4.8), we have C.A. JOHNSON AND A. SOFER where r 2 R (x) v can be computed exactly without incurring signicant computational expense. The forward-and-back-transformation operation in (4.17) dominates the computational cost of a CG iteration. This operation is computationally similar to computing the gradient, or one EM iteration. Some authors advocate solving simultaneously for p x and p , using the full unsymmetric primal-dual equations (4.5), or an equivalent symmetrized system [13, 14, 52, 61]. The unsymmetric primal-dual matrix in particular remains nonsingular, and its condition number remains bounded as ! 0 [12, 41], when the standard conditions of a constraint qualication, strict complementarity, and the second-order su-cient conditions are satised at the solution. In our application, due to the size of our problem, we must use an iterative method. We believe that solving a symmetric system via a symmetric solver such as the CG would be more e-cient than solving the full unsymmetric system via an unsymmetric iterative solver such as GMRES (even though our symmetric system is ill-conditioned), since the the amount of work and storage required per iteration in GMRES increases linearly with the iteration count. An advantage of using the condensed system (4.6){(4.7) is that although the primal search direction is computed inexactly, the equation for maintaining complementarity (4.7) is maintained. In practice we nd that the resulting primal and dual direction vectors are both well scaled, and that x and are typically close to 1. 4.1.1. Preconditioning. The use of a preconditioner with the CG is essential for a competitive algorithm. Since every CG subiteration is as costly as a gradient evaluation or EM iteration, it is highly desirable to obtain a quality direction vector in as few CG iterations per subproblem as possible. We have investigated a number of preconditioners, including FFT-based preconditioners that model the approximately Toeplitz-block-Toeplitz nature of CC T with a circulant-block-circulant approximation [2, 3], high-pass lter approximations to the FFT-based preconditioner [4], the EM preconditioner XQ 1 [34], the exact diagonal of M , and diagonal Hessian approximations [46]. Of the above preconditioners, by far the best-performing was the exact diagonal of M , which can be computed at reasonable cost: Note that the rst right-hand side term in (4.18) is similar in form to a backward transformation, although a bit more expensive due to the squaring operations. We have found that the preconditioned CG method using an exact diagonal preconditioner in the form of (4.18) almost always requires using fewer than 10 iterations to achieve (4.16), regardless of the size of the problem. In many cases, only 3 or 4 CG iterations are required. Moreover, the directions produced using an exact diagonal preconditioner are well scaled (usually resulting in primal step sizes of near 1), and lead to rapid descent. In contrast, the other preconditioners did not perform well. Already in the initial subproblems they tended to yield a poorly-scaled search direction, which in turn, resulted in small steplengths. Subsequent calls to the CG suered further from this problem, and the algorithm made little progress. This behavior was particularly surprising for the block-circulant FFT-based preconditioners. These preconditioners perform very well in other reconstruction methods, especially in least-squares methods where the block-circulant approximation is well matched to the Hessian structure. We were motivated to try them for our problem because the ML Hessian is almost block circulant. But because of the strong diagonal component in M and its spatially- variant dependence on y, ^ y, x, and , shift-invariant Toeplitz models of M yield a poor approximation in our method. 4.1.2. Line search. For ML and MAP reconstructions, knowledge of the structure of the objective function can lead to a substantial reduction in the cost of implementing a line search over a more naive approach. Specically, after the search direction p x has been found, and once a forward transformation ^ been computed, it is possible to compute the objective function and rst and second directional derivative values at the trial points x k at nearly negligible cost. To see this, note that ^ and therefore [27, 28] Similar expressions exist for the directional rst and second derivatives [24]. After the initial forward transformation to compute ^ no further forward- or back-transformation operations are required during the line search at any of the trial points. The forward transformation ^ w can be re-used, so that only one backward transformation is subsequently required to update the gradient. The above observations and the well behaved convex nature of the objective function have permitted us to implement a highly accurate but low-cost Newton line search. Due to the low cost of each step we have chosen a relatively strict tolerance of 0:05 on the Wolfe condition for termination of the line search: We nd this line search technique to be highly eective and, in no small part, responsible for the positive results we report. 4.2. Stabilization. A well known property of the Hessian of the primal barrier function is its increasingly ill-conditioned nature as ! 0 [45]. Analogous results hold for the condensed primal-dual matrix: as the solution is approached the matrix becomes increasingly ill-conditioned. (For a detailed analysis see the paper by Wright [60]). In [50], Nash and Sofer developed an approximation to the Newton direction for the logarithmic barrier, that avoids the structural ill-conditioning of the barrier Hessian and is suitable for large-scale problems. The direction is the sum of two vectors, one in the null space of the Jacobian of the active constraints, and the other orthogonal to it. The associated decoupling is based on a prediction of the binding set at the solution. We have recently adapted this approximation to the condensed Newton equations arising in primal-dual methods. Although our derivation is valid for general nonlinear constraints, we present it here for the special case of bound constraints in the context of (4.6). We will assume in the following that strict complementarity holds at the solution, that is, 0g to be the index set of binding constraints at the solution, and ^ n to be the number of binding constraints at the solution. We will assume that 0 < always the case in reconstructions of practical interest. Dene I 0g the set of nonbinding constraints. Let x I be the subvector of variables that are positive at the optimal solution, and x J the subvector of variables that are zero at the optimal solution. Assume also that the 12 C.A. JOHNSON AND A. SOFER variables are ordered so that the positive variables are rst, i.e., x I x J The Hessian of the objective function will then be similarly partitioned, as will the condensed primal-dual matrix I;J MJ ;J I I H I;J where X I , XJ , I and J are the diagonal matrices of the associated components of x and . We will assume that the sequence of iterates (x; ) generated by the primal-dual satises the following properties, when is su-cently small: Here we dene there exist constants 0 < l < u so that l kk u for all su-ciently small > 0. We say that a vector or matrix is () if its norm is (). We also dene = O() if there exists some positive constant u so that kk u for all su-ciently small > 0. We will also assume that near the solution the Hessian is reasonably well condi- tioned, so that Now the diagonal terms of MJ ;J are O(1=), and become unbounded as ! 0. In contrast, the diagonal terms of M I;I dier from those of the reduced Hessian H I;I by O(), and the condition of M I;I thus re ects that of the constrained problem. The condensed primal-dual matrix M can then be shown to have \large" eigenvalues of magnitude (1=), and \small" eigenvalues that dier from those of H I;I by O(), and have magnitude (1). The condensed primal-dual matrix thus suers from the same structured ill-conditioning as the barrier Hessian. For small values of we propose approximating the primal Newton direction p x , by a direction ~ whose null- and range-space components are computed as follows: ~ The system for computing the component ~ I x involves the well conditioned matrix I;I , and can be solved exactly or inexactly via the conjugate gradient method. The computation of ~ x is straightforward. Thus, the ill-conditioning of the condensed primal-dual is avoided. We will show now that under the assumptions above, ~ so that the accuracy of the approximation increases as the solution is approached and the potential harm from ill-conditioning increases. Using the well known formula for the inverse of a partioned matrix (see e.g. [51, 61]) it follows that ~ I I;I rF I +M I;J (XJ 1 I;I rF I (XJ 1 where I;I Now by denition so that G Note further, that I whereas It follows that ~ I and ~ so that ~ In [50], Nash and Sofer prove (for the case of the Newton direction arising from the logarithmic barrier objective function) that, for su-ciently small ; the vector computed using an approximation similar to (4.19) and (4.20) yields a descent direction with respect to the logarithmic barrier objective function. The proof is readily extended to the present primal-dual case; thus p x is a descent direction for the merit function F (x; ). We have found that, for the present problem, the above approximation to the Newton direction is useful for values of of order 10 4 or less. Recently Wright [60] showed that the errors generated by backward-stable numerical methods (various Cholesky factorizations and Gaussian elimination with partial pivoting) for solving (4.6) are not magnied by the structured ill-conditioning. These methods are inappropriate for our large problems which involve potentially millions of variables. Instead we nd an approximate solution using a CG iteration. When working in inexact arithmetic with large numbers of variables, the convergence rate of the CG method depends on the condition of M [15]. Thus the structural ill-conditioning in M can lead the CG iteration to spend an unnecessary amount of work in computing Further, as we have observed, the criterion for terminating the CG may be overly optimistic in an ill-conditioned system, so that the resulting direction is poorly scaled as The potential eect of ill-conditioning is illustrated through an example in Table 4.1. This example was encountered during development and motivated the incorporation of stabilization into the algorithm. Starting at the subproblem the primal steplength, dual steplength, and ncg (the number of CG it- erations), are listed for both the non-stabilized and stabilized cases. This test was terminated at T x=n 7:5 10 5 . Note that in the non-stabilized case, the number of CG iterations from the rst subproblem in the test to termination is signicantly lower in the stabilized test than the non-stabilized test. Note also that in many of the non-stabilized subproblems, either the primal or dual steplength is small, indicating a poorly scaled direction or loss of accuracy. 14 C.A. JOHNSON AND A. SOFER Table An example of the eect of stabilization. The number of CG iterations, ncg, is counted from the beginning of the subproblem. The termination condition in this example is non-stabilized stabilized 7.08E-5 0.392 1.000 46 3.25E-5 5.83E-5 1.000 0.166 51 4.77E-5 1.000 62There has been much recent interest in stabilization methods that do not require a prediction of the active set [13, 14, 59]. These approaches are based on factorization methods which are unsuitable for a problem as large as the present one. The argument against stabilization methods that require a prediction set is that the active set is unknown in interior-point methods. We argue that, close to the solution in the emission tomography reconstruction problem, an accurate prediction of the active set can be made. In our problem, the constraints have a simple interpretation. The positive variables correspond to those voxels containing at least some radioactive tracer, while the zero-valued variables correspond to those voxels that lack any tracer activity. Close to the solution, when becomes su-ciently small that stabilization is appropriate, the set of binding constraints is obvious and can be conservatively identied with a -dependent threshold. 4.3. Extrapolation. Fiacco and McCormick showed that the solutions x () at the perturbed KKT solutions form a unique dierentiable trajectory in [12]. The perturbed KKT conditions (4.1){(4.3) dene a \central path" as ! 0. Thus, a successful algorithm may be able to move both \along" and \toward" the path. As discussed in [12], from the subproblem solutions fx ( l the trajectory can be approximated as a polynomial l=k r c l l ; where r is the degree of the approximating polynomial and c k r are r vectors of coe-cients. Using the approximation in (4.21), we nd a direction x such that l=k r c l l x and set to be a prediction to the next subproblem's primal solution. Here x k is the computed (approximate) subproblem solution for Primal feasibility is maintained by the steplength is the maximum steplength that does not violate non-negativity in x. Then, in the manner of (4.7), we compute a dual direction vector according to The dual vector is then moved according to 4x; which requires another dual line search to minimize (4.10). The resulting point serves as a starting point for the 1)st subproblem, a prediction to the solution at k+1 . The extrapolated primal-dual method can be viewed as a predictor-corrector algorithm, with the extrapolation (4.22 and 4.24) serving as the \predictor" step, and the subproblem minimization serving as the centering or \cor- rector" step [23]. The degree r of the approximating polynomial is 1 when predicting the 3rd subproblem, 2 for the 4th, and 3 for the 5th and beyond. We have experimented with line searches in conjunction with (4.22), but often 1, and hence the line search just yields . For this reason, we have found that (4.24) yields a more eective dual direction than does the equivalent of (4.7) in the context of extrapolation. Although the extrapolated search direction x can often be poorly scaled (i.e., 1), we have observed that the directions produced are always descent directions to the merit function and lead to a signicant decrease in the objective function f . A number of reconstructions were performed in which was computed by extrapolating the dual solution vector (rather than computing it via (4.24)); the discouraging nature of the results led us to abandon direct extrapolation of the dual vector in favor of (4.24) which is highly eective in comparison. Following extrapolation, a gradient evaluation is required to update the vector the primal-dual algorithm requires between 12 and 25 subproblems to perform a 3-D MAP reconstruction, extrapolation adds that many gradient evaluation operations to the computational cost. So extrapolation is only economical if it reduces the computational burden by at least as much as it adds. Our experience has been that for some data sets, the cost of extrapolation is well worthwhile but for other data sets the benets were only marginal. Extrapolation thus appears to serve as somewhat of a safeguard against di-cult problems. In an extrapolated primal-dual reconstruction, the convergence measure does not decrease as monotonically as in a primal dual reconstruction without extrapolation. Certain extrapolated steps seem to cause the algorithm to \get ahead of itself," but this eect is transient. On the studies we've performed, the algorithm does ultimately converge to an accurate solution with extrapolation. 4.4. Initialization. The choice of the initial barrier parameter may have a substantial eect on the algorithm. If the parameter is too small, the rst subproblem may have extreme di-culty due to ill conditioning; if the parameter is too large, then many (unnecessary) subproblems will be required to solve the problem. Proper initialization of the barrier parameter involves nding the most suitable point on the barrier trajectory based on the initial solution x and the measurement data y. Recalling the perturbed necessary conditions in (4.1), if the initial solution were to be on the central path, it would satisfy C.A. JOHNSON AND A. SOFER Pre-multiplying by ^ T we arrive at r This suggests the following rule for initialization, which we nd quite eective: r Another, similar, initialization rule is motivated by the goal of nding an initial value 0 so that While (4.26) cannot be solved exactly, we can try to nd a 0 that results in a point ^ that is close to the barrier trajectory according to, say, the 2-norm. This motivation leads to an alternative initialization rule [51] During the course of development, both initialization rules were tried on certain data sets. Although both initialization rules performed well, reconstructions initialized with (4.25) usually reached the optimal solution in slightly less overall work than those initialized with (4.27). The initial estimate for ^ used most frequently was in each case a positive uniform eld. A discussion on the rationale of using a uniform eld for and on criteria for choosing the constant value of the primal initial solution may be found in [24]. Alternative choices for the initial dual vector may be preferable, and an investigation into this question may be worthwhile. 4.5. Termination. Given that subproblem termination is based on the -criti- cality conditions (4.11) and (4.12), the closeness of each subproblem solution can be measured by . If subproblems are solved exactly, jf [12]. The -criticality conditions, however, are designed for a \short-step" algorithm in which one truncated-Newton step should satisfy each subproblem for su-ciently small . To ensure the accuracy of the nal solution, nal termination is based on the following two requirements: We have found that reasonably accurate solutions are ensured when " The traditional view in tomographic reconstruction is that a highly accurate solution is unnecessary. This view stems in part from the ill-posedness of the problem and the computational cost of taking a reconstruction to full convergence. From empirical evidence in our studies, the ability to perform certain imaging tasks such as \cold spot detectability" improves with accuracy of the solution. Although the termination criteria we propose above may not appear particularly strict, they are from a tomographic reconstruction perspective. Table Properties aecting computation, memory, and storage costs for two dierent-sized reconstruction problems. Gradient evaluation costs are based on a 2.5M-count study on 10 120-MHz IBM RISC/6000 SP processors. size class n N elements density storage cost of in C in C cost of C gradient thick-slice 376,882 5:36 thin-slice 5. Computational studies. To test our algorithm we have performed a number of reconstructions on data acquired from a small animal scanner, and on data generated by Monte Carlo simulations on the same animal scanner. 5.1. Size of the problem. Our studies involved two dierent-sized problems. Raw coincidence data from the scanner can be binned into either \thick-slice" or \thin-slice" measurement spaces, or both. \Thick-slice" reconstructions, in which minutes for a gradient evaluation using processors (120 MHz) on a 2.5M-count study. For a \thin-slice" reconstruction with on the same data and processors, a gradient evaluation requires 6.75 minutes. These properties are summarized in Table 5.1. The cost of storing the full n N system matrix is prohibitive, even for thick-slice reconstructions. Extensive exploitation of the sparsity and symmetries inherent in the system matrix makes its storage and retrieval possible [24, 25]. The dominant computational operations of the reconstruction problems are the forward- and back-transformation operations that underlie EM iterations, gradient evaluations, Hessian-vector products, and diagonal Hessian calculations. These operations have been implemented in parallel via a data decomposition strategy that partitions the \measurement-space" vectors y and ^ y across the processors. The \image- space" vectors such as x and are replicated over all processors. Our data decomposition is justiable under the observation that N >> n. On a data set with 2.5M counts, at most 47% of the elements of y will be nonzero in the thick-slice case; at most 4% in the thin-slice case. (The thin-slice conguration has over 10 times as many lines of response as the thick-slice.) The dominant computational operations have been implemented in such a way to exploit sparsity in y and further conserve computation [24]. 5.2. Cost metrics. We have devised metrics to measure the cost of an interior point reconstruction. Dene the number of subproblems to be npr, the number of truncated-Newton iterations nit, the number of conjugate gradient subiterations ncg. The cost of one CG iteration (dominated by the Hessian-vector product) is equivalent to the cost of one gradient calculation or EM iteration. One truncated-Newton iteration requires, in addition to the ncg operations, one diagonal Hessian evaluation plus one forward transformation and one backward transformation. The exact cost of these operations varies depending on the size of the problem and number of counts, but we shall approximate the cost of one truncated-Newton iteration to be the equivalent of two gradient calculations beyond the cost of the conjugate gradients. Using this approximation, the total cost of unextrapolated interior-point reconstructions can be measured in units of equivalent number of gradient calculations (or C.A. JOHNSON AND A. SOFER Table Summary of thick-slice primal-dual results and comparison with MAP-EM and LSEM. Extrapolation was not used, and in all cases 2. study f npr nit ncg ngr MAP-EM LSEM A 2,465,770 19 19 110 148 1000 344 G 3,660,344 24 24 127 175 > 1000 724 average ngr 183 Table Summary of thick-slice extrapolated primal-dual results and comparison with MAP-EM and LSEM; in all cases 2. study f npr nit ncg ngr MAP-EM LSEM A 2,465,772 17 17 94 145 960 332 F 3,296,029 G 3,660,384 20 20 100 160 >1000 430 average ngr 156 EM iterations): Extrapolation requires an additional gradient calculation following the extrapolation in order to update the gradient vector. With extrapolation we modify the formula to 5.3. Computational results. We have performed a number of 3-D reconstructions on data acquired from a small animal scanner and data generated by a Monte Carlo simulation of the same small animal scanner. Reconstructions of seven datasets were taken to full convergence, as dened by the termination criteria (4.28) and (4.29) with . The various datasets used in our computational studies represent a fairly diverse sample of the types of scans that might be encountered in practice. The number of counts in the datasets used in these studies ranged from 850K to 5.1M. The number of binding constraints at the optimal solution ranged from approximately 20% to 80%. Our main results are summarized in Tables 5.2 and 5.3 for the non-extrapolated and extrapolated primal-dual cases, respectively. Studies A through D are reconstructions of data acquired from a small animal PET scanner, while studies E through G are reconstructions of Monte Carlo simulated data. These reconstructions were performed in \thick-slice" mode (376,832 variables) with the regularization parameter set at In these tables, the column "MAP-EM" indicates the number of DePierro MAP-EM iterations that were required to achieve the value of f in the same row. The column "LSEM" indicates the number of iterations required for an EM algorithm, where the search direction on the last term of (3.2) is enhanced by a Table Summary of thick-slice logarithmic barrier results and comparison with MAP-EM and LSEM. Extrapolation was used on all data sets, and in all cases study f npr nit ncg ngr MAP-EM LSEM A 2,465,832 5 28 159 218 880 194 average ngr 265 line search. (To avoid excessive computation, the function values were only calculated every iterations, and the nal count was rounded down, to favor this method.) Since the cost of one gradient evaluation is equivalent to the cost of one EM iteration, the numbers in the columns ngr and MAP-EM and LSEM can be compared directly. We nd that the primal-dual method consistently reaches convergence much more rapidly than either MAP-EM or LSEM. Another interesting observation can be made in the comparison between Tables 5.2 and 5.3. Consider the number of EM iterations required to reach f for study C. In Table 5.2, the LSEM algorithm reached iterations. In Table 5.3 on the same data set, the LSEM algorithm reached tions. Thus, the algorithm took 64 iterations to reduce the function value by only 10 units near the solution. MAP-EM did even worse, requiring 180 iterations to reduce the function value by 10. This is in fact a typical example of the slow limit behavior of the EM algorithm. In all studies, the EM method did not achieve the same convergence results obtained by the primal-dual method at termination. The Lagrangian gradient norm and complementary slackness values of the terminated MAP-EM and LSEM iterates were consistently much higher than those of the terminated primal-dual solution. We have also performed these reconstructions using a stabilized logarithmic barrier algorithm based on the method presented in [50] and specialized to the present reconstruction problem. Many of the computational features of our logarithmic barrier implementation are identical to our primal-dual implementation, e.g., truncated Newton, line search, computation of the gradient, Hessian-vector product, etc. For a more detailed discussion, see [24]. The logarithmic barrier results are summarized and compared against MAP-EM in Table 5.4. Termination of the logarithmic barrier was dened by (4.29) and These termination criteria for the logarithmic barrier correspond to roughly the same accuracy as (4.28) and (4.29) do for the primal-dual method. Being a \long-step" method, the logarithmic barrier gives the user less control over the exact stopping point than does the \short-step" primal-dual. All of the logarithmic barrier reconstructions in Table 5.4 used extrapolation. In all logarithmic barrier reconstructions, was reduced by a factor of 10 between subproblems. The eect of extrapolation is illustrated in Figures 5.1 and 5.2. In Figure 5.1, the equivalent number of gradient evaluations (ngr) to reach termination is plot- C.A. JOHNSON AND A. SOFER 100 150 200 250 300 350 PD barrier f-f* ngr Fig. 5.1. \Distance" from optimal solution at termination, as measured by dierence in objective function f f (where f is here dened to be the lowest objective function obtained per study), versus work required to reach termination, as measured by ngr, the equivalent number of gradient evaluations. The studies included are those listed in Table 5.2. PD stands for non-extrapolated primal-dual, PDX for extrapolated primal-dual. ngr PD barrier Fig. 5.2. Average value of at subproblem termination versus average ngr (equivalent number of gradient evaluations) for the seven studies listed in Table 5.2. PD stands for non-extrapolated prmal-dual, PDX for extrapolated primal-dual. ted against objective function \distance" f f , the dierence between the function value of the terminated solution and the lowest function value obtained for that recon- struction. In all seven test cases (those listed in Tables 5.2{5.4), the unextrapolated primal-dual method achieved the lowest objective function value. Thus, f f is zero for all unextrapolated primal-dual (PD) results but greater than zero for the extrapolated primal-dual (PDX) and barrier results. The PDX results are clustered in a region of lower ngr than the PD results. This indicates that extrapolation lowers the computational expense to the solution at a slight deterioration in the nal objective. Compared with the barrier method, either extrapolated or unextrapolated primal-dual produces equivalent or better accuracy with less computation required. In Figure 5.2, the average number of equivalent gradient evaluations at subproblem termination is plotted against the average value of for each subproblem. Both averages (ngr and ) were taken from the same seven test cases of Tables 5.2{5.4. Compared with either unextrapolated primal-dual (PD) or extrapolated primal-dual Study C MAP-EM LSEM PD |f-f*| ngr Fig. 5.3. Improvement in objective function as a function of gradient evaluations, Study F. PDX denotes extrapolated primal-dual, PD denotes unextrapolated primal-dual, both using Study F MAP-EM LSEM PD |f-f*| ngr Fig. 5.4. Improvement in objective function as a function of gradient evaluations, Study C. PDX denotes extrapolated primal-dual, PD denotes unextrapolated primal-dual, both using (PDX), the logarithmic barrier is clearly on a slower trajectory. The PD and PDX trajectories are quite similar until approximately 0:01, at which point the PD curve \swings out", while the PDX curve continues to descent log-linearly. This result conrms that the prediction (extrapolation) step becomes more accurate near the so- lution, resulting in more rapid convergence. However, a comparison of the objective functions indicates that the value of PDX is perhaps one step \ahead of itself," compared with the unextrapolated case. The progress of the reconstruction on a study of a rat skull, Study C, is compared for the various algorithms in Figure 5.3. The measure used is kf f k (plotted on a logarithmic scale). In the initial iterations DePierro Map EM and LSEM progress rapidly and are ahead of the primal dual method. However the interior-point methods rapidly reach the DePierro and LSEM objective values, and hence on, surpass them. In the primal-dual methods depicted the value of is 2. The methods achieve faster initial progress using however the overall computational eort for full convergence with this parameter setting is greater. The progress of the reconstruction in another example, Study F, is compared in Figure 5.4. We have also reconstructed a number of very large-scale \thin-slice" reconstructions involving variables. Table 5.5 summarizes a number of properties of 22 C.A. JOHNSON AND A. SOFER Table Summary of thin-slice extrapolated results, including convergence measures and computational costs to optimal solution. study f kr'k F 1E-5 7,721,001 1.29E-9 2.87E-11 14 14 71 113 average ngr 115 these extrapolated primal-dual reconstructions at the converged solution. A smaller group of datasets (the more visually \interesting" studies) were selected for the thin- slice work, and certain reconstructions were repeated with dierent values of the prior strength . Thin-slice reconstructions seem to require a lower prior strength than the corresponding thick-slice reconstructions. The most visually pleasing results were from reconstructions using which is 1/30 the prior strength that was generally found to be most satisfactory in thick-slice reconstructions. The total amount of work (as measured in ngr) required to reach termination in Table 5.5 is also quite pleasing. The number of variables in a thin-slice reconstruction is approximately 3.7 times the number in thick-slice. The number of nonzero-valued measurements in thin-slice mode is only marginally greater than in thick-slice mode, however, since the number of counts is the same in both cases. These thin-slice reconstructions may thus be better conditioned than their thick-slice counterparts. In closing, we should comment that the tolerance we have used in our tests is stricter than that usually necessary. Indeed, less accurate solutions may still give acceptable images. When the EM method is applied to the (unregularized) ML ob- jective, it is usually terminated after 50 or 100 iterations, and the images produced are often good. Thus EM-ML remains a practical method that can sometimes reach a solution of desirable image quality faster than an interior-point method. The difculty with EM-ML is that its convergence is object-dependent [1]. Convergence in areas of high activity amidst low activity or vice versa is notoriously slow, and a xed termination rule based on (say) 50 or 100 iterations cannot guarantee acceptable image quality. This has been observed in a number of reconstructions, including some of high biomedical interest. In contrast to ML-EM, the primal-dual algorithm has object-independent convergence characteristics. Furthermore it is exible, and can be adapted to solve a problem e-ciently both to the strict tolerance in the studies above by setting a modest rate of decrease for the barrier parameter, say 2, and to a looser tolerance by setting a more aggressive reduction rate such as 6. Conclusion. From the results of the previous Section, it is clear that the primal-dual method can converge signicantly faster than the EM algorithm for regularized ML reconstructions in emission tomography. The results also indicate that the primal-dual method converges faster than the logarithmic barrier method. The use of extrapolation in conjunction with the primal-dual method further reduces the amount of computation required to achieve convergence. Given that the negative regularized ML objective function that we minimize is convex, approximately solving the reduced unsymmetric primal-dual Newton equa- tions is appropriate. Symmetrizing the unsymmetric system, while potentially useful for nonconvex problems, would in this case require solving for 2n variables without avoiding the potential for ill-conditioning. Our stabilization technique avoids the structural ill-conditioning of the condensed primal-dual matrix, and therefore solving the reduced system poses no asymptotic di-culty as the barrier parameter approaches zero. The computational e-ciency and relative simplicity of formation of the reduced system of equations pose such a strong advantage that our choice of primal-dual method almost seems obvious for this problem. Since Newton's method converges quadratically near the solution, for a well-conditioned system in the limit as ! 0, one truncated-Newton step per subproblem should yield an increasingly accurate and well scaled direction to the subproblem solution for k . As is decreased, the subproblem solutions should become \close" to each other for a convex problem [14]. Yet, the example in Table 4.1 illustrates that the direction produced by the early-terminated CG can in fact become less accurate for smaller due to the structured ill-conditioning in M . In practice, we do not require the accuracy of the test example in Table 4.1. Our termination conditions are dened to be near the point on the trajectory where the stabilization approximation becomes accurate enough to guarantee descent. These termination criteria are quite accurate by the standards of the tomography community. Thus, although most reconstruction problems are unlikely to be severely aected by ill-conditioning, the potential for slow convergence near the solution due to ill-conditioning does exist. Our experience has been that stabilization has been an eective safeguard against poor performance for small values of the barrier parameter. 7. Acknowledgments . The study utilized the high-performance computational capabilities of the IBM RISC/6000 SP system at the Division of Computer Research and Technology, National Institutes of Health, Bethesda, MD. We are grateful to Jurgen Seidel of the Department of Nuclear Medicine, National Institutes of Health, for kindly providing us with the small animal data and Monte Carlo simulation data. Our thanks go to two anonymous referees and the associate editor for their careful reading and helpful comments. --R Noise properties of the EM algorithm: I. Conjugate gradient methods for Toeplitz systems A general class of preconditioners for statistical iterative reconstruction of emission computed tomography Preconditioning methods for improved convergence rates in iterative reconstructions A primal-dual algorithm for minimizing a nonconvex function subject to bound and linear equality constraints Image reconstruction and restoration: overview of common estimation structures and problems Maximum likelihood from incomplete data via the EM algorithm On the convergence of an EM-type algorithm for penalized likelihood estimation in emission tomography Numerical stability and e-ciency of penalty algorithms Sequential Unconstrained Minimization Techniques Stability of symmetric ill-conditioned systems arising in interior methods for constrained optimization Matrix Computations On use of the EM algorithm for penalized likelihood estimation A generalized EM algorithm for 3-D Bayesian reconstruction from Poisson data using Gibbs priors Image Reconstruction from Projections: the Fundamentals of Computerized Tomography Methods of conjugate gradients for solving linear systems Accelerated image reconstruction using ordered subsets of projection data A practical interior-point method for convex programming Nonlinear optimization for Volume A system for the 3D reconstruction of retracted-septa PET data using the EM algorithm Evaluation of 3D reconstruction algorithms for a small animal PET camera Implementing and accelerating the EM algorithm for positron emission tomog- raphy PET regularization for envelope guided conjugate gradients Constrained reconstruction by the conjugate gradient method A primal-dual interior point algorithm for linear programming Probability measure estimation using The importance of preconditioners in fast Poisson-based iterative reconstruction algorithms for SPECT Practical tradeo EM Reconstruction Algorithms for Emission and Transmission Tomography A theoretical study of some maximum likelihood algorithms for emission and transmission tomography Convergence of EM image reconstruction algorithms with Gibbs smoothing A maximum a posteriori probability expectation maximization algorithm for image reconstruction in emission tomography The superlinear convergence of a nonlinear primal-dual algorithm Statistical Modeling and Fast Bayesian Reconstruction in Positron Tomog- raphy Fast gradient-based methods for Bayesian reconstruction of transmission and emission PET images Bayesian reconstruction of PET images: methodology and performance analysis Analytic expressions for the eigenvalues and eigenvectors of the Hessian matrices of barrier and penalty functions Preconditioning of truncated-Newton methods Block truncated-Newton methods for parallel optimization Barrier methods for large-scale quadratic programming Maximum Likelihood Reconstruction for Emission Tomography An infeasible interior-point method for linear complementarity problems A statistical model for positron emission tomography Noise properties of Interior methods for constrained optimization Stability of linear equation solvers in interior point methods. --TR
applications of nonlinear programming;parallel computation;tomography;primal-dual methods;estimation;large-scale problems
589052
On the Convergence Theory of Trust-Region-Based Algorithms for Equality-Constrained Optimization.
In a recent paper, Dennis, El-Alem, and Maciel proved global convergence to a stationary point for a general trust-region-based algorithm for equality-constrained optimization. This general algorithm is based on appropriate choices of trust-region subproblems and seems particularly suitable for large problems.This paper shows global convergence to a point satisfying the second-order necessary optimality conditions for the same general trust-region-based algorithm. The results given here can be seen as a generalization of the convergence results for trust-regions methods for unconstrained optimization obtained by Mor and Sorensen. The behavior of the trust radius and the local rate of convergence are analyzed. Some interesting facts concerning the trust-region subproblem for the linearized constraints, the quasi-normal component of the step, and the hard case are presented.It is shown how these results can be applied to a class of discretized optimal control problems.
Introduction . Trust-region algorithms have been proved to be efficient and robust techniques to solve unconstrained optimization problems. An excellent survey in this area is Mor'e [22]. Other classical references for convergence results are Carter [3], Mor'e and Sorensen [23], Powell [26], and Shultz, Schnabel, and Byrd [29]. The standard techniques to handle the trust-region subproblems are the dogleg algorithm (Powell [25]), conjugate gradients (Steihaug [32] and Toint [33]), and Newton-like methods for the computation of locally constrained optimal steps (Gay [15], Mor'e and Sorensen [23], and Sorensen [30]). See also the book of Dennis and Schnabel [9]. Recent new algorithms to compute a locally constrained optimal step (in other words a step that satisfies a fraction of optimal decrease on the trust-region subproblem) that are very promising for large problems have been proposed by Rendl and Wolkowicz [28] and Sorensen [31]. Since the mid eighties a significant effort has been made to address the equality- constrained optimization problem. References are Celis, Dennis, and Tapia [4], Vardi [34] (see also El-Hallabi [14]), Byrd, Schnabel, and Shultz [2], Powell and Yuan [27], and El-Alem [13]. The fundamental questions associated with the application of trust-region algorithms to equality-constrained optimization are the decomposition of the step, the choice of the trust-region subproblems, and the choice of the merit function. During the first stages of the research conducted in this area it was not clear how to answer these questions properly. However, if we examine carefully the most recent Department of Computational and Applied Mathematics, Rice University, Houston, Texas 77005- USA. E-Mail: dennis@rice.edu. Support of this author has been provided by DOE contract DOE-FG03-93ER25178, NSF cooperative agreement CCR-9120008, and AFOSR contract F49620- 9310212. y Departamento de Matem'atica, Universidade de Coimbra, 3000 Coimbra, Portugal. This work was developed while the author was a graduate student at the Department of Computational and Applied Mathematics of Rice University. E-Mail: lvicente@mat.uc.pt. Support of this author has been provided by INVOTAN (NATO scholarship), CCLA (Fulbright scholarship), FLAD, and NSF cooperative agreement CCR-9120008. references (Byrd and Omojokon [24], Dennis, El-Alem, and Maciel [7], El-Alem [12], [13], and Lalee, Nocedal, and Plantenga [21]) we can observe the same decomposition of the step (in its normal, or quasi-normal, and tangential components) and the same trust-region subproblems (the trust-region subproblem for the linearized constraints and the trust-region subproblem for the Lagrangian reduced to the tangent subspace). This is explained in great detail in Section 2 of this paper. As in the unconstrained case, the conditions that each component has to satisfy and the way they are computed might of course differ from algorithm to algorithm. We can see also in these most recent references that the merit function used is either the ' 2 penalty function without constraint term squared (cases of [21], [24]) or the augmented Lagrangian function (in Consider now the equality-constrained optimization (ECO) problem minimize f(x) subject to n. The functions f and c i , are assumed to be at least twice continuously differentiable in the domain of interest. In [7], Dennis, El-Alem, and Maciel have considered a general trust-region-based algorithm for the solution of the ECO problem (1.1). This general algorithm is very much like the algorithm proposed by Byrd and Omojokon [24] 1 . As mentioned before, each step s is decomposed as s n is the quasi-normal component of the step, associated with trust-region subproblem for the linearized constraints and s t is the tangential component, associated with the trust-region subproblem for the Lagrangian reduced to the tangent subspace. Each component of the step is only required to satisfy a fraction of Cauchy decrease on the corresponding trust-region subproblem. Another key feature of this general algorithm is the choice of the augmented Lagrangian as a merit function and the use of the El-Alem's scheme [11] to update the penalty parameter. Under appropriate assumptions, it can be shown that there exists a subsequence of iterates driving to zero the norm of the residual of the constraints and the norm of the gradient of the Lagrangian reduced to the tangent sub-space (see [7][Section 8]). It is important to remark that this global convergence result is obtained under very mild conditions on the components of the step, on the multipliers estimates, and on the Hessian approximations. Thus, the Dennis, El-Alem, and Maciel [7] result is similar to the global result given by Powell [26] for unconstrained optimization. One of the purposes of this paper is to show global convergence to a point satisfying the second-order necessary optimality conditions for this class of algorithms. Our result is similar to the results established by Mor'e and Sorensen [23], [30] for trust-region algorithms for unconstrained optimization. We accomplish this here by imposing a fraction of optimal decrease on the tangential component s t of the step, by using exact second-order information, and by imposing conditions on the quasi- normal component s n and on the Lagrange multipliers. 1 The Thesis [24] was directed by Professor R. H. Byrd. The trust-region algorithm proposed here is usually referred as the Byrd and Omojokon algorithm. In [2], Byrd, Schnabel, and Shultz have proposed a trust-region algorithm for equality-constrained optimization and established global convergence to a point satisfying the second-order necessary optimality conditions. However this algorithm does not belong to the class of trust-region algorithms considered here and their result is obtained with the use of the (exact) normal component and the least-squares multipliers update which we do not require in this paper. Other differences are that they use the ' 1 penalty function as the merit function and the analysis is carried out by using an orthogonal null-space basis. In recent papers, Coleman and Yuan [6] and El-Alem [12] have proposed trust-region algorithms for which they prove global convergence to points satisfying first-order and second-order necessary optimality conditions. Their algorithms use the (exact) normal component, an orthogonal null-space basis, and the least-squares multipliers update. The conditions we need to impose to assure that a limit point of the sequence of iterates satisfies the second-order necessary optimality conditions are k is the quasi-normal component of the step s k , and k is the trust-region radius. In the case where kC(x k )k is small compared with the first condition implies that any increase of the quadratic model of the Lagrangian from x k to x k +s n k is O(ffi 2 To see why this is relevant recall that a fraction of optimal decrease is being imposed on the tangential component s t k yielding a decrease of O(ffi 2 on the quadratic model. The second condition is needed for the same reasons because appears in the definition of predicted decrease. We show that both conditions are satisfied when either (i) the (exact) normal component and the least-squares multipliers are used; or (ii) the most reasonable choices of quasi-normal component and multipliers are made for a class of discretized optimal control problems. The former result is in agreement with the result obtained by El-Alem [12]. Gill, Murray, and Wright [17] and El-Alem [10] considered in their analyses that k). In the latter work this assumption is used to prove local convergence results, and in the former to establish properties of an augmented Lagrangian merit function. We point out that this assumption implies that r x '(x k is since s k is O(ffi k ) and we assume that s n k is O(kC(x k )k). We also prove that if the algorithm converges to a point where the reduced Hessian is positive definite, then the penalty parameter ae k is uniformly bounded and the trust-region radius ffi k is uniformly bounded away from zero, a desired property of a trust-region algorithm. In this case, particular choices of the multipliers and of the components s n and s t lead us to a q-quadratic rate of convergence in x. A detailed treatment of the global convergence theory is given in Vicente [35]. The structure of the trust-region subproblem for the linearized constraints can be exploited to obtain some interesting results. We introduce a quasi-normal component that satisfies a fraction of optimal decrease on the trust-region subproblem for the linearized constraints. We show that the (exact) normal component shares this property. We also prove that if the algorithm is well behaved (for instance if the trust radius is uniformly bounded away from zero), then this subproblem has a natural tendency to fall into the so-called hard case. We review the notation used in this paper. The Lagrangian function associated with the ECO problem (1.1) is defined by '(x; is the Lagrange multiplier vector. The matrix rC(x) is given by , where rc i (x) represents the gradient of the function c i (x). Let be the Hessian matrices of f(x) and c i (x) respectively. We use subscripted indices to represent the evaluation of a function at a particular point of the sequences fx k g and f k g. For instance, f k represents f(x k ) and ' k is the same as The vector and matrix norms used are the ' 2 norms, and I l represents the identity matrix of order l. Finally, 1 (A) denotes the smallest eigenvalue of the symmetric matrix A. The structure of this paper is as follows. In Section 2, we introduce the trust-region subproblems and outline the general trust-region algorithm and the general assumptions. In Section 3, we present the global convergence theory. A class of discretized optimal control problems is introduced in Section 4 as a justification for the general form of our algorithms and theory. In Sections 5 and 6, we analyze respectively the behavior of the trust radius and the local rates of convergence. The trust-region subproblem for the linearized constraints is studied in Section 7. We end this paper with some summary conclusions. 2. Algorithm and general assumptions. The trust-region algorithm analyzed by Dennis, El-Alem, and Maciel [7] for the solution of the ECO problem (1.1), consists of computing, at each iteration k, a step s k decomposed as s the components s n k and s t are required to satisfy given conditions. If the step s k is accepted, the algorithm continues by setting x k+1 to x . If the step is rejected then x 2.1. Decomposition of the step. Suppose that kC k k 6= 0. The component k is called the quasi-normal (or quasi-vertical) component of s k and is required to satisfy a fraction of Cauchy decrease on the trust-region subproblem for the linearized constraints defined by subject to ks n k where r 2 (0; 1) and ffi k is the trust radius. In other words, s n k has to satisfy k is the so-called Cauchy point for this trust-region subproblem, i.e. c n k is the optimal solution of subject to c n 2 spanf\GammarC k C k and therefore The component s n k also is required to satisfy the condition ks n where 1 is a positive constant independent of the iterate k of the algorithm. This condition is saying that close to feasibility the quasi-normal component has to be small. In this paper, we require the quasi-normal component s n k also to satisfy where 2 is a positive constant independent of the iterates. The important consequence of this condition is that if kC k k is small compared with ffi k , then any increase of the quadratic model of the Lagrangian along the quasi-normal component s n k is of O(ffi 2 The two choices of s n k given in Sections 4.1 and 4.2 satisfy conditions (2.1), (2.2), and (2.3). Other choices have been suggested in [7], [20]. The component s t k is the tangential (or horizontal) component, and it must satisfy i.e. it must lie in the null space N (rC T k ) of rC T k . Let W k be an n \Theta (n \Gamma m) matrix whose columns form a basis for N (rC T be a quadratic model of ' at is an approximation to r 2 Since for any s t 2 N (rC T k ), there exists a s t 2 IR n\Gammam such that s consider also which is given by with If kg k k 6= 0, s t k is required to satisfy a fraction of Cauchy decrease for the trust-region subproblem minimize subject to ks n Note that this is not a standard trust-region subproblem because s n k might not be orthogonal to N (rC T might not be the center of the trust region. The steepest-descent direction at associated with in the ' 2 norm is \Gamma g k . If we take into account the scaling matrix W k , then the steepest-descent direction in the kW k \Delta k norm is given by \Gamma(W T k . We consider the steepest-descent direction \Gamma g k for ks n and require k to satisfy where oe t ? 0, and k is the Cauchy point for the ' 2 norm given by with ks n This is equivalent to saying that max is the maximum step length along s n allowed inside the trust region defined by ffi k . It is easy to verify that The results given in this paper hold also if c t k is defined along \Gamma(W T provided the sequence fk(W T are valid also if the coupled trust-region constraint ks n is decoupled as ks t k ffi k . In this latter case the parameter r defining the quasi-normal component s n k can have any positive value. A step k that satisfies this requirement can be computed by using Powell's dogleg algorithm [25] or by the conjugate-gradient algorithm adapted for trust regions by Steihaug [32] and Toint [33] (see also [7], [8], [21]). In order to establish global convergence to a point satisfying the second-order necessary optimality conditions, we need k to satisfy a fraction of optimal decrease on the following trust-region subproblem minimize subject to kW k where In other words, we require k to satisfy the following conditions: s k is the optimal solution of the trust-region subproblem (2.5). This can be accomplished by applying the GQTPAR routine of Mor'e and Sorensen [23] or by using the algorithms recently proposed by Rendl and Wolkowicz [28] and Sorensen [31]. If s t k satisfies a fraction of optimal decrease on the trust-region subproblem (2.5), then ks k k ks n If k is only required to satisfy a fraction of Cauchy decrease, then ks k ks n . We can combine both cases and write ks ks n It is also important to note that the definition of ~ assures that the fraction of optimal decrease (2.6) implies the fraction of Cauchy decrease (2.4) provided 2.2. General trust-region algorithm. We introduce now the merit function and the corresponding actual and predicted decreases. The merit function used is the augmented Lagrangian where ae is the penalty parameter. The actual decrease ared(s k ; ae k ) at the iteration k is given by The predicted decrease (see [7]) is the following: pred(s To update the penalty parameter ae k we use the scheme proposed by El-Alem [11]. The Lagrange multipliers k are required to satisfy where 3 is a positive constant independent of k. This condition is not necessary for global convergence to a stationary point. The general trust-region algorithm is given below. Algorithm 2.1 (General trust-region algorithm). and r such that ae ? 0, and r 2 (0; 1). do 2.1 If kC is given in (2.10), stop the algorithm and use x k as a solution for the ECO problem (1.1). 2.2 Set s n satisfying (2.1), (2.2), (2.3), and ks n If kW T satisfying (2.6). k . 2.3 Compute k+1 satisfying (2.8). 2.4 Compute pred(s If pred(s then set ae Otherwise set ae: 2.5 If ared(s k ;ae k ) pred(s k ;ae k ) ks k k and reject s k . Otherwise accept s k and choose ffi k+1 such that 2.6 If s k was rejected set x . Otherwise set x It is important to understand that the role of ffi min is just to reset ffi k after a step s k has been accepted. During the course of finding such a step the trust radius can be decreased below ffi min . To our knowledge Zhang, Kim, and Lasdon [37] were the first to suggest this modification. We remark that the rules to update the trust radius in the previous algorithm can be much more complicated but those given suffice to prove convergence results and to understand the trust-region mechanism. As a direct consequence of the way the penalty parameter is updated, we have the following result. Lemma 2.1. The sequence fae k g satisfies ae k ae pred(s In order to establish global convergence results, we use the general assumptions given in [7]. These are Assumptions A.1-A.4. However for global convergence to a point that satisfies the second-order necessary optimality conditions, we also need Assumption A.5. We assume that for all iterations k, x k and x are in \Omega\Gamma where \Omega is an open subset of IR n . General assumptions A.1 The functions f , c i , are twice continuously differentiable in \Omega\Gamma A.2 The gradient matrix rC(x) has full column rank for all x 2 \Omega\Gamma A.3 The functions f , rf , r are bounded in \Omega\Gamma The matrix (rC(x) T rC(x)) \Gamma1 is uniformly bounded in \Omega\Gamma A.4 The sequences fW k g, fH k g, and f k g are bounded. A.5 The Hessian approximation H k is exact, i.e. H are Lipschitz continuous in \Omega\Gamma Assumptions A.3 and A.4 are equivalent to the existence of positive constants 9 such that jf(x)j 0 , krf(x)k 1 , kr 2 f(x)k 2 , kC(x)k 3 , 2.3. Predicted decrease along the tangential component. Consider again the trust-region subproblem (2.5). We can use the general assumptions and the structure of this subproblem to obtain a lower bound on the predicted decrease along the tangential component of the step. It follows from the Karush-Kuhn-Tucker conditions that there exists a fl k 0 such that positive semi-definite, s ~ s 0: (It turns out that these conditions are also sufficient for s k to solve the trust-region subproblem (2.5), see Gay [15] and Sorensen [30].) As a consequence of this we can s where Hence we have 3. Global convergence. Dennis, El-Alem, and Maciel [7] have proved under Assumptions A.1-A.4 and conditions (2.1), (2.2), and (2.4) that lim inf 0: In this section we assume that s t k satisfies the fraction of optimal decrease (2.6) on the trust-region subproblem (2.5), as well as conditions (2.3), (2.8), and A.5 on respectively, and show that (3.1) can be extended to 0: The proof of (3.2) although simpler has the same structure as the proof given in [7]. We prove the result by contradiction, under the supposition that for all k. We start by analyzing the fraction of Cauchy and optimal decrease conditions. Lemma 3.1. Let the general assumptions hold. Then and and, moreover, since s t k satisfies a fraction of optimal decrease for the trust-region subproblem (2.5), are positive constants independent of the iterate k. Proof. The conditions (3.4) and (3.5) are an application of Powell's result (see [26, Theorem 4], [22, Lemma 4.8]) followed by the general assumptions. The condition (3.6) is a restatement of (2.11) with The following inequality is needed in the forthcoming lemmas. Lemma 3.2. If the general assumptions hold, then positive constant independent of k. Proof. The term q k k ) can be bounded using (2.2), (2.3), and Assumption A.4, in the following way: ks n On the other hand, it follows from (2.8) and krC T that If we combine these two bounds we get (3.7) with The following technical lemma gives us upper bounds on the difference between the actual decrease and the predicted decrease. The proof follows similar arguments as the proof of Lemma 6.3 in [11]. Lemma 3.3. Let the general assumptions hold. There exist positive constants independent of k, such that ks k k 3 ks k ae k 3 ks k k 3 ks k k 2 and ks k ae k 6 ks k k 3 ks k k 2 Proof. If we add and subtract '(x for some 1 1). Again using the Taylor expansion we can write 1). Now we expand c i This and the general assumptions give us the estimate (3.8) for some positive constants The inequality (3.9) follows from (3.8) and ae k 1. The following three lemmas bound the predicted decrease. They correspond respectively to Lemmas 7.6, 7.7, and 7.8 given in [7]. Lemma 3.4. Let the general assumptions hold. Then the predicted decrease in the merit function satisfies pred(s (3. and also pred(s for any ae ? 0. Proof. The two conditions (3.10) and (3.11) follow from a direct application of (3.7) and from the two different lower bounds (3.5) and (3.6) on q k (s n Lemma 3.5. Let the general assumptions hold, and assume that kW T ff min ae ffl tol min ae 7 ffl tol oe 9 ffl tol oe then the predicted decrease in the merit function satisfies either pred(s or pred(s for any ae ? 0. Proof. From kW T and the first bound on ff given by (3.12), we get Thus either kW T us first assume that kW T ffl tol . Then it follows from the second bound on ff given by (3.12) that Using this, (3.10), and the third bound on ff given by (3.12), we obtain pred(s Now suppose that To establish (3.14), we combine (3.11) and the last bound on ff given by (3.12) and get pred(s We can set ae to ae k\Gamma1 in Lemma 3.5 and conclude that, if kW T ffl tol and kC k k ffffi k , then the penalty parameter at the current iterate does not need to be increased. See Step 2.4 of Algorithm 2.1. The proof of the next lemma follows the argument given in the proof of Lemma 3.5 to show that either kg k k ? 1ffl tol or fl k ? 1ffl tol holds. Lemma 3.6. Let the general assumptions hold, and assume that kW T (3.12), then there exists a constant pred(s Proof. By Lemma 3.5 we know that either (3.13) or (3.14) holds. Now we set . In the first case we use kg pred(s In the second case we use pred(s Hence (3.15) holds with ae 6 ffl tol min ae 7 ffl tol oe 9 ffl toloe Next we prove under the supposition (3.3), that the penalty parameter ae k is bounded. Lemma 3.7. Let the general assumptions hold. If kW T for all k, then ae k ae ; where ae does not depend on k, and thus fae k g and fL k g are bounded sequences. Proof. If ae k is increased at iteration k, then it is updated according to the rule ae: We can write ae ki By applying (3.4) to the left hand side and (3.5) and (3.7) to the right hand side, we obtain aei If ae k is increased at iteration k, then from Lemma 3.5 we certainly know that kC k k ? Now we use this fact to establish that We have proved that fae k g is bounded. From this and the general assumptions we conclude that fL k g is also bounded. We can prove also under the supposition (3.3), that the trust radius is bounded away from zero. Lemma 3.8. Let the general assumptions hold. If kW T for all k, then where does not depend on k. Proof. If s k\Gamma1 was an acceptable step, then ffi k ffi min . If not then ks and we consider the cases kC (3.12). In both cases we use the fact ared(s pred(s Case I. kC From Lemma 3.6, inequality (3.15) holds for Thus we can use ks ared(s pred(s )ks Thus ks Case II. kC . In this case from (2.9) and (3.4) with pred(s where rg. Again we use ae and this time the last two lower bounds on pred(s pred(s ks ae ks ae ks Hence ks The result follows by setting ffi g. The next result is needed also for the forthcoming Theorem 3.1. Lemma 3.9. Let the general assumptions hold. If kW T for all k, then an acceptable step is always found in finitely many trial steps. Proof. Let us prove the assertion by contradiction. Assume that for a given k, k. This means that lim k!+1 all steps are rejected after iteration k. See Steps 2.5 and 2.6 of Algorithm 2.1. We can consider the cases and appeal to arguments similar to those used in Lemma 3.8 to conclude that in any case pred(s where 15 is a positive constant independent of the iterates. Since we are assuming that lim k!+1 ared(s k ;ae k ) pred(s k ;ae k ) 1 and this contradicts the rules that update the trust radius. See Step 2.5 of Algorithm 2.1. Now we finally can state our first asymptotic result. Theorem 3.1. Under the general assumptions, the sequence of iterates fx k g generated by the Algorithm 2.1 satisfies lim inf 0: Proof. Suppose that there exists an ffl tol ? 0 such that kW T for all k. At each iteration k either kC k k ffffi k or kC k k ? ffffi k , where ff satisfies (3.12). In the first case we appeal to Lemmas 3.6 and 3.8 and obtain pred(s we have from ae k 1, (2.9), (3.4), and Lemma 3.8, that pred(s Hence there exists a positive constant 16 not depending on k such that pred(s k From Lemma 3.9, we can ignore the rejected steps and work only with successful iterates. So, without loss of generality, we have Now, if we let k go to infinity, this contradicts the boundedness of fL k g. From this result we can state our global convergence result: existence of a limit point of the sequence of iterates generated by the algorithm satisfying the second-order necessary optimality conditions. This result generalizes those obtained for unconstrained optimization by Sorensen [30] and Mor'e and Sorensen [23]. Theorem 3.2. Let the general assumptions hold. Assume that W (x) and (x) are continuous functions and If fx k g is a bounded sequence generated by Algorithm 2.1, then there exists a limit point x such that positive semi-definite on N (rC(x ) T ). Moreover, if (x ) is such that r x '(x ; (x satisfies the second-order necessary optimality conditions. Proof. Let fk i g be the index subsequence considered in (3.16). Since fx k i g is bounded, it has a subsequence fx k j g that converges to a point x and for which lim 0: Now from this and the continuity of C(x), we get C(x we use the continuity of W (x) and rf(x) to obtain Since 1 (\Delta) is a continuous function, we can use (2.10), lim j!+1 the continuity of W (x), (x), and of the second derivatives of f(x) and c i (x), to obtain 0: This shows that r 2 positive semi-definite on N (rC(x ) T ). The continuity of an orthogonal null space basis has been discussed in [1], [5], [16]. A class of nonorthogonal null space basis is described in Section 4.1. The equation r x '(x ; (x consistent updates of the Lagrange multipliers like the least-squares update (4.7) or the adjoint update (4.3). 4. Examples. 4.1. A class of discretized optimal control problems. We now introduce an important class of ECO problems where we can find convenient matrices W k , quasi- normal components s n k , and multipliers k satisfying all the requirements needed for our analysis. The numerical solution of many discretized optimal control problems involves solving the ECO problem subject to C(y; y (see [8], [19], [20]). The variables in y are the state variables and the variables in u are the control variables. Other applications include parameter identification, inverse, and flow problems and design optimization. In many situations there are bounds on the control variables, but this is not considered here. Another interesting aspect of these problems is that rC(x) T can be partitioned as where C y (x) is a square matrix of order m. In this class of problems the following assumption traditionally is made: The partial Jacobian C y (x) is nonsingular and its inverse is uniformly bounded in \Omega\Gamma As a consequence of this, the columns of \GammaC I n\Gammam form a basis for the null space of rC(x) T . The usual choice for k in these problems is the so-called adjoint multipliers It follows directly from the continuity of rC(x) and the uniformly boundedness of continuously with x. Furthermore is a continuous function of x with bounded derivatives. Using the structure of the problem we can define the quasi-normal component s n (see references [8], [19], [20]) as where kCy As we will see in Section 7, the quasi-normal component (4.4) satisfies a fraction of optimal decrease and hence a fraction of Cauchy decrease on the trust-region subproblem for the linearized constraints. Other choices for quasi-normal components are given in [20]. All these quasi- normal components are of the form Lemma 4.1. If s n verifies (4.5) and k is given by (4.3), then conditions (2.3) and (2.8) are satisfied. Proof. From (4.3) and (4.5) we can see that /r and condition (2.3) is trivially satisfied. Condition (2.8) follows from the existence of bounded derivatives for 4.2. The normal component and the least-squares multipliers. Consider again the general ECO problem (1.1). If the component s n k of the step s k is orthogonal to the null space of rC T k , then it is a multiple of rC k (rC T . If we also require that s n lies inside the trust region of radius rffi k , then it is given by . If the quasi-normal component s n k of the step is given by (4.6), then it is called normal. As we will see in the Section 7, the normal component (4.6) satisfies a fraction of optimal decrease and hence a fraction of Cauchy decrease on the trust-region subproblem for the linearized constraints. Lemma 4.2. The quasi-normal component (4.6) and the least-squares update satisfy conditions (2.3) and (2.8). Proof. It can be easily confirmed that r x ' T The condition (2.8) holds since bounded derivatives in \Omega\Gamma 5. The behavior of the trust radius. In Sections 5 and 6 we no longer need to consider that the tangential component s t k satisfies a fraction of optimal decrease on the trust-region subproblem (2.5). It suffices to assume the fraction of Cauchy decrease condition (2.4). We assume that the component s n k satisfies conditions (2.1) and (2.2). We need to strengthen conditions (2.3) and (2.8) in the following way: ks k k; ks k k; ks n ks k k; 3 , and 0 4 are positive constants independent of the iterates. The choices of s n k and k suggested in Section 4 satisfy these requirements as well. See Lemmas 4.1 and 4.2 for the first two conditions. It is obvious that the normal component (4.6) satisfy (5.3). The quasi-normal component (4.4) also satisfies (5.3) (see [35][Lemma 5.6.1]). The next theorems show that if lim k!+1 x positive definite on N (rC(x ) T ), then the penalty parameter ae k is uniformly bounded and the trust radius ffi k is uniformly bounded away from zero. Theorem 5.1. Let the general assumptions hold and W (x) and (x) be continu- ous. If fx k g converges to x and r 2 positive definite on N (rC(x then fae k g is a bounded sequence. Proof. First since r 2 positive definite on N (rC(x are continuous functions of x, there exists a neighborhood N (x ) of x and a fl ? 0 such that for any x in N (x ), xx '(x; (x))W (x) fl: Since ks t Thus for all k such that x k 2 N (x ) we flks t ks t and this implies ks t Now by using (3.5) and (5.4), we have for all k such that x k 2 N (x ), that 17 ks t where 1+r g. Now let kC k k ff 0 ks k k where the positive constant ff 0 is defined later. Using similar arguments as in Lemma 3.2, it follows from (2.2), (5.1), (5.2), kC k k ff 0 ks k k, and Assumption A.4 that ks k k; 3 . From (2.2) and kC k k ff 0 ks k k we also get ks ks n ks t 2ks n ks which together with (5.5) and (5.6) implies pred(s ks ks ks k k for all ae ? 0. We now need to impose the following condition on ff Now we set ae = ae k\Gamma1 in (5.7) and conclude that the penalty parameter does not need to be increased if kC k k ff 0 ks k k (see Step 2.4 of Algorithm 2.1). Hence, if ae k is increased then kC ks k k holds, and by using (5.1)-(5.3) we obtain: ks k k; with 00 3 . Recall from the proof of Lemma 3.7 that if ae k is increased then ae r ks k k oe ae 4 )ks k k kC k k; which in turn implies ae r ae k 00 ae 4 This completes the proof of the Theorem. Theorem 5.2. Let the general assumptions hold and W (x) and (x) be continu- ous. If fx k g converges to x and r 2 positive definite on N (rC(x then ffi k is uniformly bounded away from zero and eventually all iterations are successful Proof. The proof of the theorem is based on the boundedness of fae k g. We consider the cases kC ks k k and kC k k ff 0 ks k k, where ff 0 satisfies (5.8). ks k k, then from (2.7), (2.9), and (3.4), we find that pred(s ks where g. In this case it follows from (3.9), (5.10), and ae k 1 that pred(s ks Now, suppose that kC k k ff 0 ks k k. From (5.7) with pred(s ks k Now we use (3.9) and ae k ae to get pred(s ks It follows from Theorem 8.4 in [7] that lim inf 0: From this result, the continuity of C(x), and the convergence of fx k g we obtain Finally from (5.11), (5.12), and the limits lim k!+1 x finally get lim pred(s which by the rules for updating the trust radius in Step 2.5 of Algorithm 2.1 shows that ffi k is uniformly bounded away from zero. 6. Local rate of convergence. In order to obtain q-quadratic local rates of convergence, we require the reduced tangential component s t k to satisfy (2.4) and the following if H k is positive definite and k 6.1. Discretized optimal control formulation. Consider again problem (4.1) and assume that this problem has the structure described in Section 4.1. We can now use Theorem 5.2 to obtain a local rate of convergence. Theorem 6.1. Suppose that the ECO problem is of the form (4.1). Let the general assumptions and assumption (4.2) hold and assume that fx k g converges to x . In addition to this, let s t k , and k be given by (6.1), (4.4) and (4.3). If r 2 xx '(x ; ) is positive definite on N (rC(x then x k converges q-quadratically to x . Proof. It can be shown by appealing to Theorem 8.4 in [7] that r x '(x ; It follows from Theorem 5.2 that ffi k is uniformly bounded away from zero. Thus there exists a positive integer k such that for all k k, \GammaC y . Now the rate of convergence follows from [19]. 6.2. Normal component and least-squares multipliers. Consider the general ECO problem (1.1) again and suppose that the quasi-normal component is the normal component (4.6) and k is given by (4.7). We can now use Theorem 5.2 to obtain the desired local rate of convergence. It is assumed that the orthogonal null-space basis is a continuous function of x. Theorem 6.2. Let the general assumptions hold and assume that fx k g converges to x . In addition to this, let k , and k be given by (6.1), (4.6), and (4.7). If r 2 xx '(x ; ) is positive definite on N (rC(x then x k converges q-quadratically to x . Proof. It can be shown by appealing to Theorem 8.4 in [7] that r x '(x ; It follows from Theorem 5.2 that ffi k is uniformly bounded away from zero. Thus there exists a positive integer k such that for all k k, . The q-quadratic rate of convergence follows from [18], [36]. 7. The trust-region subproblem for the linearized constraints. In this section we investigate a few aspects of the trust-region subproblem for the linearized constraints subject to ks n k First we prove that the normal component (4.6) and the quasi-normal component always give a fraction of optimal decrease on this trust-region subproblem. Theorem 7.1. Let the general assumptions hold. Then: (i) The normal component (4.6) satisfies a fraction of optimal decrease on the trust-region subproblem for the linearized constraints, i.e. there exists a positive constant fi n 1 such that where s k is the optimal solution of (7.1). (ii) In addition, assume assumption (4.2). The quasi-normal component (4.4) satisfies the fraction of optimal decrease (7.2). Proof. (i) If krC k (rC T solves (7.1), and the result holds for any positive value of fi n 1 in (0; 1]. If this is not the case, then since krC k (rC T We also have ks ks ks since krC k (rC T ks k. Combining this last inequality with (7.3) we get and this completes the proof of (i). (ii) If kC y solves (7.1), and (7.2) holds for any positive value of fi n 1 . If this is not the case, we have is the uniform bound on kC y Now the rest of the proof follows as in (i). As a consequence of this theorem, we have immediately that the normal component (4.6) and the quasi-normal component (4.4) give a fraction of Cauchy decrease on the trust-region subproblem for the linearized constraints. To compute a step s n k that gives a fraction of optimal decrease on the trust-region subproblem for the linearized constraints we can also use the techniques proposed in [23], [28], [31]. In the next theorem we show that the trust-region subproblem (7.1), due to its particular structure, tends to fall in the hard case in the latest stages of the algorithm. This result is relevant in our opinion since the algorithms proposed in [23], [28], [31] deal with the hard case. The trust-region subproblem (7.1) can be rewritten as subject to ks n k The matrix rC k rC T k is always positive semi-definite and, under the general assump- tions, has rank m. Let E k (0) denote the eigenspace associated with the eigenvalue 0g. The hard case is defined by the two following conditions: (a) (v k (b) k(rC k rC T Theorem 7.2. Under the general assumptions, if lim k!+1 exists a k h such that, for all k k h , the trust-region subproblem (7.5) falls in the hard case as defined above by (a) and (b). Proof. First we show that (a) holds at every iteration of the algorithm. If v k 2 Multiplying both sides by (rC T k gives us Thus (v k Now we prove that there exists a k h such that (b) holds for every k k h . Since is a monotone strictly decreasing function of lim is equivalent to g k () ! rffi k , for all ? 0. Also, from the singular value decomposition of rC k , we obtain lim Hence holds for all ? 0 if and only if krC k (rC T Now since lim k!+1 there exists a k h such that kC k , and this completes the proof of the theorem. We complete this section with the following corollary. Corollary 7.1. Under the general assumptions, if lim k!+1 kC k and the trust radius is uniformly bounded away from zero, then there exists a k h such that, for all k k h , the trust-region subproblem (7.5) falls in the hard case as defined above by (a) and (b). Proof. If lim k!+1 kC k and the trust radius is uniformly bounded away from zero then lim k!+1 Theorem 7.2 can be applied. 8. Concluding remarks. In Theorems 3.1 and 3.2 we have established global convergence to a point satisfying the second-order necessary optimality conditions for the general trust-region-based algorithm considered in this paper. In Theorem 5.2 we have proved that the trust radius is, under sufficient second-order optimality conditions, bounded away from zero. With the help of this result we analyzed local rates of convergence for different choices of steps and multipliers. We believe that these results complement the theory developed by Dennis, El-Alem, and Maciel in [7] that proves global convergence to a stationary point. We have also given a detailed analysis of the trust-region subproblem for the linearized constraints. Acknowledgments . We thank Mahmoud El-Alem with whom we had many discussions about the contents of this paper. We are also grateful to our referees for their careful and insightful reading of this paper. --R Continuity of the null space basis and constrained optimiza- tion A trust region algorithm for nonlinearly constrained optimization On the global convergence of trust region algorithms using inexact gradient information A trust region strategy for nonlinear equality constrained optimization A note on the computation of an orthonormal basis for the null space of a matrix A new trust region algorithm for equality constrained optimiza- tion A global convergence theory for general trust-region-based algorithms for equality constrained optimization Numerical Methods for Unconstrained Optimization and Nonlinear Equations A Global Gonvergence Theory for a Class of Trust Region Algorithms for Constrained Optimization A global convergence theory for arbitrary norm trust-region algorithms for equality constrained optimization Computing optimal locally constrained steps Properties of a representation of a basis for the null space Some theoretical properties of an augmented Lagrangian merit function Newton's method for constrained optimization Projected sequential quadratic programming methods Analysis of inexact trust-region interior-point SQP algorithms On the implementation of an algorithm for large-scale equality constrained optimization Trust Region Algorithms for Optimization with Nonlinear Equality and Inequality Constraints A new algorithm for unconstrained optimization A trust region algorithm for equality constrained optimization A semidefinite framework for trust region subproblems with applications to large scale minimization A family of trust-region-based algorithms for unconstrained minimization with strong global convergence properties Newton's method with a model trust region modification The conjugate gradient method and trust regions in large scale optimization Towards an efficient sparsity exploiting Newton method for minimization A trust region algorithm for equality constrained minimization: convergence properties and implementation Numerical Methods for Nonlinearly Constrained Optimization An improved successive linear programming algorithm --TR --CTR Zhensheng Yu , Changyu Wang , Jiguo Yu, Combining trust region and linesearch algorithm for equality constrained optimization, Journal of Computational and Applied Mathematics, v.14 n.1-2, p.123-136, 1 January 1986 Detong Zhu, Nonmonotonic back-tracking trust region interior point algorithm for linear constrained optimization, Journal of Computational and Applied Mathematics, v.155 n.2, p.285-305, 15 June
equality-constrained optimization;second-order necessary optimality conditions;trust regions;local rate of convergence;hard case;SQP methods
589060
Algorithms for Constrained and Weighted Nonlinear Least Squares.
A hybrid algorithm consisting of a Gauss--Newton method and a second-order method for solving constrained and weighted nonlinear least squares problems is developed, analyzed, and tested. One of the advantages of the algorithm is that arbitrarily large weights can be handled and that the weights in the merit function do not get unnecessarily large when the iterates diverge from a saddle point. The local convergence properties for the Gauss--Newton method are thoroughly analyzed and simple ways of estimating and calculating the local convergence rate for the Gauss--Newton method are given. Under the assumption that the constrained and weighted linear least squares subproblems attained in the Gauss--Newton method are not too ill conditioned, global convergence towards a first-order KKT point is proved.
Introduction . Assume that f : R n continuously differentiable function and that diagonal matrix with weights We will discuss the Gauss-Newton method and a second order method for solving the problem min x2R denotes the 2-norm. For simplicity, and without loss of generality, we assume that the weights are normalized and sorted such that The normalization is easily done by first sorting out the zero weights, reducing the problem and then dividing the remaining nonzero weights with the smallest positive weight. To our knowledge all existing algorithms for solving (1.1) are based on the un-weighted problem min Assume that the ordinary Gauss-Newton method is used to solve (1.3). The search direction, p, is then got by solving min p2 (1. rg. Note that (1.4) is solved as an unweighted problem and thus the condition of this problem is determined by kKk kK y k where K y is the pseudo inverse of K. If we on the other hand linearize (1.1), without explicitly multiplying with the weights, we solve the weighted linear least squares problem to obtain the search direction p. The condition for the problem (1.5) is mainly determined by kBk kJk where . For a more detailed discussion on condition numbers for (1.5) see [11]. The problem (1.4) may be very ill conditioned (regarded as an unweighted linear least squares problem) despite the fact that (1.5) is well conditioned (regarded as a weighted linear least squares problem). Obviously it is very important to look at (1.1) as the class of weighted nonlinear least squares problem. Another important advantage of using (1.1) instead of (1.3) is that the former defines a more general problem class than the latter. This is evident if we allow the weights to be infinitely large. To be more precise, we define the vector - the equations weights correspond to zero elements in M . Note that is the Lagrange multiplier corresponding to the ith constraint and consequently - i is not defined by (1.6). We will return to the proper way of calculating these Lagrange multipliers. Problem (1.1) is rewritten, using (1.6), as Hence, by allowing infinite weights, our original problem formulation (1.1) defines the class of weighted nonlinear least squares problems with nonlinear equality constraints. To be even more specific we assume that we have p infinite weights such that Problem (1.7) can now be stated as \Theta f T An equivalent formulation of problem (1.8) using - is min 2 . Of course, we could have started by defining our problem as the one in (1.9) instead of (1.5) (without the need of (1.7) and (1.8)) but then the notations would get unnecessarily complicated. In the next section we describe the Gauss-Newton method for solving (1.1). The local convergence properties of the Gauss-Newton method is analyzed in Section 3 and in Section 4 we show that, under certain assumptions on nondegeneracy, global convergence is achieved. If the Gauss-Newton method is too slow or does not converge and second derivatives are available at a reasonable cost then the Newton method may be used to solve (1.1). However, when there are large and possibly infinite weights a pure Newton method based on forming the Hessian of g(x) may not work or, with infinite weights, is not even defined. The natural approach is then to use the Perturbation method [9] that we will call the generalized Newton-Raphson method (the gNR method). In Section 5 we construct and analyze an algorithm for solving (1.1) based on the gNR method. Computational experiments are presented in Section 6 and finally we discuss our results and give hints of possible future work. 2. The Gauss-Newton Method Using the System Equations. In the Gauss-Newton method, the nonlinear least squares problem (1.1) is linearized around the current iteration point, x k , and the search direction, p k , is computed as the solution to The next iterate is x the steplength. In the presence of large weights, possibly infinite, it is adequate to reformulate (2.1) as \Gammaf# where we for simplicity have dropped the iteration index k. There are several names to the linear system of equations in (2.2) such as the equilibrium equations, the system equations or the augmented system equations. We call (2.2) the system equations and the matrix in (2.2) is called the system matrix. A less obvious reason for using (2.2) is that the elements in - corresponding to infinite weights are approximations to the Lagrange multipliers and that - can be used in a second order method as described in Section 5. The following lemma gives the relevant conditions for the system matrix to be nonsingular. Lemma 2.1. The system matrix in (2.2) is nonsingular if and only if the rows in J that correspond to infinite weights are linearly independent and J has full column rank. There exist several stable algorithms that solve (2.2), see e.g. [5] for further references. We have chosen to use the modified QR decomposition, see [5], and the reasons are the following. The modified QR decomposition is simple and easy to compute and it is identical to the ordinary QR decomposition when the weights are equal. The modified QR decomposition is also easily reused in the second order gNR method, see Section 5. The modified QR decomposition of J 2 R m\Thetan is defined as R# n\Thetan is an upper triangular matrix and \Pi is a permutation matrix. The decomposition in (2.3), with Q and R nonsingular, exists if and only if the system matrix in (2.2) is nonsingular (see Lemma 2.1). The system equations are solved with the modified QR decomposition in the following way. Using the decomposition (2.3) in (2.2) we get6 6 4 R# If we make the partition Mm\Gamman ) the solution to (2.4) is m\Gamman 3. The local rate of convergence for the Gauss-Newton method. In this section we will describe the local convergence properties of the Gauss-Newton method described in the previous section. Our analysis depends much upon the perturbation analysis of the constrained and weighted linear least squares problem done in [11, 12]. After having defined the inverse of the system matrix, using the same notation as in [11], we state and prove two important theorems on the local convergence rate for projected residual). In fact, the local convergence properties of these two quantities are, as we shall see, very similar. Finally we show that J k p k and the local convergence rate for x x and J k p k are independent of the parametrization in R n . Assuming that b x is a solution of (1.1), we define b and the corresponding notation for other quantities evaluated at b x. A necessary condition for our algorithm to converge without regularization is that the system matrix in (2.2) has full rank and it is convenient to make the following definition. Definition 3.1. If the system matrix in (2.2) is nonsingular at x we say that x is a nondegenerate point. At a nondegenerate point the inverse of the system matrix in (2.2) is given by is a generalized inverse of J , see [11]. From (2.5) we immediately get The following theorem describes the local behaviour of x Theorem 3.1. Assume that fp k g are generated by solving (2.2) and that all points x are nondegenerate. If b x is the solution of (1.1) and b - is the vector - from (2.2) at b x then R 1f 00 Proof. From x f) Using the Taylor expansion the first term in (3.4) can be expressed as To express the second term, \GammaB k f , in (3.4) we use the perturbation identity (2.2) p. in [11] which says that Using the identity Z 1[f 00 equation (3.6) becomes The equations (3.5) and (3.7) inserted into (3.4) gives the theorem. The Gauss-Newton method can be written as and with b From Theorem 3.1 we conclude that and from [8] we get the following theorem. Theorem 3.2. Define and - i as the eigenvalues of H x . Then lim sup xk xk It is easy to get an estimation of the local convergence rate if we use the matrix defined by (3.2), because then b R R \GammaT \Pi T . A useful quantity for estimating how close x k is to the solution, b x, is the projected residual is the oblique projection of f k onto R(J k ). The following theorem shows that J k p k locally has the same convergence behaviour as Theorem 3.3. Assume that fp k g are generated by solving (2.2) and that all points x are nondegenerate. If -(x k ) is the vector - from (2.2) at x k then R 1f 00 Proof. Denote the projection J k B k by P k . Then we have Using the Taylor expansion multiplying with P k+1 we obtain I , the equality B k s holds, and we can identify the last term in equation (3.14) as From (3.1) we get and hence From the perturbation identity (2.1) p. 16 in [11] we get Using (3.18) and the fact that Y k+1 ffiJ k B k the equation (3.17) becomes where the last equality follows from (3.16). The identities - \Gammap k together with a Taylor expansion of (ffiJ k ) T give The theorem follows by inserting (3.15) and (3.19) into (3.14). The matrix corresponding to H x for the projected residual, s k , is B: and it is easy to show that H x and H s has the same nonzero eigenvalues. Hence, we have from Theorem 3.3 the following corollary. Corollary 3.1. Define B k from the inverse of the system matrix in (3.1). If lim sup ks are the eigenvalues of the matrix H x defined in (3.11). The relation (3.12) can also be used to determine when ks k+1 k=ks k k reflects the linear convergence rate and if a second order method should be used. If the convergence of the Gauss-Newton method is slow we use a higher order method if2 see also Algorithm 6.1. Several of the above quantities are invariant under a change of parametrization and as an example we have the following theorem. Theorem 3.4. The matrix B: is independent of the parametrization in R n . Proof. Assume that f(x(')). We want to show that y B y is the generalized inverse of ry( b '). Now, consider the Taylor expansion where \Delta' T x 00 \Delta' T x 00 By comparing the Taylor expansion (3.21) above with the Taylor expansion using that J T that From (3.23) we finally get y which proves the theorem. A consequence of Theorem 3.4 is that the local convergence for is the same as for x The main argument for choosing Jp as a measure of the closeness to the solution is the following theorem which is a direct consequence of (3.23). Theorem 3.5. The projection of f on R(J), is independent of the parametrization in R n . 4. Global convergence. In this section we assume that x k , where k is the iteration index, is nondegenerate and that p k is the solution of (2.2) at x k . If nothing else is stated we assume that all limits denoted by ! are when k ! 1, and that all sums with no explicitly stated upper or lower limit are from one to infinity. 4.1. The merit function. As a merit function we have chosen The goal is to find a matrix D k of merit weights and a step length ff k , at each iteration, such that global convergence towards a first order Kuhn-Tucker point can be proved. To compute D k we will use the approximation of D). For a fixed matrix D, we define Obviously a sufficient condition on p k to be a descent direction to \Phi(x; D) at x k is that OE 0 We realize that we can determine a good matrix, \Upsilon(x k ), of merit weights by solving min k\Upsilonk s:t: where ffi is a small positive constant and - i is a lower limit for the weights determined by some previously computed weights, see below. There is always a solution to (4.2) because lim argf min ff Note that keeping the weights not too large is important in practice but for the global convergence it is only the constraints in (4.2) that must be satsified. We will now describe the algorithm for computing the merit weights D k , using \Upsilon(x k ), such that does not become unnecessarily large. We first describe a method for solving (4.2) and then an algorithm for computing the actual merit weights D k . When solving (4.2) we have chosen to use the max-norm since this gives a simple algorithm. The problem (4.2) can be rewritten as min kuk1 where u is the diagonal in \Upsilon(x k ), ! is the diagonal in W , and y with that when Jp is given the problem (4.3) consists of only vectors and no matrices. The first step in our algorithm is to reduce (4.3) such that u We then get a new problem y are the corresponding parts of u; -; ! and y left after the reduction ae we are ready with the solution - -. Otherwise we choose - y, where e is a vector of ones and thus attain equality in the constraints. If - respectively. Again we can reduce the problem to a copy of (4.4) but where the vectors are shorter and ae is smaller. The procedure is then repeated until the whole of u is found. It is easily realized that the infinite weights in ! do not change the algorithm and the algorithm will terminate with a solution of (4.4). We determine the actual merit weights D k from the solution \Upsilon(x k ) of (4.2). The weights may get large close to a saddle point and when the iterates diverge from this saddle point (that is always the case with the Gauss-Newton method) we would like the weights to decrease. This is accomplished by saving, say t, older versions, the merit weight matrices. Initially, at iteration and at the kth iteration we update m ), as in Algorithm 4.1. Algorithm 4.1. Solve (4.2) for the vector u(x k ). If d (k) i be the new sequence - (1) In Algorithm 4.2 our Gauss-Newton algorithm is described with line search and quadratic merit function. Algorithm 4.2. Initiate the start vector x k . while not convergence Compute Compute using the modified QR decomposition of J k . Determine D k from Algorithm 4.1. Determine the step length ff k such that 4.2. Proving global convergence. We will need the following two technical lemmas to prove that our algorithm is globally convergent. In the lemmas we use d k as an arbitrary diagonal element in D k . Lemma 4.1. Assume that d k - 0; and that fd k g is bounded. Let be the subsequence of fd k g such that d k j+1 ? d k j . Then the positive series converges if and only if converges. Proof. Take a a where a converges too. Now assume that b + N converges to b N . Hence, k g is a bounded sequence that increases to a limit b \Gamma and converges to b Lemma 4.2. Assume that an arbitrary component, d k , in the diagonal of D k stays bounded as k ! 1 and let v k be the corresponding diagonal element in V t . Then lim and the series converges. Proof. Let us first exclude the trivial case that v k becomes equal to the upper bound ! for a finite k. The sequence fv k g is an increasing infinite sequence. Hence, lim v k exists and is denoted v. Take ffl be an arbitrary small but fixed positive number. Then d k ? - k-values. Hence, v ? - and since ffl ? 0 was arbitrary this implies that v - d. From d k - v k it follows that thus we have v - d - d - v and consequently d k ! v. Let fd i k g be the subsequence of fd k g with d i k+1 ? d i k . From lemma 4.1 we know that the series converges if and only if converges. Let us now prove that the latter series converges. From v i k - d i k it follows that and hence Since a subserie of increases to v, the series converges. Since positive series it is sufficient to prove that it is bounded. Hence, it only remains to prove that the series converges. Since d the saved older weights are updated in step i 1 . When we reach d i 1+t there have been t updates and v i 1+t equals one of the earlier d t. In this way we can eliminate both this v i 1+t and the corresponding d i j . In the same way it is seen that v i 1+t+1 equals one of the d i j . That pair can also be eliminated from the series. We go on and eliminate elements in this way to get t. Thus the positive series bounded and so converges. That completes the proof. Our main global convergence theorem covers both bounded and unbounded sequences of merit weights. Theorem 4.3. Let fx k g and fD k g be generated by Algorithm 4.2. Assume that is bounded and that the system matrix in (2.2) is nonsingular in the closure of g. Then the sequence fx k g has either finite termination at a KKT point or an accumulation point that is a KKT point of (1.1). Proof. It is trivial that there is finite termination just at KKT points. Let us now assume that we have an infinite sequence. Algorithm 4.2 implies that it is sufficient to consider the following two cases : These cases will now be treated separately. There exist a subsequence fx i k g of fx k g such that kD i is bounded it its possible to choose a subsequence fx j k g of fx i k g such that x x for some e x. From Algorithm 4.2 it follows that kD k k ! 1 only when is continuous for all points in the closure of fx k g except KKT points, e x is both an accumulation point of fx k g and a KKT point. ii) From the inequality one can prove that a point e x cannot be an accumulation point of fx k g if there exist (The proof of (4.5) is a trivial extension of a similar proof in [7] pp. 21-22.) From Lemma 4.2 we know that converges and from the Goldstein- Armijo condition in Algorithm 4.2 then, for a given D k , it follows that for every point e x in the closure of fx k g, that is not a KKT point, there exists constants ffl ? 0 that (4.5) is satisfied. Hence, only KKT points remain as possible accumulation points. That proves the theorem in case ii). 4.3. Line search. We have chosen to keep things simple and therefore we use a standard cubic interpolation from [3] to approximate the minimum of our merit function OE(ff). Another, more efficient, line search algorithm can be found in [6]. 4.4. Regularization. We use a simple form of subspace minimization described for the unweighted and constrained case in [7]. We have not been able to prove a general global convergence result as the one in Theorem 4.3 but as we shall see in the computational experiments our regularization seems to work appropriately. 5. The generalized Newton-Raphson method. A constrained Newton method for solving (1.9) can be based on the quadratic subproblem G)p are first order approximations of the Lagrange multipliers. The solution, - p, to (5.1) is given by the linear system of equations G \Gammaf# The main disadvantage with using (5.2) is that for very large weights in W 2 the quadratic subproblem (5.1) and the matrix in (5.2) may be very ill conditioned. To avoid the ill conditioning due to large weights in W 2 we solve \Gammaf# and - is from (2.2). This method is the generalized Newton-Raphson method [9], or just the gNR method. The gNR method has an interesting theoretical motivation. Assume that we have reached a point x k . From the first order approximation (1.5) it is known that x k solves the perturbed problem min is a projection onto R(J k ). Hence, we know the solution x k of (5.4) and want to compute the solution of the perturbed problem min Then we can use the quadratic approximation of z(x) at x k to compute a solution of problem (5.5) whose error is O(kP k f k k 2 ). If we change back to the original notations in f(x), this perturbed solution is found by solving problem (5.3) for From (5.3) it is seen that there exists a matrix N k such that x as the solution to (1.1) and Take x . Then from the quadratic approximation in (5.6) we get From (5.6) it is also seen that J k N k only depends on the surface and not on the parameterization in x and consequently J k p k is independent of the parameterization in R n . The generalized Newton-Raphson method is in fact the only quadratically convergent method with J k p k independent of the parametrization. To see this we assume that there exists another method which computes e e . The series expansion (5.6) is unique and we have J k e which implies that e If we define Z 1 as a matrix whose columns span the null space of J 1 we call p a descent direction if p T Z T drawback with both the constrained Newton method based on (5.2) and the gNR method is that a nonsingular matrix in (5.2) or (5.3) is not sufficient for p to be a descent direction. However, we use the gNR method only when we are close to the solution, see (3.20) and Algorithm 6.1, and therefore we use the gNR method undamped. From (2.2) we get -, needed for G, and the Gauss-Newton search direction and if the matrix in (5.3) is singular we use the already available Gauss-Newton direction. If we use the modified QR decomposition to solve (2.2) it is possible to reduce the size of the system in (5.3). Ignoring the permutation matrix it is possible to rewrite R# Now implies that Q and we can reduce (5.7) to R T \GammaG \Gamma- are the first n elements in Q T - and Q respectively. The matrix in (5.8) may be indefinite and we must either use a stable method for indefinite systems, see e.g. [4], or add some condition on the submatrices in (5.8). One possibility of the latter kind is to assume that R is well conditioned and use R T to reduce (5.8) to R T \GammaG "\Gamma- The solution is the matrix R+MnR \GammaT G is nonsingular, if not we take a Gauss-Newton step. 6. Computational experiments. The algorithm we use in our tests is shown below. Algorithm 6.1. Initialize while Determine the Jacobian J and the vector f . Compute the GN direction p and - by solving (2.2). If regularization was needed then Second := false. If Close and Second and Rate ? 0:5 Compute the gNR direction, p gNR , by solving (5.3). If the matrix in (5.3) is nonsingular then If GN Compute the merit weights by Algorithm 4.1. Determine the step length ff using the line search described in Section (4.3) with the merit function OE(ff). x To use a pure GN method then the variable Second has a fixed value of false. We have tested our algorithm on three different problems described in the Ap- pendix; Schittkowski 308 [10], Boggs 2 and Boggs 8 [2]. The intention with the tests is not to show that the algorithms are faster than other existing algorithms but to show how our algorithms handles large weights and inadequate models (ill conditioning in the linear problems). Another important aim with the tests is to verify our theoretical results on the local convergence rate. Therefore it has been natural to use small and simple test problems. We define as two different measures of the convergence rate for the Gauss-Newton method. We emphasize that k is an excellent way of estimating the convergence rate when regularization is not needed and when b x is not known. The first problem, Schittkowski 308, is first solved with the Gauss-Newton method and the result is in Tab. 1. The largest weight is 10 20 and if the weights are multiplied explicitly with f , forming then the algorithm breaks down because of numerical instability. Note the slow growth of the merit weights. The first problem Schittkowski 308 with the Gauss-Newton method. 5 7.3e-3 6.6e-5 2.7e-3 4.6e-2 4.4e-2 3.0 99 1.0 6 3.4e-4 3.0e-6 1.2e-4 4.6e-2 4.6e-2 2.9 3.9 1.0 8 7.2e-7 6.5e-9 2.7e-7 4.6e-2 4.7e-2 3.0 1.0e+2 1.0 9 3.3e-8 3.0e-10 1.2e-8 4.6e-2 4.6e-2 3.0 1.0e+2 1.0 Table Schittkowski 308 with the gNR method. 5 5.9e-5 5.6e-12 2.2e-5 1.0 6 2.9e-11 2.3e-16 1.1e-11 1.0 solved with the gNR method is showed in Tab. 2. The asterisk indicates that the gNR method was used in that step. The second problem, Boggs 2, is a constrained problem and it has been solved with the Gauss-Newton method, Tab. 3, and the gNR method, Tab. 4. All the merit weights for the Gauss-Newton method were equal to one and are not shown in the Tab. 3. The remaining two test problems illustrate the regularization. The rank of the problem is shown under the headline Rank. In Tab. 5 the second test problem, Boggs 2, is solved with the Gauss-Newton method when the Jacobian is rank deficient at the starting point. In the third problem, Boggs 8, the Jacobian at the solution is rank deficient and the result is shown in Tab. 6. 7. Discussion. We claim that we have developed an efficient and fairly robust algorithm for solving (1.1) (with possibly infinite weights as discussed in the intro- duction). However, it is difficult for us to measure the effectiveness of the algorithm Boggs 2 with the Gauss-Newton method. 3 1.2e-2 2.8e-2 2.3 0.13 0.50 2.6 1.0 5 2.4e-4 2.5e-4 7.3e-3 0.10 1.9e-2 1.6e-2 1.0 6 6.4e-5 6.6e-5 3.0e-4 0.27 4.2e-2 1.6e-2 1.0 Table Boggs 2 with the gNR method. 3 1.2e-2 2.8e-2 2.3 1.0 5 2.4e-4 2.5e-4 7.3e-3 1.0 6 6.4e-5 6.6e-5 3.0e-4 1.0 9 8.2e-16 1.0e-15 1.9e-15 1.0 because there are, to our knowledge, no other algorithm that can solve such a general problem as (1.1). The local convergence properties are well understood for the Gauss-Newton algo- rithm. It is especially interesting that the local convergence results are valid for the whole problem class defined by (1.1) and that they are independent of the parametrization in R n . The merit function is especially suited for our weighted and constrained problem and our technique for choosing the merit weights is effective and do not lead to unnecessary large weights. Boggs 2, Gauss-Newton and rank deficient at the starting point. 9 0.47 0.71 0.14 1.0 0.12 3 As for robustness, we have shown that our algorithm is globally convergent when the iteration points are nondegenerate. It remains to find a way to regularize when the rows in J corresponding to very large weights become (almost) linearly dependent. We believe that this is a difficult and challenging problem to solve. Appendix . Test problems. In this appendix we define our three test problems and the weight sequences. We also give the starting points, x start ; solutions, b the residuals f(bx). The examples are from [10] and [2] and includes unconstrained as well as constrained problems. Schittkowski 308 [10] An unconstrained problem which we have modified by incorporation of weights. A constrained problem where the Jacobian is rank deficient at the second Boggs 8, Gauss-Newton and rank deficient at the solution. 28 1.1 1.0 0.65 1.0 3.4e-10 5 29 1.1 0.99 1.6 3.5 9.5e-7 4 53 4.2e-9 0.25 0.50 3.5 1.0 4 54 1.0e-9 0.25 0.50 3.5 1.0 3 starting point, x start2 . Boggs 8 [2] A constrained problem where the Jacobian is rank deficient at the solution. --R A Strategy for Global Convergence in a Sequential Quadratic Programming Algorithm Numerical methods for unconstrained optimization and nonlinear equations Matrix Computations Modifying the QR decomposition to weighted and constrained linear least squares Iterative solution of nonlinear equations in several variables A comparision of some algorithms for the nonlinear least squares problem More Test Examples for Nonlinear Programming Codes Perturbation theory and condition numbers for generalized and constrained linear least squares problems --TR --CTR Hiroshi Hosobe, Hierarchical nonlinear constraint satisfaction, Proceedings of the 2004 ACM symposium on Applied computing, March 14-17, 2004, Nicosia, Cyprus
nonlinear least squares;parameter estimation;optimization;weights
589066
On the Convergence of Pattern Search Algorithms.
We introduce an abstract definition of pattern search methods for solving nonlinear unconstrained optimization problems. Our definition unifies an important collection of optimization methods that neither compute nor explicitly approximate derivatives. We exploit our characterization of pattern search methods to establish a global convergence theory that does not enforce a notion of sufficient decrease. Our analysis is possible because the iterates of a pattern search method lie on a scaled, translated integer lattice. This allows us to relax the classical requirements on the acceptance of the step, at the expense of stronger conditions on the form of the step, and still guarantee global convergence.
Introduction . We consider the familiar problem of minimizing a continuously di#erentiable function f : R n # R. Direct search methods for this problem are methods that neither compute nor explicitly approximate derivatives of f . Our interest is in a particular subset of direct search methods that we will call pattern search methods. Our purpose is to generalize these methods and to present a global convergence theory for them. To our knowledge, this is the first convergence result for some of these methods and the first general convergence theory for all of them. Examples of pattern search methods include such classical direct search algorithms as coordinate search with fixed step sizes, evolutionary operation using factorial designs (first proposed by G. E. P. Box [2, 3, 13]), and the original pattern search algorithm of Hooke and Jeeves [7]. A more recent example is the multidirectional search algorithm of Dennis and Torczon [6, 15]. For some time, it has been apparent to us that the unifying theme that distinguishes these algorithms from other direct search methods is that each of them performs a search using a "pattern" of points that is independent of the objective function f . This informal insight is the basis for our general definition of pattern search methods-it turns out that each of the above pattern search methods is an instance of our general model. Formally, our definition of pattern search methods requires the existence of a lattice T such that if {x 1 , . , xN } are the first N iterates generated by a pattern search method, then there exists a scale factor #N such that the steps {x 1 lie in the scaled lattice #N T . The lattice depends on the pattern that defines the individual method and on the initial choice of the step length control parameter, but it is independent of the objective function f . The scaling # Received by the editors June 23, 1993; accepted for publication (in revised form) September 20, 1995. This research was sponsored by Air Force O#ce of Scientific Research grants 89-0363, F49620- 93-1-0212, and F49620-95-1-0210; United States Air Force grant F49629-92-J-0203; and Department of Energy grant DE-FG005-86ER25017. This research was also supported in part by the Geophysical Parallel Computation Project under State of Texas contract 1059. http://www.siam.org/journals/siopt/7-1/25078.html Department of Computer Science, The College of William &Mary, Williamsburg, VA 23187-8795 (va@cs.wm.edu). This work was completed while the author was in the Department of Computational and Applied Mathematics and the Center for Research on Parallel Computation, Rice University, Houston, depends solely on the sequence of updates that have been applied to the step length control parameter. Despite isolated convergence results [4, 11, 16] for certain individual pattern search methods, a general theory of convergence for the class of such methods remained elusive for some time. The standard convergence theory for line search and trust region methods depends crucially on some notion of su#cient decrease, but pattern search methods do not enforce any such notion. Therefore, attempts such as [18] to apply the standard theory to pattern search methods arbitrarily introduce some notion of su#cient decrease, thereby modifying the original algorithms. Thus, the challenge was to develop a general convergence theory for pattern search methods without redefining what they are. Our convergence analysis is guided by that found in Torczon [16] for the multidirectional search algorithm; however, the present level of abstraction makes the important elements of that analysis easier to appreciate. The present paper also includes a correction to the specification of the scaling factors found in [16]. There are three key points to our analysis. First, we show that pattern search methods are descent methods. Second, we prove that pattern search methods are gradient-related methods in the sense of [10]. Finally, we demonstrate that pattern search methods cannot terminate prematurely due to inadequate step length control mechanisms. The crucial element of this analysis is the fact that pattern search methods are able to relax the conditions on accepting a step by enforcing stronger conditions on the step itself. The lattice T , together with the way in which the step length control parameter is updated, prevent a pathological choice of steps: steps of arbitrary lengths along arbitrary search directions are not permitted. We are able to guarantee that, if the function f is continuously di#erentiable, then an explicit representation of the gradient or the directional derivative. In particular, we prove global convergence for pattern search methods despite the fact that they do not explicitly enforce a notion of su#cient decrease on their iterates, such as fraction of Cauchy decrease, fraction of optimal de- crease, or the Armijo-Goldstein-Wolfe conditions. However, our convergence analysis does share certain characteristics with the classical convergence analysis of both line search and trust region methods. This connection is both subtle and unexpected. Our convergence analysis for pattern search methods makes it clear why these methods are as robust as their proponents have long claimed, while clarifying some of the limitations that have long been ascribed to them. In addition, having identified the common structure of these methods, it is now possible to develop new pattern search methods with guaranteed global convergence. In section 2 we establish the notation and general specification of pattern search methods. In section 3 we prove that if the function to be minimized is continuously dif- ferentiable, then pattern search methods guarantee that lim inf k#f(x k In addition, we identify the modifications that must be made to pattern search methods to obtain the stronger result lim k#f(x k In section 4 we show that the classical pattern search methods mentioned above, as well as the newer multidirectional search algorithm of Dennis and Torczon, conform to the general specification for pattern search methods. In section 5, we give some concluding remarks; section 6 contains technical results needed for the proofs of section 3. Notation. We denote by R, Q, Z, and N the sets of real, rational, integer, and natural numbers, respectively. All norms are Euclidean vector norms or the associated operator norm. We define ON THE CONVERGENCE OF PATTERN SEARCH ALGORITHMS 3 2. Pattern search methods. We begin by introducing the following abstraction of pattern search methods. We defer to section 4 demonstrations that the pattern search methods mentioned above fall comfortably within this abstraction. 2.1. The pattern. To define a pattern we need two components, a basis matrix and a generating matrix. The basis matrix can be any nonsingular matrix B # R n-n . The generating matrix is a matrix C k # Z n-p , where p > 2n. We partition the generating matrix into components (1) We require that M k # M # Z n-n , where M is a finite set of nonsingular matrices, and that L k # Z n-(p-2n) and contains at least one column, the column of zeros. A pattern P k is then defined by the columns of the matrix P Because both B and C k have rank n, the columns of P k span R n . For convenience, we use the partition of the generating matrix C k given in (1) to partition P k as follows: (2) Given # k # R, # k > 0, we define a trial step s i k to be any vector of the form denotes a column of C ]. Note that Bc i determines the direction of the step, while # k serves as a step length parameter. At iteration k, we define a trial point as any point of the form x i x k is the current iterate. 2.2. The exploratory moves. Pattern search methods proceed by conducting a series of exploratory moves about the current iterate before declaring a new iterate and updating the associated information. These moves can be viewed as sampling the function about the current iterate x k in a well-defined deterministic fashion in search of a new iterate x with a lower function value. The individual pattern search methods are distinguished, in part, by the manner in which these exploratory moves are conducted. To allow the broadest possible choice of exploratory moves and yet still maintain the properties required to prove convergence for the pattern search methods, we place two requirements on the exploratory moves associated with any particular pattern search method. These requirements are given in the following Hypotheses on exploratory moves. (Please note an abuse of notation that is nonetheless convenient: means that the vector y is contained in the set of columns of the matrix A.) Hypotheses on exploratory moves. 1. 2. If min{f(x k The choice of exploratory moves must ensure two things: 1. The direction of any step s k accepted at iteration k is defined by the pattern its length is determined by # k . 2. If simple decrease on the function value at the current iterate can be found among any of the 2n trial steps defined by # k B# k , then the exploratory moves must produce a step s k that also gives simple decrease on the function value at the current iterate. In particular, f(x k need not be less than or equal to min{f(x k Thus, a legitimate exploratory moves algorithm would be one that somehow guesses which of the steps defined by # k P k will produce simple decrease and then evaluates the function at only one such step. (And that step may be contained in # k BL k rather than in # k B# k .) At the other extreme, a legitimate exploratory moves algorithm would be one that evaluates all p steps defined by # k P k and returns the step that produced the least function value. These are the properties of the exploratory moves that enable us to prove lim inf even though we only require simple decrease on f . Thus we avoid the necessity of enforcing either fraction of Cauchy decrease, fraction of optimal decrease, or the Armijo-Goldstein-Wolfe conditions on the iterates. To obtain lim we need to place stronger hypotheses on the exploratory moves as well as place a boundedness condition on the columns of the generating matrices. These extensions will be discussed further in section 3.3.2. 2.3. The generalized pattern search method. Algorithm 1 states the generalized pattern search method for unconstrained minimization. Algorithm 1. The Generalized Pattern Search Method. For (a) Compute (b) Determine a step s k using an exploratory moves algorithm. (c) Compute (d) If # k > 0 then x Update C k and # k . To define a particular pattern search method, it is necessary to specify the basis matrix B, the generating matrix C k , the exploratory moves to be used to produce a step s k , and the algorithms for updating C k and # k . 2.4. The updates. Algorithm 2 specifies the requirements for updating # k . The aim of the updating algorithm for # k is to force # k > 0. An iteration with otherwise, the iteration is unsuccessful. Again we note that to accept a step we only require simple, as opposed to su#cient, decrease. Algorithm 2. Updating # k . Given # Q, let # w0 and # k # w1 , . , # wL (a) If # k # 0 then # (b) If # k > 0 then # The conditions on # and # ensure that 0 < # < 1 and # i # 1 for all # i #. Thus, if an iteration is successful it may be possible to increase the step length parameter k is not allowed to decrease. Not surprisingly, this is crucial to the success of the analysis. Also crucial to the analysis is the relationship (overlooked in [16]) between # and the elements of #. ON THE CONVERGENCE OF PATTERN SEARCH ALGORITHMS 5 The algorithm for updating C k depends on the pattern search method. For theoretical purposes, it is su#cient to choose the columns of C k so that they satisfy (1) and the conditions we have placed on the matrices M k # M # Z n-n and L k # Z n-(p-2n) . 3. The convergence theory. Having set up the machinery to define pattern search methods, we are now ready to analyze these methods. This analysis produces theorems of several types. The first, developed in section 3.1, demonstrates an algebraic fact about the nature of pattern search methods that requires no assumption on the function f . This theorem is critical to the proof of the convergence results for it shows that we only need require simple decrease in f to ensure global conver- gence. The second theorem, developed in section 3.2, describes the limiting behavior of the step length control parameter # k if we place only a very mild condition on the function f and exploit the interaction of the simple decrease condition for the generalized pattern search method with the algorithm for updating # k . Finally, the third and fourth theorems, developed in section 3.3, give the global convergence results. The first theorem guarantees lim inf k#f(x k generalized pattern search method that satisfies the specifications given in section 2. This is significant since the theorem applies to all the pattern search methods we discuss in section 4 without the need to impose any modifications on the methods as originally stated. The second theorem is equivalent to convergence results for line search and trust-region globalization strategies. We can guarantee lim k#f(x k but to do so requires placing stronger conditions on the specifications for generalized pattern search methods. We could certainly impose these stronger conditions on the pattern search methods presented in section 4-none of them are unreasonable to suggest or to enforce-but we would do so at the expense of attractive algorithmic features found in the original methods. 3.1. The algebraic structure of the iterates. The results found in this section are purely algebraic facts about the nature of pattern search methods; they are also independent of the function to be optimized. It is the algebraic structure of the iterates that allows us to prove global convergence for pattern search methods without imposing a notion of su#cient decrease on the iterates. We begin by showing in what sense # k is a step length parameter. Lemma 3.1. There exists a constant # > 0, independent of k, such that for any trial step s i produced by a generalized pattern search method (Algorithm 1) we have Proof. From (3) we have s i k . The conditions we have placed on the generating matrix C k ensure that c i the smallest singular value of B. Then The last inequality holds because at least one of the components of c i k is a nonzero integer, and hence #c i k # 1. From Lemma 3.1 we can see that the role of # k as a step length parameter is to regulate backtracking and thus prevent excessively short steps. Theorem 3.2. Any iterate xN produced by a generalized pattern search method 6 VIRGINIA TORCZON (Algorithm 1) can be expressed in the following form: z k , where . x 0 is the initial guess, . # , with # N and relatively prime, and # is as defined in the algorithm for updating # k (Algorithm 2), . r LB and r UB depend on N , . # 0 is the initial choice for the step length control parameter, . B is the basis matrix, and . z k # Z n , Proof. The generalized pattern search algorithm, as stated in Algorithm 1, guarantees that any iterate xN is of the form s k . (We adopt the convention that s iteration k is unsuccessful.) We also know that the step s k must come from the set of trial steps s i p. The trial steps are of the form s i k . Consider the step length parameter # k . For any k # 0, the update for # k given in Algorithm 2 guarantees that # k is of the form (Recall that We have also placed the following restrictions on the form of # and L. We can thus rewrite (5) as: where r k # Z. Let Then from (4) and (6) we have Since # is rational, we can express # as # , where # N are relatively prime. Then, using (7), z k , where z k # Z n . Theorem 3.2 synthesizes the requirements we have placed on the pattern, the definition of the trial steps, and the algorithm for updating # k . Note that this means that for a fixed N , all the iterates lie on a translated integer lattice generated by x 0 and the columns of # rLB # -rUB # 0 B. ON THE CONVERGENCE OF PATTERN SEARCH ALGORITHMS 7 3.2. The limiting behavior of the step length control parameter. The next theorem combines the strict algebraic structure of the iterates with the simple decrease condition of the generalized pattern search algorithm, along with the algorithm for updating # k , to give us a useful fact about the limiting behavior of # k . Theorem 3.3. Assume that L(x 0 ) is compact. Then lim inf k# Proof. The proof is by contradiction. Suppose 0 < #LB # k for all k. From (6) we know that # k can be written as # The hypothesis that #LB # k for all k means that the sequence {# rk } is bounded away from zero. Meanwhile, we also know that the sequence {# k } is bounded above because all the iterates x k must lie inside the set L(x 0 and the latter set is compact; Lemma 3.1 then guarantees an upper bound #UB for {# k }. This, in turn, means that the sequence {# rk } is bounded above. Consequently, the sequence {# rk } is a finite set. Equivalently, the sequence {r k } is bounded above and below. Let Then (8) now holds for the bounds given in (9), rather than (7), and we see that for all k, x k lies in the translated integer lattice G generated by x 0 and the columns of The intersection of the compact set L(x 0 ) with the translated integer lattice G is finite. Thus, there must exist at least one point x # in the lattice for which x for infinitely many k. We appeal to the simple decrease condition in the generalized pattern search method (Algorithm 1 (d)), which guarantees that a lattice point cannot be revisited infinitely many times since we accept a new step s k if and only if f(x k ) > Thus there exists an N such that for all k # N , x which implies that # We now appeal to the algorithm for updating # k (Algorithm 2 (a)) to see that thus leading to a contradiction. 3.3. Global convergence. Throughout the discussion in this section, we assume that f is continuously di#erentiable on a neighborhood of L(x 0 ); however, this assumption can be weakened, using the same style of argument found in [16]. 3.3.1. The general result. To prove Theorem 3.5 we need Proposition 3.4. We defer the proof of Proposition 3.4 to section 6 in part because we wish to discuss there several other issues that are tangential to the proof of Theorem 3.5. It is also the case that the proofs for the results in section 6 are similar to those given for the equivalent results found in [16], though now restated more succinctly in terms of the machinery developed in section 2. Proposition 3.4. Assume that L(x 0 ) is compact, that f is continuously di#er- entiable on a neighborhood of L(x 0 ), and that lim inf k#f(x k )#= 0. Then there exists a constant #LB > 0 such that for all k, # k > #LB . We emphasize that the existence of a positive lower bound #LB for # k is guaranteed only under the null hypothesis that lim inf k#f(x k )#= 0. Theorem 3.5. Assume that L(x 0 ) is compact and that f is continuously di#er- entiable on a neighborhood of L(x 0 ). Then for the sequence of iterates {x k } produced by the generalized pattern search method (Algorithm 1), lim inf Proof. The proof is by contradiction. Suppose that lim inf k#f(x k )#= 0. Then Proposition 3.4 tells us that there exists #LB > 0 such that for all k, # k #LB . But this contradicts Theorem 3.3. 3.3.2. The stronger result. We can strengthen the result given in Theorem 3.5 at the expense of wider applicability. To begin with, we must add three further restrictions: one on the pattern matrix, one on the Hypotheses on exploratory moves, and one on the limiting behavior of the step length control parameter # k . First, we must ensure that the columns of the generating matrix C k are bounded in norm, i.e., that there exists a constant C > 0 such that for all k, C > #c i k # for all p. Given this bound, we can place an upper bound, in terms of # k , on the norm of any trial step s i k . Lemma 3.6. Given a constant C > 0 such that for all k, C > #c i k # for all there exists a constant # > 0, independent of k, such that for any trial step s i k produced by a generalized pattern search method (Algorithm 1) we have Proof. From (3) we have s i k . Then #s i C||B|| . Note that the columns of M k # M are bounded by the assumption that |M| < +#; we use this fact in the proof of Proposition 6.4. The stronger boundedness condition on the columns of C is needed to monitor the behavior of L k . Second, we must replace the original Hypotheses on exploratory moves with a stronger version, as given below. Together, Lemma 3.6 and the Strong hypotheses on exploratory moves allow us to tie decrease in f to the norm of the gradient when the step sizes get small enough. This is the import of Corollary 6.5, which is given in section 6. Strong hypotheses on exploratory moves. 1. 2. If min{f(x k Third, we require that lim k# We can use the algorithm for updating to ensure that this condition holds. For instance, we can force # k to be nonincreasing by requiring w which when taken together with Theorem 3.3 guarantees that lim k# All the algorithms we consider in section 4, except the multidirectional search algorithm, enforce this condition by limiting However, it is not necessary to force the steps to be nonin- creasing; we need only require that in the limit the step length control parameter goes to zero, which, in conjunction with Lemmas 3.1 and 3.6, has the e#ect of ultimately forcing the steps to zero. Theorem 3.7. Assume that L(x 0 ) is compact and that f is continuously differentiable on a neighborhood of L(x 0 ). In addition, assume that the columns of the generating matrices are bounded in norm, that lim k# and that the generalized pattern search method (Algorithm 1) enforces the Strong hypotheses on exploratory moves. Then for the sequence of iterates {x k } produced by the generalized pattern search method, lim ON THE CONVERGENCE OF PATTERN SEARCH ALGORITHMS 9 Proof. The proof is by contradiction. Suppose lim sup k#f(x k )#= 0. Let be such that there exists a subsequence #f(xm i )#. Since lim inf given any 0 < #, there exists an associated subsequence l i such that )#. Then, since # k # 0, we can appeal to Corollary 6.5 to obtain for su#ciently large, Then the telescoping sum l i gives us Since f is bounded below, f(xm i because #f is uniformly continuous, for i su#ciently large. However, )# 2#. Since equation (10) must hold for any #, 0 < #, we have a contradiction (e.g., try The proof of Theorem 3.7 is almost identical to that of an equivalent result for trust-region methods that was first given by Thomas [14] and which is included, in a more general form, in the survey by Mor-e [8]. One final note: the hypotheses of Theorem 3.7 suggest that in the absence of any explicit higher-order information about the function to be minimized, it makes sense to terminate a generalized pattern search algorithm when # k is less than some reasonably small tolerance. In fact, this is a common stopping condition for algorithms of this sort and the one implemented for the multidirectional search algorithm [17]. 4. The particular pattern search methods. In section 2 we stated the conditions an algorithm must satisfy to be a pattern search method. We now illustrate these conditions by considering the following specific algorithms: . coordinate search with fixed step lengths, . evolutionary operation using factorial designs [2, 3, 13], . the original pattern search method of Hooke and Jeeves [7], and . the multidirectional search algorithm of Dennis and Torczon [6, 15]. We will show that these algorithms satisfy the conditions that define pattern search methods and thus are special cases of the generalized pattern search method presented as Algorithm 1. Then we can appeal to Theorem 3.5 to claim global convergence for these methods. There are other algorithms for which the abstraction and accompanying analysis holds-including various modifications to the algorithms presented-but we shall confine our investigation to these, the best known of the pattern search methods, to illustrate the power of our abstract approach to pattern search methods. 4.1. Coordinate search with fixed step lengths. The method of coordinate search is perhaps the simplest and most obvious of all the pattern search methods. Davidon describes it concisely in the opening of his belated preface to Argonne National Laboratory Research and Development Report 5990 [5]: Enrico Fermi and Nicholas Metropolis used one of the first digital computers, the Los Alamos Maniac, to determine which values of certain theoretical parameters (phase shifts) best fit experimental data (scattering cross sections). They varied one theoretical parameter at a time by steps of the same magnitude, and when no such increase or decrease in any one parameter further improved the fit to the experimental data, they halved the step size and repeated the process until the steps were deemed su#ciently small. Their simple procedure was slow but sure. This simple search method enjoys many names, among them alternating direc- tions, alternating variable search, axial relaxation, and local variation. We shall refer to it as coordinate search. Perhaps less obvious is that coordinate search is a pattern search method. To see this, we begin by considering all possible outcomes for a single iteration of coordinate search when shown in Fig. 1. We mark the current iterate x k . The x i 's denote trial points considered during the course of the iteration. The next iterate x k+1 is marked. Solid circles indicate successful intermediate steps taken during the course of the exploratory moves while open circles indicate points at which the function was evaluated but that did not produce further decrease in the value of the objective function. Thus, in the first scenario shown a step from x k to x 1 k resulted in a decrease in the objective function, so the step from x 1 k to x k+1 was tried and led to a further decrease in the objective function value. The iteration was then terminated with a new point x k+1 that satisfies the simple decrease condition f(x k+1 ) < f(x k ). In the worst case, the last scenario shown, 2n trial points were evaluated k , and producing decrease in the function value at the current iterate x k . In this case, x and the step size must be reduced for the next iteration. We now show this algorithm is an instance of a generalized pattern search method. 4.1.1. The matrices. Coordinate search is usually defined so that the basis matrix is the identity matrix; i.e., However, knowledge of the problem may lead to a di#erent choice for the basis matrix. It may make sense to search using a di#erent coordinate system. For instance, if the variables are known to di#er by several orders of magnitude, this can be taken into account in the choice of the basis matrix (though, as we will see in section 6.2, this may have a significant e#ect on the behavior of the method). The generating matrix for coordinate search is fixed across all iterations of the method. The generating matrix C contains in its columns all possible combi- ON THE CONVERGENCE OF PATTERN SEARCH ALGORITHMS 11 Fig. 1. All possible subsets of the steps for coordinate search in R 2 . #k z }| { Fig. 2. The pattern for coordinate search in R 2 with a given step length control parameter # k . nations of {-1, 0, 1}. Thus, C has columns. In particular, the columns of C contain both I and -I, as well as a column of zeros. We define of the remaining 3 n columns of C. Since C is fixed across all iterations of the method, there is no need for an update algorithm. For . Thus, when possible trial points defined by the pattern given step length # k , can be seen in Fig. 2. Note that the pattern includes all the possible trial points enumerated in Fig. 1. 4.1.2. The exploratory moves. The exploratory moves for coordinate search are given in Algorithm 3, where the e i 's denote the unit coordinate vectors. Algorithm 3. Exploratory Moves Algorithm for Coordinate Search. Given x k , # k , f(x k ), and B, set s For do (a) s i . Compute (b) If f(x i Otherwise, k . Compute (ii) If f(x i k . Return. The exploratory moves are executed sequentially in the sense that the selection of the next trial step is based on the success or failure of the previous trial step. Thus, while there are 3 n possible trial steps, we may compute as few as n trial steps, but we compute no more than 2n at any given iteration, as we saw in Fig. 1. From the perspective of the theory, there are two conditions that need to be met by the exploratory moves algorithm. First, as Figs. 1 and 2 illustrate, all possible trial steps are contained in # k P . The second condition on the exploratory moves is the more interesting; coordinate search demonstrates the laxity of this second hypothesis. For instance, in the first scenario shown in Fig. 1, decrease in the objective function was realized for the first trial step so the second trial step was tried and accepted. It is certainly possible that greater decrease in the value of the objective function might have been realized for the trial step which is defined by a column in the matrix M (the step s 2 k is defined by a column in the matrix L), but s # k is not tried when simple decrease is realized by the step s 1 k . However, in the worst case, as seen in Fig. 1, the algorithm for coordinate search ensures that all 2n steps defined by # k B# k B[M -M are tried before returning the step s In other words, the exploratory moves given in Algorithm 3 examine all 2n steps defined by # k B# unless a step satisfying 4.1.3. Updating the step length. The update for # k is exactly as given in Algorithm 2. As noted by Davidon, the usual practice is to continue with steps of the same magnitude until no further decrease in the objective function is realized, at which point the step size is halved. This corresponds to setting Thus, This su#ces to verify that coordinate search with fixed step length is a pattern search method. Theorem 3.5 thus holds. The exploratory moves algorithm for coordinate search would need to be modified to satisfy the Strong hypotheses on exploratory moves for the conditions of Theorem 3.7 to be met. ON THE CONVERGENCE OF PATTERN SEARCH ALGORITHMS 13 4.2. Evolutionary operation using factorial designs. In 1957 G. E. P. Box [2] introduced the notion of evolutionary operation as a method for increasing industrial productivity. The ideas were developed within the context of the on-line management of industrial processes, but Box recognized that the technique had more general applicability. Subsequent authors [3, 13] argued that the basic technique was readily applicable to general unconstrained optimization and it is within this context that we examine the ideas here. In its simplest form, evolutionary operation is based on using two-level factorial designs: evaluate the function at the vertices of a hypercube centered about the current iterate. (G. E. P. Box refers to this as one of a variety of "pattern of variants" [2].) If simple decrease in the value of the objective function is observed at one of the vertices, it becomes the new iterate. Otherwise, the lengths of the edges in the hypercube are halved and the process is repeated. 4.2.1. The matrices. As with coordinate search, the usual choice for the basis matrix is though, as with coordinate search, other choices may be made to reflect information known about the problem to be solved. The generating matrix for evolutionary operation is fixed across all iterations of the method. The generating matrix C contains in its columns all possible combinations of {-1, 1}; to this we append a column of zeros. Thus C has columns. We take M to be any linearly independent subset of n columns of C; -M necessarily will be contained in C. Once again, L is fixed and consists of the remaining columns of C. There is no need for an algorithm to update C since the generating matrix is fixed. 4.2.2. The exploratory moves. The exploratory moves given in Algorithm 4 are simultaneous in the sense that every possible trial step s i computed at each iteration. It is then the case that every trial step s i k is contained in . The second observation of note is that since of our choice of M (and thus, by extension, our choice of #). Furthermore, we are guaranteed that the Strong hypotheses on exploratory moves are satisfied. Algorithm 4. Exploratory Moves Algorithm for Evolutionary Operation Given x k , # k , f(x k ), B, and , set s For do (a) s i k . Compute (b) If f(x i k . Return. 4.2.3. Updating the step length. The algorithm for updating # k is exactly as given in Algorithm 2, with # usually set to 1/2 and Since we have shown that evolutionary operation satisfies all the necessary re- quirements, we can therefore conclude that it, too, is a pattern search method, so Theorem 3.5 holds. The algorithm, as stated above, also satisfies the conditions of Theorem 3.7. 14 VIRGINIA TORCZON Fig. 3. The pattern step in R 2 , given x k #= x k-1 , k > 0. 4.3. Hooke and Jeeves' pattern search algorithm. In addition to introducing the general notion of a "direct search" method, Hooke and Jeeves introduced the pattern search method, a specific kind of search strategy [7]. The pattern search of Hooke and Jeeves is a variant of coordinate search that incorporates a pattern step in an attempt to accelerate the progress of the algorithm by exploiting information gained from the search during previous successful iterations. The Hooke and Jeeves pattern search algorithm is opportunistic. If the previous iteration was successful (i.e., # k-1 > 0), then the current iteration begins by conducting coordinate search about a speculative iterate x k the current iterate x k . This is the pattern step. The idea is to investigate whether further progress is possible in the general direction x k - x k-1 (since, if x k #= x k-1 , then x k - x k-1 is clearly a promising direction). To make this a little clearer, we consider the example shown in Fig. 3. Given x k-1 and x k (we assume, for now, that k > 0 and that x k #= x k-1 ), the pattern search algorithm takes the step x k - x k-1 from x k . The function is evaluated at this trial step and the trial step is accepted, temporarily, even if f(x k The Hooke and Jeeves pattern search algorithm then proceeds to conduct coordinate search about the temporary iterate x k Thus, in R 2 , the exploratory moves are exactly as shown in Fig. 1, but with x k substituted for x k . If coordinate search about the temporary iterate x k then the point returned by coordinate search about the temporary iterate is accepted as the new iterate x k+1 . If not, i.e., then the pattern step is deemed unsuccessful, and the method reduces to coordinate search about x k . For the two dimensional case, then, the exploratory moves would simply resort to the possibilities shown in Fig. 1. If the previous iteration was not successful, so x the iteration is limited to coordinate search about x k . In this instance, though, the updating algorithm for # k will have reduced the size of the step (i.e., # The algorithm does not execute the pattern step when To express the pattern search algorithm within the framework we have developed, we use all the machinery required for coordinate search. Once again, the basis matrix is usually defined to be We append to the generating matrix another set of columns to capture the e#ect of the pattern step and we change the exploratory moves algorithm, as detailed below. 4.3.1. The generating matrix. Recall that the generating matrix for coordinate search consists of all possible combinations of {-1, 0, 1} and is never changed. For the Hooke and Jeeves pattern search method, we allow the generating matrix to change from iteration to iteration to capture the e#ect of the pattern step. We append ON THE CONVERGENCE OF PATTERN SEARCH ALGORITHMS 15 another set of 3 n columns, consisting of all possible combinations of {-1, 0, 1}, to the initial generating matrix for coordinate search. Thus C k has columns. The additional 3 n columns allow us to express the e#ect of the pattern step with respect to x k , rather than with respect to the temporary iterate x k which is how the Hooke and Jeeves pattern search method usually is described. The matrix M is unchanged; is allowed to vary, though only in the 3 n columns associated with the pattern step. For . For notational convenience, we require that the last column of C 0 , which we denote as c p 0 , be the column of zeros. In both the algorithm for updating C k (Algorithm 5) and the algorithm for the exploratory moves (Algorithm 6), we use the column c p k to measure the accumulation of a sequence of successful pattern steps. This can be seen, in (12), for our example from Fig. 3. In this example, we have the generating matrix . The pattern step represented by the vector (1 1) T , seen in the last column of C k . Note that the only di#erence between the columns of C 0 given in (11) and the columns of C k given in (12) is that (1 1) T has been added to the last 3 2 columns of C k . The algorithm for updating the generating matrix updates the last 3 n columns of C k ; the first 3 n columns remain unchanged, as in coordinate search. The purpose of the updating algorithm is to incorporate the result of the search at the current iteration into the pattern for the next iteration. This is done using Algorithm 5. Note the distinguished role of c p k , the last column of C k , which represents the pattern step Algorithm 5. Updating C k . For do k . Return. Since (1/# k )s k is necessarily a column of C k and C 0 # Z n-p , an argument by induction shows that the update algorithm for C k ensures that the columns of C k always consist of integers. 4.3.2. The exploratory moves. In Algorithm 6, the e i 's denote the unit co-ordinate vectors and c p k denotes the last column of C k . We set is defined when A useful example for working through the logic of the algorithm can be found in [1], though the presentation and notation di#er somewhat from that given here. Algorithm 6. Exploratory Moves Algorithm for Hooke and Jeeves. Given x k , # k , f(x k ), B, and # k-1 , set # For do (a)s i k . Compute (b)If f(x i k . Otherwise, k . Compute k . For do (a)s i k . Compute (b)If f(x i k . Otherwise, k . Compute k . Return. All possible steps are contained in # k P k since C k contains columns that represent the "pattern steps" tried at the beginning of the iteration. And, once again, the exploratory moves given in Algorithm 6 examine all 2n steps defined by # k B# unless a step satisfying Since we have shown that the pattern search algorithm of Hooke and Jeeves satisfies all the necessary requirements, we can therefore conclude that it, too, is a special case of the generalized pattern search method and Theorem 3.5 holds. 4.4. Multidirectional search. The multidirectional search algorithm was introduced by Dennis and Torczon in 1989 [15] as a first step towards a general purpose optimization algorithm with promising properties for parallel computation. While subsequent work led to a class of algorithms (based on the multidirectional search algorithm) that allows for more flexible computation [6, 17], one of the unanticipated results of the original research was a global convergence theorem for the multidirectional search algorithm [16]. The multidirectional search algorithm is a simplex-based algorithm. The pattern of points can be expressed as a simplex (i.e., points or vertices) based at the current iterate; as such, multidirectional search owes much in its conception to its predecessors, the simplex design algorithm of Spendley, Hext, and Himsworth [12] and the simplex algorithm of Nelder and Mead [9]. However, multidirectional search is a di#erent algorithm-particularly from a theoretical standpoint. Convergence for the Spendley, Hext, and Himsworth algorithm can be shown only with some modification of the original algorithm, and then only under the additional assumption that the function f is convex. There are numerical examples to demonstrate that the Nelder-Mead simplex algorithm may fail to converge to a stationary point of the function because the uniform linear independence property (discussed in section 6.2), which plays a key role in the convergence analysis, cannot be guaranteed to hold [15]. The multidirectional search algorithm is described in detail in both [6] and [16]. The formulation given here is di#erent and, in fact, introduces some redundancy that can be eliminated when actually implementing the algorithm. However, the way of expressing the algorithm that we use here allows us to make clear the similarities between this and other pattern search methods. 4.4.1. The matrices. It is most natural to express multidirectional search in terms of multiple basis matrices B k and a fixed generating matrix C, which is at odds with our definition for generalized pattern search methods. As we shall see, however, ON THE CONVERGENCE OF PATTERN SEARCH ALGORITHMS 17 it is possible to convert the more natural specification to one that conforms to our requirements for a pattern search method. The multidirectional search algorithm centers around a family of basis matrices B that consists of all matrices representing the edges adjacent to each vertex in a nondegenerate n-dimensional simplex that the user is allowed to specify. Since the ordering of the columns is not unique and typically not preserved in the implementation of the method, we consider all possible representations of the columns of the matrices associated with the edges adjacent to the (n+1) vertices of the simplex. We then add the negatives of these (n basis matrices to account for the e#ect of the reflection step allowed by the multidirectional search algorithm. Thus the cardinality of the set B is Fortunately, there is no need to construct this unwieldy number of basis matrices to initialize the method. We can update the basis matrix after each iteration k by reconstructing the new basis matrix B k+1 , given the outcome of the exploratory moves, from the trial points x i during the course of the exploratory moves. This procedure is given in Algorithm 7. The scalar scale is chosen during the course of the exploratory moves (see Algorithm 8) to ensure that factoring out any change in the size of the simplex introduced by a change in # k . This has the further e#ect of preserving the role of # k as a step length parameter. Algorithm 7. Updating B k . Given scale, best, and x i For (best - 1) do best For (best do best Otherwise For do Return. Given this use of a family of basis matrices to help define the multidirectional search algorithm, the generating matrix is then the fixed matrix Thus, C contains To ensure that C # Z n-p , we Z. Furthermore, to ensure that the role of # k as a step length parameter is not lost with the introduction of the expansion step represented by -I, we require - #. The algorithm is defined so that # w1 , # w2 This requires the further restriction that # N. Again, this is not an onerous restriction. Multidirectional search usually is specified so that 2. Now, to bring this notation into conformity with our definition for a generalized pattern search method, observe that we can represent all possible basis matrices B # in terms of a single reference matrix B # B so that A convenient feature of using the edges of a simplex to form the set of basis matrices is that the matrices - consist only of elements from the set {-1, 0, 1}. The matrices are necessarily nonsingular because of the nondegeneracy of the simplex. We use to represent the set of matrices - and observe that since B is a finite set, the set B is also finite. We then observe that Thus we can define the pattern in terms of the single reference matrix B and the redefined generating matrix with B. We also have L k # [- 0] and since - # Z, 4.4.2. The exploratory moves. The exploratory moves for the multidirectional search method are given in Algorithm 8; the e i 's denote the unit coordinate vectors. We use the notion of B k # B for consistency with the update algorithm given in Algorithm 6, but we could just as easily substitute B - in the algorithm given below. Algorithm 8. Exploratory Moves Algorithm for Multidirectional Search. Given # N, set s For do (a) s i k . Compute (b) If f(x i k , and best = i. For do (a) s i k . Compute (b) If f(x i k , and best = i. For do (a) s i (b) If f(x i Return. Clearly, s k # k P k . Since the exploratory moves algorithm considers all steps of the form # k B# k , unless simple decrease is found after examining only the steps defined by # k BM k , this guarantees we satisfy the condition that if min{f(x k +y), y # 4.4.3. Updating the step length. The algorithm for updating # k is that given in Algorithm 2. In this case, while # usually is set to 1/2 so that and include an expansion factor usually equals one. Thus usually 2. The choice of # k # is made during the execution of the exploratory moves. Since we have shown that the multidirectional search algorithm satisfies all the necessary requirements, we conclude that it is also a pattern search method and thus Theorem 3.5 applies. Note that since we allow - > 1, which is a useful algorithmic feature, we cannot guarantee that lim k# and so Theorem 3.7 does not automatically apply. ON THE CONVERGENCE OF PATTERN SEARCH ALGORITHMS 19 5. Conclusions. We have presented a framework in which one can analyze pattern search methods. This framework abstracts and quantifies the similarities of the classical pattern search methods and enables us to prove lim inf k#f(x k for this class of algorithms. We also specify the conditions under which the limit can be shown to hold. These convergence results are perhaps surprising, given the simplicity of pattern search methods, but derive from the algebraic rigidity imposed on the iterates produced by pattern search methods. This is gratifying, since while this rigidity originally was introduced as a heuristic for directing the exploratory moves, it turns out to be the key to proving convergence as well. This analysis also highlights just how weak the conditions on the acceptance of the step can be and yet still allow a global convergence analysis, an observation that may prove useful in the analysis of other classes of optimization methods. 6. Technical results. We deferred the proof of Proposition 3.4 for several rea- sons. First, many of the results in this section are generalizations of similar results to be found in [16]. The abstraction in section 2 leads to more succinct proofs. Second, the proof of Proposition 3.4 is closely related to that of several other results presented in this section and requires us to introduce several additional notions. We return to our definition of the pattern as to show that the pattern contains at least one direction of descent whenever #f(x k ) #= 0. Recall that we require the columns of C k to contain both M k and -M k . Thus, can be partitioned as follows: We now elaborate on these requirements. Since M k is an n-n nonsingular matrix and B is nonsingular, we are guaranteed that BM k forms a basis for R n . Further, we are guaranteed that at any iteration k, if #f(x k ) #= 0, x k - Bc i is a direction of descent for at least one column c i k contained in the block # k . 6.1. Descent methods. Of course, the existence of a trial step in a descent direction is not su#cient to guarantee that decrease in the value of the objective function will be realized. To guarantee that a pattern search method is a descent method, we need to guarantee that in a finite number of iterations the method produces a positive step size # k that achieves decrease on the objective function at the current iterate. We now show that this is the case. Lemma 6.1. Suppose that f is continuously di#erentiable on a neighborhood of there exists q # Z, q # 0 such that # k+q > 0 (i.e., the q)th iteration is successful). Proof. A key hypothesis placed on the exploratory moves is that if descent can be found for one of the trial steps defined by # k B# k , then the exploratory moves returns a step that produces descent. Because BC k has rank n, if there exists at least one trial direction d i loss of generality. Thus, there exists an h k > 0 such that for 0 < h # h k , If at iteration k, # k > h k , then the iteration may be unsuccessful; that is, # When the iteration is unsuccessful, the generalized pattern search method sets x and the updating algorithm sets # is strictly less than one, there exists q # Z, q # 0, such that # q # k # h k . Thus we are guaranteed descent, i.e., a successful iteration, in at most q iterations. 6.2. Uniform linear independence. The pattern P k guarantees the existence of at least one direction of descent whenever #f(x k ) #= 0. We now want to guarantee the existence of a bound on the angle between the direction of descent contained in B# k and the negative gradient at x k (whenever #f(x k ) #= 0). We will show, in fact, that this bound is uniform across all iterations of the pattern search algorithm. To do so, we use the notion of uniform linear independence [10]. Lemma 6.2. For a pattern search algorithm, there exists a constant # > 0 such that for all k # 0 and x #= 0, Proof. To demonstrate the existence of #, we first consider the simplest possible I and and use this to derive a bound for any choice of B and C k that satisfies the conditions we have imposed. Lemma 6.3. Suppose where the e j 's are the unit coordinate vectors. I and min cos #(y) =# n Proof. We have |y T e j are guaranteed that |y j | # 1/ # n for some j, so |y T e j | # 1/ # n for some j. Thus cos #(y) # 1/ # n. Now note that cos #(y) attains this lower bound for any y Thus, if the pattern search is restricted to the coordinate directions defined by gives the lower bound on the absolute value of the cosine of the angle between the gradient and a guaranteed direction of descent. We now use the bound for this particular case to derive a bound for the general case. Assume a general basis matrix B and a general matrix M k # M, where |M| < +#. We adopt the notation BM k ]. Then for any x #= 0 we have the If we set -T w, we have ON THE CONVERGENCE OF PATTERN SEARCH ALGORITHMS 21 where #(BM k ) is the condition number of the matrix BM k . Thus, we have To ensure a bound # that is independent of the choice of any particular matrix simply observe that the set M is required to be finite. Thus, # is taken to be M#M . The bound given in (14) points to two features that explain much about the behavior of pattern search methods. Since we never explicitly calculate-or approximate- the gradient, we are dependent on the fact that in the worst case at least one of our search directions is not orthogonal to the gradient; # gives us a bound on how far away we can be. Thus, as either the condition number of the product BM k increases, or the dimension of the problem increases, our bound on the angle between the search direction and the gradient deteriorates. This suggests two things. First, we should be very careful in our choice of B and M for any particular pattern search method. Second, we should not be surprised that these methods become less e#ective as the dimension of the problem increases. Nevertheless, even though pattern search methods neither require nor explicitly approximate the gradient of the function, the uniform linear independence condition demonstrates that the pattern search methods are, in fact, gradient-related methods, as defined by Ortega and Rheinboldt [10], which is one reason why we can establish global convergence. 6.3. The descent condition. Having introduced the notion of uniform linear independence with the bound #, we are now ready to show that pattern search methods reduce # k only when necessary to find descent. To do this we will show that once the steps s i are small enough, then a successful step must be returned by the exploratory moves algorithm. Lemma 3.1 allows us to restate this condition in terms of # k . We use the result to prove Proposition 3.4. Proposition 6.4. Suppose that L(x 0 ) is compact and f is continuously di#er- entiable on a neighborhood of L(x 0 ). Given # > 0, Suppose also that x 0 # . Then there exists # > 0, independent of k, such that if # and # k < #, then the kth iteration of a generalized pattern search method (see Algorithm 1) will be successful (i.e., # thus Proof. We restrict our attention to the steps defined by the columns of # k B# k . This is su#cient since the Hypotheses on exploratory moves ensure that a step s k satisfying the simple decrease condition # k > 0 must be returned if a trial step defined by a column of # k B# k satisfies the simple decrease condition. If s i is a step defined by # k B# k (we assume that P k is partitioned as in (2) so that the first 2n columns of P k contain the columns of B# k # independent of k, 22 VIRGINIA TORCZON since M k # M # Z n-n and M is a finite set of matrices. Together, (15) and Lemma 3.1 yield allows us to define dist (L(xN are compact and disjoint, we know that d > 0. If # k < d/2# , then #s i Thus x i k lies in the interior of L(x 0 ) for all precisely, for all lies in the ball B(x k , d/2) # L(x 0 ). x#f(x)#. By design, # > 0. Since #f is continuous on a neighborhood of L(x 0 ), #f is uniformly continuous on a neighborhood of L(x 0 ). Thus, there exists a constant r > 0, depending only on # and the # from (13), such that We define min d, r . We are now assured that if then and We are ready to argue that if at any iteration k # N , x k # and (17) is satisfied, then an acceptable step will be found. Choose a trial point x i and The definitions of# and the pattern P k , together with Lemma 6.2, guarantee the existence of at least one such x i k . Since (17) holds by assumption, (18) also holds. We can apply the mean value theorem to obtain f(x i Consider the first term on the right-hand side of (20). Our choice of x i k gives us ON THE CONVERGENCE OF PATTERN SEARCH ALGORITHMS 23 Furthermore, since #f(x k ) T Now consider the second term on the right-hand side of (20). The Cauchy-Schwarz inequality gives us Combine (21) and (22) to rewrite (20) as holds by assumption, (19) also holds. We then have Thus, when # k < #, f(x i for at least one s i k defined by # k Bc i 2n. The Hypotheses on exploratory moves guarantee that if min{f(x k and the algorithm for updating # k (Algorithm 2) ensures that # k+1 # k . Proposition 6.4 guarantees that if # k is small enough, a generalized pattern search method realizes simple decrease because there exists at least one step among the 2n steps defined by # k B# k that gives decrease as a function of the norm of the gradient at the current iterate, as shown in (23); the Hypotheses on exploratory moves then ensure that the exploratory moves algorithm must return a step that satisfies at least simple decrease. However, there are no guarantees that the step returned by an exploratory moves algorithm satisfies more than the simple decrease condition. To tie the amount of actual decrease to the norm of the gradient, we must place much stronger conditions on the generalized pattern search method, as discussed in section 3.3.2. Once we have done so, Corollary 6.5 follows more or less immediately from Proposition 6.4. Corollary 6.5. Suppose that L(x 0 ) is compact and f is continuously di#eren- tiable on a neighborhood of L(x 0 ). Suppose that the columns of the generating matrix are bounded in norm and that the generalized pattern search method (Algorithm 1) enforces the Strong hypotheses on exploratory moves. Given # > 0, let Suppose also that x 0 # . Then there exist # > 0 and # > 0, independent of k, such that for all but finitely many k, if x k # and # k < #, then Proof. From Proposition 6.4, (23) says that for k # (Lemma 6.1 guarantees the existence of N), there exists at least one trial step s i such that once # k < #, where # is as defined in (16), we have The Strong hypotheses on exploratory moves give us Lemma 3.1 ensures that Lemma 3.6, which holds only when the columns of the generating matrix are bounded in norm, gives us We define # 2 # to complete the proof. We now prove Proposition 3.4. Proof. By assumption, lim inf k#f(x k )#= 0. Then we can find N 1 and such that for all k # N 1 , x k guarantees the existence of N From Proposition 6.4 we are assured of # > 0 such that if # k #, then the iteration will be successful. Given # 0 , there exists a constant q # Z, q # 0, such that # q # 0 #, where # (0, 1) and is as defined in the algorithm for updating # k (Algorithm 2). Thus, for k # N , # q+1 # 0 < # k . Acknowledgments . This paper benefited from conversations with J. E. Dennis, Stephen Nash, Michael Trosset, Lu-s Vicente, and especially Michael Lewis. In partic- ular, discussions with Michael Lewis were critical in the distillation of the abstraction for pattern search methods found in section 2 and in the development of the analytic arguments for the algebraic structure of the iterates found in section 3.1. The review and comments made by an anonymous referee, Danny Ralph (the second referee, who agreed to reveal his identity for this acknowledgment), and Jorge Mor-e, the editor, are gratefully acknowledged. In particular, the observation by Danny Ralph that the pattern contains only vectors in a fixed lattice led to a more general result and a much more elegant presentation. --R Analysis and Methods Variable metric method for minimization "Direct search" A simplex method for function minimization Iterative Solution of Nonlinear Equations in Several Variables Computational Methods in Optimization: A Unified Approach Sequential application of simplex designs in optimisation and evolutionary operation in Numerical Methods for Unconstrained Optimization Sequential Estimation Techniques for Quasi-Newton Algorithms A Direct Search Algorithm for Parallel Machines On the convergence of the multidirectional search algorithm PDS: Direct Search Methods for Unconstrained Optimization on Either Sequential or Parallel Machines Positive basis and a class of direct search techniques --TR --CTR M. Hintermller, Solving nonlinear programming problems with noisy function values and noisy gradients, Journal of Optimization Theory and Applications, v.114 n.1, p.133-169, July 2002 D. Byatt , I. D. Coope , C. J. Price, Conjugate Grids for Unconstrained Optimisation, Computational Optimization and Applications, v.29 n.1, p.49-68, October 2004 Stefano Lucidi , Marco Sciandrone, A Derivative-Free Algorithm for Bound Constrained Optimization, Computational Optimization and Applications, v.21 n.2, p.119-142, February 2002 rpd Brmen , Janez Puhan , Tadej Tuma, Grid Restrained Nelder-Mead Algorithm, Computational Optimization and Applications, v.34 n.3, p.359-375, July 2006 J. M. Gablonsky , C. T. Kelley, A Locally-Biased form of the DIRECT Algorithm, Journal of Global Optimization, v.21 n.1, p.27-37, September 2001 Peter Buchholz , Dennis Mller , Peter Kemper , Axel Thmmler, OPEDo: a tool framework for modeling and optimization of stochastic models, Proceedings of the 1st international conference on Performance evaluation methodolgies and tools, October 11-13, 2006, Pisa, Italy Lennart Frimannslund , Trond Steihaug, A generating set search method using curvature information, Computational Optimization and Applications, v.38 n.1, p.105-121, September 2007 Todd A. Sriver , James W. Chrissis, Combined pattern search and ranking and selection for simulation optimization, Proceedings of the 36th conference on Winter simulation, December 05-08, 2004, Washington, D.C. Michael Heitzer, Structural Optimization with FEM-based Shakedown Analyses, Journal of Global Optimization, v.24 n.3, p.371-384, November 2002 Hong-Xuan Huang , Panos M. Pardalos , Zuo-Jun Shen, A Point Balance Algorithm for the Spherical Code Problem, Journal of Global Optimization, v.19 n.4, p.329-344, April 2001 Convergence Analysis of Unconstrained and Bound Constrained Evolutionary Pattern Search, Evolutionary Computation, v.9 n.1, p.1-23, January 2001 Rommel G. Regis , Christine A. Shoemaker, Constrained Global Optimization of Expensive Black Box Functions Using Radial Basis Functions, Journal of Global Optimization, v.31 n.1, p.153-171, January 2005 Hart, Locally-adaptive and memetic evolutionary pattern search algorithms, Evolutionary Computation, v.11 n.1, p.29-51, Spring Matthias Ihme , Alison L. Marsden , Heinz Pitsch, Generation of optimal artificial neural networks using a pattern search algorithm: Application to approximation of chemical systems, Neural Computation, v.20 n.2, p.573-601, February 2008 Genetha Anne Gray , Tamara G. Kolda , Ken Sale , Malin M. Young, Optimizing an Empirical Scoring Function for Transmembrane Protein Structure Determination, INFORMS Journal on Computing, v.16 n.4, p.406-418, Fall 2004 Jack Dongarra , Ian Foster , Geoffrey Fox , William Gropp , Ken Kennedy , Linda Torczon , Andy White, References, Sourcebook of parallel computing, Morgan Kaufmann Publishers Inc., San Francisco, CA,
coordinate search;evolutionary operation;pattern search;direct search methods;local variation;multidirectional search;convergence analysis;axial relaxation;unconstrained optimization;globalization strategies;downhill simplex search;alternating variable search
589078
Degenerate Nonlinear Programming with a Quadratic Growth Condition.
We show that the quadratic growth condition and the Mangasarian--Fromovitz constraint qualification (MFCQ) imply that local minima of nonlinear programs are isolated stationary points. As a result, when started sufficiently close to such points, an $L_\infty$ exact penalty sequential quadratic programming algorithm will induce at least R-linear convergence of the iterates to such a local minimum. We construct an example of a degenerate nonlinear program with a unique local minimum satisfying the quadratic growth and the MFCQ but for which no positive semidefinite augmented Lagrangian exists. We present numerical results obtained using several nonlinear programming packages on this example and discuss its implications for some algorithms.
Introduction Recently, there has been renewed interest in analyzing and modifying sequential quadratic programming (SQP) algorithms for constrained nonlinear optimization for cases where the traditional regularity conditions do not hold [14,13, 20,25]. This research has been motivated by the fact that large-scale nonlinear programming problems tend to be almost degenerate (have large condition numbers for the Jacobian of the active constraints). It is therefore important to establish to what extent the convergence properties of the SQP methods are dependent on the ill-conditioning of the constraints. In this work, we term as degenerate those nonlinear programs (NLPs) for which the gradients of the active constraints are linearly dependent. In this case there may be several feasible Lagrange multipliers. Many of the previous analysis and rate of convergence results for degenerate NLP are based on the validity of second-order conditions. These are essentially equivalent to the condition in unconstrained optimization that, for a critical point of a function f(x) to be a local minimum, f xx 0 is a necessary condition and f xx 0 is a sufficient condition. Here is the positive semidefinite ordering. The place of f xx in constrained optimization is taken for these conditions by L xx , the Hessian of the Lagrangian, which is now required to be positive definite on the critical cone for one or all of the Lagrange multipliers [6,21]. Department of Mathematics, University of Pittsburgh This work was completed while the author was the Wilkinson Fellow at the Mathematics and Computer Science Division, Argonne National Laboratory, 9700 South Cass Avenue, Argonne, IL 60439. This work was supported by the Mathematical, Information, and Computational Sciences Division subprogram of the Office of Advanced Scientific Computing, U.S. Department of Energy, under Contract W-31-109-Eng-38. This work differs from previous approaches in that we assume only that 1. At a local solution x of the constrained nonlinear program, the first-order Mangasarian-Fromowitz constraint qualification holds. 2. The quadratic growth condition (QG) [4,16] is satisfied: for some oe ? 0 and all x feasible in a neighborhood of x . 3. The data of the problem are twice continuously differentiable. These assumptions are equivalent to a weaker form of the second-order sufficient conditions [15,4] which do not require the positive semidefinitenes of the Hessian of the Lagrangian on the entire critical cone. We prove that these conditions guarantee that x is the only local stationary point (3) of the nonlinear program. This is an important issue because it guarantees that descent-like algorithms will not stop arbitrarily close to x , except at x . This extends a result from [21] that required some second-order sufficient conditions to be satisfied for all multipliers. In particular, our work implies that if MFCQ holds and the second-order sufficient conditions hold for one multiplier, then x is a strict local minimum and an isolated stationary point. We also show that, under the same assumptions, the L1 exact penalty sequential quadratic program (SQP) induces at least Q linear convergence [19] of the penalized objective to f(x ) and R-linear convergence of the iterates. Finally, we provide an example of a nonlinear program that satisfies our assumptions for which it is not possible to construct an augmented Lagrangian such that x will be an unconstrained local minimum. This may present an adverse case to algorithms based on this assumption, such as Lagrange multiplier methods. However, we show that it is possible to construct a nondifferentiable function that has x as its minimum, namely the L1 penalty function (which can also be inferred from the results in [4]). We describe our computational experience with several nonlinear programming packages applied to this example and discuss the expected and observed behavior of Lagrangian multiplier methods. Our convergence analysis for the L1 exact penalty function suggests that it is possible to construct a convergence theory with much more general second-order conditions. This may result in algorithms with superior robustness, because their properties depend on significantly fewer assumptions. 1.1. Previous Work, Framework, and Notations We deal with the NLP problem f(x) subject to g(x) 0; (2) are twice continuously differentiable. We call x a stationary point if the following conditions hold for some 2 IR Degenerate Nonlinear Programming with a Quadratic Growth Condition 3 Here L is the Lagrangian function If certain regularity conditions hold (discussed below), then a local solution x of (2) is a stationary point. In that case (3) are referred to as the KKT (Karush-Kuhn-Tucker) conditions. Since our analysis will be limited to a neighborhood of a point x that is a strict minimum, we will assume that all constraints are active at x , or g(x Such a situation can be obtained by simply dropping the constraints i for which 0, since this relationship holds in an entire neighborhood of x . This does not reduce the generality of our results, but it simplifies the notation because now we do not have to refer separately to the active set. The regularity condition, or constraint qualification, ensures that a linear approximation of the feasible set in the neighborhood of x captures the geometry of the feasible set. Often in local convergence analysis of constrained optimization algorithms, it is assumed that the constraint gradients rg i are linearly independent, so that the Lagrange multiplier in (3) is unique. We assume instead the Mangasarian-Fromowitz constraint qualification (MFCQ): n . (5) It is well known [9] that MFCQ is equivalent to boundedness of the set M(x ) of Lagrange multipliers that satisfy (3), that is, Note that M(x ) is certainly polyhedral in any case. The critical cone at x is [6,22] We briefly review the some of the second-order conditions in the literature, although they are not an assumption for our analysis but only a basis for com- parison. In the framework of [6], the second-order sufficient conditions for x to be an isolated local solution of (2) are: 9 2 M(x If these conditions hold at x for some , then the quadratic growth condition is satisfied, irrespective of the validity of the first-order constraint qualification [6,7]. However, this does not imply that x is an isolated stationary point, as shown by a simple example [21], which may prevent an optimization algorithm that uses only first derivative information from reaching x even when started arbitrarily close to x . In [21] it is shown that if MFCQ holds, and the relation (8) is satisfied for all 2 M(x ) then x is an isolated stationary point and a minimum of (2). Also, with these conditions, the exact solution is Lipschitz stable with respect to perturbations. By compactness of M(x ), we can choose oe independently of in this case. In [1] it is proven that, under these assumptions, the L1 exact penalty SQP will converge Q-linearly to f(x ), when the descent direction is computed by a QP using only first-order information. A refinement of the second-order conditions was introduced in [15]. In the presence of MFCQ, those conditions require that Further analysis shows that, in presence of MFCQ, these conditions are necessary and sufficient for the quadratic growth condition to hold [4,15,16,22]. Also, the exact solution is Lipschitz stable with respect to certain classes of perturbations [22], though not to any perturbation (see an example in [10, p.308]). In this paper we assume only the quadratic growth condition and MFCQ, and thus we do not use the perturbation results. If the condition (9) holds, but (8) does not, then there is no positive semidefinite augmented Lagrangian, as we will show with an example. This is an interesting aspect since it invalidates the usual working assumption of Lagrange multiplier methods [3]. Finally, we review some of the facts concerning the L1 nondifferentiable exact penalty function: We are looking for an unconstrained minimum of the function where c OE is a sufficiently large constant. Descent directions d of OE(x) at the point x can be obtained by solving the following quadratic program (QP) [3]: subject to g j (x) where H is some positive definite matrix and g 0 In this paper the analysis will be restricted to the case although the same results apply for any other positive definite matrix. At the current point x k of an iterative procedure that attempts to determine x , the QP (11) generates the descent direction d k . The next iterate is x obtained by a line search procedure. Usual stepsize rules are the minimization rule, the limited minimization rule, and the Armijo rule [3]. For these rules, any limit point of fx k g is a stationary point of OE(x), and the descent procedure is therefore globally convergent in this sense [3]. If, in addition, Degenerate Nonlinear Programming with a Quadratic Growth Condition 5 for some 2 M(x ), then x is a stationary point of OE(x) [2]. A suitable value for c OE is not available in the early stages of the algorithm, but a good estimate can be found via an update scheme [2]. Here we assume that c is constant and for all 2 M(x ), where fl is some prescribed safety factor. Consider the quadratic program subject to g j (x) We denote the unique solution of this program by d(x) and the set of its multipliers by M(x). At x (14) has the same multiplier set as (2), which are both denoted by M(x ). Since MFCQ is satisfied at x , this QP is feasible in a neighborhood of x . The KKT conditions for this QP require With these notations, d(x If the QP (14) were unconstrained, then its solution would be We name a descent-like algorithm a sequential quadratic program that solves instances of the above QP. At x , the QP (14) satisfies MFCQ and some second-order sufficient con- ditions. From [21] there exists c d such that, in a neighborhood of x we have there exists 2 M(x ) such that Therefore, from the definition of c OE , there exists a neighborhood of x such that for all multipliers 2 M(x). For such x, it can be verified by inspection that is a solution of (11) [2, p. 195]. We therefore concentrate on the QP (14), because, if c OE is large enough and we are sufficiently close to x , it generates the same descent direction as (11), thus sharing its global convergence property. For some function h : IR k we denote by c 1h , c 2h bounds depending on the first and second derivatives of h. The positive and negative parts of h(x) are h componentwise. With this notation 6 Mihai Anitescu 2. Stationary Points of NLPs Satisfying MFCQ In this section, we assume that x is in a sufficiently small neighborhood of x , whose size or properties are specified in each of the following results. In particular the standing assumptions hold on all neighborhoods considered here and Here p with one of the vectors satisfying (5), suitable neighborhood of x . Lemma 1. There exist a neighborhood W (x ) such that Here P (x) is the usual L1 penalty function (10). Proof. We have by Taylor's theorem We choose For 0 ff ff P we have The claim follows after choosing c proof of the following lemma can be inferred from [4]. We include it here for completeness. Lemma 2. There exists a c OE such that for all x in a neighborhood of x . Proof. Let r ? 0 be such that B(x ; r) ae W (x ). We choose r 2 such that This is always possible because We then have that, for any x 2 B(x and thus x r). By the intermediate value theorem, we have that implying in turn that rg i Degenerate Nonlinear Programming with a Quadratic Growth Condition 7 Take now If x is infeasible, then ff 1 ? 0 and there exists i such that g i applies to give If x is feasible, then ff and the previous bound still applies. From the quadratic growth assumption (1) and the feasibility of x+ ff 1 p, we must have that or By (23) and Taylor's theorem we have c P Choose c P c P Then by (23) c P c P c P c P Using (25), (24) and (26) we get The conclusion follows, because from the Cauchy-Schwartz inequality. We can assume that c OE from the previous lemma satisfies (17), or otherwise we replace it with the right-hand side of (17) and the conclusion of the lemma still holds for the new c OE . To prove the following results, we will use the results from [12] concerning sets defined by linear inequalities: For such a set, denote by d(x; P) the distance from a point x 2 IR n to the set P. Also, denote by dP (x) the maximum value of the infeasibility: Then there exists a number (P) ? 0 such that If we have equality constraints, we recast them as two inequality constraints. The following lemma uses the fact that M(x ) is polyhedral and can thus be expressed in the form (27). Lemma 3. Let I be an index set such that there exists a multiplier with lambda there exists a constant c I such that 8 2 M(x ) there exists a 2 M(x ) with and such that jj \Gamma jj c I jj I jj 1 . For a vector we have denoted by I the restriction of the vector to the index set I. Proof. Let M I (x ) be the set of all 2 M(x ) such that I = 0. Then 0: (32) From our assumptions, M I (x ) is not empty. By eventually rescaling the x space, we can assume, without loss of generality, that the vectors defining the equality constraints in (30) are of norm 1; otherwise, if all entries are 0, we remove that row, and the feasible set remains unchanged. M I can be described in the alternative, way: I 0; (35) I 0; (36) 0; (37) where each row is described by a unit vector, which puts the set in the form (27). Thus from [12] there exists a (M I ) ? 0 such that (M I )d(; M I ) dMI (): (38) However, since 2 M(x ) is a valid multiplier set, we have that only the constraints I 0, (35), are violated. Thus . The conclusion follows from (38) by taking c I . The proof is complete. We define Iaef1;::;mg c I ; for feasible M I (x Lemma 4. There exists a neighborhood W of x such that, 8x 2 W; 2 M(x), implies that there exists a 2 M(x ) with I = 0. Degenerate Nonlinear Programming with a Quadratic Growth Condition 9 Proof. Assume the contrary. Then there exists a sequence x k ! x such that there exists k 2 M(x) and an index set I for which I = 0, but I 6= 0, there is only a finite set of index sets, we can extract an infinite subsequence for which the above happens for a fixed set I. By extracting another subsequence, we can assume that k is convergent, from (16) and the fact that M(x ) is compact. But then k ! 2 M(x ) and I = 0, a contradiction. From here on we will use extensively that, for h twice continuously differen- tiable, we have is a continuous function with / 3h Indeed by Taylor's theorem we have that there exist continuous functions / 1 3h and 3h 3h 3h and The relation (40) now follows by comparing the last two equations. Theorem 1. There exists a constant c oe ? 0 such that in a neighborhood of x we have that where (d; ) is the solution of the QP (14). Proof. From (16), there exists a 2 M(x ) such that jj \Gamma jj c d jjx \Gamma x jj. Let I be the set of indices i for which We have jj I jj c d jjx \Gamma x jj. From (39) and Lemmas 3 and 4 there exists a ~ 2 M(x ) with ~ As a result and The important consequence of this fact, using the complementarity relations from (15), is that \Gamma( ~ Indeed, all the above equalities are 0. From Lemma 2 we have that Here (d; ) is a solution of (15), and ~ We also used (40). We now employ the identity ab (42), and Taylor's theorem for rg(x) to get, by continuing from the previous equation, We denote From sufficiently close to x . By using 1 , (40), (43) and (42), we get (x). Using the above bound in (52), together with \Gammad T Degenerate Nonlinear Programming with a Quadratic Growth Condition 11 We can now choose a sufficiently small neighborhood of x such that /(jjx \Gamma x jj) oeand subtract the last term of the last relation from the lower bound We take and with this new notation, we get that We treat jjx \Gamma x jj as a variable and, by using the formulas for the quadratic equation, we get that oe By using the arithmetic-quadratic mean inequality, we get that Choosing oe c OE g (67) we prove the claim. Corollary 1. x is an isolated stationary point. Proof. Let x be another stationary point of the NLP in the neighborhood of x where the above theorem holds. Therefore there exists a 2 M(x) satisfying (3). Hence solution of (15) and is the unique solution of the strictly convex QP (14). Since is feasible from (14) and P (x). Now from the complementarity conditions in (15) we get T From the previous theorem we get which proves the claim. Corollary 2. If the second-order sufficient condition (8) is satisfied for one multiplier, and if MFCQ holds at x , then x is an isolated stationary point. Proof. Since x is satisfies the quadratic growth condition (1) under these assumptions [6,7] and MFCQ holds, Corollary 1 applies. 3. An Example Without a Locally Convex Augmented Lagrangian Consider the matrix Take We then have that u T 1. Since the vector corresponds to the positive eigenvalue, we have that for any u at an angle of at most 6 from u 0 , u T Qu 1jjujj 2 . Consider now the rotation matrix since Q 0 and Q 2 have the same axes of symmetry, but with the eigenvalues switched. Also, for any u 2 IR 2 , there exists a k such that u T Q k u 1jjujj 2 , since the wide cones centered at the axis of the positive eigenvalues of Q k now sweep the entire IR 2 . Consider now the optimization problem minz subject to z By the previous observation, we have that z 1(x 2 +y 2 ) on the feasible set; thus z 0. Clearly, the only solution of the problem is (0; 0; 0). Since z z 2, if 4, we have that z 1(x 2 +y 2 +z 2 ), for all x; Therefore at x the quadratic growth condition is satisfied for the above NLP, with constant 1. Obviously, MFCQ holds at (0; 0; 0), and a simple calculation shows that a multiplier of (70). In particular, at least one multiplier has to be positive. Also, at (0; 0; 0), all constraints are active and their gradients are (0; 0; \Gamma1) for any of them. As a result, the linear constraints in (8) now become either z 0 or z = 0, with at least one being Therefore the critical cone at x is 0g. Also, from (3), if 2 M(x ), then Assume that there is a choice 2 M(x ) such that L xx , the Hessian of the Lagrangian, is positive semidefinite on the critical cone: y zA 0; 8(x; y; z); such that z = 0: (71) This is equivalent to X Since our construction is invariant to rotations with (U T that the positive semi-definiteness holds for any circular permutation oe of this multiplier set: X We denote by A c (4) the set of circular permutations of four elements. Since the set of positive definite matrices is a convex cone, and we must have which is impossible. Therefore L xx cannot be positive semidefinite on the critical cone for any choice 2 M(x ). Hence the second-order conditions from [6,21] will not hold for any choice of the multipliers. Degenerate Nonlinear Programming with a Quadratic Growth Condition 13 3.1. Augmented Lagrangian Approaches Here we discuss the expected behavior of augmented Lagrangian techniques when applied to this example. For these methods, the inequalities of the NLP (2) are converted into equalities [3,5]. The feasible set can be represented as [5] The NLP is replaced by a bound-constrained optimization problem. The equality constraints are incorporated in the objective function based on an estimate of the multipliers and a penalty term, subject to t i 0; Here is the barrier parameter. The objective function in (76) is the augmented Lagrangian. The problem is subjected to an additional trust-region constraint [5] to enforce global convergence. The desired outcome is to have bounded away from zero and the trust-region inactive as approaches M(x ) and the solution of the above problem approaches x . If that happens for our example, then, by a continuity argument following the lower boundedness of , should be a solution of (76) for an appropriate choice of ; . Since (76) has linearly independent gradients of the constraints, both the first and second order necessary conditions must hold [7]. The first order necessary condition results in where , with components i 0, are the multipliers associated with the variables As a result 2 M(x ). The second order necessary conditions require that I 4 be positive semidefinite, at least on the subspace of (ffix; ffi t) with results in or We proved that the last matrix cannot be positive semidefinite for our example and we thus get a contradiction. This shows that, either the trust region will be active arbitrarily close to x , or ! 0. 14 Mihai Anitescu This also shows that the Hessian of the augmented Lagrangian of the equality constrained problem F xx +X is not positive semidefinite and thus the augmented Lagrangian of the equality constrained problem cannot be locally convex. 4. Linear Convergence of the SQP with Nondifferentiable Exact Penalty P (x) The points x considered in thus subsection are assumed to be sufficiently close to x . The notation d and 2 M(x) will refer to the solutions of (14) and (15). Also, P (x) is the L1 penalty function (10) and 4.1. Proof of the Technical Results Lemma 5. Proof. Since d is a feasible point of (14), we have that rg i (x) T d \Gammag i (x); 8i 2 mg. By Taylor's remainder theorem Hence This completes the proof. Lemma 6. There exist ff]: Proof. Writing the KKT conditions for (14), we obtain and, hence, (d) T d +rf(x) T d (d) Degenerate Nonlinear Programming with a Quadratic Growth Condition 15 since, by the complementarity conditions satisfied by the solution of (14), T rg(x) T m. Therefore, since g i \Gamma(d) \Gamma(d) by (10), (17). By Taylor's remainder theorem, Hence, for ff 2 [0; 1], ff(\Gamma(d) from (79) and Lemma 5. Therefore, for ff 2 [0; 1], The result of the statement follows by choosing and Lemma 7. There exists a constant c 5 such that, 8() 2 M(x), Proof. From (15) and the definition of the Lagrangian (4) it follows, using Tay- lor's theorem, that, for a sufficiently small neighborhood of x, g. Also, by (16), we can choose 2 M(x ) such that we have that and, thus Therefore The conclusion of the lemma follows by choosing c 1g. 4.2. Nondifferentiable Exact Penalty Algorithms and the Linear Convergence Theorem The linearization algorithm [3, p.372] has the following form: 1. 2. Compute d k from (11). 3. Choose ff k from a line search procedure, and set x 4. return to Step 2. The stepsize ff k is chosen by one of the following procedures [3, p.372]. (a) Minimization rule Here ff k is chosen such that (b) Limited minimization rule Here a fixed scalar s ? 0 is selected, and ff k is chosen such that (c) Armijo rule Here fixed scalars s, , and oe with s ? 0, 2 (0; 1), and oe 2 (0; 1) are chosen and we set ff is the first nonnegative integer m for which It can be shown that the Armijo rule yields a stepsize after a finite number of iterations. The following theorem establishes the convergence properties of the linearization algorithm. The global convergence properties, established in [2, Prop. 4.3.3], are also stated here for completeness. Theorem 2. Let x k be a sequence generated by the linearization algorithm, where the stepsize ff k is chosen by the minimization rule, limited minimization rule or the Armijo rule. Then any accumulation point of the sequence x k is a stationary point of is a strict local minimum of the problem (2) satisfying the local quadratic growth (1) and the Mangasarian-Fromowitz constraint qualification (5), then OE(x k Q-linearly and x k ! x R-linearly. Proof. The first part is an immediate consequence of [2, Prop. 4.3.3]. We prove the linear convergence statement only for the Armijo rule, the proof being similar for the other stepsize selection mechanisms. By Lemma 6 for all ff 2 [0; ff]. Since m k is the smallest integer m for which Degenerate Nonlinear Programming with a Quadratic Growth Condition 17 it follows that m s ff. This therefore ensures that the stepsize is at least ff for k sufficiently large. As a result of Lemma 6, we have that On the other hand, by Lemma 7 we have that By Theorem (1) and the previous relation it follows that there exists c c6 by using Lemma 6 and where . After some obvious manipulation, it follows that which proves the Q-linear convergence [19] of the sequence OE(x k ) to OE(x ) with a linear rate of at most ffi\Gamma1 . Therefore lim sup From Lemma 2 Therefore lim sup which proves the R-linear convergence [19] to 0 of the sequence x . The proof is complete. Following the techniques from [1], we can extend the result for the case where the matrix H of the QP is not I but changes from iteration to iteration. The only condition is that the sequence of strictly convex H k be uniformly upper and lower bounded. Iteration OE(x k )\GammaOE(x ) 9 4.00 Table 1. Rates of convergence for the L1 penalty algorithm Iteration (New) Penalty Parameter Trust Region Radius jjjj 1 43 1e-4 1.1 e-02 268 1e-14 1.93 283 1e-16 4.41 e02 336 STOP Table 2. Reduction of the penalty parameter for LANCELOT 5. Numerical Experiments with Degenerate NLP We experimented with several nonlinear programming packages on the example from Section 3. Certainly, comparing the behavior of NLP algorithms on a unique degenerate example cannot result in a complete characterization. Nev- ertheless, it may be of interest to determine whether methods using augmented Lagrangians will really encounter problems when solving an example without a positive semidefinite augmented Lagrangian. We also desire to validate the theoretical conclusions of the preceding sections. We have shifted the origin for our example, to avoid one step convergence of algorithms that start at 0; 0; 0 by default. The algebraic form of the example is minz From our analysis, we have that w is a minimum satisfying the quadratic growth condition (1) with z \Gamma 0 feasible (x; . The feasible set is described in Figure 5. In the lateral view, the quadratic growth at (1; 1; 0) is fairly obvious from the curvature of the ridges that appear at the intersection of two constraints. From the shape of the feasible set it is also clear that (1; 1; 0) is the unique stationary point of the NLP. Among the solvers we used, MINOS [17] and SNOPT [11] use quasi-Newton methods that do not require second-order derivatives of the constraints. They Degenerate Nonlinear Programming with a Quadratic Growth Condition 19 Feasible set: view from above. Center: (1,1,0) lateral view. Fig. 1. Feasible set of the nonlinear program (89) (1,1,0) is the local minimum satisfying the quadratic growth condition (1). The jagged edges in the lateral view are a meshing effect. also use an augmented Lagrangian as a merit function. DONLP2 [23] solves a linear system instead of a Quadratic Program at each iteration and uses an L 1 penalty function. LANCELOT [5] uses an augmented Lagrangian technique in conjunction with a trust-region. FilterSQP [8] also uses a trust region approach but with a special classification of the relative merits of the iterates instead of Nonlinear solver jjx final \Gamma x jj 2 Iterations Message at termination FilterSQP 5.26e-09 28 Convergence LANCELOT 8.65e-07 336 Step size too small LINF 1.05e-08 28 Step size too small LOQO 1.60e-07 200 Iteration limit LOQO 5.50e-07 1000 Iteration limit MINOS 4.76e-06 27 Current point cannot be improved SNOPT 3.37e-07 3 Optimal Solution Found Table 3. Runs with various nonlinear solvers on the problem (89) a penalty or merit function. LOQO [24] is an interior-point approach. Finally, LINF is an ad hoc Matlab implementation of the L1 exact penalty function described in the preceding section, with an Armijo rule. The latter algorithm is started at (0; 0; 0). All runs, except for the L1 penalty and FilterSQP algorithms, were done on the NEOS server [18], where additional documentation can be found for all of the above solvers. For such a small example the time of execution is not relevant in comparing the behavior of the solvers. Since the solution of the problem is known, we chose as a criteria for comparison the best achievable solution. We set all relevant tolerances to 1e \Gamma 16, via the AMPL interface of NEOS. Smaller tolerances may interfere with the machine precision, though most of the solvers gave comparable answers even when the tolerances are set to 1e \Gamma 20. Larger tolerances (1e \Gamma 12- resulted in very similar results. Whenever allowed, we also changed other limiting parameters until an intrinsic stopping decision was issued. The only exception was DONLP2 which converged to all digits in the mantissa with the default settings. Table 1 shows the ratios OE(x k )\GammaOE(x ) at various iterations for our implementation LINF. All are close to 4:00, which is consistent with the Q-linear convergence claim for OE(x). Table 2 shows that LANCELOT decreases succesively the value of the penalty parameter (by 16 orders of magnitude), until it stops with the message 'Step size too small'. This was indeed one of the alternatives allowed by our analysis in Subsection 3.1 0). This is an undesirable outcome since the subproblems may become harder to solve. The results for all runs are illustrated in Table 3. It can be seen that the solvers that use augmented Lagrangians MINOS, SNOPT, LANCELOT exhibit an error of at least one order of magnitude larger compared to all other algo- rithms. However, one would expect that SNOPT and MINOS would have had at least as good a behavior as LINF if they would use a different merit function, since the nature of the QP solved is very similar to (14). Increasing the iteration limit in LOQO did not result in a better outcome. It is interesting to note that the outcome in FilterSQP and LINF differ by only a factor of 2 in the same number of iterations, though FilterSQP uses second-order information whereas LINF does not. Both LINF and FilterSQP solve quadratic programs at each Degenerate Nonlinear Programming with a Quadratic Growth Condition 21 iteration. DONLP2 has a remarkable behavior, though further investigation is necessary to determine whether this has some general implications. It is impossible to draw a general conclusion from one example. However, there seems to be an adverse bias for methods using augmented Lagrangians on degenerate NLPs as the one above. We are not advocating the use of LINF on general NLP, since its similarity to steepest descent makes it very sensitive to ill-conditioning. But the fact that it gives an outcome comparable to the one of solvers using second-order information shows that, for better results, a different way of incorporating second-order derivatives may be necessary. 6. Conclusions In this work we analyze the behavior of nonlinear programs in presence of constraint degeneracy: linear dependence of the gradients of the active constrains. The problems of interest exhibit minima with a quadratic growth property that satisfy the Mangasarian-Fromowitz constraint qualification. The novelty of our approach is that, while studying the SQP convergence properties, we do not assume the positive semidefiniteness of the Hessian of the Lagrangian on the critical cone for any of the feasible Lagrange multipliers. Our conditions are equivalent to a weak second-order sufficient condition [15,22]. We prove that, under these assumptions, if the data of the problem are twice continuously differentiable, the target minimum will be an isolated stationary point of the NLP. We also show that, when started sufficiently close to the minimum, the L1 exact penalty SQPs induce Q-linear convergence of the values of the penalized objective R-linear convergence of the iterates. This shows that such methods are robust with respect to constraint degeneracy. We give an example of a nonlinear program with a unique minimum that satisfies our conditions for which the Hessian of the Lagrangian is not positive semidefinite on the critical cone for any feasible choice of the multipliers. The direct consequence of this fact is that there is no augmented Lagrangian that will be positive semidefinite at the solution. Therefore, Lagrange multipliers algorithms will have to drive the penalty parameter to zero for such examples unless the trust region is active even at convergence. We provide our computational experience with this small nonlinear program. As a criteria for comparison we used the best achievable solution, which was obtained after tuning the parameters of the algorithms. We observed that, for this example, algorithms that use augmented Lagrangians resulted in errors of one order of magnitude or larger when compared to the other approaches. The Lagrange multiplier package that we used (LANCELOT [5]), was confined to decrease substantially the value of the penalty parameter (16 orders of magnitude), which is one of the outcomes allowed by our analysis. The linear convergence results concerning the L1 penalty function were also validated by our experiments 22 Mihai Anitescu Undoubtedly, such a small experiment is insufficient to draw any conclusions, especially about the approaches for which we have no theory under these assump- tions, such as interior-point algorithms. However, both from our theory and our experiments, it does appear that methods that use augmented Lagrangians are less robust with respect to constraint degeneracy when compared to SQP. We believe that attempting to develop a convergence theory in absence of the usual second-order conditions is interesting because it may result in algorithms that are more robust by virtue of the fact that their properties depend on fewer assumptions. However, how to improve on the current results, and especially how to define reliable variants of the Newton method (if possible) for this case, is a subject of future research. Acknowledgments Thanks to Stephen Wright, Jorge Mor'e and Danny Ralph for the many discussions on the subject. David Gay, Sven Leyffer and Chi-Jen Lin have kindly provided me information and support for the numerical examples. --R "On the rate of convergence of sequential quadratic programming with non-differentiable exact penalty function in the presence of constraint degeneracy" New York "Second-order sufficiency and quadratic growth for nonisolated minima" LANCELOT: A Fortran Package for Large-Scale Nonlinear Optimization Introduction to Sensitivity and Stability Analysis in Practical Methods of Optimization "Nonlinear programming without a penalty function" "A necessary and sufficient regularity condition to have bounded multipliers in nonconvex programming" "Differential stability in nonlinear programming" "User's guide for SNOPT 5.3: A Fortran package for large-scale nonlinear programming" "The relaxation method for solving systems of linear inequalities" "Stabilized sequential quadratic programming" "Stability in the presence of degeneracy and error estima- tion" "Necessary and sufficient conditions for a local minimum.3: Second order conditions and augmented duality" "On Sensitivity analysis of nonlinear programs in Banach spaces: the approach via composite unconstrained optimization" The NEOS Guide. Iterative Solutions of Nonlinear Equations in Several Vari- ables "Superlinear convergence of an interior-point method despite dependent constraints" "Generalized equations and their solutions, Part II: Applications to non-linear programming" "Sensitivity analysis of nonlinear programs and differentiability properties of metric projections" "An SQP method for general nonlinear programs using only equality constrained subproblems" "An interior-point code for quadratic programming" "Superlinear convergence of a stabilized SQP method to a degenerate so- lution" --TR --CTR Jin-Bao Jian, A Superlinearly Convergent Implicit Smooth SQP Algorithm for Mathematical Programs with Nonlinear Complementarity Constraints, Computational Optimization and Applications, v.31 n.3, p.335-361, July 2005
degeneracy;sequential quadratic programming;nonlinear programming;quadratic growth
589079
Bounds for Linear Matrix Inequalities.
For iterative sequences that converge to the solution set of a linear matrix inequality, we show that the distance of the iterates to the solution set is at most \( O(\epsilon ^{2^{-d}}) \). The nonnegative integer d is the so-called degree of singularity of the linear matrix inequality, and $\epsilon $ denotes the amount of constraint violation in the iterate. For infeasible linear matrix inequalities, we show that the minimal norm of $\epsilon $-approximate primal solutions is at least \( 1/O(\epsilon ^{1/(2^{d}-1)}) \), and the minimal norm of $\epsilon $-approximate Farkas-type dual solutions is at most \( O(1/ \epsilon ^{2^{d}-1}) \). As an application of these error bounds, we show that for any bounded sequence of $\epsilon $-approximate solutions to a semidefinite programming problem, the distance to the optimal solution set is at most \( O(\epsilon ^{2^{-k}}) \), where k is the degree of singularity of the optimal solution set.
Introduction Linear matrix inequalities play an important role in system and control theory, see the book by Boyd et al. [3]. Recently, considerable progress has been made in optimization over linear matrix inequalities, i.e. semi-definite programming, see [1, 6, 8, 9, 16, 19, 18, 23, 25] and the references cited therein. We study the linear matrix inequality (LMI) ae means positive semi-definiteness, B is a given (real) symmetric matrix and A is a linear subspace of symmetric matrices. The LMI (1) is in conic form, see e.g. [17, 23]. Since we leave complete freedom as to the formulation of A, it is in general not difficult to fit a given LMI into conic form. Consider for instance a linear matrix inequality are given symmetric matrices. This is a conic form LMI and A is the span of fF g. Recently developed interior point codes for semi-definite programming make it possible to solve LMIs numerically. Such algorithms generate sequences of increasingly good approximate solutions, provided that the LMI is solvable. For a discussion of interior point methods for semi-definite programming, see e.g. [8, 23]. A typical way to measure the quality of an approximate solution, is by evaluating its constraint violation. For instance, if we denote the smallest eigenvalue of an approximate solution ~ X), then we may say that ~ X violates the constraint 'X - 0' by an amount of [\Gamma- min ( ~ X)]+ , where the operator [\Delta] + yields the positive part. In fact, X)]+ is the distance, measured in the matrix 2-norm, of the approximate solution ~ X to the cone of positive semi-definite matrices. The matrix 2-norm is a convenient measure for the amount by which the positive semi-definiteness constraint is violated, but other matrix norms can in principle be used as well. Similarly, we say that ~ X violates the constraint 'X 2 B +A' by an amount of dist( ~ denotes the distance function (for a given norm). The total amount of constraint violation in ~ X, i.e. is called the backward error of ~ X with respect to the LMI (1). The backward error indicates how much we should perturb the data of the problem, such that ~ X is an exact solution to the perturbed problem. However, the backward error does not (immediately) tell us the distance from ~ X to the solution set of the original LMI; this distance is called the forward error of ~ X . knowing any exact solution, there is no straightforward way to estimate the forward error. For linear inequality and equation systems however, the forward error and backward error are of the same order of magnitude, see Hoffman [7]. The equivalence between forward and backward errors holds also true for systems that are described by convex quadratic inequalities, if a Slater condition holds, see Luo and Luo [12]. In these cases, we have a relation of the which is called a Lipschitzian error bound. For systems of convex quadratic inequalities without Slater's condition, an error bound of the form was obtained by Wang and Pang [26]. They also showed that d - where n is the dimension of the problem. Error bounds for systems with a nonconvex quadratic inequality are given in Luo and Sturm [14], and references cited therein. An error bound of the form (3) is called a H-olderian error bound. A H-olderian error bound has been demonstrated for analytic inequality and equation systems, if the size of the approximate solutions is bounded by a fixed constant, see Luo and Pang [13]. However, there are no known positive lower bounds on the exponent fl, except in the linear and quadratic cases that are mentioned above, or when a Slater condition holds [4], For a comprehensive survey of error bounds, we refer to Pang [20]. Some issues on error bounds for LMIs and semi-definite programming were recently addressed by Deng and Hu [4], Goldfarb and Scheinberg [5], Luo, Sturm and Zhang [16] and Sturm and Zhang [24]. Deng and Hu [4] derived upper bounds on the Lipschitz constant (or condition number) for LMIs, if Slater's condition holds. Luo Sturm and Zhang [16] and Sturm and Zhang [24] prove some Lipschitzian type error bounds for central solutions for semi-definite programs under strict complementarity. Goldfarb and Scheinberg [5] prove Lipschitz continuity of the optimal value function for semi-definite programs. In this paper, we show for LMIs in n \Theta n matrices, that (3) holds for a certain the so-called degree of singularity, provided that the size of the approximate solutions is bounded. We interpret the degree of singularity in the context of Ramana-type regularized duality. It is basically the number of elementary regularizations that are needed to obtain a fully regularized dual. Under Slater's constraint qualification, the irregularity level d is zero. (Notice that this is also true for convex quadratic systems, see Wang and Pang [26].) The degree of singularity of the optimal solution set of a semi-definite programming problem is at most one, if strict complementarity holds. The concept of singularity degrees thus embeds the Slater and strict complementarity conditions in a hierarchy of singularity for LMIs. This paper is organized as follows. In Section 2, we introduce the concept of regularized backward errors, which is closely related to the concept of minimal cones [2]. In this section, we also show that there is a close connection between the regularized backward error and the forward error. We will then estimate in Section 3 how the regularized backward error depends on the usual backward error. In Section 4, we apply the error bound for LMIs to semi-definite programming problems. The paper is concluded in Section 5. Notation. Let S n\Thetan denote the space of n \Theta n real symmetric matrices. The cone of all positive semi-definite matrices in S n\Thetan is denoted by S n\Thetan we . The interior of S n\Thetan is the set of positive definite matrices S n\Thetan ++ , and we write X - 0 if and only if X 2 S n\Thetan ++ . We let N := n(n + 1)=2 denote the dimension of the real linear space S n\Thetan . The standard inner product for two symmetric matrices X and Y is tr XY . The matrix norm kXk 2 is the operator 2-norm that is associated with the Euclidean norm for vectors, namely For symmetric matrices, kXk 2 is the eigenvalue of X that has the largest absolute value. 2 The regularized backward error A denote the smallest linear subspace containing B +A, i.e. We are naturally interested in the intersection of this linear subspace with the cone of positive semi-definite matrices. It holds that A " S n\Thetan the above characterization is a special case of a duality theorem for convex cones. The general theorem states that given a linear subspace L and a convex cone !+ g, it holds that see Corollary 2 in Luo, Sturm and Zhang [15] and Corollary 2.2 in [23]. This result generalizes a classical duality theorem of Gordon and Stiemke for linear inequalities. To see why (5) is a special case of (6), we must interpret S n\Thetan as a convex cone in ! N . This can be established by choosing an orthonormal basis of S n\Thetan , say an orthonormal set of symmetric matrices fS[1]; 1)=2 is the dimension of S n\Thetan . We can then associate with any matrix n\Thetan a coordinate vector x 2 ! N into this basis, and vice versa. Namely, we let x Due to the orthonormality of the basis, we have y, for all matrices X;Y 2 S n\Thetan with coordinate vectors x; y 2 ! N . As a convention, we use upper-case symbols, like X and B, for symmetric matrices, and we implicitly define the corresponding lower-case symbols, like x and b, to be the associated coordinate vectors, as described above. Furthermore, we use calligraphic letters, such as S n\Thetan , to denote sets. With the established one-to-one correspondence between S n\Thetan and ! N in mind, we do not only use S n\Thetan for the set of positive semi-definite matrices in S n\Thetan , but also for the set of coordinate vectors of positive semi-definite matrices, which is a convex cone in the Euclidean space ! N . We will also use such a convention for other sets of symmetric matrices. In particular, we reformulate (4) as where Img b ae ! N is the line of all multiples of b. The orthogonal complement of - A is The all-zero matrix is obviously the only matrix that is both positive and negative semi-definite, i.e. S n\Thetan " \GammaS n\Thetan f0g. Also, the cone of positive semi-definite matrices is self-dual, i.e. (S n\Thetan . Thus, taking and A ? in (6) yields (5). Relation (5) states that if - A and S n\Thetan intersect only at the origin, then there exists a positive definite matrix Z 2 - A ? . Consider now a sequence of increasingly accurate solutions fX(ffl) notice that the parameter ffl measures the backward error in X(ffl). It follows that since Z?(B + A), we must have jZ ffl O(ffl). Using the fact that positive definite, this implies that O(ffl). The above reasoning establishes the relation A " S n\Thetan which is an error bound for the case that - A intersects the semi-definite cone only at the origin. Assume now that - A " S n\Thetan A " S n\Thetan applying a basis transformation if necessary, we may assume without loss of generality that we can partition X as Using this notation, we can partition an arbitrary matrix X 2 S n\Thetan as U XN A " S n\Thetan suppose without loss of generality that X is of the form (9). Then it holds for all A " S n\Thetan and Proof. Suppose to the contrary that XN is not the all-zero matrix, and let \Theta 0 y T be such that XN yN 6= 0. Then for any ff 2 !, where we used the fact that X is positive semi-definite. Consequently, we have for all ff ? 0 that A " S n\Thetan which contradicts the fact that by definition, X is in the relative interior of A " S n\Thetan . We have now shown by contradiction that positive semi-definite, it follows that also A face of S n\Thetan is by definition a cone of the form n\Thetan where Z is a given positive semi-definite matrix. In particular, if then n\Thetan ae oe and X is in the relative interior of face(S n\Thetan 0). The facial structure of S n\Thetan has been studied in detail by Lewis [11] and Pataki [21]. A " S n\Thetan suppose without loss of generality that X is of the form (9). Then relint S n\Thetan I Proof. The lemma holds trivially true if (B +A) " S n\Thetan now that there exists A, there exists t 2 ! such that X \Gamma tB 2 A. However, for all ff ? 0 satisfying fft ? \Gamma1, we S n\Thetan I where we used Lemma 1. This shows that S n\Thetan I Using Lemma 1 once again, the lemma follows from the above relation. Q.E.D. Due to the above result, the face face S n\Thetan I is sometimes called the minimal cone [2] or the regularized semi-definite cone [15] for the affine space B +A. The backward error of X(ffl) with respect to the regularized system is naturally defined as The following lemma states, among others, that if fX(ffl) then the regularized backward error is of the same order as the forward error A " S n\Thetan suppose without loss of generality that X is of the form (9). If fX(ffl) is such that for all ffl ? 0, then (B+A)"S n\Thetan ;. Moreover, there exists such that Proof. As is well known, the backward and forward error for a system of linear equations are of the same order [7]. Therefore, the relations imply that This bound implies the existence of fY (ffl) 0g, such that Using also the fact that X B is positive definite, it follows that with Notice that A, there must exist t 2 ! such that be such that and hence dist Under Slater's condition, i.e. if (B Hoffman's error bound [7] for systems of linear inequalities and equations to LMIs. Notice in particular that no boundedness assumptions are made, i.e. the error bound holds globally over S n\Thetan . However, the lemma requires a scaling which is not needed in case of linear inequalities and equations. The following example shows that this scaling factor is essential in the case of LMIs. Example 1 Consider the LMI in S 2\Theta2 with ae oe i.e. we want to find find x 11 and x 12 such that x 11 - jx 12 j 2 . This LMI obviously has positive definite solutions (the identity matrix for instance). Therefore, the regularized backward error is identical to the usual backward error. The approximate solution has backward error ffl ? 0. However, X(ffl) if and only if y 22 which shows that the distance of X(ffl) to (B is bounded from below by a positive constant as ffl # 0. However, we have X(ffl)=(1+ ffl) 2 (B+A)"S 2\Theta2 which agrees with the statement of Lemma 3. Below are more remarks on the regularized error bound of Lemma 3. states that the mere existence of fX(ffl) (12) for all ffl ? 0 implies that (B even though X(ffl) is not necessarily bounded for ffl # 0. In the case of weak infeasibility, i.e. if dist(B +A;S n\Thetan we can therefore conclude that if X(ffl) satisfies (7) then lim inf ffl#0 is a bounded sequence with then also kX (k) as follows from Lemma 1. Letting it follows from Lemma 3 and the boundedness of the sequence fX 3 Regularization steps In order to bound the regularized backward error (11) in terms of the original backward error (2), we use a sequence of regularization steps. In the preceding, we have partitioned n \Theta n matrices according to the structure of X , given by (9). In this section, we will also partition n \Theta n matrices into blocks, but with respect to a possibly different eigenvector basis; the sizes of the blocks can be different as well. We will denote the blocks by the subscripts We will also encounter the dual cone of a face of S n\Thetan face S n\Thetan I ae - oe Obviously, we have relint face S n\Thetan I ae - oe In the following, we will allow the possibility that are X 22 are non-existent. For this case, we use the convention that kX 12 A be a linear subspace of S n\Thetan , and suppose that fX(ffl) is such that for all S n\Thetan I It holds that ffl Z 11 - 0 if and only if S n\Thetan I only if S n\Thetan I ffl For the remaining case that 0 6= Z 11 6- 0, let \Theta be an orthogonal matrix such that Z 11 Proof. The first two cases, i.e. Z are immediate applications of (6). It remains to consider the case that Z 11 is a nonzero but singular, positive semi-definite matrix. - ffl, there must exist Y (ffl), such that for all ffl ? 0. This implies that Z?(X(ffl)+Y (ffl)) because Z 2 - A ? , and therefore Z '- where we used the Cauchy-Schwartz inequality. Recall now that so that we further obtain Z Since Z 11 is positive semi-definite and - min (X(ffl)) - \Gammaffl, we have where we used Z 11 in the first identity, and (15) in the last identity. Recalling that Q T easily follows from the above relation that Finally, since - min (X(ffl)) - \Gammaffl, we know that X 11 (ffl) + fflI is positive semi- definite, and hence where we used (16). This completes the proof. Q.E.D. For a given linear subspace - A ' S n\Thetan , we define the level of singularity by recursively applying the construction of Lemma 4. This procedure is outlined below: Procedure 1 Definition of the level of singularity of a linear subspace - S n\Thetan . Otherwise, proceed with Step 2. be such that Z (0) ae \Theta A oe A ? S n\Thetan I If Z (d) proceed with Step 4. be such that Z (d) I Let ae \Theta - A d oe return to Step 3. In the above procedure, we start with the full dimensional cone S n\Thetan , and in the first iteration we determine a face of this cone. Next, we arrive at a face of this face, and so on. We claim that this procedure finally arrives at the minimal cone. To see this, notice that at any given step above, we perform a regularization step as described in Lemma 4. Recall from (5) that A " S n\Thetan f0g, and this case has already been treated in Section 2. In any other case, we have Z (d( - It is easily seen from Lemma 4 that if X 2 - A " S n\Thetan This means that all nonzeros of X must be contained in the (final) 11 block for - A . On the other hand, since Z (d( - in the above procedure, it follows from (6) that there exists ~ A " S n\Thetan such that ~ we just showed that X A " S n\Thetan , we must have ~ A " S n\Thetan Hence, the face in the final iteration is the minimal cone. For - (9). By applying a basis transformation if necessary, we may assume without loss of generality that there is a (d( - partition, such that 0: Above, we used a Matlab-type 1 notation, thus means denotes the block on the ith row and jth column in the (d( - is symmetric, we only specified the upper block triangular part of Z. The relation between the (d( - partition in (18) and the 2 \Theta 2 partition in iteration is that The minimal cone is the set of matrices X for which In iteration of Procedure 1, we arrive at the face where which indeed includes the minimal cone. Remark that the 3rd row and column in the 3 \Theta 3 block form of (18) are non-existent for A), i.e. for Based on Lemma 4, we can now estimate the regularized backward error. A " S n\Thetan loss of generality that X is of the form (9). If is such that for all then with is the degree of singularity of - A. Proof. Applying Lemma 4 in iteration of Procedure 1, we have that 1 MATLAB is a registered trademark of The MathWorks, Inc. where we used X ffl as a synonym for X(ffl). Suppose now that in iteration d 2 where It then follows from Lemma 4 that (19) also holds for By induction. we obtain that (19) holds for (ffl), the lemma follows. Q.E.D. We arrive now at the main result of this paper, namely an error bound for LMIs. Theorem 1 Let - is such that kX(ffl)k is bounded and then Proof. For the case that d( - the theorem follows by combining Lemma 3 with Lemma 5. If there are two cases, either - A " S n\Thetan A " S n\Thetan In the former case, we have hence the error bound holds, see Section 2. In the latter case, we have that X the error bound follows from Lemma 3. Q.E.D. An LMI is said to be weakly infeasible if 1. there is no solution to the LMI, i.e. (B +A) " S n\Thetan 2. dist(B +A;S n\Thetan For weakly infeasible LMIs, there exist approximate solutions with arbitrarily small constraint violations. However, the following theorem provides a lower bound on the size of such approximate solutions to weakly infeasible LMIs. Theorem 2 Let - and suppose that is such that small enough, we have X(ffl) 6= 0 Proof. Suppose to the contrary that there exists a sequence ffl 1 Applying Lemma 5, it follows that Together with Lemma 3, we obtain that (B +A) " S n\Thetan Q.E.D. There is an extension of Farkas' lemma from linear inequalities to convex cones, which states that where K ae S n\Thetan is a convex cone, and K is the associated dual cone. See e.g. Lemma 2.5 in [23]. If dist(B +A;S n\Thetan we say that the LMI is strongly infeasible. Relation (20) states that strong infeasibility can be demonstrated by a matrix Z 2 A ? " S n\Thetan 0, and such Z is called a dual improving direction. For weakly infeasible LMIs, infeasibility cannot be demonstrated by a dual improving direction. However, an LMI is infeasible if and only if there exist approximate dual improving directions with arbitrarily small constraint viola- tions. See e.g. Lemma 2.6 in [23]. The next theorem gives an upper bound for the minimal norm of such approximate dual improving directions in the case of infeasibility. Theorem 3 Let - there exist 0g such that for all holds that and Proof. Let X 2 relint ( - A " S n\Thetan suppose without loss of generality that X is of the form (9). Using the same 2 \Theta 2 partition as in (9), it follows from Lemma 3 that dist S n\Thetan I -" 0: Applying (20), it thus follows that there exists a matrix Y (0) such that S n\Thetan I Partitioning Y (0) , we have Y (0) U We shall now construct fY A)g such that Y 1. Remark from (21)-(22) that (23) holds for construct Y (k) for k 2 in such a way that it satisfies (23), provided that Y (k\Gamma1) satisfies (23). We can then use induction. Let immediately obtain from (23) that irrespective of t. Furthermore, since Y positive semi-definite if and only if the Schur-complement is positive semi-definite. From (18) and the definition of Y t , we have and hence positive semi-definite if we choose t as where we used that kY (k\Gamma1) Setting Y The theorem follows by letting We remark from the proof of Theorem 3 that the matrices Y (0) and Z (k) , finite certificate of the infeasibility of the LMI. Together, these matrices form essentially a solution to the regularized Farkas- type dual of Ramana [22], see also [10, 15]. Thus, the degree of singularity is the minimal number of layers that are needed in the perfect dual of Ramana. As discussed in the introduction, it is easy to calculate the backward error of an approximate solution. However, the error bound for the forward error of an LMI, as given in Theorem 1, does not only involve the backward error, but also the degree of singularity. We will now provide some easily computable upper bounds on the degree of singularity. Lemma 6 For the degree of singularity of a linear subspace - A ' S n\Thetan , it holds that A ? g: Proof. If A " S n\Thetan by definition of A). For this case, we have defined the (d( - partition (18), where each of the diagonal blocks is at least of size 1 \Theta 1. Thus, Furthermore, Lemma 4 defines a matrix Z A ? , for each regularization and it is easily verified that these matrices are mutually independent. Therefore, Finally, using the (d( - partition (18), we claim that there exists X A with Namely, if such X (k) does not exist, then by (6), there must exist \DeltaZ 2 - A ? such that and this contradicts the fact that Z (d( - of maximal rank, see its definition in Lemma 4. Again, it is easy to see that the matrices A), are mutually independent, and hence A. Q.E.D. The bounds of Theorem 1 and Theorem 2 quickly become inattractive as the singularity degree increases. However, the next two examples show that these bounds can be tight. This means that problems with a large degree of singularity can be very hard to solve numerically. Example 2 Consider the LMI Due to the restriction 'X 22 = 0' and the positive semi-definiteness, we have which further implies and inductive argument, we have X we can construct a sequence fX(ffl) j ffl ? 0g with a constraint violation ffl, but , viz. Notice that 'X 22 = 0' is the only constraint that is violated by X(ffl). To see how unfortunate this example is, consider a backward error Then, already for any solution - X of the LMI. Example 3 Extending Example 2 with the restriction 'X we obtain a (weakly) infeasible LMI: However, we may construct a sequence fX(ffl) violation ffl and Namely, we let This example shows that (in)feasibility can be hard to detect. Namely, for which is not unusually large; yet, the problem is infeasible. 4 Application to semi-definite programming bounds for LMIs can be applied to semi-definite optimization models as well. A standard form semi-definite program is where B and C are given symmetric matrices. Associated with this optimization problem is a dual problem, viz. (D) An obvious property of the primal-dual pair (P) and (D) is the weak duality relation. Namely, if X 2 (B +A) " S n\Thetan Clearly, if X ffl must be optimal solutions to (P) and (D) respectively; we say then that (X; Z) is a pair of complementary solutions. In general, such a pair may not exist, even if both (P) and (D) are feasible. (We say that (P) is feasible if (B+A)"S n\Thetan and (D) is feasible if (C+A ? )"S n\Thetan ;.) A sufficient condition for the existence of a complementary solution pair is that (P) and (D) are feasible and satisfy the primal-dual Slater condition, in which case Based on (25), we can formulate the set of complementary solutions as the In principle, we can apply our error bound results for LMIs directly to the above system. But, tighter bounds can be obtained by exploring its special structure. Consider a bounded trajectory of approximate primal and dual solutions C) be a complementary solution pair, i.e. Such a pair must exist, since in particular any cluster point of f(X(ffl); Z(ffl) j ffl ? 0g for ffl # 0 is a complementary solution pair. Notice that B and similarly C +A , from which we easily derive that for feasible solutions X and Z, and for (X(ffl); Z(ffl)) satisfying (26). This means that X(ffl) has an O(ffl) constraint violation with respect to the LMI Notice that (27) describes the set of optimal solutions to (P). Letting A := Img the Theorems 1 and 2 are applicable to the LMI (27) and hence to the semi-definite program (P). Specifically, given a bounded trajectory fX(ffl); Z(ffl) 0g satisfying (26), we know that the distance from X(ffl) to the set of optimal solutions to (P) is O(ffl 2 \Gammad( - is the degree of singularity of the linear subspace defined in (28). 0, we can move the parentheses in definition (28) to get from which we get Noticing the primal-dual symmetry, we conclude that the distance from Z(ffl) to the set of optimal solutions to (D) is O(ffl 2 \Gammad( - A ? ) is the degree of singularity of - A ? . Concluding remarks Theorem 1 provides a H-olderian error bound for LMIs. For weakly infeasible LMIs, we have derived relations between backward errors and the size of approximate solutions, see Theorems 2 and 3. In Section 4, we applied the error bound of Theorem 1 to semi-definite programming problems (SDPs). If the SDP has a strictly complementary solution, then its degree of singularity can be at most 1, and the bound becomes backward error): For this case, Luo, Sturm and Zhang [16] obtained a Lipschitzian error bound if the approximate solutions (X(ffl); Z(ffl)) are restricted to the central path. The sensitivity of central solutions with respect to perturbations in the right-hand side was studied by Sturm and Zhang [24]. Acknowledgment . Tom Luo's comments on an earlier version of this paper have resulted in substantial improvements in the presentation. --R Regularizing the abstract convex program. Linear matrix inequalities in system and control theory Computable error bounds for semidefinite program- ming On parametric semidefinite programming. An interior-point method for semidefinite programming On approximate solutions of systems of linear inequalities. Interior Point Methods for Semidefinite Programming. Perfect duality in semi-infinite and semidefinite programming Extensions of Hoffman's error bound to polynomial systems. bounds for analytic systems and their applications. bounds for quadratic systems. Duality results for conic convex programming. Superlinear convergence of a symmetric primal-dual path following algorithm for semidefinite programming Interior point polynomial methods in convex programming bounds in mathematical programming. On the rank of extreme matrices in semidefinite programs and the multiplicity of optimal eigenvalues. An exact duality theory for semidefinite programming and its complexity implications. On sensitivity of central solutions in semidefinite programming. programming. Global error bounds for convex quadratic inequality systems. --TR --CTR Dominique Az , Jean-Baptiste Hiriart-Urruty, Optimal Hoffman-Type Estimates in Eigenvalue and Semidefinite Inequality Constraints, Journal of Global Optimization, v.24 n.2, p.133-147, October 2002
regularized duality;error bounds;linear matrix inequality;semidefinite programming
589089
A Feasible BFGS Interior Point Algorithm for Solving Convex Minimization Problems.
We propose a BFGS primal-dual interior point method for minimizing a convex function on a convex set defined by equality and inequality constraints. The algorithm generates feasible iterates and consists in computing approximate solutions of the optimality conditions perturbed by a sequence of positive parameters $\mu$ converging to zero. We prove that it converges q-superlinearly for each fixed $\mu$. We also show that it is globally convergent to the analytic center of the primal-dual optimal set when $\mu$ tends to 0 and strict complementarity holds.
Introduction . We consider the problem of minimizing a smooth convex function on a convex set defined by inequality constraints. The problem is written as # R is the function to minimize and c(x) # 0 means that each component m) of c must be nonnegative at the solution. To simplify the presentation and to avoid complicated notation, the case when linear equality constraints are present is discussed at the end of the paper. Since we assume that the components of c are concave, the feasible set of this problem is convex. The algorithm proposed in this paper and the convergence analysis require that f and c are di#erentiable and that at least one of the functions f , -c (1) , . , -c (m) is strongly convex. The reason for this latter hypothesis will be clarified below. Since the algorithm belongs to the class of interior point (IP) methods, it may be well suited for problems with many inequality constraints. It is also more e#cient when the number of variables remains small or medium, say, fewer than 500, because it updates n - n matrices by a quasi-Newton (qN) formula. For problems with more variables, limited memory BFGS updates [39] can be used, but we will not consider this issue in this paper. Our motivation is based on practical considerations. During the last 15 years much progress has been realized on IP methods for solving linear or convex minimization problems (see the monographs [29, 10, 38, 44, 23, 42, 47, 49]). For nonlinear convex problems, these algorithms assume that the second derivatives of the functions used to define the problem are available (see [43, 35, 36, 12, 38, 26]). In practice, how- # Received by the editors September 15, 1998; accepted for publication (in revised form) January 26, 2000; published electronically August 3, 2000. http://www.siam.org/journals/siopt/11-1/34472.html des Sciences, 123 av. A. Thomas, 87060 Limoges Cedex, France (Paul.Armand@ unilim.fr). # INRIA Rocquencourt, BP 105, 78153 Le Chesnay Cedex, France (Jean-Charles.Gilbert@inria.fr). - MIP, UFR MIG, Universit-e Paul Sabatier, 118 route de Narbonne, 31062 Toulouse Cedex 4, France (jan@mip.ups-tlse.fr). 200 P. ARMAND, J. CH. GILBERT, AND S. JAN-J - EGOU ever, it is not uncommon to find situations where this requirement cannot be satis- fied, in particular for large scale engineering problems (see [27] for an example, which partly motivates this study and deals with the estimation of parameters in a three phase flow in a porous medium). Despite the possible use of computational di#eren- tiation techniques [8, 19, 3, 28], the computing time needed to evaluate Hessians or Hessian-vector products may be so large that IP algorithms using second derivatives may be unattractive. This situation is familiar in unconstrained optimization. In that case, qN tech- niques, which use first derivatives only, have proved to be e#cient, even when there are millions of variables (see [32, 20] and [9] for an example in meteorology). This fact motivates the present paper, in which we explore the possibility of combining the IP approach and qN techniques. Our ambition remains modest, however, since we confine ourselves to the question of whether the elegant BFGS theory for unconstrained convex optimization [41, 6] is still valid when inequality constraints are present. For the applications, it would be desirable to have a qN-IP algorithm in the case when f and -c are nonlinear and not necessarily convex. We postpone this more di#cult subject for future research (see [21, 48] for possible approaches). Provided the constraints satisfy some qualification assumptions, the Karush- Kuhn-Tucker (KKT) optimality conditions of problem (1.1) can be written (see [17], for example) as follows: there exists a vector of multipliers # R m such that where #f(x) is the gradient of f at x (for the Euclidean scalar product), #c(x) is a matrix whose columns are the gradients #c (i) (x), and is the diagonal matrix, whose diagonal elements are the components of c. The Lagrangian function associated with problem (1.1) is defined on R n Since f is convex and each component c (i) is concave, for any fixed # 0, #) is a convex function from R n to R. When f and c are twice di#erentiable, the gradient and Hessian of # with respect to x are given by Our primal-dual IP approach is rather standard (see [24, 36, 35, 11, 12, 1, 26, 25, 15, 7, 5]). It computes iteratively approximate solutions of the perturbed optimality system for a sequence of parameters - > 0 converging to zero. In (1.2), is the vector of all ones whose dimension will be clear from the context. The last inequality means that all the components of both c(x) and # must be positive. By perturbing the complementarity equation of the KKT conditions with the parameter A BFGS INTERIOR POINT ALGORITHM 201 -, the combinatorial aspect of the problem, inherent in the determination of the active constraints or the zero multipliers, is avoided. We use the word inner to qualify those iterations that are used to find an approximate solution of (1.2) for fixed -, while an outer iteration is the collection of inner iterations corresponding to the same value of -. The Newton step for solving the first two equations in (1.2) with fixed - is the solution of the linear system d #f(x) +#c(x)# in which xx #(x, #) and This direction is sometimes called the primal-dual step, since it is obtained by linearizing the primal-dual system (1.2), while the primal step is the Newton direction for minimizing in the primal variable x the barrier function log c (i) (x) associated with (1.1) (the algorithms in [16, 33, 4] are in this spirit). The two problems are related since, after elimination of #, (1.2) represents the optimality conditions of the unconstrained barrier problem c(x) > 0. As a result, an approximate solution of (1.2) is also an approximate minimizer of the barrier problem (1.4). However, algorithms using the primal-dual direction have been shown to present a better numerical e#ciency (see, for example, [46]). In our algorithm for solving (1.2) or (1.4) approximately, a search direction d is computed as a solution of (1.3) in which M is now a positive definite symmetric matrix approximating # 2 xx #(x, #) and updated by the BFGS formula (see [14, 17] for material on qN techniques). By eliminating d # from (1.3) we obtain Since the iterates will be forced to remain strictly feasible, i.e., (c(x), #) > 0, the positive definiteness of M implies that d x is a descent direction of # - at x. Therefore, to force convergence of the inner iterates, a possibility could be to force the decrease of # - at each iteration. However, since the algorithm also generates dual variables #, we prefer to add to # - the function (see [45, 1, 18]) log # (i) c (i) (x) # to control the change in #. This function is also used in [30, 31] as a potential function for nonlinear complementarity problems. Even though the map (x, # - V(x, #) is not necessarily convex, we will show that it has a unique minimizer, which is the solution of (1.2), and that it decreases along the direction this primal-dual merit function can be used to force the convergence of the pairs (x, #) to the solution of (1.2), using line-searches. It will be shown that the additional P. ARMAND, J. CH. GILBERT, AND S. JAN-J - EGOU function V does not prevent unit step-sizes from being accepted asymptotically, which is an important point for the e#ciency of the algorithm. Let us stress the fact that our algorithm is not a standard BFGS algorithm for solving the barrier problem (1.4), since it is the Hessian of the Lagrangian that is approximated by the updated matrix M , not the Hessian of # - . This is motivated by the following arguments. First, the di#erence between # 2 involves first derivatives only. Since these derivatives are considered to be available, they need not be approximated. Second, the Hessian # 2 xx #, which is approximated by M , is independent of - and does not become ill-conditioned as - goes to zero. Third, the approximation of # 2 # obtained at the end of an outer iteration can be used as the starting matrix for the next outer iteration. If this looks attractive, it has also the inconvenience of restricting the approach to (strongly) convex functions, as we now explain. After the computation of the new iterates the step-size given by the line-search), the matrix M is updated by the BFGS formula using two vectors # and #. Since we want the new matrix M+ to be an approximation of # 2 it satisfies the qN equation M+ # (a property of the BFGS formula), it makes sense to define # and # by The formula is well defined and generates stable positive definite matrices provided these vectors satisfy # > 0. This inequality, known as the curvature condition, expresses the strict monotonicity of the gradient of the Lagrangian between two successive iterates. In unconstrained optimization, it can always be satisfied by using the Wolfe line-search, provided the function to minimize is bounded below. If this is a reasonable assumption in unconstrained optimization, it is no longer the case when constraints are present, since the optimization problem may be perfectly well defined even when # is unbounded below. Now, assuming this hypothesis on the boundedness of # would have been less restrictive than assuming its strong convexity, but it is not satisfactory. Indeed, with a bounded below Lagrangian, the curvature condition can be satisfied by the Wolfe line-search as in unconstrained optimization, but near the solution the information on # 2 collected in the matrix M could come from a region far from the optimal point, which would prevent q-superlinear convergence of the it- erates. Because of this observation, we assume that f or one of the functions -c (i) is strongly convex, so that the Lagrangian becomes a strongly convex function of x for any fixed # > 0. With this assumption, the curvature condition will be satisfied independently of the kind of line-search techniques actually used in the algorithm. The question whether the present theory can be adapted to convex problems, hence including linear programming, is puzzling. We will come back to this issue in the discussion section. A large part of the paper is devoted to the analysis of the qN algorithm for solving the perturbed KKT conditions (1.2) with fixed -. The algorithm is detailed in the next section, while its convergence speed is analyzed in sections 3 and 4. In particular, it is shown that, for fixed - > 0, the primal-dual pairs (x, #) converge q-superlinearly toward a solution of (1.2). The tools used to prove convergence are essentially those of A BFGS INTERIOR POINT ALGORITHM 203 the BFGS theory [6, 13, 40]. In section 5, the overall algorithm is presented and it is shown that the sequence of outer iterates is globally convergent, in the sense that it is bounded and that its accumulation points are primal-dual solutions of problem (1.1). If, in addition, strict complementarity holds, the whole sequence of outer iterates converges to the analytic center of the primal-dual optimal set. 2. The algorithm for solving the barrier problem. The Euclidean or # 2 norm is denoted by #. We recall that a function # : R n # R is said to be strongly convex with modulus # > 0, if for all (x, y) # R n equivalent definitions, see, for example, [22, Chapter IV]). Our minimal assumptions are the following. Assumption 2.1. (i) The functions f and -c (i) (1 # i # m) are convex and di#erentiable from R n to R and at least one of the functions f , -c (1) , . , -c (m) is strongly convex. (ii) The set of strictly feasible points for problem (1.1) is nonempty, i.e., there exists x # R n such that c(x) > 0. Assumption 2.1(i) was motivated in section 1. Assumption 2.1(ii), also called the (strong) Slater condition, is necessary for the well-posedness of a feasible interior point method. With the convexity assumption, it is equivalent to the fact that the set of multipliers associated with a given solution is nonempty and compact (see [22, Theorem VII.2.3.2], for example). These assumptions have the following clear consequence. Lemma 2.2. Suppose that Assumption 2.1 holds. Then, the solution set of problem (1.1) is nonempty and bounded. By Lemma 2.2, the level sets of the logarithmic barrier function # - are compact, a fact that will be used frequently. It is a consequence of [16, Lemma 12], which we recall for completeness. Lemma 2.3. Let f : R n # R be a convex continuous function and c : R n be a continuous function having concave components. Suppose that the set {x # R c(x) > 0} is nonempty and that the solution set of problem (1.1) is nonempty and bounded. Then, for any # R and - > 0, the set log c (i) (x) # is compact (and possibly empty). Let x 1 be the first iterate of our feasible IP algorithm, hence satisfying c(x 1 ) > 0, and define the level set Lemma 2.4. Suppose that Assumption 2.1 holds. Then, the barrier problem (1.4) has a unique solution, which is denoted by - x - . Proof. By Assumption 2.1, Lemma 2.2, and Lemma 2.3, L P 1 is nonempty and compact, so that the barrier problem (1.4) has at least one solution. This solution is also unique, since # - is strictly convex on {x # R Indeed, by Assumption 2.1(i), # 2 # - (x) given by (1.6) is positive definite. To simplify the notation we denote by z := (x, #) a typical pair of primal-dual variables and by Z the set of strictly feasible z's: 204 P. ARMAND, J. CH. GILBERT, AND S. JAN-J - EGOU The algorithm generates a sequence of pairs (z, M ), where z # Z and M is a positive definite symmetric matrix. Given a pair (z, M ), the next one (z obtained as follows. First is a step-size and is the unique solution of (1.3). The uniqueness comes from the positivity of c(x) and from the positive definiteness of M (for the unicity of d x , use (1.5)). Next, the matrix M is updated into M+ by the where # and # are given by This formula gives a symmetric positive definite matrix M+ , provided M is symmetric positive definite and # > 0 (see [14, 17]). This latter condition is satisfied because of the strong convexity assumption. Indeed, since at least one of the functions f or -c (i) is strongly convex, for any fixed # > 0, the function x #(x, #) is strongly convex, that is, there exists a constant # > 0 such that Since # sizes the displacement in x and #, the merit function used to estimate the progress to the solution must depend on both x and #. We follow an idea of Anstreicher and Vial [1] and add to # - a function forcing # to take the value -C(x) The merit function is defined for z = (x, # Z by where log # (i) c (i) (x) # . Note that # . Using # - as a merit function is reasonable provided the problem z # Z A BFGS INTERIOR POINT ALGORITHM 205 has for unique solution the solution of (1.2) and the direction descent direction of # - . This is what we check in Lemmas 2.5 and 2.6 below. Lemma 2.5. Suppose that Assumption 2.1 holds. Then, problem (2.4) has a unique solution - z -x - ), where - x - is the unique solution of the barrier problem # - has its ith component defined by ( - ) (i) := -/c (i) (-x - ). Furthermore, - has no other stationary point than - z - . Proof. By optimality of the unique solution - x - of the barrier problem (1.4) x such that c(x) > 0. On the other hand, since log t is minimized at c (i) (- x - ) (i) for all index i, we have z # Z. Adding up the preceding two inequalities gives # -z - (z) for all z # Z. Hence z - is a solution of (2.4). It remains to show that - z - is the unique stationary point of # - . If z is stationary, it satisfies Canceling # from the first equality gives #f(x) -#c(x)C(x) thus x - is the unique minimizer of the convex function # - . Now, by the second equation of the system above. Lemma 2.6. Suppose that z # Z and that M is symmetric positive definite. Let be the solution of (1.3). Then so that d is a descent direction of # - at a point z z - , meaning that # - (z) # d < 0. Proof. We have # - (z) # d. Using (1.5), which is nonpositive. On the other hand, when d satisfies the second equation of (1.3), one has (see [1]) which is also nonpositive. The formula for # - (z) # d given in the statement of the lemma follows from this calculation. Furthermore, # - (z) # d < 0, if z #= - z - . We can now state precisely one iteration of the algorithm used to solve the perturbed KKT system (1.2). The constants # ]0, 1[ and 0 < # < 1 are given independently of the iteration index. 206 P. ARMAND, J. CH. GILBERT, AND S. JAN-J - EGOU Algorithm A - (for solving (1.2); one iteration). At the beginning of the iteration, the current iterate supposed available, as well as a positive definite matrix M approximating the Hessian of the Lagrangian # 2 xx #(x, #). 1. Compute d := (d x , d # ), the solution of the linear system (1.3). 2. Compute a step-size # by means of a backtracking line search. 2.0. 2.1. Test the su#cient decrease condition: 2.2. If (2.5) is not satisfied, choose a new trial step-size # in [#] and go to Step 2.1. If (2.5) is satisfied, set z 3. Update M by the BFGS formula (2.1) where # and # are given by (2.2). By Lemma 2.6, d is a descent direction of # - at z, so that a step-size # > 0 satisfying (2.5) can be found. In the line-search, it is implicitly assumed that (2.5) is not satisfied if z holds for the new iterate z + . We conclude this section with a result that gives the contribution of the line-search to the convergence of the sequence generated by Algorithm A - . It is in the spirit of a similar result given by Zoutendijk [50] (for a proof, see [6]). We say that a function is C 1,1 if it has Lipschitz continuous first derivatives. We denote the level set of # - determined by the first iterate z Lemma 2.7. If - is C 1,1 on an open convex neighborhood of the level set L PD there is a positive constant K such that for any z # L PD determined by the line-search in Step 2 of Algorithm A - , one of the following two inequalities holds: It is important to mention here that this result holds even though - may not be defined for all positive step-sizes along d, so that the line-search may have to reduce the step-size in a first stage to enforce feasibility. 3. The global and r-linear convergence of Algorithm A- . In the convergence analysis of BFGS, the path to q-superlinear convergence traditionally leads through r-linear convergence (see [41, 6]). In this section, we show that the iterates generated by Algorithm A - converge to - z -x - ), the solution of (1.2), with that convergence speed. We use the notation Our first result shows that, because the iterates (x, #) remain in the level set L PD the sequence {(c(x), #)} is bounded and bounded away from zero. A BFGS INTERIOR POINT ALGORITHM 207 Lemma 3.1. Suppose that Assumption 2.1 holds. Then, the level set L PD 1 is compact and there exist positive constants K 1 and K 2 such that 1 . Proof. Since # c(x) - log(# (i) c (i) (x)) is bounded below by m-(1 - log -), there is a constant K 1 . By Assumption 2.1 and Lemma 2.3, the level set L # 1 } is compact. By continuity, c(L # ) is also compact, so that c(x) is bounded and bounded away from zero for all z # L PD 1 . What we have just proven implies that {# - 1 } is bounded below, so that there is a constant K # 1 . Hence the #-components of the z's in L PD are bounded and bounded away from zero. We have shown that L PD 1 is included in a compact set. Now, it is itself compact by continuity of # - . The next proposition is crucial for the technique we use to prove global convergence (see [6]). It claims that the proximity of a point z to the unique solution of (2.4) can be measured by the value of # - (z) or the norm of its gradient # - (z). In unconstrained optimization, the corresponding result is a direct consequence of strong convexity. Here, # - is not necessarily convex, but the result can still be established by using Lemma 2.5 and Lemma 3.1. The function # - is nonconvex, for example, when is minimized on the half-line of nonnegative real numbers. Proposition 3.2. Suppose that Assumption 2.1 holds. Then, there is a constant a > 0 such that for any z # L PDa#z - z - # 2 Proof. Let us show that # - is strongly convex in a neighborhood of - z - . Using (2.3) and the fact that - # -e, the Hessian of # - at - z - can be written as # . From Assumption 2.1, for fixed # > 0, the Lagrangian is a strongly convex function in the variable x. It follows that its Hessian with respect to x is positive definite at us show that the above matrix is also positive definite. Multiplying the matrix on both sides by a vector (u, v) # R n positive definite and c(-x - ) > 0, this quantity is nonnegative. If it vanishes, one deduces that next that positive definite. Let us now prove a local version of the proposition: there exist a constant a # > 0 and an open neighborhood N # Z of - z - such that a #z - z - # 2 208 P. ARMAND, J. CH. GILBERT, AND S. JAN-J - EGOU The inequality on the left comes from the fact that # -z - and the strong convexity of # - near - z - . For the inequality on the right, we first use the local convexity of for an arbitrary z near - z -z - (z) z). With the Cauchy-Schwarz inequality and the inequality on the left of (3.2), one gets a # # 1Simplifying and squaring give the inequality on the right of (3.2). To extend the validity of (3.2) for all z # L PD su#ces to note that, by virtue of Lemma 2.5, the ratios and - (z) -z - ) are well defined and continuous on the compact set L PD z - is the unique minimizer of - on L PD 2.5), the ratios are respectively bounded away from zero and bounded above on L PD 1 \ N , by some positive constants K # 1 and K # 2 . The conclusion of the proposition now follows by taking The proof of the r-linear convergence rests on the following lemma, which is part of the theory of BFGS updates. It can be stated independently of the present context (see Byrd and Nocedal [6]). We denote by # k the angle between M k # k and and by # the roundup operator: Lemma 3.3. Let {M k } be positive definite matrices generated by the BFGS formula using pairs of vectors {(# k , # k )} k#1 , satisfying for all k # 1 where a 1 > 0 and a 2 > 0 are independent of k. Then, for any r # ]0, 1[, there exist positive constants b 1 , b 2 , and b 3 , such that for any index k # 1, for at least #rk# indices j in {1, . , k}. The assumptions (3.3) made on # k and # k in the above lemma are satisfied in our context. The first one is due to the strong convexity of one of the functions f , -c (1) , . , -c (m) , and to the fact that # is bounded away from zero (Lemma 3.1). When f and c are C 1,1 , the second one can be deduced from the Lipschitz inequality, the boundedness of # (Lemma 3.1), and the first inequality in (3.3). Theorem 3.4. Suppose that Assumption 2.1 holds and that f and c are C 1,1 functions. Then, Algorithm A - generates a sequence {z k } converging to - z - r-linearly, meaning that lim sup k#z k - z - # 1/k < 1. In particular, z - #. Proof. We denote by K # positive constants (independent of the iteration index). We also use the notation A BFGS INTERIOR POINT ALGORITHM 209 The bounds on (c(x), #) given by Lemma 3.1 and the fact that f and c are C 1,1 imply that # - is C 1,1 on some open convex neighborhood of the level set L PD 1 , for example, on where O is an open bounded convex set containing L PD 1 (this set O is used to have #c bounded on the given neighborhood). Therefore, by the line-search and Lemma 2.7, there is a positive constant K # 1 such that either or Let us now apply Lemma 3.3: fix r # ]0, 1[ and denote by J the set of indices j for which (3.4) holds. Using Lemma 2.6 and the bounds from Lemma 3.1, one has for # . Let us denote by K # 4 a positive constant such that #c(x)# K # 4 for all x # L PD 1 . By using (2.3), (1.5), and the inequality (a and also, by (1.3), #d x P. ARMAND, J. CH. GILBERT, AND S. JAN-J - EGOU Combining these inequalities with (3.5) or (3.6) gives for some positive constant K #and for any j # J The end of the proof is standard (see [41, 6]). Using Proposition 3.2, for j # J , # [0, 1[. On the other hand, by the line-search, # - (z k+1 # 1. By Lemma 3.3, |[1, k] # J | #rk# rk, so that the last inequality gives for any k # 1 where K # 7 is the positive constant using the inequality on the left in (3.1), one has for all k # 1 from which the r-linear convergence of {z k } follows. 4. The q-superlinear convergence of Algorithm A- . With the r-linear convergence result of the previous section, we are now ready to establish the q-superlinear convergence of the sequence {z k } generated by Algorithm A - . By definition, {z k } converges q-superlinearly to - z - if the following estimate holds: z z - #), which means that #z k+1 - z - #z k - z - # 0 (assuming z k #= - z - ). To get this result, f and c have to be a little bit smoother, namely twice continuously di#erentiable near x - . We use the notation We start by showing that the unit step-size is accepted asymptotically by the line-search condition (2.5), provided the updated matrix M k becomes good (or su#ciently large) in a sense specified by inequality (4.1) below and provided the iterate z k is su#ciently close to the solution - z - . Given two sequences of vectors {u k } and {v k } in some normed spaces and a positive number #, we write u k # o(#v k # ), if there exists a sequence of {# k } # R such that # k # 0 and u k # k #v k # for all k. Proposition 4.1. Suppose that Assumption 2.1 holds and that f and c are twice continuously di#erentiable near - x - . Suppose also that the sequence {z k } generated by Algorithm A - converges to - z - and that the positive definite matrices M k satisfy the estimate (d x when k #. Then the su#cient decrease condition (2.5) is satisfied with # k su#ciently large provided that # < 1 A BFGS INTERIOR POINT ALGORITHM 211 Proof. Observe first that the positive definiteness of - - with (4.1) implies that (d x for some positive constant K # and su#ciently large k. Observe also that d k # 0 (for d x use (1.5), (4.2), and # - Therefore, for k large enough, z k and z k are near - z - and one can expand - (z k . A second order expansion gives for the left-hand side of (2.5) We want to show that this quantity is negative for k large. Our first aim is to show that # - (z k smaller than a term of order o(#d k # 2 ). For this purpose, one computes . On the other hand, using one gets from Lemma 2.6 k . With these estimates, (4.1), and the fact that # 2 with Lemma 3.1 and the boundedness of {#c k }, (4.3) becomes (d x (d # clear that the result will be proven if we show that, for some positive constant K and k large, # - (z k ) # d k # -K#d k # 2 . To show this, we use the 212 P. ARMAND, J. CH. GILBERT, AND S. JAN-J - EGOU last expression of # - (z k ) # d k and an upper bound of |(d x k |, obtained by the Cauchy-Schwartz inequality: (d x (d # k . It follows that (d x (d # k . Therefore, using (4.2) and Lemma 3.1, one gets for some positive constant K and k large. Proposition 4.1 shows in particular that the function V, which was added to # - to get the merit function - , has the right curvature around - z - , so that the unit step-size in both x and # is accepted by the line-search. In the following proposition, we establish a necessary and su#cient condition of q-superlinear convergence of the Dennis and Mor-e [13] type. The analysis assumes that the unit step-size is taken and that the updated matrix M k is su#ciently good asymptotically in a manner given by the estimate (4.5), which is slightly di#erent from (4.1). Proposition 4.2. Suppose that Assumption 2.1 holds and that f and c are twice di#erentiable at - x - . Suppose that the sequence {z k } generated by Algorithm A - converges to - z - and that, for k su#ciently large, the unit step-size # accepted by the line-search. Then {z k } converges q-superlinearly towards - z - if and only if Proof. Let us denote by M the nonsingular Jacobian matrix of the perturbed KKT conditions (1.2) at the solution - z -x - # . A first order expansion of the right-hand side of (1.3) about - z - and the identities # -e give Subtracting Md k from both sides and assuming a unit step-size, we obtain z - #). A BFGS INTERIOR POINT ALGORITHM 213 Suppose now that {z k } converges q-superlinearly. Then, the right-hand side of (4.6) is of order o(#z k - z - #), so that Then (4.5) follows from the fact that, by the q-superlinear convergence of {z k }, z k - Let us now prove the converse. By (4.5), the left-hand side of (4.6) is an o(#d k #) and due to the nonsingularity of M, (4.6) gives z k+1 - z With a unit step-size, d z - (z k - z - ), so that we finally get z k+1 - z For proving the q-superlinear convergence of the sequence {z k }, we need the following result from the BFGS theory (see [40, Theorem 3] and [6]). Lemma 4.3. Let {M k } be a sequence of matrices generated by the BFGS formula from a given symmetric positive definite matrix M 1 and pairs (# k , # k ) of vectors verifying < #, where M is a symmetric positive definite matrix. Then, the sequences {M k } and k } are bounded and By using this lemma, we will see that the BFGS formula gives the estimate #). Note that the above estimate implies (4.5), from which the q-superlinear convergence of {z k } will follow. A function #, twice di#erentiable in a neighborhood of a point x # R n , is said to have a locally radially Lipschitzian Hessian at x, if there exists a positive constant L such that for x # near x, one has Theorem 4.4. Suppose that Assumption 2.1 holds and that f and c are C 1,1 functions, twice continuously di#erentiable near - x - with locally radially Lipschitzian Hessians at - x - . Suppose that the line-search in Algorithm A - uses the constant # <2 . Then the sequence {z k generated by this algorithm converges to z -x - ) q-superlinearly and, for k su#ciently large, the unit step-size # accepted by the line-search. Proof. Let us start by showing that Lemma 4.3 with M - can be applied. First, # k # k > 0, as this was already discussed after Lemma 3.3. For the convergence of the series in (4.7), we use a Taylor expansion, assuming that k is large enough (f and c are C 2 near - 214 P. ARMAND, J. CH. GILBERT, AND S. JAN-J - EGOU With the local radial Lipschitz continuity of # 2 f and # 2 c at - x - and the boundedness of {# k+1 }, there exist positive constants K # and K # such that z - # . Hence the series in (4.7) converges by Theorem 3.4. Therefore, by (4.8) with and the fact that # k is parallel to d x #). By the estimate (4.9) and Proposition 4.1, the unit step-size is accepted when k is large enough. The q-superlinear convergence of {z k } follows from Proposition 4.2. 5. The overall primal-dual algorithm. In this section, we consider an overall algorithm for solving problem (1.1). Recall from Lemma 2.2 that the set of primal solutions of this problem is nonempty and bounded. By the Slater condition (As- sumption 2.1(ii)), the set of dual solutions is also nonempty and bounded. Let us denote by - primal-dual solution of problem (1.1), which is also a solution of the necessary and su#cient conditions of optimality Our overall algorithm for solving (1.1) or (5.1), called Algorithm A, consists in computing approximate solutions of the perturbed optimality conditions (1.2), for a sequence of -'s converging to zero. For each -, the primal-dual Algorithm A - is used to find an approximate solution of (1.2). This is done by so-called inner iterations. Next - is decreased and the process of solving (1.2) for the new value of - is repeated. We call an outer iteration the collection of inner iterations for solving (1.2) for a fixed value of -. We index the outer iterations by superscripts j # N\{0}. Algorithm A (for solving problem (1.1); one outer iteration). At the beginning of the jth outer iteration, an approximation z j # Z of the solution - z of (5.1) is supposed available, as well as a positive 1 approximating the Hessian of the Lagrangian. A value is given, as well as a precision threshold # j > 0. 1. Starting from z j use Algorithm A - until z j := 2. Choose a new starting iterate z j+1 for the next outer iteration, as well as a positive definite matrix M j+1 1 . Set the new parameters - j+1 > 0 and # j+1 > 0, such that {- j } and {# j } converge to zero when j #. A BFGS INTERIOR POINT ALGORITHM 215 To start the (j+1)th outer iteration, a possibility is to take z j+1 , the updated matrix obtained at the end of the jth outer iteration. As far as the global convergence is concerned, how z are determined is not important. Therefore, on that point, Algorithm A leaves the user much freedom to maneuver, while Theorem 5.1 gives us a global convergence result for such a general algorithm. Theorem 5.1. Suppose that Assumption 2.1 holds and that f and c are C 1,1 functions. Then Algorithm A generates a bounded sequence {z j } and any limit point of {z j } is a primal-dual solution of problem (1.1). Proof. By Theorem 3.4, any outer iteration of Algorithm A terminates with an iterate z j satisfying the stopping criteria in Step 1. Therefore Algorithm A generates a sequence {z j }. Since the sequences {- j } and {# j } converge to zero, any limit point of {z j } is a solution of problem (1.1). It remains to show that {z j } is bounded. Let us first prove the boundedness of {x j }. The convexity of the Lagrangian implies that Using the positivity of # j and c(x 1 ) and next the stopping criteria of Algorithm A, it follows that #). If {x j } is unbounded, setting t j := #x j # and y j := x j one can choose a subsequence J such that lim From the last inequality we deduce that Moreover, since c(x j ) > 0, we have (-c (i) It follows that (see, for example, [22, Proposition IV.3.2.5] or [2, Formula (1)]). Therefore, the solution set of problem (1.1) would be unbounded, which is in contradiction with what is claimed in Lemma 2.2. To prove the boundedness of the multipliers, suppose that the algorithm generates an unbounded sequence of positive vectors {# j subsequence J # . The sequence #)} j#J # is bounded and thus has at least one limit point, say, Dividing the two inequalities in (5.2) by # j # and taking limits when j #, we deduce that # 0, #c(x Using the concavity of the components c (i) , one has where the inequality on the right follows from the strict feasibility of the first iterate. Multiplying by # , we deduce that (# c(x 1 contradiction with P. ARMAND, J. CH. GILBERT, AND S. JAN-J - EGOU In the rest of this section, we give conditions under which the whole sequence } converges to a particular point called the analytic center of the primal-dual optimal set. This actually occurs when the following two conditions hold: strict complementarity and a proper choice of the forcing sequence # j in Algorithm A, which has to satisfy the estimate meaning that # j /- j Let us first recall the notion of analytic center of the optimal sets, which under Assumption 2.1 is uniquely defined (see Monteiro and Zhou [37], for related results). We denote by opt(P ) and opt(D) the sets of primal and dual solutions of problem (1.1). The analytic center of opt(P ) is defined as follows. If opt(P ) is reduced to a single point, its analytic center is precisely that point. Otherwise, opt(P ) is a convex set with more than one point. In that case, f is not strongly convex and, by Assumption 2.1(i), at least one of the constraint functions, -c (i 0 ) say, is strongly convex. It follows that the index set is nonempty (it contains i 0 ). The analytic center of opt(P ) is then defined as the unique solution of the following problem: log c (i) (-x) # . The fact that this problem is well defined and has a unique solution is the matter of Lemma 5.2 below. Similarly, if opt(D) is reduced to a single point, its analytic center is that point. In case of multiple dual solutions, the index set is nonempty (otherwise opt(D) would be reduced to {0}). The analytic center of opt(D) is then defined as the unique solution of the following problem: log - # (i) # . Lemma 5.2. Suppose that Assumption 2.1 holds. If opt(P ) (resp., opt(D)) is not reduced to a singleton, then problem (5.3) (resp., (5.4)) has a unique solution. Proof. Consider first problem (5.3) and suppose that opt(P ) is not a singleton. We have seen that B is nonempty. By the convexity of the set opt(P ) and the concavity of the functions c (i) , there exists - Therefore the feasible set in (5.3) is nonempty. On the other hand, let - x 0 be a point satisfying the constraints in (5.3). Then the set log c i (-x) # i#B log c i (-x is nonempty, bounded (Lemma 2.2), and closed. Therefore, problem (5.3) has a solution. Finally, by Assumption 2.1(i), we know that there is an index i A BFGS INTERIOR POINT ALGORITHM 217 that -c (i 0 ) is strongly convex. It follows that the objective in (5.3) is strongly concave and that problem (5.3) has a unique solution. By similar arguments and the fact that the objective function in (5.4) is strictly concave, it follows that problem (5.4) has a unique solution. By complementarity (i.e., C(-x) - and convexity of problem (1.1), the index sets B and N do not intersect, but there may be indices that are neither in B nor in N . It is said that problem (1.1) has the strict complementarity property if {1, . , n}. This is equivalent to the existence of a primal-dual solution satisfying strict complementarity. Theorem 5.3. Suppose that Assumption 2.1 holds and that f and c are C 1,1 functions. Suppose also that problem (1.1) has the strict complementarity property and that the sequence {# j } in Algorithm A satisfies the estimate # Then the sequence {z j } generated by Algorithm A converges to the point - z x 0 is the analytic center of the primal optimal set and - # 0 is the analytic center of the dual optimal set. Proof. Let (- x, - #) be an arbitrary primal-dual solution of (1.1). Then - x minimizes #) and - # so that Using the convexity of # j ) and the stopping criterion (5.2) of the inner iterations in Algorithm A, one has x#, because . By Theorem 5.1, there is a constant C 1 such that m 1 adding the corresponding sides of the two inequalities above leads to We pursue this by adapting an idea used by McLinden [34] to give properties of the limit points of the path -x - ). Let us define # has for all indices i c (i) c (i) Substituting this in (5.5) and dividing by - j give c (i) (-x) c (i) By assumptions, # Now supposing that (-x 0 , - # 0 ) is a limit point of {(x j , # j )} and taking the limit in the preceding estimate provide c (i) (-x) P. ARMAND, J. CH. GILBERT, AND S. JAN-J - EGOU Necessarily, now that, by strict complementarity, there are exactly m terms on the left-hand side of the preceding inequality. Hence, by the arithmetic-geometric mean inequality c (i) (-x) or c (i) (-x) # One can take - in this inequality, so that c (i) (-x) # c (i) (-x 0 ) and # This shows that - x 0 is a solution of (5.3) and that - # 0 is a solution of (5.4). Since the problems in (5.3) and (5.4) have unique solutions, all the sequence {x j } converges to x 0 and all the sequence {# j } converges to - # 0 . 6. Discussion. By way of conclusion, we discuss the results obtained in this paper, give some remarks, and raise some open questions. Problems with linear constraints. The algorithm is presented with convex inequality constraints only, but it can also be used when linear constraints are present. Consider the problem min f(x), obtained by adding linear constraints to problem (1.1). In (6.1), A is a p - n matrix with p < n and b # R p is given in the range space of A. Problem (6.1) can be reduced to problem (1.1) by using a basis of the null space of the matrix A. Indeed, let x 1 be the first iterate, which is supposed to be strictly feasible in the sense that Let us denote by Z an n - q matrix whose columns form a basis of the null space of A. Then, any point satisfying the linear constraints of (6.1) can be written With this notation, problem (6.1) can be rewritten as the problem in u # R which has the form (1.1). Thanks to this transformation, we can deduce from Assumption 2.1 what are the minimal assumptions under which our algorithm for solving problem (6.2) or, equivalently, problem (6.1) will converge. A BFGS INTERIOR POINT ALGORITHM 219 Assumption 6.1. (i) The real-valued functions f and -c (i) (1 # i # m) are convex and di#erentiable on the a#ne subspace X b} and at least one of the functions f , -c (1) , . , -c (m) is strongly convex on X. (ii) There exists an x # R n such that With these assumptions, all the previous results apply. In particular, Algorithm A - converges r-linearly (if f and c are also C 1,1 ) and q-superlinearly (if f and c are also C 1,1 , twice continuously di#erentiable near - x - with locally radially Lipschitzian Hessian at - Similarly, the conclusions of Theorem 5.1 apply if f and c are also C 1,1 . Feasible algorithms and qN techniques. In the framework of qN methods, the property of having to generate feasible iterates should not be only viewed as a restriction limiting the applicability of a feasible algorithm. Indeed, in the case of problem (6.2), if it is sometimes di#cult to find a strictly feasible initial iterate, the matrix to update for solving this problem is of order q only, instead of order n for an infeasible algorithm solving problem (6.1) directly. When q # n, the qN updates will approach the reduced Hessian of the Lagrangian Z # 2 #)Z more rapidly than the full Hessian # 2 #, so that a feasible algorithm is likely to converge more rapidly. About the strong convexity hypothesis. Another issue concerns the extension of the present theory to convex problems, without the strong convexity assumption (Assumption 2.1(i)). this hypothesis, the class of problems to consider encompasses linear programming (f and c are a#ne). It is clear that for dealing properly with linear programs, our algorithm needs modifications, since then # and the BFGS formula is no longer defined. Of course, it would be very ine#ective to solve linear programs with the qN techniques proposed in this paper (M is the desired matrix), but problems that are almost linear near the solution may be encountered, so that a technique for dealing with a situation where # k # k # can be of interest. To accept # can look at the limit of the BFGS formula (2.1) when possible update formula could be The updated matrix satisfies M k+1 # positive semidefinite, provided is already positive semidefinite. The fact that M k+1 may be singular raises some di#culties, however. For example, the search direction d x may no longer be defined (see formula (1.5), in which the matrix M +#c(x)C(x) -1 #c(x) # can be singular). Therefore, the present theory cannot be extended in a straightforward manner. On the other hand, the strong convexity assumption may not be viewed as an important restriction, because a fictive strongly convex constraint can always be added. An obvious example of fictive constraint is "x # x # K." If the constant K is large enough, the constraint is inactive at the solution, so that the solution of the original problem is not altered by this new constraint and the present theory applies. Better control of the outer iterations. Last but not least, the global convergence result of section 5 is independent of the update rule of the parameters # j and In practice, however, the choice of the decreasing values # j and - j is essential for the e#ciency of the algorithm and would deserve a detailed numerical study. From a theoretical viewpoint, it would be highly desirable to have an update rule that would allow the outer iterates of Algorithm A to converge q-superlinearly. Along P. ARMAND, J. CH. GILBERT, AND S. JAN-J - EGOU the same lines, an interesting problem is to design an algorithm in which the barrier parameter would be updated at every step, while having q-superlinear convergence of the iterates. Such extensions would involve more di#cult issues. The global convergence result proved in this paper gives us some reasons to believe that it is not unreasonable to tackle these open questions. Acknowledgments . We would like to thank the referees for their valuable com- ments. One of them has shown us a direct argument for the last part of the proof of Proposition 3.2, which is the one we have finally chosen to give in the paper. The other referee has brought McLinden's paper to our attention, which led us to Theorem 5.3. --R On the convergence of an infeasible primal-dual interior-point method for convex programming Asymptotic analysis for penalty and barrier methods in convex and linear programming Computational Di A trust region interior point algorithm for linearly constrained optimization A. Trust Region Method Based on Interior Point Techniques for A tool for the analysis of quasi-Newton methods with application to unconstrained minimization A Primal-Dual Algorithm for Minimizing a Non-convex Function Subject to Bound and Linear Equality Constraints Principles and Techniques of Algorithmic Di Interior Point Approach to Linear A Potential Reduction Method for a Class of Smooth Convex Programming Problems On the classical logarithmic barrier function method for a class of smooth convex programming problems Numerical Methods for Unconstrained Optimization and Nonlinear Equations On the formulation and theory of the Newton interior-point method for nonlinear programming Sequential Unconstrained Minimization Techniques Practical Methods of Optimization Feasible direction interior-point technique for nonlinear optimization Interior Point Techniques in Optimization-Complementarity On the method of analytic centers for solving smooth convex problems A practical interior-point method for convex programming A Unified Approach to Interior Point Algorithms for Linear Complementarity Problems A new continuation method for complementarity problems with uniform P Limiting behavior of trajectories generated by a continuation method for monotone complementarity problems On the limited memory BFGS method for large scale optimization The projective SUMT method for convex programming problems An analogue of Moreau's proximation theorem An interior point algorithm for solving smooth convex programs based on Newton's method An extension of Karmarkar-type algorithms to a class of convex separable programming problems with global linear rate of convergence On the existence and convergence of the central path for convex programming and some duality results Updating quasi-Newton matrices with limited storage On the convergence of the variable metric algorithm Some global convergence properties of a variable metric algorithm for minimization without exact line searches Theory and Algorithms for Linear Optimization-An Interior Point Approach "analytical center" Interior Point Methods of Mathematical Programming Computational experience with a primal-dual interior-point method for smooth convex programming Why a pure primal Newton barrier step may be infeasible Superlinear and quadratic convergence of some primal-dual interior point methods for constrained optimization Interior Point Algorithms-Theory and Analysis in Integer Nonlinear Pro- gramming --TR --CTR Richard H. Byrd , Jorge Nocedal , Richard A. Waltz, Feasible Interior Methods Using Slacks for Nonlinear Optimization, Computational Optimization and Applications, v.26 n.1, p.35-61, October Paul Armand, A Quasi-Newton Penalty Barrier Method for Convex Minimization Problems, Computational Optimization and Applications, v.26 n.1, p.5-34, October Dingguo Pu , Weiwen Tian, The revised DFP algorithm without exact line search, Journal of Computational and Applied Mathematics, v.154 n.2, p.319-339, 15 May
convex programming;superlinear convergence;constrained optimization;interior point algorithm;analytic center;BFGS quasi-Newton approximations;primal-dual method;line-search
589092
Auction Algorithms for Shortest Hyperpath Problems.
The auction-reduction algorithm is a strongly polynomial version of the auction method for the shortest path problem. In this paper we extend the auction-reduction algorithm to different types of shortest hyperpath problems in directed hypergraphs. The results of preliminary computational experiences show that the auction-reduction method is comparable to other known methods for specific classes of hypergraphs.
Introduction The shortest hyperpath problem is the extension to directed hypergraphs [11] of the classical shortest path problem (SPT) in directed graphs. Though not as pervasive as SPT, shortest hyperpaths have several relevant applications. In particular, they are at the core of traffic assignment algorithms for transit networks [13, 14, 18]. Shortest hyperpath models have been constructed for the SPT problem in stochastic and time-dependent networks [17] and for production planning in assembly lines [12]. Moreover, shortest hyperpath algorithms are used as building blocks of enumerative algorithms for hard combinatorial problems [10]. As a consequence, there is a growing interest for efficient shortest hyperpath algorithms. This provides motivations for further investigating known methods [11, 15], both from a theoretical and a practical point of view, and for developing new ones. Auction algorithms were first proposed by Bertsekas [1, 2] for the assignment problem and later extended to general transportation problems [5, 6]. A survey of the auction algorithms for network optimization problems is contained in [4, Chapter 4]. Auction algorithms for shortest path problems on graphs were proposed in [3]. For the single-origin single-destination case the method can be viewed as an application of the auction method (with to a specifically constructed assignment problem, and finite termination of the procedure can be established. Furthermore, the algorithm is a dual coordinate ascent method. Strongly polynomial version of the auction method were proposed by Pallottino and Scutell'a [16], who define a pruning procedure that reduces the graph to the shortest path tree. Further improvements to this method are given in [7], where the pruning method is strengthened, and the structure of the reduced graph is exploited to obtain a better time complexity. A variant of the auction algorithm with pruning is proposed in [8]. In this paper, we devise an auction method for shortest hyperpaths with nonnegative hyperarc weights, by slightly modifying the SPT algorithm given in [7]. Our method can be tailored to solve several types of shortest hyperpath problems; for the various cases, we provide a worst case complexity bound. Finally, we report the results of a preliminary computational experience. In Section 2 we give the basic definitions on hypergraphs and shortest hyper- paths. The proposed auction method for shortest hyperpaths is presented in Section 3. Computational results and conclusions are presented in Sections 4 and 5, respectively. Shortest Hyperpaths in Directed Hypergraphs A directed hypergraph H is a pair (V; E), where V is a set of nodes and E is a set of directed hyperarcs; a hyperarc is a pair the tail of e, and h(e) 2 V n T (e) is its head. A detailed introduction to directed hypergraphs can be found in [11], where a more general definition of hypergraphs is introduced; the particular class of hypergraphs considered in this paper are called B-graphs in [11]. The size jej of hyperarc e is the number of nodes it contains in its tail and head: The hyperarc e is an arc if 2, and a proper hyperarc if jej ? 2. Denote by m a and m h the number of arcs and proper hyperarcs, j. The size of H is the sum of the cardinalities of its hyperarcs: Given a node u, the Forward Star of u, FS(u), is the set of hyperarcs e such that and the Backward Star of u, BS(u), is the set of hyperarcs e such that A path P st , of length q, in the hypergraph E) is a sequence: st nodes s and t are the origin and the destination of the path P st , respectively. We say that node t is connected to node s in H if a path P st exists in H. If t 2 T (e 1 ), then the path P st is a cycle. A path is cycle-free if it does not contain any subpath which is a cycle, Given a hypergraph E) and two nodes s; t 2 V, a hyperpath - st is a minimal hypergraph (with respect to deletion of nodes and hyperarcs) H such 2. s; 3. connected to s in H - by a cycle-free path. Observe that for each there exists a unique hyperarc e 2 E - such that is the predecessor hyperarc of u in -, and is denoted by e - (u). We say that node t is hyperconnected to s in H if there exists a hyperpath from s to t in H. Given a hyperarc a, we say that a hyperarc a r is contained in a, or is a reduction of a, if h(a r (a). Note that a is contained in itself, and a r is strictly contained in a if T (a r Given a and u 2 T (a), we denote by anu the reduction of a obtained deleting u from T (a). We say that a hypergraph is full when it contains all the possible reductions of each of its proper hyperarcs. A full hypergraph can be represented by its support hypergraph H s , obtained by deleting all the strictly contained hyperarcs. Conversely, given any hypergraph H, we can obtain the corresponding full hypergraph H f by adding all the strictly contained hyperarcs. 2.1 Shortest Hyperpaths A weighted hypergraph is such that each hyperarc e is assigned a nonnegative real weight w(e). The weight of a hyperpath in a weighted hypergraph can be defined in several ways. It is known that some definitions lead to intractable shortest hyperpath problems [9]. Here we restrict to definitions that are known to be tractable [11]. A weighting function is a node function that, given a hyperpath - st assigns a value W - (u) to its nodes, depending on the weights of its hyperarcs. The value W - (t) is the weight of - under the chosen weighting function. An additive weighting function satisfies the properties that W - a function of the predecessor e - (u) only. Formally, an additive weighting function can be defined by means of the following recursive equations: where F (e) is a non-decreasing function of the weights of the nodes in T (e). Clearly, many different additive weighting functions can be defined (see [11]); consider first the value function, which is obtained by defining F (e) in (1) as follows: The minimum value (shortest hyperpath) problem consists of finding a set of minimum value hyperpaths from the origin node s to each node u 6= s hyperconnected to s. We denote by V (u) the minimum value of a hyperpath - su in H; we assume if u is not hyperconnected to s. The distance function is obtained from (1) defining F (e) as follows: The minimum distance problem asks for the minimum distance hyperpaths from origin s. Denote by D the vector of minimum distances, where again if u is not hyperconnected to s. Since arc weights are nonnegative, the minimum value and minimum distance problems can be solved efficiently by procedure SBT [11]. A computational analysis of several variants of SBT can be found in [15]. 2.1.1 Minimum Time Problems in Transit Hypergraphs The problem of finding the passenger's expected travel time is at the core of several urban transit networks models. This problem has been formulated in terms of hyper- paths in transit networks [13] and in F-graphs [11, 14]. Here, introducing the time weighting function in transit hypergraphs, we define a particular shortest hyperpath problem that, though using a different (and slightly more general) terminology, is equivalent to the formulations found in [13, 14]. A transit hypergraph is a weighted support hypergraph E) where a positive parameter OE u is associated with each node u 2 V. Let H the full hypergraph represented by H. Consider an e contained in a proper hyperarc a 2 H: the time weighting function is obtained from (1) by defining the weight w(e) and the function F (e) as follows: OE u In practice, F (e) is the weighted average (with weights OE:) of the values W - (\Delta) in T (e), while w(e) is the inverse of the sum of the weights OE: in T (e). For an arc corresponding to an arc a 2 E the function F (e) is defined in the same way, which gives F however, in this case w(e) = w(a) can be any nonnegative value. The minimum time problem consists of finding the minimum time hyperpaths, from a given origin s, in the full hypergraph H f . Note that H f may be considerably larger than its support H: in practice, solving the minimum time problem efficiently requires to work directly on H. This is the aim of the following observations. Denote by E the vector of minimum times, where is not hyperconnected to s in H f . For each e 2 OE u Consider a proper hyperarc a 2 E , and let R(a) be the set of the reductions of a; denote by e(a) the reduction of a yielding the minimum time t(e): The following relations hold ([13], Proposition 6): Consider the nodes in T (a) in increasing order of E, i.e. let T with be the reduction of a such that T (a g. It follows from (2) that e(a) = a h , where In practice, working on the support hypergraph requires finding the hyperarc e(a) without considering the whole set of reductions of a. Indeed, according to the previous observations, this can be done by processing the nodes in T (a) in the order . For each u compute the value t(a i This technique has been used to compute expected travel times efficiently [13, 14], and will be adopted in our auction algorithm for the minimum time problem. 2.1.2 Reductions and Shortest Hyperpaths A hyperarc reduction operation on a proper hyperarc a consists of replacing a by a hyperarc a r contained in a, returning a reduced hypergraph. Clearly, if the hyper-graph is weighted, a nonnegative weight w(a r ) must be assigned to a r . The following propositions show that by suitably choosing the weight on a r a reduction operation does not modify the optimal solution of the shortest hyperpath problem. Proofs are rather straightforward and are omitted. Suppose we are given a weighted hypergraph H, and the corresponding vectors of optimal solutions V , D for the value and distance weighting functions. Given a proper hyperarc a and u 2 T (a), consider replacing a by a r = a n u. Proposition 2.1 If w(a r is the vector of optimal values in the reduced hypergraph. Proposition 2.2 If D(u) - is the vector of optimal distances in the reduced hypergraph. Now suppose we are given a transit hypergraph H, the corresponding full hypergraph and the optimum times E. Let e be a proper hyperarc in H f , with t(e) ? E(u) for each u 2 T (e), and consider replacing e by an arc e Proposition 2.3 If w(e r is the vector of optimal times in the reduced full hypergraph. Auction Algorithms for Shortest Hyperpaths In this section we propose an auction method for the minimum value problem and discuss the adaptation to other weighting functions. Before introducing our ap- proach, we briefly recall some relevant features of the auction algorithms for SPT; the reader is referred to the cited literature for further details. The auction algorithm for shortest path problems on graphs maintains a path P (the candidate path) starting at the origin s and a set of dual node prices p satisfying the following complementary slackness (CS) conditions: is the cost of arc (i; j). The algorithm consists of three basic operations: path extension, path contraction and dual price raise. At each iteration, the candidate path P is possibly extended, by adding a new node at the end of the path without violating (3). When no extensions are possible, the dual price of the terminal node i in P is raised, and if i 6= s the path is contracted by deleting node i. For the single-origin, single-destination case the algorithm terminates when the destination node is reached; several variants have been devised, also for the multiple-destination case. Consider the case of nonnegative costs and dual prices initially set to zero. At the first scan of a node i (i.e. when node i becomes the last node in P for the first time) the optimal distance of node i is determined, indeed, it is equal to p(s). As a consequence, since p(s) is never decreased during the algorithm, the sequences of first scan operations ranks the nodes in increasing order of distance from the origin s. Based on the above property, the auction-reduction method [16] introduces the following reduction operation: at the first scan of node i, delete each arc entering i, except the last arc in P . By means of these reduction operations, the graph is transformed into the shortest path tree, and a strongly polynomial time complexity can be obtained. Further reduction operations, including deletion of nodes, have been proposed in [7], improving the complexity bound. The following observation is at the core of our auction shortest hyperpath method: Observation 3.1 According to the definition of value, distance, and time weighting functions, for an arc a = (fug; v) we have F (a) In other words, the weighting functions above define a standard SPT problem if the hypergraph is a directed graph. This suggests the following technique: ffl apply the auction-reduction SPT algorithm to the arcs of the hypergraph; ffl at the first scan of node i, apply hyperarc reduction operations according to Properties (2.1),. ,(2.3), possibly generating new arcs. Note that the hypergraph is modified during the algorithm; each step is applied to the current hypergraph, as returned by the previous reduction operations. In practice, during the execution the proper hyperarcs lay in the not yet explored part of the hypergraph, and they are not considered until they are replaced by arcs as a result of successive reductions. Our auction algorithm for the minimum value problem is described in procedure MinValue. Remark that procedure MinValue applied to a graph becomes the auction algorithm with graph reduction described in [7]. For each node i, the predecessor pred(i) gives the last arc in the best s-i path determined so far; for notational convenience, we consider pred(i) as a set; initially, ;. The label l(i) is the minimum s-i hyperpath value determined so far, which becomes the optimum s-i hyperpath value V (i) at the first scan of node i. We denote by FSA (i) and FSH (i) the arcs and proper hyperarcs in FS(i), respectively; thus Replacing hyperarc a by its reduction ani is denoted by a := an i; note that an i may be an arc. The last node in P is denoted by last(P ). During the execution of the algorithm, the contained graph is the directed graph defined by the nodes and arcs in the current hypergraph Proposition 3.1 At each step of the algorithm, for each node i 2 VA such that gives the shortest s-i path length in the current contained graph HA . Proof: The property follows from the correctness of the auction algorithm for SPT, observing that a new arc (i; j) is created only before the first scan of node i. Theorem 3.1 The vector V determined by the algorithm gives the minimum hy- perpath values in the original hypergraph. Proof: The theorem can be proved by induction considering nodes in order of first scan, that is, in increasing order of value V (\Delta). The claim is clearly true at the beginning because to node s is assigned V Assume that all the previously assigned V are correct at the first scan of node i. It follows from Proposition 3.1 that l(i) is a lower bound on the length of any path in HA from node s to each node j such that Therefore, in the current hypergraph, the value of any hyperpath containing a proper hyperarc cannot be less than l(i). This implies that is correct; as a consequence, Step 1(c) does not change the optimal solution (Proposition 2.1). 3.1 Other Weighting Functions The auction algorithm for minimum value can be easily adapted to the minimum distance problem. To this aim, it suffices to skip the weight update w(a) := w(a) V (i) in Step 1(c). This follows from Property (2.2) since, when a is replaced by an i, for each node j 2 T (a n i). At the end of the algorithm, the vector of optimal distances. The proof of correctness for the distance function is similar to the one of Theorem 3.1. Procedure MinValue(H; s) for each u 2 V: pred(u) := ;, p(u) := 0, V (u) := l(u) := +1; steps (a). (d) (first scan of i) (a) (set value) V (i) := l(i); (b) (delete (c) (reduce hyperarcs) for each a 2 FSH (i): a := a n i, w(a) := w(a) (d) (update labels, delete arcs) for each a = (i; otherwise, deletion, contraction, or expansion) if FSA (i)g: go to Step 4; a=(i;j)2FSA (i) go to Step 1; Step 4 (expansion) expand P by node j i , where: a=(i;j)2FSA (i) go to Step 1. The situation is slightly more complex for travel times. Recall that our goal is to work with the support transit hypergraph, thus we must deal with the corresponding full hypergraph implicitly. To this aim, we replace hyperarc reductions by arc insertion operations, as described below. Consider a proper hyperarc a in the support, with T and We know that it suffices to consider the k reductions a (see Section 2.1.1). At the first scan of node u we compute the value t(a i ); if t(a generate an arc according to Proposition 2.3. Otherwise, i.e., that a i\Gamma1 is the reduction e(a) of a yielding minimum time, and we delete a. If necessary, a is removed at the first scan of node u k . In conclusion, for a proper hyperarc a, up to jT (a)j arcs can be generated. In order to compute each t(a i ) efficiently, for each proper hyperarc a in the support hypergraph we keep two values, initially set to 0: u2T (a) u2T (a) At first scan of u i , it is t(a We also keep a counter k(a) of visited nodes in T (a). We rewrite Step 1(c) as follows: Step 1(c) (reduce hyperarcs) for each a 2 FSH (i): oe(a) k(a) Observe that a new arc is added only if it can be used to improve the label of h(a). In this case, the current predecessor pred(h(a)) will be deleted in Step 1(d); therefore, at most one arc generated from a belongs to the contained graph at the end of Step 1. Lemma 3.1 If the values V (\Delta) assigned before the first scan of node i are correct then the value V (i) assigned at the first scan of node i is correct. Proof: Let S be the set of nodes in the current hypergraph whose first scan occurred before first scan of node i. We know that, for gives the SPT distance from s in the contained graph; moreover, V (i) is a lower bound on the SPT distance for each node u 62 S. Consider a proper hyperarc a in the current support hyper- graph. The reductions of a containing nodes in S only have been already considered by the algorithm, possibly adding new arcs. Moreover, it is from (2) that V (i) is a lower bound on t(e(a)). Therefore V (i) is a lower bound on the minimum time for each node u 62 S in the current hypergraph. The thesis follows. Using Lemma 3.1 we can proof the correctness of the auction algorithm for minimum times by induction, as we did for Theorem 3.1. 3.2 Computational Complexity The auction-reduction algorithm presented in [7] solves the SPT problem on a graph E) in O(jV j minfjEj; jV j log jV jg) time. It is easy to see that the maximum number of arcs generated during the execution of MinValue is m, for the value and distance weighting functions, and O(size(H)) for the time function. Moreover, the total time spent in first scans (Step 1) is O(size(H)). Therefore, we can state the following proposition: Proposition 3.2 The running time of the auction shortest hyperpaths algorithm is log ng), for value and distance, while for the time function it is O(size(H) Two techniques for improving the running time of the auction-reduction method are presented in [7]: path scanning ant multipath restructuring. The resulting complexity is O(jV In fact, the total computation time between two successive first scan operations is O(jV j), and clearly there are at most jV j first scans. The above techniques can be easily applied within our shortest hyperpath algorithm; the next proposition follows: Proposition 3.3 The auction shortest hyperpath algorithm with path scanning or multipath restructuring takes O(size(H)+n 2 ) time, for the value, distance, and time functions. Computational Results In this section we present the preliminary computational results for auction methods for shortest hyperpath problems. Our main goal here is to compare a few variants of the basic method; a complete experimental evaluation of auction shortest hyperpath methods would require a much larger effort. Our basic shortest hyperpath algorithm, denoted by HAR, is an implementation of procedure MinValue where we used the last data structure [7]. A variant of this algorithm, denoted by HAR2, makes use of the "second best" device [4, Chapter 4] too. We implemented a third version, denoted by HARn, where the "second best" device is not used, and a node contraction operation is introduced. A node contraction deletes a node k with indegree and outdegree equal to one: the arcs incident with node k, say (i; are replaced by an arc (i; j), where contraction operations simplify the current graph, and may help keeping the current path shorter. We compared our auction algorithms to an implementation of procedure SBT- heap [11], denoted by SBTh. All algorithms were coded in C language, and run on an IBM RISC-6000 P43 workstation, with 64M RAM, using the AIX operating system. In general, devising a reasonable experimental setup for shortest hyperpaths is not a trivial task, since hypergraphs show many more degrees of freedom than graphs (see e.g. [15]). Here, we restricted ourselves to one weighting function, namely the distance, and we considered two different hypergraph topologies: random and grid. Random hypergraphs do not show any special structures, except that the origin s is a distinguished node, and FS(s) contains only arcs. The size of proper hyperarcs is chosen randomly in the interval [d min ; d max ]. In our experiments, we set d and jF and we defined five classes of random hypergraphs with different values of d a . For proper hyperarcs, and for arcs exiting the root, weights were generated randomly in the interval [0; 1 for the remaining arcs, weights belong to [ 1 This choice has been motivated by the attempt to increase the relevance of hyperarcs. The results for random hypergraphs are shown in Table 1. For each class, the value ffi is the expected size of FS(u) for u 6= s. Execution times are given in milliseconds; each entry is the (rounded) average of 20 instances. m a 2n 25n 2n 25n 50n HARn 176 123 135 150 277 Table 1. Random Hypergraphs In a grid hypergraph nodes are arranged in a b \Theta h grid; a node is identified by its cohordinates Nodes with the same x cohordinate form a level; for each pair (x; y) and (x; y 0 ), with y h, there are two vertical arcs and . Hyperarcs connect nodes in successive levels; for each (x; y) with there exists a hyperarc: h, and y h, In addition, there is an origin node s, and arcs s; (1; y) for each 1 - y - h. We generated three classes of grid hypergraphs: square, where h, long, where b AE h, and high, where h AE b. Parameters b and h where choosen in order to have the same number of nodes in the three classes. Hyperarc weights lay in the interval vertical arcs weights lay in [1; 2]; weights of arcs leaving s lay in [0; 1 Execution times are reported in Table 2. Each entry is the (rounded) average of 5 instances; times are given in seconds. high square long HAR 44 88 86 172 370 785 HARn 44 85 86 176 370 783 SBT 2:22 3:85 1:18 2:06 :65 1:15 Table 2. Grid Hypergraphs Though clearly incomplete, the above results allow us to draw some conclu- sions. For what concerns random hypergraphs, our auction algorithms are comparable to standard label-setting methods, that are the most efficient for this class of hypergraphs[15]. Auction methods become more and more competitive as the density increases; in one case, HAR2 gives the best results. On the other hand, auction methods do not seem to be suitable for large grid hypergraphs. This result (that matches the computational results for auction methods for long grid graphs) is not surprising, since the auction algorithm must maintain a long current path P in order to connect nodes in the last layers. The "second best" data structure gives the best results for random hypergraphs, and for high grids, but it is not suitable for square and long grids. Again, this result is not surprising, since in a grid hypergraph there exist at most two hyperarcs (plus two vertical arcs) leaving each node; it is conceivable that the good results for high grids are due to savings obtained when scanning the origin node. On the contrary, the node contraction operation is almost useless, also for grid hypergraphs. This result is rather disappointing, since in some preliminary experiments this operation proved to be very effective on some classes of grid graphs. A possible explanation may be the following: if a node has the highest distance in the tail of a hyperarc, it is likely to have the highest distance also in the tail of the other hyperarc it belongs to; in this case, hyperarc reduction may create two arcs leaving the node, so that node contraction cannot be applied. This observation may suggest some guidelines for improving our algorithms. Conclusions In this paper, we proposed an auction method for shortest hyperpath problems, that can be adapted to several types of weighting functions. Our method is derived, with minor changes, from the auction-reduction SPT algorithm. Indeed, an appealing feature of our approach is that several techniques originally developed for graphs could be easily exported to hypergraphs. From a practical point of view, auction shortest hyperpath methods are comparable to other known methods, at least in favourable cases. As one would expect, their behaviour can be dramatically affected by the structure of the underlying hy- pergraph; however, this seems to resemble closely what happens for graphs. We can conclude that auction shortest hyperpath methods deserve more inves- tigation, both on the theoretical and the practical side. A possible direction could be adapting some of the variants proposed in the literature, such as the price raise technique devised in [8], and the forward-reverse approach for the single-origin single-destination case [4, Chapter 4]. --R A distributed algorithm for the assignment problem. The auction algorithm: A distributed relaxation method for the assignment problems. An auction algorithm for shortest paths. Linear Network Optimization: Algorithms and Codes. The auction algorithm for the minimum cost network flow problem. The auction algorithm for transportation problems. Polynomial auction algorithms for shortest paths. A modified auction algorithm for the shortest path problem. Dynamic maintenance of directed hypergraphs. Max Horn SAT and the minimum cut problem in directed hypergraphs. Directed hypergraphs and applications. Hypergraph models and algorithms for the assembly problem. Equilibrium traffic assignment for large scale transit networks. Implicit enumeration of hyperpaths in logit models for transit networks. A computational study of shortest hyperpath algo- rithms Strongly polynomial auction algorithms for shortest paths. A hypergraph model for stochastic time dependent shortest paths. A simplicial decomposition method for the transit equilibrium assignment problem. --TR
auction algorithms;shortest paths;directed hypergraphs;hyperpaths
589094
Constraint Qualifications for Semi-Infinite Systems of Convex Inequalities.
We introduce and study the Abadie constraint qualification, the weak Pshenichnyi--Levin--Valadier property, and related constraint qualifications for semi-infinite systems of convex inequalities and linear inequalities. Our main results are new characterizations of various constraint qualifications in terms of upper semicontinuity of certain multifunctions. Also, we give some applications of constraint qualifications to linear representations of convex inequality systems, to convex Farkas--Minkowski systems, and to formulas for the distance to the solution set. Some of our concepts and results are new even in the particular case of finite inequality systems.
Introduction Let I) be a family of convex functions, where I is an arbitrary (but nonempty) index set, and let us consider the system of \convex inequalities" Throughout this paper we shall consider only the above framework, which is sucient for many applications. However, let us mention that some of our results and proofs can be extended to arbitrary (nite or innite dimensional) normed linear spaces X and to inequality systems (1) with convex functions In the sequel we shall assume, without any special mention, that the solution set S of the system (1) is nonempty, i.e., We shall often consider the important particular case when each g i is ane, say where a denotes the dot product of vectors in IR n . In this case, (1) becomes a system of linear inequalities, and (2) becomes Note that one can formally convert (1) to one convex inequality: where G() is the sup-function [7] of (1), dened as i2I The system (1) is said to be a system with nite-valued sup-function, if In this paper, we always assume that (8) holds. For the inequality system (1) and for any x in IR n , we shall denote by I(x) the set of \active indices" at x, i.e., Note that if is the classical denition of active indices. One of the reasons of the diculty of extending the results from nite inequality systems to semi-innite inequality systems is that in the semi-innite case for x 2 bd S the set I(x) may be empty or may be innite. As we shall see in the sequel, some other reasons, which explain why many results cannot be extended at all, or can be extended only under some additional assumptions (and sometimes only with dierent proofs), are the following: while in the nite case the index set I is compact, in our main results on the general semi-innite case we shall assume no topology on I; also, while in the nite case for each x 2 IR n the set A x := fg i (x)j i 2 Ig is closed in IR; and the sup-function G(x) := sup i2I g i (x) is always nite-valued on IR n ; in the general semi-innite case these are no longer true. Furthermore, it is well-known that for a linear inequality system (4) with a nite index set I we have where N S (x) and bd S denote the normal cone of S at x and the boundary of S; respectively. In general, (10) does not hold for a linear inequality system (4) if I is innite. Given a convex system (1), another important well-known property for a nite I is (with the convention [ i2; A where co(A) denotes the convex hull of a set A and @g(x) denotes the subdierential of a convex function g at x : In general, (11) does not hold if I is innite. In the present paper we shall give a detailed discussion of constraint qualications for semi- innite systems of convex inequalities and linear inequalities and the relations among them. We shall introduce and study the Abadie constraint qualication, the weak Pshenichnyi-Levin-Valadier property and related constraint qualications. Our main results are new characterizations of various constraint qualications in terms of upper semicontinuity of certain multifunctions. Also, we shall give some applications of constraint qualications to linear representations of convex inequality systems, convex Farkas-Minkowski systems, and formulas for the distance to the solution set. Moreover, some of our concepts and results on semi-innite convex inequality systems will yield new contributions even when applied to the particular case of nite inequality systems (such as Corollary 2). Let us describe now, brie y, the sections of our paper. It is well-known (see e.g. [7, pp. 307-309]) that in the theory of convex minimization over the solution set of a nite system of convex inequalities the so-called basic constraint qualication, or brie y, the BCQ, which requires that the normal cone at each point of the boundary of the solution set should coincide with \the cone of the active constraints" at that point, plays an important role; for example, it is satised if and only if the Karush-Kuhn-Tucker (KKT) sucient optimality conditions are also necessary for optimality (see e.g. [7, Proposition 2.2.1, page 308]). Recently, the BCQ has been extended to semi-innite linear inequality systems by Puente and Vera de Serio [14], who have used the term \locally Farkas-Minkowski systems", or brie y, LFM systems, and further extended to semi-innite systems of convex inequalities by Goberna and Lopez [5, page 162], who have used the term \convex locally FM systems", or brie y, CLFM systems. In section 2 we shall introduce a weaker constraint qualication than the BCQ, which is dierent from the BCQ even in the particular case of nite convex inequality systems, and which we shall call the Abadie CQ, requiring only that the normal cone at each point of the boundary of the solution set should coincide with the closure of the cone of the active constraints at that point. We shall give new characterizations of the Abadie CQ and the BCQ in terms of upper semicontinuity of certain associated convex cone-valued multifunctions. In section 3 we shall introduce and study the PLV (Pshenichnyi-Levin-Valadier) property and the weak PLV property of a semi-innite convex inequality system at a point x, requiring that the subdierential of the sup-function G() at x should coincide with (respectively, with the closure of) the convex hull of the subdierentials of constraints corresponding to the active indices at that point; when this property holds for all points in the boundary of the solution set, we shall simply use the terms PLV property or weak PLV property, respectively. In the particular case of nite linear inequality systems the BCQ (and hence the Abadie CQ) is always satised, and for nite convex inequality systems so is the PLV property (whence also the weak PLV property) at all points of IR n , but for semi-innite inequality systems the situation is dierent. We shall give new characterizations of the PLV and weak PLV properties in terms of upper semicontinuity of certain associated multifunctions. We shall also show some connections among the PLV, weak PLV properties, the BCQ, and Abadie CQ. In section 4 we shall be concerned with Slater conditions. In contrast with the case of nite systems of convex inequalities, in the semi-innite case two dierent Slater conditions appear in a natural way: the usual one, requiring the existence of a point in the solution set, at which all inequalities of the system are satised as strict inequalities, and the so-called strong Slater condition (following the terminology of [5, page 128]), in which the inequalities of the system are required to be satised uniformly strictly, that is, in which the sup-function of the system is required to satisfy the usual Slater condition. We shall study the connections between the Slater conditions and the constraint qualications discussed in sections 2 and 3; it will turn out that the situation concerning these connections is dierent from that occurring in the case of nite inequality systems. Also, we shall see that in the general semi-innite case the usual Slater condition is too weak. The nal section 5 is devoted to some applications of constraint qualications. Given a system of convex inequalities, we recall that any equivalent system of linear inequalities (i.e., a system of linear inequalities with the same solution set) is called a linear representation of the given convex system. It is well known that linear representations and, especially a certain simple one, which we shall call the \standard" linear representation, are useful tools in the study of convex inequality systems (see e.g. [5] and the references therein). In subsection 5.1 we shall give a new linear representation of (nite or semi-innite) convex inequality systems satisfying the Abadie CQ, which uses a much smaller subset of inequalities of the standard linear representation. Also, we shall show the connections between some properties of the initial convex inequality system and its representation. In subsection 5.2 we shall extend from semi-innite linear inequality systems to semi-innite convex inequality systems the concepts of consequence relations and FM (Farkas-Minkowski) systems and, using the standard linear representation, we shall extend a known relation between linear FM systems and the BCQ, given in [14] and [5], to the case of convex FM systems. We shall also give a direct proof of this result, which, in contrast with the known proof for the linear case, does not use any subset of The exact formulas for the distance of a point to the solution set of a convex inequality system are important, among other reasons, for their connection with \asymptotic constraint qualications" and for obtaining results on error bounds for such a system (see [11]). The well-known general formulas for the distance of a point to a closed convex set are not suciently useful for this purpose, since they do not exploit the special structure of the constraints of the inequality system. Up to the present, only a formula for the distance to the solution set of a semi-innite linear inequality system has been essentially known (for a dual version, see [4] and [19], and for the nite case see [2]). In subsection 5.3, assuming the Abadie CQ or the BCQ, we shall give the rst formulas for the distance of a point to the solution set of a semi-innite system of convex inequalities, which are new even in the nite case. Also, using this result, we shall show that the distance of a point to the solution set of a convex inequality system (1) satisfying the BCQ is equal to the distance of that point to some nite subsystem of (1). We conclude this section by introducing some notations which we shall use in this paper. We shall consider IR n endowed with the usual scalar product h; i; the Euclidean norm kk ; and the topology induced by this norm. For an index set J , jJ j denotes the cardinality of J . Let A be a subset of IR n . Then A; int(A) and bd(A) denote the closure, the interior, and the boundary of are the convex hull and the closed convex hull of A; respectively; are the convex cone and the closed convex cone, generated by vectors in A; is the polar of A, i.e., and A is the bipolar of A; in the particular case when A is a cone, A 0 coincides with the \negative polar of A", i.e., we have The results of the present paper and of [11] have been presented at the "Workshop on Error Bounds and Applications in Mathematical Programming" in Hong Kong, December 8-14, 1998. We wish to thank A. Auslender and A. M. Rubinov for the references [13] and [8], in connection with Theorem 3. We also thank M. A. Lopez for sending us the manuscript [3], which we have received after the present paper had been completed; in order to compare our results with those of [3], we have inserted Remarks 2, 3(a) and 5(a). Finally, we wish to express our gratitude to M. A. Lopez and M. D. Fajardo for their careful reading of the manuscript of the present paper and for their valuable remarks which contributed to its improvement. Constraint Qualication and Basic Constraint Qualication For characterizations of constraint qualications for (1) we consider the convex cone generated by the subdierentials of the active members of G() at x: We use N 0 (x) to denote the closure of N 0 (x). If every g i is an ane function as dened in (3), then is the cone generated by the \active constraints" at x. Thus, in [5], N 0 (x) is called \the cone of active constraints at x". Let T S (x) be the tangent cone of S at x, i.e., T S x). The normal cone of S at x 2 S is dened as It is well-known that N S e.g. [7, Proposition 5.2.4, page 137]). Denition 1 We shall say that the convex inequality system (1) satises (a) the Abadie constraint qualication, or brie y, the Abadie CQ, at a point x 2 bd or equivalently, N S (b) the basic constraint qualication, or brie y, the BCQ, at a point x 2 bd S, if (c) the Abadie CQ (respectively, the BCQ), if it satises the Abadie CQ (respectively, the BCQ) at all points x 2 bd S. Remark 1 (a) The equivalence of the two formulas in (18) follows from the bipolar theorem. In fact, if N S since T S (x) is a closed convex cone, by the bipolar theorem and by e.g. [7, Proposition 5.2.4, page 137]) we have On the other hand, if and the bipolar theorem, and since is a cone, we obtain (b) Since we always have (see e.g. [7, Lemma 4.4.1, page 267] and [7, Lemma 2.1.3, page 305]) the system (1) satises the Abadie CQ at x 2 bd S if and only if and it satises the BCQ at x 2 bd S if and only if Moreover, for x 2 bd S, by (20), (1) satises the Abadie CQ at x if and only if N 0 and (6) satises the Abadie CQ at x. Similarly, by (21), (1) satises the BCQ at x if and only if the BCQ at x. Clearly, (1) satises the BCQ at x if and only if it satises the Abadie CQ at x and N 0 (x) is closed. Thus, (1) satises the BCQ at x if and only if T S closed. The points x 2 bdS with the latter property have been called \Lagrangian regular points" in [12, Denition 3.3]. (c) When I is nite, Denition 1 is the classical denition of the Abadie CQ introduced by Abadie (see [1], also [10]) and, respectively, of the basic constraint qualication (see [7, page 207]). If I is nite and each g i is a dierentiable convex function, then @g i is the gradient of g i at x) and the cone N 0 (x) of (15) is closed (see e.g. [7, Lemma 4.3.3, page 130]). In this case, the Abadie CQ and the BCQ coincide. (d) The BCQ, in an equivalent form, has been introduced for semi-innite linear inequality systems (4) in [14] (called locally Farkas-Minkowski systems, or brie y, LFM systems) and extended to convex inequality systems (1) in [5, pp. 162-163] (called convex LFM constraint qualication). When I is nite and each g i is an ane function, the BCQ, and hence also the Abadie CQ, are satised (see e.g. [7, Example 5.2.6(b), page 138]). When each g i is ane, for semi-innite systems of linear inequalities, the Abadie CQ may not hold and the Abadie CQ is not the same as the BCQ, as shown by the following example. Example 1 Let 2: (a) Semi-Innite Linear Systems Without the Abadie CQ: x 2i 0g. Also, Thus the family (24) does not satisfy the Abadie CQ. (b) Semi-Innite Linear Systems With the Abadie CQ but Without BCQ: Also, Thus the family (25) satises the Abadie CQ, but not the BCQ. Next we give new characterizations of the Abadie CQ and the BCQ for (1), in terms of the upper semicontinuity of the multifunctions N 0 (); cone(@G()), N 0 () and cone(@G()). We recall (see e.g. [16, page 55]) that a multifunction (i.e., a set-valued (the collection of subsets of IR n ) is said to be upper semicontinuous in the sense of Kuratowski, or brie y, upper semicontinuous, at x 2 IR n , if the relations lim k!+1 x Clearly, the graph of Q (i.e., the set f(x; y)j x 2 is closed if and only if Q is upper semicontinuous at all x 2 IR n . We shall rst prove a lemma, in which we shall use the convex hull of the subdierentials of the active members of G(x), that is, the set be a multifunction such that, for all z in a neighborhood of x; Q(z) is a convex set, I(z) 6= ;, and D 0 (z) Q(z). If Q is upper semicontinuous at x, then @G(x) Q(x). Proof. Assume the contrary that there exists y 2 @G(x) n Q(x): Since Q is upper semicontinuous at x and Q(x) is a convex set, Q(x) is a closed convex set. Then, by the strict separation theorem [7, Theorem 4.1.1, page 121], there exists u 2 IR n nf0g such that Let G 0 (x; u) be the directional derivative of G at x in the direction u. By y 2 @G(x) and a well-known formula for G 0 (x; u) (see e.g. [7, page 240]), we get z in a neighborhood of x, without loss of generality we can assume I(x+t k u) 6= ;. Let Then whence, by t k is bounded, the set fy is bounded as well (see e.g. [7, Proposition 6.2.2, page 282]). Hence, we may assume, without loss of generality, that y y. Then, letting k ! +1 in (28) and using (27), we obtain hb But, since lim and since Q is upper semicontinuous at x, we have b which contradicts (29). 2 Theorem 1 Let x 2 bd S and I(z) 6= ; for z in a neighborhood of x. Then the following two statements are equivalent. (a) (1) satises the Abadie CQ at x. (b) Both cone(@G()) and N 0 () are upper semicontinuous at x. Proof. (a))(b): Let x y. We claim that To prove the claim, let We consider two cases: Case 1: (6) does not satisfy the Slater condition, that is, G(z) 0 for all z 2 IR n . Then . Thus, (31) implies Letting in the above inequality we obtain That is, y 2 N S (x). Case 2: (6) satises the Slater condition, that is, G(^x) < 0 for some ^ S. Then there is k 0 > 0 such that z 2 S k for k k 0 By (31), we have Since (32) holds for any z 2 int S, it also holds for z 2 int If (1) satises the Abadie CQ at x, we get y 2 N S Therefore, both N 0 () and cone(@G()) are upper semicontinuous at x. y. Then x 0 62 S (since x would imply hy; xi hy; x which is impossible) and hence and assume k large enough so that 1 be such that k. Then, since x 2 S k , we have Thus, fx k g is a bounded sequence. Without loss of generality we may assume that x in (34) we obtain On the other hand, since x 0 x is the projection of x 0 onto S. Hence, imply that Consequently, x y. We claim that x 0 x k k we have x and x 0 x satises the Slater condition and x e.g. [7, Theorem 1.3.5, page 245]). Thus, x 0 x which proves the claim. by the upper semicontinuity of cone(@G()) at x we get y 2 cone(@G(x)). Hence, since y was an arbitrary nonzero element in On the other hand, by the upper semicontinuity of N 0 () at x and (applied to we have @G(x) N 0 (x), which implies cone(@G(x)) N 0 (x). Therefore, we have (22), and thus (1) satises the Abadie CQ at x. 2 Theorem 2 Let x 2 bd S and I(z) 6= ; for z in a neighborhood of x. Then the following two statements are equivalent. (a) (1) satises the BCQ at x. (b) Both cone(@G()) and N 0 () are upper semicontinuous at x. Proof. (a))(b): By Theorem 1, both cone(@G()) and N 0 () are upper semicontinuous at x. Since closed, we know that N 0 (x) and cone(@G(x)) must be closed. Thus, both cone(@G()) and N 0 () are upper semicontinuous at x. (b))(a): By Theorem 1, (1) satises the Abadie CQ at x. But (b) also implies that N 0 (x) is a closed set. So N 0 satises the BCQ. 2 From Theorems 1 and 2 we obtain the following characterizations of the Abadie CQ and the BCQ. Corollary 1 Suppose that I(z) 6= ; for z in a neighborhood of bdS. Then the following two statements are true. (a) (1) satises the Abadie CQ if and only if both cone(@G()) and N 0 () are upper semicontinuous at every x in bdS. (b) (1) satises the BCQ if and only if both cone(@G()) and N 0 () are upper semicontinuous at every x in bd S. Remark 2 Fajardo and Lopez [3, Theorem 3.1(i)] proved that if (1) satises the BCQ, then the multifunction A(x) := fy 2 N 0 (x)j kyk 1g (B)-upper semicontinuous on S (which is equivalent to the upper semicontinuity of N 0 () on S) and that if (6) satises the Slater condition, then the converse is also true [3, Theorem 3.1(iib)]. Here a mapping is said to be (B)-upper semicontinuous at x 2 IR n , if for every open set W in IR n containing Q(x) there exists a neighbourhood V (x) of x such that Q(z) W for each z 2 V (x); furthermore, Q is said to be (B)-upper semicontinuous on a set M IR n if it is (B)-upper semicontinuous at each x 2 M: It is well-known (see e,g, [5, p. 128]) that if Q is (B)-upper semicontinuous (at x) then it is upper semicontinuous (at x); but the converse is not true. One can show that if the set fzj z 2 Q(y) for y in some neighborhood of xg is bounded, then the (B)-upper semicontinuity of Q() at x is equivalent to the upper semicontinuity of Q() at x. 3 Subdierentials of the Sup-Function and Its Active Member In this section, we study the relations between the subdierential of the sup-function G and the subdierentials of its active member functions fg I(x)g. We shall use the set D 0 (x) dened by (26). Note that We shall denote by D 0 (x) the closure of the set D 0 (x): One important property of (1) when I is nite is the equality (11) (see e.g. [13, Theorem 1.4]), which can be rewritten as The above equality means that the subdierential of the sup-function G() is the convex hull of subdierentials of its active members. In general, (37) does not hold if I is innite. The following sucient (but not necessary) condition that guarantees , and is satised when I is nite, has been given by Levin [8, Theorem 2] (and, at the same time, Valadier [18, Theorem 2] has obtained, in an arbitrary topological linear space instead of IR n ; the weaker conclusion in which D 0 is replaced by D 0 Theorem 3 (Pshenichnyi-Levin-Valadier Theorem [13, 8, 18]) If I is a compact set (in some metric space), and is a family of convex functions such that for each xed upper semicontinuous on I, then (37) holds at all x 2 IR n . Here a real-valued function f(t) is said to be upper semicontinuous at lim i.e., lim sup Note that (39) is equivalent to the upper semicontinuity of the multifunction F fy 2 IRj y f(t)g at t 0 . For convenience, we shall introduce the following denition related to (37) (where PLV stands for \Pshenichnyi-Levin-Valadier"). Denition 2 Let (1) be a convex inequality system with a nite-valued sup-function. We shall say that the family fg or the system (1), has (a) the weak PLV property at a point x 2 (b) the PLV property at a point x 2 (c) the weak PLV property (respectively, the PLV property), if it has the weak PLV property (re- spectively, the PLV property ) at all x 2 bd S: One major problem with an innite I is the possibility of (which can not happen when I is nite). For example, given a family of convex functions holds if and only if However, the active index set for (42) is empty at every x 2 bd S and so there is no (weak) PLV property for (42). One can avoid the inconvenience of having by requiring the closedness of fg i (x)j i 2 Ig. Proposition 1 Let x 2 bd S: If the set A x := fg i (x)j i 2 Ig is closed in IR, then I(x) 6= ;: Proof. Since x 2bd S; we have (by the continuity of G): Taking any sequence fi k g I such that g i k by the closedness of A x we obtain 0 2 A x ; so I(x) 6= ;: 2 However, even if I(x) 6= ;, (4) (and hence, in general, (1)) might not have the weak PLV property. Also, in general, the weak PLV property is not the same as the PLV property. (a) A Semi-Innite Linear System Without Weak PLV Property: Then k1 4x if x 0: f3g. Therefore, the system does not have the weak PLV property at (b) A Semi-Innite Linear System With Weak PLV Property, But Without PLV Property: Then k1 x if x < 0; 5x if x 0: 5). Therefore, the system has the weak PLV property at but not the PLV property. There is no relation of implication between the CQs and the PLV properties. Indeed, in Example 2(a), the system actually satises the BCQ at but the PLV property does not hold at while for any I with one index the PLV property is trivially true, but (1) may not satisfy the Abadie CQ. However, when (1) satises the PLV property (in particular, when I is nite), and the characterizations for the Abadie CQ and the BCQ can be simplied. Corollary 2 Suppose that (1) satises the PLV property and I(x) 6= ; for x in a neighborhood of bd S (e.g., this happens when I is nite). Let x 2 bd S. Then the following two statements are true. (a) (1) satises the Abadie CQ at x if and only if N 0 () is upper semicontinuous at x. (b) (1) satises the BCQ at x if and only if N 0 () is upper semicontinuous at x. Proof. By the assumption, D 0 all x, which implies N 0 x. Thus Corollary 2 follows from Corollary 1. 2 The following proposition shows that if (1) satises the (weak) PLV property, then (1) and (6) are \equivalent" in terms of CQ's. Proposition 2 Let x 2 bdS. (a) Suppose that (1) satises the weak PLV property at x. Then (1) satises the Abadie CQ at x if and only if (6) satises the Abadie CQ at x. (b) Suppose that (1) satises the PLV property at x. Then (1) satises the BCQ at x if and only if (6) satises the BCQ at x. Proof. (a) If (1) satises the weak PLV property at x; then, by (40) and (36), only if N S (b) If (1) satises the PLV property at x; then, by (41) and (36), only if N S It turns out the the weak PLV property or the PLV property at a point x 2 IR n can be characterized by the upper semicontinuity at x of the multifunction D 0 () or D 0 (), respectively. Theorem 4 Let x 2 IR n and I(z) 6= ; for z in a neighborhood of x. Then only if D 0 () is upper semicontinuous at x. Proof. If D 0 () is upper semicontinuous at x, then, by Lemma 1, @G(x) D 0 (x). Since @G(x) holds, we have Next we assume that prove the upper semicontinuity of D 0 () at x. Let lim k!+1 x we have y k 2 @G(x k G is a nite convex function, @G(z) 6= ; for all z 2 IR n and @G is upper semicontinuous (see e.g. [7, Proposition 6.2.1, page 282]). Thus, y 2 @G(x). Since upper semicontinuous at x. 2 Theorem 5 Let x 2 IR n and I(z) 6= ; for z in a neighborhood of x. Then only if D 0 () is upper semicontinuous at x. Proof. Note that D 0 () is upper semicontinuous at x if and only if D 0 () is upper semicontinuous at x and D 0 (x) is a closed set. If D 0 () is upper semicontinuous at x, then, by Theorem 4, On the other hand, if then by the upper semicontinuity of @G(), D 0 (x) is a closed set and Hence, by Theorem 4, D 0 () is upper semicontinuous at x. Therefore, D 0 () is upper semicontinuous at x. 2 Using Theorems 4 and 5 we obtain the following characterizations of the weak PLV and PLV properties at all x 2 IR n . Theorem 6 The following statements are true. (a) semicontinuous on only if D 0 (x) 6= ; for all x 2 IR n and D 0 () is upper semicontinuous on Proof. Since G() is a nite convex function, @G(x) 6= ; for any x. Therefore, every condition in the above theorem implies I(x) 6= ; for all x 2 IR n . Consequently, Theorem 6 follows from Theorems 4 and 5. 2 Remark 3 (a) Fajardo and Lopez [3, Theorem 4.1(i)] proved that if D 0 (x) 6= ; for all x 2 IR n and (B)-upper semicontinuous on IR n (see Remark 2 above), then Since the (B)-upper semicontinuity of D 0 () is equivalent to the upper semicontinuity of D 0 () (see Remark 2), the \if" part of Theorem 6(a) is equivalent to [3, Theorem 4.1(i)]. (b) Using Theorem 6, we can give a new proof of Theorem 3, which seems simpler and more natural than the proofs known in the literature (see e.g. the proof of Theorem 4.4.2 [7, page 267 ]. To this end, let us rst prove the following fact: If I is a compact metric space and if is a family of convex functions such that for each x 2 IR n the function upper semicontinuous on I, then the set-valued mapping x ! I(x) is upper semicontinuous on IR n and the set-valued mapping upper semicontinuous on W; where W := f(x; i)j i 2 I(x)g. i.e., if lim k!+1 x y with First we prove by contradiction. In fact, if by the upper semicontinuity of g i (x) with respect to i, there exist a positive constant and a neighborhood O( ^ i) of ^ i in I such that (x). By the assumptions, ^ are continuous convex functions. Thus, But (45) implies that ^ G(^x) G(^x) , a contradiction to (46). This proves that Now, for any z 2 IR n , we have Thus, letting k ! +1 in (47) and using lim k!+1 g i k (by the assumption of upper semicontinuity of the mapping Since (49) holds for any z, we have ^ Finally, in order to prove Theorem 3 it will be sucient, by Theorem 5, to prove that under the assumptions of Theorem 3 the mapping D 0 () has closed graph and I(x) 6= ; for all x 2 IR n . Let By the denition of D 0 there exist theorem (see e.g. [7, Theorem 1.3.6, page 98]) we may assume, without loss of generality, that m k n + 1. Since G() is nite-valued and x x, the set fy 2 @G(x k )j bounded (see e.g. [7, Proposition 6.2.2, p. 282]). By y j;k 2 @g i j we know that fy j;k a bounded set. Since I is compact and fm k g, fy j;k are all bounded with respect to k, by repeatedly selecting subsequences we may assume, without loss of generality, that m +1. By the fact proved above, we know that and y (x). Thus, which proves that D 0 () has closed graph. Finally, since I is compact and i ! g i (x) is upper semicontinuous, we have I(x) 6= ; for all x 2 IR n . This provides an alternative proof of Theorem 3. 4 Slater Conditions If there exists x 2 IR n such that then (1) is said to satisfy the Slater condition. Let us recall the following well-known result, which gives a sucient condition for the BCQ of (1) or (6). Proposition 3 ([7, Theorem 1.3.5, page 245] and [7, Remark 1.3.6, page 246]) If I is nite and (1) satises the Slater condition, then (1) satises the BCQ. In particular, if (6) satises the Slater condition, then (6) satises the BCQ. Remark 4 (a) If (6) satises the Slater condition, then (1) also satises the Slater condition. But the converse is not true. The Slater condition for (6) is sometimes called the strong Slater condition for the convex system (1) (see e.g. [5, page 128]). However, the term \strong Slater condition" is also used in the literature in other senses (see e.g. Lewis and Pang [9], where \strong Slater condition" means that 0 does not belong to the closure of the set @G(G 1 (0)); and, for a dierent sense, see [7, Denition 2.3.1, page 311]). (b) When I is nite, (1) satises the Slater condition if and only if (6) satises the Slater condition. Proposition 4 Suppose that (6) satises the Slater condition. (a) If (1) has the weak PLV property, then (1) satises the Abadie CQ. (b) If (1) has the PLV property, then (1) satises the BCQ. Proof. (a) By Proposition 3, the Slater condition for (6) implies the BCQ for (6). Hence, by Proposition 2(a), we have the Abadie CQ at all x 2 bd S. (b) The proof is similar, using Proposition 2(b). 2 Remark 5 (a) Fajardo and Lopez [3, Theorem 4.1(ii)] proved that if (6) satises the Slater condi- tion, D 0 () is (B)-upper semicontinuous (see Remark 2), D 0 (x) closed for each x in S, then (1) satises the BCQ. But the (B)-upper continuity of D 0 () is equivalent to the upper continuity of D 0 () (see Remark 2), so this result also follows from Theorem 6(a) and Proposition 4(a). (b) The assumptions in (a) and (b) of Proposition 4 cannot be omitted, as shown by Example 1(a), in which the Slater condition for (1) or (6) is satised (in fact, for but the Abadie CQ is not satised. When I is a nite set, the PLV property always holds. In this case, the Slater condition, the BCQ, and the Abadie CQ are all dierent. We recall that for a convex system (1) a solution x 2 S is called a Slater point if we have (50). Proposition 5 If for each x 2 S the active index set I(x) 6= ;, then every Slater point of (1) is a Slater point of (6) (and hence, in this case, the Slater condition for (1) and the Slater condition for are equivalent). Proof. Let x be a Slater point of (1), i.e., let x be a point such that (50) holds. If x were not a Slater point of (6), i.e., if we had sup i2I g i since I(x) 6= ;, there would exist i 0 2 I such that g i 0 a contradiction to the assumption (50). 2 Using Theorem 3, one can give a stronger condition which ensures the BCQ. Indeed, combining Theorem 3 and Proposition 4, we obtain the following result, which has been proved with a more complicated method by Lopez and Vercher [12, Theorem 3.8]. Corollary 3 If I is a compact set (in some metric space), is a family of convex functions such that for each xed x 2 IR n the function upper semicontinuous on I, and (1) satises the Slater condition, then the BCQ holds for (1). Proof. Since I is compact and i ! g i (x) is upper semicontinuous, I(x) 6= ; for any x. By Proposition 5, (6) satises the Slater condition. By Theorem 3, the PLV property holds. Thus, the corollary follows from Proposition 4. 2 Even though we stated Corollary 3 in terms of the Slater condition of (1), obviously the Slater condition of (6) is also satised. In general, the Slater condition for (1) is not very meaningful if (6) does not satisfy the Slater condition. One might wonder whether we should use the following stronger version of (50): where is a positive constant. In the case that G(x) < +1 for x 2 IR n , (51) is nothing more than the Slater condition for (6). If one allows G(x) to be +1, then (51) does not provide any useful information about the system as shown in the following example. Example 3 (a) For fg Then (51) holds with g i and being replaced by g i and 1, respectively. Note that x 2 S if and only if g i (x) 0 (i 2 I). Also the sets I(x) and N 0 (x) remain unchanged for x 2 bd S. Thus, (1) satises the Abadie CQ (respectively, the BCQ) if and only if so does the system This example shows that replacing (50) by (51) without requiring a nite-valued G does not give any new information about the underlying system. (b) Or we could make the situation worse. For example, Then we always have It is easy to see that x 2 S if and only if 1. In this case, the active index set is always empty for any x 2 bdS. Thus, it is not possible to study S by using N 0 (x). This example shows that (51) without requiring a nite-valued G(x) could be a meaningless condition; while (51) with a nite-valued G(x) means that (6) satises the Slater condition, which is useful for constraint qualication properties of (1) (see Proposition 4). Applications 5.1 Linear Representation of Convex Systems Given a semi-innite convex inequality system (1), we recall that a semi-innite linear inequality system, a is said to be a linear representation of the system (1), provided that x is a solution of (1) if and only if x is a solution of (54) (i.e., provided that the systems of inequalities (1) and (54) are equivalent). Each linear representation (54) of (1) is also called a linear system associated to the convex system (1). It is well-known (see e.g. the proof of Theorem 5.2 in [6]) that the system, is a linear representation of the convex system (1), which we shall call the standard linear representation of (1). For the sake of completeness, we include here the proof. If x 2 so x satises (55). Conversely, if x satises (55), then Taking here which completes the proof. A natural question is whether we can use a smaller subsystem of (55) to get a linear representation of (1). In particular, we study when the following semi-innite linear system, is a linear representation of (1). Note that (56) is indeed a subsystem of (55), since for z 2 bd S, Theorem 7 (a) If the convex system (1) satises the Abadie CQ, then the system (56) is a linear representation of (1). (b) If the convex system (1) satises the Slater condition (50) and I(x) 6= ; for all x 2 S, then the system (56) is a linear representation of (1). Proof. Obviously, if x 2 S, then, since @g i (z) N S (z) for i 2 I(z) and z 2 bd S, (56) follows from (z). Thus, every solution of (1) is a solution of (56). In order to prove that (56) is a linear representation of (1), it is sucient to show that if x 62 S, then (56) does not hold. (a) First we prove the following more general result: Suppose that for any x 2 bd S and y 2 N S (x) n f0g, there is a vector ^ that linear representation of (1). Let x 62 S and let z be the projection of x onto S. Then z 2 bdS, x z 6= 0, and x z 2 N S (z) [7, Theorem 3.1.1, page 117]. If hy for all y 2 N 0 (z), a contradiction to the assumption there is ^ Hence, x does not satisfy (56). This proves that x 2 S if and only if x satises (56), i.e., (56) is a linear representation of (1). Now, if the convex system (1) satises the Abadie CQ, i.e., N S then it is trivially true that for any x 2 bd S and y 2 N S (x) n f0g, there is a vector that (indeed, it is enough to take ^ suciently close to y): So (56) is a linear representation of (1). (b) By Proposition 5, (6) satises the Slater condition. Let x 62 S, and let x 2 S be such that contains exactly one point, say z. Then there is a positive constant such that by the denition of @G(z), we have xi G(z) Since > 0, it follows from (57) and (58) that 1 Thus, for any x 62 does not hold. So (56) is a linear representation of (1).Remark 6 (a) In [12, the proof of Theorem 4.5], it has been observed that if (1) satises the assumptions of Corollary 3, then the linear inequality system, is a linear representation of (1). Let us observe that this follows also from Corollary 3 and Theorem under the assumption of Corollary 3, hence (59) is equivalent to (56). (b) In the particular case when all are convex and dierentiable, Theorem 7(b) (even with a smaller subsystem of (56), obtained by choosing for each x 2bd S; with the aid of the axiom of choice, an index i(x) 2 I(x) and an y i(x) 2 @g i (x)) has been shown, essentially, in the proof of Theorem 5.4 in [6]. Some connections between the inequality systems (1) and (56) are given in the following proposition Proposition 6 Let (56) be a linear representation of the convex inequality system (1). Then (a) (1) satises the Abadie CQ (respectively, the BCQ) if and only if so does (56). (b) Denoting by G and G 0 the sup-functions of (1) and (56) respectively, we have Proof. (a) Let (56) be a linear representation of (1) and let x 2 bd S: Since (1) and (56) have the same solution set and hence the same normal cone N S (x) at x; it will be enough to show that Since the linear system (56) is a subset of the linear system (55), we have Furthermore, by [5, proof of Theorem 10.7], there holds Finally, from the denitions it is obvious that which, together with (62) and (63), yields (61). (b) By the denitions of the sup-function and of @g i (z) and I(z); we have, for any x 2 z2bd sup i2I Remark 7 (a) The inequality in Proposition 6(b) may be strict. (b) From Proposition 6(b) above it follows that if (1) satises the Slater condition, then so does also, if (6) satises the Slater condition, then G 0 (x) < 0 for some x 2 However, the converse statements are not true. 5.2 Convex Farkas-Minkowski Systems Related to the BCQ are the convex Farkas-Minkowski systems, dened as follows. Denition 3 (a) A linear inequality where a will be called a consequence relation of the convex inequality system (1), if every x 2 S satises (65): (b) The system (1) will be called a convex Farkas-Minkowski (or brie y, a convex FM) system, if every linear consequence relation of system (1) is also a consequence relation of some nite subsystem of (1). Remark 8 In the particular case of a linear inequality system (4), the above denition reduces to the usual denition of consequence relations and FM systems [5]. One can extend some results on linear FM systems to convex FM systems. For example, the fact that a linear inequality system (4) satisfying the BCQ and with bounded solution set S is an FM system (see [5, Exercise 5.7]), admits the following extension. Proposition 7 A convex inequality system (1) satisfying the BCQ and with bounded solution set S is a convex FM system. Proof. By the above proof of Proposition 6(a), (1) satises the BCQ (if and) only if so does its standard linear representation (55). Furthermore, since (55) is a linear inequality system having the same bounded solution set it is an FM system (see e.g. [5, Exercise 5.6]). Finally, let us show that if (55) is an FM system, then so is (1). Indeed, let where a since (55) is an FM system and has the same solution set there exists a nite subsystem of (55), say where jJ j < +1; such that (65) is a consequence relation of (67). Let S J be the solution set of the nite subsytem of (1) and let x so x is a solution of (67). Hence, since (65) is a consequence relation of (67), we obtain ha which, since x 2 S J was arbitrary, proves that (1) is an FM system. 2 Remark 9 The proofs of some results on linear systems, given in [5], use certain cones of IR n+1 associated to (4). However, let us observe that one can give, directly for the extensions of those results to convex systems, proofs which are new even for the case of linear inequality systems, and do not use any subsets of us give such a proof of Proposition 7. Assume that S is bounded and (1) satises the BCQ, and let (65) be a consequence relation of (1). Let c be the smallest number such that ha 0 ; xi c is still a consequence relation of (1) (such a number exists, since otherwise S fx 2 a contradiction to the general assumption made in this paper). We claim that there exists z 2 S such that ha Indeed, by the denition of c; we have cg and for each Then, since S is bounded and closed, hence compact, fx k g has a subsequence converging to some which proves our claim. By the above, we have S fx 2 Hence, by the BCQ, there exists a nite subset J of I(z); such that a Let Then, by (70) and J I(z), we have a 0 2co([ i2J @g i (z)) N S J (z); that is, ha Thus, the inequality ha 0 ; xi c; whence also (65), is a consequence relation of the nite subsystem of (1), which completes the proof. Combining Proposition 7 and Corollary 3, there results the following corollary, which has been proved with more complicated methods in [12, Theorem 4.5]. Corollary 4 If I is a compact set, is a family of convex functions such that for each xed x 2 IR n the function upper semicontinuous on I, (1) satises the Slater condition, and the solution set S is bounded, then (1) is an FM system. The denition of a convex FM system given in [12] referees to its standard linear representation being a FM system. Under the assumptions of Corollary 4, both denitions are equivalent. 5.3 The Distance to the Solution Set of a Convex Inequality System Theorem 8 Let x 2 IR n nS and ^ x be the projection of x onto S. (a) If (1) satises the Abadie CQ, then I 0 I(^x) sup xi: (72) (b) If (1) satises the BCQ, then I 0 I(^x) sup xi: Proof. By [17, Remark 8b)], we have (a) By the Abadie CQ at ^ x, we have y 2 N S (^x) with only if there exist I k 1 such that y. Thus, (72) is equivalent to (74). (b) By the BCQ at ^ x, we have y 2 N S (^x) with only if there exist I 0 I(^x), 1 such that y. Thus, (73) is equivalent to (74). 2 Remark 11 The assumption of Abadie CQ may be too strong, but at least the assumption I(b x) (which is implied by the Abadie CQ) is necessary in order to have (72). Indeed, if I(b the right hand side of (72) is meaningless. When applied to the semi-innite linear system (4), @g j means that and thus Theorem 8 reduces to the following form. Corollary 5 Let x 2 IR n nS and ^ x be the projection of x onto S. (a) If (4) satises the Abadie CQ, then I 0 (b) If (4) satises the BCQ, then Remark 12 (a) By a well-known theorem of Caratheodory (see e.g. [15, Corollary 7.1(i), page 94]), in each positive combination may assume that fy j jj 2 I 0 g is linearly independent. That is, fy is linearly independent; i 0 (i 2 I 0 ) fy j jj 2 I 0 g is linearly independent: (78) and hence we also have the following representation of N 0 (x): One could rewrite the distance formulas (72), (73), (75), and (76), based on either (77) or (79). (b) In the particular case when I is nite, (4) satises the BCQ (see the observation before Example 1), and hence Corollary 5(b) reduces to [17, Remark 8(a)]. The following theorem shows that if the BCQ holds, then the distance of a point to the solution set S of an arbitrary convex inequality system (1) is equal to the distance of that point to the solution set of some nite subsystem of (1). Theorem 9 If (1) satises the BCQ and b x is the projection of x onto S, then there exists with Jj < +1; such that where J)g: Proof. Choose any J I(b x) for which the rst max in (73) is attained. Then, applying Theorem 8 to the inequality system and to its solution set S J (of (81)), we obtain dist (x; S which, since S S J (by (2) and (81)), yields (80). 2 Remark 13 In the particular case when I is nite and each g i is an ane function, (1) satises the BCQ, and hence Theorem 9 yields [2, Corollary 1.1]. --R On the Kuhn-Tucker theorem The distance to a polyhedron. On systems of linear inequalities. Application of a theorem of E. bounds for convex inequality systems. Abadie's constraint quali Asymptotic constraint quali Convex programming in a normed space Theory of linear and integer programming. The theory of best approximation and functional analysis. Duality for optimization and best approximation over Generalizations of some fundamental theorems in linear inequalities. --TR --CTR Chong Li, Strong uniqueness of the restricted Chebyshev center with respect to an RS-set in a Banach space, Journal of Approximation Theory, v.135 n.1, p.35-53, July 2005 Mara J. Cnovas , Marco A. Lpez , Juan Parra, Stability in the Discretization of a Parametric Semi-Infinite Convex Inequality System, Mathematics of Operations Research, v.27 n.4, p.755-774, November 2002
convex Farkas-Minkowski systems;distance formulas;constraint qualifications;semi-infinite inequality systems
589096
Differential Stability of Two-Stage Stochastic Programs.
Two-stage stochastic programs with random right-hand side are considered. Optimal values and solution sets are regarded as mappings of the expected recourse functions and their perturbations, respectively. Conditions are identified implying that these mappings are directionally differentiable and semidifferentiable on appropriate functional spaces. Explicit formulas for the derivatives are derived. Special attention is paid to the role of a Lipschitz condition for solution sets as well as of a quadratic growth condition of the objective function.
Introduction Two-stage stochastic programming is concerned with problems that require a here- and-now decision on the basis of given probabilistic information on the random data without making further observations. The costs to be minimized consist of the direct costs of the here-and-now (or first stage) decision as well as the costs generated by the need of taking a recourse (or second stage) decision in response to the random environ- ment. Recourse costs are often formulated by means of expected values with respect to the probability distribution of the involved random data. In this way, two-stage models and their solutions depend on the underlying probability distribution. Since this distribution is often incompletely known in applied models, or it has to be approximated for computational purposes, the stability behaviour of stochastic programming models when changing the probability measure is important. This problem is studied in a number of papers. We only mention here the surveys [13], [37] and the papers [1], This research is supported by the Deutsche Forschungsgemeinschaft [12], [17], [24], [25], [31] and [32]. The paper [1] contains general results on continuity properties of optimal values and solutions when perturbing the probability measures with respect to the topology of weak convergence. Quantitative continuity results of solution sets to two-stage stochastic programs with respect to suitable distances of probability measures are derived in [24] and [25]. Asymptotic properties of statistical estimators of values and solutions to stochastic programs are derived in [17], [31], [32]. They are based on directional differentiability properties of the underlying optimization problems with respect to the parameter that carries the randomness ([17], [32]) or the probability measure ([31]). These directional differentiability results for values (in [32]) and solutions (in [17], [31]) lead to asymptotic results via the so-called delta-method . For a description of the delta-method we refer to Chapter 6 in [26], [32], to [33] for an up-to-date presentation and to [15] for a set-valued variant. These papers illuminate the importance of the Hadamard directional differentiability (for single-valued functions) and of the semidifferentiability (for set-valued mappings) in the context of asymptotic statistics. The present paper aims at contributing to this line of differential stability studies. The results in [17], [31] apply to fairly general stochastic optimization models, but impose conditions that are rather restrictive in our context. The present paper deals with special two-stage models and, using structural properties, avoids certain assumptions that complicate or even prevent the applicability of those general results to two-stage stochastic programs. Such assumptions are the (local) uniqueness of solutions and differentiability properties of perturbed problems, which are indispensable in [17], [31]. Before discussing this in more detail, let us introduce the class of two-stage stochastic programs, we want to consider: is a nonempty closed convex set, A is a (s; m)-matrix and Q - is the expected recourse function with respect to the (Borel) probability measure - on IR s , Z ~ ~ Here m are the recourse costs, W is an (s; - m)-matrix and called the recourse matrix, and ~ corresponds to the value of the optimal second stage decision for compensating a possible violation of the (random) constraint To have the problem (1.1) - (1.3) well-defined, we assume Z first moment). The assumptions (A1) and (A2) imply that ~ Q is finite, convex and polyhedral on . Due to (A3) also Q - is finite and convex on IR s (cf. [14], [36]). Observe that, in general, an expected recourse function Q - may be nondifferentiable on a certain union of hyperplanes in IR s and that, indeed, differentiability properties of Q - depend on the degree of smoothness induced by the measure - (cf. [14], [19], [35], [36] and Remark 4.8). Another observation shows that the uniqueness of solutions to (1.1) is guaranteed only if the constraint set C picks just one element from the relevant level set of g(\Delta) +Q - \Delta). This set may be large since Q - \Delta) is constant on translates of the null space of the matrix A (see Example 1.1 in [25]). Proposition 2.1 below provides some more insight into the structure of the solution set to (1.1) and elucidates the role of the set-valued mapping oe(y) yg in this respect. Note that assumption (A1) could be relaxed by introducting the set +1g. Then (A2) and (A3) imply that K is a closed convex polyhedron and that Q - is convex and continuous on K (cf. [36]). Now (A1) can be replaced by the condition K ' A(C) (relatively complete recourse), and much of the work done in this paper carries over to this more general setting by using spaces of functions defined on K instead of IR s . Let K C denote the set of all convex functions on IR s which forms a convex cone in the space C 0 (IR s ) of all continuous functions on IR s . K C will serve as the set of possible perturbations of the given expected recourse function Q - 2 K C . We define and regard ' and / as mappings from K C into the extended reals and the set of all closed convex subsets of IR m , respectively. In this paper we develop a sensitivity analysis for the mappings ' and / at some given function Q - . The stochastic programming origin of the model (1.1) takes a back seat and our results are stated in terms of general conditions on Q - and its perturbations Q. We identify conditions such that the value function ' has first- and second-order directional derivatives and the solution-set mapping / is directionally differentiable at admissible directions. Here, admissibility means that the direction belongs to the radial tangent cone to K C at Q - , i.e., ensuring that the difference quotients are well-defined. For v belonging to T r (K C the Gateaux directional derivatives of ' and / at Q - and (Q - tively, are defined as if the limits exist. The third limit is understood in the sense of (Painlev'e-Kuratowski) set convergence (e.g. [2]). Recall that the lower and upper set limits of a family (S t ) t?0 of subsets of a metric space (X; d) are defined as lim inf lim sup Both sets are closed and the lower set limit is contained in the upper limit. If both limits coincide, the family (S t ) t?0 is said to converge and its limit set is denoted by lim For sequences of sets (S n ) n2IN the definitions of set limits are modified correspondingly. We also derive conditions implying that the limits defining the directional derivatives exist uniformly with respect to directions v belonging to compact subsets of certain functional spaces. The limits are then called (first- or second-order) Hadamard directional derivatives and semiderivatives for set-valued maps, respectively. The corresponding directional derivatives are defined on tangent cones to the cone of convex functions in certain functional spaces. For more information on concepts of directional differentiability and multifunction differentiability we refer to [5], [30] and to [2], [4], [21], [23], respectively. Let us fix some notations used throughout the paper. k \Delta k and h\Delta; \Deltai denote the norm and scalar product, respectively, in some Euclidean space IR n ; B(x; r) denotes the open ball around x 2 IR n with radius r ? 0; d(x; D) denotes the distance of x 2 IR n to the set D ' IR n ; for a real-valued function f on IR n , rf denotes its gradient in IR n and the its Hessian; if f is locally Lipschitzian near x 2 IR n , @f(x) denotes the Clarke subdifferential of f at x; f 0 (x; d) denotes the directional derivative of f at x in direction d if it exists; for denotes the tangent cone to C at x, i.e., cl stands for closure; for denotes the second order tangent set to C at x in direction -, i.e., T 2 (C; x; closed and convex; see [10] for further properties). In our paper, we use the following linear metric spaces of real-valued functions on The space C 0 (IR s ) of continuous functions on IR s equipped with the distance d1 (f; ~ f) =X \Gamman kf \Gamma ~ , where kyk-r jf(y)j, for f; ~ the space C 0;1 (IR s ) of locally Lipschitzian functions on IR s with the metric d L (f; ~ f) =X \Gamman kf \Gamma ~ , where y the space C 1 (IR s ) of continuously differentiable functions on IR s with the metric d(f; ~ d1 (f; ~ and the space C 1;1 (IR s ) of functions in C 1 (IR s ) whose gradients are locally Lipschitzian on IR s equipped with the distance d(f; ~ d1 (f; ~ f) for all f 2 C 1;1 (IR s ). The sensitivity analysis of the mappings ' and / is carried out by exploiting structural properties of the optimization model (1.1). We obtain novel differentiability properties of solution sets and extend our earlier results on directional differentiability of optimal values in [12] considerably. As one might expect, the basic ingredients of our analysis are a Lipschitz continuity result for solution sets with respect to the distance in (Theorem 2.3) and a quadratic growth condition near solution sets (Theo- rem 2.6). Both theorems extend earlier results in [25] to more general situations for the first stage costs g and constraint set C. All results in the paper apply to the linear-quadratic case, i.e., to linear or convex quadratic g and polyhedral C. Indeed, all results are formulated as general as possible and most of them are accompanied by illustrative examples. The second-order analysis of ' in Section 3 utilizes some ideas from [28] and [29], but its proof is entirely different and its Gateaux differentiability part is valid for nondifferentiable directions (Theorem 3.4). It is also elaborated that the Hadamard directional differentiability properties require the C 0 -topology for the first-order result and the C 1 -topology for the second-order one (Theorem 3.8), while the C 1;1 -topology is needed for the semidifferentiability of the solution-set mapping / (Theorem 4.7). All results on differentiability properties of / in Section 4 are new and do not follow from recent sensitivity results (as e.g. [3], [6], [7], [16], [29]; see also the survey [8] for further references). The results of Sections 3 and 4 have direct implications to asymptotic properties of values and solution sets of two-stage stochastic programs when applying nonparametric estimation procedures to approximate Q - . For a discussion of some of the related aspects we refer to [11], where the delta-method is utilized and a central limit theorem for all selections belonging to a Castaing representation of the approximate solution sets is derived. Further applications to asymptotics are beyond the scope of this paper and will be done elsewhere. Basic directional properties The first step in our analysis of directional properties consists in establishing results on the lower Lipschitz continuity of / and on the directional uniform quadratic growth of the objective near its solution set. Both results become important for our method of deriving directional differentiability properties for the optimal value function ' and the solution set mapping / at some given expected recourse function Q - . Their proofs are based on a decomposition of the program with Q belonging to K C , into two auxiliary problems. The first one is a convex program with decisions taken from A(C) and the second represents a parametric convex program which does not depend on Q. Proposition 2.1 Let Q 2 K C and /(Q) be nonempty. Then we have Moreover, - is convex on A(C) and dom oe is nonempty. Proof. Let - For the converse inequality, let " ? 0 and - be such that Then there exists a - "- is arbitrary, the first statement has been shown. In particular, x 2 oe(Ax) and Ax 2 Y (Q) for any x 2 /(Q) . Hence, it holds that /(Q) ' oe(Y (Q)). Conversely, implying Since the convexity of - is immediate, the proof is complete. 2 In the following, it will turn out that Lipschitzian properties of the solution set mapping y 7! oe(y) and a quadratic growth property of g near oe(y) are essential. For the linear-quadratic case we are in a comfortable situation in this respect. Namely, we have the following Proposition 2.2 Let g be linear or convex quadratic, C be convex polyhedral and assume dom oe to be nonempty. Then oe is a polyhedral multifunction which is Hausdorff Lipschitzian on its domain dom there exists a constant L ? 0 such that yk; for all where dH denotes the (extended) Hausdorff distance on subsets of IR m . Moreover, for each r ? 0 there exists a constant j(r) ? 0 such that (Here - and oe are defined as in Proposition 2.1). Proof. The Lipschitz property of oe is shown in [18], Theorem 4.2. To prove the second statement, let g be of the form positive semidefinite and c 2 IR m . For each y 2 A(C) we fix some z(y) 2 oe(y). An elementary characterization of solution sets to convex quadratic programs with linear constraints yields that Due to the Lipschitz behaviour of convex polyhedra (cf. [34]), there exists a constant for all y 2 A(C) and x 2 C with y. Using the decomposition 2 denotes the square root of H, and the representation one arrives at the estimate for all y 2 A(C) and x 2 C with y. us fix some element - r) and a corresponding oe(A-x). For each y 2 A(C) we now select z(y) 2 oe(y) such that Hausdorff Lipschitzian on A(C), this implies A(C). Hence, there exists a constant r). Thus our estimate continues to d(x; and some constant - Furthermore, the equation implies kH 1 y. Therefore, we finally obtain for all x Due to the above proposition, the main results in this section apply to the linear-quadratic case. Although this case represents the main application of our results, the assumptions of the following theorems are formulated in terms of general conditions on the mapping oe in order to gain generality and clarity. The first theorem states (lower) Lipschitz continuity of / at Q - and supplements Theorem 2.4 in [25]. Theorem 2.3 Let Q nonempty, bounded and Q - be strongly convex on some open, convex neighbourhood of A/(Q - ). Let - assume that there exist a constant L ? 0 and a neighbourhood U of - y with Then there exist constants - Proof. We may assume that U is open, convex and that Q - is strongly convex on U . Let V be an open, convex, bounded subset of IR m such that /(Q - It follows from Proposition 2.3 in [25] (where a slightly different terminology is used) that there exists a constant cl A(V chosen such that cl A(V r). Hence, we have ; 6= Proposition 2.1 yields the relation strongly convex on U , there exists a constant - ? 0 such that belongs to A(V ) ae U , we obtain and, hence, yk -- -- The proof can now be completed as follows. Let Q 2 K C be such that Then Remark 2.4 The proof shows that a Lipschitz modulus of / can be chosen as the quotient of a Lipschitz constant to oe and a strong convexity constant to Q - . From the proof it is immediate that replacing the local Lipschitz condition on oe by stronger conditions like sup leads to corresponding stronger Lipschitz continuity properties of solution sets. Because of Proposition 2.2, all of this applies to the linear-quadratic case. However, it is worth mentioning that the theorem also applies to more general problems such that the corresponding solution sets oe(y) enjoy Lipschitzian properties. Conditions ensuring Lipschitz behaviour of oe can be derived from stability results for the corresponding parametric generalized equation which describes the first order necessary optimality condition. Here L(x; -; y) := g(x)+ is the Lagrangian function, rL(x; -; , where g is assumed to be continuously differentiable, and N C \ThetaIR s is the normal cone map of convex analysis. Such stability results are presently available for broad classes of parametric generalized equations (e.g. [16], [20], [22]). A typical recent result in this direction, which applies to our situation for twice continuously differentiable g, is Theorem 5.1 in [20]. It says that the solution set mapping of the parametric generalized equation (2.2) is pseudo-Lipschitzian around (-x; - y) if the adjoint generalized equation has only the trivial solution w Here D N C \ThetaIR s (-x; -; \GammarL(-x; - y)) is the Mordukhovich coderivative ([20]) of the normal cone multifunction at the point (-x; -; \GammarL(-x; - belonging to the graph of . Translating this into our framework, we obtain that the mapping oe is pseudo- Lipschitzian around (-x; - y) if the following two conditions are satisfied: (a) There exists an element - x belonging to the relative interior of C such that y (Slater (b) the equations Aw have only the trivial solution w is a solution of (2.2) for y.) The next example shows that the theorem also applies to instances of two-stage stochastic programs with nonpolyhedral convex constraint sets C. Example 2.5 In (1.1) - be the uniform distribution on [\Gamma 1; 1] and x g. Then we have ~ R jyj otherwise , strongly convex on (\Gamma 1; 1). For y we have and, hence d((0; 0); . Thus Theorem 2.3 applies for - Example 2.8 shows that Theorem 2.3 gets lost if Q - fails to be strongly convex on some neighbourhood of A/(Q - ). Our next result establishes a sufficient condition for the uniform quadratic growth near solution sets. Theorem 2.6 Let Q nonempty, bounded and Q - be strongly convex on some open convex neighbourhood U of A/(Q - ). Assume that there exists a constant yk; for all and, for each r ? 0 there exists a constant j(r) ? 0 such that Then, for some open, bounded neighbourhood V of /(Q - ) and each there exist constants c ? 0 and ffi ? 0 such that the following uniform growth condition holds: for all x Proof. Let be an open, bounded subset of IR m such that As in Theorem 2.3 we choose ffi ? 0 such that ; 6= and, in addition, that strongly convex on U for all t 2 [0; ffi) (with a uniform constant - ? 0). For each t 2 [0; ffi) Proposition 2.1 then yields that is the unique minimizer of the strongly convex function - +tv on A(C) and, moreover, we have -ky \Gammay t k 2 -(y)+(Q - +tv)(y)\Gamma'(Q - +tv), for all y . Now, we choose r ? 0 such that V ' B(0; r) and continue for each Putting c completes the proof. 2 The following examples show that the quadratic growth condition gets lost even for the original problem, i.e. either the Lipschitz condition for oe or the strong convexity property for Q - are violated. Example 2.7 Consider again the set-up of Example 2.5. It holds that dH (oe(y); oe is not Hausdorff Lipschitzian on A(C). Supposed there exists a neighbourhood V of /(Q - such that the growth condition is satisfied. Since the sequence (( 1 belongs to C " V for sufficiently large n 2 IN , this would imply %( 1 for large n, which is a contradiction. Example 2.8 In (1.1) - - be the probability distribution on IR having the density R jyj otherwise , there is no neighbourhood of /(Q - ) where Q - is strongly convex. It is clear that the quadratic growth condition fails to hold, since the inequality %x 2 - cannot be true for some % ? 0 and all x belonging to some neighbourhood of With the linear function we obtain for all t 2 [0; 1] that /(Q - f (cf. Example 3.7). Hence, the lower Lipschitz property of / has got lost, too. Since the strong convexity and later also the strict convexity of the expected recourse function Q - (on certain convex subsets of IR s ) form essential conditions in most of our results, we record a theorem (Theorem 2.2 in [27]) that provides a handy criterion to check these properties for problem (1.1) - (1.3). Proposition 2.9 Let V ae IR s be open convex and assume (A1), (A3). Consider the following conditions: absolutely continuous on IR s ; there exist a density f - for - and a constant Then (A2) and (A4) imply that Q - is strictly convex on V if V is a subset of the support of -, and (A2) , (A4) imply that Q - is strongly convex on V . In addition, it is shown in [27] that under (A1) - (A4) the condition (A2) is also necessary for the strict convexity of Q - . For extended simple recourse models (i.e. is equivalent to q (componentwise), where This may be used to check strict or strong convexity properties in the Examples 2.5 and 2.8. Directional derivatives of optimal values In this section, we study first- and second-order directional differentiability properties of the optimal value function ' on its domain K C . We begin with the first-order analysis and show that ' as a mapping from K C to the extended reals is Hadamard directionally differentiable at some given expected recourse function Q - 2 K C . Here K C is regarded as a subset of C 0 (IR s ). Recall that ' is Hadamard directionally differentiable at Q - on K C iff for all sequences (v n ) converging to some v in C 0 (IR s ) and all sequences t n ! 0+ such that the elements belong to K C the limit exists. Since the condition means that v the limit v belongs to the tangent cone T (K C ; Q - ) to K C at Q - in C 0 (IR s ). In [32], [33] this property is also called Hadamard directional differentiability tangentially to K C . Proposition 3.1 Let Q - 2 K C and assume that /(Q - ) is nonempty, bounded. Then ' is Hadamard directionally differentiable at Q - on K C and it holds for all v 2 If, in addition, Q - is strictly convex on some open convex neighbourhood of A/(Q - ), we have Proof. Arguing similarly as in the proof of Propostion 2.1 in [24] there exists a neighbourhood N of Q - in C 0 (IR s ) such that /(Q) is nonempty for all Q sequences such that t n ! 0+, belongs to K C for all n 2 IN . Then sufficiently large n 2 IN . upper semicontinuous at Q - ([24]), the sequence has an accumulation point x 2 /(Q - ) and we obtain lim sup where the last inequality follows from the uniform convergence of (v n ) to v on bounded subsets of IR s . In order to show the reverse inequality for lim inf, let x 2 /(Q - ). Then lim inf This completes the proof of the first part. The second part is an immediate conclusion, since A/(Q - ) is a singleton whenever Q - is strictly convex on some of its open, convex neighbourhoods. 2 The preceding result can also be proved by using the methodology of Theorem 6.4.1 in [26]. There the compactness of the constraint set is assumed and Gateaux directional differentiability of ' at Q - together with its Lipschitz continuity is shown. Here we prefer a direct two-sided argument, which will also be used in the subsequent second-order analysis of '. Namely, we will first derive an upper bound for the second-order Hadamard directional derivative of ' at some Q - 2 K C , where K C is equipped with the C 0;1 -topology. Secondly, we identify conditions implying that the upper bound coincides with the Gateaux directional derivative of ' at Q - for all directions taken from T r (K C Lemma 3.2 Let y sequence in K C such that sequence converging to - in IR s . Then we have lim sup (v Proof. Each function v n is locally Lipschitzian on IR s and, hence, Lebourg's mean value theorem for Clarke's subdifferential ([9]) implies the existence of elements ~ y n belonging to the segments (v The convergence v n ! v in C 0;1 (IR s ) implies that n!1holds for any r ? 0. This yields 0: Here dH denotes the Hausdorff distance and the inequality is a consequence of general properties of the subdifferential (cf. Lemma 2.1 in [25]). Hence, there exist elements ~ belonging to @v(~y n ) such (v and, for some ~ lim sup (v Here, the identity follows from the upper semicontinuity of @v(\Delta). This completes the proof. 2 Proposition 3.3 Let Q - 2 K C and assume that /(Q - ) is nonempty, bounded. Let g be twice continuously differentiable, Q - be strictly convex on some open convex neighbourhood of A/(Q - ) and twice continuously differentiable at - y, where sequence in K C such that v n := 1 in C 0;1 (IR s ). Then lim sup x) is the tangent cone to C at - x and T 2 (C; - x; -) the second order tangent set to C at - x in direction -. Proof. Let -). Then there exists a sequence (z n ) such that Using Proposition 3.1, this allows for the following estimate After dividing by t 2 n and using Lemma 3.2 the limes superior as of the right-hand side can be bounded above by Taking the infimum on the right-hand side yields the assertion. 2 We notice that the upper second-order Hadamard directional derivative lim sup nonpositive, since ' is concave on K C and, hence, the inequality '(Q - +t n v n We also note that the upper bound is nonpositve, since (0; belongs to S(-x) \Theta T 2 (C; - x; Next we consider particular perturbations Q n of Q - , namely, Q for some Q 2 K C sufficiently large n 2 IN . Then v In the following result we give conditions implying that the second-order directional derivative exists and coincides with the upper bound of the previous proposition. The result extends those in [12] although its proof parallels in parts that of Theorem 3.6 in [12]. Theorem 3.4 Let Q - 2 K C and assume that /(Q - ) is nonempty, bounded. Let g be twice continuously differentiable, Q - be strictly convex on some open convex neighbourhood of A/(Q - ) and twice continuously differentiable at - y, where assume that (ii) the second-order set S 2 (-x; -) := fz 2 T 2 (C; - is nonempty for each - 2 S(-x). Then the second-order Gateaux directional derivative of ' at Q - in direction v exists and it holds that Moreover, the infimum is attained at some - having the property that -). (Here S(-x) and T 2 (C; - are defined as in the previous result, v 0 (-y; j) is the directional derivative of v at - y in direction j and O(t) denotes a real quantity such that 1 jO(t)j is bounded as t ! 0+.) Proof. (i) implies that there exist constants L ? 0, expanding g and Q - and using Proposition 3.1 we obtain xi +2 xi Moreover, we have that denotes a real quantity having the property 1 the optimality of - x implies for any t 2 (0; ffi), we Now take a sequence (t n ) tending to 0+ in such a way that lim inf and that - n := 1 -. The latter is possible since k 1 sufficiently large. Then - x) and Proposition 3.1 yields This implies - From (3.1) and (3.2) we obtain lim inf Here we have used the fact that v is Hadamard directionally differentiable and Clarke regular ([9]), i.e. v 0 (-y; ji. From Proposition 3.3 we obtain lim sup The latter equality is due to (ii) and to the fact that the necessary optimality condition for - x yields Hence, the limit lim exists and is equal to the infimum subject to - 2 S(-x). Moreover, this infimum is attained at - S(-x). For the remainder of the proof we put a(-) := v 0 (-y; A-) and Since S(-x) is a (convex) cone, we have -)); for all - ? 0: In case of B( - -) ? 0, the quadratic function f vanishes at with the property and the final assertion is shown. If B( - the fact that -) holds for any - ? 0, implies a( - -) and the proof is complete. 2 The proof shows that the previous theorem remains true when replacing condition (ii) by the condition that both infima in (3.3) coincide. Next we state a more handy criterion implying that ' 00 (Q - ; v) exists for any direction v 2 T r (K C Corollary 3.5 Let Q - 2 K C and assume that /(Q - ) is nonempty, bounded. Let g be twice continuously differentiable, Q - be strongly convex on some open convex neighbourhood of A/(Q - ) and twice continuously differentiable at - y where assume that (i) 0 there exist a constant L ? 0 and a neighbourhood U of - y such that (ii) the second-order set S 2 (-x; is nonempty for each - 2 S(-x). Then the second-order Gateaux directional derivative of ' at Q - exists for any direction the formula for ' 00 (Q - ; v) in Theorem 3.4 holds true. Moreover, condition (ii) is satisfied if C is polyhedral and (i) 0 is satisfied for any - x 2 in addition to the polyhedrality of C, g is linear or (convex) quadratic. Proof. Let Theorem 2.3 then says that there exist constants - Hence, the strong convexity of Q - and condition (i) 0 imply that condition (i) of the previous theorem is satisfied and that the first part of the assertion is shown. If C is polyhedral, we have T 2 (C; - x). Hence, (ii) is satisfied. If C is polyhedral and g is linear or (convex) quadratic, Proposition 2.2 implies (i) 0 to hold for any - Let us consider two illustrative examples to provide some insight into the benefit and limits of the previous results. Example 3.6 We revisit Example 2.5 and know that condition (i) 0 is satisfied for Furthermore, it holds that T (C; - Hence, (ii) and the general assumptions of Corollary 3.5 are satisfied and ' 00 (Q - ; v) exists for any v 2 T r (K C ; Q - ). It holds that ' 00 (Q - IRg. Let us finally replace the function g(x) j 0 by IR, and condition (ii) is violated. But, since we have both infima in (3.3) coincide, the result holds true and we have Example 3.7 Here we revisit Example 2.8, and have For the function 2 . Hence, ' has no second-order directional derivative at Q - in direction v. Note that there is no neighbourhood of - strongly convex. Finally, we aim at showing that ' is even second-order Hadamard directionally differentiable at equipping K C with a suitable topology. To this end we need a certain counterpart of Lemma 3.2 for the corresponding limes inferior. Since this is not available for nonsmooth functions, it is a natural idea to consider the space C 1 (IR s ), to restrict ' to the subset K C " C 1 and to equip K C " C 1 with the C 1 -topology. Then we are able to show that the assumptions of Corollary 3.5 even imply the second-order Hadamard directional differentiability of ' at Q - . Theorem 3.8 Let Q assume that /(Q - ) is nonempty, bounded. Let g be twice continuously differentiable, Q - be strongly convex on some open convex neighbourhood of A/(Q - ) and twice continuously differentiable at - y where assume the conditions (i) 0 and (ii) of Corollary 3.5 to hold. Then the second-order Hadamard directional derivative of ' at Q - exists in any direction v belonging to the tangent cone T (K any such v, and all sequences t n ! 0+ and (Q n ) in K C such that v n := 1 exists, and it holds 0g. Proof. Let sequence in K C such that together with Theorem 2.3 then imply that there exist constants L ? such that Since the sequence (v n ) converges in C 1 (IR s ), the norms kv n k L;r are uniformly bounded and we have Expanding g and Q - as in the proof of Theorem 3.4 we obtain analogously to (3.1), for all n - n 0 :t 2 Putting - n := 1 and using the mean value theorem for v n we may continue with some - Arguing as in the proof of Theorem 3.4 and using v n ! v in C 1 (IR s ) we arrive at the estimate lim inf for some element - Furthermore, we conclude from (ii) and Proposition 3.3 that lim sup Hence, the desired limit exists and the proof is complete. 2 Let us finally note that all minimization problems appearing as bounds or formulas for second-order directional derivatives represent convex programs. Those in the results 3.4, 3.5 and 3.8 have convex cone constraints, which are polyhedral if C is polyhedral. Moreover, the solution sets of the convex minimization problems in 3.4, 3.5 and 3.8 are nonempty. Indeed, we show next that these solution sets represent certain derivatives of the set-valued mapping / at the pair (Q - x). 4 Differentiability of solution sets It is well-known that second-order differentiability properties of optimal values in perturbed optimization are intrinsic for establishing the differentiability of solutions (see e.g. [8]). We also pursue this approach and derive conditions implying directional differentiability properties of the solution set mapping by exploiting the results of the previous section. Our first results in this direction concern Gateaux directional differ- entiability, and complement Theorem 3.4 and its corollary. Theorem 4.1 Assume that the general conditions on g, Q - and C of Theorem 3.4 are suppose the conditions (i) and (ii) of Theorem 3.4 to be satisfied. In addition, assume that (iii) there exist a neighbourhood V of such that the uniform growth condition for all x Then the Gateaux directional derivative of / at the pair (Q - x) into direction v exists and it holds that Proof. Let M(-x; v) denote the solution set in the assertion. First we show that lim sup x). Then there exists a sequence (t converging to (0+; -) such that - n 2 1 Hence, analogously to the proof of Theorem 3.4 we deduce that - belongs to S(-x). In view of Theorem 3.4 it remains to show that 1hr 2 g(-x)-i expanding g and Q - as in the proof of Theorem 3.4, we obtain analogously to (3.1): After dividing by t 2 n and taking the lim on both sides of the inequality, we obtain the desired estimate. In a second step we show that or, equivalently, that it holds for any - 2 M(-x; v), lim t!0t sequence with t n ! 0+. We have to show that lim there exists an element z and a sequence (z n ) converging to z with - suffices to show that lim Condition (iii) implies the following estimate for all sufficiently large n By expanding g and Q - as in the proof of Theorem 3.4 and using the fact that - belongs to S(-x), we may continue After dividing by t 2 n and taking the lim sup on both sides of the latter inequality, we obtain lim sup c where we made use of z Theorem 3.4. This completes the proof. 2 Complementing Corollary 3.5 we provide a result on the directional differentiability of / at Q - into any direction v Theorem 4.2 Assume that the general conditions on g, Q - and C of Corollary 3.5 are satisfied. Let - assume that (i) 00 there exists a constant L ? 0 such that and, for each r ? 0, there exists a constant j(r) ? 0 such that (ii) the second-order set S 2 (-x; is nonempty for each - 2 S(-x). Then the Gateaux directional derivative / 0 (Q - x; v) of / at the pair (Q - exists for any direction v the formula in Theorem 4.1. Moreover, condition (ii) is satisfied if C is polyhedral, and (i) 00 is satisfied if C is polyhedral and g is linear or (convex) quadratic. Proof. Let strongly convex on some open convex neighbourhood of A/(Q - ), we infer from condition (i) 00 and Theorem 2.6 that condition (iii) of Theorem 4.1 is satisfied. Moreover, condition (i) 00 implies (i) 0 and, thus, Corollary 3.5 says that the second-order directional derivative ' 00 (Q - ; v) exists. Hence, the first part of the assertion follows from the proof of the previous theorem. If C is polyhedral, we have 0 2 S 2 (-x; -) for any - 2 S(-x), and if, in addition, g is convex quadratic, Proposition 2.2 implies condition (i) 00 to hold. 2 We note that Example 3.7 shows that, in general, the directional differentiability property of / gets lost at those pairs (Q - is not strongly convex on some neighbourhood of A/(Q - ). Finally, we turn to directional differentiability properties of / where the derivatives exist uniformly with respect to directions taken from compact sets of certain functional spaces. For our first result we consider the space C 1 (IR s ) and equip the set K with the C 1 -topology. Proposition 4.3 Let Q assume that the general conditions on g, and C in Proposition 3.3 are satisfied. In addition, we suppose condition (ii) of Theorem 3.4 to be satisfied. Let - sequence in K C such that v n := 1 Then the upper set limit of the sequence ( 1 of closed convex subsets in IR m , i.e., lim sup x)), is contained in the closed convex set argmin Proof. Let D n := 1 x) for all n 2 IN and let - - belong to the upper set limit lim sup . Then there exist a subsequence (again denoted by (D n )) and elements -. Since - we have that - As in the proof of Theorem 3.4 we deduce that hrg(-x); - thus, - expanding g and Q - as in the proof of Theorem 3.4, we also obtain analogously to (3.1): After dividing by t 2 n and taking the lim sup on both sides of the inequality, we obtain as in the proof of Theorem 3.8 lim sup -i: Hence, we may conclude from (ii) and Proposition 3.3 that - - belongs to the set and we are done.Remark 4.4 The upper limit of the sequence ( 1 in Proposition 4.3 is nonempty if the mapping d(-x; /(\Delta)) from K C into the extended reals has the Lipschitzian property of Theorem 2.3 at Q - . Indeed, we may select x for large n 2 IN , such that for some constants - . Hence, the sequence ( 1 is bounded and has a convergent subsequence whose limit belongs to lim sup x). If the Lipschitz property of d(-x; /(\Delta)) is violated, the upper set limit may be empty. This is illustrated by Example 3.7, in which we have - g. In order to establish the semidifferentiability of / at a pair (Q - x) belonging to the graph of /, it remains to show, according to Proposition 4.3, that the solution set argmin is contained in the lower set limit lim inf converges to v. To this end, a uniform quadratic growth condition of the objective functions g(\Delta) is significant. In view of Theorem 2.6, the uniform strong convexity of Q - and its approximations Q n , for large n 2 IN , is decisive for the growth condition. The next example and the following result show that the approximations Q n do not maintain the strong convexity property of Q - in general if the sequence (Q n ) converges to Q - in C 1 (IR s ), but that the situation is much more advantageous when considering the C 1;1 -topology. Example 4.5 Let Q be the following differentiable convex functions \Gammay \Gamman 0; y \Gamman Note that Q and Q n is not strongly convex for each n 2 IN , but (Q n ) converges to Q - in C 1 (IR s ). strongly convex on some bounded convex set (with some constant - ? 0). Then there exists a neighbourhood N of Q - in C 1;1 (IR s ) such that each function Q belonging to N is strongly convex on U with constant -Proof. The strong convexity of Q - on U (with constant - ? 0) is equivalent to the condition such that cl U ' B(0; r) and let N be a neighbourhood of Q - in C 1;1 (IR s ) having the property y. Then we obtain for any Q 2 N , and, hence This means that Q is strongly convex on U with constant -. 2 Now we are able to show that the solution set mapping / is semidifferentiable on at some pairs (Q - direction v from the tangent cone in C 1;1 (IR s ). The assumptions are essentially the same as in Theorem 4.2. Theorem 4.7 Let Q assume that /(Q - ) is nonempty, bounded. Let g be twice continuously differentiable, Q - be strongly convex on some open convex neighbourhood U of A/(Q - ) and twice continuously differentiable at - y, where Assume that, for each r ? 0, there exist constants L ? 0 and j(r) ? 0 such that the following condition (i) 00 is satisfied for Then the solution set mapping / from K C " C 1;1 into IR m is semidifferentiable at any such that S 2 (-x; -) is nonempty for each - 2 S(-x), and into any direction v 2 T (K any such - x and v, t n ! 0+, and (Q n ) in exists. The semiderivative D/(Q - x; v) is equal to the set argmin Moreover, / is semidifferentiable at any pair (Q - direction polyhedral. Condition (i) 00 is satisfied if C is polyhedral and g is linear or (convex) quadratic. Proof. Let be such that S 2 (-x; -) is nonempty for each - 2 S(-x), and (Q n ) is a sequence in K C " C 1;1 . We may assume that U is bounded. Since converges to Q - in C 1;1 (IR s ), we obtain from Lemma 4.6 that there exists an , such that Q n is strongly convex on U for each n - n 0 with a uniform constant sufficiently large such that /(Q n ) is nonempty, for each Arguing as in the proof of Theorem 2.6, we obtain a constant c ? 0 and a neighbourhood V of /(Q n ) such that the growth condition holds for all x be a minimizer of the function 1hr 2 g(-x)-i subject to - 2 S(-x). Because of Proposition 4.3 it remains to show that - belongs to the lower limit lim inf there exists an element z -) and a sequence (z n ) converging to z with As in the proof of Theorem 4.1 it suffices to show that lim By using the above growth condition and by expanding the function g and Q - , we obtain as in the proof of Theorem 4.1 and lim sup c This implies - x) and the semidifferentiability of / at (Q - in direction v is shown. The remaining part of the assertion follows as in the proof of Theorem 4.2. 2 For the linear-quadratic case, the essential assumptions in Theorem 4.7 are the strong convexity of Q - , and the smoothness properties of Q - and its perturbations Q, respec- tively. While criteria for strong convexity were already discussed in Section 2, we now close this section by adding some comments on C 1;1 - and C 2 -properties of expected recourse functions. Remark 4.8 Assume (A1) - (A3) and - to have a density with respect to the Lebesgue measure on IR s . Then the function Q - in (1.2) is continuously differentiable on IR s and its gradient is of the form rQ - )), for all y are certain basis submatrices of the recourse matrix W such that the simplicial cones B i (IR s are linearity regions of ~ Q and d i is the gradient of ~ Denoting by F - the distribution function of - and using the formula -(y +B(IR s for any nonsingular (s; s)-matrix B, C 1;1 - and C 2 -properties of Q - may thus be formulated in terms of Lipschitz and differentiability properties of the distribution functions to the linear transforms - of the measure -. The distribution function F - of a probability measure - on IR s is locally Lipschitzian if all one-dimensional marginal distribution functions of - are locally Lipschitzian (cf. [24], [35]). F - is continuously differentiable if - has a continuous density function and all one-dimensional marginal distribution functions of - are continuously differentiable (cf. [19], [35]). If - has a continuous density function, then - ffi B has a continuous density for any nonsingular (s; s)-matrix B, too. Hence, we may conclude, for instance, that Q - belongs to C 1;1 (IR s has a (continuous) density and the above-mentioned conditions on the one-dimensional marginal distribution functions for F -ffiB belonging to C 0;1 (IR s are satisfied for any nonsingular (s; s)-matrix B. This criterion is particularly useful for probability distributions which have the property that all one-dimensional marginal distributions of - and all linear transforms - ffi B, for all nonsingular matrices B, belong to the same class of measures. For instance, all multivariate normal and all logarithmic concave probability measures (e.g. [14]) form classes having this property. Acknowledgement : The authors wish to thank Alexander Shapiro (Georgia Institute of Technology, Atlanta) and Ren'e Henrion (WIAS Berlin) for beneficial discussions --R Stability results for stochastic programs and sen- sors First and second order sensitivity analysis of nonlinear programs under directional constraint qualification conditions A comparative study of multifunction differentiability with applications in mathematical programming Directional derivatives in nonsmooth optimization Perturbed optimization in Banach spaces I: A general theory based on a weak directional constraint qualification Quadratic growth and stability in convex programming problems with multiple solutions Optimization problems with perturbations Optimization and Nonsmooth Analysis tangent sets and second-order optimality condi- tions Differentiable selections of set-valued mappings with application in stochastic programming Strong convexity and directional differentiability of marginal values in two-stage stochastic programming Stability and sensitivity analysis for stochastic programming Generalized delta theorems for multivalued mappings and measurable selections Sensitivity analysis for nonsmooth generalized equations Asymptotic theory for solutions in statistical estimation and stochastic programming bounds for solutions of linear equations and inequalities Approximationen der Entscheidungsprobleme mit linearer Ergebnis- funktion und positiv homogener Stability theory for parametric generalized equations and variational inequalities via nonsmooth analysis Differentiability of relations and differential stability of perturbed optimization problems Strongly regular generalized equations Stability of solutions for stochastic programs with complete recourse Lipschitz stability for stochastic programs with complete recourse Discrete Event Systems. Strong convexity in stochastic programs with complete recourse Second order directional derivatives in parametric optimization prob- lems Sensitivity analysis of nonlinear programs and differentiability properties of metric projections On concepts of directional differentiability On differential stability in stochastic programming Asymptotic analysis of stochastic programs Weak Convergence and Empirical Pro- cesses A Lipschitzian characterization of convex polyhe- dra Distribution sensitivity analysis for stochastic programs with complete recourse Stochastic programs with fixed recourse: the equivalent deterministic program in: Handbooks in Operations Research and Management Science. --TR --CTR Svetlozar T. Rachev , Werner Rmisch, Quantitative Stability in Stochastic Programming: The Method of Probability Metrics, Mathematics of Operations Research, v.27 n.4, p.792-818, November 2002
semide-rivatives;sensitivity analysis;two-stage stochastic programs;directional derivatives;solution sets
589099
Multiple Cuts in the Analytic Center Cutting Plane Method.
We analyze the multiple cut generation scheme in the analytic center cutting plane method. We propose an optimal primal and dual updating direction when the cuts are central. The direction is optimal in the sense that it maximizes the product of the new dual slacks and of the new primal variables within the trust regions defined by Dikin's primal and dual ellipsoids. The new primal and dual directions use the variance-covariance matrix of the normals to the new cuts in the metric given by Dikin's ellipsoid.We prove that the recovery of a new analytic center from the optimal restoration direction can be done in O(plog (p+1)) damped Newton steps, where p is the number of new cuts added by the oracle, which may vary with the iteration. The results and the proofs are independent of the specific scaling matrix---primal, dual, or primal-dual---that is used in the computations.The computation of the optimal direction uses Newton's method applied to a self-concordant function of p variables.The convergence result of [Ye, Math. Programming, 78 (1997), pp. 85--104] holds here also: the algorithm stops after $O^*(\frac{\bar p^2n^2}{\varepsilon^2})$ cutting planes have been generated, where $\bar p$ is the maximum number of cuts generated at any given iteration.
Introduction The analytic center cutting plane (ACCPM) algorithm [5, 19] is an efficient algorithm in practice [2, 4]. The complexity of related algorithms was given in [1, 13], and subsequently in [6]. Extensions to deep cuts were given in [7] and to very deep cuts in [8]. The method studied in [8] corresponds to the practical implementation of ACCPM [11] with a single cut. In practice, it often occurs that the oracle in the cutting plane scheme generates multiple cuts. The papers [12, 20, 17] show that it is possible to handle several cuts at a time provided they are central [20] or moderately shallow [12]. Although these analyses show how one can recover feasibility after introducing multiple cuts, there is no clear argument as to the choice of a feasibility restoration direction. Intuitive, but well justified, arguments about how to introduce multiple cuts were given in [2] in the context of a primal projective algorithm and two cuts (one shallow, one deep) and in [10] with an infeasible primal-dual approach to the introduction of several cuts in general position. The case of two central cuts was analyzed in [9]. It was shown that there exist explicit primal and dual directions which allow a best move towards primal and dual feasibility. An argument using the primal, dual and primal-dual potentials at this new optimal primal and dual point proves that O(1) damped Newton steps are enough to recover centrality. The updating direction depends on the cosine in the metric of Dikin's ellipsoid of the normals to the cuts. In this paper, we analyze the multiple cut generation scheme in the analytic center cutting plane method. We propose an optimal updating direction when the cuts are central. The direction is optimal in the sense that it maximizes the product of the new slacks within the trust region defined by Dikin's ellipsoid. The new primal and dual directions use the variance-covariance matrix of the normals to the new cuts in the metric given by Dikin's ellipsoid. We prove that the recovery of a new analytic center from the optimal restoration point can be done in O(p log(p+1)) damped Newton steps, where p is the number of new cuts added by the oracle. The results and the proofs are independent of the specific scaling matrix -primal, dual or primal-dual- that is used in the computations. The computation of the optimal direction uses Newton's method applied to a self-concordant function of p variables. This could be very advantageous in practice if the number of cuts p is a small multiple of n, the dimension of the space. The convergence result of [20] holds here also: the algorithm stops after O cutting planes have been generated. Analytic center cutting plane method 2.1 Cutting planes The problem of interest is that of finding a point in a convex set C ae IR n . We make the following assumptions. Assumption 2.1 The set C is convex, contains a ball of radius " ? 0 and is contained in the cube 0 - y - e. Assumption 2.2 The set C is described by an oracle. That is, the oracle either confirms that y 2 C, or answers at least one cutting plane that contains C and does not contain y in its interior. A cut at - takes the form a T y - a T - fl: the cut is deep; if - the cut is shallow; if - the cut passes through - y, and is thus a central cut. The algorithm may generate multiple cuts at a time. They take the form a T We define the matrix B by Assumption 2.3 All the cutting planes generated have been scaled so that We also assume that the cuts are central, that is - A cutting plane algorithm constructs a sequence of query points fy k g. The answers of the oracle to the queries, together with the cube 0 - y - e, define a polyhedral outer approximation of C. Since A contains the identity matrix associated with the cube, A has full row rank. Therefore there is a one-to-one correspondence between points and the slack y, leading to the equivalent definition of FD The number of columns in A is denoted as m and is equal to 2n plus the number of cutting planes generated until the k th iteration. The analytic center cutting plane method chooses as a query point an approximate analytic center of FD . 2.2 Analytic center The analytic center of FD is the unique point maximizing the dual potential log s i with formally introduce the optimization problem and the associated first order optimality conditions where x is a vector in R m . The notation xs indicates the Hadamard or componentwise product of the two vectors x and s. The analytic center can alternatively be defined as the optimal solution of where log x i denotes the primal potential. One easily checks that problem (2) shares with (1) the same first order optimality conditions. At this stage it is convenient to introduce the primal-dual potential and an associated duality relationship. Lemma 2.4 Let x 2 intF P and s 2 intF D . Then 'PD (x; s) - \Gammam; with equality if and only if Proof Consider the simple inequality log with equality if and only if Apply (3) with By summing the resulting inequalities, one gets log log with equality if and only if xs = e. Therefore, with equality if and only if Finally, we define approximate centers by relaxing the condition in the first order optimality conditions. Formally, any solution (x; s) of defines a pair of '-approximate centers, or '-centers in short. 2.3 Analytic center cutting plane method ACCPM can be shortly stated as follows. Initialization Let F 0 fy eg be the unit cube and y e be its center. The centering parameter is Basic Step y k is a '-center of F k the total number hyperplanes describing F k D . 1. The oracle returns the cuts amk+j , k , at y k . 2. Update g. 3. Compute a '-center of F k+1 D . The computation of a new '-center after adding new cuts will be discussed in a further section. 3 Some useful properties The literature on interior point methods essentially proposes three approaches for computing analytic centers. All of them are based on Newton's method. The primal (resp. dual) Newton direction is initiated at an interior primal (resp. dual) feasible point; it involves the scaling matrix (We recall the standard notation X which denotes the diagonal matrix diag(x).) The primal-dual direction is initiated at an interior primal-dual feasible pair, involves the scaling matrix Let us shortly recall the formulas. The primal direction is given by xp(x) with The dual direction is given by \Deltas = sq(s) with Finally the primal dual direction is \Deltas and 3.1 Properties of the Newton step There are two basic properties, a local one in the vicinity of the analytic center, and a global one. Since the results are well-known we state them without proofs. Missing proofs can be found in the books [18] or [21]. Let us start with the local properties. Proximity to analytic center is measured with the quantity sxk. In this definition, either Note that if and case). The Newton step defines a pair Theorem 3.1 Assume 3 . Let be the point resulting from a Newton step (primal, dual, or primal-dual). Then, intF D \Theta intF P . In the primal and dual cases, the theorem holds with any 0 ! ' ! 1. One can derive from the above theorem a useful corollary that yields lower bounds on the potentials near the analytic center. Let be the pair of exact analytic centers. Denote ' c Corollary 3.2 Assume (5)-(7) at (x; s). Then 1. ' c 2. ' c 3. \Gammam - 'PD (x; s) - \Gammam Let us now consider the global properties of a damped Newton step. The properties are consequences of the well-known inequality on the logarithm function Lemma 3.3 Let h be any point in R m such that khk ! 1. Then, The main result bounds the variation of the potentials after a damped Newton step. Theorem 3.4 Assume s+ff\Deltas. (\Deltax and \Deltas may be the primal, dual or primal-dual directions.) Then, there exists a step size ff ? 0 and absolute constants oe P , oe D and oe PD such that 1. 'P 2. 'D (x(ff)) - 'D 3. 'PD (x(ff); In the primal and dual cases the constants are oe '). The above result allows to design a potential increase algorithm based on damped Newton steps. The convergence estimate is given by the following theorem. Theorem 3.5 Let x potential increase algorithm (primal, dual, primal-dual) produces an interior feasible pair such that a number of iterations not greater than oe with or oe PD , depending on which approach (primal, dual or primal- dual) is taken. 3.2 Dikin's ellipsoids . From the observation that x \Deltax such that 1, we can define an ellipsoidal neighborhood of x that is entirely contained in FP . Formally, We shall be particularly concerned with ellipsoids around a '-center. We can extend the definition of Dikin ellipsoid to include a different scaling. Lemma 3.6 Let (x; s) be a pair of '-centers. 1. If 2. If Proof For the dual scaling the proof follows from and For the primal-dual scaling the proof follows from and We can similarly define Dikin's ellipsoids in the dual. Let s 2 intF D . The dual ellipsoid is The extension of Dikin's ellipsoid to a different scaling at a '-center is given by Lemma 3.7 1. If 2. If The proof is the same as for Lemma 3.6. It is well-known that an homothety of Dikin's ellipsoid contains the feasible set. We shall use this property in the restricted context of the set FD . Lemma 3.8 Let (x; s) be a '-centered feasible pair. Then ae oe Proof s) be a '-centered feasible pair and ~ point of FD . Since are orthogonal, Since ~ one has x T s - thus obtain the weak bound Finally, from kD(~s \Gamma s)k - Hence, ae oe 4 Multiple central cuts We assume now that a '-center (x; s; y) has been computed, i.e., The cuts are a T We define The new cuts lead to two new sets: e or e and e We shall use the notation so After adding the cuts, one has ~ FD and ~ Let us introduce the notation The primal and dual potentials at the new points (-x; fi) and (-s; fl) are: ~ log - log log and ~ log - log log The points lie on the boundary of the primal and dual sets respectively. To recover the new analytic center, one has to increase the components fi and fl. Since the terms log fi i and log fl i are dominant near maximizing those terms while limiting the variation on 'P and 'D is likely to produce a good step towards the solution. This approach requires the knowledge of the level sets of the potential, something that we don't have, but that can be approximated by Dikin's ellipsoids. Therefore, we are interested in solving the following problems log and log \Deltay Here D is one of the scaling matrices X depending whether the computations are done with the primal, the dual or the primal-dual algorithm. Let show here that the above problems are well-defined and have a finite optimum Lemma 4.1 Under Assumptions 2.1 and 2.2, Problems (11) and (12) are well defined and have a finite optimum that is uniquely defined by the first order optimality conditions. Proof Both problems have a strictly concave objective. Their optimum, if it exists, is unique in fi (resp., fl). By Assumptions 2.1 and 2.2, there exists a - \Deltay such that B T - Problem (12) is well-defined. Since \Deltay is bounded, fl is bounded and the feasible set is compact. Since the objective tends to \Gamma1 close to the boundary, the problem has a finite solution that is uniquely defined by the set of first order optimality conditions. To show that Problem (11) is also well-defined, we note that the equation A\Deltax+ has a solution for any fi ? 0 since A has full row rank. Let us show that the feasible set is bounded. Indeed, let fi - 0 and \Deltay Recalling that A has full row rank, we conclude from A\Deltax fi is bounded, since \Deltax is bounded by Problem (11) is thus well-defined and has a finite optimum. The solutions of Problems (11) and (12) define the primal dual pair of rays fffiA and \Deltay ffflA ~ FP and ~ s(ff) 2 int ~ The following positive semidefinite matrix plays a fundamental role in the analysis. V can be interpreted as variance-covariance matrix between the vectors (a m+j in the metric induced by the matrix (AD 2 A Theorem 4.2 The solution of Problems (11) and (12) is given by and with fi defined as the unique solution of log and Proof Let - 2 R n and oe 2 be the multipliers associated with the constraints of Problem (11). The optimality conditions are ?From the definition of \Deltax, one immediately sees that A\Deltax Letting This proves the second relation. To prove the first relation, we shall use the optimality condition for Problem (13). However, we must check first that (13) has a bounded optimum. In Lemma 4.1 we proved that nonegative solution. Thus, for all fi - 0, fi 6= 0, one has This proves that the objective \Gamma p log fi i is bounded above and Problem has a unique optimum. The optimality condition for Problem (13) is Replacing fi \Gamma1 by pV fi we get the identity It remains to check that and Let us now consider Problem (12). The optimality conditions are \Deltay are the multipliers associated with the two constraints We want to show that are the optimal multipliers, where fi is the optimal solution of Problem (13). Solving for \Deltay, one gets: Now Remembering the optimality condition for fi, one may replace thus check that the first optimality condition holds. Finally, \Deltay proves that with our choice of multipliers, the last optimality condition also holds. Remark 4.1 If V is nonsingular, fl is also the unique solution of log We can now give an explicit formula for the restoration direction. Noting that we have the new primal-dual pair ~ and ~ Remark 4.2 We note a significant dissymmetry between the primal and dual directions: 1. any positive value of fi, say primal feasible direction 2. but fi ? 0 does not guarantee then taking a feasible dual direction. Different stepsizes (ff could be used in the primal and dual space. Note that, by construction, D\Deltas. At the optimum direction, one has The computation of fi requires solving the nonlinear optimization problem (13). Since the function F log self-concordant, it can easily be minimized by classical Newton schemes. We postpone to a later section the discussion on the complexity estimate for getting approximate solutions. For the sake of a simpler presentation we shall assume in our analysis of ACCPM that the minimizers are exact. However, this is not the case in practice and we must be concerned with the impact of errors on fi and fl on the performance of ACCPM. This discussion is also postponed to a later section. Below, we sketch the result that enables an easy extension of our analysis of ACCPM with multiple cuts in the case of inexact computations of fi and fl. The convergence analysis of section 5 relies on the following properties: e: If we can guarantee that the solutions satisfy pfifl - e and 1 then the convergence result on ACCPM is essentially unaffected, while the proofs need only minor adjustments. We give here a theorem that stipulates the condition that must be met by fi and fl to carry the analysis with inexact minimizers. In a later section we shall show that classical interior point schemes make it possible to meet the condition. Theorem 4.3 Assume fi ? 0 and kpfi(V fi) \Gamma ek - j. Let and In particular, Proof The first set of inequalities follows directly from the assumption and the definition of fl. These inequalities also imply that Multiplying these inequalities by e T one gets 5 Convergence analysis We now assume that (x; s) is a pair of '-centers and that \Deltax and \Deltas are computed as in Section 4 with fi and fl being the exact minimizers of problems and (14). We assume that the computations are done with either the primal, the dual of the primal-dual scaling. Lemma 5.1 Independently of the specific scaling matrix D (primal, dual or primal-dual), one has, for any ff Proof By construction 1. From Lemma 3.6, for any primal, dual or primal-dual scaling D, we have The proof is the same in the dual case. Remark 5.1 The above result can be sharpened by considering separately the three different scaling matrices D. However, we prefer the weaker result since it allows a single formulation for the three cases. Lemma 5.2 The following inequalities hold: and je Proof ?From \GammaA\Deltax, one has Thus, To prove the second statement, we note that x T je In view of the above lemmas, we can bound the potentials e 'P and e 'D at the new pair of points (~x(ff); ~ s(ff)). Lemma 5.3 For any 0 ! ff the new potentials satisfy e log e log and e Proof Let us prove first the inequality on the primal potential. At the updated point ~ x(ff) the potential is e log x log log x log log 1. We can apply Lemma 3.3 to get Then, by Lemma 5.2 ffe decreasing, we can bound khP k e log Let us prove now the dual case. We have e log s log log \Deltas. By Lemma 5.1 khD k ! 1. We can apply Lemma 3.3 to get Since by Lemma 5.2 we obtain, by putting the inequalities together, the same result as in the primal case e log To conclude the proof of the theorem, we just sum the inequalities on ~ 'P and ~ 'D and use e to get e 5.1 Recovering the new analytic center Theorem 5.4 The number of Newton steps to compute the updated '-analytic center is bounded by oe where and, depending on the Newton scheme, Proof To bound the number of Newton steps, we compute the optimality gap for the sum of the primal and dual potentials. On the one hand, ~ On the other hand, we can write e Finally, Hence e Thus Using theorem 3.4 and the above bound on the potential variation we conclude the proof of the theorem. 5.2 Convergence of ACCPM with multiple cuts The next lemma is a first step on bounding the number of calls to the oracle. Theorem 5.5 For all ~ log with Proof The first inequality uses ~ 'P (~x(ff)), the duality on potential and Lemma 2.4 to yield log We now need to deal with the contribution of the new variables log Since fi solves (13), we have fi T V log f log log for any arbitrary fi 0 . Let us define the vector - by a T Note that - while the off-diagonal terms of V are The off-diagonal elements Those properties are typical of a variance-covariance matrix. Let us choose Then The correlation matrix: all its coefficient are bounded in absolute value by 1, and Thus log log log log Using corollary 3.2 we have Putting together (21), (22) and (23) yields ~ ff ff log The bound ff can be analyzed by selecting, somewhat arbitrarily, guaranteeing ff which is exactly the same result as in [20], but with a rather different derivation, as we show that this inequality is actually achieved at the iterate obtained by the restoration step. Remark 5.2 If the p cuts generated are identical, then the correlation matrix R is the rank-one matrix ee T . Otherwise for the optimal fi log fi log may be significantly greater than 0 and speed the convergence in practice, even though this does not appear to affect the worst case complexity bound. 5.3 Convergence of ACCPM The convergence analysis uses the proof given in [20], for the case of multiple cuts. Denote and let P k be the same value after k calls to the oracle, that is, after adding denotes the number of cuts added at iteration j. By Theorem 5.5 and the observation (24) the following inequality holds log Theorem 10 of [20] can be used here, with p - n denotes the maximum number of cuts generated by any call to the oracle. Theorem 5.6 The algorithm stops with a solution as soon as k satisfies: Furthermore the number of damped Newton steps per call to the oracle is O(p log(p+ 1)). The number of cutting planes generated is at most O The assumption that p - n is not required in the proof of [20], and in fact would still lead to O cutting planes (this would only impact the constant). 6 Computing the optimal direction of restora- tion The restoration direction requires the solution of the concave problem log We note that in the computation of the restoration direction a significant absence of symmetry occurs: it is easy to give a feasible value for fi, say or that gives a feasible solution to the problem of finding a feasible direction, but, in general, this is not the case for the dual side. If V is invertible, then the dual direction could also be computed by maximizing log 1 The notation O indicates that lower order terms are ignored. A good starting value for fl could also be given, say or , with - 2 D being the diagonal of V \Gamma1 . The following bounds on F (fi) will be useful in the computation of complexity estimate of a Newton method to solve (13). Theorem 6.1 For log and Proof The inequality on F (fi 0 ) was derived in the proof of theorem 5.5. (See (22).) Let us construct an upper bound on F (fi ) where fi denotes the optimal solution of problem (13), and fl . From the optimality condition log log(fi Hence, F (fi log By Lemma 3.8, an homothety of Dikin's ellipsoid contains the current set of localization, i.e., ae oe By assumption (2.1) and the fact that the algorithm has not terminated, a sphere of radius " is contained in FD . Hence, contains a sphere of radius "(1\Gamma') (m+1)(1+') . Denoting by y c the center of this sphere, and selecting one has log with And thus If V is invertible, one can derive alternative upper bounds on F (fi ) as follows. Using we have log (setting If instead of log(- D The bounds on F (fi 0 ) and F (fi ) are used to derive a complexity estimate for the computation of an approximate optimal solution. Using the fact that the function F is self-concordant [15], we can resort to a potential increase scheme. The scheme uses the Newton direction \Gamma[F Let us denote kuk the norm of an arbitrary vector u in the metric induced by the positive definite matrix H . The norm a critical role in the analysis. The potential increase scheme is based on an extension of lemma 3.3. The proof can be found in the unpublished lecture notes [14]. (The proof is also made available in [16].) Lemma 6.2 Let \Deltafi be such that k\Deltafik [F 00 (fi)] with Assume now and satisfies the condition of the above lemma. Thus, easily shows that is bounded from below by an absolute constant. The complexity estimate for the potential increase scheme follows directly from the above analysis and a bound on the achievable potential increase Theorem 6.3 Let fi . The potential increase algorithm applied to the maximization of F produces a point fi such that in a number of iterations not greater than6 6 6 Remark 6.1 The total number of Newton steps involved in the computation of all the approximate optimum directions can easily bounded by m k log(1="), using theorem (6.3), where k is the number of calls to the oracle at termination, long step argument similar to the one given in [20] could most likely be used to reduce this bound. Looking at every iteration individually, and using the fact that A T y - c contains the cutting planes 0 - y - e, we can assert that and hence This indicates that, in practice, the number of iterations needed at each iteration to compute the optimal fi should not increase with the number of cutting planes. It remains to prove that the potential increase scheme yields a solution fi that meets the proximity condition kpfi(V used in Theorem 4.3. In other words, we must show that for j small enough the condition implies To this end, we adapt some results and proofs of [3] developed for quadratic programming. We then relate a few critical norms. Lemma 6.4 Let The following inequality holds Proof Since This proves the left-hand side inequality. As V is positive semidefinite, one has Therefore We can now prove the main result of the section. Lemma 6.5 Assume Proof Since over, we get We conclude that The above lemma shows that once the condition met, one more Newton step is enough to generate point satisfying and thus, by Theorem 4.3, a point 7 Conclusion In this paper, we defined an efficient direction to restore primal and dual feasibility and centrality after adding p new central cuts simultaneously. The direction is efficient in the sense that it maximizes the the product of the new variables brought into the primal or the dual potentials, under the constraints that the other variables remains within the Dikin ellipsoid. The computation of the optimal direction takes place in a space of dimension p equal to the number of cuts added at a given iteration. If p is sufficiently smaller than n, then significant gains in efficiency can be expected. The analysis has been derived under the assumption that the cuts are central. If deep cuts are present, which is to be expected in practice, primal feasibility can always be recovered but dual feasibility appears difficult to achieve in general, except by the use of a primal Newton method. One could then extend the long step argument of [8] in the case of one deep cut to multiple deep cuts. The implementation of ACCPM [11] uses e. Other choices using the variance-covariance matrix V , if it is invertible, have been proposed in [10], and the analysis of this paper actually strengthens that line of thinking. Both the heuristic and optimal choices for fi and fl need to be tested in practice, and rigorous extensions to multiple deep cuts deserve a more thorough study. --R "A cutting plane algorithm that uses analytic centers" "A Cutting Plane Method from Analytic Centers for Stochastic Programming" Interior Point Approach to Linear "Solving Non-linear Multicommodity Flows Problems by the Analytic Center Cutting Plane Method" "Decomposition and non-differentiable optimization with the projective algorithm" "Complexity analysis of an interior cutting plane for convex feasibility problems" "Using the Primal Dual Infeasible Newton Method in the Analytic Center Method for Problems Defined by Deep Cutting Planes" "Shallow, deep and very deep cuts in the analytic center cutting plane method" "A two-cut approach in the analytic center cutting plane method" "Warm start of the Primal-Dual Method Applied in the Cutting Plane Scheme" "ACCPM - A Library for Convex Optimization Based on an Analytic Center Cutting Plane Method" "Analysis of a Cutting Plane Method That Uses Weighted Analytic Center and Multiple Cuts" "Cutting plane algorithms from analytic centers: efficiency estimates" Introductory lectures on Convex Optimization. Interior Point Polynomial Algorithms in Convex Programming Homogeneous Analytic Center Cutting Plane Methods for Convex Problems and Variational Inequalities "On Updating the Analytic Center after the Addition of Multiple Cuts," Theory and Algorithms for Linear Optimization: An Interior point Approach "A potential reduction algorithm allowing column genera- tion" "Complexity Analysis of the Analytic Center Cutting Plane Method That Uses Multiple Cuts" Interior Point Algorithms: Theory and Analysis. --TR --CTR Olivier Pton , Jean-Philippe Vial, Multiple Cuts with a Homogeneous Analytic Center Cutting Plane Method, Computational Optimization and Applications, v.24 n.1, p.37-61, January Fernanda Raupp , Clvis Gonzaga, A Center Cutting Plane Algorithm for a Likelihood Estimate Problem, Computational Optimization and Applications, v.21 n.3, p.277-300, March 2002 Shu-Cherng Fang , Soon-Yi Wu , Jie Sun, An Analytic Center Cutting Plane Method for Solving Semi-Infinite Variational Inequality Problems, Journal of Global Optimization, v.28 n.2, p.141-152, February 2004
cutting plane method;self-concordance;primal Newton algorithm;interior-point methods;analytic center;multiple cuts
589112
On the Accurate Identification of Active Constraints.
We consider nonlinear programs with inequality constraints, and we focus on the problem of identifying those constraints which will be active at an isolated local solution. The correct identification of active constraints is important from both a theoretical and a practical point of view. Such an identification removes the combinatorial aspect of the problem and locally reduces the inequality constrained minimization problem to an equality constrained problem which can be more easily dealt with. We present a new technique which identifies active constraints in a neighborhood of a solution and which requires neither complementary slackness nor uniqueness of the multipliers. We also present extensions to variational inequalities and numerical examples illustrating the identification technique.
Introduction In this paper we consider the problem of identifying the constraints which are active at an isolated stationary point - x of the nonlinear program where it is assumed that the functions f are at least continuously differentiable. More specifically, we are interested in the following question: Given an (x; -) 2 IR n+m belonging to a sufficiently small neighborhood of a Karush-Kuhn-Tucker (KKT) point of Problem (P), is it possible to correctly estimate, on the basis of the problem data in x, the set of indices I 0 := fij g i of the active constraints? The correct identification of active constraints is important from both a theoretical and a practical point of view. Such an identification, by removing the difficult combinatorial aspect of the problem, locally reduces the inequality constrained minimization problem to an equality constrained one which is much easier to deal with. In particular, the study of the local convergence rate of most algorithms for Problem (P) implicitly or explicitly depends on the fact that I 0 is eventually identified. The identification of the active constraints is not difficult if strict complementarity holds at the solution, see the discussion in the next section. However, as far as we are aware of, to date no technique can successfully deal with the case in which the complementary slackness assumption is violated, except in the case of linear programs, see [10]. In this paper we present a new technique which, under mild as- sumptions, correctly identifies active constraints in a neighborhood of a KKT point. This technique appears to improve on existing techniques. In particular, it enjoys the following properties: (i) It is simple and independent of the algorithm used to generate the point (x; -). (ii) It does not require complementary slackness. (iii) It does not require uniqueness of the multipliers. (iv) It does not rely on any convexity assumption. (v) In the case of unique multipliers it also permits the correct identification of strongly active constraints. (vi) The identification technique can be applied also to the Karush-Kuhn-Tucker system arising from variational inequalities. IDENTIFICATION OF ACTIVE CONSTRAINTS 3 Strategies for identifying active constraints are part of the optimization folklore [2, 13, 15], however, they almost invariably lack many of the good characteristics listed above. In the last ten years a special attention has been devoted to this problem in the field of interior point methods for linear programs; we refer the reader to the survey paper [10]. Recent works on the nonlinear case include [9, 11, 23], where the case of box constraints is considered, and [12, 36], where the general nonlinear case is studied. Related material can also be found in [4, 5, 6], where the problem of establishing whether or not a sequence fx k g, converging to a solution - x, in some way identifies the set I 0 is dealt with. Note, however, that in these latter papers no explicit rule is given in order to identify the active constraints from a close, but arbitrary point. We remark that, in order to identify the active set, we suppose we are given a pair (x; -) of primal and dual variables. If we think of algorithmic applications of the results in this paper, we stress that most algorithms will produce a sequence of primal and dual variables. Even in the rare cases in which this does not occur, it is usually possible, under reasonable assumptions, to generate a continuous dual estimate by using a multiplier function, see, e.g., [12] and references therein, as well as Section 4. This paper is organized as follows. In the next section we introduce the identification technique and prove its main properties. The identification technique critically depends on the definition of what we call identification function. Therefore, the more technical Section 3 is devoted to the definition of identification functions under different sets of assumptions. In Section 4 we use the results of the previous sections in order to define a local active set Newton-type algorithm for the solution of inequality constrained optimization problems for which a Q-quadratic convergence rate of the primal variables can be proved under very weak conditions. In Section 5 we give some final comments. We conclude this section by providing a list of the notation employed. Through-out the paper, k \Delta k indicates the Euclidean vector norm. The symbol B ffl denotes the open Euclidean ball with radius ffl ? 0 and center at the origin; the dimension of the space will be clear from the context. The Euclidean distance of a point y from a nonempty set S is abbreviated by dist[y; S]. We write x+ for the vector maxf0; xg; where the maximum is taken componentwise. We set I := use of the notation x J for J ' I in order to represent the jJ j-dimensional vector with components Finally, the transposed Jacobian of the vector-valued mapping g at a point x will be denoted by rg(x); i.e., the ith column of this matrix is the gradient rg i (x): Active Constraints Following the usual terminology in constrained optimization, we call a vector - a stationary point of (P) if there exists a vector - 2 IR m such that (-x; - -) solves the 4 F. FACCHINEI, A. FISCHER AND C. KANZOW Karush-Kuhn-Tucker system (1) The pair (-x; -) is called a KKT point of Problem (P). In the sequel - x will always denote a fixed, isolated stationary point, so that there is a neighborhood of - x which does not contain any further stationary point of (P). Moreover, we shall indicate by the set of all Lagrange multipliers - - associated with - x and by K the set of all KKT points associated with - x, that is, -) solves (1)g; K := f(-x; - The set is closed and convex and therefore, so is the set K. Gauvin [14] showed that is bounded (and hence compact) if and only if the Mangasarian-Fromovitz constraint qualification (MFCQ) is satisfied, i.e., if and only if On the other hand, Kyparisis [22] showed that reduces to a singleton if and only if the strict Mangasarian-Fromovitz constraint qualification (SMFCQ) holds, i.e., if and only if denotes the index set I In particular, the linear independence constraint qualification (LICQ), i.e., the linear independence of the gradients of the active constraints, implies that is a singleton. Our basic aim is to construct a rule which is able to assign to every point (x; -) an estimate A(x; -) ' I so that A(x; lies in a suitably small neighborhood of a point (-x; -) 2 K. Usually estimates of this kind are obtained by comparing the value of g i (x) to the value of - i . For example, it can easily be shown that the set I \Phi (x; -) := coincides with the set I 0 for all (x; -) in a sufficiently small neighborhood of a KKT point (-x; -) which satisfies the strict complementarity condition. If this condition is violated, then only the inclusion I \Phi (x; -) ' I 0 (2) IDENTIFICATION OF ACTIVE CONSTRAINTS 5 holds. Furthermore, if is a singleton, then we also have, in a sufficiently small neighborhood of (-x; - -), I This relation was exploited to construct locally superlinearly convergent QP-free optimization algorithms when the unique multiplier - does not satisfy the strict complementarity condition, see, e.g., [12, 20, 35]. We refer the reader to [2, 12] and references therein for a more complete discussion of these kind of results. An analysis of results established in the literature shows that this conclusion holds in general: if strict complementarity is satisfied, it is usually possible to correctly identify the active constraint set, otherwise a relation like (3) is the best result that has been established in the general nonlinear case. To overcome this situation we propose to compare g i (x) to a quantity which goes to 0 at a known rate if (x; -) converges to a point in the KKT set K. To this end, we introduce the notion of identification function. called identification function for (a) ae is continuous on K, (b) (-x; -) 2 K implies ae(-x; - (c) if (-x; - -) belongs to K, then lim ae(x; -) dist [(x; -); K] In the next section we shall give examples of how to build, under appropriate as- sumptions, identification functions. Basically, Definition 2.1 says that a function is an identification function if it goes to 0 when approaching the set K at a "slower" rate than the distance from the set K. We note that dist [(x; -); K] ? 0 whenever since K is a closed set; hence the denominator in (4) is always nonzero. Using Definition 2.1 it is easy to prove that the index set correctly identifies all active constraints if (x; -) is sufficiently close to the KKT set K. Theorem 2.2 Let ae be an identification function for K. Then, for any - 2 , an exists such that 6 F. FACCHINEI, A. FISCHER AND C. KANZOW Proof. Since g is continuously differentiable, g is locally Lipschitz-continuous. Hence there exists a constant c ? 0 such that, for all x sufficiently close to - Suppose now that g i using (4) and (7), we obviously have, for in a sufficiently small neighborhood of (-x; -), dist [(x; -); K] - ae(x; -); so that, by (5), i 2 A(x; -). If, instead, (x; -) 2 K, then we have x by the local uniqueness of - x. From the definition of an identification function, we also have ae(x; so that and also in this case i 2 A(x; -). On the other hand, if by continuity, that i 62 A(x; -) if (x; -) is sufficiently close to (-x; - -). Therefore, for any - 2 , we can find such that (6) is satisfied. 2 From the previous theorem it is obvious that there exists an open set containing K where the identification of the active constraints is correct. Using the MFCQ condition we can obtain a somewhat stronger result. Theorem 2.3 Let ae be an identification function for K. If the MFCQ condition holds, then there is an ffl ? 0 such that Proof. By the previous Theorem, for every (-x; -) 2 K, there exists a neighborhood -) such that A(x; -)). The collection of open obviously forms an open cover of K. Since the set K is compact in view of the MFCQ condition, we can extract from the infinite cover - such that (-x; -) 2 K a finite subcover \Omega\Gamma ffl( - s: Then it is easy to see that the Theorem holds with ffl := min j=1;:::;s fffl( - the SMFCQ holds, it is even possible to identify the set of strongly active constraints at - x, i.e., the set of constraints whose multipliers are positive. To this end, let be defined by The following theorem holds. IDENTIFICATION OF ACTIVE CONSTRAINTS 7 Theorem 2.4 Let ae be an identification function for K. If the SMFCQ holds at - x, then there is an ffl ? 0 such that Proof. We first recall that the SMFCQ implies that reduces to a singleton, i.e., -g. Theorem 2.2 shows that A+ (x; -) ' I 0 for all (x; -) in a certain neighborhood of (-x; -). Now, consider an index . By continuity, this implies in a sufficiently small neighborhood of (-x; -). On the other hand, let in a sufficiently small neighborhood of -), we have dist [(x; -); K] - ae(x; -)=2 ! ae(x; -): This means i 62 sufficiently close to Until now we made reference to the Karush-Kuhn-Tucker system (1) which expresses first order necessary optimality conditions for the minimization Problem (P). We showed how the active constraints associated to an isolated stationary point - x can be identified. However, the fact that the Karush-Kuhn-Tucker system (1) derives from an optimization problem plays no role in the theory developed. What we actually proved is the following: Given a solution (-x; - -) of a system with the structure of system (1) and with an isolated x-part, we can identify, in a suitable neighborhood of this solution, those inequalities which hold as equalities at the solution (-x; -). Therefore, if we consider the KKT system continuous function, the theory of this section goes through without any change. This is an important observation, since it allows us to extend the theory developed so far to the identification of active constraints for the variational inequality problem: Find - x 2 X such that F (-x) T is continuous and is continuously differentiable. It is well known that, under a standard regularity assumption [17], a necessary condition for - to be a solution of the variational inequality problem is that - 2 IR m exists such that (-x; -) solves system (8). There- fore, if we have a sequence f(x converging to a solution of system (8) which has an isolated primal part, we can apply the techniques described in this section in order to identify which of the constraints g i (x) - 0 will be active at - 8 F. FACCHINEI, A. FISCHER AND C. KANZOW 3 Defining Identification Functions From what exposed in the previous section we see that the crucial point in the identification of active constraints is the definition of an identification function. In this section we show how it is possible to define such a function for Problem (P): We consider three cases. In the first one we assume that the functions f and g are analytic, in the second case we require them to be LC 1 , but then we also need that the MFCQ condition and a second order sufficient condition for optimality are satisfied. Finally, in the third case, the functions are required to be C 2 and the KKT point is assumed to satisfy a regularity condition related to (but weaker than) Robinson's strong regularity [33] and which we call quasi-regularity. Extensions of these results to the KKT system (8) are possible. We shall point out the relevant changes in corresponding remarks. The cases considered here do not cover all the situations in which an identification function can be defined and computed, but they certainly show that the definition and computation of an identification function is possible in most of the cases of interest. 3.1 The Analytic Case Let f and each g i (i 2 I) be analytic around a point x. We recall that this means that f and each g i (i 2 I) possess derivatives of all orders and that they agree with their Taylor expansions around x. We say that f and each g i (i 2 I) are analytic on an open set X ' IR n if they are analytic around each x 2 X: We shall make use of the following result due to Lojasiewicz, Luo and Pang [25, 27]. Theorem 3.1 Let S denote the set of points in IR r satisfying are analytic functions defined on an open set Suppose that S 6= ;. Then, for each compact ae X, there exist Using this result, it is possible to define an identification function for Problem (P): Theorem 3.2 Suppose that f and g are analytic in a neighborhood of a stationary point - x. Then, the function ae defined by log(r(x;-)) if r(x; -) 2 (0; 0:9); IDENTIFICATION OF ACTIVE CONSTRAINTS 9 where is an identification function for K. Proof. It is obvious, by definition, that ae 1 is a nonnegative function such that lim so that ae 1 is also continuous on K. Hence we only have to check the limit lim dist [(x; -); K] To this end we recall that, for arbitrary - ? 0 and fl ? 0, the limit lim holds, see, e.g., [28, p. 328]. We can now apply Theorem 3.1 by considering the system (1) which defines KKT points. It is then easy to see that (9) yields, for every given compact containing (-x; -) in its interior, where - and fl are fixed positive constants. Therefore we can write lim dist [(x; -); K] from which (11) follows taking into account (12), recalling the definition of ae 1 and noting that r(x; -) is a continuous function that goes to 0 from the right as (x; -) tends to (-x; -). 2 We stress that Theorem 3.2 holds under the mere assumption that f and g are analytic. Remark 3.3 If we want to define an identification function for the solutions of the KKT system (8), we only have to substitute the definition of the residual (10) by the following one: Obviously, also in this case, we have to assume that F and each g i (i 2 I) are analytic in a neighborhood of the KKT point under consideration. F. FACCHINEI, A. FISCHER AND C. KANZOW 3.2 The Second Order Condition Case In this subsection we assume that f and g are LC 1 , i.e., that they are differentiable with Lipschitz-continuous derivatives. We denote the Lagrangian of problem (P) by and write r x L(x; -) for the gradient of L with respect to the x-variables. We further assume that the MFCQ holds along with the following second order sufficient condition for optimality: Assumption 3.4 There is Here, W ( -) denotes the cone and @ x r x L(-x; - -) denotes Clarke's [8] generalized Jacobian with respect to x of the gradient r x L, calculated at (-x; - -). We remark that, if the functions f and g are twice continuously differentiable and only one multiplier exists, then the previous definition reduces to the classical KKT second-order sufficient condition for optimality. We shall show that these two conditions allow us to define an identification function for K which, because the MFCQ holds, is a compact set. To this end consider the perturbed nonlinear program denotes the perturbation parameter. In what follows we will assign to any vector (y; -) 2 IR n \Theta IR m a particular perturbation vector For this purpose we first define the function componentwise by where, we recall, I \Phi (y; g. We can now introduce the function \Gammag i (y) if i 2 I \Phi (y; -); Although, in general, the functions - \Phi and - are not everywhere continuous, the following properties can be proved. IDENTIFICATION OF ACTIVE CONSTRAINTS 11 Lemma 3.5 (a) If - 2 , then - \Phi (-x; -. (b) The function - \Phi is continuous on K. (c) If - 2 , then -x; (d) The function - is continuous on K. Proof. (a) Since (-x; -) is a KKT point, it readily follows that I \Phi (-x; the definition of the function - \Phi yields - \Phi (-x; - for all - 2 . (b) Let (-x; - -) belong to K. According to assertion (a), in order to show continuity of - \Phi in (-x; - -), we have to show that, for every i 2 I, lim If - easily follows from the definition of - \Phi sufficiently large, continuity. Using the definition of - \Phi again, we have - \Phi i for all k large enough. Thus, (14) follows also in this case. (c) Taking into account assertion (a), the KKT conditions (1) for problem (P) yield On the other hand, analogously to point (a), we have I \Phi (-x; readily follows from the definition of - g . (d) The continuity of the function - f on K follows by its definition and assertion (b). In order to prove the continuity of - g on K; let (-x; - be given and let be any sequence converging to (-x; - -). In view of part (c), we have to show that lim To this end, first consider an index i 2 I Hence (15) follows from the definition of - On the other hand, if sufficiently large. Hence - g (y k all these indices, i.e., (15) holds also for i 62 I Using the particular perturbation vector -), we can prove the following result. Lemma 3.6 Let (y; -) 2 IR n \Theta IR m be arbitrarily chosen. Then, (y; - \Phi (y; -)) is a KKT point for problem (P(t)); where Proof. The KKT system for the perturbed program (P(t)) reads as follows: 12 F. FACCHINEI, A. FISCHER AND C. KANZOW Let (y; -) be arbitrary but fixed. Obviously, since we find that (x; -) := solves (16) and (17). Now, we will show that (y; - \Phi (y; -)) also satisfies (18) and (19). For i 2 I \Phi (y; -) the definition of - g (y; -) yields (g(y) that both (18) and (19) are fulfilled. If, instead, i 2 I n I \Phi (y; -), it follows from the definition of - \Phi (y; -) that - \Phi and (19) is satisfied. Moreover, the definition of - g (y; -) implies Thus, (18) is also valid for i 2 I n I \Phi (y; -). We therefore conclude that (y; - \Phi (y; -)) is a KKT point of (P(t)) when The next result can easily be derived from Theorem 4.5 b) and formula (3.2 f) in Klatte [18]. If the functions f and g of the program (P) are twice continuously differentiable it can also be obtained from a corresponding result in Robinson [34, Corollary 4.3]. We further note that Assumption 3:4 can be weakened by using generalized directional derivatives, see [18] for more details and references. Theorem 3.7 Let the MFCQ and Assumption 3.4 be satisfied. Then, there are for every t 2 B ffi and for every KKT point (-x(t); - -(t)) of problem (P(t)) for which Putting together the last two results, we can easily prove the following theorem. Theorem 3.8 Let the MFCQ and Assumption 3.4 be satisfied. Then, there are Proof. By Lemma 3.5 (c), we have -x; Lemma 3.5 (d) and by the compactness of K, we have that, for ffi from Theorem 3.7, we can find an " ? 0 such that, if Therefore, since ffl - j (with j from Theorem 3.7) can be assumed without loss of generality, Theorem 3.7 together with Lemma 3.6 yields the desired result. 2 We are now in the position to show that the function ae defined by can be used as an identification function. Theorem 3.9 Let the MFCQ and Assumption 3.4 be satisfied. Then ae 2 is an identification function for K. IDENTIFICATION OF ACTIVE CONSTRAINTS 13 Proof. By Lemma 3.5 we easily obtain that ae 2 is continuous on K and that any sequence with lim Using Theorem 3.8 we get, for k sufficiently large, Let z 1 2 K and z 2 2 K be the projections of tively, on the closed convex set K. Then, using the triangle inequality, we get Combining relations (21) and (22), we obtain, for k sufficiently large, ae 2 is continuous on the compact set K and since its value is 0 on K it follows from (20), (21) and Lemma 3.5 (b) that the quantity dist[(x goes to 0 for k ! 1. But then the right hand side of (23), and thus also the left hand side, tends to infinity. Therefore, we have shown that ae 2 possesses all properties of an identification function. 2 If, instead of the upper Lipschitz-continuity as stated in Theorem 3.7, the multi-function t 7! K(t) is upper H-older-continuous at with a known rate - 2 (0; 1], that is, if, for some dist [(-x(t); -(t)); K] - cktk - for every t 2 B ffi and for every KKT point (-x(t); - -(t)) of Problem (P(t)) for which then the technique presented in this subsection can easily be extended if we define ae In particular, Theorems 3.8 and 3.9 remain valid for this ae 2 if Assumption 3.4 is replaced by the upper H-older-continuity. An interesting case in which it is possible to prove, under an assumption weaker than Assumption 3.4, the upper H-older-continuity at of the multifunction 14 F. FACCHINEI, A. FISCHER AND C. KANZOW t 7! K(t) is the case of convex problems. Assume that f is convex and each g i (i 2 I) is concave, that the MFCQ holds and that the following growth condition holds (in place of Assumption 3.4): positive - exist such that Under these assumptions and using the results in [19], it is possible to show (we omit the details) that ffi ? exist such that dist [(-x(t); -(t)); K] - c ktk for every t 2 B ffi and for every KKT point (-x(t); -(t)) of Problem (P(t)) for which It may be interesting to note that the growth condition holds, in particular, if Assumption 3.4 is fulfilled. Remark 3.10 The extension of the results of this section to general KKT systems is not straightforward, since the sensitivity analysis of perturbed KKT systems re- quires, to date, stronger assumptions. The key point is to establish a result analogous to Theorem 3.7. Once this has been done, we can easily prove theorems analogous to Theorem 3.9 by substituting F to rf in every relevant formula. As an example of the kind of the results that can be obtained we cite the following one. Suppose that F is C 1 and g is C 2 . Assume also that the SMFCQ holds at - x along with Assumption 3.4. Then, according to [16, Corollary 8 (c)], Theorem 3.7 holds and therefore ae 2 is a regular identification function for the KKT system (8). 3.3 The Quasi-Regular Case In this subsection we assume that the functions f and g are C 2 . We shall introduce a condition which we call quasi-regularity. As will be clear later, this quasi-regularity is related to , but weaker than Robinson's strong regularity [33]. In order to motivate the definition of a quasi-regular KKT point we will first recall a condition which is equivalent to the notion of a strongly regular KKT point. To this end we shall use the index set I 00 := I 0 n I + of all those indices for which the strict complementarity condition does not hold at the KKT point (-x; -). For any J ' I 00 (empty set included) introduce the matrix xx L rg+ rg J \Gammarg T \Gammarg T xx L, rg+ and rg J are abbreviations for the matrices r 2 -), rg I+ (-x) and rg J (-x), respectively. The following result is due to Kojima et al. [21]. Theorem 3.11 The following statements are equivalent: IDENTIFICATION OF ACTIVE CONSTRAINTS 15 (a) (-x; -) is a strongly regular KKT point. (b) For any J ' I 00 (empty set included), the determinants of the matrices M(J) all have the same nonzero sign. Motivated by point (b) in Theorem 3.11, we introduce the following definition. Definition 3.12 The KKT point (-x; - -) is a quasi-regular point if the matrices M(J) are nonsingular for every J ' I 00 (empty set included). Note that, in view of Theorem 3.11, quasi-regularity is implied by Robinson's strong regularity condition, but the converse is not true. In fact, consider the following example: It is easy to check that - is a global minimizer and that the Lagrange multipliers of the two constraints are both zero, so that I Therefore (-x; 0; 0) is a quasi-regular KKT point, but not a strongly regular one. Note that in this example the KKT point is an isolated KKT point. This is not a chance. In fact we shall show in this section that quasi-regularity of a KKT point implies its local uniqueness. It is also worth pointing out that quasi-regularity implies the linear independence of the active constraints. This easily follows from the fact that Now let us introduce the operator r x L(x; -) Note that the KKT conditions are equivalent to the nonlinear system of equations By the differentiability assumption we have that \Phi is locally Lipschitzian. Hence, by Rademacher's Theorem, \Phi is differentiable almost everywhere. Denote by D \Phi the set of points where \Phi is differentiable. Then we can define the B-subdifferential (see, e.g., [31]) of \Phi at (x; -) as F. FACCHINEI, A. FISCHER AND C. KANZOW Note that the B-subdifferential is a subset of Clarke's generalized Jacobian [8, 31]. The next lemma illustrates the structure of the B-subdifferential of \Phi: Before stating this lemma, however, we introduce three index sets: Lemma 3.13 Let (x; -) 2 IR n+m be arbitrary. Then xx L(x; -) rg(x)D a (x; -) where D a (x; -) := diag (a 1 are diagonal matrices with a and D b (x; Proof. This follows immediately from the definition of the operator \Phi: 2 We are now in the position to prove the following result. Lemma 3.14 Let (-x; - n+m be a quasi-regular KKT point. Then all matrices are nonsingular. Proof. Let In view of Lemma 3.13, there exists an index set -) such that \Gammarg T \Gammarg T \Gammarg T \Gammarg T denotes the complement of J in the set fi(-x; -): Obviously, this matrix is nonsingular if and only if the matrixB @ xx L rg ff rg J \Gammarg T \Gammarg T IDENTIFICATION OF ACTIVE CONSTRAINTS 17 is nonsingular. In turn, this matrix is nonsingular if and only if the matrix M(J) is nonsingular. Hence the thesis follows immediately from Definition 3.12. 2 We are now able to prove the main result of this subsection. Theorem 3.15 Let (-x; - n+m be a quasi-regular KKT point of problem (P). Then, (a) (-x; -) is an isolated KKT point, (b) the function ae 3 : defined by ae 3 (x; -) := is an identification function for -)g: Proof. Obviously, ae 3 is a continuous and nonnegative function with ae 3 (-x; - Furthermore, since f and g have locally Lipschitzian gradients and the min operator is semismooth (see [29, 32] for the definition of semismoothness and [29] for the proof that the min operator is semismooth) it follows that also \Phi, which is the composite of semismooth functions, is semismooth [29, 32]. Hence it follows from Lemma 3.14 and [30, Proposition 3] that there exists a constant c ? 0 such that for all (x; -) in a neighborhood of (-x; - only if (x; -) is a KKT point, part (a) follows immediately. From (24) we also get ae 3 (x; -) c and therefore lim ae 3 (x; -) i.e., ae 3 is an identification function. 2 Remark 3.16 In the case of the KKT system (8) everything goes through. It is sufficient to assume that F is continuously differentiable and to substitute everywhere the gradient r x L(x; -) by the function F Also in this case the definition of quasi-regularity is related to and weaker than that of a strongly regular KKT point since Theorem 3.11 carries over to the KKT system (8), see Actually, the case of KKT systems of variational inequalities is probably the main case in which quasi-regularity can be applied. In fact, it is F. FACCHINEI, A. FISCHER AND C. KANZOW not difficult to see that, if strict complementarity holds and - x is a local minimum point of Problem (P), quasi-regularity implies the conditions of the previous sub- section. However, these conditions and quasi-regularity are fairly distinct if one considers variational inequalities. For example, it can easily be checked that, given the variational inequality defined by the function F and the set 0g, the point (0; is a quasi-regular solution but does not satisfy the conditions stated in Remark 3.10 of the previous subsection. 4 An Application In this section we apply the results obtained in the previous sections to a local active-set Newton algorithm for the solution of Problem (P). The algorithm to be introduced here is a simple variation of the one presented in [12]. However, using the new results obtained in this work, we are able to relax the assumptions used in [12]. The result is an algorithm which, by solving only linear systems at each iteration, guarantees Q-quadratic convergence of the sequence fx k g to the solution under very mild assumptions and without requiring strict complementarity. We remark that, as far as we are aware of, there exist only two other algorithms which ensure Q- quadratic convergence of the primal variables. The first one is due to Bonnans [3] and requires, at each iteration, the solution of a quadratic and possibly nonconvex subproblem. Furthermore the algorithm of Bonnans also requires the selection of a suitable solution of the quadratic subproblem, which appears to be a difficult task in practice. The other algorithm that guarantees Q-quadratic convergence is the one discussed in [12] which, as already said, requires stronger assumptions. We refer the interested reader to [12] for a more detailed discussion of these issues. Consider problem (P) and assume that f and g are twice continuously differen- tiable. The algorithm we consider generates a sequence fx k g as with d k being obtained by solving the linear system z k In the previous system, where N(x) is the m \Theta m matrix defined by: diag i2I while IDENTIFICATION OF ACTIVE CONSTRAINTS 19 We shall assume that the LICQ holds at - x, along with the following weak second order assumption. Assumption 4.1 It holds that We note that Assumption 4.1 is extremely weak if compared to second order assumptions usually used in the local analysis of algorithms for the solution of inequality constrained optimization problems. In particular, even if coupled with the linear independence assumption, it does not even imply that - x is an isolated local solution of problem (P). This can be checked on the following example, where a is a nonnegative constant: t. x 2 - 0: It is easy to see that - is a stationary point satisfying both the LICQ assumption and Assumption 4.1. However, if a ? 0, - x is not a local solution, while, if we have that - x is indeed a local solution but not isolated. We recall a result which illustrates the properties of the multiplier function defined by (27). Theorem 4.2 (see [26]) Let - x be a KKT point where the linear independence of the active constraints holds. Then, - and there exists ffl 1 ? 0 such that, for all (a) -(x) is well defined; (b) -(x) is continuously differentiable. We now pass to the proof that the algorithm (25)-(26) is Q-quadratically convergent in the primal variables. To this end we first need a simple Lemma. Lemma 4.3 Let (-x; - -) be a KKT pair for Problem (P) which satisfies the LICQ and Assumption 4.1. Then there exist the matrix xx L(x; -(x)) \Gammarg I 0 rg I 0 is nonsingular and kM(x) Proof. Assumption 4.1 and well known properties of quadratic forms (see, e.g., [1, p. 78]) imply that there exists a constant oe ? 0 such that the matrix (-x)rg I 0 F. FACCHINEI, A. FISCHER AND C. KANZOW is positive definite. Then, by continuity, the matrix is positive definite for all x 2 f-xg +B ffl 2 sufficiently small. This implies (see, e.g., [1, p. 78]) that, for all x 2 f-xg +B ffl 2 Therefore (29) and the linear independence assumption imply that ffl 2 ? 0 can be chosen as small as necessary so that, for all x 2 f-xg +B ffl 2 and rg I 0 using (30) and (31), it is easy to show that the matrix M(x) is nonsingular for . Hence, the remaining result follows by the continuity of M(x). 2 Theorem 4.4 Let f and g i (i 2 I) be twice continuously differentiable with locally Lipschitz-continuous Hessian matrices r 2 f and r 2 g i (i 2 I). Let (-x; - -) be an isolated KKT pair for Problem (P) which satisfies the LICQ and Assumption 4.1, and suppose that an identification function ae for known. Then there exists the system (26) is nonsingular and the sequence fx k g produced by (25) satisfies converges to - x, and the rate of convergence is Q-quadratic. Proof. For us consider the linear system with I 0 is arbitrary but fixed. Recalling that r x L(-x; I 0 0: (33) Taking into account (33) and the differentiability assumptions on f and each g i (i 2 I), it is possible to show (see [12] for the details), by repeated use of Taylor's formula, that positive numbers ffl 3 , C 1 , and C 2 exist such that, for all x 2 f-xg +B ffl 3 IDENTIFICATION OF ACTIVE CONSTRAINTS 21 By Lemma 4.3, (32) and (33), we have, for all x 2 f-xg +B ffl 3 Moreover, if ffl 3 ? 0 is small enough, it follows from Theorem 4.2 that C 3 ? 0 exists such that, for all x 2 f-xg +B ffl 3 Assume now that and suppose that ffl 3 2 (0; 1) is chosen small enough so that, according to Theorem 2.2, Therefore, setting the linear systems (26) and (32) are equivalent as long as x k . By (34)-(36), we also have, if x k 2 f-xg +B ffl 3 xk: From these relations, all the assertions of the theorem easily follow by induction. 2 We stress that the example in Section 3.3 satisfies the assumptions of the previous theorem, but not those of the corresponding Theorem 4.1 in [12]. Final Remarks In this paper we introduced a technique to accurately identify active constraints in inequality constrained optimization and variational inequality problems. The most remarkable feature of the new identification technique is that it identifies all active constraints even if strict complementarity does not hold. Furthermore, as discussed in the introduction, it also enjoys several other favorable characteristics. In particu- lar, the identification technique can be used in combination with any algorithm for the solution of inequality constrained optimization or variational inequality prob- lems. In Section 4 we gave an example of an application of the results of this paper to an active set Newton-type method; however, we believe that the techniques introduced in this paper can be useful in many other cases, especially in the theoretical analysis and in the design of optimization methods. From a practical point of view, the following questions may also be of interest: 22 F. FACCHINEI, A. FISCHER AND C. KANZOW (a) How large is the region where exact identification occurs? (b) Can we build identification functions which are scale invariant? (c) Can we relax the assumption that - x is an isolated stationary point and still obtain useful results? It is difficult to answer to these questions at the level of generality adopted in this paper. We think that an answer can come from practical experiments and from an analysis of structured classes of problems, e.g., linear or quadratic problems, box or linearly constrained problems etc. Acknowledgment . We would like to thank Professor D. Klatte for helpful discussions on the stability of KKT-systems. --R Introduction to Matrix Analysis. Constrained Optimization and Lagrange Multiplier Meth- ods Rates of convergence of Newton type methods for variational inequalities and nonlinear programming. On the identification of active constraints II: The nonconvex case. Optimization and Nonsmooth Analysis. Global convergence for a class of trust region algorithms for optimization problems with simple bounds. A study of indicators for identifying zero variables in interior-point methods "La Sapienza" Quadratically and superlinearly convergent algorithms for the solution of inequality constrained minimization problems. Practical Methods of Optimization. A necessary and sufficient regularity condition to have bounded multipliers in nonconvex programming. Practical Optimization. Stability analysis of variational inequalities and nonlinear complementarity problems Nonlinear optimization problems under data perturbations. On quantitative stability for C 1 Strongly stable stationary solutions in nonlinear programs. On uniqueness of Kuhn Tucker multipliers in nonlinear pro- gramming Convergence of trust region algorithms for optimization with bounds when strict complementarity does not hold. Strong stability in variational inequalities. Sur la probl'em de la division. New results on a continuously differentiable exact penalty function. bounds for analytic systems and their applications. Calculus I. Semismooth and semiconvex functions in constrained optimiza- tion Nonsmooth equations: motivation and algorithms. Convergence analysis of some algorithms for solving nonsmooth equa- tions A nonsmooth version of Newton's method. Strongly regular generalized equations. Generalized equations and their solution surfaces in constrained optimization. --TR --CTR Wang , Lifeng Chen , Guoping He, Sequential systems of linear equations method for general constrained optimization without strict complementarity, Journal of Computational and Applied Mathematics, v.182 n.2, p.447-471, 15 October 2005 Lus N. Vicente , Stephen J. Wright, Local Convergence of a Primal-Dual Method for Degenerate Nonlinear Programming, Computational Optimization and Applications, v.22 n.3, p.311-328, September 2002 Huang , Defeng Sun , Gongyun Zhao, A Smoothing Newton-Type Algorithm of Stronger Convergence for the Quadratically Constrained Convex Quadratic Programming, Computational Optimization and Applications, v.35 n.2, p.199-237, October 2006 A. N. Daryina , A. F. Izmailov, On the Newton-type method with admissible trajectories for mixed complementatiry problems, Automation and Remote Control, v.68 n.2, p.351-360, February 2007 Christian Kanzow , Andreas Klug, On Affine-Scaling Interior-Point Newton Methods for Nonlinear Minimization with Bound Constraints, Computational Optimization and Applications, v.35 n.2, p.177-197, October 2006 N. H. Xiu , J. Z. Zhang, Local convergence analysis of projection-type algorithms: unified approach, Journal of Optimization Theory and Applications, v.115 n.1, p.211-230, October 2002 Andreas Fischer , Houyuan Jiang, Merit Functions for Complementarity and Related Problems: A Survey, Computational Optimization and Applications, v.17 n.2-3, p.159-182, December 2000 Naihua Xiu , Jianzhong Zhang, Some recent advances in projection-type methods for variational inequalities, Journal of Computational and Applied Mathematics, v.152 n.1-2, p.559-585, 1 March
active constraints;constrained optimization;degeneracy;variational inequalities;identification of active constraints
589114
Robust Solutions to Uncertain Semidefinite Programs.
In this paper we consider semidefinite programs (SDPs) whose data depend on some unknown but bounded perturbation parameters. We seek "robust" solutions to such programs, that is, solutions which minimize the (worst-case) objective while satisfying the constraints for every possible value of parameters within the given bounds. Assuming the data matrices are rational functions of the perturbation parameters, we show how to formulate sufficient conditions for a robust solution to exist as SDPs. When the perturbation is "full," our conditions are necessary and sufficient. In this case, we provide sufficient conditions which guarantee that the robust solution is unique and continuous (Hlder-stable) with respect to the unperturbed problem's data. The approach can thus be used to regularize ill-conditioned SDPs. We illustrate our results with examples taken from linear programming, maximum norm minimization, polynomial interpolation, and integer programming.
Introduction . A semidefinite program (SDP) consists of minimizing a linear objective under a linear matrix inequality (LMI) constraint; precisely, subject to F (1) - {0} and the symmetric matrices F are given. SDPs are convex optimization problems and can be solved in polynomial time with, e.g., primal-dual interior-point methods [24, 35, 26, 19, 2]. SDPs include linear programs and convex quadratically constrained quadratic programs, and arise in a wide range of engineering applications; see, e.g., [12, 1, 35, 22]. In the SDP (1), the "data" consist of the objective vector c and the matrices In practice, these data are subject to uncertainty. An extensive body of work has concentrated on the sensitivity issue, in which the perturbations are assumed to be infinitesimal, and regularity of optimal values and solution(s), as functions of the data matrices, is studied. Recent references on sensitivity analysis include [30, 31, 10] for general nonlinear programs, [33] for semi-infinite programs, and [32] for semidefinite programs. When the perturbation a#ecting the data of the problem is not necessarily small, a sensitivity analysis is not su#cient. For general optimization problems, a whole field # Received by the editors June 21, 1996; accepted for publication (in revised form) September 22, 1997; published electronically October 30, 1998. http://www.siam.org/journals/siopt/9-1/30571.html Ecole Nationale Sup-erieure de Techniques Avanc-ees, 32, Bd. Victor, 75739 Paris, France (elghaoui@ensta.fr, oustry@ensta.fr, lebret@ensta.fr). 34 L. EL GHAOUI, F. OUSTRY, AND H. LEBRET of study (stochastic programming) concentrates on the case where the perturbation is stochastic with known statistics. One object of this field is to study the impact of, say, a random objective on the distribution of optimal values (this problem is called the "distribution problem"). References relevant to this approach to the perturbation problem include [15, 9, 29]. We are not aware of special references for general SDPs with randomly perturbed data except for the last section of [30], some exercises in the course notes of [13], and section 2.6 in [23]. The main objective of this paper is to quantify the e#ect of unknown but bounded deterministic perturbation of problem data on solutions. In our framework, the perturbation is not necessarily small, and we seek a solution that is "robust," that is, remains feasible despite the allowable, not necessarily small, perturbation. Our aim is to obtain (approximate) robust solutions via SDP. Links between regularity of solutions and robustness are, of course, expected. One of our side objectives is to clarify these links to some extent. This paper extends results given in [16] for the least-squares problem. The approach developed here can be viewed as a special case of stochastic programming in which the distribution of the perturbation is uniform. The ideas developed in this paper draw mainly from two sources: control theory, in which we have found the tools for robustness analysis [36, 17, 12] and some recent work on sensitivity analysis of optimization problems by Shapiro [31] and Bonnans, Cominetti, and Shapiro [10]. Shortly after completion of our manuscript, we became aware of the ongoing work of Ben-Tal and Nemirovski on the same subject. In [7], they apply similar ideas to a truss topology design problem and derive very e#cient algorithms for solving the corresponding robustness problem. In [8], the general problem of tractability of obtaining a robust solution is studied, and "tractable counterparts" of a large class of uncertain SDPs are given. The case of robust linear programming, under quite general assumptions on the perturbation bounds, is studied in detail in [6]. Our paper can be seen as a complement of [8], giving ways to cope with (not necessarily) tractable robust SDPs by means of upper bounds. (In particular, our paper handles the case of nonlinear dependence of the data on the uncertainties.) A unified treatment, and more results, will appear in [4]. The paper is divided as follows. Our problem is defined in section 2. In section 3, we show how to compute upper bounds on our problem via SDP. We give special attention to the so-called full perturbations case, for which our results are nonconservative. In section 4, we examine sensitivity of the robust solutions in the full perturbations case. We provide conditions which guarantee that the robust solution is unique and a regular function of the data matrices. We then consider several interesting examples in section 5, such as robust linear programming, robust norm minimization, and error-in-variables SDPs. 2. Problem definition. 2.1. Robust SDPs. Let F(x, #) be a symmetric matrix-valued function of two variables x # R m . In the following, we consider x to be the decision variable, and we think of # as a perturbation. We assume that # is unknown but bounded. Precisely, we assume that # is known to belong to a given linear subspace D of R p-q , and in addition, #, where # 0 is given. In section 2.2, we will be more precise about the dependence of F on #. We define the robust feasible set by for every # D, #, F(x, #) is well defined and F(x, # 0 # . (2) Now let c(#) be a vector-valued rational function of the perturbation #, such that We consider the following min-max problem: c(#) T x subject to x # X # . From now on, we assume that the function c(#) is independent of # (in other words, the objective vector c is not subject to perturbation). This is done with no loss of generality: introduce a slack variable # and define Problem (3) can be formulated as x subject to - # is the robust feasible set corresponding to the function - F. In the following, we thus consider a problem of the form subject to x # X # and refer to it as a robust semidefinite problem (RSDP). In general, although X # is convex, P # is not a tractable problem-in particular, it is not an SDP. Our aim is to find a convex, inner approximation of X # that is described by a linear matrix inequality constraint. This inner approximation is then used to find an upper bound on the optimal value of P # by solving an SDP. In some cases, we can prove our results are nonconservative, that is, as in the so-called "full perturbation" case. We refer to the set X 0 (resp., problem P 0 , i.e., (1)) as the nominal feasible set nominal SDP). We shall assume that the nominal SDP is feasible, that is, course, the robust feasible set X # may become empty for some # > 0; we return to this question later. 2.2. Linear-fractional representation. In this paper, we restrict our attention to functions F that are given by a "linear-fractional representation" (LFR): where F (x) is defined in (1), R(-) is an a#ne mapping taking values in R q-n , and L # R n-p and D # R q-p are given matrices. Together, the mappings F (-), R(-), the matrices L, D, the subspace D, and the scalar # constitute our perturbation model for the nominal SDP (1). The above class of models seems quite specialized. In fact, these models can be used in a wide variety of situations, for example, in the case where the (matrix) coe#cients F i in P 0 are rational functions of the perturbation. The representation lemma, given below, and the examples of section 5 illustrate this point. A constructive proof of the following result can be found in [37]. Lemma 2.1. For any rational matrix function M : R k singularities at the origin, there exist nonnegative integers r 1 , . , r k , and matrices M # R n-c , 36 L. EL GHAOUI, F. OUSTRY, AND H. LEBRET , such that M has the following linear-fractional representation (LFR): For all # where M is defined, Using the LFR lemma, we may devise LFR models for SDPs, where a perturbation vector enters rationally in the coe#cient matrices. The resulting set D of perturbation matrices # is then a set of diagonal matrices of repeated elements, as in (6). Componentwise bounds on the vector #, such as |#| i #, into a norm-bound # on the corresponding matrix #. 2.3. A special case. We distinguish a special case for which exact (nonconser- vative) results can be obtained via SDP. This is when F(x, #) is block diagonal, each block being independently perturbed-precisely, when where each F i (x, # i ) assumes the form shown in section 2.2 for appropriate L i , R i , D i , with consists of block-diagonal matrices of the form # . We refer to this situation as the block-full perturbation case. When of the full perturbation case. As will be seen later, all results given for can be generalized to the case L > 1. 3. Robust solutions for SDPs. Unless otherwise specified, we fix # > 0. 3.1. Full perturbations case. In this section, we consider the full perturbations case, that is, . We assume #D# -1 , which is a necessary and su#cient condition for F(x, #) to be well defined for every x #. The following lemma is a simple corollary of a classic result on quadratic inequal- ities, referred to as the S-procedure [12]. Its proof is detailed in [16]. Lemma 3.1. Let real matrices of appropriate size. We have det(I -D#= 0 and for every # 1, if and only if #D# < 1 and there exists a scalar # such that A direct application of the above lemma shows that, in the full perturbations case, the RSDP (4) is an SDP. Theorem 3.1. When the RSDP (4) and a corresponding solution x can be computed by solving the SDP in variables subject to # F (x) - #LL T R(x) T ROBUST SOLUTIONS TO UNCERTAIN SEMIDEFINITE PROGRAMS 37 Special barrier functions adapted to a conic formulation of the problem can be devised and yield an interior-point algorithm that has the same complexity as the nominal problem; see [24]. We may define the maximum allowable perturbation level, which is the largest number # max such that X # for every #, 0 # Computing # max is a (quasi-convex) generalized eigenvalue minimization problem [24, 11]: minimize # subject to # F (x) - #LL T R(x) T Remark. The above exact results are readily generalized to the block-full perturbation case (L > 1) as defined in section 2.2. 3.2. Structured case. We now turn to the general case (D is now an arbitrary linear subspace). In this section, we associate with D the following linear subspace: -R q-q -R q-p S#T , G# T G T for every # D # . As shown in [16], a general instance of problem (4) is NP-hard. Therefore, we look for upper bounds on its optimal value. The following lemma is a generalization of Lemma 3.1 that traces back to [17]. Its proof is detailed in [16]. Lemma 3.2. Let real matrices of appropriate size. Let D be a subspace of R p-q , and denote by B the set of matrices associated with D as in (11). We have det(I -D#= 0 and for every # D, # 1 if there exist a triple (S, T, and Using Lemma 3.2, we obtain the following result. Theorem 3.2. An upper bound on the RSDP (4) and a corresponding solution x can be computed by solving the SDP in variables x, S, T subject to (S, T, Note that when the perturbation is full, the variable G is zero and S, T are of the form #I p , #I q , resp., for some # 0. We then recover the exact results of section 3.1. As before, we may define the maximum allowable perturbation level, which is the largest number # max such that X # for every #, 0 # max . Computing a lower bound on this number is a (quasi-convex) generalized eigenvalue minimization problem: subject to (S, T, 38 L. EL GHAOUI, F. OUSTRY, AND H. LEBRET 4. Uniqueness and regularity of robust solutions. In this section, we derive uniqueness and regularity results for the RSDP in the case of full perturbations. As before, we first take . The results of this section remain valid in the general case L > 1 (several blocks). We fix #, 0 < # max . For simplicity of notation (and without loss of generality) we take 1). For well-posedness reasons, we must assume #D# < 1. We make the further assumption that #) is a#ne in #). In section 4.5, we show how the case D #= 0 can be treated. For full perturbations and the RSDP is the SDP subject to # F (x) - #LL T R(x) T 4.1. Hypotheses. We assume that the SDP (15) (with satisfies the following hypotheses: H1. The Slater condition holds, that is, the problem is strictly feasible. H2. The problem is inf-compact, meaning that any unbounded sequence feasible points (if any) produces an unbounded sequence of objectives. An equivalent condition is that the Slater condition holds for the dual problem [28, p. 317, Thm. 30.4]. H3. (a) The nullspace of the matrix x is independent of (#, x) #= (0, 0) and not equal to the whole space. (b) For every x, # is full column-rank. Hypotheses H1 and H2 ensure, in particular, the existence of optimal points for problem (15) and its dual. Hypotheses H3(a) and (b) are di#cult to check in general, but sometimes can be easily tested in practical examples, as seen in section 5. We note that H3(a) implies that R(x) #= 0 for every x. Hypothesis H1 is equivalent to Robinson's condition [27], which can be expressed in terms of # . Robinson's condition is stated in [27] as the existence of x 0 R such that where dF is the di#erential of F , and S n+q is the set of positive semidefinite matrices of q. The equivalence between H1 and Robinson's assumption is not true, in general. Here, this equivalence stems from the fact that the problem is convex and that the cone S n+q has a nonempty interior. Remark. Hypothesis H1 holds if and only if it holds for the nominal problem (1) (recall our assumption # max > 1). Also, hypothesis H2 implies L #= 0 (otherwise, we can let # without a#ecting the objective value). If H2 holds for the nominal problem and L #= 0, then H2 holds for the RSDP (15). 4.2. An equivalent nonlinear program. Let x opt , # opt be optimal for (15). Hypothesis H3(a) ensures that any # that is feasible for (15) is nonzero (otherwise, R(x) would be zero for some x). We thus have # opt > 0. We introduce some notation. For x Using Schur complements and # opt > 0, we obtain that problem (15) can be rewritten as minimize d T y subject to G(y) # 0 and that y Our aim is first to prove that the so-called quadratic growth condition [10] holds at y opt for problem (16). Then, we will apply the results of [10] to obtain uniqueness and regularity theorems. 4.3. Checking the quadratic growth condition. Following [10], we say that the quadratic growth condition (QGC) holds at y opt if there exists a scalar # > 0 such that, for every feasible y, Roughly speaking, this condition guarantees that y opt is not on a facet on the boundary of the feasible set. Define the set of dual variables associated with y opt by The following result is a direct consequence of a general result by Bonnans, Cominetti, and Shapiro [10]. Roughly speaking, this result states that, if an optimization problem satisfies Robinson's condition and has an optimal point, and if a certain "curvature" condition is satisfied, then the QGC holds at that point. Theorem 4.1. With the notation above, if H1 and H2 hold, and if then problem (16) satisfies the QGC. The following theorem is proven in appendix A. Theorem 4.2. If H1-H3 hold, problem (15) satisfies the quadratic growth condition at every optimal point y opt . Consequently, there exists a unique solution to the SDP (15). Remark. Note that the QGC is satisfied independent of the objective vector. This means that the boundary of the feasible set is strictly convex (it contains no facets). 4.4. Regularity results. In problem (15), the data consist of the matrices L, We seek to examine the sensitivity of the problem with respect to small variations in F i , L i , and R i . In this section, we consider matrices are functions of class C 1 of a (small) parameter vector u. Define x We denote by P(u) the corresponding problem (15), where F (-), R(-), and L are replaced by F (-, u), R(-, u), and L(u). We assume that F (-, and L, so that P(0) is (15). We first note that, in the vicinity of problem P(u) satisfies the hypotheses H1 and H2 if P(0) does. In this case, for every # > 0 we may define the set S # (u) of #-suboptimal points of P(u): where v(u) is the optimal value of P(u). Recall that, if P 0 satisfies hypotheses H1 and H2, the optimal value v(u) is con- tinuous, and even directionally di#erentiable, at Thm. 5.1]. With the QGC in force, and using [31, Thm. 4.1], we can give quite complete regularity results for the robust solutions. Theorem 4.3. If hypotheses H1-H3 hold for P(0), then for every there exists a # > 0 and a neighborhood V of such that for every u # V and When H1-H3 hold for P(0), the above theorem states that every (su#ciently) suboptimal solution to P(0) is H-older-stable (with coe#cient 1/2). This is true, in particular, for any optimal solution of P(u) (that is, for 0). The fact that the theorem remains true for # > 0 guarantees regularity of numerical solutions to the RSDP. The main consequence is that even if the nominal SDP is ill conditioned (with respect to variations in the F i 's), the RSDP becomes well conditioned for every # > 0. Now assume #= 1. We seek to examine the behavior of problem (10) (with when the uncertainty level # for 0 < # max varies. This is a special case of the problem examined above, with Corollary 4.1. For every #, 0 < # max , the solution to (10) (with is unique and satisfies the regularity results (written with #) of Theorem 4.3. Remark. The results of this section are all valid in the block-full perturbation case (L > 1), as defined in section 2.2. Of course, the conditions given in H3 should be understood blockwise. 4.5. Case D #= 0. When D #= 0, we can get back to the case Recall that we have #D# < 1 in order to ensure that F(x, #) is defined everywhere on D. With this assumption, we can define, for x Using Schur complements, we have, for every x and # > 0, Hypothesis H3 holds for - L, - R(-) if and only if it holds for L, R(-). We can then follow the steps detailed previously. Corollary 4.2. If the SDP (10) (with then the results of Theorem 4.3 hold. 5. Examples. 5.1. Unstructured perturbations. Assume This case corresponds to the representation in section 5, with x Using Lemma 3.2, we obtain that problem (4) is equivalent to the SDP subject to # F (x) - #I # 1 x T #I T#I #I It turns out that we may get rid of the variable # and get back to a convex problem of the same size as that of the unperturbed problem (1). To see this, first note that every feasible variable # in problem (20) is strictly positive. Use Schur complements to rewrite the matrix inequality in (20) as Minimizing (over variable #) the scalar in the left-hand side of the above inequality shows that the RSDP (1) is equivalent to subject to F (x) # Formulation (21) is more advantageous than (20), since (21) involves a (convex) matrix inequality constraint of the same size as the original problem. As noted before, special barrier functions can be devised for this problem and yield an interior-point algorithm that has the same complexity as the original problem; see [24]. We note that, with the above choice for L, R, hypothesis H3 holds, which yields the following result. Theorem 5.1. The optimal value of the RSDP (20) can be computed by solving the convex problem (21). If (21) satisfies hypotheses H1 and H2, then for every # > 0, the solution is unique and satisfies the regularity conditions of Theorem 4.3. Remark. A su#cient condition for hypotheses H1 and H2 to hold for (21) is that they hold for the nominal problem. A more restrictive su#cient condition is that the nominal feasible set X 0 is nonempty and bounded, and # max . 42 L. EL GHAOUI, F. OUSTRY, AND H. LEBRET .3 x Fig. 1. Nominal and robust solutions of an SDP, with a 5 - 5 matrix F (x). Here 5.2. Robust center of a linear matrix inequality. In this section, for # > 0, we consider the SDP (21) and corresponding feasible (convex) set X # . We assume that X 0 is nonempty and bounded, and that P 0 is strictly feasible. Then, for every #, is nonempty and bounded, and we can define a (unique) solution x(#) to the strictly convex problem (21). In view of Corollary 4.1, x(#) is a continuous function of # in ]0 # max [. Since (X # ) is a decreasing family of bounded sets, we may define x(#). Note that x # is independent on the objective vector c. Thus, to the matrix inequality F (x) # 0, we may associate the robust center, defined by (22). The robust center has the property of being the most tolerant (with respect to unstructured perturbation) among the feasible points. An example is depicted in Fig. 1. The nominal feasible set X 0 is described by a linear matrix inequality F (x) # 0, where F is a 5 - 5 matrix. For various values of #, we seek to minimize x 2 . The dashed lines correspond to the optimal objectives. As # increases, we observe that the robust feasible sets shrink. A crucial property of these robust sets is that they do not possess any straight faces, as observed in the figure. For the robust feasible set is a singleton (in this example, x the optimal solution is not unique and not continuous with respect to changes in the coe#cient matrices F i , (although the optimal value is continuous). Since the sets X # become strictly convex as soon as # > 0, the resulting robust solutions are continuous. 5.3. Robust linear programs. An interesting special case arises with linear programming (LP). Consider the LP subject to a T Assume that the a i 's and b i 's are subject to unstructured perturbations. The perturbed value of [a T We seek a ROBUST SOLUTIONS TO UNCERTAIN SEMIDEFINITE PROGRAMS 43 robust solution to our problem, which is a special case of the block-full perturbation case referred to in section 2.2, with F given by (7), and and D is the set of diagonal, L - L matrices. The robust LP is subject to a T The above program is readily written as an SDP by introducing slack variables. In fact, the robust LP is a second-order cone program (SOCP) for which e#cient special-purpose interior-point methods are available [24, 20, 23]. We note that hypothesis H3 holds blockwise. This yields the following result. Theorem 5.2. The optimal value of the robust LP can be computed by solving the convex problem (23). If the latter satisfies hypotheses H1 and H2, then for every #, 0 < # max , the solution is unique and satisfies the regularity conditions of Theorem 4.3. In [6], robust linear programming is studied in detail. For a wide class of perturbation models, where the data of every linear constraint vary in an ellipsoid, explicit robust solutions are constructed using convex SOCPs. Reference [23] also provides examples of robust linear programs solved via SOCP. 5.4. Robust eigenvalue minimization. Consider the case where the nominal problem consists of minimizing the largest eigenvalue of a matrix-valued function When F (-) is subject to unstructured perturbations (as defined in section 5.1), the robust version of the problem is subject to #I # F (x), or equivalently When written in an SDP form, the above problem satisfies the hypotheses H1-H3. From Theorem 4.3 we obtain that the solution is unique. If we consider that the data of the above problem consist of the matrices F i , then we know that the corresponding solution is H-older-stable (with coe#cient 1/2). Since the problem is unconstrained, we can use a result of Shapiro [31, Thm. 3.1], by which we conclude that the solution is actually Lipschitz stable (inequality (18) holds with the exponent 1/2 replaced by 1). Finally, using the results from Attouch [3], we can show that computing the solution for # 0 amounts to selecting the minimum norm solution among the solutions of the nominal problem. Theorem 5.3. The optimal value of the min-max problem (24) can be computed by solving the convex problem (25). For every # > 0, the solution is unique and is Lipschitz stable with respect to perturbations in F i , the solution converges to the minimum norm solution of the nominal problem (24). Remark. In this case, the RSDP is a regularized version of the nominal SDP, which belongs to the class of Tikhonov regularizations [34]. The regularization parameter 44 L. EL GHAOUI, F. OUSTRY, AND H. LEBRET is 2# and is chosen according to some a priori information on uncertainty associated with the nominal problem's data. Taking # close to zero can be used as a selection procedure for choosing a particular (minimum norm, regular) solution among the (not necessarily unique and/or regular) solutions of the nominal problem. Problem (25) is particularly suitable to the recent so-called U-Newton algorithms for solving problem (24). These methods, described in [21, 25], require that the Hessian of the "smooth part" (the so-called U-Hessian) of the objective of (24) be positive definite. For general mappings F (-), this property is not guaranteed. However, when looking at the robust problem (25), we see that the modified U-Hessian is guaranteed to be positive definite for every x and # > 0. This indicates that the RSDP approach may be used to devise robust algorithms for solving SDPs. 5.5. Robust SOCPs. An SOCP is a problem of the form subject to #C can be formulated as SDPs, but special-purpose, more e#cient algorithms can be devised for them; see [24, 5, 23]. Assuming that C i , d i , e i , f i are subject to linear-or even rational-uncertainty, we may formulate the corresponding RSDP as an SDP. This SDP can be written as an SOCP if the uncertainty is unstructured and a#ects each constraint independently. The subject of robust SOCPs is explored in [5] in detail. Explicit SDPs that yield robust counterparts to SOCPs nonconservatively are given for a wide class of uncertainty structures. In some cases, albeit not all, the robust counterpart is itself an SOCP. In [16, 14], the special case of least-squares problems with uncertainty in the data is studied at length. 5.6. Robust maximum norm minimization. Several engineering problems take the form minimize #H(x)#, where are given p-q matrices. A well-known instance of this problem is the linear least-squares problem, with Another example is a minimal norm extension problem for a Hankel operator studied in [18], in which H 0 is a given Hankel matrix and H i , is the n - n Hankel matrix associated with the polynomial 1/z i . In practice, the matrices H i , are subject to perturbation, which motivates a study of the robust version of problem (27). Note that the least-squares case is extensively studied in [16]. Consider the full perturbation case, which occurs when each H i is perturbed independently in a linear manner. Precisely, consider the matrix-valued function For a given # > 0, we address the min-max problem min x #H(x, #. This problem is an RSDP for which we can get exact results using SDP. Indeed, for every x # R m and # 0, the property is equivalent to F(x, # 0 for every #, where where F (x, #I H(x) x I # . We thus write problem (28) as (4), where the perturbation set D is R p-q . Applying Theorem 3.2, we obtain that the RSDP above is equivalent to the As in section 5.1, we may get rid of the variable # and obtain the equivalent formulation This RSDP satisfies hypotheses H1-H3, so we conclude that the results of Theorem 4.3 hold. As in robust eigenvalue minimization, we can get improved results using [31, section 3, Thm. 3.1]. Theorem 5.4. The optimal value of the min-max problem (28) can be computed by solving the convex problem (29). For every # > 0, the solution is unique and Lipschitz stable with respect to perturbations in H i , the solution converges to the minimum norm solution of the nominal problem (27). Remark. As for the RSDP arising in robust eigenvalue minimization, the robust minimum norm minimization problem is a regularized version of the nominal problem, which belongs to the class of Tikhonov regularizations. We now consider the general case where each matrix H i in (27) is perturbed in a structured manner. To be specific, we concentrate on the minimal norm extension problem mentioned above. In practice, the matrix H 0 is obtained from measurement and is thus subject to error. We may assume that this matrix is constructed from an n - 1 vector h 0 is unknown but bounded. The perturbed matrix H 0 is of the form where L, R are given matrices (the exact form of which we do not detail), and (In the above, each # i corresponds to the uncertainty associated with the ith antidiagonal of H 0 .) We address the min-max problem L. EL GHAOUI, F. OUSTRY, AND H. LEBRET This problem is amenable to the robustness analysis technique. Defining we obtain the following result. Theorem 5.5. An upper bound on the objective value of the min-max problem (30) can be computed by solving the SDP in variables x, S, G: subject to 5.7. Polynomial interpolation. This example, taken from [16], can be formulated as an RSDP with rational dependence. For given integers n # 1, k, we seek a polynomial of degree that interpolates given points (a If we assume that (a i , b i ) are known exactly, we obtain a linear equation in the unknown x, with a Vandermonde structure: 1 a 1 . a n-11 a k . a n-1 which can be solved via standard least-squares techniques. Now assume that the interpolation points are not known exactly. For instance, we may assume that the b i 's are known, while the a i 's are parameter dependent: a where the # i 's are unknown but bounded: |# i | #, We seek a robust interpolant, that is, a solution x that minimizes where 1 a k (#) . a k (#) n-1 # . The above problem is an RSDP. Indeed, it can be shown that where . a n-2 and, for each i, . a n-2 . a i . a n-3 . a i (Note that det(I - D#= 0, since D is strictly upper triangular.) With the above notation, if we define F(x, #) as in section 5, then problem (31) can be formulated as the RSDP (4). With the approach described in this paper, one can compute an upper bound for the minimizing value of (31), and a corresponding suboptimal x. We do not know if the problem can be solved exactly in polynomial time, e.g., using SDP. We conjecture (as the reviewers of this paper did) that the answer is no. To motivate this claim, note that the solution to the problem of computing (31) for arbitrary a#ne functions A is already NP-hard [16]. 5.8. Error-in-variables RSDPs. In many SDPs that arise in engineering, the variable x represents physical parameters that can be implemented with finite absolute precision only. A typical example is integer programming, where integer solutions to (linear) programs are sought. These problems (which are equivalent to integer programming) are NP-hard. We now show that we may find upper bounds on these problems using robustness analysis. Consider, for instance, the problem of finding a solution x to the feasibility SDP find an integer vector x such that F (x) # 0. Now, consider the robust SDP maximize # subject to Assume there exists a feasible pair (x feas , #) to the above problem, with # 0. By construction, x feas satisfies F Furthermore, any vector x chosen such that is guaranteed to satisfy F (x) # 0. This is true, in particular, for x int , the integer closest to x feas . Thus, if we know a positive lower bound #, and corresponding feasible point for problem (33), then we can compute an integer solution to our original problem. Finding a lower bound for (33) and an associated feasible point can be done as follows. For Let Rm # , and 48 L. EL GHAOUI, F. OUSTRY, AND H. LEBRET Problem (33) can be formulated as maximize # subject to #I # F for every # D, # 1/2. The above is a special instance of the structured problem examined in section 3.2. Theorem 5.6. A su#cient condition for an integer solution to the feasibility SDP (32) is that the constraints are feasible. If x feas is feasible for the above constraints, then any integer vector closest to x feas (in the maximum norm sense) is feasible for (32). 6. Conclusions. In this paper, we considered semidefinite programs subject to uncertainty. Assuming the latter is unknown but bounded, we have provided su#cient conditions that guarantee "robust" solutions to exist via SDPs. Under some conditions (detailed in section 4), the robust solution is unique, and not surprisingly, stable. The method can then be used to regularize possibly ill-conditioned problems. For some perturbation structures (as for unstructured perturbations), the conditions are also necessary. That is, there is no conservatism induced by the method. The paper raises several open questions. In our description, we have considered the problem of making the primal SDP robust, thereby obtaining upper bounds on an SDP subject to uncertainty. The dual point of view should be very interesting. One might be interested in applying the approach to the dual problem instead. Does this lead to lower bounds on the perturbed problem? Also, in some cases, the RSDP approach leads to a unique (and stable) primal solution. May we obtain a unique solution to the dual problem by making the latter robust? (This would lead to analyticity of the primal solution; see [32].) As seen in section 5.2 the notion of robust center has, certainly, connections with the well-known analytic center; is the latter related to some robustness characterization It seems that the RSDP method could be useful for deriving fast and robust (stable) algorithms for solving SDPs (see section 5.4), especially in connection with maximum eigenvalue minimization. Finally, as said in section 2.2 (Lemma 2.1), an SDP with coe#cient matrices depending rationally on a perturbation vector can always be represented by an LFR model. Now, this LFR model is not unique. However, the results given here (for example, Theorem 3.2) hinge on a particular linear-fractional representation for a perturbed SDP. Hence we have the question: are our results independent of the chosen representation? We partially answer this di#cult question in Appendix B. Appendix A. Proof of Theorem 4.2. We take the notation of section 4. diag(Z, -) be dual variables associated with are optimal (their existence is guaranteed by H1 and H2). Then, Y # Y(y opt ). Let us show that condition (17) holds for this choice of Y . Since the problem satisfies H1 and H2, the complementarity conditions hold; therefore, the (optimal) dual variable - associated with the constraint # opt is zero. Consequently the variable Z is nonzero (recall c #= 0). Using TrY we obtain From # opt #= 0 (implied by H3(a)), and using hypothesis H3(b), we can show that are impossible for Z # 0, Z #= 0. This yields TrR(x opt ) T R(x opt )Z > 0. Now let # R m and # R, and define We have, for every i, (R(x) (R T By summation, we have (R(#) opt opt R T R+# opt R(x opt ). We obtain finally, opt 0 with means that every column of Z 1/2 belongs to the nullspace of R(#) -R(0). Now assume #= 0. By hypothesis H3(a), we obtain that every column of Z 1/2 also belongs to the nullspace of R(x opt ), which contradicts 50 L. EL GHAOUI, F. OUSTRY, AND H. LEBRET We conclude that # 2 yy L is positive definite at (y opt , Y ). Thus, problem (15) satisfies the QCG. Appendix B. Invariance with respect to the LFR model. In this section, we show that the su#cient conditions obtained in this paper are, in some sense, independent of the LFR model used to describe the perturbation structure. Consider a function F taking values in the set of symmetric matrices having an LFR such as that in section 5. This function can be written in a more symmetric #) where we have dropped the dependence on x for convenience, and # . It is easy to check that, if an invertible matrix Z satisfies the relation Z - # for every # D, then #) LZ) T . In other words, the "scaled" triple DZ)) can be used to represent F instead of F, - L, - D in (35). By spanning valid scaling matrices Z, we span a subset of all LFR models that describe F. A valid scaling matrix Z can be constructed as follows. Let (S, T, G) # B, and define I # . It turns out that such a Z satisfies the relation Z - # for every # D. Using the above facts, we can show that if condition (13) is true for the original LFR model F, L, R, D with appropriate S, T, G, then it is also true for the scaled LFR obtained using any scaling matrix Z such as that above, for appropriate matrices - T . That is, the condition is independent of the scaling Z. In this sense, the conditions we obtained are independent of the LFR used to represent the perturbation structure. Acknowledgments . This paper has benefitted from many stimulating discussions with several colleagues, including Aharon Ben-Tal, Stephen Boyd, Arkadii Ne- mirovski, Michael Overton, and particularly, Lieven Vandenberghe (who pointed out a mistake just before the final version was sent). Last but not least, the authors would like to thank the editor and reviewers for their very helpful comments and revisions. --R Interior point methods in semidefinite programming with applications to combinatorial optimization Viscosity Solutions of Optimization Problems in Semidefinite Programming and Applications Robust Solutions to Uncertain Linear Programs Robust truss topology design via semidefinite programming Anal., <Year>1998</Year>, to appear. Some Numerical Methods in Sensitivity Analysis of Optimization Problems under Second Order Regular Constraints Method of centers for minimizing generalized eigenvalues Linear Matrix Inequalities in System and Control Theory Introduction to convex optimization with engineering applications New York Robust solutions to least-squares problems with uncertain data Robustness in the presence of mixed parametric uncertainty and unmodeled dynamics Minimal norm extensions and eigenstructures Global and Local Convergence of Predictor-Corrector- Interior-Point Algorithm for Semidefinite Programming Interior Point Polynomial Methods in Convex Program- ming: Theory and Applications On homogeneous interior-point algorithms for semidefinite pro- gramming Stability theorems for systems of inequalities Convex Analysis Discrete Event Systems Perturbation theory of nonlinear programs when the set of optimal solutions is not a singleton Perturbation analysis of optimization problems in Banach spaces First and second order analysis of nonlinear semidefinite programs. Solutions of Ill-Posed Problems The solution of certain matrix inequalities in automatic control theory Robust and Optimal Control --TR --CTR Frank Lutgens , Jos Sturm , Antoon Kolen, Robust One-Period Option Hedging, Operations Research, v.54 n.6, p.1051-1062, November 2006 Budi Santosa , Theodore B. Trafalis, Robust multiclass kernel-based classifiers, Computational Optimization and Applications, v.38 n.2, p.261-279, November 2007 B. P. Lampe , E. N. Rosenwasser, Polynomial solution to stabilization problem for multivariable sampled-data systems with time delay, Automation and Remote Control, v.67 n.1, p.105-114, January 2006 D. Goldfarb , G. Iyengar, Robust portfolio selection problems, Mathematics of Operations Research, v.28 n.1, p.1-38, February Mung Chiang, Geometric programming for communication systems, Communications and Information Theory, v.2 n.1/2, p.1-154, July 2005
uncertainty;robustness;regularization;semidefinite programming;convex optimization
589118
Towards a Practical Volumetric Cutting Plane Method for Convex Programming.
We consider the volumetric cutting plane method for finding a point in a convex set ${\cal C}\subset\Re^n$ that is characterized by a separation oracle. We prove polynomiality of the algorithm with each added cut placed directly through the current point and show that this "central cut" version of the method can be implemented using no more than 25n constraints at any time.
Introduction Let C ae ! n be a convex set. Given a point - separation oracle for C either reports that - returns a separating hyperplane a 2 ! n such that a T x ? a T - x for every x 2 C. The convex feasibility problem is to use such an oracle to find a point in C, or prove that the volume of C must be less than that of an n-dimensional sphere of radius 2 \GammaL , for given It is well known [9] that a variety of convex optimization problems can be cast as instances of the convex feasibility problem, and moreover the problem plays a fundamental role the complexity analysis of many combinatorial optimization problems. For many years the standard approach to the convex feasibility problem has been the ellipsoid algorithm; see for example [5] or [9]. In [14], Vaidya proposed an alternative algorithm for the convex feasibilty problem based on a new barrier for a polyhedral set, the volumetric barrier. On each iteration algorithm has a point x a polyhedral set P where A k is a m k \Theta n matrix. For each k the set P k is bounded, is an approximation of the volumetric center of P k , the minimizer of the volumetric barrier (see Section 2). The algorithm then either deletes one constraint that defines P k , or calls the separation oracle to see if x k 2 C. If not, the oracle returns a separating hyperplane which is used to add a constraint to P k . After the addition or deletion of a constraint, the algorithm takes a number of Newton, or Newton-like, steps for the volumetric barrier to obtain a new point x k+1 which is an approximation of the volumetric center of the new polyhedron P k+1 . Let T represent the cost, in numerical operations, of a call to the separation oracle. fundamental result is that the complexity of his volumetric cutting plane algorithm for the convex feasibility problem is O(nLT +n 4 L) operations, compared to O(n 2 LT +n 4 L) operations for the ellipsoid algorithm. (In theory, the complexity of Vaidya's method can be further reduced through the use of "fast matrix multiplication," which cannot be applied to the ellipsoid algorithm.) Although Vaidya's result is theoretically significant, the algorithm of [14] does not appear to very practical. In particular, the analysis of [14] requires that the polyhedral sets P k have up to 10 7 n constraints, and the algorithm might require thousands of Newton-like steps following the addition or deletion of a constraint. In [3], Anstreicher describes a strengthened version of Vaidya's volumetric cutting plane algorithm for the convex feasibility problem. The algorithm of [3] reduces the maximum number of constraints to 200n, while requiring no more than 5 Newton steps following a constraint addition or deletion. Although these figures represent a substantial improvement over [14], the algorithm of [3] is still not fully practical. In particular: i. For reasonable n, 200n constraints is still quite large, given that least-squares systems with this number of rows must be repeatedly solved on each iteration. ii. The algorithm of [3] uses true Newton steps, which in practice are expensive to compute compared to the Newton-like steps used in [14]. iii. As in [14], the algorithm of [3] cannot place a new constraint directly through the current point, but must rather "back off" each separating hyperplane to generate a shallow cut. Ramaswamy and Mitchell [13] describe a "central cut" version of the volumetric cutting plane algorithm that allows for the placement of each new constraint through the current point, and uses Newton-like steps following constraint additions and deletions. (The algorithm of [13] actually solves the problem of minimizing a linear function over a convex set C using a separation oracle, but most of the analysis is very similar to that required to solve the convex feasibility problem.) Unfortunately [13], which uses many results from [14], requires that the algorithm maintain up to 10 8 n constraints. The purpose of this paper is to develop a central cut volumetric cutting plane algorithm that also improves on the 200n constraints required by the algorithm of [3]. As in [13], the algorithm uses an "affine" step to move off of a cut placed through the current point. The use of such a step in the context of a cutting plane method based on analytic centers is well known [7]. (See [4], [8], [10], [11], [12], and [15] for other results on analytic center cutting plane methods.) In fact the affine step we use is based on that of [7], rather than the step used in [13]. Our analysis uses a number of results from [2] and [3], and an improved second-order expansion of the volumetric barrier, to improve upon the analysis of [13]. As in [13], the method described here requires O( n) Newton-like steps following the addition or deletion of a constraint, compared to O(1) Newton, or Newton-like, steps in [14] and [3]. Although this is certainly a disadvantage from the standpoint of theoretical complexity, the fact that the O( n) bound arises from a worst-case analysis of descent in the volumetric barrier suggests that in practice far fewer steps would likely be required. Our final result is a central cut volumetric cutting plane method that requires no more than 25n constraints at any time. In Table 1 we summarize important attributes of four papers (including this paper) on volumetric cutting plane methods. These features are the placement of added cuts (shallow or central), the number of Newton or Newton-like steps required after a constraint addition or deletion, the maximum number of constraints required, and the value of a scalar \DeltaV , defined as the difference between the mimimal increase in the volumetric barrier following a constraint addition, and the maximal decrease following a constraint deletion (see Section 3). For all four algorithms the number of oracle calls is O(nL), with a constant that is inversely proportional to \DeltaV (see for example the proof of [3, Theorem 3.2]). Reference Placement Steps after Number of \DeltaV of cut addition/deletion constraints Anstreicher [3] Shallow O(1) 200n 3:7 \Theta 10 \Gamma4 Ramaswamy and Mitchell [13] Central O( This Paper Central O( n) 25n 1:4 \Theta 10 \Gamma3 Table 1: Attributes of volumetric cutting plane methods 2 The Volumetric Barrier In this section we collect a number of properties of the volumetric barrier V (\Delta) which will be used in the subsequent analysis. To start, let A is an m \Theta n . Whenever we refer to P, we are implicitly refering to the constraint system [A; b] which defines it. The volumetric barrier for P is the function x be a point having denote the vector equal to the diagonal of the projection In other words oe m. It is easy to show (see for example the Appendix of [1]) that 0 - oe - e, e T is the vector with each component equal to one. The gradient and Hessian of V (\Delta) at x are then given by denotes the Schur, or Hadamard product of P with itself: ij . Let good approximation of H(x), in that where - denotes the ordering for positive semi-definite matrices defined by A - B () A is positive semidefinite. See for example the Appendix of [1] for a derivation of (1) and a proof of (2); these and other properties of V (\Delta) are originally due to Vaidya [14]. In the sequel we will often be interested in the behavior of V (\Delta) for a step of the form 1. For such an - Q(-x). The proof of the following is very straightforward; see for example [14, Lemma 5] or [1, Lemma 2.2]. Proposition 2.1 Let - It follows immediately from Proposition 2.1 that if - Using a Taylor series expansion, (2), and (3), it is easy to show that The bounds in (4) have been used in [1], [13], and [14]. The following theorem provides a strengthening of (4) that will be used throughout the paper. Theorem 2.2 Suppose that - Proof: We have dff To prove the lemma we will obtain lower and upper bounds on the final term in (5). We begin with the lower bound. Using (2) and (3), we have and therefore But it is straightforward to compute that d and therefore Substituting (8) into (7), we obtain Z 1ff An integration by parts shows that Z 1ff and substituting (10) into (9) produces the lower bound of the lemma. The proof of the upper bound is similar. Again using (2) and (3), we have and therefore But d and therefore Substituting (12) into (11), we obtain Z 1ff Another integration by parts shows that Z 1ff and substituting (14) into (13) produces the upper bound of the lemma. 2 Since the bounds in Lemma 2.2 involve both k-k to consider how these two quantities are related. For x with g. Note that oe min ? 0 under the trivial assumption that A contains no zero row. Define In the following lemma we give two bounds for kS \Gamma1 A-k1 in terms of k-k Q ; one involving -, and therefore oe min , and the other independent of oe. Theorem 2.3 Let x have 2. kS m)=2] 1=2 k-k Q . Proof: See [1, Theorem 3.3] for the proof of 1, and [3, Lemma 2.3] for the proof of 2. 2 Motivated by Theorem 2.3, we define It then follows from Theorem 2.3 that The fundamental proximity criterion that we employ throughout the paper is - this quantity is "large" (that is,\Omega\Gamma,/23 we will take a damped Newton-like step in an effort to reduce V (\Delta), and thus move closer to !, the volumetric center of P. When - we will be close enough to ! to adequately control the effect of adding or deleting a constraint, as required. The following theorem and corollary obtain a simple condition on - suffices to demonstrate the boundedness of P. Theorem 2.4 Let where the columns of A are independent. Let x have g. Suppose that kS Proof: P is bounded if and only if 6 9x 6= 0, Ax - 0. Since the columns of A are independent, ve T where the third equivalence uses a standard "Theorem of the Alternative" for systems of linear inequalities. But is exactly A T S \Gamma2 can be written A It follows that if kS is bounded. 2 Corollary 2.5 Let where the columns of A are independent. Let x have suppose that - Proof: This follows from Theorem 2.4, (17), and the fact that show that if - sufficiently small, then we can bound the possible remaining decrease in V (\Delta). The proximity allowed in Theorem 2.6, - weaker than in previous, similar results in [14], [2], and [3]. Theorem 2.6 Let x have where ! is the minimizer of V (\Delta), and Proof: Assume that - 1. Then there is a -, has - from (17), we have d dff where the inequality uses (6), and the final equality uses (8), both with 1. Using the fact that jg T we then obtain d dff where the last inequality uses and the assumption that - V (\Delta) is strictly convex, it follows that V (!) ? V (-x), which is a contradiction. Therefore 1, from (17), so Theorem 2.2 implies that where the second inequality uses jg T -j - kgk Q and the final inequality uses - 1, and the assumption that - In the corollary below we use Theorem 2.6 to establish bounds on V for two values of the parameter fl which are useful in the sequel. Corollary 2.7 Let x have Proof: For is straightforward to show that the minimum in (20), for 0 - ff - 1, occurs at 2. Substituting this value of ff into (20), and simplifying, then implies that It is also easy to show that for fl - 4=27, the right-hand side in (20) is monotonically decreasing for 0 - ff - 1. Substituting final topic that we consider in this section is that of reducing V (\Delta) when - \Omega\Gamma/54 To accomplish this we use a Newton-like step of the form for some ff ? 0. When - =\Omega\Gamma338 it can be shown that ff may be chosen in (21) so that an m) reduction is obtained in V (\Delta). In the following lemma we give a result for a particular value of fl used in the sequel. Lemma 2.8 Suppose that x has x(ff) be as in (21). m. Proof: Let construction we have - by (17). Applying Theorem 2.2, we obtain Substituting using - . Finally from (16), - m)=2 - m. 2 3 The Algorithm and its Complexity In this section we describe the central-cut volumetric cutting plane method, and establish its complexity using results from the two following sections. At the start of each iteration k - 0, we have an interior point x k of a bounded polyhedron P k oe C, where and A k is an m k \Theta n matrix with independent columns. We assume that C is contained in the hypercube kxk1 - 1, and set P is straightforward to show that x 0 is the volumetric center of P 0 .) The algorithm to be analyzed is as follows: Central-Cut Volumetric Cutting Plane Algorithm 1. Go to Step 1. Step 1. If V k Else go to Step 2. Step 2. If oe k min - ffl, go to Step 3. Else go to Step 4. Step 3. (Constraint Addition) Call the oracle to see if x k 2 C. If so, STOP. Otherwise the oracle returns a vector a 2 ! n such that a T x ? a T x k for all x 2 C. Let be an augmented constraint system having a k+1 ff ? 0. Go to Step 5. Step 4. (Constraint Deletion) Suppose that oe k be the reduced constraint system obtained by removing the j'th row of Go to Step 5. Step 5. (Centering Steps) Take a sequence of damped Newton-like steps of the form - -x J ). Let x to Step 1. In Step 1 of the algorithm, the value of V k max is such that V proves that the volume of P k , and therefore also C ae P k , is less than that of an n-dimensional sphere of radius 2 \GammaL . An explicit value for is given in Lemma 3.1, below. A suitable steplength ff in Step 3 is given in Theorem 4.6. Note that by construction each P k is bounded, by Corollary 2.5, since - each k. In addition, the fact that a constraint is only added if oe k For the Newton-like steps in Step 5, we assume that the steplengths ff are chosen so that each step produces an n) decrease in V k (\Delta). That this is always possible follows from \Omega\Gamma17 (see for example Lemma 2.8), and the fact that m k - O(n). Lemma 3.1 Consider the volumetric cutting plane algorithm with fl - :03. Assume that termination in Step 1 proves that the volume of C is less than that of an n-dimensional sphere of radius 2 \GammaL . Proof: See [3, Lemma 3.1]. 2 Next we consider the issue of how many iterations might be required for the algorithm to terminate. Assume that each time a constraint is added the algorithm achieves while each time a constraint is deleted it is assured that . The following theorem provides a complexity result for the algorithm under simple assumptions regarding \DeltaV and the number of Newton-like steps taken in Step 5. Theorem 3.2 Assume that the iterates of the volumetric cutting plane algorithm, using fl - :03, satisfy (23) and (24) on iterations where a constraint is added or deleted, respectively. Assume further that \DeltaV + is O(1), is\Omega\Gamma/4 , and the number of Newton-like steps in Step 5 of the algorithm is always O( n). Then for L =\Omega\Gamma47 (n)) the algorithm terminates in O(nL) iterations, using a total of O(nLT +n 4:5 L) operations, where T is the cost of a call to the separation oracle. Proof: See the proof of [3, Theorem 3.2]. 2 Compared to the algorithms of [14] and [3], Theorem 3.2 demonstrates that the central cut method of this paper has the same order number of oracle calls, O(nL), but performs more non-oracle work; O(n 4:5 L) versus O(n 4 L) operations. The reason for this is the larger number of centering steps, O( n) versus O(1), required after a constraint addition or deletion. Using results from the next two sections, we now show that the assumptions of Theorem 3.2 hold for certain choices of the parameters ffl and fl. Theorem 3.3 Let :01. Then the central-cut volumetric cutting plane method satisfies the assumptions of Theorem 3.2, with Proof: First consider an iteration where a cut is added. In Theorem 4.6, it is shown that for a particular choice of ff in Step 3 it is assured that where we are using the fact that kg k k (Q In addition, in Theorem 4.5 it is shown that for kg k k (Q is the volumetric center of P k+1 . Combining (25) and (26) we obtain Next, in Lemma 2.8 it is shown that if - in Step 5, then a steplength ff may be chosen so that together imply that after n) steps we must obtain - x J having - Finally, V k+1 implies that Next consider an iteration where a constraint is deleted. In Lemma 5.2 it is shown that dropping constraint i to obtain a new polyhedron P k+1 results in V k+1 Arguing exactly as above, it follows that after n) steps in Step 5, we must obtain - x J with - In addition, in Lemma 5.2 it is shown that V k+1 (! k+1 must have The assumptions of Theorem 3.2 thus hold with The value demonstrated in Theorem 3.3 may seem relatively small, but it should be noted that this is the largest value of \DeltaV to date for a volumetric cutting plane algorithm; see Table 1 in Section 1. 4 Adding a Central Cut Let x be an interior point of P. In this section we consider augmenting the constraint system defining P by imposing a central cut through x, to obtain a new polyhedron ~ b; a T ~ x - a T xg. Let ~ V (\Delta) be the volumetric barrier for ~ ! the volumetric center. Note that for any - x with - x ? a T x, we have ~ (a T - I (a T - (a T - We will first use (29) to establish a lower bound on ~ a cut is added through x. We will obtain two versions of this result. The first, using produces a relatively simple bound for the fundamental quantity ~ (!). Although this bound may be of some independent interest, in practice it cannot be used since Therefore we will also obtain a lower bound using x in a certain neighborhood of !. We begin with a series of lemmas. Throughout we let Lemma 4.1 Assume that - x ? a T x, and k - a T (a T - Proof: Consider the problem a T - Letting - can be written as a T - the solution of which is a T with solution value equal to ae a T It follows that a T (a T - - a T ae 2 a T :Lemma 4.2 Assume that - and -r - 1. Then a T (a T - immediately implies that oe min and also, from Lemma 2.3, that From (32) and Proposition 2.1 it follows that Then (31) and (33) together imply that oe min and the lemma follows from Lemma 4.1. 2 Lemma 4.3 Assume that - Proof: This follows immediately from (32), and the lower bound of Lemma 2.2. 2 Now let the volumetric center of P. We will use Lemmas 4.2 and 4.3 to establish a lower bound on ~ P is obtained by placing a central cut through !. Theorem 4.4 Suppose that ! is the volumetric center of P, ~ !g. Let ~ V (\Delta) be the volumetric barrier for ~ ! the volumetric center. Then ~ assume for the moment that -(!). Using (29), Lemmas 4.2 and 4.3, and the fact that ~ Next we use the fact that ln(1 to obtain where the second inequality uses - 1, and and the final inequality uses Substituting (35) into (34) then gives ~ A straightforward differentiation shows that the minimum of the right-hand side of (36), for occurs at From (36) we then have Next assume that r ? 1=-. Then there is an ff 2 (0; 1) so that - 1=-. From the convexity of V (\Delta), and Lemma 4.3, we obtain and (29) certainly implies that ~ It follows that ~ oe min )=8. 2 It is worthwhile to mention that the analysis in [13, Section 4.1] actually shows that ~ oe min ), although the authors of [13] do not note this fact. In practice the added cut a T ~ x - a T x cannot be passed through as in Theorem 4.4, but rather through a point x which is close to ! in some sense. As a result the lower bound of Theorem 4.4 must be modified to account for the use of x 6= !. In the next theorem we give a result based on particular parameter choices used throughout the paper. Theorem 4.5 Let x have ~ V (\Delta) be the volumetric barrier for ~ ! the volumetric center. Then ~ assume for the moment that Proceeding as in the proof of Theorem 4.4, but including the effect of ~ where the second inequality uses the fact that jg T We distinguish two cases. Case 1: oe min - :04725. Note that 1=- monotonically increasing in oe min , so oe min - :04 implies that 1=- 2 - 2 In addition, - 2 oe oe min =(2 monotonically increasing in oe min , so oe min - :04 also implies that Finally oe min - :04725 implies that - (2 1:606. Using these facts in (37), and the assumption that kgk Q ~ It can be verified numerically that the minimum of the right-hand side in (38), for occurs at approximately with value greater than :0340. (See Figure 1, Case 1 for a plot of the right-hand side of (38), for Case 1 Case 2 Figure 1: Lower bound on ~ Case 2: oe min - :04725. In this case we have 1=- Using these facts in (37), with - 1 and the assumption that kgk Q ~ It can be verified numerically that the minimum of the right-hand side in (39), for occurs at approximately with value greater than :0340. (See Figure 1, Case 2 for a plot of the right-hand side of (38), for This completes the proof under the assumption that r - 1=-. However, arguing as at the end of Theorem 4.4, it is easy to show that if k~! \Gamma xkQ ? 1=-, then ~ :For the final topic of the section, we consider moving off of the cut a T ~ x - a T x to a new point - x having a T - x ? a T x. Our goal is to obtain an upper bound for the quantity ~ Consider a point of the form a T 1. Note that - x in (40) is based on A T S \Gamma2 A, the Hessian of the logarithmic barrier at x, and not Q as used in [13, Section 4.1.2]. Theorem 4.6 Suppose that x has x - a T xg, and let ~ V (\Delta) be the volumetric barrier for ~ P. Then using x having ~ Proof: By construction we have a T - x \Gamma a T a T It follows that a T (a T - where the last inequality uses (41) and Proposition 2.1. Let - x. Then from (41) and Lemma 2.2 we have where the last inequality uses the facts that Q - A T S \Gamma2 A, and - T A T S \Gamma2 Combining (29), (42), and (43), we obtain ~ The proof is completed by substituting 5 Dropping a Constraint In this section we consider the effect of dropping a constraint, as in Step 4 of the algorithm. For simplicity we assume that oe P be the new constraint system obtained by deleting the mth constraint in the original system [A; b] defining P. Throughout we use the tilde (~) notation to denote quantities related to the reduced constraint system [ ~ Theorem 5.1 Suppose that x has P is obtained by deleting the mth constraint defining P. Then 1. ~ 2. oe i - ~ 3. k~gk ~ Proof: See [3, Lemma 5.1, Lemma 5.2, and Theorem 5.3]. 2 We will use Theorem 5.1 to bound the change in our fundamental proximity measure following the deletion of a constraint. We use - ~ -(x) to denote the value of - with respect to the reduced constraint system [ ~ Theorem 5.2 Assume that x has Let ~ P be obtained by deleting the mth constraint defining P. Then ~ P is bounded, ~ ~ ! is the volumetric center of ~ P. Proof: Note that - ~ -, from part 2 of Theorem 5.1, and the fact that ~ Applying part 3 of Theorem 5.1, we obtain -k~gk ~ min where the second inequality uses the assumption that - It is clear that the right-hand side of (45) is increasing in oe min , and substituting oe results in -k~gk ~ Assume for the moment that - ~ - :833-. Then (46) implies that - ~ -k~gk ~ ~ from Corollary 2.7. Alternatively assume that - ~ - :833-. Since in any case - ~ implies that - ~ -k~gk ~ In addition, oe min - :04 implies - (2 From Corollary 2.7 we then have ~ so in all cases ~ as claimed. In addition, we have ~ and part 1 of Theorem 5.1 gives ~ Then ~ :0326 follows from (47), (48), and ~ 6 Conclusion From a practical standpoint, this paper gives the best result to date for a cutting plane method for the convex feasibility problem based on the volumetric barrier. From the stand-point of theoretical complexity, the most interesting open problem is the use of central cuts with the volumetric barrier, while requiring only O(1) Newton (or Newton-like) steps following the introduction of a cut, as is possible when shallow cuts are employed ([14], [3]). Although the affine step (40) is sufficient to obtain an O(1) bound on ~ as in Theorem 4.6, this bound is too weak relative to - oe min to show that O(1) steps suffice to return to a suitable proximity of the new volumetric center ~ !. As a result it becomes necessary to use a proximity measure based on - in place of -, leading to a worst-case decrease of n) instead of \Omega\Gamma/1 in the steps on Step 5 of the algorithm. In practice the algorithm might of course do much better than these worst-case bounds indicate, but serious computational work using the volumetric barrier has not yet been conducted. For the analytic center cutting plane method it is relatively easy to show that O(1) steps suffice to return to a suitable proximity of the new analytic center following the addition of a central cut [7]. (The basic analytic center cutting plane method is not a polynomial time algorithm, however. To date the only polynomial cutting plane algorithm based on analytic centers, due to Atkinson and Vaidya [4], uses shallow cuts.) The complexity analysis for the analytic center cutting plane method can also be extended to multiple cuts ([11], [15]), and deep cuts ([6], [8]). Similar results for the volumetric cutting plane method would be desirable. In [13] a result allowing multiple cuts is developed, but in addition to the very small constants required throughout [13], the multiple cut result requires a "Selective Orthonormalization" procedure that weakens the original cuts in the interest of constructing a feasible affine step. --R "Large step volumetric potential reduction algorithms for linear programming," "Volumetric path following algorithms for linear program- ming," "On Vaidya's volumetric cutting plane method for convex programming," "A cutting plane algorithm for convex programming that uses analytic centers," "The ellipsoid method: a survey," "Using the primal dual infeasible Newton method in the analytic center method for problems defined by deep cutting planes," "Complexity analysis of an interior point cutting plane method for convex feasibility problems," "Shallow, deep, and very deep cuts in the analytic center cutting plane method," Geometric Algorithms and Combinatorial Optimization "Complexity of some cutting plane methods that use analytic cen- ters," "Analysis of a cutting plane method that uses weighted analytic centers and multiple cuts," "Complexity estimates of some cutting plane methods based on analytical barrier," "A long step cutting plane algorithm that uses the volumetric barrier," "A new algorithm for minimizing convex functions over convex sets," "Complexity analysis of the analytic center cutting plane method that uses multiple cuts," --TR
convex programming;cutting plane method;volumetric barrier
589127
An Interior-Point Approach to Sensitivity Analysis in Degenerate Linear Programs.
We consider an interior-point approach to sensitivity analysis in linear programming developed by the authors. We investigate the quality of the interior-point bounds under degeneracy. In the case of a special type of degeneracy, we show that these bounds have the same nice asymptotic relationship with the optimal partition bounds as in the nondegenerate case. We prove a weaker relationship for general degenerate linear programs.
Introduction Sensitivity analysis (or post-optimality analysis) is the study of how the optimal solution of an optimization problem changes with respect to the changes in the problem data. The possible presence of errors in the problem data often makes sensitivity analysis as important as solving the original problem itself. In the context of linear programming (LP), sensitivity analysis can be performed using an optimal basis approach (as in the simplex method) or an optimal partition approach, where the optimal partition refers to knowing, for each index, whether the corresponding component of an optimal primal solution or of an optimal dual slack vector can be positive. The latter approach has close connections with interior-point methods since such methods, when properly terminated, provide an optimal solution in the relative interior of the optimal face, from which the optimal partition is readily available. In fact, as will shortly be discussed in detail, the optimal partition approach has been developed by Adler and Monteiro [1] and Jansen, de Jong, Roos and [7] as a promising alternative in order to circumvent the drawbacks of the classical optimal basis approach in the presence of degeneracy. Later, Monteiro and Mehrotra [9] extended this approach by relaxing the requirement that the optimal partition be known. They also provided two methods to estimate the range of perturbations, each of which can be performed at any optimal solution, regardless of where it lies on the optimal face. More recently, Greenberg, Holder, Roos and Terlaky [5] related the dimension of the optimal set to the dimension of the set of objective perturbations for which the optimal partition is invariant. Greenberg [4] considered the simultaneous perturbations of the right-hand side and the cost vectors from an optimal partition perspective. Recently, the authors studied perturbations of the right-hand side and the cost parameters in linear programming [12], motivated by how interior-point methods from a near-optimal pair of strictly feasible solutions for a problem and its dual would compare with the optimal basis approach obtained from a nondegenerate optimal basic solution for such perturbations. The proposed interior-point perspective stems from the objectives of regaining feasibility and maintaining near-optimality in a single iteration of the interior-point method. This requires the setup of the "right" Newton system among many possible choices in order to achieve both objectives simultaneously. Such a perspective provides a basis for the comparison of the interior-point and the simplex approaches to sensitivity analysis. Under the assumption of a unique, nondegenerate optimal solution, the authors showed that the Newton system proposed in [12] is the "right" one in the sense that it yields asymptotically the same bounds on perturbations as those that keep the current basis optimal (after symmetrization with respect to the origin). Similar results, but changing only one of the primal or dual near-optimal solutions, were obtained by Kim, Park and Park [8]. However, most LPs arising from real-life problems are degenerate. Our goal in this paper is to investigate the quality of the bounds from the interior-point perspective in the absence of the strong assumption of nondegeneracy. This will lead to a complete analysis of the interior-point perspective proposed in [12]. In doing so, we need something to compare our interior-point bounds with. In contrast to the nondegenerate case, the presence of multiple optimal bases makes a simplex-based approach unsuit- able, as will be explained shortly. We therefore compare our bounds to those obtained from considering how much the right-hand side or the cost vector can change while maintaining the same optimal partition. Consequently, we use completely different tools for our analysis in this paper. The next section is devoted to the preliminaries including the introduction of the tools relevant for the analysis as well as the restatement of our interior-point approach. Section 3 discusses the equivalence between the primal and dual formulations and shows that it suffices to consider perturbations of the right-hand side only. We analyze the interior-point bounds under a special case of degeneracy in Section 4 and extend the analysis to the general degenerate case in Section 5. We present and discuss some computational results in Section 6 and Section 7 concludes the paper. Preliminaries We consider the LP in the following standard form: c T x; subject to The associated dual LP is given by subject to A T y constitute the data, and (x; are the decision variables. Throughout this paper, the coefficient matrix A will be fixed and we will consider one-dimensional perturbations of the right-hand side vector b and the cost vector c, i.e., b will be replaced by b + t\Deltab and c by c and \Deltac will be fixed in IR m and IR n , respectively, and t 2 IR will be the parameter. This is also called parametric analysis in the literature. We will make the following assumptions: 1. The coefficient matrix A has full row rank. 2. Both (P) and (D) have strictly feasible solutions, i.e., there exist x ? and y such that The classical approach to sensitivity analysis has been based on the simplex method. Assuming that an optimal solution exists, the simplex method terminates with a basic optimal solution along with a corresponding basis. A natural criterion for the allowable perturbations in the data is then given by the following: how much perturbation in the data can one allow so that the current basis remains optimal for the perturbed LP? Let us consider the parametric right-hand side (RHS) problem, i.e., let b be replaced 0g. It is well-known that v is a convex, piecewise linear, continuous function of t. The parametric RHS problem includes finding out all the "breakpoints" of v(t). Fixing a value of t, say at 0 for the purposes of this paper, the classical approach to sensitivity analysis then provides the set of values of t for which an optimal basis for remains optimal for the resulting LPs parametrized by t. This is called the optimality interval associated with an optimal basis. Note that the optimal basis approach indeed yields the breakpoints of v(t) around 0 under primal and dual nondegeneracy (which holds only if 0 itself is not a breakpoint of v(t)). However, the presence of primal and/or dual degeneracies is a shortcoming for this approach since, for example, multiple optimal bases might yield different optimality intervals. This shortcoming has been observed by several researchers. Adler and Monteiro [1], and Jansen, de Jong, Roos and Terlaky [7] developed an optimal partition approach to sensitivity analysis and showed that the optimality intervals associated with the optimal partitions uniquely and unambiguously identify the breakpoints of v(t) and the intervals between the consecutive breakpoints. By the symmetry between (P) and (D), which will be treated in more detail in Section 3, the same conclusions also hold for the parametric analysis of the cost vector c. The idea of the optimal partition is based on a well-known result of Goldman and Tucker [2]. The optimality conditions for (P) and (D) are given by primal and dual feasibility and complementary slackness, that is, a triple (x; y; s) is optimal for (P) and (D) if and only if it satisfies where x i and s i denote the ith components of x and s, respectively. and\Omega D denote the set of optimal solutions for (P) and (D), respectively. Then, we can define two index sets as ng 2\Omega D g: (2.2) The optimality conditions (2.1) imply that B " ;. The Goldman-Tucker result indicates that B and N actually partition the index set ng. Therefore, there exist at least one primal solution x 2\Omega P and one dual solution 2\Omega D such that x Such a solution will be called strictly complementary and B and N will be called the optimal partition. In contrast to the possibility of multiple optimal bases, the optimal partition is unique for a given LP instance. We will denote by B and N the columns of A corresponding to the indices in B and N , respectively, and we will also partition the cost vector c as c B and c N , and the variables x and s as xB and xN , and s B and s N accordingly. Note that if (x; y; s) is a strictly complementary solution, then we have xB ? 0, Let us again restrict our attention to one-dimensional perturbations of the right-hand side vector b. The optimal partition approach is based on maintaining the whole dual optimal set invariant rather than an optimal basis as in the classical simplex approach. Note that perturbations of b do not affect the dual feasible region. Conse- quently, the range of t is given by solving two auxiliary LPs. More precisely, if b is replaced by b + t\Deltab, and if\Omega D denotes the dual optimal set for (D) (i.e., the lower and upper bounds on t are given by the optimal values of subject to We will call the resulting bounds the optimal partition bounds. Note that both problems are always feasible since together with any x 2\Omega P satisfy all the constraints. Fixing the dual optimal set\Omega D is equivalent to fixing the optimal partition B and N by the Goldman-Tucker result. Therefore, the (possibly infinite) last constraint set in (AUX1) can be replaced by the equivalent single constraint x T s point in the relative interior of\Omega D (hence s This condition, in turn, is the same as setting Consequently, (AUX1) can be written in the following simplified subject to The analogous derivation for the one-dimensional perturbations of the cost vector c leads to the following auxiliary problems, whose optimal values give the optimal partition bounds for t when c is replaced by c subject to Here, \Deltac B and \Deltac N constitute the corresponding partition of \Deltac. Before getting into the symmetrized bounds we would like to recall an important result about the dimensions of the optimal solution and\Omega D . In what follows, dim(\Delta) denotes the dimension and j \Delta j denotes the cardinality of a set. The reader is referred to Lemma IV.44 in [10] for a proof. Proposition 2.1 dim(\Omega dim(\Omega 2.1 Symmetrized Bounds The auxiliary problems (AUX1) and (AUX2) can be reformulated in the following way. Let us consider (AUX1) and let x 2\Omega P . Then, the equality constraint can be rewritten as Therefore, by a change of variable, if we let then (AUX1) is equivalent to subject to Next, we will tighten the constraints in the above formulation by putting upper bounds on u as well, and our choice for the upper bound will be x B , which will give the largest L1 -box around the origin which is contained in the feasible region: subject to We will call (SA1) the symmetrized LP and the resulting optimal solutions the symmetrized bounds. The formulation of (SA1) reveals that if (u ; - ) solves the maximization problem, then (\Gammau ; \Gamma- ) solves the minimization problem. Therefore, it suffices to solve one LP as opposed to solving two LPs to obtain the optimal partition bounds from (AUX1). A similar treatment of (AUX2) gives rise to the following symmetrized LP: subject to \Gammas which is obtained by replacing y \Gamma y by v and s N by w, where (y ; s ) 2\Omega D . Finally, a similar symmetrization has been applied to w. Next, we would like to discuss the relationship between the auxiliary and the symmetrized LPs. First of all, let us assume that both (P) and (D) have unique and nondegenerate solutions. Then, Proposition 2.1 implies that B is actually a square and nonsingular matrix, hence invertible. In fact, B is the optimal basis. Conse- quently, (AUX1) and (AUX2) are trivial to solve and their optimal solutions coincide with the optimal basis bounds arising from the simplex method. With this observation, the constraints of (AUX1) reduce to -B B or -(X B is the diagonal matrix whose components are given by x B and e denotes the vector of ones in the appropriate dimension. Similarly, the constraints of (SA1) can be rewritten as \Gammae -(X is the L1 -norm. A similar treatment reveals that the constraints of are equivalent to -(S N is defined similarly, and that those of (SA2) to The derivations (2.3)-(2.6) imply the following relationship between the auxiliary and the symmetrized LPs: let the optimal partition bounds given by the optimal solutions of the auxiliary LPs (including possibly \Sigma1). Then, the symmetrized bounds for t are (\Gamma- s Therefore, the symmetrized bounds are indeed equal to the "symmetrization" of the optimal partition bounds. Next, let us assume that (P) has a unique but degenerate solution. Then, by Proposition 2.1, B is nonsquare but it has full column rank. Therefore, (AUX1) is still easy to solve. If \Deltab does not lie in the range space of B, then the optimal solutions of (AUX1) and (SA1) are all zero (which implies that is a breakpoint of v(t)). Otherwise, there exists a unique vector v such that \Deltab, and hence, the constraints of (AUX1) are equivalent to -(X Similarly, the constraints of (SA1) can be stated as Once again, we conclude that a similar symmetry as in (2.7) continues to hold between (SA1) and (AUX1). In a similar manner, one can show that such a relationship holds between (SA2) and (AUX2) if (D) has a unique but degenerate solution. The preceding discussion shows that the optimal solutions of the auxiliary and the symmetrized LPs have the nice relationship (2.7) as long as there is a unique optimal solution that one can use to symmetrize the constraints of the auxiliary LPs to obtain the symmetrized LPs. An interesting question then is whether the same nice relationship continues to hold between the auxiliary and the symmetrized LPs if there are multiple optimal solutions, that is whether the symmetrized bounds are independent of the choice of the optimal solution used to symmetrize the constraints. Unfortunately, the answer is no as shown by the following example. Let (P) be given by 0g. Then (P) has multiple optimal solutions given by with an optimal value of 0. If the right-hand side is perturbed to (0; then the reader can easily verify that (AUX1) yields (\Gamma1=3; +1) as the optimal partition bounds, whereas the symmetrized bounds are (\Gammafi; +fi) if one uses the optimal solutions with to symmetrize the constraints, and (\Gamma1=3; 1=3) if those with fi - 1=3 are used. This example illustrates that in case of multiple optimal solutions, the symmetrized bounds are dependent on the optimal solution used in the formulation of the symmetrized LPs. Therefore, the relationship (2.7) no longer holds between the symmetrized and the auxiliary LPs. However, we will keep using the symmetrized LPs for two reasons. First of all, at least in the unique solution case, they bear a nice relationship to the auxiliary LPs. For our analysis, we will always choose an optimal solution in the relative interior of the optimal set; therefore the symmetrization will hopefully allow more room for the decision variables of the symmetrized LPs. Secondly, the symmetrized LPs are easier to deal with than the auxiliary LPs and the symmetrized bounds will provide a good comparison basis for our interior-point approach proposed in [12], as will be analyzed in the subsequent sections. 2.2 Interior-Point Approach and Central Path Neighborhood We will start with a brief review of the primal-dual path-following interior-point meth- ods. The reader is referred to [11] for an extensive treatment. The central path is a path of strictly feasible points (x(-); y(-); s(-)) parametrized by a positive scalar -. Each point on the central path satisfies the following system for some - ? 0: Under the two assumptions in Section 2, such a solution exists and is unique for each positive -. Interior-point methods are iterative algorithms that generate iterates which "follow" the central path in the direction of decreasing - towards the primal-dual optimal \Theta\Omega D . The iterates generated typically lie in some neighborhood of the central path. For any given feasible iterate (x; the duality gap is given by c T and we define the duality measure - as denote the set of feasible and strictly feasible primal-dual points respectively, that is, One of the commonly used neighborhoods in interior-point methods is the so-called wide neighborhood, denoted by N where At each iteration, given (x; (fl), the algorithm determines a search direction (\Deltax; \Deltay; \Deltas). This direction is usually obtained by seeking an approximation to the point on the central path corresponding to some parameter -, and then applying Newton's method to the nonlinear system of equations (2.10). Finally, a (damped) step is taken in this direction in such a way that the resulting iterate still lies in N \Gamma1 (fl). However, as in the context of target-following methods, one might seek an approximation to a point other than the one on the central path. We will say that a Newton step from (x; targeting the feasible pair of points is the direction (\Deltax; \Deltay; \Deltas) obtained from the Newton's method applied to (2.10) with -e replaced by X e: A T \Deltay Next, we describe the interior-point approach proposed by the authors in [12]. Given a primal-dual pair of LPs (P) and (D), let us assume that b or c is perturbed in some fixed direction. Assuming strictly primal-dual feasible for (P) and (D), a full Newton step is taken from (x; targeting "a feasible point" of the perturbed LPs which satisfies X is possible that there is no such feasible point for the perturbed LPs, however, the Newton step as given above is still well-defined.) We state the results formally, referring the reader to [12] for the proofs. Note, in particular, that the duality gap of the resulting feasible iterate for the perturbed LPs is no greater than that of the original iterate. Proposition 2.2 Assume that (x; y; s) is a strictly feasible point for (P) and (D) and the right-hand side vector b is replaced by b+t\Deltab, where t 2 IR and \Deltab 2 IR m . Suppose a Newton step is taken from (x; targeting the feasible pair of points the perturbed pair of LPs that satisfies X a full Newton step will yield a feasible iterate for the new problem if and only if . Moreover, in this case the new iterate will have duality gap at most x T s. Proposition 2.3 Assume that (x; y; s) is a strictly feasible point for (P) and (D) and the cost vector c is replaced by c Suppose a Newton step is taken from (x; targeting the feasible pair of points of the perturbed pair of LPs that satisfies X a full Newton step will yield a feasible iterate for the new problem if and only if . Moreover, in this case the new iterate will have duality gap at most x T s. Under primal-dual nondegeneracy, the bounds arising from Propositions 2.2 and 2.3 computed at near-optimal solutions for (P) and (D) asymptotically equal the symmetrized bounds arising from (SA1) and (SA2) [12]. The goal of this paper is to investigate the quality of these bounds in the absence of the nondegeneracy assumption We first present a nice characterization of the distance of the strictly feasible primal-dual points strictly complementary optimal solutions in terms of the duality gap -n. Using this characterization, we derive some bounds on the components of such points. In what follows, xB , xN , s B and s N denote the partitions of x and s according to the optimal partition B and N as before. Furthermore, we will use the bounds O(-), \Omega\Gamma -) and \Theta(-) interchangeably for scalars as well as vectors and matrices by which we mean each entry satisfies the stated bounds. O(-) will indicate that the quantity (in absolute value) is bounded above by some positive multiple of -, where the multiple depends on the primal-dual instance (P) and (D) but does not depend on the particular strictly feasible point or on -. Similarly, \Omega\Gamma -) will indicate a lower bound by some positive multiple of - and \Theta(-) will mean a lower and upper bound by some positive multiples of -. The following proposition will be useful for the analysis that follows. Actually, the proposition continues to hold for any feasible solutions and even for a point where feasibility is violated by O(-). The statement below suffices for the purposes of this paper. Proposition 2.4 Let (x; y; s) be a strictly feasible point for (P) and (D) with duality gap -n. Then, there exists a strictly complementary optimal solution and (D) such that Proof: Optimal solutions of (P) and (D) satisfy the linear system strictly feasible point satisfies the same linear system with the third equality replaced by c T [6] indicates that there exists a solution (-x; - of the first system such that (-x; - +O(-). The result follows immediately if (-x; - s) is strictly complementary. If not, there exists an arbitrarily small perturbation of (-x; - s) which leads to a strictly complementary solution and (2.17) follows since - ? 0. The following corollary immediately follows from Proposition 2.4 since x s optimal solution of (P) and (D). Corollary 2.1 Let (x; y; s) be a strictly feasible point for (P) and (D) with duality gap -n. Then, Note that both Proposition 2.4 and Corollary 2.1 hold for any primal-dual strictly feasible (x; s). Next, we derive some more bounds by restricting the iterates to lie in a wide neighborhood given by (2.13). Proposition 2.5 Let (x; duality gap -n for (P) and (D). Then, Proof: similar argument shows \Omega\Gamma126 Finally, together with s N imply O(-). The proof of SB X \Gamma1 Equivalence In this section, we show that the interior-point bounds are independent of the problem formulation. It is well-known that although (P) and (D) do not look symmetric, they can easily be reformulated so that (D) is in the form of (P) and vice versa. We briefly review this reformulation. Let (-x; - s) be such that us consider (D) first. The objective function can be rewritten as where we used and the fact that every feasible pair (y; s) for (D) satisfies Note that the first term is a constant: therefore maximizing b T y is the same as minimizing - x T s. Let K 2 IR (n\Gammam)\Thetan be such that its rows form a basis for the null space of A. Then, premultiplying the equality constraints in (D) by K yields Moreover, if s satisfies (3.2), then c \Gamma s lies in the null space of K, for which the columns of A T form a basis by definition of K. Therefore, there exists y such that A T Consequently, (D) is equivalent to x T s; subject to Note in particular that K has full row rank by its definition. If we take the dual of (D'), we obtain subject to K T It is not hard to see that (P) and (P') are also equivalent by a similar argument. Therefore, the roles of (P) and (D) can be interchanged via this reformulation. Let us now focus on perturbations of c, i.e., let c be replaced by c+t\Deltac. By the above reformulation, this is the same as replacing the right-hand side of (D') by - c Therefore, Proposition 2.2 can be used to evaluate the interior-point bound at a strictly feasible primal-dual pair (s; x) (note that the roles of x and s are interchanged). We need to compute On the other hand, one can also use Proposition 2.3 to compute the interior-point bound directly at (x; s), which requires the evaluation of A simple manipulation of (3.3) gives rise to another equivalent formula: where \Psi is the orthogonal projection matrix onto the range space of X \Gamma1=2 S 1=2 K T . Similarly, (3.4) is equivalent to where \Xi is the orthogonal projection matrix onto the null space of AX 1=2 S \Gamma1=2 . There- fore, in order to prove that (3.3) and (3.4) are equivalent, it suffices to show that \Psi and \Xi project onto the same subspace, or that the null space of AX 1=2 S \Gamma1=2 equals the range space of X \Gamma1=2 S 1=2 K T . This is easily proven by an inclusion argument: if w satisfies AX 1=2 S \Gamma1=2 Thus, w is in the range space of X \Gamma1=2 S 1=2 K T . Conversely, if AX This proves the equivalence of the interior-point bounds. We next argue that the range of t resulting from the optimal partition bounds is also independent of the formulation. If the two LPs are formulated in the form of (P) and (D), then (AUX2) yields the range of t for perturbations of c. Premultiplying the equality constraints of (AUX2) by leads to (AUX1') given by min w;- w;- which exactly yields the range of t for perturbations of the right-hand side of (D') if one uses the form (D') and (P'). Similarly, if (w; -) is feasible for (AUX1'), then lies in the null space of K. Then, by our previous observation, there exists v such that which is exactly the constraints of (AUX2), completing the proof of the claim. Using this observation, we will carry out our analysis for perturbations of b only in the subsequent sections, and state the corresponding results for changes in c as corollaries. We begin with a special case of degeneracy first and then consider the most general case. 4 Unique Primal Solution Throughout this section, we assume that (P) has a unique but degenerate optimal solution x . Note that by Proposition 2.1, we have linearly independent columns. In this particular case, Proposition 2.4 provides another useful bound on xB for a strictly feasible primal-dual point (x; Corollary 4.1 Assume that (P) has a unique optimal solution x . Let (x; y; s) be primal-dual strictly feasible for (P) and (D) with duality gap -n. Then, An analogous corollary follows if (D) has a unique solution. Corollary 4.2 Assume that (D) has a unique optimal solution (y be primal-dual strictly feasible for (P) and (D) with duality gap -n. Then, Next, we will analyze one-dimensional perturbations of b. 4.1 Perturbations of b In this subsection, we assume that the right-hand side vector b is replaced by b We also assume that (x; strictly feasible point for (P) and (D) for some fl 2 (0; 1]. We will compare the interior-point bounds arising from Proposition 2.2 at with the optimal values of (SA1), i.e., the symmetrized bounds. The interior-point bounds are given by the L1-norm of where Let us now consider (SA1). Since B has full column rank, \Deltab either does not lie in the range space of B, in which case the optimal values of (SA1) as well as (AUX1) are all 0, or there exists a unique v 2 IR jBj such that \Deltab, in which case the constraints of (SA1) reduce to (2.9), from which the symmetrized bounds can be obtained easily. We will consider both situations in turn. Let us start with the second case. Without loss of generality, we can assume that \Deltab has unit L 2 -norm, which implies a bound on v. Then, we need to compute in order to obtain (4.3). However, (4.4) is equivalent to where B and N are the partitions of the coefficient matrix A with respect to B and N as before. Since B has linearly independent columns, there exists a matrix C 2 IR m\Theta(m\GammajBj) such that the augmented matrix [B C] is square and nonsingular: let W be its inverse. Therefore, premultiplying the second equality in (4.5) by W , we obtain I# ~ I# where ~ I is a jBj \Theta jBj identity matrix. Therefore, if we partition ~ N and ~ accordingly as ~ ~ ~ ~ can then be decomposed in the following way: ~ ~ ~ ~ ~ ~ v# where DB and DN are the corresponding partitions of D. By (4.3), we need to compute For notational convenience, let us define F := ~ Note that G has full row rank since A does. The bottom equality in (4.7) can be rewritten as Substituting (4.9) in the top equality in (4.7) gives ~ ~ where PG is the orthogonal projection matrix onto the range space of G T . Therefore, I \Gamma PG is the orthogonal projection matrix onto the null space of G. We briefly review the Neumann lemma now [3]. Let U be an invertible matrix and being used does not really matter: we will always use k \Delta k for the Euclidean norm or the operator norm arising from it.) Then, I +U \Gamma1 V is invertible with kI +U 2. Moreover U given by Now, we apply this result to (4.10) with U := D 2 2.5 implies that both U \Gamma1 and V are O(-) since I \Gamma PG is a projection matrix and has unit Euclidean norm. Therefore, assuming the duality gap -n is small, ~ I +D \Gamma2 It then follows that I +D \Gamma2 However, by Proposition 2.5, F is O(- 1=2 ), D \Gamma2 B is O(-) and X \Gamma1 B is O(1). Consequently, the second term on the right hand side of (4.13) is O(- 2 ) since kI \Gamma PG k - 1. Finally, Corollary 4.1 implies X \Gamma1 We have thus obtained the top part of (4.8). For the lower part, we get where we substituted (4.9) for ~ Proposition 2.5 implies (XN SN Combining these bounds with leads to Using (4.8), we conclude that the L1-norm of the quantity (4.3) we need to evaluate is given by O(-) The reciprocal of (4.17) gives the desired interior-point bound. Consequently, if the duality gap -n is small, we conclude by comparing (4.17) with (2.9) that the interior-point approach yields exactly the same bound as the optimal solution to (SA1) asymptotically in -. Next, we address the situation where \Deltab does not lie in the range space of B. In this case, the optimal values of both (AUX1) and (SA1) are clearly 0. \Deltab can be uniquely written as where [B C] is nonsingular as before and v C is a nonzero vector. Once again, we need to compute (4.3). We follow a similar treatment as before, and corresponding to (4.7) we ~ ~ ~ ~ ~ ~ The bottom part can be expanded as ~ ~ ~ However, (4.8) implies that the term in the brackets is exactly the bottom part of the quantity (4.3) we seek. Let us denote that term by p and let XN is equivalent to ~ nonzero, the norm of q is bounded below, that is, kqk - ff ? 0 where ff is the norm of the least squares solution. Therefore, e.g. [3]). (Note that jBj ! n since can happen only if case \Deltab is always in the range of B.) However, k1 kpk1 since XN This implies kXN k1 where the last equality follows from Corollary 2.1. Therefore, as - tends to 0, kpk1 tends to 1, which implies that the interior-point bound given by its reciprocal tends to 0 as desired. We remark that if is the only optimal solution of (P), which can happen only if In this case, the top part of (4.8) disappears. The interior-point bound is then given by the reciprocal of kpk1 , where p is as defined after (4.20). By the preceding argument, the interior-point bound tends to 0 as - approaches 0. This is still in agreement with the optimal partition bounds since any nonzero perturbation of b leads to a change in the optimal partition and hence, the optimal partition bounds in this case are also equal to 0. Therefore, we have proved the following theorem: Theorem 4.1 Let (x; be a primal-dual strictly feasible point for (P) and (D). Assume that (P) has a unique but degenerate optimal solution and that b is replaced by b . Then the interior-point bound evaluated at yields exactly the same value as the optimal solution of (SA1) asymptotically in -, where The following corollary of Theorem 4.1 is an immediate consequence of the equivalence between (P) and (D) as discussed in Section 3. One uses Corollary 4.2 in place of Corollary 4.1 in the preceding analysis. Corollary 4.3 Let (x; be a primal-dual strictly feasible point for (P) and (D). Assume that (D) has a unique but degenerate optimal solution and that c is replaced by c+t\Deltac where t 2 IR and \Deltac 2 IR n . Then the interior-point bound evaluated at yields exactly the same value as the optimal solution of (SA2) asymptotically in -, where It does not appear that we can obtain better results for perturbations of c in the case of a unique primal optimal solution (but not dual optimal solution) than those arising from the analysis of the general case in the next section. A similar remark holds for perturbations of b in the case of a unique dual optimal solution (but not primal optimal solution). 5 General Case In this section, we turn our attention to the most general case where both (P) and (D) may have multiple optimal solutions. As the small example given at the end of Section 2.1 reveals, some complications arise in the presence of multiple optimal solutions. For instance, unlike the previous case, the symmetrized bounds become dependent on the optimal solution of (P) used in the formulation of (SA1) if (P) has multiple optimal solutions. Furthermore, they do not necessarily coincide with the "symmetrizations" of the optimal partition bounds arising from (AUX1). Similar remarks hold for the relationship between (SA2) and (AUX2) if (D) has multiple optimal solutions. Despite this complication arising from the presence of multiple optimal solutions, we aim to be able to say something about the quality of the interior-point bounds at least in comparison with the symmetrized bounds. In the next subsection, we analyze perturbations of b in this general setting. 5.1 Perturbations of b Let (P) have multiple optimal solutions and let b be replaced by b and \Deltab 2 IR m . Suppose that (x; strictly feasible where For such a point, Proposition 2.4 guarantees the existence of a strictly complementary solution whose distance from (x; y; s) is bounded above by the duality gap n-. We will compare the interior-point bounds evaluated at with the optimal values of (SA1). Among other optimal solutions of (P), the x above will be the particular choice of the primal optimal solution to be used in the formulation of (SA1). The use of such an optimal solution in the relative interior of the primal optimal set is likely to leave more room for the decision variables of (SA1) since x Let us first consider (SA1). The constraints of (SA1) are k. Clearly we have r - m and r ! k since Proposition 2.1 implies dim(\Omega which is positive by our assumption. This, in turn, implies that r ? 0 since (assuming no columns of A are identically zero). A QR factorization of B yields orthogonal and R 2 IR m\Thetak is upper triangular with R 1# rows. Note that R 1 has full row rank. Premultiplying the equality constraints in (5.1) by Q T yields R 1# f with f that (SA1) has a nontrivial optimal solution - if and only if f First, we consider the nontrivial case. (Since f \Deltab is nonzero, this implies that k ? 0.) Let (- ; u ) be an optimal solution to the maximization problem with - 6= 0. Note that - is finite since u is bounded (this follows since B 6= ;). Then, we have f The interior-point approach, on the other hand, requires the evaluation of (4.3) at s). By (5.4), we then need to evaluate the L1-norm of Let Premultiplying the second equality in (5.6) by Q T gives ~ R T ~ ~ ~ R 1# where ~ are the appropriate partitions of ~ are those of ~ us define F := ~ can then be decomposed into two equations as Note, in particular, that both G and H have full row rank since R 1 and A do. From the second equation in (5.9), we obtain ~ Substituting (5.10) in the first equation of (5.9) leads to where I \Gamma PH is the orthogonal projection matrix onto the null space of H. Proposition 2.5 implies that the second term in parentheses in the second equation above is O(-) since kI \Gamma PH k - 1. In order to apply Neumann's lemma, we need to show that Lemma 5.1 (GG Proof: We use the "thin" QR factorization of G columns and Z is upper triangular and nonsingular. Then, (GG Therefore, it suffices to find an upper bound on Z \Gamma1 . We have Therefore, (R (R Y . However, by Proposition 2.5, D \Gamma1 which implies that Z completing the proof. We can now apply Neumann's lemma to (5.11). Using the same notation as in (4.11) we have U := GG T and V := F . Note that both U \Gamma1 and V are O(-). We obtain ~ where we used R By (5.5) and (5.6), we need Let us define For the top part of (5.14) we need to evaluate where we used (5.13), (5.15) and where PG is the orthogonal projection matrix onto the range space of G T . Consider the second term in the right hand side of the second equality. By Proposition 2.5, (SB XB ) \Gamma1=2 is O(- \Gamma1=2 ), V is O(-) and D \Gamma1 B is O(- 1=2 ). Lemma 5.1 implies that fore, the whole expression is O(- 2 ). We conclude that the top part of (5.14) is Let us next consider the lower part of (5.14). We need to compute ~ ~ By (5.13) the first term in (5.18) is given by Note that by the preceding discussion. As for the second term in brack- ets, we have both (GG are O(-), which implies the whole second term is O(- 5=2 ). Thus, the expression in brackets is O(- 1=2 ). By Proposition 2.5, (SN XN ) \Gamma1=2 is O(- \Gamma1=2 ), whereas both F T and D \Gamma1 are O(- 1=2 ). We therefore conclude that (5.19) is O(-). For the second term in (5.18), we use (5.10) together with (5.13): \Gamma(S Note that ~ O(-) by the preceding arguments. The fact that kPH k - 1 together with (SN XN ) \Gamma1=2 being O(- \Gamma1=2 ) and F T being O(- 1=2 ) implies (5.20) is O(-). Therefore, we conclude that the lower part of (5.14) is O(-). Combining this result with (5.17) yields the following: r := (1=- ) O(-) Consequently, we need to evaluate the L1-norm of (5.21) and take its reciprocal. Observe that X \Gamma1 Proposition 2.4. Using this, we derive an upper bound on the L1-norm of (5.21). Thus, . Furthermore, since e.g. [3]), where k. Finally, since u is optimal for (SA1), k(X Therefore, /s We conclude that the interior-point bound, which is the reciprocal of (5.24), is then bounded below bykrk 1 Note, in particular, that the lower bound tends to 1= k, independent of n, as if s) is on the central path. We next consider the case where \Deltab is not in the range space of B. Again, in this case, the symmetrized bounds as well as the optimal partition bounds are all 0. The QR factorization of B can be rewritten as use (5.2) and [Q 1 is the appropriate partition of Q. Since Q is orthogonal, \Deltab can uniquely be expressed as Arguing similarly to Section 4, we need to evaluate (4.3), which in turn requires the computation of Premultiplying (5.27) by Q T leads to ~ R T ~ ~ ~ which looks like (4.19). Essentially the same arguments as in Section 4 reveal that the interior-point bound tends to 0 as - approaches 0. Therefore, we have proved the following theorem. Theorem 5.1 Let (x; be a primal-dual strictly feasible point for (P) and (D) with duality gap -n. Assume that (P) has multiple optimal solutions and that b is replaced by b . If the strictly feasible solution given by Proposition 2.4 is used for symmetrization in (SA1), then the ratio of the interior-point bound evaluated at (x; y; s) to the value of the optimal solution of (SA1) is at least p Note that the presence of multiple primal optimal solutions implies k ? 0, therefore, the expression (5.29) is well-defined. As in Section 4, Theorem 5.1 leads to the following corollary by the discussion in Section 3. Due to the interchange of the roles of the basic and nonbasic variables in the reformulation given in Section 3, k in the denominator of is replaced by (n \Gamma k). Under the assumption of multiple dual optimal solutions, Proposition 2.1 indicates that m ? r, which implies k ! n since A has full row rank. Corollary 5.1 Let (x; be a primal-dual strictly feasible point for (P) and (D) with duality gap -n. Assume that (D) has multiple optimal solutions and that c is replaced by c . If the strictly feasible solution given by Proposition 2.4 is used for symmetrization in (SA2), then the ratio of the interior-point bound evaluated at (x; y; s) to the value of the optimal solution of (SA2) is at least p fl 6 Computational Results In the previous sections, we have provided a theoretical basis for the behavior of the interior-point bounds evaluated at the near-optimal solutions. We present some computational results in this section to shed some light on the performance of the interior-point bounds in practice. We have generated random LPs with 400. The input parameters are the number of basic variables (jBj) and dimension of the primal optimal set (dim(\Omega P )), which together determinedim(\Omega D ) and rank(B). This allows us to incorporate all scenarios of primal and dual degeneracies into the random LPs. We first generate a suitable matrix A, then a strictly complementary pair of solutions, and finally set b and c to make these feasible and hence optimal. Having generated a random LP with the prespecified degeneracies, we obtain a strictly feasible, near-optimal solution by perturbing the known strictly complementary optimal solution. Next, random perturbations of b and c are generated in the correct subspaces so that (AUX1) and (AUX2) have nontrivial optimal solutions. We compute the interior-point bounds evaluated at those near-optimal solutions and compare them with the optimal solutions to (AUX1) and (AUX2) as well as the optimal solutions to the symmetrized LPs (SA1) and (SA2), where the initially generated strictly complementary optimal solution is used to symmetrize the constraints. We present our results for various degeneracy scenarios in Table 6. Eight instances with various levels of primal-dual degeneracies are reported. For each instance, the interior-point bounds are evaluated at two iterates corresponding to each row. DP and DD are the dimensions of the primal and dual optimal sets, respectively. - is the duality gap measure given by x T s=n, and fl is the parameter of the narrowest wide central- path neighborhood containing the iterate. (AUX1) and (AUX2) are the minimum of the absolute values of the optimal values of the corresponding minimization and maximization problems (symmetrizations). (SA1) and (SA2) are the optimal values of the symmetrized maximization problems. Finally, IPB and IPC are the upper interior-point bounds for changes in b and c evaluated at the corresponding iterates. The predicted nice theoretical behavior of the interior-point bounds is exhibited in Instances 1,2 and 4 for changes in b and in Instances 4,6 and 8 for changes in c. Observe that the bounds converge to the symmetrized bounds even though fl is very small, which is typical in practice. For the remaining degeneracy scenarios, the interior-point bounds lie within a factor of the symmetrized bounds as discussed in the previous section. It is worth noting, however, that the actual ratio seems to be much better than the theoretical worst-case ratios (5.29) and (5.30). In our extensive computational tests, the ratio was never worse than a hundredth although the predicted lower bounds and (5.30) are on the order of 10 \Gamma5 in most of the instances. Finally, we note that the condition number of AD 2 A T blew up in all of the degenerate instances as expected. Therefore, the numerically unstable results have been discarded. Furthermore, the bound for changes in b seems to be computationally much more stable than its counterpart for c; however, this is most likely due to the fact that we use (2.15) and (2.16) to compute the bounds. By the equivalence discussed in Section 3, this problem can be overcome using (3.3) instead of (3.4) at the extra cost of computing K, which can easily be obtained by a QR factorization of A T . 7 Conclusion In this paper, we have studied the quality of the bounds arising from the interior-point perspective on sensitivity analysis developed by the authors in [12]. By relaxing the strong assumption of nondegeneracy, we have been able to consider all possible degeneracy scenarios and to investigate how our bounds compare with those arising from the optimal partition approach to sensitivity analysis. If the primal problem has a degenerate but unique optimal solution, then our approach yields the same bounds as the "symmetrized" optimal partition bounds for perturbations of b. By the equivalence discussed in Section 3, the same relationship holds for perturbations of c if the dual problem has a degenerate but unique opti- Table 1: Computational Results (m 200, Ins DD IPB IPC40160e-5e-65.47264.1633.005e-3e-63.06112080e-5e-614.61327.6653.806e-3e-62.670120140e-5e-52195.607223.65325.957e-3e-525.9572000e-6e-50.6760.008300.00829e-4e-50.00823200100e-5e-54030.46669.4843.086e-3e-53.0992800e-7e-5220.6561.5611.561e-5e-51.547280100e-4e-48512.007115.83438.813e-3e-438.5403600e-6e-4552.0142.6362.638e-4e-4 87. mal solution. This result directly extends the previous result proved in [12] under the assumption of a unique and nondegenerate solution. We then considered general degenerate LPs. In this case, we were able to show that our interior-point approach would yield bounds that are at least a certain fraction of the symmetrized bounds, where the fraction depends on certain characteristics of the problem instance and of the iterate at which the bounds are evaluated. Our extensive computational tests suggest that the ratio in practice is much better than the predicted worst-case ratio. Although this result is not as strong as the aforementioned results, our interior-point bounds still yield some useful information on the range of allowable perturbations. The fact that the cost of the evaluation of our bounds is simply the same as that of an interior-point iteration makes it more appealing given the cost of solving two LPs to obtain the range from the optimal partition approach. --R A geometric view of parametric linear program- ming Theory of linear programming. Matrix Computations. Simultaneous primal-dual right-hand-side sensitivity analysis from a strictly complementary solution of a linear program On the dimension of the set of rim pertubations for optimal partition invariance. On approximate solutions of systems of linear inequalities. Sensitivity analysis in linear programming A general parametric analysis approach and its implication to sensitivity analysis in interior point methods. Theory and Algorithms for Linear Optimization Sensitivity analysis in linear programming and semidefinite programming using interior-point methods --TR
degeneracy;sensitivity analysis;linear programming;interior-point methods
589159
On Some Properties of Quadratic Programs with a Convex Quadratic Constraint.
In this paper we consider the problem of minimizing a (possibly nonconvex) quadratic function with a quadratic constraint. We point out some new properties of the problem. In particular, in the first part of the paper, we show that (i) given a KKT point that is not a global minimizer, it is easy to find a "better" feasible point; (ii) strict complementarity holds at the local-nonglobal minimizer. In the second part of this paper, we show that the original constrained problem is equivalent to the unconstrained minimization of a piecewise quartic merit function. Using the unconstrained formulation we give, in the nonconvex case, a new second order necessary condition for global minimizers. In the third part of this paper, algorithmic applications of the preceding results are briefly outlined and some preliminary numerical experiments are reported.
Introduction In this paper we study the problem of minimizing a general quadratic function q : subject to an ellipsoidal constraint, that is where H is a symmetric positive definite n \Theta n matrix and a is a positive scalar. The interest in this problem initially arose in the context of trust region methods for solving unconstrained optimization problems. In fact, such methods require at each iteration an approximate solution of Problem (1) where q(x) is a local quadratic model of the objective function over a restricted ellipsoidal region centered around the current iterate. However, recently, it has been shown that problems with the same structure of y This work was partially supported by Agenzia Spaziale Italiana, Roma, Italy z Universit'a di Roma "La Sapienza" - Dipartimento di Informatica e Sistemistica - via Buonarroti, 12 Italy and Gruppo Nazionale per l'Analisi Funzionale e le sue Applicazioni del Consiglio Nazionale delle Ricerche. (1) play an important role not only in the field of unconstrained minimization. In fact, the solution of Problem (1) is at the basis of algorithms for solving general constrained nonlinear problems (e.g. [3, 42, 20, 27]), and integer programming problems (e.g. [21, 41, 22, 31, 19]). Many papers have been devoted to point out the specific features of Problem (1). Among the most important results there are the necessary and sufficient conditions for a point x to be a global minimizer, due to Gay [12] and Sorensen [35], and the characterization and uniqueness of the local-nonglobal minimizer due to Mart'inez [26]. The particular structure of the Problem (1) has led to the development of algorithms for finding a global solution. The first algorithms proposed in literature were those of Gay and Sorensen [12, 35]. Mor'e and Sorensen [29] developed an algorithm that produces an approximate global minimizer in a finite number of steps. More recently, it has been proved that an approximation to the global solution can be computed in polynomial time (see for example [39, 38, 40, 41, 21]). Furthermore, Mor'e [28] has considered a more general case by allowing in Problem (1) a general quadratic constraint and has extended the results of [12, 35, 29]. In spite of all these results, there is still interest in studying Problem (1). In fact, as we mentioned before, there is a growing use of Problem (1) as a tool for tackling large nonlinear programming problems and combinatorial optimization problems. This leads to the necessity of solving more and more efficiently large scale problems of the type (1) and motivates further research on theoretical properties of Problem (1) and on the definition of efficient methods for locating its global minimizers. Recently some interesting algorithms for tackling large scale trust region problems have been proposed in [36, 34, 33]. The basic idea behind these algorithms is to recast the trust region problem in term of a parametrized eigenvalue problem and then to adjust the parameter to find an optimal solution. In this paper we point out further theoretical properties of Problem (1). In par- ticular, our research develops along two lines: the study of some new properties of its Karush-Kuhn-Tucker points and its equivalence to an unconstrained minimization problem. Besides their own theorical interest, these results allows us to define new classes of algorithms for solving large scale case trust region problems. These algorithms use only matrix-vector product and do not require the solution of an eigenvalue problem at each iteration (see [24] for details). The paper is organized as follows. In Section 2 we recall some preliminary results. In Section 3 we show that (i) given a KKT point - x which is not a global minimizer, it is possible to find a new feasible point - x such that the objective function is strictly decreased, i.e. (ii) the strict complementarity condition holds at the local minimizer, hence in the nonconvex case, strict complementarity holds at local and global minimizers. In Section 4 we show that there is a one to one correspondence between KKT points (global minimizers) of Problem (1) and stationary points (global minimizers) of a piecewise quartic merit function Therefore, Problem (1) is equivalent to the unconstrained problem of P over IR n . In Section 5, by exploiting some results of the preceding sections, we give a new second order necessary condition for global minimizers of Problem (1). Finally, in Section 6, we sketch some possible applications of the results of Section 3 and Section 4 for defining new classes of algorithms for solving large scale trust region problems. In the sequel we will use the following notation. Given a vector x 2 IR n , we denote by kxk the ' 2 -norm on IR n . The ' 2 -norm of a n \Theta n matrix Q is defined by 1g. Moreover, we denote by - 1 - 2 - n the eigenvalues of Q. Preliminaries Without loss of generality we can assume that the feasible set F is defined by so that the problem under consideration is and Q is a n \Theta n symmetric matrix and c 2 IR n . In fact, since H is positive definite, we can reduce Problem (1) to the form (2) by employing the transformation we refer to [25] for the direct treatment of Problem (1)). The Lagrangian function associated with Problem (2) is the function A Karush-Kuhn-Tucker point for Problem (2) is a pair (-x; -) 2 IR n \Theta IR such that: Furthermore, we say that strict complementarity holds at a KKT pair (-x; - for It is well known that it is possible to completely characterize the global solutions of Problem (2) without requiring any convexity assumption on the objective function. In fact, the following result due to Gay [12] and Sorensen [35] holds (see also Vavasis Proposition 2.1 A point x such that kx k 2 - a 2 is a global solution of Problem (2), if and only if there exists a unique - - 0 such that the pair (x ; - ) satisfies the KKT conditions and the matrix (Q positive semidefinite. If (Q positive definite then Problem (2) has a unique global solution. Moreover, Mart'inez [26] gave the following characterization of the local-nonglobal minimizer for Problem (2). Proposition 2.2 There exists at most one local-nonglobal minimizer - x of Problem (2). Moreover we have and the KKT necessary conditions holds with - 2 are the two smallest eigenvalues of Q. 3 Further features of KKT points In this section we give some new properties of the KKT points for Problem (2). Our interest in the characterization of KKT points is due to the fact that, in general, algorithms for the solution of constrained problems, converge towards KKT points. We show that the number of different values that the objective function can take at KKT points is bounded from above by the number of negative eigenvalues of the matrix Q. First we state a preliminary result that extends one given in [38]. Lemma 3.1 Let (b x; -) and (-x; -) be KKT pairs for Problem (2) with the same KKT multiplier. Then q(b Proof We observe that the function q(x) can be rewritten at every KKT pair (x; -) as follows By using the KKT conditions we obtain Now, we can state the following proposition whose proof follows from a result of Forsythe and Golub [11] on the number of stationary values of a second degree polynomial on the unit sphere. For sake of completeness we give a sketch of the proof. Proposition 3.2 There exist at most minf2m points with distinct multipliers -, where m is the number of negative distinct eigenvalues of Q. Proof First we observe that at every KKT point (x; -) such that kxk 2 ! a 2 the value of the objective function q is constant. This easily follows from Lemma 3.1 by observing that all these pairs are characterized by the fact that Now, we consider the values of the function q(x) at all the points such that kxk there exists an orthogonal matrix V such that V T diag are the eigenvalues of Q. By considering the transformation ff we can write the first equation of the KKT condition (4) (premultiplied by V T ) as follows: diag with recalling that kxk we have that the KKT multipliers must satisfy the system where The function g(-) has poles at \Gamma2- i and it is convex on the subintervals \Gamma2- Thus there exists at most 2 roots of in each subinterval. Moreover, since a 2 has one root in each exteme subinterval. If all eigenvalues - i are positive there exists at most one non negative root; if all the eigenvalues are negative there are at most 2n non negative roots; in the case of m ! n negative eigenvalues, there are at most 2m negative roots. Hence the number of the solutions of system (6) is at most minf2m Finally, by summarizing the two cases, we can conclude that the number of distinct KKT multipliers is bounded above by minf2m Recalling Lemma 3.1, we get directly the following corollary. Corollary 3.3 The number of distinct values of the objective function q(x) at KKT points is bounded from above by minf2m 1g. Now we can state the main result of this section. In particular, we show that the peculiarity of Problem (2) can be exploited to escape from the KKT points that are not global solutions in the sense that, whenever we have a KKT point - x, either - x is a global minimizer of Problem (2), or it is possible to compute the expression of a feasible point with a strictly lower value of the objective function. This results is very appealing from a computational point of view, as discussed in Section 6. Proposition 3.4 Let (-x; -) be a KKT point for Problem (2). Let us define the point x in the following way (a) If c T - x: (b) If c T - x - 0 and a vector z 2 IR n such that z T (Q with z: with z -I)z Then we have q(-x) ! q(-x) and k-xk 2 - a 2 . Proof In case (a), the point - x is still feasible and Now consider case (b). In case (i) we have by the KKT conditions that - hence we have that z is a vector of negative curvature for q(x). Therefore, for every satisfies the inequality In particular, if we take ff - ~ ff with ~ we have that k-xk 2 - a 2 . let us consider case (ii). Let - x be the vector defined as follows z and consider the quadratic function We note that k-xk and that z is a negative curvature direction for the quadratic function L(x; -). By simple calculation, taking into account that (Q get -I)z and hence L(-x; -) ! L(-x; - -). Hence, recalling the expression (8) we can write Hence we get the result for case (ii). Let us consider the case (iii). Let us define the vector - 0: We can find a value for ff such that - s is a negative curvature direction for L(x; -) and - s T - x 6= 0, so that we can proceed as in case (ii). In fact, by simple calculation we have: and by using the KKT conditions xj: By solving the quadratic equation with respect to ff we get that - s T ff where \Gammac Hence, by proceeding as in case (ii), we get the result by introducing the point with ff ? - ff. Remark We note that the local-nonglobal minimizer can corresponds either to the case (a) with k-xk or to the case (b)(ii). The preceding proposition shows that, if the KKT point - x is not a global minimizer, it is possible to determine a feasible point - x such that q(-x) ! q(-x) by computing at most a direction z such that z T (Q The existence of such a direction is guaranteed by Proposition 2.1 and from the numerical point of view, its computation is not an expensive task. In fact, we can obtain such a direction by using, for example, the Bunch-Parlett decomposition [2, 30], modified Cholesky factorizations [10] or, for large scale problem, methods based on Lanczos algorithms [4]. Now, as last result of this section, we investigate a regularity property of the local and global minimizers. In particular, we focus our attention on the strict complementarity property, that, roughly speaking, indicates that these points are "really constrained". Also this property can be interesting from an algorithmic point of view. Proposition 3.5 At the local-nonglobal minimizer for Problem (2) the strict complementarity condition holds. Proof Since - x is a local minimizer the KKT conditions (4) hold. Moreover the second order necessary conditions require that z By Proposition 2.2 we have that - there is no local-nonglobal minimizer. Furthermore, if necessarily restrict ourselves to the case since in this case 0 2 us assume by contradiction that - (4) and Proposition 2.2 we have that z x is not a global minimizer, by Proposition 2.1 there exists a direction y such that y T Qy ! 0 and from the second order necessary conditions y T - x 6= 0. We assume, without loss of generality, that y T - us consider the point ffy with We prove that for sufficiently small values of ff the point x(ff) is feasible and produces a smaller value of the objective function, thus contradicting the assumption of local optimality. In fact, we have and hence for we obtain kx(ff)k 2 ! a 2 . Moreover, By this proposition and by Proposition 2.1 we directly obtain the following result. Proposition 3.6 In the nonconvex case at every local or global minimizer the strict complementarity holds. Unconstrained formulation In this section, we show that Problem (2) is equivalent to an unconstrained minimization problem of a piecewise quartic merit function. A general constrained optimization problem can be transformed into an unconstrained problem by defining a continuously differentiable exact penalty function by following, for example, the approach proposed in [6, 7]. However, in the special case of minimization of a quadratic function with box constraints, it has been shown in [16] and [23] that it is possible to define simpler penalty functions by exploiting the particular structure of the problem. In the same spirit of these papers we show that also for Problem (2) it is possible to construct a particular continuously differentiable penalty function. This new penalty function takes full advantage of the peculiarities of the trust region problem and enjoys distinguishing features that make its unconstrained minimization significantly simpler in comparison with the unconstrained minimization of the penalty functions proposed in [6, 7]. The main properties of the penalty function proposed in this section are: ffl it is globally exact according to the definition of [7]; ffl it does not require any shifted barrier term hence it is defined on the whole space; ffl it has a very simple expression (it is piecewise quartic); ffl it is known, a priori, for which values of the penalty parameter the correspondence between the constrained problem and the unconstrained one holds. As a first step for the definition of the exact penalty function, we recall the Hestenes- Powell-Rockafellar augmented Lagrangian function [32, 18] L a (x; -; "" ae is a given positive parameter. Now, according to the classical approach, we replace the multiplier vector - in the function L a (x; -; ") with a multiplier function which yields an estimate of the multiplier vector associated to Problem (2) as a function of the variables x. In the literature different multiplier functions have been proposed (see e.g. [9, 13, 6, 7, 23]). However, all the expression of the multiplier functions given in [9, 13, 6, 7] are not defined in the origin of the space. Here we define a new simpler multiplier function that is defined on the whole space IR n whose expression is the following Its properties are summarized in the following proposition. Proposition 4.1 (i) -(x) is continuosly differentiable with gradient (ii) If (-x; -) is a KKT point for Problem (2) then we have (iii) For every x 2 IR n we have x Proof Part (i) easily follows from the definition of the multiplier function (10). As regards part (ii), from (4) we have that a pair (-x; - It is easy to see that if k-xk corresponds exactly to the definition of the multiplier function (10). Otherwise, if k-xk 2 ! a 2 , (4) imply that hence by comparing (11) and (10) that Now let us consider part (iii). By simple calculations we have a 2 On the basis of the previous considerations we can replace the vector - in the function L a with the multiplier function -(x). Furthermore, as regards the penalty parameter ", we can select, a priori, an interval of suitable values depending on the problem data Q; c; a. Therefore, we are now ready to define our merit function P L a (x; -(x); "(Q; c; a)), that is ae where -(x) is the quadratic function given by (10) and " is any parameter that satisfies the following inequality: a First, we show some immediate properties of the merit function P . Proposition 4.2 (i) P (x) is continuosly differentiable with gradient ae (ii) P (x) is twice continuosly differentiable except at points where" (iii) P (x) is twice continuosly differentiable in a neighborhood of a KKT point - x where strict complementarity holds; (iv) for every x such that kxk 2 - a 2 we have that P (x) - q(x); (v) the penalty function P (x) is coercive and hence it admits a global minimizer. Proof Part (i), (ii) and (iii) directly follows from the expression of the penalty function P . As regards Part (iv) it follows from a classical results on penalty functions (see Theorem 2 of [7]). As regards part (v), we want to show that as kxk ! 1 the function P (x) goes to infinity. First, we observe that" hence for sufficiently large values of kxk the leading term of the preceding inequality is strictly positive since, recalling that " satisfies (13), we have that " - sufficiently large values of kxk, we can assume that ae oe By simple calculation, the expression of the penalty function becomes in this case: and the following inequalities hold: As " satisfies (13), we have that " - and hence we get lim 1. The existence of the global minimizer immediately follows from the continuity of P and the compactness of its level sets. Now, we state the first result about the exactness properties of the penalty function P . Since its proof is technical and lenghty we report it in the Appendix. Proposition 4.3 A point - is a stationary point of P (x) if and only if (-x; -x)) is a KKT pair for Problem (2). Furthermore, at this point we have P Now we prove that there is a one to one correspondence betweeen global minimizers of Problem (2) and global minimizers of the penalty function P . Proposition 4.4 Every global minimizer of Problem (2) is a global minimizer of P (x) and conversely. Proof By Proposition 4.3, the penalty function P admits a global minimizer - which is obviously a stationary point of P and hence by the preceding proposition we have On the other hand, if x is a global minimizer of Problem (2), it is also a KKT point and hence the preceding proposition implies again that P (x proceed by contradiction. Assume that a global minimizer - x of P (x) is not a global minimizer of Problem (2), then there should exists a point x , global minimizer of Problem (2), such that that contradicts the assumption that - x is a global minimizer of P . The converse is true by analogous considerations. In order to complete the correspondence between the solution of Problem (2) and the unconstrained minimization of the penalty function P we prove the following result that considers the corrispondence between local minimizers. Proposition 4.5 The function P (x) admits at most a local-nonglobal minimizer - x which is a local minimizer of Problem (2) and -x) is the associated KKT multiplier. Proof We first prove that if - x is a local minimizer of P (x) then the pair (-x; -x)) satisfies the KKT conditions for Problem (2). Moreover, by Proposition 4.3, we have that x is a local minimizer of P , there exists a neighbourhood of - x such that Thus, by using (iv) of Proposition 4.2, we obtain and hence - x is a local minimizer for Problem (2). The proof can be easily completed by recalling Proposition 2.2. 5 A new second order optimality condition The results given in Section 3 and Section 4 can be combined to state new theoretical properties of Problem (2). In this section we introduce a new second order necessary optimality condition for Problem (2) for the nonconvex case that follows from the unconstrained formulation. Proposition 5.1 Assume that Q is not positive semidefinite, if - x is a global (local) minimizer of Problem (2) then there exists a unique - such that the KKT conditions hold and a 2 a 2 is positive semidefinite for every " satisfying (13). Proof If - x is a global minimizer of Problem (2), by Proposition 3.6, we have that . Then, there exists a neighborhood \Omega\Gamma - x) of - x such that" Thus, by (ii) of Proposition 4.2, the function P (x) is twice continuously differentiable in \Omega\Gamma - x). and the Hessian matrix evaluated at - x is given by: a 2 By Proposition 4.4, - x is also a global minimizer of P (x) and therefore - x satisfies the second order necessary conditions to be a global unconstrained minimizer of P , that is positive semidefinite. Then the result follows. Recalling point (a) of Proposition 3.4, we have that in a global minimizer - x, it results Hence, the matrixa 2 a 2 is not necessarily positive semidefinite. A similar second order necessary condition was given in [1], where it has been proved, without requiring any assumptions on the matrix Q, that if the global minimum is on the boundary, the matrix a 2 is positive semidefinite where again the matrix 1 a 2 x)-x-x T is not necessarily positive semidefinite. 6 Algorithmic application Besides their own theorical interest, the results of the preceding sections are appealing also from a computational point of view. Although the study of a numerical algorithm for the solution of Problem (2) is out of the aim of this paper, in this section we give a hint of possible algorithmic applications of the results of Section 3 and Section 4. We recall that Proposition 3.4 ensures that given a KKT point which is not a global solution for Problem (2), it is possible to find a new feasible point with a lower value of the objective function and that Proposition 3.2 states that the number of KKT points with different value of the objective function is finite. These results indicate a new possibility to tackle large scale trust region problems. In fact they show that a global minimum point of Problem (2) could be efficiently computed by applying a finite number of times a constrained optimization algorithm that presents the following features: (i) given a feasible starting point, it is able to locate a KKT point with a lower value of the objective function; (ii) it presents a "good" (at least superlinearly) rate of convergence; (iii) it does not require an heavy computational burden. A possibility to ensure property (i) is to use any feasible method that forces the decrease of the objective function, following, for example, the approach of [37, 17]. Another possibility is to exploit the unconstrained reformulation of Problem (2) described in Section 4 which allows us to use any unconstrained method for the minimization of the penalty function P . In fact, starting from a point x 0 , any of this algorithm obtains a stationary point - x for P such that Then, Proposition 4.3 ensures that - x is a KKT point of Problem (2) and that P q(-x). On the other hand, if x 0 is a feasible point, part (iv) of Proposition 4.2 yields that In conclusion by using an unconstrained algorithm, we get a KKT point of Problem (2) with a value of the objective function lower than the value at the starting point. Furthermore, the possibility of transforming the trust region problem into an unconstrained one, seems to be quite appealing also as regards properties (ii) and (iii). In fact Proposition 3.6 and (iii) of Proposition 4.2 guarantees that, in the nonconvex case, the penalty function is twice continuosly differentiable in every local and global minimizer of the problem. Therefore, in this case, any unconstrained Truncated Newton algorithm (see for example [5, 37, 15]) can be easily adapted in order to define globally convergent methods which show a superlinear rate of convergence in a neighbourhood of every global or local minimizer. Nevertheless, we can define algorithm with superlinear rate of convergence without requiring that the penalty function is twice continuosly differentiable in the neighbourhood of the points of interest, that is without requiring the strict complementarity in these points. In fact we can drawn our inspiration from the results in [8]. In particular, we can define a search direction d k as follows: !/ z k The results of [8] ensure that the algorithm x locally superlinearly convergent without requiring the strict complementarity. Following the approach of truncated Newton method (see for example [5, 15]), in [24] it is shown that an approximate solution ~ d k of (15)(16) is able to preserve the local superlinear rate of convergence of the algorithmic scheme. Furthermore it is also proved that this direction ~ suitable descent conditions with respect to the penalty function P . This strict connection between the direction ~ d k and the penalty function P (x) allow us to define globally and superlinearly convergent algorithms of the type ~ where ff k can be determined by every stabilization technique and ~ d k is computed by using a conjugate gradient based iterative method for solving approximately the linear system (15)(16) . The paper [24] is devoted to a complete description of this approach with the analysis of its theoretical properties and to the definition of an efficient algorithm. Here, in order to have only a preliminary idea of the viability of this unconstrained approach for solving Problem (2), we have performed some numerical experiments with a rough implementation of algorithm (17) where ff k is determined by the line-search technique of [14] and ~ d k is computed by a conjugate gradient algorithm similar to the one proposed in [5]. We coded the algorithm in MATLAB and run the experiments on a IBM/RISC 6000. We run two sets of problems randomly generated that we take from the collection of [34]. We solved ten related problems for each of the two classes both with the easy and the hard case. According to [34], the hard case occurs when the vector c is orthogonal to the subspace generated by the smallest eigenvalue of the matrix Q. In Table 1 we report the results in terms of average number of iterations for problems with increasing dimension We run also a set of near hard-case problems (with that is with c nearly orthogonal to the subspace of the smallest eigenvalue of Q. The results are FIRST SET SECOND SET 100 11.3 21.9 10.7 25.6 Table 1: Average number of iterations NEAR HARD CASE mult. - min Table 2: Average number of iterations reported in Table 2. We tested the invariance with respect to the multiplicity of the smallest eigenvalue (mult. of - The results obtained are encouraging. The number of iteration is almost constant when the dimension increases. This feature is appealing when solving large scale problems taking into account that, at each iteration, the main effort is due to the approximate solution of a linear system of dimension n or n \Gamma 1 that requires only matrix-vector products. Furthermore the efficiency of the algorithm seems not to be seriously affected by the occurrence of the hard case, while it is completely insensible to the near-hard case. Of course, even if no final conclusion can be drawn by these limited numerical exper- iments, the results obtained encourage further research in defining new algorithms for solving large scale trust region problems which use the results described in this paper. In particular, as we said before, the possibility of defining efficient algorithms based on the unconstrained reformulation is investigated in [24]. Acknowledgments We wish to thank S. Santos, D. Sorensen, F. Rendl and H. Wolkowiz, for providing us their Matlab codes and test problems. Moreover we thank the anonymous referees for their helpful suggestions which led to improve the paper. --R New optimality conditions and algorithms for homogeneous and polynomial optimization over spheres. Direct methods for solving symmetric indefinite systems of linear equations. Computing a trust region step for a penalty function. Lanczos Algorithms for Large Symmetric Eigenvalue Computation. An exact penalty method with global convergence properties for nonlinear programming problems. Exact penalty functions in constrained optimization. Quadratically and superlinear convergent algorithms for the solution of inequality constrained optimization problems. A class of methods for nonlinear programming with termination and convergence properties. Computing modified Newton directions using a partial Cholesky factorization. On the stationary values of a second-degree polynomial on the unit sphere Computing optimal locally constrained steps. A multiplier method with automatic limitation of penalty growth. A nonmonotone line search technique for Newton's method. A truncated Newton method with nonmonotone linesearch for unconstrained optimization. A differentiable exact penalty function for bound constrained quadratic programming problems. On the solution of a two ball trust region subproblem. Multiplier and gradient methods. A continuous approach to compute upper bounds in quadratic maximization problems with integer constraints. Fast algorithms for convex quadratic programming and multicommodity flows. An interior-point approach to NP-complete problems An interior point algorithm to solve computationally difficult set covering problems. A differentiable piecewise quadratic exact penalty functions for quadratic programs with simple bound constraints. "La Sapienza" "La Sapienza" Local minimizers of quadratic functions on Euclidean balls and spheres. Trust region algorithms on arbitrary domains. Generalization of the trust region problem. Computing a trust region step. On the use of directions of negative curvature in a modified Newton method. Algorithms for the solution of quadratic knapsack problems. A method for nonlinear constraints in minimization problem. A semidefinite framework to trust region subproblems with applications to large scale minimization. A new matrix-free for the large scale trust region subproblem Newton's method with a model trust region modification. Minimization of a large scale quadratic function subject to an ellipsoidal constraint. Towards an efficient sparsity exploiting Newton method for minimiza- tion Nonlinear Optimization. Proving polynomial-time for sphere-constrained quadratic programming A new complexity result on minimization of a quadratic function with a sphere constraint. On affine scaling algorithms for nonconvex quadratic programming. An extension of Karmarkar's projective algorithm for convex quadratic programming. --TR --CTR Pasquale L. De Angelis , Immanuel M. Bomze , Gerardo Toraldo, Ellipsoidal Approach to Box-Constrained Quadratic Problems, Journal of Global Optimization, v.28 n.1, p.1-15, January 2004 G. Birgin , Jos Mario Martnez , Marcos Raydan, Algorithm 813: SPG---Software for Convex-Constrained Optimization, ACM Transactions on Mathematical Software (TOMS), v.27 n.3, p.340-349, September 2001 Immanuel M. Bomze , Laura Palagi, Quartic Formulation of Standard Quadratic Optimization Problems, Journal of Global Optimization, v.32 n.2, p.181-205, June 2005 Le Thi Hoai An , Pham Dinh Tao, A Branch and Bound Method via d.c. Optimization Algorithms andEllipsoidal Technique for Box Constrained Nonconvex Quadratic Problems, Journal of Global Optimization, v.13 n.2, p.171-206, September 1998
merit function;quadratic function;ell2-norm constraint
589162
Superlinear Convergence of a Symmetric Primal-Dual Path Following Algorithm for Semidefinite Programming.
This paper establishes the superlinear convergence of a symmetric primal-dual path following algorithm for semidefinite programming (SDP) under the assumptions that the semidefinite program has a strictly complementary primal-dual optimal solution and that the size of the central path neighborhood tends to zero. The interior point algorithm considered here closely resembles the Mizuno--Todd--Ye predictor-corrector method for linear programming which is known to be quadratically convergent. It is shown that when the iterates are well centered, the duality gap is reduced superlinearly after each predictor step. Indeed, if each predictor step is succeeded by r consecutive corrector steps then the predictor reduces the duality gap superlinearly with order 2 The proof relies on a careful analysis of the central path for SDP. It is shown that under the strict complementarity assumption, the primal-dual central path converges to the analytic center of the primal-dual optimal solution set, and the distance from any point on the central path to this analytic center is bounded by the duality gap.
Introduction Recently, there have been many interior point algorithms developed for semidefinite programming (SDP), see for example [1, 2, 5, 9, 11, 13, 17]. These algorithms differ in their choices of scaling matrix, the size of the central path neighborhoods, and stepsize rules, among others. In particular, the algorithms of Kojima-Shida-Hara [5] and Nesterov-Todd [11] are based on the primal-dual scaling and they both can be viewed as extensions of the predictor-corrector method for linear programming [8]. It has been shown [4, 6, 11, 13, 17] that these algorithms for SDP retain many important properties of the interior point algorithms for linear programming including polynomial complexity. For an overview of SDP and its applications, we refer to Vanderberghe and Boyd [15]. However, there exists considerable difficulty in extending one key property of the predictor-corrector method for linear programming to the interior point algorithms for SDP. This is the property of quadratic convergence of the duality gap (see [16] for a proof of the LCP case). In some sense, the need for superlinear convergence in solving SDP is more pronounced than that for the linear programming case. This is because for SDP there cannot exist any finite termination procedures as in the case of linear programming. Indeed, the recent papers of Kojima-Shida-Shidoh [4] and Potra-Sheng [12] are both focused on the issue of superlinear convergence for solving SDP. In particular, the latter reference provided a sufficient condition for the superlinear convergence of an infeasible path following algorithm, while the former reference [4] established the superlinear convergence of their algorithm [5] under certain key assumptions. These assumptions are: (1) SDP is nondegenerate in the sense that the Jacobian matrix of its KKT system is nonsingular; (2) SDP has a strictly complementary optimal solution; (3) the iterates converge tangentially to the central path in the sense that the size of the central path neighborhood in which the iterates reside must tend to zero. Among these three assumptions for superlinear convergence, (2) is inevitable since it is needed even in the case of LCP (see [16]). Assumption (3) is needed to ensure the duality gap is reduced superlinearly after each predictor step for all points in the central path neighborhood. In the reference [4], an example was given which showed that, without the tangential convergence assumption, the duality gap is reduced only linearly after one predictor step for certain points in the central path neighborhood. Our goal in this paper is to establish the superlinear convergence of a symmetric path following algorithm for SDP under the only assumptions of (2) and (3) (i.e., without the nondegeneracy assumption). In particular, we consider the primal-dual path following algorithm of Nesterov- Todd [11] (later discovered independently by Sturm and Zhang [13] using a V -space notion). In this paper we adopt the framework of [13] since it greatly facilitates the subsequent analysis. We show that this symmetric primal-dual path following algorithm has an order of convergence that is asymptotically quadratic (i.e., sub-quadratic). Indeed, for any given constant positive integer r, the algorithm can be set so that the duality gap decreases superlinearly with order 2 1+2 \Gamma2r after one predictor (affine scaling) step followed by (at most) r corrector steps. The cornerstone in our bid to establish this superlinear convergence result is a bound on the distance from any point on the central path to the optimal solution set (see Section 3). Specifically, it is shown that, under the strict complementarity assumption, the primal-dual central path converges to the analytic center of the optimal solution set, and that the distance to this analytic center from any point on the central path can be bounded above by the duality gap. These properties of the central path are algorithm-independent and are likely to be useful in the analysis of other interior point algorithms for SDP. The organization of this paper is as follows. At the end of this section, we describe some basic notation to be used in this paper. In Section 2, we will discuss some fundamental background notions, and we will make two assumptions concerning the solution set of the SDP. In Section 3 we will analyze the limiting behavior of the primal-dual central path. In Section 4, the notion of V -space for SDP is reviewed and a path following algorithm in the spirit of [13] is introduced. The superlinear convergence of this algorithm is established in Section 5. Finally, some concluding remarks are given in Section 6. Notation. The space of symmetric n \Theta n matrices will be denoted S. Given X and Y in ! n\Thetan , the standard inner product is defined by where tr (\Delta) denotes the trace of a matrix. The notation X ? Y denotes orthogonality in the sense that The Euclidean norm and its associated operator norm, viz. the spectral norm, are both denoted by k\Deltak. The Frobenius norm of X is kXk positive (semi-) definite, we write (X - The cone of positive semi-definite matrices is denoted by S+ and the cone of positive definite matrices is S++ . The identity matrix is denoted by I. We use the standard "big O" and "small o" notation. In particular, w(-) = O(-) with - ? 0 means that there is a positive constant \Gamma, possibly dependent on problem data but independent of -, such that w(- \Gamma- for all -; means that lim -!0 O(w(-)). For a positive definite matrix, we use "O" and "\Theta" to denote the order of all its eigenvalues. Hence, for W (-) 2 S++ , the notation W signifies the existence of \Gamma ? 0 such Problem formulation A semidefinite programming (SDP) problem is given as subject to A (i) ffl . The decision variable is S. The corresponding dual program can be formulated as subject to Denote the feasible sets of (P) and (D) by F P and FD respectively, i.e. and We make the following assumptions throughout this paper. Assumption 1 There exist positive definite solutions X 2 F P and Z 2 FD for (P) and (D) respectively. Assumption 2 There exists a pair of strictly complementary primal-dual optimal solutions for (P) and (D). Specifically, there exists (X ; Z FD such that 0, we can diagonalize X and Z simultaneously. Therefore, by applying an orthonormal transformation to the problem data if necessary, we can assume without loss of generality that X , Z are both diagonal and of the form positive scalars Here the subscripts B and N signify the "basic" and "nonbasic" subspaces (following the terminology of linear programming). Throughout this paper, the decomposition of any n \Theta n matrix X is always made with respect to the above partition B and N . In fact, we shall adhere to the following notation throughout: U XN so XU will always denote the off-diagonal block of X with size K \Theta (n \Gamma K), etc. Notice that X 2 F P is an optimal solution to (P) if and only if XZ Hence, by Assumption 2, the primal optimal solution set can be written as F Analogously, the dual optimal solution set is given by F Given - 2 !++ , the pair (X; Z) 2 F P \Theta FD is said to be the -center (X(-); Z(-)) if and only if We refer to [5, 14] for a proof of the existence and uniqueness of -centers. The central path of the problem (P) is the curve To be consistent with the above definition of the central path, we define the analytic center of F P as the unique solution X a of the system X a P and ZB - 0: In a similar fashion, we define the analytic center of F D as the unique solution Z a of the system XNZ a A (i) 3 Properties of the central path The notion of central path plays a fundamental role in the development of interior point methods for linear programming. In this section, we shall study the analytic properties of the central path in the context of semidefinite programming. These properties will be used in Section 5 where we perform convergence analysis of a predictor-corrector algorithm for SDP. For linear programming (i.e., A (i) 's and C are diagonal), it is known that the central path curve converges: (X(-); being the analytic center of the primal and dual optimal solution sets F P and F respectively ([7]). It is also known for linear programming that the central path does not approach (X a ; Z a ) tangentially to the optimal solution set, viz. it is shown in [10] that In the following we shall extend these result to the semidefinite programs (P) and (D). The following lemma shows that the set is bounded. Lemma 3.1 For any - ? 0 there holds Proof. We have where we used the property (X(-) in the second equality. Since X(1) - 0 and Z(1) - 0, we have Q.E.D. It follows from Lemma 3.1 that the central path has a limit point. We will now show that any limit point of the central path f(X(-); Z(-))g is a strictly complementary optimal primal-dual Lemma 3.2 For any - 2 (0; 1) there holds Hence, any limit point of f(X(-); Z(-))g as - ! 0 is a pair of strictly complementary primal-dual optimal solutions of (P) and (D). Proof. Let 1. For notational convenience, we will use X and Z to denote the matrices X(-) and Z(-). Let (X ; Z ) be the pair of strictly complementary primal-dual optimal solutions postulated by Assumption 2. Since A (i) ffl Therefore, we have where the last step follows from (2.1). Since (by the positive semidefiniteness of X and Z), we obtain From Now consider the identities log det log det U log det XB log det log det ZN log det By the estimates (3.1) and using Lemma 3.1, we see that Therefore each of the four logarithm terms in the preceding equation are bounded from above as these four terms sum to zero, we must have Together with (3.1), this implies This completes the proof of the lemma. Q.E.D. Lemma 3.2 provides a precise result on the order of the eigenvalues of XB (-); XN (-); ZB (-) and ZN (-). We will now prove a preliminary result on the order of the off-diagonal blocks XU (-) and ZU (-). Lemma 3.3 For - 2 (0; 1), there holds \GammaX U (-) ffl ZU Proof. By the central path definition, we have Expanding the right-hand side and comparing the upper-right corner of the above identity, we have or equivalently, Using XB Lemma 3.2), this implies that This proves the first part of the lemma. We now prove (3.2). Let f(X(- k ); Z(- k :::g be an arbitrary convergent sequence of the central path with - k ! 0. By Lemma 3.2, the limit of this sequence satisfies strict comple- mentarity. Let (X ; Z ) denote this limit point so that As before, we assume without loss of generality that X and Z are diagonal. In addition, since (3.2) holds trivially when kXU (- k First, we divide both sides of (3.3) by kXU (- k )k and let k !1 to obtain U Z U and Z 1 U are defined by U := lim U := lim (If the above limits do not exist, then we define X 1 U and Z 1 U to be any two limit points of the corresponding sequences.) Since X B and Z are both positive diagonal matrices, it follows that the nonzero entries of the matrices X 1 U must have opposite signs. By kX 1 that This establishes (3.2) along the sequence f(X(- k ); Z(- k :::g. Since this sequence is arbitrary, we see (3.2) holds. It remains to establish the last part of the lemma. Once again, we consider an arbitrary convergent sequence f(X(- k ); Z(- k on the central path with - k ! 0; we continue to use the same notation X , Z , X 1 U defined above. Since kZU (- k only need to show kXU (- k . Assume this is not the case. Using Lemma 3.2 and passing onto a subsequence if necessary, we have kXU (- k Dividing both sides of this equation by kXU (- k )k 2 and taking limit yields U Therefore, the limit in the preceding equation equals zero, implying But this contradicts (3.5), so we must have The proof is complete. Q.E.D. We now use Lemma 3.2 and Lemma 3.3 to prove that the central path f(X(-); Z(- ? 0g converges to (X a ; Z a ), and to estimate the rate at which it converges to this limit. Lemma 3.4 The primal-dual central path f(X(-); Z(- ? 0g converges to the analytic centers (X a ; Z a ) of F P and F D respectively. Moreover, if we let ffl(-) := kXU (-)k then Proof. Suppose 1. By expanding X(-)Z(-I and comparing the upper-left block, we obtain Pre-multiplying both sides with Let J be an index set of minimal cardinality such that As Z it follows from the dual feasibility and (3.6) that- Now consider the following nonlinear system of equations: A (i) By (2.3), we know that X a B is a solution of (3.8) for some - a . Using the linear independence of the matrices A (i) using the fact that X a B is positive definite, it can be checked that the Jacobian (with respect to the variables XB and - i , of the nonlinear system (3.8) is nonsingular at the solution X a Hence we can apply the classical inverse function theorem to the above nonlinear system at the point: By (3.7) we have and from X(-) 2 F P we obtain Combining this with (3.9) and (3.10) yields It can be shown with an analogous argument that The proof is complete. Q.E.D. Lemma 3.4 only provides a rough sketch of the convergence behavior of the central path as Our goal is to characterize this convergence behavior more precisely. Theorem 3.1 Let - 2 (0; 1). There holds and Proof. The estimate (3.11) is already known from Lemma 3.2, so we only need to prove (3.12). By Lemma 3.3 and Lemma 3.4, it is sufficient to show that Suppose to the contrary that there exists a sequence with kXU (- k )k ? 0 for all k and lim 0: To simplify notations, we introduce (By virtue of Lemma 3.4, we can assume the above limit exists because otherwise we can always pass onto a convergent subsequence.) From Lemma 3.3 it follows that lim Since for each Z 2 FD we have it follows We know from Lemma 3.2 that ZB (- k so that the above relation implies lim 0: Analogously, it can be shown that lim 0: (3:15) As (X(- k we have from (3.14) and (3.15) that which clearly contradicts (3.2). The proof is complete. Q.E.D. Theorem 3.1 characterizes completely the limiting behavior of the primal-dual central path as We point out that this limiting behavior was well understood in the context of linear programming and the monotone horizontal linear complementarity problem, see Megiddo [7] and Monteiro and Tsuchiya [10] respectively. Notice that under a Nondegeneracy Assumption (i.e., the Jacobian of the nonlinear system (2.2) is nonsingular at (X a ; Z a )), the estimates (3.12) follow immediately from the application of the classical inverse function theorem. Thus, the real contribution of Theorem 3.1 lies in establishing these estimates in the absence of the nondegeneracy assumption. It is known that in the case of linear programming the proof of quadratic convergence of predictor-corrector interior point algorithms required an error bound result of Hoffman. This error bound states that the distance from any vector x 2 ! n to a polyhedral set P ag can be bounded in terms of the "amount of constraint violation" at x, namely denotes the positive part of a vector. More precisely, Hoffman's error bound ([3]) states that there exists some constant - ? 0 such that Unfortunately, this error bound no longer holds for linear systems over the cone of positive semidefinite matrices (see the example below). In fact, much of the difficulty in the local analysis of interior point algorithms for SDP can be attributed to this lack of Hoffman's error bound result (see the analysis of [4, 12]). Specifically, without such error bound result, it is difficult to estimate the distance from the current iterates to the optimal solution set. In essence, what we have established in Theorem 3.1 is an error bound result along the central path. In other words, although Hoffman type error bound cannot hold over the entire feasible set of (P), it nevertheless still holds true on the restricted region "near the central path". One consequence of this restriction to the central path is that we will need to require the iterates stay "sufficiently close" to the central path in order to establish the superlinear convergence of the algorithm. Such a requirement on the iterates was called "tangential convergence to the optimal solution set" by Kojima et. al. [4]. Notice that the analysis in this reference required the additional nondegeneracy assumption to establish their superlinear convergence result. In contrast, this assumption is no longer needed in our analysis because Theorem 3.1 holds without the nondegeneracy assumption. Example. Consider the following linear system over the cone of positive semidefinite matrices in Clearly, there is exactly one solution X to the above linear system, namely For each ffl ? 0, consider the matrix Clearly, X(ffl) - 0. The amount of constraint violation is equal to ffl 2 . However, the distance \Theta(ffl). Thus, there cannot exist some fixed - ? 0 such that for all ffl ? 0. Instead, we have in this case that is, the error bound holds with an exponent of 1=2. 4 A polynomial predictor-corrector algorithm We begin by summarizing some of the results on V -space path following for SDP that were obtained in [13]. Let (X; Z) 2 F P \Theta FD with X - 0; Z - 0. Then, there exists a unique positive definite matrix S++ such that ([13, Lemma 2.1]) Let L be such that and let V := L T ZL. It follows that The quantity serves as a centrality measure, with - := X ffl Z=n. It is easy to see that the central path is the set of solutions (X; Z) with ffi(X; equivalently, those solutions for which we have In V -space path following, we want to drive the V -iterates towards the origin by Newton's method, in such a way that the iterates reside in a cone around the identity matrix. Before stating the Newton equation, we need to introduce the linear space A(L), and its orthoplement in S A Newton direction for obtaining a (fl-center, for some fl 2 [0; 1], is the solution (\DeltaX; \DeltaZ ) of the following system of linear equations ([13], equation (17)): \DeltaX For we denote the solution of (4.4) by (\DeltaX p ; \DeltaZ p ), the predictor direction. For solution is denoted by (\DeltaX c ; \DeltaZ c ), the corrector direction. If we let then we can rewrite (4.4) as It follows from orthogonality that F F The corrector direction does not change the duality gap, whereas for any t 2 !, see equation (16) of [13]. Algorithm SDP Given positive integer r. REPEAT (main iteration) Compute the largest step t k such that for all there holds and IF ffi(X; Z) - fi k THEN exit loop. Compute UNTIL convergence. Interestingly, each corrector step reduces ffi(\Delta; \Delta) at a quadratic rate as stated in the following result: Lemma 4.1 If ffi(X; Z) - 1 Proof. It follows from Lemma 4.5 in [13] that Hence, the desired result is an immediate consequence of Lemma 4.4 in [13]. Q.E.D. Also, it follows from (4.6), (4.7) and Lemma 4.1 that for any k ? 1 Furthermore, if fi only one corrector step (i.e., needed to recenter the iterate (see [13]). In other words, the iterations of Algorithm SDP are identical to that of the primal-dual predictor-corrector algorithm of [13], for all k with We can therefore conclude from Theorem 5.2 in [13] that the algorithm yields - k - fflfor - O( log(- 0 =ffl)). Thus, we have the following polynomial complexity result. Theorem 4.1 For each generate an iterate (X ffl in at most O( predictor-corrector steps. In addition to having polynomial complexity, Algorithm SDP also possesses a superlinear rate of convergence. We prove this in the next section. 5 Convergence analysis We begin by establishing the global convergence of Algorithm SDP. Notice that Algorithm SDP chooses the predictor step length t k to be the largest step such that for all there holds oe It was shown in [13] (equation (21) therein) that Combining (5.1) and (5.2), we can easily establish the global convergence of Algorithm SDP. Theorem 5.1 There holds lim i.e. Algorithm SDP is globally convergent. Proof. Due to (4.7), - . is a monotone decreasing sequence. Hence, the sequence has a limit. Suppose contrary to the statement of the lemma that Then, we obtain from (4.5), (5.1) and (5.2) that t Together with (4.7) this implies that = \Theta(1), which contradicts (5.3). Q.E.D. Next we proceed to establish the superlinear convergence of Algorithm SDP. In light of (4.7), we only need to show that the predictor step length t k approaches to 1. Hence we are led to bound t k from below. For this purpose, we note from (5.2) that, for t 2 (0; 1), Thus, if we can properly bound , then we will obtain a lower bound on the predictor step length t k . To begin, let us consider L - with Remark that Now define the predictor direction starting from the solution (X(-); Z(-)) on the central path as Z a ) be the analytic center of the optimal solution set in the L -transformed space, Z a := L T - Z a We will show in Lemma 5.1 below that \Delta - is close to the optimal step - We will bound the difference between \Delta - afterwards. Lemma 5.1 There holds X a Z a Proof. Since it follows Z a Z a X a Therefore, the matrix ( Z a ), or equivalently, the matrix is symmetric. By the property of F-norm, we obtain Z a ) where the last step follows from Theorem 3.1. Now since - we have Z a ) Z a As Z a 2 A(L - ); it follows that fl fl X a F Z a F F where last step is due to (5.5). This proves the lemma. Q.E.D. Lemma 5.1 applies only to (\Delta - namely the predictor directions for the points located exactly on the central path. What we need is a similar bound for (\Delta - at points close to the central path). This leads us to bound the difference \Delta - our next goal is to show (Lemma 5.5) that O( We prove this bound by a sequence of lemmas. Let D be given by (4.1) and define so that - Choose L by and notice that indeed LL stipulated by (4.2). Lemma 5.2 Suppose ffi(X; Z) - 1. There holds Proof. Let Clearly, \Delta x (-) and \Delta z (-) are symmetric and \Delta x (-) ? \Delta z (-). Let ae denote the smallest eigenvalue of aeIg: tr where the last step follows from (X(-) Consider tr 0: By (4.3), the matrix V is symmetric positive definite and Diagonalize the symmetric matrix the diagonal entries of QV Q T must be \Theta(1). Therefore, the preceding equation implies that the diagonal matrix must have a nonpositive eigenvalue and that its diagonal entries are of same order of magnitude. In other words, ae - 0 and O(jaej). This further implies By the definition of the central path, we have Now using the fact that the above matrix is symmetric, it follows that and therefore, Using (4.3), we obtain Combining this with (5.6) and using the fact that \Delta x (-) ? \Delta z (-), we have Q.E.D. Lemma 5.3 Suppose ffi(X; Z) - 1=2. Then there holds Proof. Notice that and Now using we have, by pre- and post-multiplying the above two equations with - D \Gamma1=2 and rearranging terms, Together with Lemma 5.2, this implies - The lemma is proved. Q.E.D. Notice that Lemma 5.4 We have Proof. We have Now using Lemma 5.3 and (4.5), we see that It can be shown in an analogous way that Q.E.D. Now we are ready to bound the difference between \Delta - Lemma 5.5 Suppose ffi(X; Z) - 1=2. We have Proof. By definition of the predictor directions, we have and Combining these two relations yields Now using Lemma 5.4 and using the fact that we obtain fl Z p ), the lemma follows from the above relation, after applying Lemma 5.4 once more. Q.E.D. Combining (5.5), Lemma 5.1 and Lemma 5.5 we can now estimate the order of and hence, using (5.4), we can estimate the predictor step length t k . Lemma 5.6 We have Proof. Combining Lemma 5.5 with Lemma 5.1, we have X a Z a so that, using (4.5), fl fl X a Z a Moreover, Z a ) X a Applying (5.5), (5.7), (5.8) and (4.5) to the above relation yields Q.E.D. Theorem 5.2 The iterates (X k ; Z k ) generated by Algorithm SDP converge to (X a ; Z a ) superlinearly with The duality gap - k converges to zero at the same rate. Proof. From (5.4) we see that for any t - 0 satisfying there holds This implies using (4.8) and Lemma 5.6 that so that This shows that the duality gap converges to zero superlinearly with order 2=(1 It remains to prove that the iterates converge to the analytic center with the same order. Notice that However, using the definition of F-norm and applying Lemma 5.3, Recall that L - k definition, so that using Lemma 3.1, Combining (5.9) and (5.10) with Lemma 5.2, we have Hence, we obtain from Theorem 3.1 that Similarly, it can be shown that This shows that the iterates converge to the analytic center R-superlinearly, with the same order as - k converges to zero. Q.E.D. 6 Conclusions We have shown the global and superlinear convergence of the predictor-corrector algorithm SDP, assuming only the existence of a strictly complementary solution pair. The local convergence analysis is based on Theorem 3.1, which states that O(-). By enforcing the iterates "inherit" this property of the central path. For the generalization of the Mizuno-Todd-Ye predictor-corrector algorithm in [13], we do not enforce ffi(X hence we cannot conclude superlinear convergence for it yet. In this respect, it will be interesting to study the asymptotic behavior of the corrector steps. Finally, it is likely that our line of argument can be applied to the infeasible primal-dual path following algorithms of Kojima-Shindoh-Hara [5] and Potra-Sheng [12]. --R "Interior point methods in semidefinite programming with applications to combinatorial optimization problems," "An interior-point method for semidefinite programming," "On approximate solutions of systems of linear inequalities" "Global and local convergence of predictor-correct infeasible-interior-point algorithm for semidefinite programming," "Interior-point methods for the monotone linear complementarity problem in symmetric matrices," "An infeasible start predictor corrector method for semi-definite linear programming," "Pathways to the optimal solution set in linear programming," "On Adaptive-Step Primal-Dual Interior-Point Algorithms for Linear Programming," "Primal-dual path following algorithms for semidefinite programming," "Limiting behavior of the derivatives of certain trajectories associated with a monotone horizontal linear complementarity problem," "Primal-dual interior-point methods for self-scaled cones," "A superlinearly convergent primal-dual infeasible-interior-point algorithm for semidefinite programming," "Symmetric primal-dual path following algorithms for semidefinite programming," "A primal-dual potential reduction method for problems involving matrix inequalities," "Semidefinite programming," "On quadratic and O( p nL) convergence of a predictor-corrector algorithm for LCP," "On extending primal-dual interior-point algorithms from linear programming to semidefinite programming," --TR
path following;central path;superlinear convergence;semidefinite programming
589163
The Sequential Knapsack Polytope.
In this paper we describe the convex hull of all solutions of the integer bounded knapsack problem in the special case when the weights of the items are divisible. The corresponding inequalities are defined via an inductive scheme that can also be used in a more general setting.
Introduction In this paper we deal with the integer bounded knapsack problem a and the numbers a i are divisible, i.e., a i a n. In this case we say that the knapsack problem has the divisibility property. It is also called the sequential knapsack problem (see [1]). Whenever we are given a knapsack problem having the divisibility property, we will assume without loss of generality that a Our main result is the construction of the system of inequalities that describes the convex hull of all solutions in this special case. Since years the knapsack polytope is of particular interest for researchers in polyhedral combinatorics. This is due to several reasons: one is the increasing number of applications like in circuit design, telecommunication, vehicle routing and scheduling that involve the knapsack problem as a subproblem. In order to apply polyhedral methods to such complex problems, a good understanding of the knapsack polytope is important. Secondly, the knapsack problem is the "easiest" case of a number dependent problem. A slight change of the weights of the items might change the inequalities that describe the polyhedron drastically. There- fore, it is important to understand "general principles" according to which valid inequalities are constructed. Examples in this direction are, for instance, Gomory cutting planes [2], covers [12], (1; k)-configurations [8], the concept of lifting [7], the weight reduction principle [10] or inequalities based on the Hilbert basis of a cone of exchange vectors [11]. The knapsack polytope is one of the very interesting and challenging polyhedra for which beautiful results can be discovered. We present an inductive scheme to construct valid inequalities for the knapsack polytope and show, in case that the weights of the items have the divisibility property, that we obtain the complete description of the associated polyhedron. The special case of the knapsack problem with the divisibility property has been studied in the literature by several authors. Hartmann and Olmstead [4] give an O(n log n) algorithm for optimizing a linear objective function whose bottleneck operation is sorting the ratios fl i a i The case of the sequential knapsack problem when s considered by Marcotte [6]. He shows that an optimum solution can be found in linear time and applies his algorithm to the cutting stock problem. Pochet and Wolsey [9] give an explicit description of the knapsack polyhedron with the divisibility property when there are no bounds on the variables. They also refer to applications in local area networking. In Section 2 we present a transformation of any given sequential knapsack problem to a special one such that in terms of feasible solutions and optimization both formulations are equivalent. In Section 3 we outline a decomposition result for all the optimal solutions of such a transformed sequential knapsack problem. Our main result is contained in Section 4. Here we present an inductive scheme to generate valid inequalities for the sequential knapsack problem. Given an objective function, we construct an inequality via this scheme whose induced face contains the set of all optimal solutions. This sufficies to show that our inductive class of inequalities describes the sequential knapsack polyhedron. How inequalities defined via our inductive scheme can be interpreted combinatorially is the issue of Section 5. The discussions end in Section 6 with some extensions. Throughout the paper we use the following notation. For vg. The constraint is called the knapsack inequality. The number a is termed the weight of item i and a 0 2 IN is called the knapsack capacity. We set N := ng and we always assume that 0 ! a 1 - a n - a 0 . An integer vector that satisfies the knapsack constraint and the lower and upper bound constraints is called feasible. We say, F c is a face of some polytope P induced by the valid inequality c T x - fl, flg. Every x 2 F c is also called a root of c T x - fl. The inequalities x are called trivial. For real numbers ng we use the notation - (I) := P In this section we present a transformation of the given sequential knapsack problem to a special sequential knapsack problem that satisfies certain requirements. We show that in terms of polyhedra and in terms of optimization both formulations are equivalent. We start by introducing the notion of blocks. l be a subset of items. B is called a block if, for every holds. Let B be a block. The above definition implies that for every number - 2 there exists a subset W ' B such that . The number uB := a is called the multiplicity of block B. We replace block B by a single item B with weight a i 1 and multiplicity (upper bound) a . The objective function coefficient of B is the number Bm be a partition of N into blocks and denote by fw , c w , uw the weight, objective function coefficient, multiplicity of block Bw , respectively, m. Now consider the knapsack problem where every block is replaced by a single item: z From the construction of the blocks it is clear that (MSKP) is a sequential knapsack problem (MSKP stands for modified sequential knapsack problem). We now show that there is a many to one correspondence between the feasible solutions of the original problem (SKP) and the feasible solutions of its modified version (MSKP). For ease of notation we assume that f 1 - f 2 - fm , and in case holds. By P SKP and PMSKP we denote the convex hull of all feasible vectors of the problem (SKP) and (MSKP), respectively. Let z 2 IR m be a feasible solution of (MSKP), i.e., 0 - z w - uw , z w integer for all m. By Definition 2.1, for every w there exist integers Bw such that P In fact, for all subsets I w of items in Bw with IN, the vector x 2 IR n defined via x otherwise, is feasible for problem (SKP). Conversely, with every vector x 2 IR n that is feasible for problem (SKP) we associate a vector z 2 IR m by setting z w := j2Bw a j x j It follows that an integer vector z with z w feasible for (MSKP) if and only if there exist feasible vectors of (SKP) with the same total weight as z. Now suppose that is a valid inequality for the polytope PMSKP . By setting fi i := b w a i fw if item i belongs to block Bw , the inequality is valid for P SKP . This statement follows from the fact that if x is feasible for (SKP) then defined via z j2Bw a j x j is feasible for (MSKP) and satisfies This shows that valid inequalities for PMSKP can be transformed into valid inequalities for P SKP . In the following we focus on a special partition of the set N into blocks For an item i of (SKP), its gain per unit is defined as fl i a i . Let the different values of gains per unit for all items of (SKP) (clearly, v - n). We partition each set V g := fi a i into blocks ng such that B g j is not a block anymore, for denote the final blocks constructed this way. Each block B i , is called a maximal block and, by definition, all items belonging to the block B i have the same gain per unit. Example 2.2. Consider the instance of the sequential knapsack problem with upper bounds s i on the variables x i as follows: s 1. The set of items is partitioned into the f7g. After transformation we obtain the instance of the sequential knapsack problem: with upper bounds on the variables For a given sequential knapsack problem, the aggregation of items into maximal blocks is unique. If V l is the set of all items in N with gain per unit equal to g, then the unique maximal block containing g and a i l+1 is defined as item in this subset fi can belong to some maximal block containing an item t, because a i j - a i t+1 . By removing 1 from V g and iteratively using the same argument, the unique partition of V g into maximal blocks B g ng , with B g can be constructed easily. This argument applies to all numbers g 2 fg g. From the above discussions follows that, if we define (MSKP) using the unique partition into maximal blocks, a vector z is feasible for (MSKP) if and only if the associated vectors x are feasible for (SKP). As each maximal block contains items in N with the same gain per unit we obtain in addition: a vector x 2 IR n is optimal with respect to (SKP) if and only if the associated vector z 2 IR m is optimal with respect to (MSKP) and vice versa, a vector z 2 IR m is optimal with respect to (MSKP) if and only if all of its associated vectors x 2 IR n are optimal solutions to (SKP). To simplify notation, we always assume, when transforming (SKP) to (MSKP) using maximal blocks, that f 1 - f 2 - fm , and in case . Moreover, the above arguments for the construction of the unique partition into maximal blocks show that for the transformed problem (MSKP) the following property always holds: fw This property will be used in the next section to derive a decomposition scheme of all optimal solutions of (MSKP). 3 Decomposition of optimum solutions In this section, we characterize the optimal solutions of a problem (MSKP) obtained by the maximal block transformation of an initial (SKP) problem presented in the previous section. Let positive rational numbers c positive integers fm be given such that We also assume that for every For every F 2 IN and we denote by P F (j) the convex hull of all solutions of the following (MSKP) problem with knapsack capacity F and restricted to the variables 1 to j. The optimization problem OP F (j) is the program Note that in this section we only consider optimization problems OP F (j) with positive objective coefficients. Using this notation we have that and (m). By O F (j) we denote the set of all optimal solutions to OP F (j). Finally, for an item oe i.e., \Delta i is the set of all items before i whose gain per unit is strictly better than the one of i. Let f be the total weight of items in \Delta i . For every F and j, we now construct a decomposition tree whose paths from the root node to the leaves contain all the optimal solutions of OP F (j). The key for this result is the next lemma showing that for every optimum solution z 2 O F (j) the component z j can attain at most two different values. Lemma 3.1. For the optimization problem OP F (j) with positive cost coefficients the following statement is true: z 2 O F (j) implies that z j - min ae and z j - min ae l Proof. We prove this result by contradiction using standard exchange arguments. Several cases are distinguished. the lemma states that z (j). By contra- diction, suppose that there exists z 2 O F (j) with z j ? 0. As f l z l the divisibility of the weights, there exist integers - l 2 all l 2 \Delta j such that P . We now define a solution z 0 with z 0 i and the solution z 0 has strictly better objective value than z by definition of \Delta j . This contradicts the optimality of z. because u j is in- tegral. In this case the lemma states that z (j). By contradiction, suppose that there exists z 2 O F (j) with z the new solution z 0 with z 0 z 0 belongs to P F (j) and has strictly better objective value than z, because c j for all contradiction. Hence, we can assume that By the divisibility of the weights, there exist integers - l 2 with The new solution z 0 with z 0 for l 2 belongs to P F (j). 0g. As (where the last inequality holds by as- sumption), there exists i 2 W with c i . Then the solution z 0 has strictly better objective value than z, again a contradiction. In the remaining cases we have that and the lemma states and z j - l F \Gammaf suppose, by contradiction, that there exists z 2 O F (j) with z similar argument as in case (ii) shows that there exists a solution z and with a strictly better objective value than z, a contradiction. suppose, by contradiction, that there exists z 2 O F (j) with z l F \Gammaf As f l z l +f j z j - l F \Gammaf similar argument as in case (i) shows that there exists a solution z z 0 l F \Gammaf and with a strictly better objective value than z, a contradiction. Lemma 3.1 can be applied inductively to build a binary decomposition tree containing all potential optimal solutions in O F (j). We illustrate this on an example. Example 2.2 Continued. The modified sequential knapsack problem P 396 (5) using the maximal block transformation was defined as with upper bounds on the variables We have c 1 0:56, and hence Z Z Z Z Z Z Z Z Z Z (2) (2) (2) (2) (1) (1) (1) (1) (1) (1) (1) Figure 1: Decomposition of Optimal Solutions for Example 2.2 Figure 1 illustrates the decomposition tree that we obtain from applying Lemma 3.1 iteratively. The node labels identify the problems P F (j) to be solved and the value of z j is fixed on the corresponding branches. For example, Lemma 3.1 applied to P 396 (5) yields z are left with problem are left with problem P 36 (4). Potential optimal solutions of problems P 396 (4) and P 36 (4) are further decomposed using Lemma 3.1. The set S 396 (5) of potential optimal solutions to P 396 (5) is defined by all the paths from the leaves to the root node in the decomposition tree, that is and, by Lemma 3.1, O 396 (5) ' S 396 (5). For a given problem P F (j) and its associated decomposition tree, we define in the next section valid inequalities that are satisfied at equality by all solutions in this decomposition tree, and thus by all optimal solutions in O F (j). 4 The convex hull of all solutions to the sequential knapsack problem Let (SKP) be a sequential knapsack problem and suppose that C is a class of valid inequalities for P SKP . The technique that we use in order to show that C describes P SKP is due to Lovasz [5]: for every objective function fl we prove that the set of optimal solutions to (SKP) belongs to the face induced by some inequality in C. This suffices to show that C describes P SKP , because when an objective function is parallel to a facet defining inequality, then the only inequality satisfied at equality by all optimal points in (SKP) is this facet defining inequality. Hence, C contains all the facet defining inequalities. We first consider the case that all objective function coefficients are positive. As outlined in Section 2, we partition N into maximal blocks construct the modified sequential knapsack problem (MSKP). Associated with the transformed problem (MSKP), we use the notations P F (j), OP F (j), O F (j), introduced in Section 3. For every knapsack capacity F 2 IN and for every mg, we now define an inequality I F (j) satisfying the conditions (i), (ii), (iii) listed below: (i) The left hand sides of inequalities I F (j) and I F 0 (j) are equal if F modulo f j holds. (ii) I F (j) is a valid inequality for P F (j). (iii) The set of optimal solutions O F (j) is contained in the face induced by the inequality I F (j). The inequalities I F (j) are defined inductively on j. We define the inequality I F (1) as z 1 - minfF; u 1 g. This inequality clearly satisfies all the properties (i) - (iii). number r between 0 and f given and assume that for every number F 2 IN with F modulo f there exists an inequality I F (j \Gamma 1) that satisfies the properties (i) - (iii). In particular, property (i) guarantees that this family of inequalities is of the are the coefficients of the inequalities I F (j \Gamma 1) for all F with F modulo f r. With the parameter r we associate a number F r . We set F r := r if f r is the largest number of residuum class r with respect to f j not exceeding the sum of weights in that have a better gain per unit than j. For every F 2 IN with F modulo f r the left hand side of the inequality I F (j) is of the form defined as in I F (j \Gamma 1) and d j defined by In order to define the corresponding right hand side - that we denote by g F;j - we need to distinguish several cases. First, is an integer. We set Under these assumptions the inequality I F (j) defined via the three properties (i), (ii) and (iii). These statements are shown below. We first illustrate this (inductive) construction on the initial example. Then three technical lemmas are proved and afterwards applied to show that I F (j) satisfies (i)-(iii). Example 2.2 continued. with upper bounds on the variables The construction of the inequalities is defined for any value of F . If we are only interested in the inequality I 396 (5), then we need not find I F (j) for all values of F . The node labels in Figure 1 represent the subproblems we have to solve in order to obtain an optimum solution for the original problem OP 396 (5). They also give the F and r values we must consider in order to construct I 396 (5). I I F (2) with I I 6 (2) I 1+5s (2) I 1+5s (2) I F (3) with I I I I F (4) with I I I F (5) with I I The inequality I 396 (5) is satisfied at equality by all solutions in S 396 (5) containing all optimal solutions in O 396 (5). Lemma 4.1. Let F and G be natural numbers such that F - G and F modulo holds. Proof. For the statement is certainly true. So assume, it holds for all numbers that are less or equal than j \Gamma 1. We show that it is true for j as well. We G, we know that s - t. We define G; if t ! 0; Checking all cases we notice that F 0 - G 0 , F 0 modulo f As F 0 modulo by assumption of the induction and the claim follows. Lemma 4.2. Let F and G be natural numbers such that F - G and F modulo holds for every Proof. g G+oef j Applying Lemma 4.1 yields [g G+oef Iterating this argument proves Lemma 4.2. Accordingly, we obtain Lemma 4.3. Lemma 4.3. Let F and G be natural numbers such that F - G and F modulo holds for every Proof. 4.1, we conclude that [g F;j \Gamma Iterating this argument proves Lemma 4.3. Using Lemmas 4.1 - 4.3 we are now able to prove the following theorem. Theorem 4.4. Given a modified sequential knapsack problem with positive objective function obtained from the maximal block transformation. If the inequalities I F (j \Gamma 1) satisfy the three conditions (i), (ii), (iii) with so do the inequalities I F (j) with (i) The left hand sides of two inequalities I F (k) and I F 0 (k) are identical whenever (ii) I F (k) is valid for P F (k); (iii) Every optimum solution to problem OP F (k) is contained in the face induced by the inequality I F (k); Proof. We write I F (j) as (i) Let F and F 0 be two natural numbers satisfying F modulo f As \Delta j and F r are uniquely defined by the residuum class r and the objective function we obtain - per definition - that the left hand sides of the two inequalities I F (j) and I F 0 (j) are the same. (ii) The inequality I F (j) is valid for the polyhedron P F (j). Let z 2 P F (j) be a feasible point, then G;j \Gamma1 is a valid inequality for all values of G with G modulo f j Again, we write cases. (ii) (a) F - F r . Then s - 0. If z it follows from the definition of g F;j \Gamma1 that the inequality is valid. Suppose that z j ? 0. Then , where the last inequality follows from Lemma 4.3, and the statement follows. it follows from the definition of that the inequality is valid. Suppose that z j ! u j . By applying Lemma 4.2, we obtain: (ii) (c) What remains is the case where F r holds and we obtain If z the inequality is valid by construction. Otherwise, if z j ? s, then Lemma 4.3 implies that g Fr \Gamma(z j \Gammas)f together with (?) we have that I F (j) is valid. Finally, if z implies that g Fr \Gamma(z j \Gammas)f which again shows that the inequality I F (j) is valid. (iii) It remains to be shown that the set of optimal solution O F (j) is contained in the face induced by the inequality I F (j). By definition of F r , we can always write (j), then by Lemma 3.1 and by definition of F r we have F \Gammaf . If r - f then F r is the unique number such that F r - f r. In this case Lemma 3.1 implies that or z In this case, z Hence, r ? f implies that z sg. Summarizing all cases yields z or z In case s - 0, i.e., F - F r , we have z in every optimum solution. Therefore by assumption of the induction, every optimum solution to problem OP F (j) is contained in the face this case, the claim follows. In case s - every element in the set O F (j) satisfies By assumption of the induction, every optimum solution z to problem F;j . This proves the claim in this case. Finally, we have 1 - s - u j . Then every optimum solution z of OP F (j) satisfies either z By assumption of the induction we obtain that (a) This yields in case (a): In case (b) we obtain: This shows that in both cases the inequality I F (j) is satisfied at equality by all optimal points. Let us now present the final theorem describing P SKP as a system of inequalities. Let W ' N be a subset of items in N , let be a partition of W into blocks and let - be a permutation of l g be the weight of block a l s l be the multiplicity of block B i and assume that f 0 m . We set f Denote by P b a 0 f 0c (m) the modified knapsack polytope defined with the block partition B of W , weights f multiplicities um and knapsack capacity b a 0 f 0c. That is z f 0c If the inequality I b a 0 f 0c (m), written as f 0c;m , denotes the valid inequality developed in this section for P b a 0 f 0c (m) using the sets \Delta j induced by the permutation -, then the inequality K(W;B;-) is defined as a i Theorem 4.5. Given an instance of (SKP), the following system of inequalities describes the polyhedron K(W;B;-); for all W ' N; all partitions W into blocks, and all permutations - of Proof. We first show validity of the inequalities K(W;B;-). Given W , and -. It is easy to check that there exists an objective function is the partition of W into maximal blocks and there exists - such that g: Then, the inequality I b a 0 f 0c (m) is valid for the polyhedron P b a 0 f 0c (m) by Theorem 4.4 (i) and (ii). By the arguments on the transformation of valid inequalities for PMSKP to valid inequalities for P SKP (see Section 2), the inequality K(W;B;-) is valid for the polyhedron a i This polyhedron is a relaxation of P SKP , because f 0 1 is the smallest weight among all items in W . As K(W;B;-) is valid for this relaxation of P SKP , it is certainly valid for P SKP . Now given any objective function construct an inequality satisfied at equality by all optimal solutions of (SKP). If then and we set W := fi 2 Bmg be the partition of W into maximal blocks and let (MSKP) denote the modified sequential knapsack problem of Section 2. From Theorem 4.4 (iii) we know that I b a 0 f 0c (m) is satisfied at equality by all optimal solutions of (MSKP). By the arguments on the equivalence of optimal solutions beween problems (SKP) and (MSKP) (see Section 2), K(W;B;-) is satisfied at equality by all optimal solutions of a i Now if x is an optimal solution of the original problem with K(W;B;-) not satisfied at equality (because some i 2 N n W has value x i ? 0), then a solution with strictly better objective function value can be found by setting x This completes the proof. 5 Explicit Inequalities In the previous section we have inductively defined a class of inequalities that depends on the choice and ordering of the blocks. Can we find a more explicit or combinatorial formulation for those inequalities? This question is addressed now. Given a sequential knapsack problem of the form where u is the vector of upper bounds on the variables and 1 A large class of inequalities for the associated polyhedron P SKP can be described as follows: Let r i denote the residuum of the capacity F modulo f i . We choose sets S i ' N i , with the following properties: Setting the inequality is valid for P SKP . This statement can be verified by applying our inductive scheme: we define a modified sequential knapsack problem and, for every item i in this modified problem, we choose a set \Delta i such that the inequality constructed via our inductive scheme coincides with (?). We first consider the case where 2. Here we define the transformation to (MSKP) by considering k. Thus, the modified sequential knapsack is of the form The ordering of blocks is defined by \Delta F;k denote the inequality I F (k) for this modified problem (MSKP) with sets now show that (?) coincides with the inequality F;k that is obtained by transforming I F (k) to a valid inequality for (SKP) (see Section 2). As we have that, for any j - 2, r thus To derive I F (k) using our inductive scheme, we have to compute the numbers g. As f Starting from d going through the inductive scheme (see Section 4) we obtain for each finally which shows that the inequality (?) is obtained via our inductive scheme and thus is valid for P SKP when When then we consider as a single block. Performing this for all j with generating the corresponding modified knapsack problem, constructing the valid inequality using our inductive scheme and transforming it to a valid inequality for (SKP) yields the inequality (?). The inequalities of the form (?) are already a strong generalization of other known inequalities: In case that if u is the vector of all ones (the 0=1 case), then the inequality (?) is of the form cl . The latter class of inequalities plus the trivial inequalities plus the cover inequality c describe the polyhedron g. This result was shown in [11] and, independently in [3]. As a special case we obtain Padberg's result on (1; k)-configurations [8]: Suppose, we are given a knapsack problem such that the set of feasible solutions is equal to The corresponding polyhedron is described by the lower and upper bound constraints plus the inequalities for all subsets Summarizing our discussions, the inequalities (?) are only a subclass of the inequalities needed to describe a sequential knapsack polyhedron. Nevertheless, this subclass is quite large and extends all the explicitly known inequalities for special cases of the knapsack problem having the divisibility property. 6 Extensions The previous sections deal exclusively with the sequential knapsack polytope which is still a restrictive assumption when considering integer programs in gene- ral. Can we use parts of this polyhedral knowledge presented so far and apply it within a more general framework? The answer is "yes" and we outline now some directions. A first question in using our inductively defined inequalities computationally is whether we have a combinatorial algorithm for solving the separation problem, i.e., given a fractional solution y: does there exists an inequality that is violated by y and if so, then what is the inequality? We did not succeed in solving this separation problem. "Only" for the subclass of inequalities cl where ; Hartmann [3] gives a linear time algorithm for solving the separation problem . The general problem is still open. However, we can use our inductive scheme as a separation heuristic. For instance, defining every item i 2 N as a single block, setting generating an inequality according to this ordering seems to be a promissing approach to end up with a violated inequality, if one exists. Other reasonable definitions of \Delta i might be to set Whether those ideas work is certainly not clear, but similar "greedy type" of procedures work pretty well for the separation of cover- and (1; k)-configuration inequalities. Given an integer programming problem Ax - b; 0 - x - u; x integer with . If there exists some row P such that a subset S of items in fj 2 ng 0g has the divisibility property, then we can investigate the polyhedron: convfx 2 and generate inequalities for this polyhedron. By computing lifting coefficients for the items in N n S, we obtain a valid inequality for the overall polyhedron convfx 2 g. This approach can always be used to apply knowledge about special integer programs to more general cases. Another idea is to try to relax a given integer program as a sequential knapsack problem. Given a row of an integer program, the easiest way to obtain a relaxation as a sequential knapsack problem is to choose, a priori, a set of divisible numbers f say. The sequential knapsack problem defined via the constraint X is certainly a relaxation of the given integer program. A more specific relaxation is obtained by generalizing the concept of (1; k)- configurations. Consider the 0=1 knapsack problem defined by the constraint with N := assume that f(S) - F , f(S) define a partition S 1 of the set N of items as Based on this partition, we define an inequality with the divisibility property that is valid for the given 0=1 knapsack problem. We set b 1 := 1 and, for we define Note that t j - We define finally If such a t -+1 exists (i.e. if f s ! f then the inequality is valid for the 0=1 knapsack problem. Before verifying this statement, let us illustrate the above construction on an example. Example 5.1. Consider the knapsack problem in 0=1 variables defined via the constraint 3. We choose This meets the requirements that the indices must satisfy, because f 4 - f 3 +f 2 and f 8 - f 7 +f 6 . In this example, the inequality (?) is of the form and it is valid for the given knapsack polytope. Let us now show that the inequality (?) is always valid under the above assump- tions. It is valid if and only if every subset T ' S with f(T equivalently if and only if the problem s has an optimal value f Setting Y j := first show that there always exists an optimal solution to this problem with Y First, observe that Y - t -+1 is infeasible for this problem because b as by construction solution obtained by decreasing Y - \Gamma1 by t - and increasing Y - by 1 is at least as good as the initial solution in terms of objective value equivalently in terms of the knapsack constraint optimal solution with can be transformed into an optimal solution with Proceeding in this way for all , we can produce an optimal solution with The objective value of such a solution satisfies because implies that there exists z 2 S j with y such that f Summing these inequalities for Hence r and the inequality is valid. By construction, b j is a multiple of b j \Gamma1 for all j - 2. It follows that (?) has the divisibility property and we can apply all of our information for the sequential knapsack polytope induced by inequality (?). In case that if we impose a "regularity condition" such as "every subset T in S with b(T then the corresponding inequality defines a facet of the 0=1 knapsack polytope [8]. For - 2 one can also derive sufficient conditions under which inequality (?) defines a facet of the corresponding polytope. Yet, such conditions are quite technical and we refrain within this paper from explaining further details. If one finds such generalized (1; k)-configurations or some subset of the items having the divisibility property with respect to some row of a given integer program all the knowledge about the sequential knapsack polytope can be used. Together with lifting this yields a powerful tool that might help solving integer programs. Acknowledgement . This research was partially supported by Science Program SC1-CT91-620 of the EEC and contract ERB CHBGCT 920167 of the EEC Human Capital and Mobility Program. --R "The sequential knapsack pro- blem" "Some polyhedra related to combinatorial problems" "Cutting planes and the sequential knapsack problem" "Solving sequential knapsack problems " "Graph theory and integer programming " "The cutting stock problem and integer rounding" "A Note on 0-1 Programming" "(1,k)-Configurations and Facets for Packing Pro- blems" "Integer knapsack and flow covers with divisible coefficients: Polyhedra, optimization and separation" "On the 0/1 knapsack polytope" "Hilbert bases and the facets of special knapsack pro- blems" "Faces of Linear Inequalities in 0-1 Variables" --TR
integer programming;linear programming formulation;knapsack problem;knapsack polytope;separation
589172
Modified Cholesky Factorizations in Interior-Point Algorithms for Linear Programming.
We investigate a modified Cholesky algorithm typical of those used in most interior-point codes for linear programming. Cholesky-based interior-point codes are popular for three reasons: their implementation requires only minimal changes to standard sparse Cholesky algorithms (allowing us to take full advantage of software written by specialists in that area); they tend to be more efficient than competing approaches that use alternative factorizations; and they perform robustly on most practical problems, yielding good interior-point steps even when the coefficient matrix of the main linear system to be solved for the step components is ill conditioned. We investigate this surprisingly robust performance by using analytical tools from matrix perturbation theory and error analysis, illustrating our results with computational experiments. Finally, we point out the potential limitations of this approach.
Introduction . Most interior-point codes for linear programming share a common feature: their major computational operation-solution of a large linear system of equations-is performed by a direct sparse Cholesky algorithm. In this algorithm, row and column orderings are determined a priori by well-known heuristics (minimum degree and enhancements, minimum local fill, nested dissection) that are based solely on the sparsity pattern and not on the numerical values of the nonzero elements. The ordering phase is followed by a symbolic factorization phase, in which the nonzero structure of the Cholesky factor is determined and storage is allocated. Finally, a numerical factorization phase fills in the numerical values of the lower triangular In interior-point codes, the first two phases usually are performed just once, during either the first interior-point iteration or computation of a starting point. In the interior-point context, the unadorned Cholesky algorithm can run into difficulties because of extreme ill conditioning. Some of the diagonal pivots encountered during the numerical factorization phase can be zero or negative, causing the standard Cholesky procedure to break down. Instead of crashing, most codes apply a "patch" to the algorithm to handle such pivots. The offending pivot element is sometimes replaced by a huge number, as in LIPSOL [17] or PCx [1]. In other codes such as IPMOS [16], the pivot is replaced by a moderate number, but the corresponding right-hand side element is set to zero, as are the off-diagonal elements in the corresponding column of the Cholesky factor. The first practical interior-point code, OB1 [6], explicitly zeroes the components of the solution vector that correspond to small pivots. All these strategies are essentially equivalent to the algorithm we describe in this paper. To date, there has been little investigation of them from a numerical analysis viewpoint. The "patches" described above have the advantage that they can be implemented by changing just a few lines in general sparse Cholesky codes. It is therefore possible to take advantage of the long-term development effort that has gone into designing such codes and their underlying algorithms. The recent codes LIPSOL [17] and PCx Mathematics and Computer Science Division, Argonne National Laboratory, 9700 South Cass Avenue, Argonne, IL 60439. This work was supported by the Mathematical, Information, and Computational Sciences Division subprogram of the Office of Computational and Technology Research, U.S. Department of Energy, under Contract W-31-109-Eng-38. [1] make explicit use of the freely available sparse Cholesky code of Ng and Peyton [8]. Other codes either modify the well-known SPARSPAK routines of George and Liu [3] or include customized linear algebra routines that implement well established algorithmic ideas. (At least one author has experimented with modifications to the standard heuristics: M'esz'aros [7] describes an inexact version of the minimum local fill ordering.) One possible remedy for small pivots is diagonal pivoting. At each iteration, a "large" diagonal element is selected from the unreduced portion of the matrix and moved to the pivot position by symmetric row and column pivoting. The algorithm is terminated when none of the remaining diagonal elements is sufficiently large, and an approximate solution is computed with the partial factors. (See Higham [4, Chapter 10] for details and error analysis.) This strategy is not particularly appealing in the context of interior-point linear programming codes because of the loss of efficiency due to shifting of data during the numerical factorization. Moreover, there is little incentive to test this strategy because the simple patches described above perform so well in practice. In this article, we use standard results from numerical analysis to explain the good performance of these patching strategies on the vast majority of problems. We also gain some insight into their limitations and into how and why they fail. Our error analysis for the modified Cholesky algorithm is rigorous, with explicitly stated assumptions and precise bounds (see Sections 3 and 4). We revert, however, to a more informal style when applying these results to the interior-point context (Section 5). The reason is pure pragmatism. A fully rigorous analysis would be impossibly notationally speaking, and unduly pessimistic. The informal analysis yields adequate insight into the typical performance of the algorithm, as our computational results in Section 6 demonstrate. A number of other papers on linear algebra operations in barrier and interior-point methods have appeared in recent years. Wright [12] has considered the Newton- logarithmic barrier method for general constrained optimization, in which the linear system to be solved for the Newton step is positive semidefinite and ill conditioned during later iterations. She uses a Cholesky factorization with diagonal pivoting to identify the subspace spanned by the active constraint Jacobian. From this infor- mation, an accurate solution of the Newton equations can be obtained, in which the components of the step in both the range space of the active constraint Jacobian and the null space of its transpose are well resolved. Our analysis has a similar flavor to Wright's, but the application is somewhat different. The unknowns in our linear system are the unconstrained dual variables rather than the primals and, since this problem is linear, we have little interest in resolving the component of the step in the near-null space of the coefficient matrix. We focus too on Cholesky algorithms that perform no pivoting during the numerical factorization, reflecting computational practice in the current generation of interior-point linear programming codes. In an earlier paper [14], we considered the stability of algorithms for the symmetric indefinite form of the step equations at each iteration of a interior-point method for linear programming. We showed that, despite the ill-conditioning of the coefficient matrix, the steps obtained by this approach are good search directions for the interior-point method. Forsgren, Gill, and Shinnerl [2] perform a similar analysis in the context of logarithmic barrier methods. The remainder of this paper is organized as follows. In Section 2, we introduce primal-dual interior-point methods and derive the linear equations to be solved at each iteration of these methods. Section 3 introduces Algorithm modchol, the modified procedure, and examines the accuracy of the solution obtained with this factorization, under certain assumptions on the eigenvalues of the factored matrix. In Section 4, we account for the effect of finite-precision floating-point arithmetic on solution accuracy. We return to the interior-point application in Section 5, showing that Algorithm modchol yields good steps for these methods until the duality gap becomes very small, even if the linear program is primal or dual degenerate. The analytical results are verified by computational experiments with an interior-point code using Algorithm modchol, which are reported in Section 6. Notation. We summarize here the notation used in the remainder of the paper. The ith singular value of a matrix A is denoted by oe i (A). We use oe i alone to denote the ith singular value of the exact Cholesky factor L in Section 3. For any matrix M and index steps I and J , M IJ denotes the submatrix formed by the elements . The ith column of M is denoted by M \Deltai , and the column submatrix consisting of columns j 2 J is denoted by M \DeltaJ . Unit roundoff error is denoted by u. Higham [4, Chapter 1] defines u implicitly by the statement that when ff and i are any two floating-point numbers, op denotes +, \Gamma, \Theta, and =, and fl(\Delta) denotes the floating-point representation of a real number, we have For any positive integer m with mu ! 1, we define (1) (see Higham [4, Lemma 2.1]). The notation k \Delta k denotes the Euclidean vector norm k \Delta k 2 and also its induced matrix norm, unless otherwise noted. For any matrix A, the matrix consisting of the absolute values of each element is denoted by jAj. We use 1 to denote the vector Finally, we mention the parameter ffl that defines the pivot threshold in the modified algorithm. A second quantity -ffl, which is related to ffl by appears frequently in the analysis because the incorporation of the scaling term 2m 2 saves notational clutter. 2. Primal-Dual Algorithms for Linear Programming. We consider the linear programming problem in standard form: subject to (2) . The dual of (2) is subject to A T - m . We assume throughout the paper that A has full row rank, so that m - n. The Karush-Kuhn-Tucker (KKT) conditions, which identify a vector triple (x; -; s) as a primal-dual solution for (2), (3), can be stated as follows: (4a) (4b) (4c) (4d) We assume throughout the paper that a primal-dual solution exists. We make no assumptions about uniqueness or nondegeneracy; our analysis in Section 5 continues to hold when the problem (2) is primal or dual degenerate. It is well known that the index set f1; ng can be partitioned into two sets B and N such that for all primal-dual solutions x Primal-dual interior-point algorithms generate a sequence of iterates (x; -; s) that satisfy the strict inequality (x; s) ? 0. They find search directions by applying a modification of Newton's method to the system of nonlinear equations formed by the first three KKT conditions (4a),(4b),(4c), namely, In general, the search direction (\Deltax; \Delta-; \Deltas) is obtained from the following linear \Delta- \Deltas5 =4 \Gammar c where the coefficient matrix is the Jacobian of (6) and the right-hand side components r b and r c are defined by In a pure Newton (affine-scaling) method, the remaining right-hand side component r xs is defined by and, in this case, we denote the solution of (7) by (\Deltax aff ; \Delta- aff ; \Deltas aff ). In a path-following method, we have where - is the duality gap defined by is a centering parameter. In the "Mehrotra predictor-corrector" al- gorithm, which is used as the basis of many practical codes, the search direction is calculated by setting where \DeltaX aff and \DeltaS aff are the diagonal matrices formed from the affine-scaling step components \Deltax aff and \Deltas aff . Hence, Mehrotra's method requires the solution of two linear systems at each iteration-the affine scaling system (7), (8), (9), and the search direction system (7), (8), (12). A heuristic based on the effectiveness of the affine scaling direction is used to determine the value of i in (12). Once a search direction has been determined, the primal-dual algorithm takes a step of the form where ff is chosen to maintain strict positivity of the x and s components; that is, In most codes, ff is chosen to be some fraction of the step-to-boundary ff max defined as A typical strategy is to set as the interior-point method approaches the solution set. By applying block elimination to (7) and using the notation we obtain the following equivalent system: (16a) (16c) In many codes, the solution is obtained from just this formulation. A sparse Cholesky factorization, modified to handle small pivots, is applied to the symmetric positive definite coefficient matrix AD 2 A T in (16a) and the solution \Delta- is obtained by triangular substitution with the computed factor. The remaining direction components are recovered from (16b) and (16c). This technique yields steps (\Deltax; \Delta-; \Deltas) that are useful search directions for the interior-point algorithm, even when the matrix happens during later iterations. This observation is somewhat surprising, since a naive application of error analysis results would suggest that the combination of ill-conditioning and roundoff would corrupt the direction hopelessly. The results of Sections 3, 4, and 5 provide an explanation for this phenomenon. The following observation is crucial to our analysis: In computing \Delta- from (16a), we are not interested so much in the error in \Delta- itself as in the effect of this error on the remaining step components \Deltas and \Deltax that are recovered from (16b) and (16c), respectively. If the relative errors in these components are large, the positivity requirement may cause the step length ff to be significantly shortened, thereby curtailing the algorithm's progress. We return to this issue in Section 5, after describing and analyzing the modified Cholesky algorithm in Sections 3 and 4. 3. A Modified Cholesky Algorithm. In this section, we describe and analyze Algorithm modchol, a modified Cholesky algorithm designed to handle ill-conditioned matrices for which small or negative pivots may arise during the factorization Algorithm modchol accepts an m \Theta m symmetric positive definite matrix M as input, together with a small positive user-defined parameter ffl, which defines a threshold of acceptability for the pivot elements. If a candidate pivot element is smaller than this threshold, the algorithm simply skips a step of factorization. Algorithm modchol outputs an approximate lower triangular factor ~ L and an index set J ae containing the indices of the skipped pivots. In the following specification, we use M (i) to denote the unfactored part of M that remains after i steps of the algorithm. Algorithm modchol Given ffl with if M (i\Gamma1) (* skip this elimination step *) im . else (* perform the usual Cholesky elimination step *) ~ ~ L ki . The ith column of ~ L is zero for each i 2 J ; that is, ~ If we denote and denote the complement of J in f1; J , it follows from (17) that That is, the row or column index of each nonzero element in E must lie in J . It follows from the algorithm that ~ L is the exact Cholesky factor of the perturbed matrix which we denote for convenience by ~ M . That is, we have ~ By partitioning this equation into its J and - J components and using ~ (19), we obtain ~ (21a) ~ The exact Cholesky factor L (whose existence is guaranteed by the assumed positive definiteness of M ) satisfies Given the linear system where M is the matrix factored by modchol, the exact solution obviously satisfies The approximate solution ~ z is chosen so that the partial vector ~ z - J solves the reduced system M - z - J , while the complementary subvector z J is set to zero. From (21a), we see that ~ z - J can be calculated by performing a pair of triangular substitu- tions; that is, ~ z - ~ z Note that z = ~ z when on the other hand, the difference between ~ z and z can be large in a relative sense. We have z - z - and there is no reason to expect z J to be small with respect to the full vector z. We can show, however, that the difference between L T z and L T ~ z is relatively small under certain assumptions; this result is the culmination of the analysis of this section (Theorem 3.6). As we see in Section 5, this difference determines the usefulness of the computed solution of (16) as a search direction for the interior-point algorithm. To simplify the analysis, we assume implicitly throughout the paper that A trivial scaling, which affects neither the algorithm nor its analysis, can always be applied to the symmetric positive definite matrix M to yield (26). We start with a sequence of three results that lead to a bound on the difference between ~ z. These results require few assumptions on the matrix M and are relatively simple to prove. Lemma 3.1. The submatrix formed by the last columns of M (i) is symmetric positive definite, for all 1. Moreover, the diagonal elements of all these submatrices are bounded by 1. Proof. This observation follows by a simple inductive argument. By assumption, the starting matrix M positive definite. Suppose that the desired property holds for M (i\Gamma1) . If i 2 J , then the lower right (m \Gamma i) \Theta (m \Gamma i) submatrix of M (i) is identical to the lower right (m \Gamma i) \Theta (m \Gamma i) submatrix of M (i\Gamma1) , which is positive definite by assumption. Otherwise, if is obtained by applying one step of Cholesky reduction to M (i\Gamma1) . It is known that the remaining submatrix resulting from this operation is positive definite; hence, the lower right in question is positive definite, and the desired property holds. The second claim follows immediately from the fact that M ii - and the fact that the diagonal elements cannot increase during Algorithm modchol. Lemma 3.2. For each i 2 J , we have Therefore, Proof. From Lemma 3.1, we have (M (i\Gamma1) l;l for each ffl. Since the diagonals of each submatrix M (i\Gamma1) are bounded by 1, we have M (i\Gamma1) i;l l;l Hence, we have thereby proving the first claim. By (18), we have thereby proving (27). In the case in which all the small pivots appear in the bottom right corner of the matrix (that is, index p), the estimate (27) can be improved to This stronger estimate applies in most instances of the interior-point application of Section 5. We are now able to derive an estimate of the difference between ~ Theorem 3.3. For the exact solution z and approximate solution ~ z defined in (24) and (25), respectively, we have that Proof. From (24) together with (21), we have JJ z J ~ ~ JJ z J while from (25), we have ~ z - J ~ z J J ~ z: By combining these two relations, we obtain ~ Since ~ the result follows immediately. The remaining analysis of this section requires some additional assumptions on the distribution of the singular values of M and on the parameter ffl. Accordingly, we introduce a little more notation. The eigenvalues of M are denoted by oe 2 We define the diagonal matrix \Sigma by It follows that there exists an orthogonal matrix Q such that Because the largest diagonal in M is 1, we have by elementary analysis that In the subsequent analysis, we assume that there is an integer p with 1 - such that ffl ffl is small relative to oe 2 there is a significant gap in the spectrum of M between oe 2 p and p+1 . (We will be more specific about these two assumptions presently.) By partitioning the spectrum at the gap, we obtain From (33), Q can be partitioned accordingly to obtain are the singular values of L. In fact, we must have orthogonal matrix U , where \Sigma and Q are defined as above. We use ~ to denote the eigenvalues of the perturbed matrix ~ M . It follows immediately from (20) that the singular values of ~ are ~ oe i , The rank of ~ J j, because ~ J is lower triangular with nonzero diagonals while ~ Therefore, we have ~ As in (36), there are orthogonal m \Theta m matrices ~ U and ~ Q such that ~ U ~ where ~ It is an immediate consequence of an eigenvalue perturbation result of Stewart and Sun [10, Corollary IV.4.13] and Lemma 3.2 that The main assumption of this section is that j - correctly identifies the numerical rank of the matrix M . One might expect that we should not have to assume this equality at all-that it should follow from the spectrum gap and from a judicious choice of ffl. Practical experience supports this expectation; the algorithm has little trouble determining the numerical rank on the vast majority of problems. In fact, part of the result-the bound j - p-follows from a minimal assumption on ffl. Lemma 3.4. If - Proof. If we have from (37) and (39) that contradicting our assumption that -ffl 1=2 ! oe 2 . However, the conditions on ffl, oe p , and oe p+1 needed to prove the other half of the result-j - rigorous to be useful. This is a consequence of the fact that poorly conditioned triangular matrices need not have particularly small diagonal elements (see Lawson and Hanson [5, p. 31] for the classic example of this phenomenon). Our next result concerns perturbation of the subspace spanned by Q 1 , which is the invariant subspace of "large" eigenvalues of M . Lemma 3.5. Suppose that j - and that the values oe p and oe p+1 from (31) and ffl from Lemma 3.2 satisfy the conditions (40a) (40b) Then there is a p \Theta p symmetric positive definite matrix ~ and an orthonormal m \Theta p matrix ~ ~ ~ ~ (The constants used in (40a) and in similar expressions should not be taken too seriously. We assign them specific values only to avoid an excess of notation.) Proof. The result is a straightforward consequence of Theorem V.2.8 of Stewart and Sun [10, p. 238]. Since ~ use (33) and partition as in (35) to obtain We now make the following identifications with the quantities in the cited result: ~ where sep(\Delta; \Delta) is the minimum distance between the spectra of its two arguments. From the given result, there is a matrix P of dimension (m \Gamma p) \Theta p such that the matrix ~ defined by ~ is an invariant subspace for ~ ~ Moreover, the representation of ~ M with respect to ~ ~ The bound (42) follows from (44), (45), and kQ 2 It follows immediately from the first equality in (46) that ~ is symmetric, and we have verifying the inequality (43). This inequality implies that the smallest singular value of ~ is no smaller than oe 2 is symmetric positive definite. The cited result states further that the matrix ~ is orthogonal to ~ defines an invariant subspace for ~ M . In fact, we have for some (m \Gamma p) \Theta (m \Gamma p) symmetric matrix - . Since ~ and ~ both have rank b, we must have - 0, so we have ~ ~ ~ Hence, (41) is also satisfied, and the proof is complete. Combining (40b) with (39), we obtain Another quantity that enters into our error bounds is the norm of ~ J . We denote J jth singular value of ~ J . (The lower bound of 1 in simplifies our analysis.) Note from (21a) that 1 , we have from (34) and (49) that Under the assumption j - the nonzero part of ~ L-the submatrix ~ J -has full rank p and singular values ~ oe oe p . Since ~ J differs from ~ J in the presence of the additional rows ~ J , we have and therefore The additional rows ~ J can have nontrivial magnitude relative to ~ J , so ~ oe p may be significantly larger than - \Gamma1 . However, ~ oe p cannot be too large, since from (39), (40b), and (34), we have that ~ For the purposes of our analysis, we make the assumption that - ~ p is moderate in size. Specifically, we assume that Because of (39) and (40b), we have ~ implies that and, in addition, 3: We can now prove the main result of this section. Theorem 3.6. Suppose that j - m, that the conditions (40) hold, and that the estimate (51) is satisfied. We then have Proof. From (36), we have since U is orthogonal. Now from the partition (35), and using the fact that kQ 2 (unless of course \Sigma 2 and Q 2 are vacuous), we obtain ~ ~ The first term in this expression is easiest to bound. From (35), we have k\Sigma \Gamma1 . Applying the relations (41), (20), (38), (47), (29), (27), and (48), respectively, we obtain ~ We therefore have ~ The second and third terms in (53) require a bound on k~z \Gamma zk. From (30) and the fact that ~ z ~ and therefore kz J k: From (47), we have k ~ while from (27), we have Substituting these estimates into (56) and using (52), we obtain Finally, using ~ z together with - 1, (34), and (57), we obtain Turning specifically to the second term in (53), we have from (34), Lemma 3.5, (47), and (40) that ~ oe 22-ffl 1=2 By combining this bound with (58) and k\Sigma \Gamma1 , we obtain ~ For the third term in (53), we have from k\Sigma 2 The result of the theorem is obtained by substituting (54), (59), and (60) into (53). Note that if m), we have ~ so the conclusion of Theorem 3.6 holds trivially in this case as well is we define oe 4. The Effect of Finite Precision Computations. In the analysis of the preceding section, we assumed for simplicity that all arithmetic was exact. In this section, we take account of the roundoff errors that are introduced when the approximate solution ~ z is calculated in a finite-precision environment. Our analysis above focused on the approximate solution ~ z obtained from (25), where the subvector ~ z J satisfies the following system: z - while the subvector ~ z J is fixed at zero. In this section, we use - z to denote the finite precision analog of ~ z. We examine errors in - z due to ffl roundoff error in Algorithm modchol, ffl error arising during the triangular substitutions in (61), and ffl evaluation error in the right-hand side r. As we see in Section 5, evaluation error in the right-hand side is a significant feature of the application to interior-point codes. We denote this error by e, so that the right-hand side r - J in the system (61) is replaced by r - J . Fortunately, our results follow in a straightforward way from existing results for the Cholesky factorization, since a close inspection of Algorithm modchol shows that it simply performs a standard Cholesky factorization on the submatrix M - J . Before stating the main results, we introduce two more assumptions. The first concerns the relative sizes of - and u, specifically, where fl m+1 is defined as in Section 1. Since - ? 1 and m - 1, it follows immediately that The second assumption is that finite precision does not affect cutoff decisions in Algorithm modchol. That is, the presence of roundoff error in each submatrix M (i\Gamma1) does not affect whether the threshold criterion M (i\Gamma1) ii - fi ffl passes or fails for each i. This assumption concerns the relative sizes of u and ffl, and it requires some ex- planation. We cannot expect to take care of the "borderline cases" in which some candidate pivots fall just to one side or the other of the threshold. Rather, we want the cases in which there is a clear distinction between small and large pivots in exact arithmetic to retain this distinction in finite precision arithmetic, and we want the threshold fi ffl to fall comfortably inside the "gap" in both settings. In finite precision, the size of rounding error introduced into M (i\Gamma1) ii by earlier steps of Algorithm mod- chol is comparable to fiu. (Each time M ii is updated by the algorithm, a positive number no larger than itself is subtracted from it. Since jM ii j - fi, the floating-point error introduced here is bounded by fiu.) We want these errors to be smaller than the threshold fi ffl, so that pivots that are tiny in exact arithmetic do not exceed the threshold in finite precision. Hence, we can state this assumption roughly as follows: The following lemma accounts for the effects of finite precision on the approximate solution ~ z obtained from Algorithm modchol and (25). Lemma 4.1. Suppose that Algorithm modchol and the triangular substitutions in (61) are performed in finite-precision arithmetic with perturbed right-hand side J to yield an approximate solution - z. Suppose, too, that (62) holds and that roundoff error does not affect the composition of J . We then have where z is the exact solution from (23). Proof. Algorithm modchol operates as a standard Cholesky factorization on the J , so we can apply a standard perturbation theorem to bound the error in the subvector - z - J . From Higham [4, Theorem 10.4], we find that - z - J satisfies where Comparing (66) with (61), we find that z - Manipulating in the usual way, we obtain z - It follows immediately from (67) that Combining (50), (62), and (63), we obtain so that the denominator in (68) is bounded below by :5. Hence, by substitution into (68), using (34), (49), (69), and (63), we have that Finally, we bound k~z - J k in terms of kzk. From (34) and (57), we have By combining this bound with (70), we obtain the result. The major results of Sections 3 and 4 can be summarized in the following theorem. Theorem 4.2. Suppose that Algorithm modchol and the triangular substitutions in (61) are performed in finite-precision arithmetic with perturbed right-hand side J to yield an approximate solution - z. Suppose, too, that (62) holds and that roundoff errors do not affect the composition of J . Finally, suppose that either m, the conditions (40) hold, and the estimate (51) is satisfied. We then have ae oe Proof. When the result is immediate from Lemma 4.1 and ~ z. For the remaining case, we obtain (71) by combining the results of Theorem 3.6 and Lemma 4.1. We need note only that kz J k - kzk and that, from (34), we have zk: 5. Application to the Interior-Point Algorithm. In this section, we return to the motivating application: primal-dual interior-point software for linear programming and, in particular, the linear system (16) that is solved at each iteration. We apply the main result-Theorem 4.2-and examine the effect of the parameter ffl and unit roundoff u on the quality of the computed search direction ( c \Deltax; c \Delta-; c \Deltas). Our focus is on the later iterations of the interior-point algorithm, during which - is small and the ill-conditioning of AD 2 A T can become acute. Our results show how and why errors arise in ( c \Deltax; c \Delta-; c \Deltas) and what effect these errors have on the step length, the convergence of the algorithm, and the accuracy that can be attained by this algorithm. They also suggest an appropriate size for the parameter ffl. In this section, we revert to an informal style of analysis, using order notation to hide constants of moderate size. Thus if j and i are two positive numbers, we write if the ratio j=i is not too large. Similarly, we O(j). Conventionally, order notation is used only when j and i are quantities that approach zero in the limit of the algorithm in question. Here, however, we use it in connection with the unit roundoff u, which is small but fixed. This slight abuse of notation results in a much clearer insight into the behavior of Algorithm modchol in the interior-point context. In the next subsection, we look closely at the affine-scaling step, for which r xs is defined by (9). This step is important because it closely approximates the steps taken by most rapidly converging algorithms during their final iterations. Subsection 5.2 shows that the steps calculated during the final stages of Mehrotra's predictor corrector algorithm (and therefore by most interior-point codes) have essentially the same properties as affine-scaling steps. 5.1. Affine-Scaling Steps. We start by estimating the sizes of the various constituents of the equations (16)-the residuals r b and r c , the B and N components of x, s, and the diagonal matrix D. In standard infeasible-interior-point algorithms (see, for example, Wright [15, Chapter 6]), we have These estimates are also observed to hold in practice on the majority of problems for values of - greater than u 1=2 . An immediate consequence of these estimates and the definition (15) is that We assume the coefficient matrix A to be well conditioned; that is, oe 1 (A) and are both \Omega\Gammah/1 We assume further that the submatrix A \DeltaB of columns A \Deltai , well conditioned. It follows from this assumption together with the estimate (73) that the matrix A \DeltaB D 2 \DeltaB has full rank min(jBj; m). In fact, since A \DeltaB is well conditioned, all nonzero singular values of A \DeltaB D 2 \DeltaB it follows from (15) and (73) that A \DeltaN D 2 so we conclude that (74a) Since the largest diagonal element of AD 2 A T is scaled coefficient matrix for (16a) is For consistency with Section 3, the singular values of the matrix in (75) are denoted by oe 2 . From this definition together with (74) and (75), we deduce that (76a) Recalling our notation p of Section 3, we have in this case that The exact Cholesky factor L (see Sections 3 and Suppose now that Algorithm modchol is used to compute the solution of (16a), where the right-hand-side component r xs is set to its affine-scaling value XS1. This process result in a computed solution c \Delta- aff for (16a). The remaining step components c \Deltas aff and c \Deltax aff are obtained by substitution into (16b) and (16c), respectively, again in finite-precision arithmetic. Our main tool for analyzing the errors in the computed step is Theorem 4.2. Consider the exact affine scaling step (\Deltax aff ; \Delta- aff ; \Deltas aff ). Standard results for methods (see, for example, [15, Theorem 7.5]), together with the conditions (72), imply that (This estimate holds only when - falls below a data-dependent threshold ffl(A; b; c) defined by Wright [15, Chapter 3].) From (16b) and (72), we have so it follows from our assumptions about the well conditioning of A that \Delta- We can be more specific about the sizes of the critical components \Deltax aff and \Deltas aff we multiply the third block row in (7) by (XS) \Gamma1 and use the definition (9), we obtain \Deltax aff \Deltas aff Therefore, from (72) and (78), we have for i 2 N that \Deltax aff and therefore, using (72) again, we have \Deltax aff In a similar way, we obtain \Deltas aff From the estimates (78), (80), and (81), we can show that a near-unit step can be taken along the direction (\Deltax aff ; \Delta- aff ; \Deltas aff ) without violating positivity of the x and s components. Substituting (\Deltax; \Delta-; have To verify this estimate, suppose that s i (81), we have so it follows from (72) that For the corresponding component x i , we have from (72) and (78) that x i and \Deltax aff O(-). Hence, for all - sufficiently small and all ff 2 [0; 1], we have logic can be applied to the remaining indices i 2 N , thereby completing our verification of (82). Returning to the computed affine-scaling step ( c \Deltax aff \Delta- aff \Deltas aff ), we now apply Theorem 4.2 after checking that its assumptions of are satisfied for small enough - and reasonable values of u and ffl. For double-precision computations, we have . Hence, since A is well conditioned, we can expect the condition (62) to hold in all nonpathological circumstances. Because of (76), our assumption (40a) on the singular value distribution clearly holds for all sufficiently small -. The condition (40b) is satisfied for any reasonable choice of ffl. The assumption that Algorithm modchol correctly identifies the numerical rank (that is, is, as we discussed in Section 3, difficult to guarantee, but it was observed to hold on all problems that we tested. The assumption that rounding errors do not interfere with the makeup of the small pivot index set J is likewise impossible to verify rigorously; but, as discussed in Section 4, it can reasonably be expected to hold when ffl - u (64). A good choice for ffl-one that satisfies the assumptions just mentioned while keeping the bound (71) as small as possible-is therefore For generality, we continue to use ffl and - ffl in the analysis that follows, substituting the specific value (83) only at the end. Having verified that we can reasonably expect Theorem 4.2 to hold for the system (16a), we now estimate the quantities on the right-hand side of (71). From (76a), we have oe 1 =oe O(1), while from (76b), we have oe O(-). The general estimate while the definition of fl m+1 gives the estimate fl We need to account, too, for the errors incurred in evaluating the right-hand side of (16a). The floating-point error in forming r xs = XS1 is only O(-u) in magnitude, since just a single floating-point multiplication is needed to calculate each component of this vector, and each such element is O(-) (see (72)). The residuals r b and r c have magnitude O(-) in exact arithmetic (see (72)), but they are calculated as differences of O(1) quantities and so contain evaluation error of absolute magnitude O(u). Specifically, componentwise errors in the computed version of r c are bounded by u, and similarly for r b . Because of the estimate (73), the errors in r c are magnified to (- \Gamma1 u) when we multiply by AD 2 in (16a). In fact, this term is the dominant one in the total right-hand-side evaluation error. The errors that occur when we perform floating-point addition of the terms r b , AD 2 r c , and AS \Gamma1 r xs are less significant; they lead to additional terms of sizes O(u) and O(- \Gamma1 u 2 ). In summary, the total right-hand-side evaluation error is O(- \Gamma1 u). Hence, after scaling by the factor ae defined in (75), we have where e is the error vector of Section 4. Substituting the estimates (76), (79), and (84) into (71), we have \Delta- aff If (a reasonable estimate when the Cholesky factorization correctly identifies the numerical rank and A \DeltaB is well conditioned), the error bound above simplifies to \Delta- aff From (77) we have that ae for some orthogonal matrix Q. Since orthogonal transformations do not affect the Euclidean norm of a vector, we can substitute ae 1=2 DA T for L T in (86) and use (75) to write \Delta- aff \Delta- aff Note too that from (58), (65), (79), and (84), we have \Delta- aff \Delta- aff \Delta- aff \Delta- aff where ~ \Delta- aff is the approximate solution that would be obtained by Algorithm mod- chol if it was used to solve (16a) in exact arithmetic. Next, we examine the effect of the error in c \Delta- aff and the evaluation error in the right-hand side of (16b) on the calculated step c \Deltas aff . From (79) and (88), we have that \Delta- aff \Delta- aff Hence, taking into account the O(u) evaluation error in the term r c , we have immediately from (16b) that \Deltas \Deltas aff \Delta- aff Clearly, for the "large" components of s-namely, the i 2 N components-errors of this magnitude do not affect the step length ff max to the boundary defined in (14). However, for the critical components i 2 B, the estimate (90) is not good enough to guarantee that ff max is close to 1. (Repeating the argument that follows (82), we find only that Fortunately, a refined estimate of the error in the B components is available. As in (90), we have \Deltas \Deltas aff \Delta- aff where from (87) we have \Delta- aff From (73), we have D B, so from (91) we obtain c \Deltas aff As in the discussion following (82), we find that s i \Deltas aff possible only if This estimate suggests that near-unit steps can be taken, at least in the c \Deltas aff com- ponents, provided that - is significantly larger that u. When all bets are off! Finally, we estimate the errors in the computed version of \Deltax aff (obtained from (16c)) and estimate their effect on the ff max . Again, we consider the components separately. For B, the O(-u) evaluation error in (r xs ) i is magnified by the term s \Gamma1 replacement of \Deltas aff by c \Deltas aff yields an additional error of size O(-ffl which is also magnified by arithmetic errors are less significant. In summary, we find that c \Deltax aff By the usual reasoning, we find that x i \Deltax aff satisfying (94). For the O(-u) evaluation error in (r xs ) i is not magnified appreciably by s , while from (90), the O(- + u) error in \Deltas aff is actually diminished after multiplication by s We find that c \Deltax aff Hence, we can have x \Deltax aff From (94) and (97), we conclude that the value of ff max defined by (14), with the calculated direction ( c \Deltax aff \Delta- aff \Deltas aff replacing the exact search direction, satisfies the estimate Note from (89), (90), and (96) that, in an absolute sense, the errors in c \Delta- aff c \Deltas aff , and c \Deltax aff are small. By contrast, the O(- \Gamma1 u) term in (95) implies that the errors in c \Deltax aff may become large as - # 0. These large errors may in turn cause the residuals r b to grow as - # 0. These expectations are confirmed by the computational experiments of Section 6. The estimate (98) and the parameter choice suggest strongly that the algorithm should be terminated when - When - reaches this threshold, all three terms in the estimate (98) are in balance. Below this threshold, the O(- \Gamma1 u) term in c \Deltax aff may cause r b to grow, making further reduction of - counterproductive. The convergence tolerances used by most interior-point codes-arrived at by practical experience rather than any theoretical considerations-are similar to (99). The code PCx is typical. It declares optimality if the following three conditions are satisfied: where the default value of tol is 10 \Gamma8 . (Note that 10 \Gamma8 - u 1=2 in double precision arithmetic on most machines.) 5.2. Mehrotra Predictor-Corrector Steps. Having analyzed the affine-scaling search direction and its calculated approximation, we turn our attention briefly to the search direction used by Mehrotra's predictor-corrector algorithm. As mentioned in Section 2, these steps are obtained by setting r xs as in (12), for some heuristic choice of the centering parameter i. We can write the search direction as where (\Deltax cc ; \Delta- cc ; \Deltas cc ) is the "corrector-centering" step component that satisfies the following linear system:4 0 A T I \Delta- cc \Deltas cc5 =4 Block elimination on this system yields the following special case of (16a): Since we assume full rank of A, and since the diagonal elements of D are all strictly positive, the coefficient matrix is invertible, and we have A result of Stewart [9] and Todd [11] states that the norm k(AD 2 A T ) bounded independently of D over the set of all positive definite diagonal matrices D (and therefore independently of x and s with (x; s) ? 0). Therefore, we have From (72), we have kX while from (78), it follows that k\DeltaX aff \DeltaS aff O(- 2 ). Hence, we have A typical heuristic for choosing the centering parameter i is to set where - aff is the value of - that results from a full step-to-boundary ff max along the affine-scaling direction. If the search direction is exact, we have - this heuristic yields Use of the calculated direction ( c \Deltax aff \Delta- aff \Deltas aff together with the estimate (98) leads us to expect - case too, provided that - u 1=2 . Hence, we have from (101) that k\Delta- cc from (100) and (79), we have where \Delta- is the - component of the Mehrotra search direction. We also can apply the Stewart-Todd result to formulae for \Deltax cc and \Deltas cc to show that k(\Deltax cc ; \Deltas cc Therefore, we have corresponding to (78). Because of the estimates (102) and (103), the analysis of the preceding subsection can be applied without modification to the calculated version of the search direction (100). In particular, if we redefine the step-to-boundary ff max in terms of this calculated \Deltax; c \Delta-; c \Deltas), we find that the estimate (98) still applies. We conclude that near-unit steps can still be taken along this direction provided that - u 1=2 . 6. Implementation and Computational Results. Most interior-point codes use modified Cholesky algorithms with essentially the same properties as Algorithm modchol. They differ slightly, however, in the implementation. The IPMOS code of Xu, Hung, and Ye [16] replaces small pivot elements by 1 and fills out the corresponding column of the Cholesky factor with zeros and also inserts a zero in the right-hand side. The criterion for identifying a small pivot is not explained in the reference [16], but otherwise this strategy is equivalent to Algorithm modchol. Zhang's LIPSOL code [17] and the PCx code of Czyzyk, Mehrotra, and Wright [1] replace small pivots by a huge number-10 128 -but otherwise leave the Cholesky algorithm unchanged. The net effect is, however, almost equivalent to Algorithm modchol and the triangular substitution procedure (25). The advantage of this approach is that it involves minimal changes to a standard sparse Cholesky code. We need only add a loop to calculate the largest diagonal element fi, and a small pivot check immediately before the point at which the computation L M ii is performed. To test that the analysis of this paper was reflected in practical computations, we coded a primal-dual algorithm that used Algorithm modchol in conjunction with the formulation (16). The code was used to solve some small random linear programs in which the amount of degeneracy-the composition of index sets B and N -was carefully controlled. At each iterate, we monitored various quantities and compared them against the estimates of Section 5. The linear programming test problems were posed in standard form (2) with 12. The matrix A is fully dense, with elements are random variables drawn from a uniform distribution on the interval [0; 1]. (Of course, the values of - 1 and - 2 are different for each element of the matrix.) We can reasonably expect this matrix A to satisfy the well-conditioning assumptions of Section 5. The user specifies the number of indices to appear in B, and we set A primal solution x is constructed with x where - is randomly drawn from the uniform distribution on [0; 1]. We choose the dual solution - to be the vector (1; fix an optimal dual slack vector s to be s where - is random as above. Finally, we set The code was an implementation of the infeasible-interior-point algorithm described by Wright [13]. The details of this algorithm are unimportant; we need note only that its iterates satisfy the estimates (72) in exact arithmetic and that the algorithm takes steps along the affine scaling direction during its later iterations. At each iteration of the algorithm, we calculated the affine scaling direction (whether or not it was actually used as a search direction) and printed the norms k c \Deltax aff k1 , k c \Delta- aff k1 , and k c \Deltas aff k1 alongside the duality measure - and residual norm k(r b ; r c )k 1 for the current point. We also kept track of the number of small pivots encountered during the factorization, that is, the number of elements in J . The parameter ffl was set to \Gamma12 , which is about 100u on the SPARCstation 5 that was used for the experiments. The results were not particularly sensitive to this parameter. Results are shown in Tables 1-4. For each iteration of the algorithm, these tables list the number of small pivots jJ j, the base-10 logarithms of -, k(r the affine-scaling step norms mentioned above. The step-to-boundary ff max along the calculated affine-scaling direction is also tabulated. A horizontal line in each table indicates the iterate at which termination occurs according to the criterion (99). In Table 1 we chose making the linear program nondegenerate and the primal-dual solution unique. It is clear that c \Delta- aff and c \Deltas aff satisfy the estimates (88) and (90), respectively, even when the algorithm is continues past the point of normal termination. The component c \Deltax aff , on the other hand, clearly shows the influence of the O(- \Gamma1 u) error term in (95) when - becomes comparable to or smaller than u. Note, too, that the error in c \Deltax aff is transmitted to the residual r b on succeeding iterations but that this effect does not become destructive until - is much smaller than its normal termination threshold. The values of ff max are also consistent with the estimate (98). This step length approaches 1 until the normal point of termination is reached, after which the errors in c \Deltax aff and r b make further progress impossible. Table 2 shows the interesting case in which we choose 4, so that the co-efficient matrix in (16a) has four singular values of and two of -). The second column shows that Algorithm modchol correctly identifies the numerical rank during the last few iterations and that the interior-point algorithm continues to generate useful steps and to make good progress even after encounters small pivots. Apart from this feature, the behavior is the same as in Table 1, with errors in c \Deltax aff causing the interior-point algorithm to behave poorly when it is permitted to run past its normal point of termination. We noted that for all iterations, the "small" pivots were at the bottom right corner of the so that (28) rather than the general estimate (27) applies to the perturbation matrix E. In this case, we can replace -ffl 1=2 by -ffl in estimates of Section 5 such as (93), (95), and (98). Table 3 illustrates another case in which 4, with the added complication that A is rank deficient. (We forced rank deficiency by setting A so that the first and second rows each contain a single nonzero in their last column.) The (2; 2) pivot is skipped at every invocation of Algorithm modchol. As - becomes small, the final pivot is skipped as well, and the numerical rank is correctly determined. Since the small pivots are not localized in the bottom right corner, the special bound (28) does not apply, so we cannot strengthen the bounds on the step components as in the previous paragraph. The computational behavior is qualitatively the same as in Tables 1 and 2. Table 4 illustrates a problem for which 8. Here, the coefficient matrices retain full numerical rank at all iterates, and the behavior is similar to that reported in Table 1. One point of difference is that the errors in c \Deltax aff , which start to increase after iteration 19, do not have an immediate effect on the residual r b . The reason is simply that this particular interior-point algorithm chose to take a path-following step at iterations 21 and 22 rather than the affine scaling step, and the \Deltax components were calculated accurately in the path following step. An affine-scaling step is, however, taken at iteration 28, and the effect of the error in c \Deltax aff on the residual r b at the following iterate is obvious. --R Technical Report OTC 96/01 Stability of symmetric ill-conditioned systems arising in interior methods for constrained optimization Computer Solution of Large Sparse Positive Definite Systems Accuracy and Stability of Numerical Algorithms Solving Least Squares Problems Computational experience with a primal-dual interior point method for linear programming Block sparse Cholesky algorithms on advanced uniprocessor com- puters On scaled projections and pseudoinverses Matrix Perturbation Theory A Dantzig-Wolfe-like variant of Karmarkar's interior-point linear programming algorithm Some properties of the Hessian of the logarithmic barrier function A simplified homogeneous and self-dual linear programming algorithm and its implementation Solving large-scale linear programs by interior-point methods under the MATLAB enviroment --TR --CTR Francis R. Bach , Michael I. Jordan, Kernel independent component analysis, The Journal of Machine Learning Research, 3, p.1-48, 3/1/2003
error analysis;interior-point algorithms and software;cholesky factorization;matrix perturbations
589182
Two-Step Algorithms for Nonlinear Optimization with Structured Applications.
In this paper we propose extensions to trust-region algorithms in which the classical step is augmented with a second step that we insist yields a decrease in the value of the objective function. The classical convergence theory for trust-region algorithms is adapted to this class of two-step algorithms. The algorithms can be applied to any problem with whose contribution to the objective function is a known functional form. In the nonlinear programming package LANCELOT, they have been applied to update slack variables and variables introduced to solve minimax problems, leading to enhanced optimization efficiency. Extensive numerical results are presented to show the effectiveness of these techniques.
Introduction In nonlinear optimization problems with expensive function and gradient evaluations, it is desirable to extract as much improvement as possible at each iteration of an algorithm. When the objective function contains a subset of variables that occurs in a predictable functional form, a second, computationally relatively inexpensive, update can be applied to these variables following a classical optimization step. The additional step provides a further reduction in the objective function and can lead to superior optimization e-ciency. The two-step algorithms have been successfully applied to the updating of slack variables and to a particular formulation of minimax problems, as is indicated by numerical results on a variety of problems. In these instances a subset of variables (slack variables and variables introduced to solve minimax problems) appears in a xed, known algebraic form in the objective function. However, since it can be applied to any problem where a subset of the variables can be optimized relatively cheaply compared with the cost of evaluating the entire function (for example if some terms require simulation and other independent terms are Department of Mathematical Sciences, IBM T. J. Watson Research Center, Route 134 and Taconic, Room 33-206, Yorktown Heights NY 10598. y Departamento de Matematica, Universidade de Coimbra, 3000 Coimbra, Portugal. This work started when this author was visiting the IBM T. J. Watson Research Center at Yorktown Heights and was supported in part by Centro de Matematica da Universidade de Coimbra, Instituto de Telecomunicac~oes, FCT, Praxis XXI 2/2.1/MAT/346/94, and IBM Portugal. z Computer Architecture and Design Automation, IBM T. J. Watson Research Center, Route 134 and Taconic, Room 33-156, Yorktown Heights NY 10598. available analytically), their applicability is really rather broad. We propose modications to existing nonlinear optimization algorithms. An alternative approach, when feasible, is to reformulate the original problem by eliminating a subset of variables and then to apply the algorithms in the remaining variables (see, for example, Golub and Pereyra [17]). This paper deals with two-step algorithms where the second step is required to yield a decrease in the value of the objective function. The analysis given here covers the global convergence of two-step trust-region algorithms and it is presented for the unconstrained minimization problem: continuously dierentiable function. For both trust regions and line searches, one can consider two versions of the two-step algorithms, one called greedy and the other called conservative. The greedy version exploits as much as possible the decrease obtained by the second step, whereas the conservative approach calculates the second step only after the rst step has been conrmed to satisfy the traditional criteria required for global convergence. We point out that the conservative two-step line-search algorithm is not new and can be found in the books by Bertsekas [1], Section 1.3.1, and Luenberger [19], Section 7.10, where the second step is called a spacer step. A description of the greedy and conservative two-step line-search algorithms can be found in [11]. In trust regions, if the second step is guaranteed to decrease the value of the objective function, global convergence of the type lim inf k !+1 krf(y k immediately attained. Further, in the cases where the rst step would be rejected, the sum of the rst and second steps has a better chance of being accepted (see Remark 3.1). To obtain lim k !+1 krf(y k either the norm of the second step has to be controlled by the trust region (see condition (13)) or the decrease on the objective function attained by the second step has to be of the order of magnitude of the norm of this step (see condition (12)). The update of the slack variables referred to above motivated the study of the local rate of convergence of a two-step Newton's method. We show that a second Newton step in some of the variables retains the q-quadratic rate of convergence of the traditional Newton's method. This paper is structured as follows. In Section 2 we introduce the two-step trust-region al- gorithms, and in Section 3 we analyze their global convergence properties. The local rate of the two-step Newton's method is studied in Section 4. The application of the two-step ideas to update slack variables and variables introduced for the solution of minimax problems is described in Section 5. Section 6 presents the numerical results obtained with LANCELOT using these updates for analytic problems and dynamic-simulation-based and analytic static-timing-based circuit optimization problems. Finally, some conclusions are drawn in Section 7. Two-step trust-region algorithms We rst consider the trust-region framework presented in the paper by More [20] for unconstrained minimization. The (classical) trust-region algorithm builds a quadratic model of the form at the current point y k , where H k is an approximation to r 2 f(y k ) (note that m k (y k Then a step s k is computed by approximately solving the trust-region subproblem subject to ksk k ; where k is called the trust-region radius and kk is an arbitrary norm. The new point y is tested for acceptance. If the actual reduction f(y k larger than a given fraction of the predicted reduction m k (y k then the step s k and the new point y k+1 are accepted. In this situation, the quadratic model m k (y is considered to be a good approximation to the function f(y) in the region . The trust radius may be increased. Otherwise, the is considered not to be a good approximation to the function f(y) in the region . In this case, the new point y k+1 is rejected, and a new trust-region subproblem of the form (2) is solved for a smaller value of the trust radius. This simple trust-region algorithm is described below. Algorithm 2.1 (Trust-region algorithm) 1. Given y 0 , the value f(y 0 ), the gradient rf(y 0 ) and an approximation H 0 to the Hessian of f at y 0 , and the initial trust-region radius 0 . Set and in (0; 1). 2. Compute a step s k based on the trust-region problem (2). 3. Compute 4. In the case where set compute H k+1 , and select k+1 satisfying k+1 k . Otherwise, set 5. Increment k by one and go to Step 2. The mechanism used to update the trust radius that is described in Algorithm 2.1 is simple and su-ces to prove convergence results. In practice, with the goal of improving optimization e-ciency, one uses updating schemes that are more complex involving several subcases according to the value of k . We propose in this paper a modication of this trust-region algorithm. We are motivated by a situation where it is desirable to update slack variables and variables introduced to solve minimax problems, at every iteration of the trust-region algorithm [7] implemented in LANCELOT [9]. See Section 5 for more details on practical applications. The two-step trust-region algorithm is quite easy to describe. Suppose that after computing a step s k based on the trust-region subproblem (2) we know some properties of the function f(y) that enables us to compute a new step ^ s k for which we can guarantee that f(y k In this situation we would certainly like to have y and to test whether this new point should be accepted or not. This modication requires a careful redenition of the actual and predicted reductions given for Algorithm 2.1. The new actual and predicted reductions that we propose are: ared(y pred(y The new predicted reduction is the predicted reduction obtained by the rst step plus the (actual) reduction obtained by the second step. The choice pred(y k ; s not appropriate since the second step ^ s k is not computed using the model m k (y k The two-step trust-region algorithm is given below. Algorithm 2.2 (Two-step trust-region algorithm { Greedy) 1. Same as in Algorithm 2.1. 2. Compute a step s k based on the trust-region problem (2). 3. If possible, nd another step ^ s k such that 4. Compute pred(y 5. In the case where set compute H k+1 , and select k+1 satisfying k+1 k . Otherwise, set 6. Increment k by one and go to Step 2. The two-step trust-region Algorithm 2.2 evaluates the new point y acceptance after both steps s k and ^ s k have been computed. We call this version \greedy" because it tries to take as much advantage as possible of the decrease obtained by the second step ^ s k . Note that although the function f is evaluated twice in Algorithm 2.2, the reevaluation is often computationally inexpensive. The context in which we are particularly interested involves relatively expensive evaluations at y k and evaluations at y k involving only a subset of the variables that are cheap to compute (see Section 5). We could also consider a two-step trust-region algorithm where rst an acceptable step s k is determined, and only afterwards a second step ^ s k is computed. This algorithm is outlined below. Algorithm 2.3 (Two-step trust-region algorithm { Conservative) 1. Same as in Algorithm 2.1. 2. Repeat (a) Compute a step s k based on the trust-region problem (2). (b) Compute (c) If k > , then set compute k+1 satisfying k+1 k , and set accepted = true. If k , set k and accepted = false. Until accepted. 3. If possible, nd another step ^ s k such that 4. Set y s k . 5. Update H k . Increment k by one and go to Step 2. The same comments about the function evaluations apply to Algorithm 2.3 after the computation of a successful step s k . However, in the case of Algorithm 2.3, the function f has to be evaluated twice only in iterations corresponding to successful rst steps s k . 3 Global convergence of the two-step trust-region algorithms We analyze rst the two-step trust-region Algorithm 2.2, i.e., the greedy version. The analysis for the conservative Algorithm 2.3 is similar. In this section we make the assumption that fH k g is a bounded sequence. So, there exists a > 0 for which We require the step s k to satisfy a fraction of Cauchy decrease on the trust-region problem (2). In other words we ask s k to satisfy for 2 (0; 1]. The step c k is called the Cauchy step, and it is dened as the solution of the scalar problem in the unknown subject to ksk k ; There is a variety of algorithms that compute steps satisfying this condition (see [3], [22], [23], [25], and [26]). Proposition 3.1 If a fraction of Cauchy decrease then: krf(y k )k where and are as in (6) and (5) respectively. Proof: See Powell [24], Theorem 4, or More [20], Lemma 4.8. 2 If we use this proposition and the fact that f(y k pred(y krf(y k )k krf(y k )k This inequality is crucial to prove global convergence of the two-step algorithm. In particular, if the iteration k is successful, then ared(y We are ready to prove the rst convergence result. Theorem 3.1 Consider a sequence fy k g generated by Algorithm 2.2 where s k satises (6). If f is continuously dierentiable and bounded below on and fH k g is a bounded sequence, then lim inf krf(y So, if the sequence fy k g is bounded, there exists at least one limit point y for which rf(y Proof: The proof is similar to the proof given in [20], Theorem 4.10. Assume by contradiction that fkrf(y k )kg is bounded away from zero, i.e., that there exists an > 0 such that krf(y k )k for all k. As in [20], Theorem 4.10, we make direct use of (9) and of the rules that update the trust radius, to obtain: and so lim k !+1 The next step is to show that lim k Note that from the denitions (3) and (4), we have ared(y which in turn, by using a Taylor series expansion and ks k k k , implies This inequality and (8) show that converges to zero. The rest of the proof follows a classical argument in trust regions: if ^ k converges to one, the rules that update the trust radius show that k cannot converge to zero. So, a contradiction is attained and the proof is completed. 2 The result of Theorem 3.1 does not require the step ^ s k to be O( k ) which may seem surprising. This result shows the appropriateness of the denitions given in (3) and (4) for the actual and predicted reductions. These denitions allow us to obtain the conditions (9) and (11) that are crucial to establish (10). Remark 3.1 It is also important to note that the denitions (3) and (4) can improve the acceptability of a step. In fact, we have before. We now note that ^ k and the function ^ strictly increasing if k < 1. In other words, in cases where a standard trust-region algorithm rejects a step the modied criterion is always better than the usual one. Further, it can be noted that ^ which indicates that all successful iterations of the the standard algorithm will also be successful in the modied two-step algorithm. In particular, The next step in the analysis is to prove that, with additional conditions on the second step, Theorem 3.2 Consider a sequence fy k g generated by Algorithm 2.2 where that f is continuously dierentiable and bounded below on L(y 0 ) and that fH k g is a bounded sequence. If rf is uniformly continuous on L(y 0 ) and if either or are positive constants independent of k, then lim krf(y So, if the sequence fy k g is bounded, every limit point y satises rf(y Proof: The proof is similar to the proof given in [20], Theorem 4.14. See also Thomas [27]. We show the result by contradiction. Assume therefore that there exists an 1 2 (0; 1) and a subsequence indexed by fm i g of successful iterates such that, for all m i in this subsequence, Theorem 3.1 guarantees the existence of another subsequence indexed by fl i g such that krf(y k )k 2 , for m i k < l i and krf(y l i )k < 2 (where fm i g is without loss of generality the subsequence previously mentioned). Here 2 is any real number chosen to be in converges to zero, for k su-ciently large corresponding to successful iterations holds if (12) is satised, and holds otherwise with . We consider the cases (12) and (13) separately. In both cases we make use of: In the sums we consider only indices corresponding to successful iterations. If (12) holds then we use (15) to obtain [ks If (13) holds then we appeal to (16) and write [ks In either case we obtain and since the right hand side of this inequality goes to zero, so does the left hand side Since the gradient of f is uniformly continuous, we have for i su-ciently large that can be any number in (0; 1 ) this inequality contradicts the supposition. 2 In the theorem above we required the norm of the step ^ s k to either be O( k ) or O (f(y k )). The former condition can be enforced in Step 2 of the Algorithm 2.2, although this might not be benecial and could lead to an inferior decrease. We can obtain global convergence to a point that also satises the necessary second-order conditions for optimality. For this purpose, we require the step s k to satisfy a fraction of optimal decrease for the trust-region problem (2). In other words we ask s k to satisfy where 2 (0; 1], and s k is an optimal solution of (2). (This condition can be weakened in several ways [20].) A step s k satisfying a fraction of optimal decrease can be computed by using the algorithms proposed in [22] and [25] in the case where the trust-region norm is Euclidean. The global convergence result is the following. Theorem 3.3 Consider a sequence fy k g generated by Algorithm 2.2 where H satises (17). If L(y 0 ) is compact and f is twice continuously dierentiable on L(y 0 ), then there exists at least one limit point y for which rf(y positive semi-denite. Proof: The proof is basically the same as the proof of Theorem 4.7 in [22]. 2 To obtain stronger global convergence results to second-order points, for instance the results in Theorems 4.11 and 4.13 in [22] (see also [21], Theorem 4.17, c and d), other conditions are required like k^s k k being of O( k ). The next results show that the second step can preserve the nice local properties of the behavior of the trust radius that are typical in trust-region algorithms. Theorem 3.4 Let fy k g be a sequence generated by Algorithm 2.2 where In addition, assume that the step ^ s k satises either condition (12) or condition (13). If f is twice continuously dierentiable and bounded below on L(y 0 ) and fy k g has a limit point y such that H positive denite, then fy k g converges to y , all iterations are eventually successful, and f k g is bounded away from zero. Proof: From Theorem 3.2 we can guarantee that lim k !+1 krf(y k the proof is basically the same as the proof of Theorem 4.19 in [20]. 2 An alternative to this result where we do not impose conditions (12) or (13) on the second step is given below. However we need to assume that fy k g converges to y . Theorem 3.5 Let fy k g be a sequence generated by Algorithm 2.2 where continuously dierentiable on L(y 0 ) and fy k g converges to a point y such that H positive denite, then all iterations are eventually successful and f k g is bounded away from zero. Proof: The rst step s k yields a decrease in the quadratic model: Thus, the assumptions made on H k and H guarantee ks for su-ciently large k, which in turn, by using (8), implies pred(y ks k (The constants c 3 and c 4 are independent of k.) A Taylor series expansion for the expression (11) gives The fact that fy k g converges and the result lim inf k !+1 krf(y k Theorem 3.1, together imply lim k !+1 krf(y k Thus, from (18) we get lim k !+1 ks k The proof is terminated with a typical argument in trust regions. From (19), (20) and lim k !+1 ks we obtain the limit lim pred(y which shows, by appealing to the rules that update the trust radius, that all iterations are eventually successful and the trust radius is uniformly bounded away from zero. 2 The global convergence analysis for Algorithm 2.3 is identical to the analysis given above for Algorithm 2.2. We point out that Algorithm 2.3 is well dened since at a nonstationary point it is always possible to nd an acceptable rst step. Also, for every k, krf(y k )k krf(y k )k Thus, the results given in Theorems 3.1-3.5 hold for Algorithm 2.3. The lim inf-type result (10) is obtained under the classical assumptions for trust-region algorithms for unconstrained optimization. To obtain the lim-type result (14) one of the two conditions (12) and (13) is required. In the case of the applications considered in Section 5, the decrease obtained by the second step s k is always guaranteed to satisfy Moreover, the objective function strictly decreases along the segment between the points y k and s k . In this case we can modify Step 3 of Algorithms 2.2 and 2.3 in such a way that we meet the requirements of Theorem 3.2. This modication is given below. It is easy to verify that either (12) or (13). Algorithm 3.1 (Step 3 for Algorithms 2.2 and 2.3 { Quadratic decrease case) 3. Compute a step ^ s k such that g so that k^s k k c 2 k and ^ s k is not enlarged. (Otherwise (12) holds with c The positive parameters and c 2 should be set a priori in Step 1 of Algorithms 2.2 and 2.3. Of course, we would like to prove the result of Theorem 3.2 for the case where the condition (12) is replaced by the condition (21). However, such a result is unlikely to be true. 4 Local rate of convergence of a two-step Newton's method In the next section we are interested in two-step algorithms where the second step is calculated as a Newton-type step in some of the variables. In this section we investigate the local rate of convergence for an algorithm where each step is composed of two Newton steps, the second being computed only for a subset of the variables. For this purpose let x Suppose the rst step s k is a full Newton step, i.e., s At the intermediate point y k , a Newton step is applied in the variables u with x k xed. This two-step Newton's method is described below. Algorithm 4.1 (Two-step Newton's method) 1. Choose y 0 . 2. For do 2.1 Compute s k . 2.2 Compute and set s s k . 2.3 Set y The proof of the local convergence rate of the two-step Newton's method requires a few modications from the standard proof of Newton's method [12], Theorem 5.2.1. Recall that that proof of Newton's method is by induction. Corollary 4.1 Let f be twice continuously dierentiable in an open set D where the second partial derivatives are Lipschitz continuous. If fy k g is a sequence generated by Algorithm 4.1 converging to a point y 2 D for which rf(y positive denite, then fy k g converges with a q-quadratic rate. Proof: If y k is su-ciently close to y , the perturbation result [12], Theorem 3.1.4, can be used to prove the nonsingularity of the Hessian matrix r 2 f(y k ). Furthermore, Now we show that r 2 First we point out that r 2 uu f(y) is Lipschitz continuous on D and r 2 uu f(y ) is positive denite. Thus, inequality (22) and the perturbation lemma cited above, together imply the nonsingularity of r 2 Hence the method is locally well-dened, and the second step yields since r u f(y) is Lipschitz continuous near y . Now we use inequalities (22) and (23), and write This last inequality establishes the q-quadratic rate of convergence. 2 Applications We begin by considering updating the slack variables in LANCELOT. Suppose the problem we are trying to solve has the form minimize f(x) subject to c i (x) are positive integers. The technique implemented in the LANCELOT package [9] is the augmented Lagrangian algorithm proposed by Conn, Gould, and Toint in [8]. For the application of the augmented Lagrangian algorithm this problem is reformulated as: minimize f(x) subject to c i (x) by adding the slack variables u i , m. This algorithm considers the following augmented Lagrangian merit function: where: i is an estimate for the Lagrange multiplier associated with the i-th constraint, is a (positive) penalty parameter, s ii is a (positive) scaling factor that is associated with the i-th constraint, and solves a sequence of minimization problems with simple bounds of the following subject to u for xed values of , s ii , and i , m. The two-step trust-region framework and analysis described in this paper for unconstrained minimization problems can be extended in an entirely straightforward way to a number of algorithms for minimization problems with simple bounds, in particular to the algorithms [7] used by LANCELOT to solve problem (25). If x is xed, the function (x; in the slack variables u. Let us denote this quadratic by q(u; x): where d(x) and e(x) depend on x but F is constant. (The dependency on i , s ii , and is not important since these are constants xed before the minimization process is started.) The key idea is to update these slack variables at every iteration k of the trust-region algorithm [7] that is used in LANCELOT to solve problem (25). The trust-region algorithm computes, at the current point y k , a rst step s k . Now, at the new point y s k we compute the step ^ s k by updating the slack variables u. So, we have where (Here f represents the objective function of Sections 1-4.) Note that the second step ^ s k is exclusively in the components associated with slack variables. This step is computed as u k+1 is the optimal solution of subject to u Due to the simple form of this quadratic, the solution is explicit: s ii It is important to remark that these updates require no further function or gradient evaluations. They have also been considered in the codes NPSOL and SNOPT [15], [16] to update slack variables after the application of a line search to the augmented Lagrangian merit function and prior to the solution of the next quadratic programming problem. Other ways of dealing with slack variables have been studied in the literature (see Gould [18] and the references therein). For the study of the impact of the slack variable update on the global convergence of the trust-region algorithm, the step in these variables is required only to decrease the quadratic q(u; u k to u k +u k . In such a case, we can always guarantee that the decrease in the objective function is larger than k^s k k 2 , that is that (21) holds. This result is shown in the following proposition. We drop x k from q( ; x k ) to simplify the notation. Proposition 5.1 There exists a positive constant c 5 such that, whenever q(u k +u k ) < q(u k ), we have Proof: First we write down a few properties of the quadratic q(u). Simple algebraic manipulations lead to: Also, since q(u) is convex: Let c be a positive constant such that c < min is the smallest eigenvalue of F . Now we consider two cases. 1. cku k k 2 . In this case we use (29), to obtain 2. . In this case we appeal to (28) and to get min The proof is completed by setting c cg. 2 Another example of the application of two-step algorithms arises in one approach to the solution of minimax problems. Consider the following where each f i is a real-valued function dened in IR n . One way of solving this minimax problem is to reformulate it as a nonlinear programming problem by adding an articial variable z. See [18] for more details. This leads to minimize z subject to z f i (x) where the slack variables have also been introduced. If LANCELOT is used to solve this nonlinear programming problem, then the augmented Lagrangian algorithm requires the solution of a sequence of problems with simple bounds of the type: subject to u where In this situation the function (x; z; in the variables u and z for xed values of x. (Again, , S, and are constants and not variables for problem (32).) The application of the two-step trust-region algorithm follows in a similar way. The Hessian of the quadratic is positive semi-denite with the following form where the last row and the last column correspond to the variable z. The solution of the quadratic program minimize q(z; u; subject to u is given by s ii where z k+1 is the solution of the equation s ii s ii with right hand side s ii The equation (35) is solved easily with O(m) oating point operations and comparisons, showing that the solution of the quadratic program (33) is a relatively inexpensive calculation. There are several nonlinear optimization problems in which some subset of the problem variables occur linearly, for example, arrival times in static-timing-based circuit optimization problems [6]. Such problems can also benet from two-step updating. 6 Numerical tests 6.1 Analytic problems We modied LANCELOT (Release A) [9] to include the slack variable update (27) and the slack and minimax variable updates (34)-(36). These updates were incorporated in LANCELOT using a greedy two-step modication of the trust-region algorithm [7] for minimization problems with simple bounds that is implemented in the subroutine SBMIN. (The greedy two-step trust-region algorithm for unconstrained minimization problems is Algorithm 2.2.) We tested the following versions of LANCELOT: 1. LANCELOT (Release A) with the default parameter conguration SPEC.SPC le, except that we increased the maximum number of iterations to 4000. 2. Version 1 with the slack and minimax variable updates (27) and (34)-(36) incorporated in SBMIN using a greedy two-step trust-region algorithm. 3. The same as Version 2 but with no update of the variable z for minimax problems, i.e., z xed in (34)-(36). We compared the numerical performance of these three versions on a set of problems 1 from the CUTE collection [2]. This set of problems is listed in Table 1, and in the case of minimax formulations in Table 2, where we mention the number of variables (including slacks and, where applicable, the minimax variable z), the number of slack variables, and the number of equality and inequality constraints (excluding simple bounds on the variables). Note that the minimax problems were reformulated as nonlinear programming problems by the introduction of an additional minimax variable z as shown above (31). The computational results are presented in Tables 3, 4, and 5. All tests were conducted on an IBM Risc/System 6000 model 390 workstation. In Table 3 we compare the results of Versions 1 and 2 for problems that are not minimax problems. In Table 4 we present the results of Versions 1 and 2 for minimax problems. In Table 5 we include the results of Versions 1 and 3 for minimax problems. In Tables 4 and 5 we include the majority of the minimax problems but not all (see Section 6.3 for numerical results on the remaining problems). In these tables we report the value of the ag INFORM, the number of iterations, the total CPU time, and the determined values (a single value if they are both the same) of the objective function. The values of INFORM have the following meaning: meaning that the norm of the projected gradient of the augmented Lagrangian function has become smaller than 10 5 . cases where the maximum number of iterations (4000) has been reached. cases where the norm of the step has become too small. Our conclusion based on these sets of problems is that the version with the slack and minimax variable updates exhibits superior numerical behavior. In fact, this version required an average of 15% fewer iterations than the version without these updates (the problems HS109, HAIFAM, and POLAK6 were excluded from this calculation, mainly because the comparison was extraordinarily favorable in the case of the rst two and worse in the last). Comparing Tables 4 and 5, updating the variable z in addition to two-step updates on just the slacks is seen to yield a signicant benet. However, there are some minimax problems where the two-step algorithm performs poorly and this situation is analyzed in detail in Section 6.3. Although CUTE contains more than 56 problems with general constraints the majority of these are equality constrained problems. We excluded all problems that took more than 4000 iterations with both Versions 1 and 2. We included the rest, with the exception of some problems that are too easy, making a total of 56 problems of which are minimax problems and 26 are non-minimax problems. Problem Name Variables Slacks Constraints CORE1 CORE2 157 26 134 CORKSCRW HADAMARD 769 512 648 HS85 26 21 21 Table 1: Non-minimax problems from the CUTE collection that were used. 6.2 Circuit optimization problems We have built extensive experience with circuit optimization problems, where { due to expensive function evaluations, modest numerical noise levels, and practical stopping criteria { the implementation is designed to terminate before many \asymptotic" iterations are taken. The algorithms described in this paper have been used in a dynamic-simulation-based circuit optimization tool called JiyTune (see [4], [5], and [10]). JiyTune optimizes transistor and wire sizes of digital integrated circuits to meet delay, power, and area goals. It is based on fast circuit simulation and time-domain sensitivity computation in SPECS (see [13] and [28]). To optimize multiple path delays through a high-performance circuit, the tuning is often formulated as a minimax problem or a minimization problem with nonlinear inequality constraints. We remark that many of the analytic problems (especially the minimax problems) are rather small and involve inexpensive function evaluations. Moreover, it is clear that two-step updating is unlikely to be helpful asymptotically in these situations. Consequently we also report numerical results with circuit optimization problems which are indicative of problems with expensive function evaluations, where termination (because of inherent noise and practical considerations) is encouraged to be before any signicant asymptotic behavior. The numerical results are presented in Table 6. As in Version 1, the second step consisted of the slack and minimax variable updates (27) and (34)-(36). However the gradient and constraint tolerances used were 10 respectively, Problem Name Variables Slacks Constraints COSHFUN 81 20 20 GOFFIN 101 50 50 HAIFAL 9301 8958 8958 HALDMADS 48 42 42 MINMAXBD Table 2: Minimax problems from the CUTE collection that were used. with some safeguards related to an expected level of numerical noise. We can clearly observe from Table 6 that the two-step algorithm leads to better nal objective function values. In practical applications where a simple function evaluation takes more than ten minutes of CPU time the eectiveness of such a simple addition is indeed signicant. (There are situations where the greedy two-step trust-region algorithm is able to take advantage of the decrease given by the slack and variable updates and, by doing so, this algorithm can accept steps that otherwise would have been rejected, see Remark 3.1.) We also applied the algorithms of this paper to analytic static-timing-based circuit optimization problems (see Table 7), where it is clear that the advantage of the two-step approach is increasingly apparent for larger problems. Problem Name Inform Iterations Total CPU Obj. Function CORE1 0/0 953/983 7.41/17 91.1 CORE2 0/0 1048/1086 25.6/25.7 72.9 CORKSCRW 0/0 41/42 0.55/0.54 1.16 HADAMARD 0/0 1709/548 2290/276 1.14/1 TFI3 0/0 23/34 0.38/0.38 4.3 Table 3: Comparison between Versions 1 and 2 for non-minimax problems (LANCELOT with- out/with two-step updating). 6.3 Further experiments with minimax problems In this section we consider those minimax problems in our test set for which the two-step algorithm not only does not improve numerically the results obtained in the one-step case, but also makes them considerably worse (see the rst part of Table 8). We analyze the reasons for the failure of the two-step updating on some minimax problems and discuss a few ways to enforce better numerical behavior. We consider the general minimax problem (30). Our aim is to show that for some types of problems the second step has a tendency to make the Hessian of ill-conditioned. Let us assume that (as happens by default for the rst LANCELOT major iteration). Under these circumstances, we have: By using the notation g i (x; z; we have the following expressions for the elements TWO-STEP ALGORITHMS FOR NONLINEAR OPTIMIZATION 20 Problem Name Inform Iterations Total CPU Obj. Function CONGIGMZ 0/0 32/19 0.04/0.05 28 COSHFUN 0/0 127/69 1.31/1.06 -0.773 GOFFIN 0/0 14/4 1.03/0.67 0 Table 4: Comparison between Versions 1 and 2 for minimax problems (LANCELOT without/with two-step updating). of the gradient of : r z r Similarly the elements of the Hessian matrix of are given by: for If the magnitudes of the products r 2 are small compared to those of the products r x then the Hessian of is given Problem Name Inform Iterations Total CPU Obj. Function CONGIGMZ 0/0 32/25 0.04/0.1 28 COSHFUN 0/0 127/92 1.31/1.08 -0.773 GOFFIN 0/0 14/8 1.03/0.66 0 Table 5: Comparison of Versions 1 and 3 for minimax problems (LANCELOT without/with two-step updating only on slacks). approximately a i1 a a i1 a in a i1 a i a in a i a in a in i a in a i a in a and the indices i in the sums go from 1 to m. This matrix is clearly singular. In fact, the n 1-st row is the negative sum of the last m rows. Moreover, any of the rst n rows is a linear combination of the last m rows. As result of these observations, the Hessian (and the projected Hessian) of is ill-conditioned if (37) happens for \many" indices j and k. This is the key point in this analysis: the second step has a tendency to produce iterates that worsen property (37) because it produces a decrease on the Problem Name Variables Ineq. Iterations Total CPU Obj. Function IOmuxpower 102 42 21/29 7230/9220 -15100/-16000 coulman cold 33 17 22/22 69.5/68.3 271/262 clkgen 22 5 25/5 35/10.8 1.98/1.82 coulman hot 33 17 16/32 46.2/100 283/253 coulman delay 33 17 26/24 72.6/73.5 116/111 Minimax: bultmann latch stall1 coulman cold minmax 34 17 61/80 184/229 69.4/66.9 coulman hot minmax 34 17 66/44 197/134 74.4/75.1 eischer northrop xor coulman delay minmax 34 17 100/100 290/306 67.4/70.5 Table updating for dynamic-simulation-based circuit optimization problems. Ineq. stands for the number of inequality constraints. values of g i (x; z; u) for some indices i. The Hessian of might very well be ill-conditioned if no second steps are applied, but there is no doubt (and the numerical results are a evidence of this claim) that the second step for some problems worsens the situation by making the Hessian of more ill-conditioned. In the presence of nonzero Lagrange multipliers i , m, the formulae for the gradient and Hessian of are the same with g i (x; z; u) substituted by g i (x; z; u)+ i and similar conclusions could be drawn. The second step may produce very bad results on some minimax problems because it points towards the set f(x; z; (where the Hessian of the augmented Lagrangian is ill-conditioned) and this eect in uences negatively the calculation of the rst step at the next iteration. Given this undesirable feature of the Hessian of at points close to this set, one possible improvement to the two-step algorithm is to make sure that the calculation of the rst step is accurate (in the LANCELOT context this could be achieved by choosing a smaller tolerance for the stopping criterion of the conjugate-gradient technique). Another possible improvement is to reduce the ill-conditioning of the Hessian of (for instance by increasing the value of the penalty parameter as can be seen in examples with a few variables). Indeed, these modications improve the bad numerical results presented before: in the second part of Table 8 we compare the results obtained by the following modications of Versions 1 and 2: 4. Version 1 with an initial value for the penalty parameter of 100 (the default value is 0:1). 5. Version 2 with an initial value for the penalty parameter of 100 and a tolerance of 10 12 in Problem Name Variables Ineq. Iterations Total CPU Obj. Function Symmetric 9 Table 7: LANCELOT without/with two-step updating for analytic (minimax) static-timing-based circuit optimization problems. Ineq. stands for the number of inequality constraints. the stopping criterion for conjugate gradients. The study of strategies that can make two-step updating more eective for minimax problems in general is the subject for future research. 7 Concluding remarks In this paper we presented and analyzed a framework under which classical algorithms for nonlinear optimization can be modied to allow second computationally e-cient steps that are not generated in the conventional way but that are guaranteed to yield decrease in the objective function. We gave as examples of the two-step algorithms the update of slack variables in LANCELOT, and the update of variables introduced to solve minimax problems. However, we emphasize that the two-step algorithms can be very generally applied, for example, whenever the functions dening the problem are in a known functional form in some of the variables. We considered trust-region algorithms for which we proposed a greedy and a conservative two-step algorithm. We analyzed the convergence properties of the trust-region two-step algorithms (see [11] for line-search two-step algorithms), deriving the conditions under which they attain global convergence. We also showed that a two-step Newton's method (for which the second step is computed only for a subset of the variables) has a q-quadratic rate of convergence. The greedy two-step algorithms are designed to exploit as much as possible the decrease attained by the second step. The trust-region framework allowed to us to design a greedy two-step trust-region algorithm that is particularly well tailored to achieve this purpose. Finally, we included numerical evidence that this technique is eective, particularly for problems with expensive function evaluations. The two-step algorithms have already found practical applications in optimization of high-performance custom microprocessor integrated circuits. Problem Name Inform Iterations Total CPU Obj. Function MINMAXBD 0/0 267/952 1.34/3.59 116 POLAK3 0/0 71/125 0.4/0.8 5.93 MINMAXBD 0/0 47/43 0.25/0.22 116 Table 8: In the rst part, comparison of Versions 1 and 2 for minimax problems (LANCELOT without/with two-step updating). In the second part, comparison of Versions 4 and 5 for minimax problems (LANCELOT without/with two-step updating). Acknowledgments We are grateful to N. I. M. Gould (Rutherford Appleton Laboratory) for his comments and suggestions on an earlier version of this paper that led to many improvements. We are also grateful to K. Scheinberg (IBM T. J. Watson Research Center) for helping with the numerical results and explanation in Section 6.3. We would like to thank I. M. Elfadel (IBM T. J. Watson Research Center) for providing the analytic static-timing-based optimization circuit problems. Finally, we are grateful to the referees for their useful comments and suggestions. --R Computer Science and Applied Mathematics CUTE: Constrained and Unconstrained Testing Environment Approximate solution of the trust-region problem by minimization over two-dimensional subspaces Optimization of custom MOS circuits by transistor sizing Global convergence of a class of trust-region algorithms for optimization problems with simple bounds Circuit optimization via adjoint Lagrangians Numerical Methods for Unconstrained Optimization and Nonlinear Equations Sensitivity computation in piecewise approximate circuit simulation Practical Methods of Optimization SNOPT: An SQP algorithm for large-scale constrained optimization User's guide for NPSOL 5.0: A Fortran package for nonlinear programming On solving three classes of nonlinear programming problems via simple di Linear and Nonlinear Programming A new algorithm for unconstrained optimization Minimization of a large-scale quadratic function subject to a spherical con- straint The conjugate gradient method and trust regions in large scale optimization Sequential Estimation Techniques for Quasi-Newton Algorithms Piecewise approximate circuit simulation --TR --CTR Tong Zhang, On the Dual Formulation of Regularized Linear Systems with Convex Risks, Machine Learning, v.46 n.1-3, p.91-129, 2002 Andreas Wchter , Chandu Visweswariah , Andrew R. Conn, Large-scale nonlinear optimization in circuit tuning, Future Generation Computer Systems, v.21 n.8, p.1251-1262, October 2005 Xiaoliang Bai , Chandu Visweswariah , Philip N. Strenski, Uncertainty-aware circuit optimization, Proceedings of the 39th conference on Design automation, June 10-14, 2002, New Orleans, Louisiana, USA Nicholas I. M. Gould , Dominique Orban , Philippe L. Toint, CUTEr and SifDec: A constrained and unconstrained testing environment, revisited, ACM Transactions on Mathematical Software (TOMS), v.29 n.4, p.373-394, December A. R. Conn , I. M. Elfadel , W. W. Molzen, Jr. , P. R. O'Brien , P. N. Strenski , C. Visweswariah , C. B. Whan, Gradient-based optimization of custom circuits using a static-timing formulation, Proceedings of the 36th ACM/IEEE conference on Design automation, p.452-459, June 21-25, 1999, New Orleans, Louisiana, United States Andrew R. Conn , Ruud A. Haring , Chandu Visweswariah, Noise considerations in circuit optimization, Proceedings of the 1998 IEEE/ACM international conference on Computer-aided design, p.220-227, November 08-12, 1998, San Jose, California, United States Andrew R. Conn , Chandu Visweswariah, Overview of continuous optimization advances and applications to circuit tuning, Proceedings of the 2001 international symposium on Physical design, p.74-81, April 01-04, 2001, Sonoma, California, United States
spacer steps;expensive function evaluations;LANCELOT;circuit optimization;trust regions;two-step algorithms;minimax problems;line searches;slack variables
589186
A Reflective Newton Method for Minimizing a Quadratic Function Subject to Bounds on some of the Variables.
We propose a new algorithm, a reflective Newton method, for the minimization of a quadratic function of many variables subject to upper and lower bounds on some of the variables. The method applies to a general (indefinite) quadratic function for which a local minimizer subject to bounds is required and is particularly suitable for the large-scale problem. Our new method exhibits strong convergence properties and global and second-order convergence and appears to have significant practical potential. Strictly feasible points are generated. We provide experimental results on moderately large and sparse problems based on both sparse Cholesky and preconditioned conjugate gradient linear solvers.
Introduction . In this paper we propose a new algorithm for solving the box- constrained quadratic programming problem The matrix H is symmetric and, in general, indefinite; l We denote the feasible region and the strict ug. When H is indefinite we are interested in locating a local minimizer. Problem (1.1) arises as a subproblem when minimizing general nonlinear functions subject to bounds and as a problem in its own right. The box-constrained quadratic programming problem represents an important class of optimization problems and has been the subject of considerable recent work (e.g., [1, 5, 12, 13, 16, 19, 21, 24, 26, 27, 32]). A special subclass deserves mention: the box-constrained least-squares problem, where A is a rectangular m-by-n matrix with m ? n. Our proposed algorithm can of course be applied to this special case if we form b. The determination of a version of our algorithm which does not involve the formation of the matrix H is an open question. We propose a new approach, a reflective Newton algorithm. The algorithm generates a sequence of strictly feasible iterates, fx k g, which converges under standard assumptions to a local solution of (1.1), x , at a quadratic convergence rate. Coleman and Li [10] establish theoretical properties of the reflective Newton approach applied to the general nonlinear box-constrained problem - as we indicate in Section 5 these results apply directly to the reflective Newton procedure proposed here for the quadratic minimization problem (1.1). In this paper we discuss the nature of the reflective transformation (Section 2); we discuss the reflective Newton approach as applied to problem with emphasis on a specialized line search to exploit the special structure of this problem (Section 3); numerical experiments involving an implementation of the reflective Newton method applied to box-constrained quadratic minimization (1.1) are discussed (Section 4). The sequence fx k g generated by the algorithm is strictly feasible: therefore, the algorithm can be regarded as an "interior-point" algorithm. However, this is a misleading classification. The algorithm differs markedly from methods commonly referred to as "interior-point" algorithms. For example, the proposed algorithm does not use a barrier function to ensure feasibility. The algorithm generates descent directions for q and then follows a piecewise linear path, reflecting off constraints as they are encoun- tered. Most interior point methods, on the other hand, generate descent directions (for some function) and then restrict the step, along this straight-line direction, to ensure feasibility. The algorithm most similar to our current proposal is probably the recent method due to Coleman and Hulbert [6]. (There is also a strong connection to previous work by Coleman and Li [7, 8, 9, 20] on various norm minimization problems.) Both are driven by the nonlinear system of equations representing first-order optimality condi- tions. Both methods require piecewise quadratic minimization. The methods differ in that our new algorithm is more general: positive definiteness of H is not required and it is not necessary to have finite upper and lower bounds on all the variables - the Coleman/Hulbert method requires both these restrictive properties. Finally, the Coleman/Hulbert method is an exterior-point method, requiring strict decrease in a piecewise quadratic "dual" function, whereas the new method generates feasible iter- ates, requiring strict decrease in the original quadratic function q. There are four key observations that underpin our new approach. First, it is possible to change the constrained problem (1.1) to an unconstrained problem without using a penalty parameter. We can replace (1.1) with an unconstrained problem, min q(y) where - q(y) is a continuous piecewise quadratic function of y, and y 2 R n is unrestricted. Details of this transformation are given in the next section including a result, Theorem 1, proving the equivalence of (1.3) with (1.1). One view of our algorithm is that it is designed to find a local minimizer of - q(y). Variables x and y are related by a piecewise linear transformation, a reflective transformation, a feasible sequence fx k g. Moreover, evaluation of - q(y) corresponds to evaluation of q(R(y)). Alternatively, one can view our approach entirely in the original variables x. Then, instead of describing the method as a descent algorithm for the transformed problem q(y), our method can be described as a method that generates feasible iterates by following a piecewise linear path induced by the reflective mapping R. We discuss this below. The second key observation is that the first-order optimality conditions for (1.1), or equivalently (1.3), can be expressed as a single system of nonlinear equations, and a Newton step for this system is a descent direction for - q in a neighbourhood of a local solution y . Moreover, in a neighbourhood of a local solution to (1.3) a full Newton step for (1.4), i.e., a unit step size in the Newton direction, yields decrease in q(y). This is a very important point because it suggests that a Newton process for (1.4) is compatible with (1.3), at least in the neighbourhood of a solution. It suggests that ultimate second-order convergence can be achieved while decreasing - q(y). The third observation leads to globalization of the Newton process. It turns out that the Newton equation for (1.4), the nonlinear system representing first-order optimality conditions, can be written in the form: where M is a symmetric matrix. 3 Moreover, it turns out that M is positive definite in a neighbourhood of a minimizer of - and M can be interpreted, loosely, as a second derivative matrix for - q(y) . This suggests the use of an ellipsoidal constraint to ensure a descent direction when far from the solution. Specifically, solve \Deltag where D is a positive diagonal scaling matrix and \Delta is positive. As we discuss in [10], a good choice for matrix D(x) is i.e., D is a diagonal matrix with the i th diagonal component equal to jv i (x)j 1 2 . Vector defined in Figure 1 where Diagonal matrix D plays an important role in this paper - henceforth we reserve the notation D, without superscript, to refer to definition (1.7). Of course D k refers to (1.7) with all quantities defined at the current point x k . (iv) If g i - 0 and l Fig. 1. Definition of v(x) Solving (1.6) involves solving a symmetric positive definite system, for a suitable -, and then s / D-s y . Thus it is easy to see that (1.6) leads to a descent direction for - q at the current point. It may be felt that solving (1.6) is an expensive way to determine a descent direction in the large-scale setting. With this is mind a restricted version of (1.6) is used in our algorithm. In particular, similar to [2] we usually restrict s to be in a low-dimensional subspace S. So (1.6) is replaced with where S is a low-dimensional subspace of ! n . Provided the ellipsoidal constraint inactive near the solution, and the low-dimensional subspace S 3 The function - q is a piecewise quadratic function of y: therefore, r y - q does not always exist. How- ever, the proposed algorithm only generates points where r y - q is defined. ultimately includes the Newton direction, the solution to (1.9) will eventually be the Newton step (1.5). The fourth major ingredient of our approach is the line search. Once a descent direction s k is determined, a one-dimensional line search is performed to approximately locate a minimizer of - q k has structure: - q k is a one-dimensional piecewise quadratic function and so an efficient specialized line search procedure can be used. (Alternative view: A one-dimensional piecewise linear search is performed along a "reflective path", p k (ff), defined by the reflective transformation R and beginning at We conclude the introduction with a short review of optimality conditions for problem (1.1). The first-order optimality conditions can be written: If a feasible point x is a local minimizer of (1.1) then Let F ree denote the set of indices corresponding to "free" variables at point x F ree Second-order necessary conditions can be written 4 : If a feasible point x is a local minimizer of (1.1) then D 2 ree is the submatrix of H corresponding to the index set F ree These conditions are necessary but not sufficient. To state practical sufficiency conditions we first need a definition of degeneracy. Definition 1. A point x nondegenerate if, for each index i: With this definition we can state second-order sufficiency conditions: If a nondegenerate feasible point x satisfies D 2 ree ? 0, then x is a local minimizer of (1.1). 2. The Reflective Transformation. One interpretation of our approach to solving the box-constrained quadratic programming problem (1.1) involves a transformation to an unconstrained piecewise quadratic minimization problem (1.3). The purpose of this section is to introduce this transformation. Since some of the ideas involved are more generally applicable, we begin our discussion at a more abstract level and gradually work our way back to the box-constrained quadratic programming situation. 4 Notation: If a matrix A is a symmetric matrix then we write A ? 0 to mean A is positive definite; means A is positive semi-definite. Consider the problem where f is a continuous and C is a closed connected region of ! n . We consider when the constrained problem (2.1) can be replaced with an unconstrained minimization problem of the form, min where R is a continuous onto mapping, What further restrictions on the mapping R make this an acceptable transfor- mation? To see that continuity is not enough consider the following one-dimensional example. Let Obviously there is only one local solution (the global solution), x = 1. However, let R(y) be any continuous function, mapping onto [0; 1], with a strict local maximizer at - y with - 1). It is easy to see that - y is a local minimizer of f(R(y)) but - x is clearly not a local solution to the original problem. The following property plays the key role in answering this question. R is an open mapping if for each ffl ? 0 and pair See Munkres [25], for example, for a discussion of open mappings. We can now answer our question. Theorem 1. C be a continuous onto mapping. Further, assume R is an open mapping. Then, (i) If y is a local minimizer of (2.2) then - local minimizer of (2.1). (ii) If x is a local minimizer of (2.1) then for each - - y is a local minimizer of (2.2). Moreover, there exists at least one - y such that (iii) Problem (2.1) is unbounded below if and only if problem (2.2) is unbounded below. Proof. (i) Assume y is a local minimizer of (2.2). Let - local minimizer, there exists ffl ? 0 such that But by (2.3) there exists Hence for each x 2 N Therefore, - x is a local minimizer of (2.1). (ii) Assume x is a local minimizer of (2.1). But R is an onto mapping and therefore there exists - y is a local minimizer, there exists ffl ? 0 such that By continuity there exists Therefore, for all y Hence, - y is a local minimizer of (2.2). (iii) Suppose fy k g is a sequence such that lim Alternatively, assume fx k g 2 C and lim But R is an onto mapping; therefore, for each x k there exists y k such that R(y k and lim and therefore (iii) is established. To illustrate, consider the problem 0g. A definition of R that clearly satisfies the open mapping property is multiplication. Note that R is differentiable and the Jacobian of R, J R (y), is nonsingular if and only if R(y) 2 int(C), where int(C) is the interior of C. Specifically, r y x where D g is the diagonal matrix diag(r x f) and D y is the diagonal matrix diag(y). (Note that r 2 y R is a tensor term and r x fr 2 y R is a matrix - diagonal in this case.) Therefore, this definition of R leads to an unconstrained twice-differentiable minimization problem and standard techniques can be used to solve (2.2). Unfortunately, our numerical experience with this approach has been mixed: In particular, as problems become large and ill-conditioning (and near-degeneracy) increases, the number of iterations required by standard minimization algorithms, to achieve good accuracy, becomes quite large. We feel this is due in part to the fact that this transformation causes an increase in the complexity of the function to be minimized: e.g., a quadratic function becomes a quartic. Our objection to this approach is largely numerical - ill-conditioning in the original problem is accentuated when the problem is transformed to a more complex form. Note also that the transformed problem may have many more local minimizers - by Theorem 1 this, in itself, is not a problem. However, along with this increase in the number of local minimizers comes an increase in negative curvature and this may cause some optimization algorithms some difficulty. In any event, our experience with this simple differentiable transformation has not been satisfactory: the subject of this paper is an alternative definition of R. For problem (2.5) consider is a vector, jvj denotes the vector whose components are the absolute values of the vector v. It is clear that the open mapping property holds. Note that R is not everywhere differentiable. In particular, R is differentiable at point y if and only if R(y) 2 int(C), i.e., y i 6= 0. In this case obviously nonsingular. Using this transformation, f(R(y)) has a piecewise differentiable nature as a function of y. For example, if f(x) is a quadratic function then f(R(y)) is a piecewise quadratic function of y. We now extend the absolute value approach, to handle the more general situation, where for each index i either u i is finite or u Similarly, for each index i either l i is finite or l For simplicity we assume that the finite values of u are all equal to unity and the finite values of l are all equal to zero (a simple translation and scaling can achieve this form 5 ). The transformation we propose, x = R(y), is a diagonal transformation, i.e., for 5 A definition of the reflective transformation applied directly to the general problem is given in [10] y Fig. 2. The 1-Dimensional Reflective Transformation (Finite Upper and Lower Bound) each index i, x i depends only on y i . This transformation, induces a piecewise linear "reflective" path in x. For example, if u Figure 2. If l is the absolute value function; if l then R i is illustrated in Figure 3. Finally, if u . The four cases are described more formally in Figure 4. It is easy to verify that R satisfies the requirements of Theorem 1; therefore, use of R does not introduce extraneous local minimizers nor does it restrict the set of local minimizers. Using the reflective transformation, problem (1.1) can be replaced with the unconstrained piecewise quadratic minimization (1.3). In principle, problem (1.3) can be solved by a descent direction algorithm, e.g., Algorithm 1 in Figure 5. An advantage of using this piecewise linear transformation R is the linear aspect of the transformation: when y is a differentiable point the local complexity of is the same as the local complexity of f(x). When q is a piecewise quadratic function. The apparent disadvantage is the piecewise nature of - f(y). This lack of differentiability means that conventional nonlinear minimization methods cannot be used. In particular, in order to guarantee convergence, restrictions on the nature of the descent direction s y must be imposed. To see this suppose that y k is very close to a y is a descent direction for - q at y k . If s y is nearly perpendicular to this hyperplane then the usual descent condition, r y - only result in a very small step since r y - q is not continuous at x In [10] we describe two properties, "constraint-compatibility" and "consistency", which help guarantee sufficiently long steps and, consequently, global convergence. We discuss this y OE x Fig. 3. The 1-Dimensional Reflective Transformation with Infinite Upper Bound briefly in Section 3. The straight-line direction s y k corresponds to a piecewise linear path in x. This piecewise linear path can be described as follows. For simplicity, and without loss of generality, assume y k . Define the vector 6 Component i of vector BR k records the positive distance form x k to the breakpoint corresponding to variable x k i in the direction s x k . The piecewise linear (reflective) path is defined by Algorithm 2 in Figure 6. Since only a single outer iteration is considered, we do not include the subscript k with the variables in our description of Algorithm 2 - dependence on k is assumed. Given the current point x k and a descent direction s x denote the piecewise linear path defined by Algorithm 2: For fi Note that it is now possible to describe Algorithm 1 entirely in x-space without explicitly introducing either the function - q or the variables y. We do this in Algorithm Figure 8. The difference between Algorithm 1 and Algorithm 3 is notational. The view presented by Algorithm 3 has the advantage that it is in the original space - visualization 6 For the purpose of computing BR we assume the following rules regarding arithmetic with infinities. If a is a finite scalar then a Case 1: (l To evaluate x then we can differentiate R to obtain the i th diagonal element of the diagonal Jacobian matrix else J R Case 2: (l To evaluate x then we can differentiate R to obtain the i th diagonal element of the Jacobian matrix Case 3: (l To evaluate x then we can differentiate R to obtain the i th diagonal element of the Jacobian matrix If y Case 4: (l In this case there are no constraints on x i and so x Fig. 4. The Reflective Transformation R Algorithm 1 Choose For 1. Determine a descent direction s y for - q(y) at y k 2. Perform an approximate line minimization of - k ), with respect to ff, to determine an acceptable stepsize ff k (such that ff k does not correspond to a breakpoint) 3. y Fig. 5. Descent dir'n algorithm for - f (y) Algorithm 2 [Let fi [i u is a finite upper bound on the number of segments of the path to be determined] For 1. Let fi i be the distance to the nearest breakpoint along 2. 3. Reflect to get new dir'n and update BR: (a) (b) For each j such that (b i Fig. 6. Determine the linear reflective path p of the reflective process is natural. The advantage of the first view, Algorithm 1, is that the algorithm is a straight line descent direction algorithm, a familiar structure. It is probably useful for the reader to keep both views in mind. In this paper we will primarily work in the x-space and Algorithm 3. For simplicity we now drop the superscripts y and x (e.g., s x becomes s). It is well known that a descent direction algorithm demands sufficient decrease at every step in order to achieve reasonable convergence properties. We use conditions suggested by Goldfarb [18] for use in the unconstrained setting: Given and a descent direction s k , ff k satisfies our approximate line search conditions if and Fig. 7. A Reflective Path Algorithm 3 Choose For 1. Determine an initial descent dir'n s x k for q at x k 2 int(F ). Determine the piecewise linear reflective path p k (ff) via Algorithm 2. 2. Perform an approximate piecewise line minimization of q(x k +p k (ff)), with respect to ff, to determine an acceptable stepsize ff k (such that ff k does not correspond to a breakpoint). 3. x Fig. 8. A reflective path algorithm is the piecewise linear path defined by (2.8). Condition (2.9) can be interpreted as restricting the step length from being too large relative to the decrease in f ; condition (2.10) can be interpreted as restricting the step length from being relatively too small. A basic reflective path algorithm for problem (1.1) can now be stated, Algorithm 4. To allow for flexibility, especially with regard to the Newton step, we do not always require that both (2.9) and (2.10) be satisfied. Instead, we demand that either both these conditions are satisfied or (2.9) is satisfied and ff k is bounded away from zero, e.g., latter conditions are used to allow for the liberal use of Newton steps and do not weaken the global convergence results. Note that Algorithm 4 generates strictly feasible points; i.e., since x 1 2 int(F ), it follows that x k 2 int(F ). Algorithm 4 [ ae is a positive scalar.] Choose For 1. Determine an initial descent dir'n s k for q at x k . Note that the piecewise linear path p k is defined by x 2. Perform an approximate piecewise line minimization of q(x k +p k (ff)), with respect to ff, to determine ff k such that: (a) ff k does not correspond to a breakpoint (b) condition (2.9) is satisfied (c) Either i. ff k satisfies condition (2.10), or ii. 3. x Fig. 9. A reflective path algorithm satisfying line search conditions 3. Algorithm Specifics. A framework for our reflective Newton approach was presented in the previous section, Algorithm 4. In this section we specify more precisely how the search directions will be generated as well as the mechanics of the line search, specialized to the quadratic problem (1.1). The convergence analysis given in [10] uses two important properties of the sequence of search directions, "constraint-compatibility" and "consistency". "Constraint- compatability" is needed to guarantee that a sufficiently long step is taken before the first constraint is encountered. The usual descent condition, that g T sufficiently negative, is not enough in the context of a reflective algorithm because this condition takes no account of the proximity of the constraints. "Consistency" is a more standard notion capturing the idea that first-order descent, represented by the term g T sisitent with first-order convergence. Following (1.7), define D 2 Definition 2. A sequence of vectors fw k g is constraint-compatible if the sequence fD \Gamma2 is bounded. Definition 3. A sequence of vectors fw k g satisfies the consistency condition Central to our approach, both in terms of achieving quadratic convergence and the satisfaction of constraint-compatibility and consistency, is the frequent use of a reduced trust region problem to determine s where S k is a subspace of R n , D k is a positive diagonal scaling matrix, and The matrix M k is defined: where J v is the Jacobian 7 of v, where v is defined in Figure 9. Matrix D g v is a diagonal matrix with component i defined D is an "extended gradient", extended to deal with possible degeneracy. In particular, where - g is a small positive constant. Clearly if x is a nondegenerate point and - g is sufficiently small then (which implies that x is degenerate) then The diagonal matrix D(x), used in (3.1), is defined by (1.7), i.e. 8 , This choice yields a well-defined trust region problem (3.1). To see this note that using (3.4), (3.1) becomes where s; and is a diagonal matrix, D M k is positive definite in a neighbourhood of a nondegenerate point satisfying the second-order sufficiency condi- tions. Moreover, unlike fM k g, f - k g is bounded. Matrix - M k is a featured performer in our reflective Newton algorithm. A Newton step is defined when - M k is positive definite: A final remark on the choice of scaling matrix (3.4). If we assume that D has the 1is the only reasonable choice. To see this suppose consider that 7 Matrix J v is a diagonal matrix with each diagonal component equal to zero or unity. For example, if all the components of u and v are finite then J has a finite lower bound and an infinite upper bound (or vice-versa) then strictly speaking v i is not differentiable at a point define J v at such a point. Note that v i is discontinuous at such a point but v i \Delta g i is continuous. 8 Notation: If z is a vector then jzj 1 2 denotes a vector with the i th component equal to jz and the calculation of DMD involves division by jv(x)j 1\Gamma2p which includes components which go to zero as x ! x . On the other hand, if approaches singularity as x ! x (consider v We will specify subspace S k below; it is important to realize that the cardinality of our implementation. Therefore, the cost of solving (3.1) is negligible. Given S k , the subspace trust region problem (3.5) can be approached in the following way. Let S k be defined by the t k independent columns of an n-by-t k be an orthonormalization of the columns of D \Gamma1 for some vector s Y k . Therefore problem (3.5) becomes ks Y k and set s . The solution to (3.7) is of negligible cost once the matrices are small (see Appendix). Algorithm 5 in Figure 10 presents a second-order reflective path algorithm. Algorithm 5 Choose For 1. 2. Determine initial descent dir'n s k for q at M k is positive definite and k-s N k . If - k is not positive definite choose choose subspace S k , and solve (3.1) to get s k . 3. Determine ff k and x k otherwise, perform an approximate piecewise line minimization of q(x k with respect to ff, to determine ff k such that (a) ff k is not a breakpoint (b) ff k satisfies (2.9) and (2.10). 4. x Fig. 10. A second-order reflective path algorithm Note: If ff accepted by the line search but corresponds to a breakpoint, then modify where ~ ff k is not a breakpoint, ~ 9 If A is a matrix then ! A ? denotes the space spanned by the columns of A. It remains to be more precise about the determination of s k and S k and to fully specify the line search. We begin with s k and S k . Algorithm 6 in Figure 11 describes our procedure. Algorithm 6 [Let positive constants.] Case 0: - M k is positive definite and k-s N Case 1: - M k is positive definite and k-s N if kr(-s N solve (3.1) to get s k . else set s Case 2: - k is not positive definite. Compute w k is a unit vector such that fw k g is constraint-compatible and z solve (3.1) to get s k . else solve (3.1) to get s k . Fig. 11. Determination of the descent direction s k Remark on Case 2: We determine an appropriate negative curvature direction in the following way. If a (sparse) Cholesky factorization of - k does not complete then k is not positive definite and a unit direction of non-positive curvature, - readily available and easily computable (e.g., [17]). Algorithm 6 can make use of - sufficient negative curvature is displayed by - and fw is constraint-compatible. A constraint-compatibility test is implemented by introducing a large constant, - cp , and requiring, If either condition (3.9) or condition (3.10) is not satisfied then - must be rejected. In this case we can turn to a Lanczos process [10] to get a unit vector - w k such that both and (3.10) are satisfied. It is interesting to note that in our extensive numerical experimentation, with results reported in Section 4, conditions (3.9) and (3.10) were always satisfied by the (partial) Cholesky factorization method - the backup Lanczos procedure was never required. The Line Search. We have designed a specialized approximate line search procedure to efficiently exploit the structure of this problem and to guarantee the line search conditions in Algorithm 5. Before describing the approximate procedure, we develop an exact line search procedure - this is possible because the problem is to find a local minimizer of a quadratic function along a piecewise linear path. In the end we do not use the exact line search per se but rather we use a truncated version of it, subroutine "improve", within an overall approximate strategy. But we begin with the exact search. The Exact Line Search. We are initially concerned with the exact determination of ff k where ff k is a local minimizer of q(x k (ff)). Note: It is convenient to describe the exact line search in terms of the y-variables, i.e., y R(y view we have a straight-line minimization of a piecewise quadratic function - Henceforth in this section we omit the major iteration subscript. The function - is a continuous piecewise quadratic function. The ray can be divided into intervals from left to right, I q(ff) is smooth on each interval. Denote the restriction of - q(ff) to the j th interval by q j (ff) - note that q j (ff) is a quadratic function of a single variable. Our exact line search algorithm visits the intervals I 1 ; I 2 ; :::; in a left-to-right fashion in an attempt to locate the first local minimizer of - Assume we have not located a local minimizer on intervals I 1 ; I 2 ; :::; I j \Gamma1 and assume that (q j There are two possibilities: either q j (ff) has a minimizer strictly within the interval I j or it does not. If it does, i.e., ff j . However, if does not admit a minimizer within I j then we must consider the possibility that fi j is a minimizer of - (ff). This is now the case if (q j+1 is not in int(I j ) and (q j+1 process is repeated on interval I j+1 . Algorithm 7 in Figure 12 presents a compact description of the exact line search algorithm we have sketched above. Step (1.2) in Algorithm 7 follows from the observation that if (ff k then s is a direction of infinite descent for - q, beginning at y and therefore, by Theorem descent for (1.1). Algorithm 7 [Exact line search along direction s beginning at point y] (0) Determine the array of breakpoints BR according to (2.7). fi 0 / 0. (1) For , the minimizer of q k , if it exists; otherwise, set ff k 1. (problem (1.1) is unbounded). , exit. exit. is the index such that according to Algorithm 2. Fig. 12. The Exact Line Search Algorithm Our final concern, with regard to the exact line search, is an efficient implementation of step (1.1). In theory this computation is straightforward. Assume q k a k . If a k unbounded below; if a k then q k (ff) is unbounded below (unless a k which case q k is constant). However, the challenge is to determine a k 1 and a k efficiently, for (Note that a k 0 is not The key to efficiency here is the observation that the reflective transformation R is linear on each interval I is constant within each interval. For each interval I k , define oe k to be the vector of diagonal elements of J R evaluated at any point in the interval (y it follows that for It follows that Therefore, in the terminology used above, a k a k A straightforward implementation for determining a k breakpoint (a k 0 is not needed). However, there is considerable structure that can be exploited; in particular, oe k+1 , a vector with each component equal to \Sigma1, differs from oe k in exactly one component. We can exploit this to reduce the work in the line search to O(n) per breakpoint. Suppose we have determined that I k does not contain ff k and we have at hand the following quantities: a k a k and x k , the value of x at the k th breakpoint. If j is the index such that then The vector w k can be updated as follows: where H j is column j of H. Coefficient a k+1 2 is simply computed: a k+1 Coefficient a k+1 1 can be efficiently computed by considering the following equalities: a k+1 Finally, x k+1 is computed: In summary, the coefficients a k+1 2 and the intermediate quantities x can be computed, given a k using (3.16),(3.17),(3.18) and (3.19). This amounts to approximately 4n work. Of course the initial quantities, w a 0must be computed from scratch requiring O(n 2 ) work. Therefore, if k br denotes the number of breakpoints crossed, the total cost of the exact line search is: initialization of w (ii) O(n) for determination of BR, (iii) O(k br n) for steps (1.0) - (1.5). The Approximate Line Search. The exact line search described above is not practical, for two reasons. First, an exact minimizer along a line may correspond exactly to a breakpoint, i.e., a boundary point, and the algorithm requires strictly feasible points. This is actually not a serious problem since a small perturbation would yield strict feasibility and the reflective Newton method is not very sensitive to boundary proximity. A more serious objection to the exact line search is economy: despite the efficient implementation described in the previous section, the relative cost can be high if the number of breakpoints crossed, k br , is large. Certainly if there are a large number of tight variables at the solution, say something close to n, then the total cost of the exact line search algorithm ultimately becomes O(n 2 ) per line search. This is unsatisfactory and unnecessary since an economical approximate line search can be just as effective. In this section we describe an approximate line search, henceforth refer to as subroutine improve, which uses the exact line search, described above, in a limited fashion- beginning at an approximate minimizer, subject to a bound, k u , on the number of breakpoints permitted to cross. In particular, improve is used in a cleanup role: after determining an initial approximate minimizer by a bisection strategy, improve is called upon to apply the exact line search strategy. Thereby the approximate minimizer is further improved at cost O(k u n), where k u is typically chosen to be small, e.g., k (In improve we also impose an approximate upper bound ff max on the size of ff. That is, the size of the improvement is bounded by ff Subroutine improve has the following calling sequence: A more precise description of the approximate line search algorithm is given in Figure 13, Algorithm 8. The basic idea is as follows. First, if the direction s k is a Newton direction s N k then a unit step is attempted. If (2.9) is satisfied then the full Newton step is accepted subject to further improvement by subroutine improve and possible (slight) adjustment to avoid a breakpoint. Second, if s k does not corrspond to a Newton direction or if it does but a unit step does not satisfy (2.9), then a bisection procedure is executed on the interval (0; ff u ) where ff u - 1 is an upper bound on the step size. A point is located satisfying both (2.9) and (2.10) and then possibly further improved with subroutine improve. Algorithm 8 [Approximate line search along direction s beginning at point y] k and a unit step along s k satisfies (2.9) ff k corresponds to a breakpoint, set ff is not a breakpoint else f s k 6= s N k or a unit step along s N k does not satisfy (2.9)g ffl use bisection to find - and (2.10) such that - ff k is not a breakpoint set ff ff k corresponds to a breakpoint, set ff is not a breakpoint Fig. 13. Approximate Line Search Algorithm 4. Numerical Experiments. We have implemented our algorithm in a version of Matlab which allows for sparse matrix data structures [15], now Matlab 4:0. In this section we present some preliminary numerical results. With the exception of the results reported on Table 12, all experiments were performed on Sun Sparc workstations in the Matlab environment [22]. Experiments reported in Table 12 were performed in a heterogeneous environment involving an Intel IPSC/860 32-node multiprocessor as the "backend", and a Sun Sparc workstation as the "frontend". Matrix factorizations and solves were performed on the "backend", in C, while the main Matlab program executed on the "frontend". Communication between "frontend" and "frontend" over ethernet was implemented through the use of Matlab "MEX" files. We used this environment to facilitate the solving of very large problems. (Details on this heterogeneous environment are given in [3].) Starting and Stopping: In all the experiments reported in this paper the starting value of x, i.e., x 1 , is as follows. For component j where both upper and lower bounds are finite, choose the midpoint, j. If both upper and lower bound corresponding to component j are infinite in size, choose choose (Note: The reflective Newton approach is not particularly sensitive to starting value. For example, we repeated many of the experiments reported here using a random (strictly) feasible starting point - very little difference in behaviour was detected.) Choosing a robust stopping rule in optimization is not easy. Our primary stopping rule is based on the relative difference in function value. This is reasonable partly because strict feasibility is always maintained, and partly because often the real objective in practical optimization is to achieve a point of relatively low function value. Specifically, are primary stopping rule is: We choose in Matlab on a Sun Sparc workstation, . We do have secondary stopping criteria as well - designed to determine when progress is deemed too slow. This secondary rule tends to kick in when solving degenerate or ill-conditioned problems and a very flat region around the solution has been entered. Parameter settings: There are a number of parameters in the algorithm: most are either in the very large or very small category. Here are the settings we used in our experiments: Used in the determination of scaling matrix D, see (3.3): - Used in the line search, see (2.9): oe l = :1 Used in the line search, see (2.10): oe A bound on the number of breakpoints crossed in subroutine improve: ffl ae: A lower bound on the stepsize, see Algorithm 8: If the line search produces a unit step which turns out to be a breakpoint, this point is perturbed by an amount bounded by - ff kD k g k k, see (3.8): - Used to test for constraint-compatibility, see (3.10): - Used in the negative curvature test, see (3.9): ffl Used in Algorithm Used in Algorithm 8: An upper bound on the bisection process used in Algorithm 8: ff 4.1. Positive Definite Problems. We have generated a number of quadratic test problems with certain properties. In the first set of results we concentrate on the case where H is symmetric positive definite. In the results reported below we use sparse matrices H with sparsity patterns representing 3-dimensional grid using a 7- point difference scheme. The Mor'e/Toraldo [24] QP-generator was adapted to generate problems with a given sparsity pattern (see also [6]). We will not review the generator characteristics here: our generator is a straightforward adaptation of the Mor'e/Toraldo scheme to the sparse setting. We use several sparse Matlab functions (e.g., "sprandsym", "sprand"). In Tables 1-3, the dimension of the test problems is in each case. The parameter "pctbnd" indicates the percentage of variables tight at the solution - approximately evenly divided between upper and lower bounds. Parameter "deg" reflects the degree to which the solution is (nearly) degenerate - the larger the value of "deg", the greater the amount of (near) degeneracy. Technical details of "deg" are discussed in [6]. Parameter "cond" reflects the conditioning of the matrix H: the condition number of H is approximately 10 cond . The upper and lower bound vectors, u and l, were generated as follows. Approximately 75% of the components of l were chosen to be finite and assigned the value of zero - the index assignment was made in a random fashion. Similarly, approximately 75% of the components of u were chosen to be finite and assigned the value of unity. Again, the index assignment was made in a random fashion, independent of the assignment of l. Each row of Tables 1-3 reflects the results of 10 independent runs with the same parameter settings. The third column, labelled "max", indicates the maximum number of iterations required, over the set of 10 independent runs, to achieve the stopping criteria; the fourth column, labelled "avg" records the average number of iterations required to reach the stopping criteria over the 10 problems; the last column, labelled "acc", records the number of digits of accuracy achieved in the function value (the true solution is known). Table Positive deg cond max avg acc 9 6 15 15 15 9 9 Observations on Tables 1-3: First, we observe the remarkable consistency of our reflective Newton method on these problems. In terms of iterations required to achieve the stopping criteria and accuracy attained in the function value, there is apparently very little sensitivity to degeneracy, conditioning, or number of variables tight at the solution. Of course we do not claim that accuracy in x is independent of condition/degeneracy it surely is not. However, it is usually acceptable in optimization to attain a point with nearly optimal function value and we have been quite successful in that (on this test collection). Second, the absolute number of iterations required to obtain a very accurate solution (in terms of the function value q) is modest in every case, i.e., less than 20. Positive deg cond max avg acc 9 6 17 17 15 6 9 17 17 15 9 9 17 16.3 15 Table Positive deg cond max avg acc 9 3 17 16.3 15 9 6 Positive Definite Problems: Timing Breakdown 1000 This is very encouraging considering the dimension of the problems (n = 1000) and the spectrum of problem characteristics being considered. It is important to know where the algorithm spends its time. To this end we generated larger problems, with the same structure, and we have broken down the timing information. In Table 4 we consider a representative positive definite problem with "av- erage characteristics", i.e., vary the problem dimension n. (The sparsity structure remains the same.) The second column, labelled "it", records the number of iterations required to achieve the stopping criteria; "totM" records the total number of flops used by the (partial) Cholesky factorization ("m" represents a million); "totls" records the number of flops used in the approximate line search algorithm. Over 95% of the total flop count on these problems is represented by the sum of the "totM" and "totls" columns - the remaining work in the algorithm, such as the 2-dimensional trust region solution, is negligible in comparison (see Appendix for more detail on the solution of trust region problems). Observations on Table 4: First, there is no significant growth in number of iterations as the problem dimension n increases. High accuracy is maintained for larger values of n as well. As n increases the sparse matrix factorization work, "totM", increases relative to the lines search cost, "totls". Therefore, speedup of the (partial) sparse factorization aspect of the algorithm (e.g., use of parallelism, exploitation of specific particular structure) will have significant impact on the overall computing time. Conversely, improving the approximate line search (in terms of cost) is not a crucial computing issue, at this point, for large-scale problems. In addition to these randomly generated, but structured, positive definite problems, we have experimented with three specific test cases. Two of these problems are from the literature (e.g., [12, 24]) and the third example is new. In Tables 5 and 6 we report on the "obstacle" problem - in the first case there are lower bounds only, in the second case there are lower and upper bounds. In defining the specific example used we have chosen the same parameter settings and specific functions used in [24]. Table 7 reports on the elastic-plastic torsion problem. Again we used the same parameters as reported in [24] to define the problem. In Table 8 we report on a linear spline approximation problem. This type of problem arises, for example, in a particle method approach to turbulent combustion simulation [28]. The problem results in a large sparse least-squares problem subject to nonnegativity constraints on the variables. To set up a sample problem we assume an m-by-m-by-m 3-dimensional grid. Within each cell are a set of particles randomly located (we use approximately 10 particles per cell in our experiments). Each particle p has a known function value, OE(p). Associate with each grid intersection point a linear basis function and determine the best set of coefficients, x, for the basis functions, in the least-squares sense, subject to nonnegativity constraints on x. The function OE we used in our experiments is defined: given a point in 3-space, Table Obstacle Problem: Lower Bounds Only its norm 50 2500 14 13 100 10,000 15 12 Table Obstacle Problem: Lower and Upper Bounds its norm 50 2500 14 12 100 Table Elastic-plastic Torsion Problem its norm 100 10,000 11 12 Observations on Tables 5-8. The most noteworthy observation is the apparent insensitivity of our method to problem size for each of these problems. The number of Linear Spline Approximation its norm 22 10,648 17 11 iterations does not grow, for a given problem class, as the dimension of the problem increases. For example, for the linear spline problem, 16 iterations are required when iterations are required when Moreover, the number of iterations is always modest, on this test set, i.e., less than 20. High accuracy is achieved in all cases. 4.2. Indefinite Problems. We have adapted the Mor'e/Toraldo QP generation scheme, in combination with sparse matrix functions in Matlab 4.0, to generate large sparse indefinite matrices with a given sparsity pattern and given approximate set of approximate eigenvalues. In the indefinite case we chose finite upper and lower bound vectors, 1. (This is to avoid the generation of unbounded problems.) In each of the problems in Tables 9-12 roughly 10% of the eigenvalues of H are neg- ative. The column labels are the same as before however here "acc" does not represent the number of accurate digits compared with the true solution since the true solution is unknown due to indefiniteness of H. Instead, "acc" records the number of matching digits in the objective function q in the last 2 iterations. (In each case the optimality conditions were verified to hold at the final point.) Table Indefinite problems. deg cond max avg acc 9 3 23 19.3 15 9 6 26 21.7 15 9 9 Observations on Tables 9-11: Iteration counts indicate that our method is not quite as consistent or efficient on indefinite problems compared to the performance on positive definite problems. Still, the overall efficiency seems very good - the average number of Indefinite Problems. deg cond max avg acc 9 3 9 6 19 17.7 15 3 9 14 11.3 15 9 9 25 18.3 15 Table Indefinite problems. deg cond max avg acc 9 3 9 6 6 9 15 13. iterations required for any problem category is always less than 23. In Table 12 we indicate where the algorithm spends its time on indefinite problems by considering a representative example and increasing the dimension. Table Indefinite Problems: Timing Breakdown Remark on Table 12: We see no apparent growth in required iterations as n increases. Clearly the "totM" column dominates the "totls" column as n increases. Recall that "totM" represents the matrix factorization flop count while "totls" represents the number of total flops required by the line search procedure. Therefore, to obtain further improvements in efficiency for this type of approach it is best to focus on the matrix factorization aspect of the overall procedure. 5. Theory and Conclusions. The numerical results obtained to date strongly support the notion that a reflective Newton method represents an efficient way to accurately locate local minimizers of large-scale (indefinite) quadratic functions subject to bounds on some of the variables. The theory is supporting also: our reflective Newton method is globally and quadratically convergent. Coleman and Li [10] present important theoretical properties of reflective Newton methods for general nonlinear functions, subject to bounds on some of the variables. The method in this paper is a specialization of the general method to the quadratic case. Therefore, the general theory applies. We make a compactness assumption before formally stating the main result. Compactness Assumption: Given initial point x 1 2 F , it is assumed that the level set Theorem 2. Let fx k g be generated by Algorithm 5 with fs k g generated by Algorithm 6 and with fff k g determined by the approximate line search algorithm (Algorithm 7). If - ffl Every limit point of fx k g is a first-order point. ae M is the maximum spectral radius of - M (x) on )g. Since ae( - is continuous on L, a compact set, the upper bound ae M exists. Recall that - 3 is a constant used in Algorithm 6. ffl Every nondegenerate limit point satisfies the second-order necessary conditions. ffl If a nondegenerate limit point x satisfies second-order sufficiency conditions is sufficiently small, fx k g is convergent to x ; the convergence rate is quadratic, i.e., Proof. This algorithm is in the class of algorithm described in [10] and all the assumptions of Theorem 20 in [10] are satisfied. The result follows from Theorem 20 in [10]. In conclusion, strong theoretical and computational results indicate that a reflective Newton method is an efficient and reliable way to solve problem (1.1) to high accuracy. The computational results reported in this paper support this claim. 6. Acknowledgement . We thank our colleagues Jianguo Liu and Danny Ralph for many helpful discussions on this work. Danny Ralph drew our attention to the topology reference [25]. 7. Appendix : The trust region problem. The trust region problem is where A is a real symmetric matrix and k \Delta k denotes the 2-norm. The purpose of this section is to review the nature of problem (7.1) and discuss a possible solution suitable for low-dimensional problems. (In the context of our reflective Newton method for problem (1.1), A is matrix - M(x), a symmetric matix of order 2. The computational cost of the procedure we describe to solve (7.1) is negligible compared to the other required computations in the reflective Newton algorithm we propose.) For larger problems a more approximate procedure is usually preferred, e.g., [14, 23, 30, 31]. Much of the material in this section can be found elsewhere, e.g., [2, 4, 11, 14, 23, 29, 30, 31]. Diagonalization. We begin with an extremely useful characterization of the global solution to (7.1). Theorem 3. Vector s solves (7.1) if and only if there exists a scalar - 0 such that (a) positive semidefinite; (c) ksk - \Delta; Proof. A proof is given, for example, in [30]. The usefulness of this result is best revealed after diagonalization. Suppose where the columns of V are the orthonormal eigenvectors of A and Obviously then so (a) is equivalent to s: By (b), - \Gamma- 1 , and so all vectors s satisfying (a, b) are of the form The vector fi is arbitrary with respect to (a,b) but can sometimes help with respect to satisfying (c,d). A basis for an algorithmic approach to this problem is to assume the form given in (7.4) and strive to satisfy (c,d) by choosing -, and in some cases fi, appropriately. (fi plays a role only if - is the value of - at the solution.) The situation where (c,d) can be satisfied with easily dispensed with (first half of Case 1 below). Therefore the primary task, assuming form (7.4), is to determine -, and sometimes fi, to satisfy We divide our approach into three possibilities. g. Case 1: In this case either the Newton step is within the sphere ksk 2 - \Delta or it is not. If kA then the optimal solution s Case 2: Figure Obviously ks(-)k !1 as - ! \Gamma- 1 , and 1. Moreover, ks(-)k is convex; therefore, ks(-)k intersects \Delta in exactly one place for - ? \Gamma- 1 . Case 3: ks 1 (\Gamma- 1 )k is finite. Consider figures 3,4. There are now two possibilities. If ks 1 (\Gamma- 1 then there is a solution to (7.5) to the right of \Gamma- 1 , chosen to ensure (7.5), and - . Note that if jIj ? 1 then the null space component of s may not be unique. The Reciprocal Secular Equation. In theory we can build an algorithm on the remarks given above. However it is better, numerically, to replace condition (7.5) with \Gammaks(-)k 0: Equation (7.6) is more linear in shape than equation (7.5); therefore, equation (7.6) is more amenable to solution via Newton's method. Considering the definition of rsec(-), -n+- it is easy to verify the following: 1. rsec(-) is convex on (\Gamma- 1 ; 1] and lim -!1 2. lim -!\Gamma- +rsec(-) is finite. 3. If ff i 6= 0, for some i 2 I, then rsec(\Gamma- . Obviously a single zero of rsec exits to the right of \Gamma- 1 in this case. 4. If 8i 2 I; ff Algorithmic and Numerical Concerns. We assume that a solution to (7.1) is sought and we are willing and able to compute full eigenvalue information, (If this is not the case, perhaps due to the cost, then it is possible to approximately solve (7.6) using an iterative scheme involving the Cholesky factorization of A [14, The method using appears to be straightforward. Case 1: then the Newton step, \GammaA \Gamma1 g, is the solution with then we can determine the zero of rsec(- ? 0. Case 2: and rsec admits a single solution to the right of \Gamma- 1 . Case 3: If ks 1 (\Gamma- 1 )k - \Delta then there is a zero of rsec(-) to the right of \Gamma- 1 with Otherwise, a solution to (7.1) is given by (7.4), ks Unfortunately, the situation is not quite so clean from a numerical point of view: there is fuzziness between the second and third cases. In particular, if ff 1 is small the equation very ill-conditioned for - near \Gamma- 1 and it can be quite difficult (nigh impossible) to compute - such that rsec(-) is small. This extreme ill-conditioning is due to the following "disagreement". Assume I = f1g, for simplicity, and note that, if ff On the other hand, if which is not, in general, the limiting value of (7.8). Therefore nearby problems (ff can yield very different solutions to (7.6), and this is the cause of the ill-conditioning of (7.6). Our solution to this ill-conditioning problem (trust.m) is to first attempt to find a solution to However, if jrsec( -)j is not small, where - - is the computed "zero" returned by the zero-finder, then we set - solution to (7.1) via (7.4). This strategy works because the solution to (7.6) with ff i small for i 2 I is close to a solution of (7.1) with the corresponding ff i at zero. To see this, initially assume that Now first consider the case where ff the solution to (7.1) is: where otherwise, the result is obvious). Define s the solution to (7.1) can be written Next consider ff small number. We can write the solution to (7.1) as But as which implies Therefore, the solution to (7.1) with ff is near to the solution to (7.1) with ff In general, if jIj ? 1 a solution to (7.1), with some components ff i near to zero, is near to a solution of (7.1) with those components set to zero. In this case, where several coefficients ff i equal to zero, i 2 I, problem (7.1) does not enjoy a unique solution [see (7.4)] but the range space component is unique. --R Approximate solution of the trust region problem by minimization over two-dimensional subspaces Advanced Computing Research Institute Computing a trust region step for a penalty function A direct active set algorithm for large sparse quadratic programs with simple bounds A quadratically-convergent algorithm for the linear programming problem with lower and upper bounds Global convergence of a class of trust region algorithms for optimization with simple bounds On the minimization of quadratic functions subject to box constraints Minimization of a quadratic function of many variables subject only to lower and upper bounds Computing optimal locally constrained steps Sparse matrices in matlab: Design and implemen- tation Minimization subject to bounds on the variables Curvilinear path steplength algorithms for minimization algorithms which use directions of negative curvature A globally convergent method for l p problems Solving the minimal least squares problem subject to bounds on the variables A generalized conjugate gradient algorithm for solving a class of quadratic programming problems Application of the velocity-dissipation PDF model to inhomogeneous turbulent flows A family of trust-region-based algorithms for unconstrained minimization with strong global convergence properties Trust region methods for unconstrained optimization The conjugate gradient methods and trust regions in large scale optimization A class of methods for solving large convex quadratic programs subject to box constraints --TR --CTR Kamin Whitehouse , David Culler, Calibration as parameter estimation in sensor networks, Proceedings of the 1st ACM international workshop on Wireless sensor networks and applications, September 28-28, 2002, Atlanta, Georgia, USA Kamin Whitehouse , David Culler, Macro-calibration in sensor/actuator networks, Mobile Networks and Applications, v.8 n.4, p.463-472, August D. C. Jamrog , R. A. Tapia , Y. Zhang, Comparison of two sets of first-order conditions as bases of interior-point Newton methods for optimization with simple bounds, Journal of Optimization Theory and Applications, v.113 n.1, p.21-40, April 2002 Keiji Yanai , Nikhil V. Shirahatti , Prasad Gabbur , Kobus Barnard, Evaluation strategies for image understanding and retrieval, Proceedings of the 7th ACM SIGMM international workshop on Multimedia information retrieval, November 10-11, 2005, Hilton, Singapore V. G. Domrachev , O. M. Poleshuk, A Regression Model for Fuzzy Initial Data, Automation and Remote Control, v.64 n.11, p.1715-1723, November
interior-point method;interior Newton method;quadratic programming
589192
Convergence Properties of Minimization Algorithms for Convex Constraints Using a Structured Trust Region.
In this paper, we present a class of trust region algorithms for minimization problems within convex feasible regions in which the structure of the problem is explicitly used in the definition of the trust region. This development is intended to reflect the possibility that some parts of the problem may be more accurately modelled than others, a common occurrence in large-scale nonlinear applications. After describing the structured trust region mechanism, we prove global convergence for all algorithms in our class.
Introduction Trust region algorithms have enjoyed a long and successful history as tools for the solution of non- linear, nonconvex, optimization problems. They have been studied and applied to unconstrained problems (see [7], [17], [25], [28], [29], [30], [31], [34], [35], [38]) and to problems involving various classes of constraints, including simple bounds ([6], [10], [11], [27], [32]), convex constraints ([2], [3], [14], [41]), and nonconvex ones ([5], [8], [16], [36], [44]). This long lasting interest is probably justified by the attractive combination of a solid convergence theory, a noted algorithmic robustness, the existence of numerically efficient implementations and an intuitively appealing motivation. The main idea behind trust region algorithms is that, if a nonlinear function (ob- jective and/or constraints) is expensive to compute or difficult to handle explicitly, it should be replaced by a suitable model. This model is deemed to be trustworthy within a certain trust region around the current point. The trust region is defined by its shape and its radius. The minimization involving the difficult nonlinear function(s) is then replaced by a sequence of minimizations of the simpler model(s) within appropriate trust regions. The trust region radii are adjusted to reflect the agreement between the model and true functions as the process proceeds. It is remarkable that, up to now, all algorithms that we are aware of use a single trust region radius to measure the degree of trustworthiness of the models employed, even if several This research was supported in part by the Advanced Research Projects Agency of the Department of Defense and was monitored by the Air Force Office of Scientific Research under Contract No F49620-91-C-0079. The United States Government is authorized to reproduce and distribute reprints for governmental purposes notwithstanding any copyright notation hereon. This work was also supported by the Belgian national Fund for Scientific Research. different functions are involved. This choice is somewhat surprising if one admits that some of the modelled functions could be substantially "better behaved" than others in the same problem, as this implies that the region in which their models can be trusted might also be substantially larger. In this context, the unstructured trust region choice might be viewed as a conservative strategy ensuring that all models may be trusted in what amounts to a "safe minimal" region. While this strategy might be reasonable for small problems, where each involved function depends on all the problem's variables, it is clearly questionable for large-scale applications, where each of the problem's function typically depends only on a small number of variables. For instance, one might consider the minimization of an unconstrained objective function consisting of the sum of many quadratic and a few highly nonlinear terms, the latter involving a small subset of the variables. If a classical unstructured trust region algorithm, with a quadratic model, is used, the quadratic terms are perfectly modelled, but the steps that one can make at each iteration are (unnecessarily) limited by the highly nonlinear behaviour of a small subset of the variables. It is the purpose of this paper to present and analyze a class of algorithms that use the problem's structure in the definition of the trust region, allowing large steps in directions in which the model has proved to be adequate while restricting the movement in directions where the model seems unreliable. To be more precise, we will consider the problem of minimizing a partially separable objective function subject to convex constraints; we will then use the decomposition of the objective function into element functions as the basis for our structured trust region definition. The choice of the partially separable structure, a concept introduced in [21], is motivated by the very general geometric nature of this structure and by the increasing recognition of its practical use (see [4], [9], [12], [13], [18], [19], [20], [22], [26], [39], [42], [43], amongst others). More significantly, partial separability provides a decomposition of the considered nonlinear function into a linear combination of smaller element functions, each of which may then be modelled separately (see [40]). It is then quite natural to assign one trust region radius per element function and to decide on its increase or decrease separately. Because different element functions typically involve different sets of variables, each element trust region only restricts the components of the step corresponding to its elemental variables. An obvious approach is to use the norm-scaling matrices allowed in the theory for unstructured trust region methods ([10], for instance) to account for differences in model adequacy among elements when constructing the trust region. This would be satisfactory if the existing theory did not require that the scaling matrices be of uniformly bounded condition number. Unfortunately, it is easy to conceive of instances where this is a severe handicap. For example, it would prevent the trust region radius of a well-modelled (perhaps linear or quadratic) element from increasing to infinity while at the same time ensuring that that of a badly behaved nonlinear element function remains of modest size. Moreover, this strategy may well cause numerical difficulties when attempting to solve the trust region problem. In fact, as we will shortly see, additional algorithmic safeguards are important when simultaneously handling trust regions of vastly different sizes. Thus, we do not consider such an approach further in this paper. Section 2 of the paper presents the problem in more detail and the new class of algorithms using the principle of structured trust regions. Global convergence for all algorithms in the class is proved in Section 3. We briefly discuss the identification of active constraints in Section 4. We examine in Section 5 some extensions of the results of the previous sections. We finally give some comments and perspectives in Section 6. Structured trust region for partially separable problems 2.1 A structured model of the objective and the corresponding structured trust region 2.1.1 The problem The problem we consider is that of minimizing a smooth objective function subject to convex constraints. That is, we wish to solve the problem minimize where X is a closed convex subset of R n . We denote the Euclidean inner product on R n by h\Delta; \Deltai, and the associated ' 2 -norm by k \Delta k. Given Y a closed convex subset of R n , we define the operator Y (\Delta) to be the orthogonal projection onto Y . We now list our additional assumptions on (2.1). AS.1 X has a non-empty interior. AS.2 f is bounded below on X . AS.3 f is partially separable, which means that and that, for each i 2 there exists a subspace N i 6= f0g such that, for all w 2 N i and all x 2 X , AS.4 For each continuously differentiable in an open set containing X and its gradient is uniformly bounded on X . Note that we admit the case where X is unbounded or even identical to R n itself, in which case we obtain an unconstrained problem. In relation to the partial separability of the objective function, we also consider the range subspace (see [23]) associated with each element function f i , which is defined as We are mostly interested in the case where the dimension of each R i is small compared to n. A commonly occurring case is when each element function f i only depends on a small subset of the problem's variables: R i is then the subspace spanned by the vectors of the canonical basis corresponding to the variables that occur in f i (the elemental variables). The range of the projection operator PR i (\Delta) is therefore of low dimensionality. The reader is referred to [12] for a more detailed introduction to partially separable functions. We note that f is invariant for any translation in the subspace ( We may therefore restrict our attention to the case where without loss of generality. 2.1.2 The element models The algorithm we have in mind is iterative and generates feasible iterates (in the sense that all iterates belong to X). At iteration k, we will associate a model m i;k with each element function f i . This model, defined on R i in a neighbourhood of the projection of the k-th iterate x k on this subspace, is meant to approximate f i for all x in the element trust region where is the i-th trust region radius at iteration k and the norm k \Delta k is chosen to be the usual Euclidean norm in order to simplify the exposition. In what follows, we will slightly abuse notation by writing m i;k (x) for an x 2 R n , instead of the more complete m i;k (PR i (x)). We will furthermore assume that each model m i;k (i differentiable and has Lipschitz continuous first derivatives on an open set containing B i;k , and that Moreover, we assume that g i;k in the sense that, for all where e i;k is a constant and where \Delta min;k is defined by i2f1;:::;pg Condition (2.8) is quite weak, as it merely requires that the first order information be reasonably accurate whenever some trust region radius is small (i. e. the corresponding model fits badly). Indeed, one expects the coherency of this first order behaviour to be of crucial importance in such cases. Further arguments supporting a choice similar to (2.8) for problems with convex constraints are presented in [14]. Amongst the most commonly used element models, linear or quadratic approximations are pre-eminent. One can, for instance, consider the quadratic model given by the first three terms of the element function Taylor series around the current iterate. Another popular choice is a quadratic model where the second derivative matrix is recurred using quasi-Newton formulae. 2.1.3 The overall model and trust region With all the element models at hand, we are now in position to define the overall model at iteration k, denoted m k , whose purpose is to approximate the overall objective function f in a neighbourhood of the current iterate x k . From (2.2), it is natural to use the overall model for all x in the overall trust region defined by i2f1;:::;pg Indeed B k is the intersection of all element trust regions, that is the region in which all element models may be trusted, irrespective of the additional limitation possibly imposed by the feasible set X . Of course, the actual shape of the trust region B k is determined by the choice of the Euclidean norm: it corresponds to the intersection of cylinders whose axis are aligned with the subspaces N i and whose radii reflect the quality of the element models: large in subspaces where the element models predict the element function correctly and smaller in subspaces where the prediction is poorer. In practice, one might wish to choose other norms, such as the ' 1 -norm. In this case, and assuming that the subspaces R i are spanned by subsets of the canonical basis vectors, the shape of the trust region is that of a box, the length of whose sides again reflects the quality of the element models. The extension of the theory to more general norms is considered in Section 5.4. 2.1.4 Curvature We now follow [14] and [41] and define the generalized Rayleigh quotient of f at x along s 6= 0 by Obviously, this definition is valid only if s is such that x belongs to the domain of definition of f . Note that, by convention, If we assume that f is twice continuously differentiable, the mean-value theorem (see [24]) implies that Z 1Z 1t dv dt: (2:14) Furthermore, if f is quadratic, then one easily verifies that !(f; x; s) is independent of x and is equal to the Rayleigh quotient of the matrix r 2 f in the direction s. We note that, because of is bounded by some constant L i - 0 (see [24]). Hence we obtain that i2f1;:::;pg for all pg. The quantity that we need in our algorithm statement and analysis is a monotonically increasing upper bound on the magnitude of the generalized Rayleigh quotient !(m i;k defined by q2f0;:::;kg i2f1;:::;pg where s i;k the actual trial step computed by the algorithm, as defined below. The quantity !(m i;k measures the curvature of the model m i;k in the direction of the trial step s k . If quadratic models m i;k are considered, an upper bound on fi k is given by the largest singular value of all Hessian matrices, plus one. We will assume that our choice of models is such that this curvature does not increase too fast, which could lead to premature convergence of the algorithm to a non-critical point (see [41]). More precisely, we make the following assumption, as in [14], [10], [35] and [41]. This condition is weaker than the common assumption that the model's second derivative matrices are uniformly bounded [32], which holds, for instance, for the classical Newton's method, where quadratic models using analytical second derivatives are used on a compact domain. It is also weaker than the condition for some constant c 0 ? 0, which holds in the case where quadratic element models are used and updated using either the BFGS or the safeguarded Symmetric Rank One quasi-Newton formulae. 2.1.5 Criticality Before we can describe our algorithm in detail, we also need a criticality criterion for our problem. A critical point of our problem is a feasible point x where the negative gradient of the objective function \Gammarf (x) belongs to the normal cone of X at x 2 X , which is defined by fy The associated tangent cone of X at x 2 X is the polar of N (x), that is Thus every measure of criticality has to depend on the (differentiable) objective f and on the geometry of the feasible set at the current point. We will use the symbol ff(x; f; X) to denote such a criticality measure. AS.6 The criticality measure ff(x; h; X) is non-negative for all x 2 X and all functions h differentiable in an open neighbourhood of x. Moreover ff(x; h; only if x is critical for the problem minimize x2X h(x): (2:21) But, within the algorithm, only approximate gradient vectors might be available, namely the vectors g k and g i;k , the gradients of the models. It is therefore natural to use the criticality measure for the problem as an "approximate" criticality measure for (2.1). Note that ff k ? 0 implies that g k 6= 0. In unconstrained optimization, one typically chooses the obvious criticality measure (see [31] or [34]). When bound constraints are present, the choice is made in [10]. For the infinite dimensional case, the definition is used in [41]. For the case where convex constraints are considered, is chosen in [32], where t C is the line coordinate of the so-called "generalized Cauchy point" to be discussed below. In a similar context, is used in [14]. 2.2 Ensuring sufficient model decrease 2.2.1 An overview of the classical sufficient decrease condition A key to trust region algorithms is to choose a step s k at iteration k that is guaranteed to provide a sufficient decrease on the overall objective function model m k . In other words, a step such that is sufficiently positive, given the value of a suitable criticality measure ff k satisfying AS.6. This concept of "sufficient decrease" is usually made more formal by introducing the notion of the (generalized) Cauchy point. This remarkable point, denoted x C k , is typically computed by trust region algorithms as a point on (or close to) the projected gradient path PX that is also within the trust region and sufficiently reduces the overall model in the sense that is a constant and ff k a criticality measure satisfying AS.6. However, such a point may not exist when the trust region radius \Delta k is small compared with ff 2 . In this case, the generalized Cauchy point is chosen as (or close to) the intersection of the projected gradient path with the boundary of the trust region, yielding an inequality of the form A point on the projected gradient path satisfying (2.30) may also fail to exist because the projected gradient path itself ends on the boundary of X , well inside the trust region. In that case, this end point (or another feasible point close to it) is typically chosen as generalized Cauchy point, and it is then typically shown that One then ensures the "sufficient decrease" by requiring that the chosen step s k produces at least a fixed fraction of the overall model reduction achieved by the generalized Cauchy point, which is to say that ae ff k oe Many variants on the above scheme exist in the literature for the unstructured trust region case. All of these variants ensure that a suitable step is found after a finite number of trials. The best known is for unconstrained problems when the ' 2 -norm is used to define the trust region shape. In that case, the projected gradient path is simply given by all negative multiples of the gradient g k and the Cauchy point is simply the point that minimizes the model m k in the intersection of the steepest descent direction and the trust region (see, for instance, [34] and [37]). When other norms are used, for example the ' 1 -norm, one can then choose either to minimize the model in the intersection of this steepest descent direction and the trust region, as before (see [10]), or to "bend" the projected gradient path onto the boundary of the trust region and to choose the generalized Cauchy point as a point which satisfies classical Goldstein-type linesearch conditions along that path while staying within the trust region (see [33] and [41]). Both these latter strategies are used in the LANCELOT software [13]. When additional convex constraints are present, the projected gradient path is additionally "bent" to follow the boundary of the feasible domain. Thus the philosophy is the same, in that (2.33) is guaranteed in the above cases. Indeed satisfaction of this condition has been derived for each of the choices (2.24)-(2.28) for ff k in the papers where they were respectively introduced. 2.2.2 Sufficient decrease for structured model and trust region We will use a similar approach in our structured model and trust region framework to determine what is a sufficient decrease of the overall model m k within the region B k , whose shape is chosen to reflect the structure of the problem. Special care is needed because this region might be very "asymmetric" in the sense that it may allow very large steps in some directions and only very short ones in others. As a consequence, we have to adapt the notion of trust region "radius" to our context and adequately reformulate condition (2.33). From a practical point of view, one might use a two-stage approach. In this, one first aims to find a step producing a sufficient model decrease in a smaller, but more symmetric, region. Following this, one then allows the step to increase within the trust region while maintaining control over the model decrease. To be specific, let be the trust region whose radius is determined by the possibly most nonlinear part of the model. Applying the results discussed in the previous section after condition (2.33), one may deduce that it is possible to find, in a finite number of trials, a step s min;k such that x k and ae ff k oe for some suitably chosen criticality measure ff k satisfying AS.6 and some constant - However, the restriction that the length of s min;k is bounded by \Delta min;k makes the whole exercise of shaping B k to reflect the problem's structure entirely irrelevant. One might therefore be prepared to accept a larger step provided it remains feasible, within the trust region B k , and produces a further significant model decrease. More specifically, we allow our algorithm to choose any step s k such that x k which guarantees that ae ff k ks k k]; 1 oe Note that, since (2.36) holds for s this condition can therefore be achieved in practice after a finite number of trials. Observe also that (2.36) is fundamentally different from an angle test of the form as (2.36) does not prevent s k from being orthogonal to the steepest descent direction, so long as a sufficient model reduction is obtained. This is useful because such a step may occur when moving away from a saddle point of the objective function. Finally note that, as expected, (2.36) reduces to (2.33) in the case where only one trust region is considered. 2.3 A class of structured trust region algorithms We now describe the class of algorithms that we consider for solving (2.1). Besides - 1 used in used in (2.36), it depends on the constants and In addition to the above conditions, we also require a compatibility condition between the j i 's and the - i 's. Specifically, we request that Typical values for these constants are - Algorithm step 0: initialization. The starting point x together with the element function values ff i and the initial trust region radii step 1: model choice. For choose the model m i;k of the element function f i in the trust region B i;k centered at x k (as defined in (2.6)), satisfying (2.7) and (2.8). step 2: determination of the step. Choose a step s k such that the sufficient decrease condition (2.36) holds and step 3: measure overall model fit. If then else step 4: update the element trust region radii. Denote the achieved changes in the element functions and their models by ffif i;k and respectively. Then define the set of negligible elements at iteration k as and the set of meaningful elements as its complement, that is Then, for each i 2 perform the following. Case 1: ffl If and (2.43) both hold, then choose ffl If (2.50) holds but (2.43) fails then choose ffl If (2.50) fails, but holds, then choose ffl If (2.53) fails, then choose Case 2: ffl If and (2.43) both hold, then choose ffl If (2.56) holds but (2.43) fails, then choose ffl If (2.56) fails, then choose Increment k by one and return to step 1. End of Algorithm As is traditional in trust region algorithms, we will call an iteration successful if the test (2.43) is satisfied, that is when the achieved objective reduction ffif k is large enough compared to the reduction predicted by the overall model. If (2.43) fails, the iteration is said to be unsuccessful. In what follows, we will denote by S the set of all successful iterations. We now comment on various aspects of the algorithm. 1. The algorithm is constructed in such a way that a successful step is always possible, for sufficiently small trust region radii, if the current iterate x k is not critical. This result is formally proved in Corollary 8. 2. The choice of the element models m i;k is left rather open in the above description. It clearly needs to be made precise for any practical implementation of the algorithm. One common choice would be to set where H i;k is a symmetric approximation to r 2 f i nullspace contains the subspace In particular, Newton's method corresponds to the choice g which is guaranteed to satisfy this latter condition. Another possible choice is which may be attractive for the simpler element functions. In this case, the model's fit to the true function is always good for the i-th element, and the algorithm guarantees that the \Delta i;k form a non-decreasing sequence. 3. If the model change for an element is negligible, that is small compared to the overall predicted change, we do not need to restrict its element trust region size unless the true element change is relatively large compared with the same overall predicted change. We can therefore afford to ignore negligible items until they stop being relatively negligible, something which is inevitable when convergence occurs. Hence our distinction between "negligible" elements (in N k ) and "meaningful" ones (in M k ). Condition (2.41) can be viewed in this context as a guarantee that a new iterate will be accepted in (2.43) whenever the model reduction obtained for all meaningful elements is also acceptable (i.e. (2.53) holds for all irrespective of the contribution of the negligible ones. This interpretation is clarified in Lemma 2. 4. The apparent intricacy of (2.50) and (2.53) is caused by two complications which arise in the context of multiple elements. The first is that, although (2.36) ensures that always positive, we may not assume in general that the same is true for ffim i;k . The second is that possible cancellation between elements makes it necessary to consider the "accuracy of model fit" for an element to be relative to the overall model fit. Indeed, requiring small relative errors for models with very large values may result in large absolute errors. If these large errors will then cause to be a poor prediction of ffif k and the iteration might be unsuccessful. This explains why the perhaps more intuitive tests cannot be used instead of (2.53) (j = 2) and (2.50) (j = 3). Observe also that conditions (2.50) and (2.53) reduce to the familiar when 5. Note again the consistency between the trust region radii updates in step 4 and the case 1. In this latter case, the set N k is always empty and (2.50) then implies (2.43), because of (2.39). Equation (2.52) is thus never invoked. stopping criterion has been explicitly included in our algorithm description. This is adequate for the theoretical analysis that we consider in the present paper, where we are interested in the asymptotic behaviour of the method, but it should be completed for any practical use. The choice of a particular stopping criterion will depend on the type of models being used. 7. The mechanism that we specified for updating the trust region radii does not exclude the additional requirement that the radii be uniformly bounded, if that is judged suitable for the type of models used. In practice, keeping the radii bounded is essential to prevent numerical overflow. 8. One possible implementation of Step 2 first computes a feasible step s C k that minimizes trust region of radius \Delta min;k . Note that s C satisfies (2.35) and by construction. This step may then subsequently be increased by progressing further along the arc PX long as the overall model m k continues to decrease and holds. Additional decrease in m k may then be obtained (for instance by applying conjugate-gradient steps) provided condition (2.36) is maintained. Before starting our global convergence analysis, we first state, for future reference, some properties that result from the mechanism of the algorithm. Assume that AS.3 holds. At each iteration k of the algorithm, 1. M k contains at least one element. Furthermore 2. for all pg. Proof. The first result immediately follows from the definition of N k and the inequality 1. One then deduces that N k contains at most from which the first part of (2.63) may be deduced. The second inequality in this result is obtained from X the relation (2.48) and jN k 1. The bound (2.64) results from (2.51), (2.54), (2.55), (2.57) and (2.59). 2 We also investigate the coherency between the measure of fit for individual elements and that for the overall model. Assume AS.3 holds and that, at iteration k of the algorithm, (2.53) holds for all and that (2.56) holds for all i 2 N k . Then iteration k is successful, i.e. k 2 S. Proof. Because (2.53) holds for , one has that for all such i, where we used the inequality jM k j - p and Lemma 1 to deduce the second inequality. On the other hand, since (2.56) holds for i 2 N k , one obtains for these i that where we used item 1 of Lemma 1 to bound jN k j. Now, jffif i;k j: (2:69) Combining this last inequality with (2.67) and (2.68) gives that which then yields (2.43) because of (2.41). 2 We observe from this proof that the weaker condition could be imposed instead of (2.41). However (2.71), and hence the setting of the algorithm's constants, would then be problem dependent, which one might consider to be undesirable. Of course, (2.53) holds whenever (2.50) holds because of (2.39). Lemma 2 therefore shows that (2.43) is coherent with the measure of the fit between the element models and element functions. 3 Global convergence We now study the convergence properties of the class of algorithms that we introduced in the preceding section. Our analysis follows the pattern of similar proofs with an unstructured trust region (see [14] or [41]). The central idea in the proof is that the algorithm will continue to make progress as long as a critical point is not reached. We first start by bounding the error between the true element functions and their models. We next derive a lower bound on the size of the smallest trust region radius at a non-critical point. This lower bound ensures that the trust region constraint will not prevent further progress towards a critical point. Only with this bound can we then prove that limit points of the sequence of iterates produced by the algorithm are indeed critical for the models used. We close the section by deriving some simple consequences of these results on the criticality of the limit points for the true objective function. We first start by bounding the error made between the model of any element function and the element function itself at x k Lemma 3 Assume that AS.4 holds and consider a sequence fx k g of iterates generated by the algorithm. Then there exists a positive constant c 1 - 1 such that for all Proof. We first observe that, for each i 2 the definition (2.12), (2.7) and the Cauchy-Schwarz inequality imply that ks i;k k But ks i;k k - \Delta i;k because of (2.6), and hence we obtain from (2.8), (2.15) and (2.16) that Using (2.9), this then yields (3.1) with where the last inequality results from (2.15). 2 We now derive an upper bound on the change predicted for an element at a non-critical point, as a function of the size of the step in the corresponding range subspace. Lemma 4 Assume that AS.1, AS.3 and AS.4 hold. Consider iteration k of the algorithm and assume that, for some Then one has that ks i;k k (3:6) for some constant c 2 ? 0 independent of i and k. Proof. We first note that (2.9), (2.16) and (3.5) imply that Using (2.12) and (2.16), we also obtain that ks i;k ks i;k Remembering now (2.8), (2.6), (3.5) and (3.7), we can deduce that ks ks ks i;k k 2 ks i;k k: Inequality (3.9) then gives (3.6) with i2f1;:::;pg next prove the important fact that, so long as a critical point has not been determined, the trust region radii stay sufficiently bounded away from zero, therefore allowing further progress to be made. Lemma 5 Assume that AS.1-AS.4 hold. Consider a sequence fx k g of iterates generated by the algorithm and assume that there exists a constant ffl ? 0 such that for all k. Then there is a constant c 3 ? 0 such that for all k. Proof. Assume, without loss of generality, that In order to derive a contradiction, assume that there exists a k such that define r to be the smallest iteration number such that (3.14) holds. (Note that r - 1 because of (3.13) and the inequality The monotonic nature of the sequence ffi k g and the bound (2.64) then ensure that where we used (3.14) and the inequality (3.13). We note that the definitions of i and r give that which in turn implies that \Delta because of the monotonic nature of the sequence g. Using this inequality with (2.36), (3.11), and (3.15), we obtain that ae ffl ks oe ae ffl oe which ensures, because of (2.64), that But (3.15) guarantees that fi We may thus apply Lemma 4 and deduce that ks where we also used (2.6) and (3.18). Assume first that i 2 M r\Gamma1 , which guarantees that using (2.48) and (3.18), Because of (2.7), (3.1) and (3.20), we therefore obtain that ffif But (3.14) and (3.15) together give that which, with (3.21), implies that ffif Consider first the case where ffim may then apply (3.19) and deduce that Using (3.23), we now deduce that ffif and therefore, because of (3.24), that which implies that (2.50) holds for element i at iteration r \Gamma 1. Now turn to the case where Because of (3.19), we deduce that As above, we use (3.23) to obtain that ffif and therefore, because of (3.27), that which again implies that (2.50) holds for element i at iteration r \Gamma 1. Assume now that i 2 N r\Gamma1 . Then, because of (2.7), (2.48) and (3.1), we have that multiplying (3.18) by \Delta i;r\Gamma1 , we obtain that Combining (3.30) and (3.31), we deduce that Observing now that (3.14) and (3.15) imply that we obtain from (3.32) that But this inequality implies that (2.56) holds for element i at iteration r \Gamma 1. Thus either (2.50) or (2.56) holds for element i at iteration r \Gamma 1 and the mechanism of the algorithm then implies that But we may deduce from this inequality that which contradicts the assumption that r is the smallest iteration number such that (3.14) holds. The inequality (3.14) therefore never holds and we obtain that (3.12) is satisfied for all k. 2 We now turn to one of the main results in this section, which proves a weak form of global convergence. The technique is inspired by [35]. Theorem 6 Assume that AS.1-AS.6 hold. Consider a sequence fx k g of iterates generated by the algorithm. Then lim inf Proof. Assume, for the purpose of obtaining a contradiction, that there exists an ffl 2 (0; 1) such that (3.11) holds for all k - 0. Then ks k k]; 1 where we used successively (2.43), (2.36), (3.11) and Lemma 5. We note that (3.37) and AS.2 then imply that X Now let r be an integer such that and define the number of successful iterations up to iteration 1). Then define We now wish to show that both sums and are finite. Consider the first. If it has only finitely many terms, its convergence is obvious. Otherwise, we may assume that F 1 has an infinite number of elements, and we then construct two subsequences. The first consists of the indices of F 1 in ascending order and the second, F 3 say, of the set of indices in S (in ascending order) with each index repeated r times. Hence the j-th element of F 3 is no greater than the j-th element of F 1 . This gives that because of the nondecreasing nature of the sequence ffi k g and (3.38). Now turn to the second sum in (3.42). Lemma 2 and the mechanism of the algorithm imply that, at each unsuccessful iteration, at least one element trust region radius satisfies (2.55) or (2.59) and none of them is allowed to increase. Hence p Y Y which immediately implies that where We deduce from this inequality that, for k 2 F 2 , where we have also used Lemma 5 and the definition of F 2 in (3.41). Using (3.39), this gives that and the second sum is convergent. Therefore the sumX is finite, which contradicts AS.5. Hence condition (3.11) is impossible and (3.36) follows. 2 Notice that the relation between ff k , the criticality measure for problem (2.23), and ff(x k ; f; X), the criticality measure for problem (2.1), has been left rather unspecified up to this point. It is indeed remarkable that we can prove Theorem 6 assuming so little on ff. In order to derive convergence properties for the original problem from Theorem 6, we have to be slightly more specific and request that, if both function and model have the same first order information, then the criticality measures on the original problem and on the model problem agree. AS.7 Let h 1 and h 2 be two continuously differentiable functions in the intersection of X with a neighbourhood of the feasible point x, such that h 1 Then, the difference tends to zero. In other words, we require the criticality measure to be continuous (near zero) in the gradient of its second argument. Again, this is true for the choices (2.24)-(2.25) and (2.28). With this additional assumption, we are now ready to examine the criticality of the limit points of the sequence of iterates generated by the algorithm for the original problem (2.1). Corollary 7 Assume that AS.1-AS.7 hold. Consider a sequence fx k g of iterates generated by the algorithm and assume that lim for all pg. Then this sequence has at least one critical limit point x . Proof. From AS.7 and (3.49), we obtain that lim which, with (3.36), guarantees lim inf The desired conclusion then follows by taking a subsequence of fx k g if necessary. 2 Condition (3.49) is important, otherwise the situation might arise that an iterate is critical for the current overall model (because its gradient is inexact) while not being critical for the original problem. There are various ways in which (3.49) can be achieved in a practical algorithm, the simplest being to make the size of e i;k also depend on ff k itself, ensuring that the first goes to zero if the latter does. Corollary 8 Assume that AS.1-AS.7 hold. If S, the set of successful iterations generated by the algorithm is finite, then all iterates x k are equal to some x for k large enough, and x is critical. Proof. Assume indeed that S is finite. It is then clear from (2.45) that x k is unchanged for large enough, and therefore that x is the largest index in S. Note now that Lemma 2 implies that, if k 62 S, then (2.53) or (2.56) must be violated for at least one element. Hence we obtain that \Delta min;k converges to zero. But (2.8) then implies that e i;k also converges to zero for all k converges to rf(x k ). Thus AS.7 and Corollary 7 then guarantee the criticality of x . 2 As in existing theories for the unstructured trust region case, it is possible to replace the limit inferior in (3.36) by a true limit, therefore ensuring (if the gradients are asymptotically exact) that all limit points are critical. As in these theories, a slight strengthening of our assumptions is however necessary. AS.8 We assume that lim This assumption is similar to that used in [14] and [41], where it is motivated in detail. We only mention here that (3.52) holds for Newton's method on bounded domains, because fi k is bounded above in that case. With this additional assumption, we are now able to replace the limit inferior by a true limit. Theorem 9 Assume that AS.1-AS.8 hold. Consider the sequence fx k g of iterates generated by the algorithm and assume that there are infinitely many successful iterations. Then lim where S is, as above, the set of successful iterations. Proof. We again proceed by contradiction. Assume therefore that there exists an ffl 1 2 (0; 1) and a subsequence fq j g of successful iterates such that, for all q j in this subsequence Theorem 6 guarantees the existence of another subsequence fl j g such that where we have chosen ffl 2 2 (0; ffl 1 ). We may now restrict our attention to the subsequence of successful iterations whose indices are in the set where q j and l j belong, respectively, to the two subsequences defined above. Applying now (2.36) we obtain from (2.43), (2.16) and ffl ks k k]; 1 oe ks k k] oe But AS.8, along with (3.57), imply that lim ks and, because of (2.16), that lim ks Therefore, we can deduce from (3.57) and (3.58), that, for j sufficiently large, ks k k where the sums with superscript (K) are restricted to the indices in K, and But AS.2 and the decreasing nature of the sequence ff(x k )g imply that the last right-hand side of (3.60) converges to zero as j tends to infinity. Hence the continuity of rf and AS.7 give that sufficiently large. On the other hand, the second part of (3.59) and (2.8) imply that g q j is arbitrarily close to rf(x q j large enough, and AS.7 hence guarantees that sufficiently large. We note also that, because of (2.8), But the mechanism of the algorithm guarantees that no \Delta i;k can increase between iterations is the largest integer in K that is smaller than l j . This yields that We now deduce from the second part of (3.59) that the left-hand side of (3.65) tends to zero when j tends to infinity, and therefore that, for j sufficiently large, because of AS.7. Combining (3.62), (3.63) and (3.66), we obtain, using (3.55), that which is impossible because of (3.54). Hence our initial assumption cannot hold and the theorem is proved. 2 As above, we now consider the case where we impose that the element gradients are asymptotically exact. Assume that AS.1-AS.8 hold. Consider the sequence fx k g of iterates generated by the algorithm and assume furthermore that (3.49) holds for all pg. Then all limit points of this sequence are critical. Proof. If the set S is finite, the conclusion immediately follows from Corollary 8. If, on the other hand, S has an infinite number of elements, (3.49) implies that g k is arbitrarily close to rf(x k ) and the combination of AS.7 and Theorem 9 ensures the criticality of any limit point of the sequence of successful iterates. 2 Of course, (3.49) might be impossible to achieve in practice, and one might consider the case where we can only assert that lim sup i2f1;:::;pg for some small constant - 3 ? 0. This is the case, for instance, if gradients are approximated by finite differences. Corollary 11 Assume that AS.1-AS.6 and AS.8 hold. Consider the sequence fx k g of iterates generated by the algorithm. Assume furthermore that (3.68) holds and that, for some constant the criticality measure ff satisfies for all x 2 X and all functions h 1 and h 2 continuously differentiable in a neighbourhood of x such that h 1 for each limit point x of the sequence, Proof. As in Corollary 10, the desired conclusion immediately follows from Corollary 8 if S is finite. Assume therefore that S has infinitely many elements. We then deduce that, for all Taking the limit for k tending to infinity in S and using Theorem 9 and (3.68) then gives the desired conclusion. 2 Finally observe that although (3.69) is stronger than AS.7, it is not a very strong condition. For instance, it is satisfied with L for the choices (2.24), and also for (2.25) and (2.26) because of the non-expansive character of the projection operator PX (see [41], for example). The same property also holds for the choice (2.28), as discussed in [14]. Finite identification of the correct active set When applied to constrained problems, trust region algorithms typically use the notion of projected gradient or projected gradient path in order to identify a subset of inequality constraints that are satisfied as equalities. Ultimately, the aim thereby is to identify the constraints satisfied as equalities at the solution well before the solution is reached. The methods then reduce to an unconstrained calculation in the manifold defined by the currently "active" constraints. As a consequence, it is possible to guarantee fast asymptotic rates of convergence when using accurate models, as is the case when analytical second order information of the objective and constraint functions is available. It is possible to show that structured trust regions do not upset the theory developed in the unstructured case: it can indeed be shown that the constraints active at a particular limit point of the sequence of iterates are identified after a finite number of iterations, provided the normals of the active constraints are linearly independent and strict complementarity holds, and provided the step s k+1 satisfies the inequality ks ks k k (k) (4:1) for each k 62 S and for some constant This latter condition is meant to avoid a situation where the successful iterates converge to a critical point while a subsequence of unsuccessful iterates converges to another point with a different active set. It does not constitute a severe restriction in the step selection procedure and is automatically verified if s k is determined by a succession of steps of increasing norm such that they remain feasible, within the trust region ensure (2.36). This is the case, for instance, if truncated conjugate gradients are used for computing the step in the solution of an unconstrained problem (see [37] or [38]). The theory considers the active constraint identification problem from a quite general point of view. The main observation is that a number of the existing theories for active constraint identification are based on the definition of a special criticality measure that satisfies AS.6 while not satisfying AS.7 (see [2] or [3], for instance). Let us denote this measure at iteration k by - ff k . The steps leading to constraint identification are then as follows. 1. The first step is to prove that a sufficient decrease condition of the type (2.33) also holds with - ff k instead of ff k . 2. One then proceeds to prove that lim inf much in the same way as for (3.36). 3. The measure - ff k is also constructed to ensure that it is asymptotically bounded away from zero for all points such that their active set is not identical to that of a (close) critical point. (This, in particular, prevents AS.7 from holding.) 4. Some contradiction is then deduced from these last two properties. However, since this development is rather technical and lengthy, we do not include it in the present paper, but refer the interested reader to [15] for details of the results and additional assumptions. This reference also contains the theory concerning the convergence of the iterates to a single limit point, adapted from [14]. Our experience with the solution of practical problems however indicates that the identification of active constraints is seldom observed in practice before the very last iterations of the algorithm, which makes the results discussed in this section mainly of theoretical interest. Extensions We examine in this section some extensions and variants of the results presented above. 5.1 A hybrid technique One of the possible drawbacks of the algorithm of Section 2.3 is that steps might be constrained to be unnecessarily small in directions corresponding to highly nonlinear element functions. Indeed, the negative effect of inaccurate models for these elements might be compensated by a successful step in directions corresponding to less nonlinear elements. This compromise between the different parts of the objective is, of course, inherent to the classical method using an unstructured trust region. We might try to obtain the best of both classical and structured approaches by using a hydrid technique. In this technique, a global trust region radius \Delta k is recurred for the objective function considered as a single element (using the algorithm analyzed above, which is then equivalent to the classical one), along with the individual radii \Delta i;k . We then define the individual "hybrid" radii by for each i 2 i;k g: (5:2) We can then apply our algorithm with these new quantities, to the effect that well-modelled elements have their associated trust regions possibly extended without having to contract those corresponding to badly-modelled ones, as long as the global agreement is satisfactory. It is not difficult to verify that the theory presented above still holds for this hybrid mod- ification. The key points are to observe that the revised definition of our trust region implies that ae ff k oe which is the classical sufficient decrease condition (2.33), that the inequalities (2.64) are still valid with \Delta i;k replaced by \Delta h i;k , and also that an analogous result to Lemma 5 also holds for the global trust region radius, as is already well-known from the unstructured trust region case (see [14], for instance). 5.2 An alternative definition of success An immediate consequence of inequality (2.63) in Lemma 1 is that it would be possible to replace the condition (2.43) for an iteration to be successful by without altering the developments presented above. Indeed, (2.63) shows the equivalence between (2.43) and (5.4). We have chosen to use seems natural to consider the same collection of elements on both sides of the inequality. 5.3 Weaker sufficient decrease conditions It is remarkable to note that Lemma 5 and Theorem 6 can be proved in a weaker context. Indeed, we could require the weaker sufficient decrease condition ae ff k oe instead of (2.36), and still prove Lemma 5 and Theorem 6. However, we have not been able to prove Theorem 9, nor active constraint identification, with these assumptions, because (5.5) only involves the length of the step in a possibly small subspace of R n . 5.4 Using uniformly equivalent norms Another possible generalization of the theory developed above allows the use of different norms for each element and for each iteration. Let us denote these norms by the . The element trust region definition (2.6) then becomes while the gradient approximation condition (2.8) may be written as where the norm k \Delta k [i;k] is any norm that satisfies for all x; y 2 R n . In particular, one can choose the dual norm of k \Delta k (i;k) defined by With iteration k, we may also associate an overall norm k \Delta k (k) defined on the whole of R n , whose purpose is to reflect the relative weighting of the different elemental norms k \Delta k (i;k) in a global measure. If we assume that all the considered norms are uniformly equivalent, that is if there exists a constant oe - 1 such that, for all x,oe is any pair of the above defined norms, then the theory developed in all the preceding sections is still valid without any substantial modification. Again the details of the proofs in this more general setting are provided in [15]. Note that this extension covers the possible introduction of iteration dependent scaling in a practical implementation of our algorithm, which can be highly desirable for some difficult problems. 6 Conclusions We have shown in this paper that the trust region concept, one of the most powerful tools for building efficient and robust algorithms for optimization, can be extended in a very natural way to reflect the structure of the underlying problem. The algorithm proposed above is indeed a direct generalization of the more usual case where only an unstructured uniform trust region is considered. Similar global convergence properties can be proved for the new algorithm, including the case where dynamic scaling is performed on the variables and the situation where the gradients are only known approximately. It remains to see if this modification of a trust region algorithm will prove efficient in practice and justify the slight additional complexity of the method. Note that the results of preliminary numerical experiments (based on a modification of LANCELOT using the implementation described after the algorithm) have been encouraging. Tests on unconstrained problems from the collection [1] have shown that the new method, although very comparable to LANCELOT in many cases, sometimes produces substantial improvements. However, we anticipate the real power of the concept to appear when minimizing augmented Lagrangians or other penalty-like because scaling is much more critical there than in many of the classical unconstrained test examples. The authors are planning to include the new technique described in this paper within the next release of LANCELOT. One of the nice features of the partially separable functions considered in the present theory is that the objective is a linear combination of its elements. While group partially separability, as used in [12] or [13], has computational advantages in terms of economy of derivative calculation, this structure involves a nonlinear relationship between the elements and the overall function. This seems to make exploiting the link between local and global models much harder. While we would be interested in deriving structured trust region methods for group partially separable functions, the methods would undoubtedly be more complicated and less amenable to analysis. Thus, we are content, in the present paper, to consider the simpler, but nonetheless very general, partially separable structure. Finally, there might be other ways to introduce structure in trust region methods than considering (group) partially separable objective functions. In particular, trust region methods for nonlinearly constrained problems seems attractive candidates for an alternative approach that would separate the trust region(s) on the objective from those on the constraints. Acknowledgments The authors are indebted to Johara Shahabuddin for twice pointing out an unsuitable definition of the sufficient decrease condition (2.36) in Section 2.2.2. --R CUTE: Constrained and Unconstrained Testing Environment. On the identification of active constraints. Convergence properties of trust region methods for linear and convex constraints. Parallel global optimization: numerical methods A trust region algorithm for nonlinearly constrained optimization. Projected gradient methods for linearly constrained problems. On the global convergence of trust region methods using inexact gradient information. A trust region strategy for nonlinear equality constrained optimization. Performance of a multifrontal scheme for partially separable optimization. Global convergence of a class of trust region algorithms for optimization with simple bounds. Testing a class of methods for solving minimization problems with simple bounds on the variables. An introduction to the structure of large scale nonlinear optimization problems and the lancelot project. LANCELOT: a Fortran package for large-scale nonlinear optimization (Release Global convergence of a class of trust region algorithms for optimization using inexact projections on convex constraints. Convergence properties of minimization algorithm for convex constraints using a structured trust region (revised). A global convergence theory for the Dennis-Celis-Tapia trust-region algorithm for constrained optimization Practical Methods of Optimization: Unconstrained Optimization. Is exploiting partial separability useful? The global convergence of partitioned BFGS on problems with convex decompositions and Lipschitzian gradients. On the unconstrained optimization of partially separable functions. Numerical experiments with partially separable optimization problems. On the existence of convex decomposition of partially separable functions. Algorithmic Methods in Optimal Control. An algorithm for minimization using exact second derivatives. Partially separable optimization and parallel computing. Convergence of trust region algorithms for optimization with bounds when strict complementarity does not hold. A method for the solution of certain problems in least squares. An algorithm for least-squares estimation of nonlinear parameters The Levenberg-Marquardt algorithm: implementation and theory Recent developments in algorithms and software for trust region methods. Trust regions and projected gradients. On the solution of large scale quadratic programming problems with bound constraints. A new algorithm for unconstrained optimization. On the global convergence of trust region algorithms for unconstrained optimization. A trust region algorithm for equality constrained optimization. The conjugate gradient method and trust regions in large scale optimization. Towards an efficient sparsity exploiting Newton method for minimization. Global convergence of the partitioned BFGS algorithm for convex partially separable optimization. On large scale nonlinear least squares calculations. Global convergence of a class of trust region methods for nonconvex minimization in Hilbert space. On large scale nonlinear network optimization. LSNNO: a Fortran subroutine for solving large scale nonlinear network optimization problems. A trust region algorithm for equality constrained minimization: convergence properties and implementation. --TR --CTR Nicholas I. M. Gould , Dominique Orban , Philippe L. Toint, GALAHAD, a library of thread-safe Fortran 90 packages for large-scale nonlinear optimization, ACM Transactions on Mathematical Software (TOMS), v.29 n.4, p.353-372, December
partial separability;trust region methods;large-scale optimization;convex constraints;structured problems
589230
Tensor Methods for Large, Sparse Unconstrained Optimization.
Tensor methods for unconstrained optimization were first introduced by Schnabel and Chow [SIAM J. Optim., 1 (1991), pp. 293--315], who described these methods for small- to moderate-sized problems. The major contribution of this paper is the extension of these methods to large, sparse unconstrained optimization problems. This extension requires an entirely new way of solving the tensor model that makes the methods suitable for solving large, sparse optimization problems efficiently. We present test results for sets of problems where the Hessian at the minimizer is nonsingular and where it is singular. These results show that tensor methods are significantly more efficient and more reliable than standard methods based on Newton's method.
Introduction In this paper we describe tensor methods for solving the unconstrained optimization problem where D is some open set containing x . We assume that f is at least twice continuously differentiable, and n is large. Tensor methods for unconstrained optimization are general purpose methods primarily intended to improve upon the performance of standard methods, especially on problems where deficiency. They are also intended to be at least as efficient as standard methods on problems where r 2 f(x ) is nonsingular. Tensor methods for unconstrained optimization base each iteration upon the fourth order model of the objective function f(x) is the current iterate, rf(x c ) and r 2 f(x c ) are the first and second analytic derivatives of f at x c , or finite difference approximations to them, and where the tensor terms at x c , T c 2 ! n\Thetan\Thetan and V c 2 ! n\Thetan\Thetan\Thetan , are symmetric. (We use the notation rf(x c ) \Delta d for )d to be consistent with the tensor notation T c \Delta d 3 and Also, for simplicity, we abbreviate terms of the form dd; ddd, and dddd by d 2 ; d 3 , and d 4 , respectively.) Before proceeding, we define the tensor notation used above. n\Thetan\Thetan . Then for n\Thetan\Thetan\Thetan . Then for The tensor terms are selected so that the model interpolates a small number of function and gradient values from previous iterations. This results in T c and V c being low-rank tensors, which is crucial for the efficiency of the tensor method. The tensor method requires no more function or derivative evaluations per iteration and hardly more storage or arithmetic operations, than a standard method based on Newton's method. Standard methods for solving unconstrained optimization problems are widely described in the literature; general references on this topic include Dennis and Schnabel [9], Fletcher [11], and Gill, Murray, and Wright [13]. In this paper, we propose extensions to standard methods that use analytic or finite difference gradients and Hessians. The standard method for unconstrained optimization, Newton's method, bases each iteration upon the quadratic model of f(x) This method is defined when r 2 f(x c ) is nonsingular, and consists of setting the next iterate x+ to the minimizer of (1.3), i.e., A distinguishing feature of Newton's method is that if r 2 f(x c ) is nonsingular at a local minimizer x , then the sequence of iterates produced by (1.4) converges quadratically to x . However, Newton's method is generally linearly convergent at best if r 2 f(x ) is singular [14]. Methods based on (1.2) have been shown to be more reliable and more efficient than standard methods on small to moderate size problems [18]. In the test results obtained for both non-singular and singular problems, the improvement by the tensor method over Newton's method is substantial, ranging from 30% to 50% in iterations, and function and derivative evaluations. The improvement is even more dramatic for singular problems. Furthermore, the tensor method solves several problems that Newton's method fails to solve. The tensor algorithms described in [18] are QR-based algorithms involving orthogonal transformations of the variable space. These algorithms are very effective for minimizing the tensor model when the Hessian is dense because they are very stable numerically, especially when the Hessian is singular. However, they are not efficient for sparse problems because they destroy the sparsity of the Hessian due to the orthogonal transformation of the variable space. To preserve the sparsity of the Hessian, we have developed an entirely new way of solving the tensor model that employs a sparse variant of the Cholesky decomposition. This makes our new algorithms very well suited for sparse problems. The remainder of this paper is organized as follows. In x2 we briefly review the techniques used to form the tensor model, that were introduced in Schnabel and Chow [18]. In x3 we describe efficient algorithms for minimizing the tensor model when the Hessian is sparse. xx4 and 5 discuss the globally convergent modifications for tensor methods for unconstrained optimiza- tion. These consist of line search backtracking and model trust region techniques. A high level implementation of the tensor method is given in x6. In x7 we describe comparative testing for an implementation based on the tensor method versus an implementation based on Newton's method, and present summary statistics of the test results. Finally, a summary of our work and a discussion of future research is given in x8. 2. Forming the Tensor Model In this section, we briefly review the techniques for forming the tensor model for unconstrained optimization that were introduced in [18]. As was stated in the previous section, the tensor method for unconstrained optimization bases each iteration upon the fourth order model of the nonlinear function f(x) given by (1.2). The choices of T c and V c in (1.2) cause the third order term T c \Delta d 3 and the fourth order to have simple and useful forms. These tensor terms are selected so that the tensor model interpolates function and gradient information at a set of p not necessarily consecutive past iterates x In the remainder of this paper, we restrict our attention to 1. The reasons for this choice are that the performance of the tensor version that allows p - 1 is similar overall to that constraining p to be 1, and that the method is simpler and less expensive to implement in this case. (The derivation of the third and fourth order tensor terms for p - 1 is explained in detail in [18].) The interpolation conditions at the past point x are given by and where Schnabel and Chow [18] choose T c and V c to satisfy (2.1) and (2.2). They first show that the interpolation conditions (2.1) and (2.2) uniquely determine T c \Delta s 3 and V c \Delta s 4 . Multiplying (2.2) by s yields Let ff, fi 2 ! be defined by Then from (2.1) and (2.3) they obtain the following system of two linear equations in the two unknowns ff and fi:2 ff are defined by The system (2.4)-(2.5) is nonsingular, therefore the values of ff and fi are uniquely determined. Hence, the interpolation conditions uniquely determine T c \Delta s 3 and V c \Delta s 4 . Since these are the only interpolation conditions, the choice of T c and V c is vastly underdetermined. Schnabel and Chow [18] choose T c and V c by first selecting the smallest symmetric V c , in the Frobenius norm, for which where fi is determined by (2.4)-(2.5). Then they substitute this value of V c into (2.2), obtaining where This is a set of n linear equations in n 3 unknowns T c (i; j; k), 1 - Finally, Schnabel and Chow [18] choose the smallest symmetric T c and V c , in the Frobenius norm, which satisfy the equations (2.6)-(2.7). That is, min Vc2! n\Thetan\Thetan\Thetan subject to V c \Delta s and min n\Thetan\Thetan subject to T c \Delta s The solution to (2.8) is (s\Omega s\Omega s\Omega s); where the tensor V n\Thetan\Thetan\Thetan is called a fourth order rank-one tensor for which use the notation\Omega to be consistent with [18].) The solution to (2.9) is s\Omega s\Omega b; (2:10) where the notation n\Thetan\Thetan , is called a third order rank-one tensor for which T (i; j; is the unique vector for which (2.10) satisfies (2.6), and is given by determined by the minimum norm problems (2.9) and (2.8) have rank 2 and 1, respectively. This is the key to form, store, and solve the tensor model efficiently. The whole process of forming the tensor model requires only O(n 2 ) arithmetic operations. The storage needed for forming and storing the tensor model is only a total of 6n. For further information we refer to [18]. 3. Solving the Tensor Model when the Hessian is Sparse In this section we give efficient algorithms for finding a minimizer of the tensor model (1.2), when the Hessian is sparse. The substitution of the values of T c and V c into (1.2) results in the tensor model As we stated in x2, we only consider the case where the tensor model interpolates f(x) and rf(x) at the previous iterate, i.e., 1. The generalization for p - 1 is fairly straightforward. This constraint is mainly motivated by our computational results. When we allow p - 1, our test results showed almost no improvement over the case where 1. The tensor method is therefore considerably simpler, and cheaper in terms of storage and cost per iteration. 3.1. Case 1: the Hessian is Nonsingular We show that the minimization of (3.1) can be reduced to the solution of a third order polynomial in one unknown, plus the solution of three systems of linear equations that all involve the same coefficient matrix r 2 f(x c ). For conciseness, we use the notation A necessary condition for d to be a local minimizer of (3.1) is that the derivative of the tensor model with respect to d must be zero. That is, which yields If we first premultiply equation (3.2) by s T on both sides, we obtain a cubic equation (in fi) in the unknowns If we then premultiply equation (3.2) by b T on both sides, we obtain another cubic equation (in fi) in the unknowns fi and ', Thus, we obtain a system of two cubic equations in the two unknowns fi and ' which can be solved analytically. We now show how to compute the solutions of this system of two cubic equations in two unknowns by computing the solutions of a single cubic equation in the unknown fi. Let first calculate the value of ' as a function of fi using equation (3.3), i.e., Note that the denominator of equation (3.5) is equal to zero if either assume that fi 6= 0, otherwise the tensor model would be reduced to the Newton model. Now, would be quadratic in fi, therefore Thus, real valued minimizers of the tensor model (3.1) may exist only if 0: It is easy to check that in order for ' to have a defined cannot be zero. If fi 6= 0 and w 6= 0, we substitute expression for ' into equation (3.4) and obtain which is a third order polynomial in the one unknown fi. The roots of equation (3.6) are computed analytically. We substitute the values of fi into equation (3.5) to calculate the values of '. Then we simply substitute the values of fi and ' into equation (3.2) to obtain the values of d. The major cost in this whole process is the calculation of H After we compute the values of d, we determine which of them are potential minimizers. Our criterion is to select those values of d which guarantee that there is a descent path from x c to x c + d for the model M T among the selected steps, we choose the one that is closest to the current iterate x c in the euclidean norm sens. If the tensor model has no minimizer we use the standard Newton step as the step direction for the current iteration. 3.2. Case 2: the Hessian is Rank Deficient If the Hessian matrix is rank deficient we transform the tensor model given in (3.1) by the following procedure. Let d, and ffi is the new unknown. Substituting this expression for d into (3.1) yields the following tensor model which is a function of ffi, d) 2 d) d)s d) (b T - If we let - d, - d, - we obtain the modified tensor model, d) The advantage of this transformation is that the matrix - H is likely to be nonsingular if the rank of (r 2 f(x c )) is at least n \Gamma 1. A necessary and sufficient condition for - H to be nonsingular is given in the following lemma. Let g and H denote rf(x c ) and r 2 f(x c ), respectively. Lemma 3.1. Let H 2 ! n\Thetan , s css T is nonsingular if and only if M =6 6 6 6 6 4 H cs cs is nonsingular: (Note that the submatrix was premultiplied by the constant c to symmetrize the augmented matrix M .) Proof. We prove that there exists only if there exist H cs cs Suppose first that (H Conversely, if there exists (-v; w) satisfying (3.9), then s T - otherwise, contradicts (3.9). Thus (H singular if and only if M is singular. Corollary 3.2. Let H 2 ! n\Thetan , s css T is nonsingular then H cs has full row rank Proof. follows from Lemma 3.1. Lemma 3.3. Let H 2 ! n\Thetan , css T is nonsingular if and only if H cs has full row rank. Proof. The only if part follows from Corollary 3.2. Now assume H cs has full row rank. Since H has rank n\Theta(n\Gamma1) have full column rank. Since H cs has full row rank, (v From T and H 2 has full column rank, (3.10) is equivalent to (v Thus the n \Theta n matrix is nonsingular. Analogously, the n \Theta n matrix is nonsingular. Therefore is nonsingular. 2 For ffi to be a local minimizer of (3.8) the derivative of the tensor model (3.8) with respect to ffi must be zero. That is, which yields Premultiplying equation (3.12) by s T on both sides results in a cubic equation (in fi) in the two unknowns fis T - fis T - fis T - The premultiplication of equation (3.12) by b T on both sides yields another cubic equation (in fi) in the two unknowns fi and ' Therefore, we obtain a system of two cubic equations in the two unknowns fi and ' which we can solve analytically. Since equation (3.13) is linear in ', we can compute ' as a function of fi, and then substitute its expression into equation (3.14) to obtain an equation in the one unknown fi. Let g, and The denominator of equation (3.15) is equal to zero if either - then equation (3.13) would be quadratic in fi, therefore Hence, real valued minimizers of the tensor model (3.8) may exist only if (1 It is straightforward to verify from (3.14) that for ' to be defined reduces to the following cubic equation in fi Once we calculated the expressions for fi from equation (3.16), we substitute them into the following equation for ' obtained from equation (3.14) If neither - equation (3.14) and obtain fiw \Gamma2 which is a third order polynomial in the one unknown fi. The roots of equation (3.17) are then computed analytically. After we determine the values of fi, we substitute them into equation (3.15) to calculate the corresponding values of '. then, we simply substitute the values of fi and ' into equation (3.12) to obtain the values of ffi . The dominant cost in this whole process is the computation of - Similar to the nonsingular case, a minimizer ffi is selected such that there exists a descent path from the current point x c to x c that it is closest to x c . To obtain the tensor step d we set d to - An appropriate choice of - d is the step used in the previous iteration simply because it has the right scale. To solve linear systems of the form - n\Thetan sparse and we use the augmented matrix M defined in Lemma 3.1. That is, we write H cs cs x The (n in (3.18) is sparse and can be factorized efficiently as long as the last row and column are not pivoted until the last few iterations. In fact, we can combine the nonsingular and singular cases by factorizing H , but we shift to a factorization of the augmented matrix if H is discovered to be singular with rank n \Gamma 1. However, we use a Schur complement method to obtain the solution of the augmented matrix by updating the solution from the system b. This choice was motivated by the fact that the Schur complement method was simpler and more convenient to use than the factorization of the augmented matrix M . We describe this updating scheme in x6. If the Schur complement method shows that M is rank deficient (a case that is very rare in practice), or H has rank less than use the standard Newton step as the step direction for the current iteration. 4. Line Search Backtracking Techniques The line search global strategy we used in conjunction with our tensor method for large sparse unconstrained optimization is similar to the one used for nonlinear equations [4, 6]. This strategy has shown to be very successful for large sparse systems of nonlinear equations. We also found that it is superior to the approach used by Schnabel and Chow [18]. The main difference between the two approaches is that ours always tries the full tensor step first. If this provides enough decrease in the objective function then we terminate, otherwise we find acceptable next iterates in both the Newton and tensor directions and select the one with the lower function value as the next iterate. Schnabel and Chow on the other hand, always find acceptable next iterates in both the Newton and tensor directions and choose the one with the lower function value as the next iterate. In practice, our approach almost always requires fewer function evaluations while retaining the same efficiency in iteration numbers. The global framework for line search methods for unconstrained minimization is given in Algorithm 4.1. Algorithm 4.1. Global Framework for Line Search Methods for Unconstrained Minimization. Let x c be the current iterate d t the tensor step d n is the Newton step and if (minimizer of the tensor model was found) then else Find an acceptable x n in the Newton direction d n using Algorithm A6.3.1 page 325 ([9]) Find an acceptable x t in the tensor direction d t using Algorithm A6.3.1 page 325 ([9]) if f(x n else endif endif else Find an acceptable x n in the Newton direction d n using Algorithm A6.3.1 page 325 ([9]) endif 5. Model Trust Region Techniques The two computational methods that are generally used for approximately solving the trust region problem based on the standard model, subject to jj d jj where ffi c is the current trust region radius, are the locally constrained optimal (or "hook") step, and the dogleg step. When ffi c is shorter than the Newton step, the locally constrained optimal step [16] finds a - c such that jj d(- c takes The dogleg step is a modification of the trust region algorithm introduced by Powell [17]. However, rather than finding a point on the curve d(- c ) such that it approximates this curve by a piecewise linear function in the subspace spanned by the Newton step and the steepest descent direction \Gammarf (x c ), and takes x+ as the point on this approximation for which jj x+ \Gamma x c e.g. [9] for more details.) Unfortunately these two methods are hard to extend to the tensor model, which is a fourth order model. Trust region algorithms based on (5.19) are well defined because it is always possible to find a unique point x+ on the curve such that jj x+ \Gamma x c . Additionally, the value of f(x c )+rf(x c ) along the curve d(- c ) is monotonically decreasing from x c to x n which makes the process reasonable. These properties do not extend to the tensor model which is a fourth order model that may not be convex. Furthermore, the analogous curve to d(- c ) is more expensive to compute. For these reasons, we consider a different trust region approach for our tensor methods. The trust region approach that is discussed in this section is a two-dimensional trust region step over the subspace spanned by the steepest descent direction and the tensor (or standard) step. The main reasons that lead us to adopt this approach is because it is easy to construct, closely related to dogleg type algorithms over the same subspace. This step may be close to optimal trust region step algorithms in practice. Byrd, Schnabel, and Shultz [7] have shown that for unconstrained optimization using a standard quadratic model, the analogous two-dimensional minimization approach produces nearly as much decrease in the quadratic model as the optimal trust region step in almost all cases. The two-dimensional trust region approach for the tensor model computes an approximate solution to subject to jj d jj by performing a two-dimensional minimization, subject to jj d jj where d t and g s are the tensor step and the steepest descent direction, respectively, and ffi c is the trust region radius. This approach will always produce a step that reduces the quadratic model by at least as much as a dogleg type algorithm which reduces d to a piecewise linear curve in the same subspace. At each iteration of the tensor algorithm, the trust region method either solves (5.20), or minimizes the standard linear model over the two-dimensional subspace spanned by the standard Newton step and the steepest descent direction. The decision of whether to use the tensor or standard model is made using the following criterion: if (no minimizer of the tensor model was found) or (rf(x c then selected by trust region algorithm else selected by trust region algorithm endif Before we define the two-dimensional trust region step for tensor methods, we show how to convert the problem subject to jj d jj to an unconstrained minimization problem. First, we make g s orthogonal to d t by performing the Householder transformation: then, we normalize both - g s and d t to obtain: ~ ~ s Since d is in the subspace spanned by the tensor step ~ d t and the steepest descent direction ~ s , it can be written as If we square the l 2 norm of this expression for d and set it to ffi 2 , we obtain the following equation for fi as a function of ff Substituting this expression for fi into (5.25) and then the resulting d into (5.21), yields the global minimization problem in the one variable ff, given by (5.26) bellow. Thus, problems and (5.21) are equivalent. Let g hg = ~ s . c )ff To transform the problem subject to jj d jj to an unconstrained minimization problem, we use the same procedure described above to show that (5.27) is equivalent to the following global minimization problem in the one variable ff: c Algorithm 5.1. Two-Dimensional Trust Region for Tensor Methods Let d t be the tensor step d n the standard step x c the current iterate x+ the next iterate steepest descent direction and ffi c the current trust region radius. ~ are given by (5.23) and (5.24), respectively. ~ d n is obtained in an analogous way to ~ applying transformations (5.22) and (5.23) to it. 1. if tensor model selected then Solve problem (5.26) using the procedure described in Algorithm 3.4 [6] else fstandard Newton model selectedg Solve problem (5.28) using the procedure described in Algorithm 3.4 [6] endif 2. if tensor model selected then ~ s where ff is the global minimizer of (5.26) else fstandard Newton model selectedg ~ s where ff is the global minimizer of (5.28) endif 3. f Check new iterate and update trust region radius.g pred the global step d is successful else decrease trust region go to step 1 endif where pred pred standard Newton model selected. The methods used for adjusting the trust radius during and between steps, are given in Algorithm page 338 ([9]). The initial trust radius can be supplied by the user, if not, it is set to the length of the initial Cauchy step. 6. A High Level Algorithm for the Tensor Method In this section, we present the overall algorithm for the tensor method for large sparse unconstrained optimization. Algorithm 6.1 is a high level description of an iteration of the tensor method, that was described in xx3-5. A summary of the test results for this implementation is presented in x7. Algorithm 6.1. An Iteration of the Tensor Method for Large Sparse Unconstrained Optimization Let x c be the current iterate d t the tensor step and d n the Newton step. 1. Calculate rf(x c ) and decide whether to stop. If not: 2. Calculate r 2 f(x c ). 3. Calculate the terms T c and V c in the tensor model, so that the tensor model interpolates f(x) and rf(x) at the past point. 4. Find a potential minimizer d t of the tensor model (3.1). 5. Find an acceptable next iterate x+ using either a line search or a two-dimensional trust region global strategy. go to step 1. In step 1, the gradient is either computed analytically or approximated by the algorithm A5.6.3 given in Dennis and Schnabel [9]. In step 2, the Hessian matrix is either calculated analytically or approximated by a graph coloring algorithm described in [8]. Note that it is crucial to supply an analytic gradient if the finite difference Hessian matrix requires many gradient evaluations. Otherwise, the methods described in this paper may not be practical, and inexact type of methods may be preferable. The procedures for calculating T c and V c in step 3 were discussed in x2. Step 4 calculates d t as described in xx 3-4. The Newton step d n is also computed as a by product of the minimization of the tensor model. The Newton step d n is the modified Newton step (r 2 f(x c safely positive definite, and - ? 0 otherwise. To obtain the perturbation - we use a modification of MA27 [10] advocated by Gill, Murray, Ponceleon, and Saunders in [12]. In this method we first compute the LDL T of the Hessian matrix using the MA27 package, then change the block diagonal matrix D to D + E. The modified matrix is block diagonal positive definite. This guarantees that the E)L T is positive definite as well. Note, that the Hessian matrix is not modified if it is already positive definite. The tensor and Newton algorithms terminate if jj rf(x c ) jj Another implementation issue that deserves some attention is how to find a solution to the augmented system (3.18), when the Hessian matrix is rank deficient. To do this, we use a Schur complement method to update the solution x obtained from solving Hx = b. This requires that H must have full rank. Thus, some modifications are necessary in order for this method to work. We have modified the factorization phase of MA27 to be able to detect the row and column indices of the first pivot that is less or equal than some given tolerance tol. Note that if the rank of the Hessian matrix is less than we skip this whole updating scheme and perturb the matrix as described in the previous paragraph. We also modified the solve phase of MA27 such that whenever there is a zero pivot, the corresponding solution component is set to zero. This way the solution of is the same as the solution of H e is the matrix H minus the row and column at which singularity occurred. Since y has components, the remaining one, which is also the component corresponding to the zero pivot, is set to 0.) Afterwards, we obtain the solution of an augmented system using a Schur complement method, where the coefficient matrix is the matrix H augmented by two rows and columns, i.e., the (n+ 1)-st row and column are the ones at which singularity was detected, and the (n+2)-nd row and column are cs T and cs, respectively. The Schur complement method is implemented by first invoking MA39AD [1] to form the Schur complement H in the extended matrix, where D is the 2 by 2 lower right submatrix, C is the lower left 2 by n submatrix, and B is the upper right n by 2 submatrix, of the augmented matrix. The Schur complement is then factorized into its QR factors. Next, MA39BD [1] solves the extended system (3.18) using the following well-known scheme: 1. Solve 2. Solve y. 3. Solve 4. 7. Test Results We tested our tensor and Newton algorithms on a variety of nonsingular and singular test problems. In the following we present and discuss summary statistics of the test results. All our computations were performed on a SUN using double precision arithmetic. First, we tested our program on the set of unconstrained optimization problems from the [3] and the MINPACK-2 [2] collections. Most of these problems have nonsingular Hessians at the solution. We also created singular test problems as proposed in [4, 19] by modifying the nonsingular test problems from the CUTE collection as follows. Let be the function to minimize, where f is the number of element functions, and In many cases, F at the minimizer x , and F 0 (x ) is nonsingular. then according to [4, 19], we can create singular systems of nonlinear equations from (7.1) by forming n\Thetak has full column rank with 1 - k - n. Hence, - k. For unconstrained optimization, we simply need to define the singular function From (7.3) and - and we know that r 2 - By using (7.2) and (7.3), we created two sets of singular problems, with r 2 - respectively, by using and respectively. The reason for choosing unit vectors as columns for the matrix A is mainly to preserve the sparsity of the Hessian during the transformation (7.2). For all our test problems we used a standard line search backtracking strategy. All the test problems with the exception of rank problems were ran with analytic gradients and Hessians provided by the CUTE and MINPACK-2 collections. For rank test problems, we have modified the analytic gradients provided by the CUTE collection to take into account the modification (7.2). On the other hand, we used the graph coloring algorithm [8] to evaluate the finite difference approximation of the Hessian matrix. A summary for the test problems whose Hessians at the solution have ranks n, presented in Table 1. The descriptions of the test problems and the detailed results are given in the Appendix. In Table 1 columns "better" and "worse" represent the number of times the tensor method was better and worse, respectively, than Newton's method by more than one gradient evaluation. The "tie" column represents the number of times the tensor and standard methods required within one gradient evaluation of each other. For each set of problems, we summarize the comparative costs of the tensor and standard methods using average ratios of three measures: gradient evaluations, function evaluations, and execution times. The average gradient evaluation ratio (geval) is the total number of gradients evaluations required by the tensor method, divided by the total number of gradients evaluations required by the standard method on these problems. The same measure is used for the average function evaluation (feval) and execution time (time) ratios. These average ratios include only problems that were successfully solved by both methods. We have excluded all cases where the tensor and standard methods converged to a different minimizer. However, the statistics for the "better", "worse", and "tie" columns include the cases where only one of the two methods converges, and exclude the cases where both methods do not converge. We also excluded problems requiring a number of gradient evaluations less or equal than 3 by both methods. Finally, columns "t/s" and "s/t" show the number of problems solved by the tensor method but not by the standard method and the number of problems solved by the standard method but not by the tensor method, respectively. The improvement by the tensor method over the standard method on problems with rank averaging 48% in function evaluations, 52% in gradient evaluations, and 59% in execution times. This is due in part to the rate of convergence of the tensor method being faster than that of Newton's method, which is known to be only linearly convergent with constant3 . On problems with rank the improvement by the tensor method over the standard method is also substantial, averaging 30% in function evaluations, 37% in gradient evaluations, and 34% in execution times. In the test results obtained for the nonsingular problems, the tensor method is 9% worse than the standard method in function evaluations, but 31% and 33% better in gradient evaluations and in execution times, respectively. The main reason for the tensor method requiring on the average more function evaluations than the standard method is because on some problems, the full tensor step does not provide sufficient decrease in the objective function, and therefore the tensor method has to perform a line search in both the Newton and tensor directions, which causes the number of function evaluations required by the tensor method to be inflated. As a result, we intend to investigate other possible global frameworks for line search methods that could potentially reduce the number of functions evaluations for the tensor method. To obtain an experimental indication of the local convergence behavior of the tensor and Newton methods on problems where rank(r 2 f(x examined the sequence of ratios produced by the Newton and tensor methods on such problems. These ratios for a typical problem are given in Table 2. In almost all cases the standard method exhibits local linear convergence with constant near 2, which is consistent with the theoretical analysis. The local convergence rate of the tensor method is faster with a typical final ratio of around 0.01. Whether this is a superlinear convergence remains to be determined. We have done similar experiments for problems with rank(r 2 f(x and the tensor method did not show a faster-than-linear convergence rate, because it did not have enough information since The tensor method solved a total of four nonsingular problems, five rank and 7 rank that Newton's method failed to solve. The reverse never occurred. This clearly indicates that the tensor method is most likely to be more robust than Newton's method. The overall results show that having some extra information about the function and gradient in the past step direction is quite useful to achieve the advantages of tensor methods. 8. Summary and Future Research In this paper we presented efficient algorithms for solving large sparse unconstrained optimization using tensor methods. We described new methods for minimizing the tensor model, that are efficient for problems where the Hessian matrix is large and sparse. Implementations using these tensor methods have been shown to be considerably more efficient especially on problems where Rank tensor/standard pbs solved Average Ratio-Tensor/standard Table 1: Summary of the CUTE and MINPACK-2 test problems using line search Iteration (k) Standard method Tensor method 9 0.600 0.126 14 0.969 22 0.896 26 0.667 28 0.666 Table 2: Speed of convergence on the BRYBND problem with rank(r 2 f(x modified by (7.2), started from x 0 . The ratios in second and third columns are defined by (7. the Hessian matrix has a small rank deficiency at the solution. Typical gains over standard Newton methods range from 40% to 50% in function and gradient evaluations, and in computer time. The size and consistency of the efficiency gains indicate that the tensor method may be preferable to Newton's method for solving large sparse unconstrained optimization problems where analytic gradients and/or Hessians are available. To firmly establish such a conclusion, additional testing is required, including test problems of very large size. On sparse problems where the function or the gradient is expensive to evaluate, the finite difference approximation of the Hessian matrix by the graph coloring algorithm [8] may be very costly. Hence, Quasi-Newton methods may be preferable to use in this case. These are methods, which involve low-rank corrections to a current approximate Hessian matrix. We are currently attempting to extend our tensor methods to Quasi-Newton methods for large sparse unconstrained minimization problems. We also considered solving large sparse structured unconstrained optimization problems using tensor methods. In this variant, we explore the possibility of using exact third and fourth order derivative information. The calculation of these derivatives is simplified using the concept of partial separability, a structure which has already proven to be useful when building quadratic models for large scale nonlinear problems [15]. However, the calculation of the minimizer of this exact tensor model is more problematic because we need to solve a sparse system of nonlinear equations. An obvious approach to solve these equations is to use a Newton-like method. Such a method is characterized by the approximation of the Jacobian used in the Newton process. A simple idea is to use a fixed Jacobian at each step. This has the advantage that the Jacobian will have already been obtained in the current tensor iteration. However, potential slow convergence of such a scheme may make the cost of a tensor iteration prohibitive. We are currently investigating other possible approaches such as a modified Newton's method in which the approximated Jacobian matrix will incorporate more useful information, or an iterative method such as a nonlinear GMRES. This work, a cooperation with Nick Gould [5], will be reported in the near future. We are almost done with the implementation and testing of the two-dimensional trust region global strategy described in x5. This work will be reported in a forthcoming paper. We are also implementing the algorithms discussed in this paper in a software package. This package uses one past point in the formation of the tensor terms, which makes the additional cost and storage of the tensor method over the standard method very small. The package will be available soon. Acknowledgments . We thank Professor Bobby Schnabel for his suggestions on how to minimize the tensor model when the Hessian is rank deficient, Nick Gould for discussing a number of implementation issues, Ta-Tung Chow for reviewing the first draft of the paper, and my CERFACS colleage Jacko Koster for his numerous suggestions. --R The Minpack-2 test problem col- lection CUTE: Constrained and Unconstrained Testing Environment. Solving large sparse systems of nonlinear equations and nonlinear least squares problems using tensor methods on sequential and parallel computers. Tensor methods for large-scale unconstrained opti- mization TENSOLVE: a software package for solving systems of nonlinear equations and nonlinear least squares problems using tensor methods. Approximation solution of the trust region problem by minimization over two-dimensional subspaces Estimating sparse hessian matrices. Numerical methods for unconstrained optimization and nonlinear equations. A set of Fortran subroutines for solving sparse symmetric sets of linear equations. Practical method of optimization Preconditioners for indefinite systems arising in optimization and nonlinear least squares problems. Practical Optimization. Analysis of Newton's method at irregular singularities. On the unconstained optimization of partially separable functions. The Levenberg-Marquardt algorithm: implementation and theory A new algorithm for unconstrained optimization. Tensor methods for unconstrained optimization using second derivatives. Tensor methods for nonlinear equations. --TR
large-scale optimization;tensor methods;unconstrained optimization;sparse problems;singular problems
589234
A Superlinearly Convergent Primal-Dual Infeasible-Interior-Point Algorithm for Semidefinite Programming.
A primal-dual infeasible-interior-point path-following algorithm is proposed for solving semidefinite programming (SDP) problems. If the problem has a solution, then the algorithm is globally convergent. If the starting point is feasible or close to being feasible, the algorithm finds an optimal solution in at most $O(\sqrt{n}L)$ iterations, where n is the size of the problem and L is the logarithm of the ratio of the initial error and the tolerance. If the starting point is large enough, then the algorithm terminates in at most O(nL) steps either by finding a solution or by determining that the primal-dual problem has no solution of norm less than a given number. Moreover, we propose a sufficient condition for the superlinear convergence of the algorithm. In addition, we give two special cases of SDP for which the algorithm is quadratically convergent.
Introduction In this paper we consider the semidefinite programming (SDP) problem: and its associated dual problem: are given data, and are the primal and dual variables, respectively. By G ffl H we denote the trace of (G T H). Without loss of generality, we assume that the matrices C and are symmetric (otherwise, replace C by (C+C T )=2 and A i by Also, for simplicity we assume that A i are linearly independent. Throughout this paper we assume that both (1.1) and (1.2) have finite solutions and their optimal values are equal. Under this assumption, X and (y ; S ) are solutions of (1.1) and (1.2) if and only if they are solutions of the following nonlinear system: Some primal-dual interior-point methods for linear programming have been successfully extended to solve the SDP problems (1.3). For a survey of results obtained before 1993 in this field see the paper of Alizadeh [1]. More recent results can be found in [4, 2, 3, 7, 9, 15]. Kojima, Shindoh and Hara [9], Nesterov and Todd [13], and Monteiro [12] extended some interior-point methods for LP to SDP. In the latter paper Monteiro developed a new formulation of the primal-dual search direction originally introduced in [9]. All above mentioned methods, with the exception of the infeasible-interior-point potential-reduction method of Kojima, Shindoh and Hara [9], require a strictly feasible starting point and therefore are feasible-interior-point methods. More recently, Zhang [16], Kojima, Shida and Shindoh [8] and the present authors (in the first version of the paper) independently proposed new path-following algorithms for SDP. In this version, we have corrected some flaws in the local convergence analysis contained in the first version. Our algorithm is a predictor-corrector method generalizing the interior-point method for linear programming proposed by Mizuno, Todd and Ye [11]. We note that the algorithm of [11] has been also generalized for linear complementarity problems with feasible starting points in [6] and with infeasible starting points in [10, 14]. We also mention that the Mizuno-Todd-Ye predictor-corrector method has been extended to self-scaled cones, which includes SDP, by Nesterov and Todd[13] under the assumption that the starting point is strictly feasible. The algorithm to be presented in the present paper is globally convergent whenever the problem (1.3) has a solution. The starting point does not have to be feasible. In particular we can take as starting point any positive multiples of the identity matrix. If the starting point is feasible or close to being feasible our algorithm finds a solution in at most O( iterations. If the starting point is large enough, then the algorithm terminates in at most O(nL) steps either by finding a solution or by determining that the primal-dual problem has no solution of norm less than a given number. Moreover, we propose a sufficient condition for the superlinear convergence of the algorithm. The sufficient condition is satisfied under conditions (A), (B) and (C) of Kojima, Shida and Shindoh [8]. In addition, we give two special cases of SDP for which the algorithm is quadratically convergent. Superlinearly convergent algorithms tend to perform in practice much better than indicated by their iteration complexity which is based on global linear convergence estimates. Indeed, superlinear convergence has been observed experimentally in efficient practical algorithms. The following notation and terminology are used throughout the paper: the p-dimensional Euclidean space; nonnegative orthant of IR the positive orthant of IR the set of all p \Theta q matrices with real entries; the set of all p \Theta p symmetric matrices; : the set of all p \Theta p symmetric positive semidefinite matrices; : the set of all p \Theta p symmetric positive matrices; the (i; j)-th entry of a matrix M; Tr(M the trace of a p \Theta p matrix, equals 0: M is positive semidefinite; 0: M is positive definite; n: the eigenvalues of M 2 S the largest, smallest, eigenvalue of M 2 S Euclidean norm of a vector and the corresponding norm of a matrix, i.e., Frobenius norm of a matrix; An infeasible-interior-point algorithm We denote the feasible set of the problem (1.3) by and its solution set by F , i.e., The residues of (1.3a) and (1.3b) are denoted by: For any given ffl ? 0 we define the set of ffl-approximate solutions of (1.3) as In what follows we present an algorithm that finds a point in this set in a finite number of steps (provided the problem has a solution). The algorithm will perform in a neighborhood of the infeasible central d g: In our algorithm the positive parameter - will be driven to zero and therefore the residues will also be driven to zero at the same rate as - . We use the following neighborhood of the above central ++ \Theta IR m \Theta S n where fl is a constant such that 1. Throughout the paper we also use the notation: The algorithm depends on two positive parameters ff; fi satisfying the inequalities For example, verify (2.3). At a typical step of our algorithm we are given (X; and obtain a predictor direction (U; w; V ) 2 S n \Theta IR m \Theta S n by solving the linear system We will see later on that the above linear system has a unique solution, which we call the affine scaling direction. If we take a steplength ' along this direction we obtain the points Theoretically we would like to compute the step length However this involves computing the root of a complicated nonlinear equation. In Lemma 2.5 we will show that (2. where and In what follows we assume that a steplength ' satisfying is computed, and we consider the predicted points In case of (which is very unlikely), it is easily seen that a solution (X; is at hand and the algorithm terminates. Now suppose that ' ! 1. Then X and S are symmetric positive definite matrices since - i (X(')S(') - Therefore we can define the corrector direction (U ; w; V ) as the solution of the following linear system We will prove later on that the above linear system has a unique solution. By taking a unit steplength along the corrector direction we obtain a new point Correspondingly, we define Le us note that in setting up the linear systems (2.4) and (2.11) it is not necessary to compute square roots of matrices. Indeed, as pointed out by Monteiro [12], it is easily seen that an equation of the form can be written equivalently under the form Summarizing, we can formally define our algorithm as follows: Algorithm 2.1 Choose (X For do A1 through A5: A3 Find the solution U; w; V of the linear system (2.4), define X; as in (2.10), and set terminate. Find the solution U of the linear system (2.11) and define as in (2.12) and (2.14). In analysing our algorithm we need the following technical results. Lemma 2.2 Suppose that M 2 IR p\Thetap is a nonsingular matrix and E 2 IR p\Thetap has at least one real eigenvalue. Then, Proof. See Lemma 2.6 of Monteiro [12]. Lemma 2.3 (Monteiro [12], Lemma 2.2) Let (X; n\Thetan \Theta IR m \Theta IR n\Thetan is a solution of the linear system: \Deltay for some H 2 IR n\Thetan , and let s F The following corollary is essentially Corollary 2.3 of Monteiro [12]. Corollary 2.4 Let (X; the system of linear equations (2.18) has a unique solution (D Therefore, both linear systems (2.4) and (2.11) have unique solutions. Lemma 2.5 If (X; defined by (2.5) satisfies (2.6) where b ' is given by (2.7) and (2.8). Proof. By definition, we have If we set then, in view of (2.19) and (2.4a), we obtain Therefore,2 Hence, for any given parameter - 2 [0; 1) for all ' 2 [0; min( b '; -)), we must have X(') - -)). Otherwise, there must exist such that X(' 0 )S(' 0 ) is singular, which means However, using (2.16) with (from (2:20)) which contradicts (2.21). Since X(') - 0, its square root X(') 1=2 exists and is uniquely defined. Applying (2.17) of Lemma 2.2 with noting that P (from (2:20)) Therefore, choose which gives ' - b '. Finally, if b for all ' 2 [0; 1), which implies X(1) - 0; '. Before stating our main result let us note that the standard choice of starting points is perfectly centered and satisfies (X required in the algorithm. We will see that if the problem has a solution, then for any ffl ? 0 Algorithm 2.1 terminates in a finite number (say K ffl ) of iterations. If then the algorithm is likely to generate an infinite sequence. However it may happen that at a certain iteration (let us say at iteration which implies that an exact solution is obtained, and therefore the algorithm terminates at iteration K 0 . If this (unlikely) phenomenon does not happen we set Theorem 2.6 For any integer 0 - k ! K 0 , Algorithm 2.1 defines a triple and the corresponding residuals satisfy where and ' j is defined by (2.9). Proof. The proof is by induction. For are clearly satisfied. Suppose they are satisfied for some k - 0. As in Algorithm 2.1 we will omit the index k. Therefore we can write (X; The fact that (2.23) and (2.24) hold for immediately from (2.13) and (2.14). From (2.12) and (2.11a) we have Then, recalling (2.26) and (2.11a), we obtain Hence by applying Lemma 2.3 we deduce Using Lemma 2.3 again, we have which implies that I exists. Using (2.26), (2.27), applying Lemma 2.2 with noting that (from (2:27)) The above inequality implies that Hence which gives S In view of (2.29), this shows that (2.22) holds Finally, (2.25) is an immediate consequence of (2.22). 3 Global convergence and polynomial complexity In this section we assume that F is nonempty. Under this assumption we will prove that Algorithm 2.1, with globally convergent in the sense that lim lim In the sequel, we will frequently use the following inequality: for any M 1 ; M 2 2 IR n\Thetan , (see exercise 20 in section 5.6 of [5]). Lemma 3.1 For any ( f Proof. Let From (2.1), (2.23), it is easily seen that (X which implies X 0 ffl S By expanding (3.2) we obtain the desired result. Lemma 3.2 Assume that F is nonempty. Then for any (X ; y N (ff; -) we have where Proof. The results follow by using Lemma 3.1 with ( f Theorem 2.6 and the fact that S ffl X Lemma 3.2 shows that the pair (X k generated by Algorithm 2.1 is bounded. More precisely, we have the following corollary, which can easily be deduced from Lemma 3.2 and Theorem 2.6. Corollary 3.3 (3. Lemma 3.4 Suppose ( f Then the quantity ffi defined by (2.8) satisfies the inequality: !/ Proof. It is easily seen that (U Hence, according to Lemma 2.3, we have Therefore, and the lemma follows by virtue of (2.8). Lemma 3.5 Under the hypothesis of Lemma 3.4 we have where Proof. Using the notation of Lemma 3.4, and Lemma 3.3, we can write Also, and Then (3.11) follows from Lemma 3.4. According to Lemma 2.5 and Lemma 3.5, it follows that if F is not empty, then the step length ' k defined by (2.9) is bounded away from 0. This implies global convergence as shown in the following theorem. Theorem 3.6 If F is not empty, then Algorithm 2.1 is globally convergent at a linear rate. Moreover, the iteration sequence (X bounded and every accumulation point of belongs to F (i.e., is a primal dual optimal solution of the SDP problem). Using Lemma 3.5, we can easily deduce the following result. Theorem 3.7 Suppose that F is nonempty and that the starting point is chosen such that there is a constant - independent of n satisfying the inequality Then Algorithm 2.1 terminates in at most O( d k; jR 0 Theorem 3.8 Suppose constant such that kX k - . Then the step length ' k defined by (2.9) satisfies the inequality with Proof. According to Lemma 3.2, we have i.e., In the sequel we will frequently use the fact that ff ! 0:5. Since X ffl S we get the relation which implies Applying (3.16), (3.17), Lemma 3.3, and Lemma 3.4 with ( f F F In view of (3.18), (3.19) and Lemma 3.3, we obtain F F Therefore, !/ Consequently, ' ? 1=(!n). In the following corollary we summarize the complexity results for standard starting point of the form X Corollary 3.9 Assume that in Algorithm 2.1 we choose a starting point of the form X constant. Let ffl 0 be given by (3.12) and let ffl ? 0 be arbitrary. Then the following statements hold: (i) If F 6= ;, then the algorithm terminates with an ffl-approximate solution (X F ffl in a finite number of steps (iii) For any choice of ae ? 0 there is an index such that either or, and in the latter case there is no solution (X kg. 4 Superlinear convergence The next two lemmas are well known and can be easily proved. Lemma 4.1 Let A 1 . Then A 1 ffl A . Then, there exists an orthogonal matrix such that Q T X Q and Q T S Q are diagonal matrices. In other words, q are eigen-vectors of X and S . Definition 4.3 A triple (X ; y is called a strictly complementary solution of In this section we investigate the asymptotic behavior of Algorithm 2.1. We will propose a sufficient condition for the superlinear convergence of Algorithm 2.1. Assumption 1. The SDP problem has a strictly complementary solution (X be an orthogonal matrix such that q are eigenvectors of X and S , and define It is easily seen that IB [ ng. For simplicity, let us assume that where B and N are diagonal matrices. Here and in the sequel, if we write a matrix M in the block form then we assume that the dimensions of M 11 and M 22 are jIBj \Theta jIBj and jINj \Theta jINj, respectively. In the next lemma we use the following notations: Lemma 4.4 Under Assumption 1, ks ks Proof. Because the sequence f(X k ; S k )g is bounded, we have ks In view of (3.3b), we get Tr(X where Tr(X and q T Similarly, ks From Theorem 2.6, we obtain which implies i.e., Therefore, Hence, for any i 2 IB we can write ks Also, for any i 2 IN, ks Therefore, kb x Similarly, by considering we can show that kb s Using Lemma 4.4, we can write O( -) O( O( O( Let us define a linear manifold: It is easily seen that if (X Lemma 4.5 Under Assumption 1, F ae M. Proof. For any (X ; y Hence, which implies 4.1, we have which implies i.e., If i or j 2 IB, then q T is positive, which implies q T according to (4.4). Similarly, we can show that q T Therefore, which gives F ae M. Lemma 4.6 Under Assumption 1, every accumulation point of (X strictly complementary solution of (1.3). Proof. Suppose (X "; y"; S") 2 F is an accumulation point. Let us assume, without loss of generality, that (X k ; y k according to Lemma 4.5, for some symmetric positive semidefinite matrices MB and MN . In order to show 0, it remains to prove that MB and MN are nonsingular and therefore positive definite. From Lemma 4.4 or (4.1), we have O( it has an accumulation point. Without loss of generality, we may assume we obtain which implies Hence f must be nonsingular. Obviously f so that MB is nonsingular. Similarly, we can show that MN is nonsingular. In the next theorem, we propose a sufficient condition for the superlinear convergence of Algorithm 2.1. Let us define is the solution of the following minimization problem: and \Gamma is a constant such that k(X k ; S k )k F - \Gamma; 8k. Note that every accumulation point of belongs to the feasible set of the above minimization problem and the feasible set is bounded. Therefore ( - exists for each k. Theorem 4.7 Under Assumption 1, if Algorithm 1.3 is superlinearly convergent. Moreover, if there exists a constant oe ? 0 such that k ), then the convergence has Q-order at least 1 oe in the sense that - Proof. By Lemma 2.5, it remains to prove that us omit the index k. It is easily seen that (U with where Here we have used the relation - Denoting and applying Lemma 2.3, we obtain which implies Similarly, By Lemma 4.4 and the fact that ( - Similarly, Let us observe that Then from (4.8), (4.9), (4.10), (4.11), (4.12) and Corollary 3.3, we have Hence, Finally, if k ) for some constant oe ? 0, then we have Therefore, Recalling (2.25), we obtain - We mention that our sufficient condition in the above theorem is satisfied under conditions (A), (B) and (C) of Kojima, Shida and Shindoh [8]. We end this paper by giving two special cases of SDP for which Algorithm 2.1 is quadratically convergent. Proposition 4.8 If ng or Algorithm 2.1 is quadratically convergent. Proof. Let S) be the solution of the minimization problem: min We will show that We will prove (4.14) only for ng and a similar proof applies to the case ng. The manifold M reduces to Assume by contradiction that (4.14) does not hold, i.e., there exists a subsequence such that Let us define It is easily seen that (\Deltay k \Deltay k depends linearly on \DeltaS k (cf. (4.18)), we deduce that there exists a convergent subsequence of (\DeltaX k ; \Deltay k ; \DeltaS k ). Without loss of generality we can write Letting k !1 in (4.18), we obtain (\Deltay 0 From Lemma 4.4, we have for each i 2 ks which implies that \DeltaS Since is not the solution of the minimization problem (4.13) for sufficiently large k, which is a contradiction. It is easily seen from (4.14) that ( - Hence we can choose \Gamma in (4.6) such that Therefore, From Lemma 4.4 we deduce that if ng or In view of (4.14), (4.20) and (4.21), we have and the result follows from Theorem 4.7. Acknowledgment The authors would like to thank Professor Masakazu Kojima, Masayuki Shida, and Susumu Shindoh for sending us their paper and kindly pointing out some errors in the first version of the present paper. --R Interior point methods in semidefinite programming with applications to combinatorial optimization. Complexity and nondegeneracy in semidefinite programming. An interior-point method for semidefinite programming Matrix Analysis. A predictor-corrector method for linear complementarity problems with polynomial complexity and superlinear convergence Linear algebra for semidefinite programming. Global and local convergence of predictor-corrector infeasible-interior-point algorithms for semidefinite programs A unified approach to infeasible-interior-point algorithms via geometrical linear complementarity problems On adaptive-step primal-dual interior-point algorithms for linear programming A modified O(nL) infeasible-interior-point algorithm for LCP with quadratic convergence Positive definite programming. On extending primal-dual interior-point algorithms from linear programming to semidefinite programming --TR --CTR S. J. Li , S. Y. Wu , X. Q. Yang , K. L. Teo, A relaxed cutting plane method for semi-infinite semi-definite programming, Journal of Computational and Applied Mathematics, v.196 n.2, p.459-473, 15 November 2006 Zhensheng Yu, Solving semidefinite programming problems via alternating direction methods, Journal of Computational and Applied Mathematics, v.193 n.2, p.437-445, 1 September 2006 Stefania Bellavia , Sandra Pieraccini, Convergence Analysis of an Inexact Infeasible Interior Point Method for Semidefinite Programming, Computational Optimization and Applications, v.29 n.3, p.289-313, December 2004
path-following;superlinear convergence;infeasible-interior-point algorithm;polynomiality;semidefinite programming
589237
BFGS with Update Skipping and Varying Memory.
We give conditions under which limited-memory quasi-Newton methods with exact line searches will terminate in n steps when minimizing n-dimensional quadratic functions. We show that although all Broyden family methods terminate in n steps in their full-memory versions, only BFGS does so with limited-memory. Additionally, we show that full-memory Broyden family methods with exact line searches terminate in at most n steps when p matrix updates are skipped. We introduce new limited-memory BFGS variants and test them on nonquadratic minimization problems.
Introduction . The quasi-Newton family of algorithms remains a standard workhorse for minimization. Many of these methods share the properties of finite termination on strictly convex quadratic functions, a linear or superlinear rate of convergence on general convex functions, and no need to store or evaluate the second derivative matrix. In general, an approximation to the second derivative matrix is built by accumulating the results of earlier steps. Descriptions of many quasi-Newton algorithms can be found in books by Luenberger [16], Dennis and Schnabel [7], and Golub and Van Loan [11]. Although there are an infinite number of quasi-Newton methods, one method surpasses the others in popularity: the BFGS algorithm of Broyden, Fletcher, Goldfarb, and Shanno; see, e.g., Dennis and Schnabel [7]. This method exhibits more robust behavior than its relatives. Many attempts have been made to explain this robustness, but a complete understanding is yet to be obtained [23]. One result of the work in this paper is a small step toward this understanding, since we investigate the question of how much and which information can be dropped in BFGS and other quasi-Newton methods without destroying the property of quadratic termination. We answer this question in the context of exact line search methods, those that find a minimizer on a one-dimensional subspace at every iteration. (In practice, inexact line searches that satisfy side conditions such as those proposed by Wolfe, see x4.3, are substituted for exact line searches.) We focus on modifications of well-known quasi-Newton algorithms resulting from limiting the memory, either by discarding the results of early steps (x2) or by skipping some updates to the second derivative approximation (x3). We give conditions under which quasi-Newton methods will terminate in n steps when minimizing quadratic functions of n variables. Although all Broyden family methods (see x2) terminate in n steps in their full-memory versions, we show that only BFGS has n-step termination under limited-memory. We also show that the methods from the Broyden family terminate in n steps even if p updates are skipped, but termination is lost if we both skip updates and limit the memory. y Applied Mathematics Program, University of Maryland, College Park, MD 20742. gibson@math.umd.edu. This work was supported in part by the National Physical Science Con- sortium, the National Security Agency, and the University of Maryland. z Department of Computer Science and Institute for Advanced Computer Studies, University of Maryland, College Park, MD 20742. oleary@cs.umd.edu. This work was supported by the National Science Foundation under grant NSF 95-03126. x Department of Pure and Applied Mathematics, Washington State University, Pullman, WA 99164. nazareth@amath.washington.edu. T. Gibson, D. P. O'Leary, L. Nazareth In x4, we report the results of experiments with new limited-memory BFGS variants on problems taken from the CUTE [3] test set, showing that some savings in time can be achieved. Notation. Matrices and vectors are denoted by boldface upper-case and lower-case letters respectively. Scalars are denoted by Greek or Roman letters. The superscript "T" denotes transpose. Subscripts denote iteration number. Products are always taken from left to right. The notation spanfx 1 denotes the subspace spanned by the vectors x Whenever we refer to an n-dimensional strictly convex quadratic function, we assume it is of the form where A is a positive definite n \Theta n matrix and b is an n-vector. 2. Limited-Memory Variations of Quasi-Newton Algorithms. In this section we characterize full-memory and limited-memory methods that terminate in n iterations on n-dimensional strictly convex quadratic minimization problems using exact line searches. Most full-memory versions of the methods we will discuss are known to terminate in n iterations. Limited-memory BFGS (L-BFGS) was shown by Nocedal [22] to terminate in n steps. The preconditioned conjugate gradient method, which can be cast as a limited-memory quasi-Newton method, is also known to terminate in n iterations; see, e.g., Luenberger [16]. Little else is known about termination of limited-memory methods. Let f(x) denote the strictly convex quadratic function to be minimized, and let g(x) denote the gradient of f . We define is the kth iterate. Let denote the change in the current iterate and denote the change in gradient. Let x 0 be the starting point, and let H 0 be the initial inverse Hessian approximation. For 1. Compute 2. Choose ff k ? 0 such that f(x k 3. Set s 4. Set x 5. Compute 7. Choose H k+1 . Fig. 2.1. General Quasi-Newton Method We present a general result that characterizes quasi-Newton methods, see Figure 2.1, that terminate in n iterations. We restrict ourselves to methods with an update of the form mk Here, L-BFGS Variations 3 1. H 0 is an n \Theta n symmetric positive definite matrix that remains constant for all k, and fl k is a nonzero scalar that can be thought of as an iterative rescaling of 2. P k is an n \Theta n matrix that is the product of projection matrices of the form where is an n \Theta n matrix that is the product of projection matrices of the same form where u is any n-vector 3. m k is a nonnegative integer, w ik n-vector, and z ik vector in spanfs g. We refer to this form as the general form. The general form fits many known quasi-Newton methods, including the Broyden family and the limited-memory BFGS method. We do not assume that these quasi-Newton methods satisfy the secant condition, nor that H k+1 is positive definite and symmetric. Symmetric positive definite updates are desirable since this guarantees that the quasi-Newton method produces descent directions. Note that if the update is not positive definite, we may produce a d k such that d T which case we choose ff k over all negative ff rather than all positive ff. Example. The method of steepest descent [16] fits the general form (2.1). For each k we define Note that neither w nor z vectors are specified since m Example. The 1)st update for the conjugate gradient method with preconditioner fits the general form (2.1) with Example. The L-BFGS update, see Nocedal [22], with limited-memory constant m can be written as k\Gammam k+1;k Y L-BFGS fits the general form (2.1) if at iteration k we choose 4 T. Gibson, D. P. O'Leary, L. Nazareth Observe that P k ; Q k and z ik all obey the constraints imposed on their construction. BFGS is related to L-BFGS is the following way: if we were to use every (s; y) pair in the formation of each update (i.e. we have unlimited memory), we would be creating the same updates as BFGS. In practice, however, one would never do that because it would take more memory than storing the BFGS matrix. Example. We will define limited-memory DFP (L-DFP). Our definition is consistent with the definition of limited-memory BFGS given in Nocedal [22]. Let m - 1 mg. In order to define the L-DFP update we need to create a sequence of auxiliary matrices for where U DFP (H; s; ss T The matrix - k+1 is the result of applying the DFP update m k times to the matrix H 0 with the m k most recent (s; y) pairs. Thus, the 1)st L-DFP matrix is given by To simplify our description, note that - k+1 can be rewritten as k\Gammam k+i k\Gammam k+i k\Gammam k+i k\Gammam k+i y k\Gammam k+i k\Gammam k+j k\Gammam k+j y k\Gammam k+j Y y k\Gammam k+l k\Gammam k+l Thus H k+1 can be written as mk ik k\Gammam k+i k\Gammam k+i y k\Gammam k+i where mk Y y k\Gammam k+j H (j \Gamma1) k\Gammam k+j H (j \Gamma1) L-BFGS Variations 5 Equation (2.7) looks very much like the general form given in (2.1). L-DFP fits the general form with the following choices: k\Gammam k+i y k\Gammam k+i ); and z Except for the choice of P k , it is trivial to verify that the choices satisfy the general form. To prove that P k satisfies the requirements, we need to show Proposition 2.1. For limited-memory DFP, the following two conditions hold for each value of k: Proof. We will prove this via induction. Suppose We have (Recall that spanfs 0 g is trivially equal to spanfH 0 g 0 g.) Furthermore, So we can conclude, and so the base case holds. Assume that We will use induction on i to show (2.10) for the 1)st case. For Using the induction assumptions from the induction on k, we get that 6 T. Gibson, D. P. O'Leary, L. Nazareth Assume that - (induction assumption for i). Next, For values of maps any vector v into and so - is in Using the induction assumptions for both i and k, we get and we can continue the induction on i. If so Hence the induction on i is complete and this proves (2.10) in the (k 1)st case. consider mk k\Gammam k+i k\Gammam k+i Using the structure of V jk and (2.10) we see that Hence, (2.11) also holds in the (k 1)st case. Example. The Broyden Class or Broyden Family is the class of quasi-Newton methods whose matrices are linear combinations of the DFP and BFGS matrices: see, e.g., Luenberger [16, Chap. 9]. The parameter OE is usually restricted to values that are guaranteed to produce a positive definite update, although recent work with SR1, a Broyden Class method, by Khalfan, Byrd and Schnabel [14] may change this practice. No restriction on OE is necessary for the development of our theory. The Broyden class update can be expressed as Variations 7 We sketch the explanation of how the full-memory version fits the general form given in (2.1). The limited-memory case is similar. We can rewrite the Broyden Class update as follows: Hence, where hi y s y It is left to the reader to show that H k y k is in spanfs thus the Broyden Class updates fit the form in (2.1). 2.1. Termination of Limited-Memory Methods. In this section we show that methods fitting the general form (2.1) produce conjugate search directions (The- orem 2.2) and terminate in n iterations (Corollary 2.3) if and only if P k maps spanfy into spanfy for each n. Furthermore, this condition on P k is satisfied only if y k is used in its formation (Corollary 2.4). Theorem 2.2. Suppose that we apply a quasi-Newton method (Figure 2.1) with an update of the form (2.1) to minimize an n-dimensional strictly convex quadratic function. Then for each k before termination (i.e. g k+1 6= 0), As if and only if Proof. (() Assume that (2.15) holds. We will prove (2.12)-(2.14) by induction. Since the line searches are exact, g 1 is orthogonal to s 0 . Using the fact that P 0 y 8 T. Gibson, D. P. O'Leary, L. Nazareth from (2.15), and the fact that z i0 2 spanfs 0 g implies g T see that s 1 is conjugate to s 0 since z i0 w T 0: Lastly, spanfs 0 g, and so the base case is established. We will assume that claims (2.12)-(2.14) hold for that they also hold for The vector g - k+1 is orthogonal to s - k since the line search is exact. Using the induction hypotheses that g - k is orthogonal to fs is conjugate to g, we see that for Hence, (2.12) holds for To prove (2.13), we note that As so it is sufficient to prove that g T We will use the following facts: - k+1 since the v in each of the projections used to form Q - k is in k+1 is orthogonal to that span. since each z i - k is in spanfs is orthogonal to that span. (iii) Since we are assuming that (2.15) holds true, for each there exists can be expressed as P- (iv) For is orthogonal to H 0 y i because g - k+1 is orthogonal to spanfs from (2.14). Thus, Variations 9 0: Thus, (2.13) holds for k. Lastly, using (i) and (ii) from above, maps any vector v into spanfv; s by construction, there exist Hence, so To show equality of the sets, we will show that H 0 g - k+1 is linearly independent of g. (We already know that the basis fH 0 is linearly independent since it spans the same space as the linearly independent set fs and has the same number of elements.) Suppose that H 0 g - k+1 is not linearly indepen- dent. Then there exist OE k , not all zero, such that Recall that g - k+1 is orthogonal to fs g. By our induction assumption, this implies that g - k+1 is also orthogonal to fH 0 g. Thus for any j between 0 and - k, positive definite and g j is nonzero, we conclude that OE j must be zero. Since this is true for every j between zero and k, we have a contradiction. Thus, the set fH 0 is linearly independent. Hence, (2.14) holds for k. Assume that (2.12)-(2.14) hold for all k such that g k+1 6= 0 but that (2.15) does not hold; i.e., there exist j and k such that g k+1 6= 0, j is between 0 and k, and (2. T. Gibson, D. P. O'Leary, L. Nazareth This will lead to a contradiction. By construction of P k , there exist - that By assumption (2.16), - k must be nonzero. From (2.13), it follows that g T Using facts (i), (ii), and (iv) from before, (2.14) and (2.17), we get mk z ik w T ik mk Thus since neither fl k nor - k is zero, we must have but this is a contradiction since H 0 is positive definite and g k+1 was assumed to be nonzero. When a method produces conjugate search directions, we can say something about termination. Corollary 2.3. Suppose we have a method of the type described in Theorem 2.2 satisfying (2.15). Suppose further that H j Then the scheme reproduces the iterates from the conjugate gradient method with preconditioner H 0 and terminates in no more than n iterations. Proof. Let k be such that are all nonzero and such that H i we have a method of the type described in Theorem 2.2 satisfying (2.15), conditions (2.12) - (2.14) hold. We claim that the (k 1)st subspace of search directions, spanfs is equivalent to the 1)st Krylov subspace, g. From (2.14), we know that spanfs g. We will show via induction that spanfH 0 g. This base case is trivial, so assume that for some L-BFGS Variations 11 and from (2.14) and the induction hypothesis, which implies that Hence, the search directions span the Krylov subspace. Since the search directions are conjugate (2.13) and span the Krylov subspace, the iterates are the same as those produced by conjugate gradients with preconditioner H 0 . Since we produce the same iterates as the conjugate gradient method and the conjugate gradient method is well-known to terminate within n iterations, we can conclude that this scheme terminates in at most n iterations. Note that we require that H j g j be nonzero whenever g j is nonzero; this requirement is necessary since not all the methods produce positive definite updates and it is possible to construct an update that maps g j to zero. If this were to happen, we would have a breakdown in the method. The next corollary defines the role that the latest information (s k and y k ) plays in the formation of the kth H-update. Corollary 2.4. Suppose we have a method of the type described in Theorem 2.2 satisfying (2.15). Suppose further that at the kth iteration P k is composed of p projections of the form in (2.2). Then at least one of the projections must have is a single projection (p = 1), then v must be of the form Proof. Consider the case of p = 1. We have where k+1g. We will assume that for some scalars oe i and ae i . By (2.15), there exist - Then and so (2. 12 T. Gibson, D. P. O'Leary, L. Nazareth From (2.13), the set fs is conjugate and thus linearly independent. Since we are working with a quadratic, y As i for all i; and since A is symmetric positive definite, the set fy is also linearly independent. So the coefficient of the y k on the left-hand side of (2.18) must match that on the right-hand side, thus Hence, and y k must make a nontrivial contribution to P k . Next we will show that ae Assume that j is between 0 As j As j Now s j As j is nonzero because A is positive definite. If ae j is nonzero then the coefficient of u is nonzero and so y k must make a nontrivial contribution to P k y j , implying that g. This is a contradiction. Hence, ae To show that ae k 6= 0, consider P k y k . Suppose that ae As k+1 and so This contradicts P k y k 2 spanfy must be nonzero. Now we will discuss that p ? 1 case. Label the u-components of the p projections as for some scalars fl 1 through fl p . We know that L-BFGS Variations 13 and that Thus and since u we can conclude that at least one u i must have a nontrivial contribution from y k . 2.2. Examples of Methods that Reproduce the CG Iterates. Here are some specific examples of methods that fit the general form, satisfy condition (2.15) of Theorem 2.2, and thus terminate in at most n iterations. Example. The conjugate gradient method with preconditioner H 0 , see (2.4), satisfies condition (2.15) of Theorem 2.2 since Example. Limited-memory BFGS, see (2.6), satisfies condition (2.15) of Theorem 2.2 since ae 0 for Example. DFP (with full memory), see (2.8), satisfies condition (2.15) of Theorem 2.2. Consider P k in the full memory case. We have Y For full-memory DFP, H i y 1. Using this fact, one can easily verify that P k y Therefore, full-memory DFP satisfies condition (2.15) of Theorem 2.2. The same reasoning does not apply to the limited-memory case as we shall show in x2.3. The next corollary gives some ideas for other methods that are related to L-BFGS and terminate in at most n iterations on strictly convex quadratics. Corollary 2.5. The L-BFGS (2.5) method will terminate in n iterations on an n-dimensional strictly convex quadratic function even if any combination of the following modifications is made to the update: 1. Vary the limited-memory constant, keeping m k - 1. 2. Form the projections used in V k from the most recent along with any set of other pairs from f(s 3. Form the projections used in V k from the most recent along with linear combinations of pairs from f(s 0 ; y 4. Iteratively rescale H 0 . Proof. For each variant, we show that the method fits the general form in (2.1), satisfies condition (2.15) of Theorem 2.2 and hence terminates by Corollary 2.3. 1. Let m ? 0 be any value which may change from iteration to iteration, and define Y 14 T. Gibson, D. P. O'Leary, L. Nazareth Choose These choices fit the general form. Furthermore, so this variation satisfies condition (2.15) of Theorem 2.2. 2. This is a special case of the next variant. 3. At iteration k, let (- s (i) y (i) k ) denote the ith choice of any linear combination from the span of the set and let (- s (m) Y (- y (i) Choose These choices satisfy the general form (2.1). Furthermore, ae k for some i; and Hence, this variation satisfies condition (2.15) of Theorem 2.2. 4. Let fl k in (2.1) be the scaling constant, and choose the other vectors and matrices as in L-BFGS (2.6). Combinations of variants are left to the reader. Remark. Part 3 of the previous corollary shows that the "accumulated step" method of Gill and Murray [10] terminates on quadratics. Remark. Part 4 of the previous corollary shows that scaling does not affect termination in L-BFGS. In fact, for any method that fits the general form, it is easy to see that scaling will not affect termination on quadratics. 2.3. Examples of Methods that Do Not Reproduce the CG Iterates. We will discuss several methods that fit the general form given in (2.1) but do not satisfy the conditions of Theorem 2.2. L-BFGS Variations 15 Example. Steepest descent, see (2.3), does not satisfy condition (2.15) of Theorem 2.2 and thus does not produce conjugate search directions. This fact is well- known; see, e.g., Luenberger [16]. Example. Limited-memory DFP, see (2.8), with does not satisfy the condition on P k (2.15) for all k, and so the method will not produce conjugate directions. For example, suppose that we have a convex quadratic with Using a limited-memory constant of exact arithmetic, it can be seen that the iteration does not terminate within the first 20 iterations of limited-memory DFP with I. The MAPLE notebook file used to compute this example is available on the World Wide Web [9]. Remark. Using the above example, we can easily see that no limited-memory Broyden class method except limited-memory BFGS terminates within the first n iterations. 3. Update-Skipping Variations for Broyden Class Quasi-Newton Algo- rithms. The previous section discussed limited-memory methods that behave like conjugate gradients on n-dimensional strictly convex quadratic functions. In this sec- tion, we are concerned with methods that skip some updates in order to reduce the memory demands. We establish conditions under which finite termination is preserved but delayed for the Broyden Class. 3.1. Termination when Updates are Skipped. It was shown by Powell [26] that if we skip every other update and take direct prediction steps (i.e. steps of length one) in a Broyden class method, then the procedure will terminate in no more than 2n+1 iterations on an n-dimensional strictly convex quadratic function. An alternate proof of this result is given by Nazareth [21]. We will prove a related result. Suppose that we are doing exact line searches using a Broyden Class quasi-Newton method on a strictly convex quadratic function and decide to "skip" p updates to H (i.e. choose H occasions). Then, the algorithm terminates in no more than n iterations. In contrast to Powell's result, it does not matter which updates are skipped or if multiple updates are skipped in a row. Theorem 3.1. Suppose that a Broyden Class method using exact line searches is applied to an n-dimensional strictly convex quadratic function and p updates are skipped. Let the update at iteration j is not skippedg: Then for all As Furthermore, the method terminates in at most n iterations at the exact minimizer Proof. We will use induction on k to show (3.1) and T. Gibson, D. P. O'Leary, L. Nazareth Then (3.2) follows easily since for all j 2 J(k), As 0: be the least value of k such that J(k) is nonempty; i.e., J(k 0 g. Then g k0+1 is orthogonal to s k0 since line searches are exact, and H k0+1 y since all members of the Broyden Family satisfy the secant condition. Hence, the base case is true. Now assume that (3.1) and (3.3) hold for all values of We will show that they also hold for Case I. Suppose that - k 62 J( - k). Then H - k and J( - any j 2 J( - k), As j and Case II. Suppose that - k 2 J( - k). Then H - k+1 satisfies the secant condition and kg. Now g - k+1 is orthogonal to s k since the line searches are exact, and it is orthogonal to the older s j by the argument in (3.4). The secant condition guarantees that H - k+1 y k , and for we have !/ As j !/ As j In either case, the induction result follows. Suppose that we skip p updates. Then the set J(n cardinality n. Without loss of generality, assume that the set fs i g i2J(n\Gamma1+p) has no zero elements. From (3.2), the vectors are linearly independent. By (3.1), and so gn+p must be zero. This implies that xn+p is the exact minimizer of f . L-BFGS Variations 17 3.2. Loss of Termination for Update Skipping with Limited-Memory. Unfortunately, updates that use both limited-memory and repeated update-skipping do not produce n conjugate search directions for n-dimensional strictly convex qua- dratics, and the termination property is lost. We will show a simple example, limited-memory skipping every other update. Note that according to Corollary 2.4, we would still be guaranteed termination if we used the most recent information in each update. Example. Suppose that we have a convex quadratic with We apply limited-memory BFGS with limited-memory constant and skip every-other update to H. Using exact arithmetic in MAPLE, we observe that the process does not terminate even after 100 iterations [9]. 4. Experimental Results. The results of x2 and x3 lead to a number of ideas for new methods for unconstrained optimization. In this section, we motivate, de- velop, and test these ideas. We describe the collection of test problems in x4.2. The test environment is described in x4.3. Section 4.4.1 outlines the implementation of the L-BFGS method (our base for all comparisons) and xx4.4.2-4.4.7 describe the varia- tions. Pseudo-code for L-BFGS and its variations is given in Appendix B. Complete numerical results, many graphs of the numerical results, and the original FORTRAN code are available [9]. 4.1. Motivation. So far we have only given results for convex quadratic func- tions. While termination on quadratics is beautiful in theory, it does not necessarily yield insight into how these methods will do in practice. We will not present any new results relating to convergence of these algorithms on general functions; however, many of these can be shown to converge using the convergence analysis presented in x7 of [15]. In [15], Liu and Nocedal show that a limited-memory BFGS method implemented with a line search that satisfies the strong Wolfe conditions (see x4.3 for a definition) is R-linearly convergent on a convex function that satisfies a few modest conditions. 4.2. Test Problems. For our test problems, we used the Constrained and Unconstrained Testing Environment (CUTE) by Bongartz, Conn, Gould and Toint. The package is documented in [3] and can be obtained via the world wide web [2] or via ftp [1]. The package contains a large collection of test problems as well as the interfaces necessary for using the problems. The test problems are stored as "SIF" files. We chose a collection of 22 unconstrained problems. The problems ranged in size from to 10,000 variables, but each took L-BFGS with limited-memory constant at least 60 iterations to solve. Table 4.1 enumerates the problems, giving the SIF file name, the dimension (n), and a description for each problem. The CUTE package also provides a starting point 4.3. Test Environment. We used FORTRAN77 code on an SGI Indigo 2 to run the algorithms, with FORTRAN BLAS routines from NETLIB. We used the compiler's default optimization level. Figure 2.1 outlines the general quasi-Newton implementation that we followed. For the line search, we use the routines cvsrch and cstep written by Jorge J. Mor'e T. Gibson, D. P. O'Leary, L. Nazareth No. SIF Name n Description & Reference Extended Rosenbrock function (nonseparable version) [30, Problem 10]. problem [17, Problem 20]. 3 TOINTGOR 50 Toint's operations research problem [29]. 4 TOINTPSP 50 Toint's PSP operations research problem [29]. 5 CHNROSNB 50 Chained Rosenbrock function [29]. 6 ERRINROS 50 Nonlinear problem similar to CHNROSNB [28]. 7 FLETCHBV 100 Fletcher's boundary value problem [8, Problem 1]. 8 FLETCHCR 100 Fletcher's chained Rosenbrock function [8, Problem 2]. 9 PENALTY2 100 Second penalty problem [17, Problem 24]. Problem 5]. 11 BDQRTIC 1000 Quartic with a banded Hessian with band- diagonal variant of the Broyden tridiagonal system with a band away from diagonal [29]. First penalty problem [17, Problem 23]. 14 POWER 1000 Power problem by Oren [25]. MSQRTALS 1024 The dense matrix square root problem by Nocedal and Liu (case 0) seen as a nonlinear equation problem [4, Problem 204]. MSQRTBLS 1025 The dense matrix square root problem by Nocedal and Liu (case 1) seen as a nonlinear equation problem [4, Problem 201]. 17 CRAGGLVY 5000 Extended Cragg & Levy problem [30, Problem test problem [5, Problem 57]. 19 POWELLSG 10000 Extended Powell singular function [17, Problem 13]. Another function with nontrivial groups and repetitious elements [12]. tridiagonal matrix square root problem [4, Problem 151]. 22 TRIDIA 10000 Shanno's TRIDIA quadratic tridiagonal problem [30, Problem 8]. Table Test problem collection. Each problems was chosen from the CUTE package. and David Thuente from a 1983 version of MINPACK. This line search routine finds an ff that meets the strong Wolfe conditions, see, e.g., Nocedal [23]. We used 0:9. Except for the first iteration, we always attempt a step length of 1.0 first and only use an alternate value if 1.0 does not satisfy the Wolfe conditions. In the first iteration, we initially try a step length equal to kg . The remaining line search parameters are detailed in Appendix A. We generate the matrix H k by either the limited-memory update or one of the variations described in x4.4, storing the matrix implicitly in order to save both memory and computation time. We terminate the iterations if any of the following conditions are met at iteration L-BFGS Variations 19 k: 1. The inequality is satisfied, 2. the line search fails, or 3. the number of iterations exceeds 3000. We say that the iterates have converged if the first condition is satisfied. Otherwise, the method has failed. 4.4. L-BFGS and its variations. We tried a number of variations to the standard L-BFGS algorithm. L-BFGS and these variations are described in this subsection and summarized in Table 4.2. 4.4.1. L-BFGS: Algorithm 0. The limited-memory BFGS update is given in (2.5) and described fully by Byrd, Nocedal and Schnabel [22]. Our implementation and the following description come essentially from [22]. Let H 0 be symmetric and positive definite and assume that the m k pairs each satisfy s T We will let and m is some positive integer. We will assume that I and that H 0 is iteratively rescaled by a constant fl k as is commonly done in practice. Then, the matrix H k obtained by k applications of the limited-memory BFGS update can be expressed as \GammaU where U k and D k are the m k \Theta m k matrices given by ae and We will describe how to compute d k in the case that k ? 0. Let x k be the current iterate. Let m Given s , the matrices and the vectors S T 1. Update the n \Theta m k\Gamma1 matrices S k\Gamma1 and Y k\Gamma1 to get the n \Theta m k matrices using s k\Gamma1 and y 2. Compute the m k -vectors S T 3. Compute the m k -vectors S T by using the fact that We already know components of S k g k\Gamma1 from S k\Gamma1 g k\Gamma1 , and likewise for . We need only compute s T and do the subtractions. 20 T. Gibson, D. P. O'Leary, L. Nazareth No. Reference Brief Description x4.4.1 L-BFGS with no options. Allow m to vary iteratively basing the choice of m of kgk and not allowing m to decrease. Allow m to vary iteratively basing the choice of m of kgk and allowing m to decrease. Allow m to vary iteratively basing the choice of m of kg=xk and not allowing m to decrease. Allow m to vary iteratively basing the choice of m of kg=xk and allowing m to decrease. 5 x4.4.3 Dispose of old information if the step length is greater than one. 6 x4.4.4, Variation 1 Back-up if the current iteration is odd. 7 x4.4.4, Variation 2 Back-up if the current iteration is even. 8 x4.4.4, Variation 3 Back-up if a step length of 1.0 was used in the last iteration. 9 x4.4.4, Variation 4 Back-up if kg k k ? kg Back-up if a step length of 1.0 was used in the last iteration and we did not back-up on the last iteration. and we did not back-up on the last iteration. neither of the two vectors to be merged is itself the result of a merge and the 2nd and 3rd most recent steps taken were of length 1.0. 13 x4.4.5, Variation 2 Merge if we did not do a merge the last iteration and there are at least two old s vectors to merge. 14 x4.4.6, Variation 1 Skip update on odd iterations. update on even iterations. Alg. 5 & Alg. 8 Dispose of old information and back-up on the next iteration if the step length is greater than one. Alg. 13 & Alg. 1 Merge if we did not do a merge the last iteration and there are at least two old s vectors to merge, and allow m to vary iteratively basing the choice of m of kgk and not allowing m to decrease. 19 Alg. 13 & Alg. 3 Merge if we did not do a merge the last iteration and there are at least two old s vectors to merge, and allow m to vary iteratively basing the choice of m of kg=xk and not allowing m to decrease. Alg. 13 & Alg. 2 Merge if we did not do a merge the last iteration and there are at least two old s vectors to merge, and allow m to vary iteratively basing the choice of m of kgk and allowing m to decrease. Alg. 13 & Alg. 2 Merge if we did not do a merge the last iteration and there are at least two old s vectors to merge, and allow m to vary iteratively basing the choice of m of kg=xk and allowing m to decrease. Table Description of Numerical Algorithms 4. Compute k . Rather than recomputing U k , we update the matrix from the previous iteration by deleting the leftmost column and topmost row if m and appending a new column on the right and a new row on the bottom. Let ae 1=s T be the (m lower right submatrix of U and let (S T be the upper L-BFGS Variations 21 Note that s T so is already computed. 5. Assemble Y T We have already computed all the components. 6. Update D k using D k\Gamma1 and s T 7. Compute Note that both y T 8. Compute two intermediate values 9. Compute The storage costs for this are very low. In order to reconstruct H k , we need to store diagonal matrix) and a few m-vectors. This requires only 2mn Assuming m !! n, this is much less storage than the n 2 storage required for typical implementation of BFGS. Step Operation Count 9 Table Operations Count for Computation of H k g k . Steps with no operations are not shown. The computation of Hg takes at most O(mn) operations assuming n ?? m. (See Table 4.3.) This is much less than the O(n 2 normally needed to compute Hg when the whole matrix H is stored. We are using L-BFGS as our basis for comparison. For information on the performance of L-BFGS see Liu and Nocedal [15] and Nash and Nocedal [19]. 4.4.2. Varying m iteratively: Algorithms 1-4. In typical implementations of L-BFGS, m is fixed throughout the iterations: once m updates have accumulated, m updates are always used. We considered the possibility of varying m iteratively, preserving finite termination on convex quadratics. Using an argument similar to that presented in [15], we can also prove that this algorithm has a linear rate of convergence on a convex function that satisfies a few modest conditions. We tried four different variations on this theme. All were based on the following linear formula that scales m in relation to the size of kgk. Let m k be the number of iterates saved at the kth iteration, with Here, think of m as the maximum allowable value of m k . Let the convergence test be given by kg k k=kx k k ! ffl. Then the formula for m k at iteration k is ae log oe 22 T. Gibson, D. P. O'Leary, L. Nazareth Alg. No. Table The number of failures of the algorithms on the 22 test problems. An algorithm is said to have "failed" on a particular problem if a line search fails or the maximum allowable number of iterations (3000 in our case) is exceeded. We have two choices for ffi k , and a choice of whether or not we will allow m k to decrease as well as increase. The four variations are 1. 2. 3. 4. We used four values of m: 5,10,15 and 50, for each algorithm. The results are summarized in Tables 4.4 - 4.8. More extensive results can be obtained [9]. Table 4.4 shows that these algorithms had roughly the same number of failures as L-BFGS. Table 4.5 compares each algorithm to L-BFGS in terms of function evaluations. For each algorithm and each value of m, the number of times that the algorithm used as few or fewer function evaluations than L-BFGS is listed relative to the total number of admissible problems. Problems are admissible if at least one of the two methods solved it. We observe that in all but three cases, the algorithm used as few or fewer function evaluations than L-BFGS for over half the test problems. Table 4.6 compares each algorithm to L-BFGS in terms of time. The entries are similar to those in Table 4.5. Observe that Algorithms 1-4 did very well in terms of time, doing as well or better than L-BFGS in nearly every case. For each problem in each algorithm, we computed the ratio of the number of function evaluations for the algorithm to the number of function evaluations for L- BFGS. Table 4.7 lists the means of these ratios. A mean below 1.0 implies that the algorithm does better than L-BFGS on average. The average is better for the algorithms in 6 out of 16 cases for the first four algorithms. Observe, however, that all the means are close to one. L-BFGS Variations 23 Alg. No. m= 5 5 19/22 20/22 20/22 21/22 7 8/22 12/22 10/22 10/22 8 12/22 14/22 12/22 15/22 9 6/22 13/22 12/22 16/22 13 3/22 4/22 4/22 4/22 14 2/21 2/22 2/22 2/21 19 2/22 3/22 4/22 4/22 Table Function Evaluations Comparison. The first number in each entry is the number of times the algorithm did as well as or better than normal L-BFGS in terms of function evaluations. The second number is the total number of problems solved by at least one of the two methods (the algorithm and/or L-BFGS). We experience savings in terms of time for the first four algorithms. These algorithms will tend save fewer vectors than L-BFGS since m k is typically less than m; and so less work is done computing H k g k in these algorithms. Table 4.8 gives the mean of the ratios of time to solve for each value of m in each algorithm. Note that most of the ratios are far below one in this case. These variations did particularly well on problem 7. See [9] for more information. 4.4.3. Disposing of old information: Algorithm 5. We may decide that we are storing too much old information and that we should stop using it. For example, we may choose to throw away everything except for the most recent information whenever we take a big step, since the old information may not be relevant to the new neighborhood. We use the following test: If the last step length was bigger than 1, dispose of the old information. The algorithm performed nearly the same as L-BFGS. There was substantial deviation on only one or two problems for each value of m, and this seemed evenly divided in terms of better and worse. From Table 4.4, we see that this algorithm successfully converged on every problem. Table 4.5 shows that it almost always did as well or better than L-BFGS in terms of function evaluations. However, Table 4.7 shows that the differences were minor. In terms of time, we observe that the algorithm generally was faster than L-BFGS (Table 4.6), but again, considering the mean ratios of time (Table 4.8), the differences were minor. The method also does particularly well on problem 7 [9]. 4.4.4. Backing Up in the Update to H: Algorithms 6-11. As discussed in x2.2, if we always use the most recent s and y in the update, we preserve quadratic termination regardless of which older values of s and y we use. T. Gibson, D. P. O'Leary, L. Nazareth Alg. No. 5 15/22 13/22 14/22 15/22 7 11/22 11/22 10/22 7/22 9 9/22 10/22 7/22 8/22 13 5/22 10/22 13/22 17/22 14 2/21 2/22 2/22 3/21 19 11/22 11/22 17/22 19/22 Table Time Comparison. The first number in each entry is the number of times the algorithm did as well as or better than normal L-BFGS in terms of time. The second number is the total number of problems solved by at least one of the two methods (the algorithm and/or L-BFGS). Using this idea, we created some algorithms. Under certain conditions, we discard the next most recent values of s and y in the H although we still use the most recent s and y vectors and any other vectors that have been saved from previous iterations. We call this "backing up" because it as if we back-up over the next most recent values of s and y. These algorithms used the following four tests to trigger backing up: 1. The current iteration is odd. 2. The current iteration is even. 3. A step length of 1.0 was used in the last iteration. 4. kg k k ? kg In two additional algorithms, we varied situations 3 and 4 by not allowing a back-up if a back-up was performed on the previous iteration. The backing up strategy seemed robust in terms of failures. In 4 out of the 6 variations we did for this algorithm, there were no failures at all. See Table 4.4 for more information. It is interesting to observe that backing up on odd iterations (Algorithm backing up on even iterations (Algorithm 7) caused very different results. Backing up on odd iterations seemed to have almost no effect on the number of function evaluations (Table 4.7) and little effect on the time (Table 4.8). However, backing up on even iterations causes much different behavior from L-BFGS. It does worse than L-BFGS on most problems, but better on a few. Algorithms were two variations of the same idea: backing up if the previous step length was one. This wipes out the data from the previous iteration after it has been used in one update. Both show improvement over L-BFGS in terms of function evaluations; in fact, these two algorithms have the best function evaluation ratio for the case (Table 4.7). Unfortunately, these algorithms did not compete with L-BFGS in terms of time (Table 4.8). There is little difference between L-BFGS Variations 25 Alg. No. m= 5 6 1.000 1.000 1.000 1.000 9 1.035 1.371 1.005 0.947 14 7.521 7.917 8.288 8.502 19 1.212 1.959 1.242 1.387 Table Mean function evaluations ratios for each algorithm compared to L-BFGS. Problems for which either method failed are not used in this mean. Algorithms probably because there were rarely many steps of length one is a row. Algorithms 9 and 11 are also two variations of the same idea: back-up on iteration the norm of g k is bigger than the norm of g k+1 . There is a larger difference between the results of 9 and 11 than there was between 8 and 10. In terms of function evaluation ratios (Table 4.7), Algorithm 11 did better, indicating that it may not be wise to back-up twice in a row. Both of these did poorly in terms of time as compared with L-BFGS (Table 4.8). 4.4.5. Merging s and y information in the update: Algorithms 12 and 13. Yet another idea is to "merge" s data so that it takes up less storage and computation time. By merging, we mean forming some linear combination of various s vectors. The y vectors would be merged correspondingly. Corollary 2.5 shows that as long as the most recent s and y are used without merge, old s vectors may be replaced by any linear combination of the old s vectors in L-BFGS. We used this idea in the following way: if certain criteria were met, we replaced the second and third newest s vectors in the collection by their sum, and did similarly for the y vectors. We used various tests to determine when we would do a merge: 1. Neither of the two vectors to be merged is itself the result of a merge and the second and third most recent steps taken were of length 1.0. 2. We did not do a merge the last iteration and there are at least two old s vectors to merge. The first variation (Algorithm 12) performs almost identically to L-BFGS, especially in terms of time (Table 4.5). Occasionally it did worse in terms of time Table 4.6). These observations are also reflected in the other results in Table 4.7 and Table 4.8. It is likely that very few vectors were merged. The second variation (Algorithm 13) makes gains in terms of time, especially for 26 T. Gibson, D. P. O'Leary, L. Nazareth Alg. No. 6 1.007 0.983 0.977 0.995 9 1.032 1.220 1.043 1.173 14 4.585 3.703 3.228 2.417 Table Mean time ratios for each algorithm compared to L-BFGS. Problems for which either method failed are not used in this mean. the larger values of m (Table 4.6 and Table 4.8). Unfortunately, this reflects only a saving in the amount of linear algebra required. The number of function evaluations generally is larger for this algorithm than L-BFGS (Table 4.5 and Table 4.7). 4.4.6. Skipping Updates to H: Algorithms 14-16. If every other update to H is skipped and a step length of one is always chosen, BFGS will terminate in 2n iterations on a strictly convex quadratic function. The same holds true when doing an exact line search. (See x3.) Unfortunately, neither property holds in the limited-memory case. We will, however, try some algorithms motivated by this idea. The idea is that, every so often, we do not use the current s and y to update H, and instead just use the old H. There are three variations on this theme. 1. Skip update on odd iterations. 2. Skip update on even iterations. 3. Skip update if kg k+1k ? kg k k. As with the algorithms that did back-ups, the results of the skipping on odd or even iterations were quite different. Skipping on odd updates (Algorithm 14) did extremely well for every value of m on only two problems: 1 and 12. Otherwise, it did very badly. Skipping on even updates (Algorithm 15) performed somewhat better. It did extremely well on problem 7 but not on problems 1 and 12. It also did better than L-BFGS in terms of time on more occasions than Algorithm 14 (Table 4.6). Neither did well in terms of function evaluations, but the mean ratios for function evaluations Table 4.7) and time (Table 4.8) were usually far greater than one. Skipping the update if the norm of g increased (Algorithm 16) did not do well at all. It only did better in terms of function evaluations for one problem for each value of m ( Table 4.5) and rarely did better in terms of time (Table 4.6). It ratios were very bad for function evaluations (Table 4.7) and time (Table 4.8) L-BFGS Variations 27 4.4.7. Combined Methods: Algorithms 17-21. We did some experimentation with combinations of methods described in the previous sections. In Algorithm 17, we combined Algorithms 5 and 8: we dispose of old information and back-up on the next iterations if the step length is greater than one. Essentially we are assuming that we have stepped out of the region being modeled by the quasi-Newton matrix if we take a long step and we should thus rid the quasi-Newton matrix of that information. This algorithm did well in terms of function evaluations, having mean ratios of less than one for three values of m (Table 4.7), but it did not do as well in terms of time. In Algorithms 19-21, we combined merging and varying m. These algorithms did well in terms of time for larger m (Table 4.8) but not in terms of function evaluations Table 4.7). 5. Conclusions. There is a spectrum of quasi-Newton methods, ranging from those that require the storage of an n \Theta n approximate Hessian (e.g. the Broyden fam- ily) to those that require only the storage of a few vectors (e.g. conjugate gradients). Limited-memory quasi-Newton methods fall in between these extremes in terms of performance and storage. There are other methods that fall into the middle ground; for example, conjugate gradient methods such as those proposed by Shanno [27] and Nazareth [20], the truncated-Newton method [24, 6] and the partitioned quasi-Newton method [13]. We have characterized which limited-memory quasi-Newton methods fitting a general form (2.1) have the property of producing conjugate search directions on convex quadratics. We have shown that limited-memory BFGS is the only Broyden family member that has a limited-memory analog with this property. We also considered update-skipping, something that may seem attractive in a parallel environment. We show that update skipping on quadratic problems is acceptable for full-memory Broyden class members in that it only delays termination, but that we lose the property of finite termination if we both limit memory and skip updates. We have also introduced some simple-to-implement modifications of the standard limited-memoryBFGS algorithm that seem to behave well on some practical problems. Appendix A. Line Search Parameters. Table A.1 give the line search parameters used for our code. Note that in the first iteration, the initial steplength is rather than 1.0. Variable Value Description STP 1.0 Step length to try first. \Gamma4 Value of ! 1 in Wolfe conditions. GTOL 0.9 Value of ! 2 in Wolfe conditions. Relative width of interval of uncertainty. Maximum number of function evaluations. Table A.1 Line Search Parameters Appendix B. Pseudo-Code. B.1. L-BFGS: Algorithm 0. The pseudo-code for the computation of d \GammaH k g k at iteration k for L-BFGS is given in Figure B.2. The update of H is also handled implicitly in this computation. 28 T. Gibson, D. P. O'Leary, L. Nazareth if (sze == else % This is needed for Step 3 before we overwrite Stg and Ytg Fig. B.1. MATLAB pseudo-code for the computation of d = Hg in L-BFGS. sze is the number of s vectors available for the update this iteration and oldsze is the number of s vectors that were available the previous iteration. For L-BFGS, sze is chosen as the minimum of oldsze (the limited-memory constant). B.2. Varying m iteratively: Algorithms 1-4. Suppose that m k denotes the number of (s; y) pairs to be used in the kth update. Then simply chose sze as the minimum of oldsze computing d k . B.3. Disposing of old information: Algorithm 5. If the disposal criterion is met at iteration k, set oldsze to zero and sze to one before computing d k . B.4. Backing Up in the Update to H: Algorithms 6-11. If we are to back-up at iterations k, set oldsze to the one less than the previous value of sze and set sze as the minimum of oldsze m, as usual. B.5. Merging s and y information in the update: Algorithms 12 and 13. Merging is the most complicated variation to handle. Before we determine the newest sze and before we compute d k , we execute the pseudo-code given in Figure B.1. We then set oldsze to one less than the previous value of sze and set sze as the minimum of oldsze m, as usual. We are assuming we are at iteration k, but that the Variations 29 newest values of s and y have not yet been added to S and Y. Execute before choosing new value for sze and before computing d Fig. B.2. MATLAB pseudo-code for the merge variation. This fixes the values of the components that are used in the computation of d k . B.6. Skipping Updates to H: Algorithms 14-16. To skip the update at iteration k, set sze to oldsze. Compute Stg and Ytg before Step 0 and then skip to Step 8 and continue. --R ftp://thales. Test functions for unconstrained minimization Performance of a multifrontal scheme for partially separable optimization Numerical Methods for Unconstrained Optimization and Nonlinear Equations An optimal positive definite update for sparse Hessian matrices http://www. Matrix Computations Private communication to authors of Partitioned variable metric updates for large structured optimization problems A theoretical and experimental study of the symmetric rank one update On the limited memory BFGS method for large scale optimization Linear and Nonlinear Programming A numerical study of the limited memory BFGS method and the truncated-Newton method for large scale optimization A relationship between BFGS and conjugate gradient algorithms and its implications for new algorithms Updating quasi-Newton matrices with limited storage A discrete Newton algorithm for minimizing a function of many variables Quadratic termination properties of minimization algorithms I. Conjugate gradient methods with inexact line searches An error in specifying problem CHNROSNB. --TR --CTR Sun Linping, Updating the Self-Scaling Symmetric Rank One Algorithm with Limited Memory for Large-Scale Unconstrained Optimization, Computational Optimization and Applications, v.27 n.1, p.23-29, January 2004 Adi Ditkowski , Gadi Fibich , Nir Gavish, Efficient Solution of A, Using A-1, Journal of Scientific Computing, v.32 n.1, p.29-44, July 2007 M. Al-Baali, Extra-updates criterion for the limited memory BFGS Algorithm for large scale nonlinear optimization, Journal of Complexity, v.18 n.2, p.557-572, June 2002
update skipping;limited-memory;broyden family;BFGS;minimization;quasi-Newton
589240
Interior Point Trajectories in Semidefinite Programming.
In this paper we study interior point trajectories in semidefinite programming (SDP) including the central path of an SDP. This work was inspired by the seminal work of Megiddo on linear programming trajectories [ Progress in Math. Programming: Interior-Point Algorithms and Related Methods, N. Megiddo, ed., Springer-Verlag, Berlin, 1989, pp. 131--158]. Under an assumption of primal and dual strict feasibility, we show that the primal and dual central paths exist and converge to the analytic centers of the optimal faces of, respectively, the primal and the dual problems. We consider a class of trajectories that are similar to the central path but can be constructed to pass through any given interior feasible or infeasible point, and study their convergence. Finally, we study the derivatives of these trajectories and their convergence.
Introduction The purpose of this paper is to study properties of the trajectories associated with interior point methods for semidefinite programming (SDP) prob- lems. Since many aspects of semidefinite programming find close analogs in linear programming, several interior point methods designed for linear programming (LP) have been successfully extended to apply to semidefinite programming (e.g., see [2], [4], [10], [12], [17], [19], [20], [21], [25]). Many Research supported in part by NSF Grants DMS 91-06195, DMS 94-14438 and DMS 95-27124 and DOE Grant DE-FG02-92ER25126 y This author was supported in part by an IBM Cooperative Fellowship of these aspects have also been studied in the more general framework of self-scaled cones in [20], [21]. Many interior point methods can be viewed as iterative approximations to continuous path-following methods. Our aim is to provide a theoretical basis for such methods for SDP by describing the limiting behavior of the continuous central path and related trajectories for such problems. This work is an extension of the linear programming results in [15] to semidefinite programming. We characterize the optimal face of an SDP problem and prove that the central path converges to the analytic center of the optimal face. Unlike LP problems, an SDP problem does not always have a strictly complementary primal-dual pair of solutions (e.g, see [3], [12]). Thus the SDP central path cannot be guaranteed to converge to such a pair as it does in LP. However, we show that it converges to a "least nonstrictly complementary" pair, in the sense that the sum of the ranks of the primal and the dual solutions (viewed as matrices) is as large as possible. Another issue that makes SDP different from LP is the absence (at least as far as we know) of a suitable concept of a weighted central path. Given that it is difficult in practice to obtain a point on the central path, it is important to have a class of trajectories that have properties similar to the properties of the central path and that pass through any given pair of interior primal and dual solutions. Such trajectories for linear programming are introduced in [6] and [1] and are called primal affine scaling (PAS) trajectories due to the fact that they correspond to continuous versions of primal affine scaling iterative algorithms. We study the SDP analogs of PAS-trajectories and prove that the main convergence results of [1] hold. We show that under the assumptions of primal and dual nondegeneracy and strict complementarity defined in [3], the first order derivatives of the central path are bounded in the limit. We also provide formulae for the limit of these derivatives and show that the factorization of only one matrix is required to compute these and all higher order derivatives of a solution on the central path. The paper is organized as follows. In Section 2 we describe the central path for a primal-dual pair of SDP problems and introduce our basic assumptions and some notation. In Section 3 we characterize the optimal faces of the primal and the dual SDP problems, and prove our main convergence result for the primal-dual central path in Section 4. We extend the results of Section 4 to the shifted central path (an analog of the PAS-trajectory) in Section 5. Finally, in Section 6 we analyze the limiting properties of the derivatives of the central path and show that computation of the derivatives requires factorizing a single matrix for all orders of the derivatives. 2 The Central Path In this paper we consider the semidefinite programming problem, henceforth referred to as the primal problem, n\Thetan denotes the space of real . The inner product on S n\Thetan is we mean that X is positive semidefinite (positive definite). The problem dual to (P ) is the semidefinite programming problem: (D) s:t: Throughout the paper the following are assumed to hold: Assumption 2.1 The matrices A are linearly independent; i.e., implies that u Assumption 2.2 Both the primal and the dual problem have interior feasible solutions, i.e. and Under Assumption 2.2 both primal and dual problems have finite optimal solutions, - X and (-y; - Z), and the duality gap - [19]. The optimal solutions also satisfy - The central path for the problem (P ) is a trajectory of the solutions n\Thetan to the following parametric family of problems for values of the parameter - ? 0 ([13],[19],[25]): From Assumption 2.2 and the strict convexity of the logarithmic barrier objective function for any - ? 0, problem (P - ) has a unique solution that satisfies the Karush-Kuhn-Tucker optimality conditions for (P - The central path for the dual problem can be defined in an analogous manner and is the trajectory (y(-); Z(-)) n\Thetan whose points satisfy the same system (CP - ) as the points X(-) on the primal central path. Hence it makes sense to refer to the trajectory (X(-); y(-); Z(- ? 0 of solutions to (CP - ) as the primal-dual central path. Under Assumption 2.2, not only does this path exist, but also it converges to an optimal primal-dual solution (e.g, see [13], [19], [26]). To conclude this section we introduce some notation that we will use later in the paper. First, we note that the variables X and Z can be viewed both as symmetric matrices and as vectors (obtained from these matrices by stacking their columns one after the other), lying in a n(n+1)=2-dimensional subspace of R n 2 . Whenever we refer to the matrix X as a vector, we denote it by vec(X). By the constraint matrix A we denote the m \Theta n 2 matrix, the i-th row of which equals . Note that C ffl the usual inner product. The Kronecker product M\Omega N of matrices M 2 R n\Thetan and N 2 R n\Thetan is defined as There are two properties of the Kronecker product that we will need later: T\Omega M)vec(N) and (M\Omega P (MN)\Omega (PS) [7]. If X is a positive semidefinite symmetric matrix, then X has a spectral is an orthogonal matrix of eigenvectors of X and is a diagonal matrix with the eigenvalues of X on the diagonal. Throughout this paper the upper case letter Q will always denote a matrix with orthonormal columns and and\Omega will always denote diagonal matrices of eigenvalues. Lastly, from properties of the trace we have Property 2.3 Let A 2 R n\Thetan , X 2 R r\Thetar and P 2 R n\Thetar . Then A ffl Property 2.4 Let A 2 R n\Thetan , A - 0 and B 2 R n\Thetan , B - 0. Then 3 Optimal Faces of the Primal and Dual Problem Properties of the faces of the cone of positive semidefinite matrices, are studied in [5]. The facial structure of semidefinite programming problems (i.e., the intersection of the cone of positive semidefinite matrices with an affine subspace) is studied in general terms in [22], [23]. Here we derive a particular system which describes the optimal face of an SDP problem. Let us introduce some more notation and recall some well-known facts. Let R(X) denote the range (column space) of X . If X is a positive semidefinite symmetric matrix, it can be factorized as where is a diagonal matrix whose diagonal elements are the positive eigen-values of X and Q is a matrix with orthogonal columns that are eigenvectors corresponding to these eigenvalues. Clearly, span(Q), the subspace spanned by the columns of Q, and the dimension of this subspace (i.e., the number of positive eigenvalues of X) equals the rank of X . Z be an optimal primal-dual pair of solutions. It is well known that they can be represented as - and\Omega are diagonal matrices with the positive eigenvalues of - Z, respectively, on their diagonals and Q T and R( - X). O P denote the primal optimal face, i.e., the set of primal optimal solutions, and let OD denote the dual optimal face. Note that both O P and OD are convex subsets of affine subspaces of S n\Thetan . By ri O P (ri OD ) we denote the relative interior of O P (OD ). Then the following lemma holds [5]: Lemma 3.1 For any - any ~ riO D ), R( - Z)). This lemma shows that any ~ riO P is an optimal solution of maximum rank. Moreover, if both - X and ~ X are in ri O P , it follows from Lemma 3.1 that R( - X). Let us denote this subspace by R P . Analogously, let RD be the subspace spanned by the eigenvectors corresponding to the positive eigenvalues of Z for any dual solution (y; Z) in the relative interior of OD . Let dim R From the complementarity of any primal-dual pair of optimal solutions R P ?RD . Hence, r we say that the primal-dual pair of problems does not satisfy strict complementarity. Note that this can never happen in linear programming. If we define we have a partition of R n into three mutually orthogonal subspaces. be any n \Theta r matrix whose columns form an orthonormal basis for R P . Then any solution X O P can be written as so the optimal face of (P ) is given by the set of the solutions to the following system: Indeed, for any U feasible for (1), Q P is feasible for (P ), and since and Z satisfy complementary slackness. Therefore, Q is an optimal solution to (P ). Similarly, let the columns of Q D form an orthonormal basis for RD . Then any optimal dual solution can be written as the optimal face of (D) is given by the set of solutions to the system: Notice that the definitions of the primal and dual optimal faces are invariant with respect to the choices of Q P and Q D as long as their columns form orthonormal bases for the subspaces R P and RD , respectively. The following lemma shows that under Assumptions 2.1 and 2.2 both primal and dual optimal faces are bounded. Thus their analytic centers are well defined, which is important for the results of the next section. Lemma 3.2 Let Assumptions 2.1 and 2.2 hold, then the optimal sets of the primal and the dual problems are bounded. Proof. Suppose that the set of optimal dual solutions is unbounded. That is, there exists a nonzero direction (u; V ), satisfying: Multiplying the second equation by an interior feasible primal solution X which exists by Assumption 2.2, we obtain Therefore It then follows that which by Assumption 2.1 implies that The boundedness of the set of optimal primal solutions can be proved in a similar manner. 2 4 Convergence of the Central Path We prove in this section that the primal central path converges to the analytic center of the optimal face O P . First we show that the limit - X of the central path is in the relative interior of the optimal face. Then we show that - is, in fact, the analytic center of the optimal face. We then extend these results to the dual central path. In [27] it is shown that in the case of convex homogeneous self-dual cones, which includes the case of the cone of positive semidefinite matrices, the central path converges to a strictly complementary solution provided that one exists. In [14], under the assumption of strict complementarity, it is shown that the primal-dual central path of an SDP problem converges to the analytic center of the optimal face. We obtain the same results without assuming strict complementarity. X be the limit of the primal central path as - ! 0. Lemma 4.1 There exists a spectral factorization - and a sequence such that X(- k , where spectral factorization of X(- k ). Proof. The proof follows trivially from the compactness of the set of the orthogonal matrices. Notice that the limit - is uniquely defined by - the limit - Q, generally speaking, depends on the sequence f- k g. 2 We know that - X and (-y; - Z) are optimal solutions to the primal and dual problems, respectively. We first want to prove that each is in the relative interior of the optimal face for its respective problem. Lemma 4.2 - Z)) belongs to the relative interior of the primal (dual) optimal face. Proof. Let (X(-); y(-); Z(-)) be a point on the central path. Let ~ and (~y; ~ riO D . It is trivial to verify that and since ~ ~ Now both terms on the left side of (3) are nonnegative by Proposition 2.4; hence ~ Consider the sequence f- k g as defined in Lemma 4.1, such that X(- k X and the spectral factorizations ~ ~ (The columns of ~ Q are eigenvectors of ~ X that span R P .) Let us order the columns of Q and partition - into two parts [ - QND ] so that - Q P has r columns and - ~ Q P is nonsingular. This is always possible since - Q P has full column rank. Let us order the columns of Q(- k ) and the columns and rows of - and (- k ) and partition them according to the column order and partitioning of - Q. Then - ND (- k ), and from (4) ~ ~ ~ ~ ~ Since both terms in this sum are nonnegative by Proposition 2.4, it follows from Property 2.3 that ~ ~ ~ ~ ~ U P ~ ~ ~ U P ~ ~ It then follows from (5) that r the sum of the ratios ( - finite, it follows that X has rank r proving that - ri O P . Similarly, it can be shown that (-y; - ri OD . 2 From Lemma 3.2 it follows that the analytic center of the optimal face O P is well defined. We now show that - is, in fact, this analytic center. Theorem 4.3 Let - X be the limit of the primal central path as - ! 0. Then U depends on he choice of the orthonormal basis Q P for R P and is the unique solution to the problem X is the analytic center of the primal optimal face. Proof. Problem (6) can be rewritten in an equivalent form: where c(0) is the optimal objective function value. The solution of this problem is unique and satisfies the following system of optimality conditions: For any fixed - ? 0, let C ffl is the solution to the problem (CP - ). Then X(-) is a solution to the following problem: Using notation of Lemma 4.2 - QND ] is a matrix of eigenvectors of - P is a diagonal matrix of positive eigenvalues of - X (r of them) and - 4.2 it follows that - As in Lemma 4.2 consider the convergent sequence X(- k to - Q and (- k ) converges to - as and (- k ) is diagonal, requiring X in (9) to be of the form where U - 0 and V is equal to ND (- k ), does not affect the solution of problem (9). We restrict V and not U because, as we have already shown, the sequence of the solutions converges to an optimal solution, where and U is a positive definite matrix. Also, . Therefore from (9) and Property 2.3 we obtain the following maximization problem: The unique optimal solution satisfies the following system: As Q and the system (11) converges to (8) with Q Since P (- k ) is the solution to (11), then the limit - has to satisfy (8). This proves that - X is the analytic center of the primal optimal face. 2 As in LP, problems (P ) and (D) can be written in a "symmetric" form. Specifically, we can use the "conic" formulation given in Chapter 3 of [19] (see also [25]). Let L be the subspace of S n\Thetan spanned by A and D 2 S n\Thetan be such that A i ffl and (D) can be formulated as and (D Consequently, all the results in this section extend to the dual problem, and in particular, in terms of formulation (D) of the dual, we have the following: Theorem 4.4 Let (-y; - Z) be a limit point of the dual central path. Then W depends on the choice of the orthogonal basis Q D for RD and is the unique solution to the problem Z) is the analytic center of the dual optimal face. 5 Shifted Central Paths In this section we present a class of primal affine scaling trajectories analogous to those introduced by Bayer and Lagarias [6] and later studied by Adler and Monteiro [1]. Affine scaling vector fields associated with semidefinite programs are studied in [8] and [9]. Here we analyze the limiting behavior of affine scaling trajectories in SDP. As far as we know there is no suitable concept of a scaled or weighted central path, defined as a trajectory of solution of a class of minimization problems, passing through any given pair of primal and dual interior solu- tions. Therefore, we do not consider weighted trajectories as in [15]. How- ever, we can consider "shifted" central paths, or primal affine scaling (PAS) trajectories, as they are called in [1] or A-trajectories as they are called in [6]. We study the properties and convergence of these trajectories, using the same techniques that we used for the central path. In [1] it is shown that the tangent to a PAS trajectory at any given point has the same direction as the primal affine scaling step. The same is true in semidefinite programming [9]. Consider the family of problems dependent on a parameter - ? 0 n\Thetan is some arbitrary fixed symmetric matrix. If problem (SP - ) has a solution for some - ? 0 then that solution is unique and satisfies the Karush-Kuhn-Tucker necessary conditions The trajectory of dual solutions (y(-); Z(-)) defined by the system (SCP - ), parametrized by -, is generally different from the dual shifted central path associated with T . Thus, when referring to the shifted central path, we mean the primal and dual trajectories defined by (SCP - ). For any given - ? 0 and T 2 S n\Thetan if there exists a (y n\Thetan such that -T and Z 0 - 0 then (SCP - ) has a unique solution. Using the notation of [1], let Y (T ; and I(T ;g. By Assumption 2.2, Y (T ; T , and it is easy to see that I(T ) is an open interval, which is nonempty as long as C \Gamma -T is not spanned by the matrices A i m). Thus Lemma 5.1 If the feasible set of problem (P ) is bounded, then I(T (0; 1) for any T 2 S n\Thetan ; i.e., (SCP - ) has a unique solution for all 1. This lemma is proved in [24] and a similar, but more general, result is proved as Theorem 2.4 in [9]. We want to choose T so that the shifted central path passes through an arbitrary given primal-dual pair of interior solutions. Specifically, given (X then it is easy to verify that I(T consequently, that I(T ) 3 In other words this choice guarantees the existence of a trajectory passing through the primal-dual point (X It is shown in [1] (Proposition 2.4) that in linear programming the choice of the initial dual solution (y does not affect the PAS trajectory. The same is true in the case of the shifted central path (i.e., PAS trajectory) for a semidefinite program. The proofs are identical. We are ready to discuss the limiting behavior of the shifted central path. Z) be a limit point of the solution (X(-); y(-); Z(-)) to (SCP - ) In [9] it is shown that - X is an optimal solution to (P ) and (-y; - is an optimal solution to (D). As in the case of the central path, we can show that - X is in the relative interior of the primal optimal face and (-y; - is in the relative interior of the dual optimal face. The proof is analogous to the proof of Lemma 4.2 with the exception that X(-) ffl Z(- N- for some large number N . More importantly, it is trivial to extend the proof of Theorem 4.3 to give Theorem 5.2 - U is the unique solution to the problem X is the "shifted" analytic center of the primal optimal face. Just as in the case of LP [1], the dual solutions of the shifted central path converge to the analytic center (not shifted) of the dual optimal face. Theorem 5.3 Let (-y; - Z) be a limit dual point of the shifted central path, then - W is the unique solution to the problem Z) is the analytic center of the dual optimal face. Proof. First we observe that the dual solution (y(-); Z(-)) associated with the system of optimality condinions (SCP - ) is the unique optimal solution to the system s:t: Given a sequence Z(- k Z, we know that Z(- k Z. Let be a sequence of dual solutions on the shifted central path, which converges to (-y; - Z) and for which the sequence of spectral factorizations converges. By a similar argument to that used in the proof of Theorem 4.3, =\Omega D (- k ) is a solution to the following problem When the solution to the above problem converges to the solution to the problem which is equivalent to Problem (14) defining the analytic center of the dual optimal face. Thus (-y; - W ) is the unique solution of Problem (14). 2 Thus, the choice of the initial point (X affects the limit of the trajectory of the primal solutions (obviously, only if the optimal face is of dimension greater than zero), but does not affect the limit of the dual trajectory. Remark. Notice, that the dual problem (15) is in fact a shifted barrier problem for the original dual (D). We now consider the tangent to a shifted central path at an arbitrary point on it. Our results apply to the central path as a special case Let (X; (we omit the argument - for simplicity) be on the shifted central path corresponding to a given shift T . Differentiating the system respect to - for any - ? 0, yields In [9] it is shown that this system of differential equations is generated by the generalized primal affine scaling vector field. In our terms, this is equivalent to the fact that we can rewrite the above system as y, - X\Omega X and H (i.e., the rows of - A equal and vec( - can be viewed as an orthogonal projection of a scaled objective vector onto the kernel of a scaled constraint matrix, where the scaling depends only on the primal solution. The direction of the tangent is and the directions - y and - Z can be calculated from the dual estimates y E and Z , which in turn, can be computed from the projection operator. Remark. We would like to study the limiting behavior of the dual estimates Z that are computed at every step of an algorithm that uses an affine scaling direction. In the next section we show that under assumptions of strict complementarity and primal and dual nondegeneracy the limit of (y the limit of (y(-); Z(-)) as - ! 0. 6 Derivatives of the Central and Shifted Central We begin this section by showing that as in case of linear programming [15] the computation of derivatives of any order of solutions on the central path or a shifted central path involves inverting a single matrix. In contrast with LP, the Schur complement of this matrix, which we must factorize, is fully dense, even if the constraint matrices are sparse. Consequently, this factorization step is very expensive in SDP and it is desirable to use as much information (e.g. higher order derivatives) as possible from it, as in the interior point LP methods proposed in [18] and [16]. Let us consider solutions X and (y; Z) on the shifted central path corresponding to a shift T . ( X , y and Z depend on -, but we omit the argument.) As shown in the previous section (see (17)) the derivatives - y in vector form satisfy y For the second derivatives we have y and it can be easily shown by induction that the k-th derivatives X (k) and y (k) on the central path must satisfy a system of equations of the form y where R is a function of (T ; X; - We now turn to the limiting properties of the first order derivatives of the central path. As earlier, let ( - Z) be the limit of the central path. We need the following assumption: Assumption 6.1 i) The primal and dual solutions - X and (-y; - are strictly ii) The primal solution - X is nondegenerate; i.e., the matrices QD are linearly independent in S n\Thetan . iii) The dual solution (-y; - Z) is nondegenerate; i.e., the matrices span S r\Thetar . For these and other equivalent definitions of primal and dual nondegeneracy see Alizadeh, Haeberly, and Overton [3]. They also prove that under Assumption 6.1 the optimal primal-dual solution is unique. Hence if Assumption 6.1 holds it makes sense to say that Problems (P ) and (D) are nondegenerate and strictly complementary. For the remainder of this section we shall assume that Assumption 6.1 holds. To show that the first order derivatives of the central path converge to finite limits as - ! 0, we consider the following system which, as shown in [4], is equivalent to (CP - ):2 (XZ Viewing all symmetric matrices in the above system as vectors in R n(n+1)=2 and differentiating, we obtain the following system6 4 Z I 0 X I y where svec maps a matrix in S n\Thetan into a vector in R n(n+1)=2 and denotes "symmetric" Kronecker product. See [4] for definitions and properties of svec and . Alizadeh, Haeberly, and Overton in [4] prove that under Assumption 6.1 the coefficient matrix of (21) is nonsingular at the limit Therefore, - y and - Z are bounded and converge as - ! 0. Geometrically this means that the central path approaches the boundary of the feasible region at a strictly positive angle; i.e., the angle between the tangent to the central path and the tangent to the boundary at the optimal solution is strictly positive. From X, and the boundedness of the derivative of X we conclude that - X, where - X(0) is the right derivative of X(-) at Similar arguments prove that the derivatives of the primal and dual solutions on the shifted central path are bounded and converge as - ! 0. We obtain system similar to (21), with the same coefficient matrix and the right hand side, which depends on the shift T and on X and is uniformly bounded as - ! 0. Let us consider the derivatives of the eigenvalues of X and Z on the central path. Since is symmetric and is a smooth function of a single parameter, its eigenvalues can be ordered so that - of X is in C 1 [11]. If - i is an eigenvalue of multiplicity k then - the vector of eigenvalues of the matrix where Q i is a matrix of eigenvectors of X corresponding to - i . If and - does not generally converge; however, the subspace spanned by the columns of Q i converges as - ! 0 and therefore any accumulation point of the matrices (22) has the same eigenvalues. Thus the limit of - exists. The same can be shown for the derivatives - i of the of the the dual slack matrix Z. Let us consider the eigenvalues of X and Z that converge to zero. The multiplicity of the zero eigenvalue of - s. The multiplicity of the zero eigenvalue of - Z is r. We know that for X and Z on the central path, and therefore Z as - ! 0, Considering the above system in more detail, we obtain QD QD QD QD From the diagonal blocks of this matrix equa- tion, it then follows that and Therefore, we can conclude that there are orderings of the eigenvalues of - and - Z , for which - D and P . This generalizes the result on the limits of the derivatives of nonbasic variables (i.e., variables converging to zero) in linear programming. Complete information on - Z can be obtained by solving the system (21). We can now conclude, that the dual estimates y y and Z Z that appear in (18) converge to the same limits as y and Z, since - y and - Z are bounded as - ! 0. Remark. To prove the convergence of the derivatives of the central path as - ! 0 we assumed primal and dual nondegeneracy and strict com- plementarity. These assumptions imply the uniqueness of the primal and dual solutions [3]. However, we conjecture that it is sufficient to only assume strict complementarity. Note that strict complementarity is necessary for boundedness of the derivatives, since if there is an index i such that both primal and dual eigenvalues - i (- then from hence that both - cannot be finite at the limit as - ! 0. In [14] it is shown that assuming strict complementarity, O(-) and which implies that the derivatives of the central path are bounded as - ! 0. Acknowledgement . The authors are grateful to Yurii Nesterov for bringing the question of the convergence of the central path in SDP to their attention, to Michael Overton for many stimulating discussions on SDP, and to Jean-Pierre Haeberly and two anonymous referees for helpful comments on earlier versions of the paper. --R "Limiting behavior of the affine scaling continuous trajectories for linear programming problems" "Interior point methods in semidefinite programming with applications to combinatorial optimization" "Complementarity and nondegeneracy in semidefinite programming" "Primal-Dual Interior-Point Methods for Semidefinite Programming: Convergence Rates, Stability and Numerical results" "Cones of diagonally dominant matri- ces" "The nonlinear geometry of linear pro- gramming. I. Affine and projective scaling trajectories" Kronecker Products and Matrix Calculus: with Applica- tions "On a matrix generalization of affine-scaling vector fields" "A Hamiltonian structure of generalized affine scaling vector fields" "An interior-point method for semidefinite programming" A Short Introduction to Perturbation Theory for Linear Op- erators "Local convergence of predictor-corrector Infeasible-Interior-Point Algorithms for SDPs and SDLCPs" "Interior-point methods for monotone semidefinite linear complementarity problem in symmetric matri- ces" "Superlinear convergence of a symmetric primal-dual path following algorithms for semidefinite program- ming" "Pathways to the optimal set in linear programming" "Higher order methods and their performance" "Primal-dual path following algorithms for semidefinite programming" "A polynomial time primal-dual affine scaling algorithm for linear and convex quadratic programming and its power series extension" Interior point Methods in Convex Programming: Theory and Application "Self-scaled cones and interior-point methods in nonlinear programming" "Primal-dual interior point algorithms for self-scaled cones" "On the facial structure of cone-LP's and semi-definite pro- grams" "Strong duality for semidefinite programming" "Issues in interior point methods in semidefintie and linear programming" "A primal-dual potential reduction method for problems involving matrix inequalities" "Positive-definite programming" "Convergence behavior on central paths for convex homogeneous self-dual cones" "On Extending primal-dual interior-point algorithms from linear programming to semidefinite programming" --TR --CTR Anthony Man-Cho So , Yinyu Ye, Theory of semidefinite programming for sensor network localization, Proceedings of the sixteenth annual ACM-SIAM symposium on Discrete algorithms, January 23-25, 2005, Vancouver, British Columbia
interior point methods;central path;semidefinite programming
589254
Optimality Conditions for Optimization Problems with Complementarity Constraints.
Optimization problems with complementarity constraints are closely related to optimization problems with variational inequality constraints and bilevel programming problems. In this paper, under mild constraint qualifications, we derive some necessary and sufficient optimality conditions involving the proximal coderivatives. As an illustration of applications, the result is applied to the bilevel programming problems where the lower level is a parametric linear quadratic problem.
Introduction . The main purpose of this paper is to derive necessary and su#cient optimality conditions for the optimization problem with complementarity constraints (OPCC) defined as follows: y, u) s.t. #u, #(x, y, y, u) # 0 y, y, u) # 0, (x, y, u) and# is a nonempty subset of R n+m+q . (OPCC) is an optimization problem with equality and inequality constraints. However, due to the complementarity constraint (1.1), the Karush-Kuhn-Tucker (KKT) necessary optimality condition is rarely satisfied by (OPCC) since it can be shown as in [9, Proposition 1.1] that there always exists a nontrivial abnormal multi- plier. This is equivalent to saying that the usual constraint qualification conditions, such as the Mangasarian-Fromovitz condition, will never be satisfied (see [8, Proposition 3.1]). The purpose of this paper is to derive necessary and su#cient optimality conditions under mild constraint qualifications that are satisfied by a large class of OPCCs. To motivate our main results, we formulate problem (OPCC), as the following optimization problem with a generalized equation constraint: (GP) min f(x, y, u) y, u) +N(u,R q y, y, u) # 0, where N(u, C) := # the normal cone of C at y if # Received by the editors May 26, 1997; accepted for publication (in revised form) May 4, 1998; published electronically March 17, 1999. This work was supported by the Natural Sciences and Engineering Research Council of Canada and a University of Victoria internal research grant. http://www.siam.org/journals/siopt/9-2/32188.html Department of Mathematics and Statistics, University of Victoria, Victoria, BC V8W 3P4, Canada (janeye@uvic.ca). OPTIMIZATION PROBLEMS WITH COMPLEMENTARITY CONSTRAINTS 375 is the normal cone operator in the sense of convex analysis. y, - u) be a solution of (OPCC), single-valued and smooth, then the generalized equation constraint (1.2) would reduce to an ordinary equation. Using the KKT condition, we could deduce that if a constraint qualification is satisfied for (GP) and the problem data are smooth, then there exist KKT multipliers # R l , # R d , # R q such that y, - u) +#L(-x, - y, - y, - y, - (-u) #, y, - u)# 0, where # denotes the usual gradient, M # denotes the transpose of the matrix M , and NC denotes the map y # N(y, C). However, u # N(u, R q in general a set-valued map. Naturally, we hope to replace #N R q (-u) # by the image of some derivatives of the set-valued map u # N(u, R q acting on the vector #. The natural candidate for such a derivative of set-valued maps is the Mordukhovich coderivative (see Definition 2.3) since the Mordukhovich coderivatives have a good calculus, and in the case when the set-valued map is single-valued and smooth, the image of the Mordukhovich coderivative acting on a vector coincides with the usual gradient operator acting on the vector (see [6, Proposition 2.4]). Indeed, as in [7], we can show that if (-x, - y, - u) is an optimal solution of (OPCC) and a constraint qualification holds, then there exist q such that y, - y, - y, - y, - y, - u))(#), y, - u)# 0, where D # denotes the Mordukhovich coderivative (see Definition 2.3). Recall from [7, Definition 2.8] that a set-valued R q with a closed graph is said to be pseudo-upper-Lipschitz continuous at (-z, - v) with - v #(-z) if there exist a neighborhood U of - z, a neighborhood V of - v, and a constant - > 0 such that The constraint qualification for the above necessary condition involving the Mor- dukhovich coderivative turns out to be the pseudo-upper-Lipschitz continuity of the set-valued map y, u)+N(u,R q y, u)+v 3 # 0} at (-x, - y, - u, 0). This constraint qualification is very mild since the pseudo-upper- Lipschitz continuity is weaker than both the upper-Lipschitz continuity and the pseudo- Lipschitz continuity (the so-called Aubin property). However, the Mordukhovich normal cone involved in the necessary condition may be too large sometimes. For ex- ample, in [7, Example 4.1], both (0, 0) and (1, 1) satisfy the above necessary conditions, but only (1, 1) is the unique optimal solution. Can one replace the Mordukhovich normal cone involved in the necessary condition by the potentially smaller proximal normal cone? The answer is negative in general, since the proximal coderivative as defined in Definition 2.3 usually has only a "fuzzy" calculus. Consider the following 376 J. J. YE optimization problem: min -y s.t. y - The unique optimal solution (0, does not satisfy the KKT condition but satisfies the necessary condition involving the Mordukhovich coderivatives. It does not satisfy the necessary condition with the Mordukhovich normal cone replaced by the proximal normal cone. This example shows that some extra assumptions are needed for the necessary condition involving the proximal coderivatives to hold. In this paper such a condition is found. Moreover, we show that the proximal normal cone involved in the necessary condition can be represented by a system of linear and nonlinear equations, and the necessary optimality conditions involving the proximal coderivatives turn out to be su#cient under some convexity assumptions on the problem data. Although the optimization problems with complementarity constraints are a class of optimization problems with independent interest, the incentive to study (OPCC) mainly comes from the following optimization problem with variational inequality constraints (OPVIC), where the constraint region of the variational inequality is a system of inequalities: R,# is a nonempty subset of R m+n and S(x) is the solution set of a variational inequality with parameter x; i.e., . The recent monograph [4] by Luo, Pang, and Ralph has an extensive study for (OPVIC). The reader may find the references for the various optimality conditions for (OPVIC) from [4]. (OPCC) is closely related to OPVICs and bilevel programming problems. Indeed, if is C 1 and quasi convex in y and a certain constraint qualification condition holds at - y for the optimization problem min #F (-x, - then by the KKT necessary and su#cient optimality condition, (-x, - y) is a solution of (OPVIC) if and only if there exists - u # R q such that (-x, - y, - u) is a solution of the following optimization problem: (KS) min f(x, y) s.t. #u, (x, which is a special case of (OPCC). In the case where F (x, # R is di#erentiable and pseudoconvex in y, (KS) is equivalent to the following bilevel programming problem (BLPP), or so-called Stackelberg game: (BLPP) min f(x, y) OPTIMIZATION PROBLEMS WITH COMPLEMENTARITY CONSTRAINTS 377 where S(x) is the set of solutions of the problem (P x We organize the paper as follows. Section 2 contains background material on nonsmooth analysis and preliminary results. In section 3 we derive the necessary and su#cient optimality conditions for (OPCC). As an illustration of applications, we also apply the result to (BLPP), where the lower level is a linear quadratic programming problem. 2. Preliminaries. This section contains some background material on non-smooth analysis and preliminary results which will be used later. We give only concise definitions that will be needed in the paper. For more detailed information on the subject, our references are Clarke [1, 2], Loewen [3], and Mordukhovich [6]. First we give some concepts for various normal cones and subgradients. Definition 2.1. Let# be a nonempty subset of R n . Given - z # cl# , the closure of set# , the convex cone #z is called the proximal normal cone to set# at point - z, and the closed cone N(-z, := { lim is called the limiting normal cone to# at point - z. Remark 2.1. It is known that if# is convex, then the proximal normal cone and the limiting normal cones coincide with the normal cone in the sense of convex analysis. Definition 2.2. Let f : R n R #} be lower semicontinuous and finite at z # R n . The limiting subgradient of f at - z is defined to be the set N(-z, where denotes the epigragh of f . Remark 2.2. It is known that if f is a convex function, the limiting subgradient coincides with the subgradient in the sense of convex analysis. For a locally Lipschitz function f , #f(x), where # denotes the Clarke generalized gradient and co denotes the convex hull. Hence the limiting subgradient is in general a smaller set than the Clarke generalized gradient. For set-valued maps, the definition for limiting normal cone leads to the definition of coderivative of a set-valued map introduced by Mordukhovich (see, e.g., [6]). Definition 2.3. Let # : R n R q be an arbitrary set-valued map (assigning to each z # R n a set #(z) # R q which may be empty) and (-z, - v) # cl Gr#, where Gr# denotes the graph of #; i.e., (z, v) # Gr# if and only if v #(z). The set-valued maps from R q into R n defined by v), Gr#)} are called the proximal and Mordukhovich coderivatives of # at point (-z, - v), respectively 378 J. J. YE Proposition 2.4. Suppose B is closed, - # B. Then Proof. Since - is closed, there exists a neighborhood of - x that is not contained in B. Therefore, from the definition of the proximal normal cone, we have In the following proposition we show that the proximal normal cone of a union of a finite number of sets is the intersection of the proximal cones. Proposition 2.5. are closed. Then Proof. Let # N # (-x, . Then, by definition, there exists a constant M > 0 such that #, x - x# M #x - #x the above inequality implies that # m Conversely, suppose # m there exists #, x - #x That is, there exists #, x - x# M #x - #x which implies that # N # The above decomposition formula for calculating the proximal normal cones turns out to be very useful, since when a set can be written as a union of some convex sets, the task of calculating the proximal normal cones is reduced to calculating the normal cone to convex sets which are easier to calculate. The following proposition is a nice application of the decomposition formula and will be used to calculate the proximal normal cone to the graph of the set-valued map N R q for general q in Proposition 2.7. Proposition 2.6. Proof. It is easy to see that GrNR+ We discuss the following three cases. Case 1. - In this case, (-x, - y) closed, by Proposition 2.4 we have in this case OPTIMIZATION PROBLEMS WITH COMPLEMENTARITY CONSTRAINTS 379 Case 2. - y < 0. In this case, (-x, - y) closed, by Proposition 2.4 we have in this case Case 3. - In this case, (-x, - . By Proposition 2.5 we have (R - [0, #)) Now we are in a position to give an expression for the proximal normal cone to the graph of the set-valued map N R q for general q. Proposition 2.7. For any (-x, - y) #GrN R q , define I I 0 := I 0 (-x, - Then y), GrN R q Proof. Since we have if and only if GrNR+ . Hence from the definition, it is clear that if and only if The rest of the proof follows from Proposition 2.6. It turns out that we can express any element of N # ((-x, - y),GrN R q by a system of nonlinear equations as in the following proposition. J. J. YE Proposition 2.8. if and only if there exist # R 2q such that Proof. By Proposition 2.7, (# N # ((-x, - y), GrN R q if and only if By the definition for the index sets I 0 , I + , L in Proposition 2.7, we have Since for any (-x, - y) # GrN R q y # 0, for nonnegative vectors # and #, (2.1) is equivalent to Hence the existence of nonnegative vectors # and # satisfying (2.1)-(2.2) is equivalent to the following condition: Consequently, it is equivalent to The proof of the proposition is therefore complete. Finally, we would like to recall the following definition of a very mild constraint qualification called "calmness," introduced by Clarke [1]. Definition 2.9. Let - x be a local solution to the following mathematical programming problem: minimize f(x) OPTIMIZATION PROBLEMS WITH COMPLEMENTARITY CONSTRAINTS 381 and C is a closed subset of R d . The above mathematical programming problem is said to be calm at - x provided that there exist positive # and M such that for all (p, q) #B, for all x in - x+#B satisfying g(x)+p # where B is the open unit ball in the appropriate space. It is well known that the calmness condition is a constraint qualification for the existence of a KKT multiplier and the su#cient conditions for the calmness condition include the linear independence condition, the Slater condition, and the Mangasarian- Fromowitz condition. Moreover, the calmness condition is satisfied automatically in the case where the feasible region is a polyhedron. 3. Optimality conditions for OPCC. Let (-x, - y, - # and g(-x, - y, - y, - I y, y, - u) < 0}, I 0 (-x, - y, y, Where there is no confusion, we simply use L, I + , I 0 instead of L(-u), I y, - u), I 0 (-x, - y, - u), respectively. It is clear that {1, 2, . , y, - y, - u). Let y, u) y, y, u) # 0 #u, #(x, y, y, u) # 0 be the feasible region of (OPCC). For any I # {1, 2, . , q}, let F I := y, u) y, y, u) # 0 y, y, u) # 0 #i # {1, 2, . , q}\I denote a piece of the feasible region F . Taking the "piecewise programming" approach in the terminology of [4], as in Corollary 2 of [5], we observe that the feasible region of the problem (OPCC) can be rewritten as a union of all pieces I . Therefore, a local solution y, - u) for (OPCC) is also a local solution for each subproblem of minimizing the objective function f over a piece which contains the point (-x, - y, - u). Moreover, if y, - u) is contained in all pieces and all subproblems are convex, then it is a global minimum for the original problem (OPCC). Hence the following proposition follows from this observation. Proposition 3.1. Let (-x, - y, - u) be a local optimal solution to (OPCC). Suppose that f , g, , L are locally Lipschitz near (-x, - y, - and# is closed. If for any given index set # I 0 , the problem of minimizing f over F#L is calm in the sense of Definition 2.9 at (-x, - y, - u), then there exist # R l , # R d , # R q , # R q such that y, - l y, d y, y, - y, - (3. J. J. YE y, - Conversely, let (-x, - y, - u) be a feasible solution for (OPCC), and for all index sets I 0 , there exist # R l , # R d , # R q , # R q such that (3.1)-(3.3) are satisfied. If f is either convex or pseudoconvex, g is convex, , L are a#ne, and# is convex, then (-x, - y, - u) is a minimum of f over all (x, y, u) #I0 F#L. If in addition to the above assumptions I y, - u) is a global solution for (OPCC). Proof. It is obvious that the feasible region of (OPCC) can be represented as the union of pieces I . Since - y, - u) < 0 y, - u), and y, u) y, y, u) # 0 y, y, y, u) # 0 #i # I 0 \# we have y, - u) #I0 F#L and y, - Hence if (-x, - y, - u) is optimal for (OPCC), then for any given index set # I 0 , (-x, - y, - is also a minimum for f over F#L . Since this problem is calm, by the well-known nonsmooth necessary optimality condition (see, e.g., [1, 2, 3]), there exist # R l , q such that (3.1)-(3.3) are satisfied. Conversely, suppose that for each # I 0 there exist # R l , # R d , # R q , # R q such that (3.1)- are satisfied and the problem is convex. By virtue of Remarks 2.1 and 2.2, the limiting subgradients and the limiting normal cones coincide with the subgradients and the normal cone in the sense of convex analysis, respectively. Hence, by the standard first-order su#cient optimality conditions, (-x, - y, - u) is a minimum of f over F#L for each # I 0 and hence is a minimum of f over #I0 F#L . In the case when I and the feasible region y, - is a global optimal for (OPCC) in this case. The proof of the proposition is now complete. Remark 3.1. The necessary part of the above proposition with smooth problem data is given by Luo, Pang, and Ralph in [4] under the so-called "basic constraint qualification." Note that the multipliers in Proposition 3.1 depend on the index set # through (3.3). However, if for some pair of index sets # I 0 ) and I 0 \#, the components of the multipliers are the same, then we would have a necessary condition that does not depend on the index set #. In this case the necessary condition turns out to be the necessary condition involving the proximal coderivatives as in (b) of the following theorem. Theorem 3.2. Suppose f, g, L, are continuously di#erentiable. Then the following three conditions are equivalent: OPTIMIZATION PROBLEMS WITH COMPLEMENTARITY CONSTRAINTS 383 (a) There exist # R l , # R d , # R q such that y, l y, - d y, - u) y, y, - (b) There exist # R l , # R d , # R q such that y, - l y, d y, - y, - y, - u))(#), y, - (c) There exist # R l , # R d , # R q , # R 2q such that (3.4) and (3.5) are satisfied and y, - y, - Let (-x, - y, - u) be a local optimal solution to (OPCC), that there exists an index set # I 0 such that the problem of minimizing f over F#L and the problem of minimizing f over F (I 0 \#L are calm. Furthermore, suppose that l y, - d y, - y, y, - implies that # the three equivalent conditions (a)-(c) hold. Conversely, let (-x, - y, - u) be a feasible solution to (OPCC), let f be pseudoconvex, g be convex, #, L be a#ne. If one of the equivalent conditions (a)-(c) holds, then (-x, - y, - u) is a minimum of f over all (x, y, u) #I0 F#L . If in addition to the above assumptions I y, - u) is a global solution for (OPCC). Proof. By the definition of the proximal coderivatives (Definition 2.3), y, - u))(#) if and only if 384 J. J. YE Hence the equivalence of condition (a) and condition (b) follows from Proposition 2.7. The equivalence of condition (b) and condition (c) follows from Proposition 2.8. Let (-x, - y, - u) be a local optimal solution to (OPCC), it is also a local optimal solution to the problem of minimizing f over F#L and the problem of minimizing f over F (I 0 \#L . By the calmness assumption for these two problems, there exist # i (3.1)-(3.3), which implies that l y, - d y, - u) y, - y, - By the assumption we arrive at # 1 I0 . Since by (3.3), # 1 I0 \# 0, we have That is, condition (a) holds. The su#cient part of the theorem follows from the su#cient part of Proposition 3.1. As observed in [4, Proposition 4.3.5], the necessary optimality conditions (3.4)- (3.6) happen to be the KKT condition for the relaxed problem (RP) minf(x, y, u) s.t. y, y, - u), y, - u), y, y, u) # 0, and (#) satisfies (3.4)-(3.6) if and only if it satisfies the KKT condition for the subproblem of minimizing f over the feasible region F#L , i.e., (3.1)-(3.3) with the smooth problem data y, - u). Conse- quently, if the strict Mangasarian-Fromovitz constraint qualification (SMFCQ) holds for problem (RP) at (#) which satisfies (3.4)-(3.6), then (#) is the unique multiplier which satisfies (3.4)-(3.6). Since the index sets # only a#ect the (# I0 , # I0 ) components of the multiplier (#), we observe that the existence of multipliers satisfying (3.4)-(3.6) is equivalent to the existence of multipliers satisfying (3.1)-(3.3) for all index sets # I 0 (-x, - y, - u) with the components (# I0 , # I0 ) having the same sign. From the proof of Theorem 3.2, it is easy to see that the condition that no nonzero vectors satisfy (3.9)-(3.10) is a su#cient condition for the existence of common (# I0 , # I0 ) components of the multiplier (#) for all index sets # I 0 (-x, - y, - u). Hence this condition refines the su#cient condition of a unique multiplier such as the SMFCQ for the relaxed problem proposed in [4, Proposition 4.3.5]. We now give an example which does not have a unique multiplier satisfying (3.4)- but does satisfy the condition proposed in Theorem 3.2. OPTIMIZATION PROBLEMS WITH COMPLEMENTARITY CONSTRAINTS 385 Example 3.1 (see [4, Example 4.3.6]). Consider the following OPCC: s.t. x 2 are any real numbers, are obviously solutions to the above problem. As pointed out in [4, Example 4.3.6], SMFCQ does not hold for this problem. However, we can verify that it satisfies our condition. Indeed, the equation (3.9) for this problem is which implies that Moreover, the calmness condition is satisfied since the constraint region for each subproblem F#L is a polyhedron due to the fact that # and g are both a#ne. Hence by Theorem 3.2, if (-x, - u) is a local minimum to the above problem, then there exist # such that which implies # is a global optimal solution according to Theorem 3.2 and (-x, 0, 0) with are local optimal solutions. To illustrate the application of the result obtained, we now consider the following bilevel programming problem (BLQP), where the lower level problem is linear quadratic: (BLQP) min f(x, y) s.t. y # S(x), where G and H are l - n and l - m matrices, respectively, a # R l , and S(x) is the solution set of the quadratic programming problem with parameter x: where Q # R m-m is a symmetric and positive semidefinite matrix, p # R n , q are q - n and q -m matrices, respectively, and b # R q . Replacing the bilevel constraint by the KKT condition for the lower level problem, it is easy to see that (BLQP) is equivalent to the problem (KKT) min f(x, y) 386 J. J. YE which is an OPCC. Let (-x, - y) be an optimal solution of (BLQP) and - u a corresponding u, Then I The feasible region of problem (KKT) is and for any I # {1, 2, . , q}, y, u) # R Since F#L for any index set # I 0 has linear constraints only, the problem of minimizing f over F#L is calm. Hence the following result follows from Proposition 3.1. Corollary 3.3. Let (-x, - y) be an optimal solution of (BLQP) and - u a corresponding multiplier. Suppose that f is locally Lipschitz near (-x, - y). Then for each I 0 , there exist # R m , # R d , # R q such that If f is either convex or pseudoconvex, then the above necessary condition is also su#cient for a feasible solution (-x, - y, - u) of (KKT) to be a minimum of f over all y, u) #I0 F#L . In particular, if f is either convex or pseudoconvex and I {1, 2, . , q}, then the above condition is su#cient for a feasible solution (-x, - y) to be a global optimum for (BLQP). The following result follows from Theorem 3.2. Corollary 3.4. Let (-x, - y) be an optimal solution of (BLQP) and - u a corresponding multiplier. Suppose that f is C 1 and implies # there exist # R m , # R d , # R q such that OPTIMIZATION PROBLEMS WITH COMPLEMENTARITY CONSTRAINTS 387 Equivalently, there exist # R l , # R d , # R q such that (3.16)-(3.17) are satisfied and Equivalently, there exist # R l , # R d , # R q , # R 2q such that (3.16)-(3.17) are satisfied and Conversely, let (-x, - y) be any vector in R n+m satisfying the constraints G-x+H - be pseudoconvex. If there exists - (3.11)-(3.12) such that one of the above equivalent conditions holds, then (-x, - y, - u) is a minimum of f over all (x, y, u) #I0 F#L . In addition to the above assumptions, if I y) is a global minimum for (BLQP). Acknowledgments . The author would like to thank Dr. Qing Lin for a helpful discussion of Proposition 2.8. --R Optimization and Nonsmooth Analysis Optimal Control via Nonsmooth Analysis Mathematical Programs with Equilibrium Constraints Piecewise Sequential Quadratic Programming for Mathematical Programs with Nonlinear Complementarity Constraints Generalized di Necessary optimality conditions for optimization problems with variational inequality constraints Optimality conditions for bilevel programming problems Exact penalization and necessary optimality conditions for generalized bilevel programming problems --TR --CTR Jin-Bao Jian, A Superlinearly Convergent Implicit Smooth SQP Algorithm for Mathematical Programs with Nonlinear Complementarity Constraints, Computational Optimization and Applications, v.31 n.3, p.335-361, July 2005 Houyuan Jiang , Daniel Ralph, Extension of Quasi-Newton Methods to Mathematical Programs with Complementarity Constraints, Computational Optimization and Applications, v.25 n.1-3, p.123-150
optimality conditions;optimization problems;bilevel programming problems;complementarity constraints;proximal normal cones
589256
Solving the Trust-Region Subproblem using the Lanczos Method.
The approximate minimization of a quadratic function within an ellipsoidal trust region is an important subproblem for many nonlinear programming methods. When the number of variables is large, the most widely used strategy is to trace the path of conjugate gradient iterates either to convergence or until it reaches the trust-region boundary. In this paper, we investigate ways of continuing the process once the boundary has been encountered. The key is to observe that the trust-region problem within the currently generated Krylov subspace has a very special structure which enables it to be solved very efficiently. We compare the new strategy with existing methods. The resulting software package is available as HSL_VF05 within the Harwell Subroutine Library.
Introduction Trust-region methods for unconstrained minimization are blessed with both strong theoretical convergence properties and a good reputation in practice. The main computational step in these methods is to find an approximate minimizer of some model of the true objective function within a "trust" region for which a suitable norm of the correction lies within a given bound. This restriction is known as the trust-region constraint, and the bound on the norm is its radius. The radius is adjusted so that successive model problems mimic the true objective within the trust region. The most widely-used models are quadratic approximations to the objective function, as these are simple to manipulate and may lead to rapid convergence of the underlying method. From a theoretical point of view, the norm which defines the trust region is irrelevant so long as it "uniformly" related to the ' 2 norm. From a practical perspective, this choice certainly affects the subproblem, and thus the methods one can consider when solving it. The most popular practical choices are the ' 2 - and ' 1 -norms, and weighted variants thereof. In our opinion, it is important that the choice of norm reflects the underlying geometry of the problem; simply picking the may not be adequate when the problem is large, and the eigenvalues of the Hessian of the model widely spread. We believe that weighting the norm is essential for many large-scale problems. In this paper, we consider the solution of the quadratic-model trust-region subproblem in a weighted ' 2 -norm. We are interested in solving large problems, and thus cannot rely solely on factorizations of the matrices involved. We thus concentrate on iterative methods. If the model of the Hessian is known to be positive definite and the trust-region radius sufficiently large that the trust-region constraint is inactive at the unconstrained minimizer of the model, the obvious way to solve the problem is to use the preconditioned conjugate-gradient method. Note that the role of the preconditioner here is the same as the role of the norm used for the trust-region, namely to change the underlying geometry so that the Hessian in the rescaled space is better conditioned. Thus, it will come as no surprise that the two should be intimately connected. Formally, we shall require that the weighting in the ' 2 -norm and the preconditioning are performed by the same matrix. When the radius is smaller than a critical value, the unconstrained minimizer of the model will no longer lie within the trust-region, and thus the required solution will lie on the trust-region boundary. The simplest strategy in this case is to consider the piecewise linear path connecting the conjugate-gradient iterates, and to stop at the point where this path leaves the trust region. Such a strategy was first proposed independently by Steihaug (1983) and Toint (1981), and we shall refer to the terminating point as the Steihaug-Toint point. Remarkably, it is easy to establish the global convergence of a trust-region method based on such a simple strategy. The key is that global convergence may be proved provided that the accepted estimate of the solution has a model value no larger than at the Cauchy point (see Powell, 1975). The Cauchy point is simply the minimizer of the model within the trust-region along the preconditioned steepest-descent direction. As the first segment on the piecewise-linear conjugate-gradient path gives N. I. M. Gould, S. Lucidi, M. Roma and Ph. L. Toint precisely this point, and as the model value is monotonically decreasing along the entire path, the Steihaug-Toint strategy ensures convergence. If the model Hessian is indefinite, the solution must also lie on the trust-region boundary. This case may also be simply handled using preconditioned conjugate gradients. Once again the piecewise linear path is followed until either it leaves the trust-region, or a segment with negative curvature is found (a vector p is a direction of negative curvature if the inner product is the model Hessian). In the latter case, the path is continued downhill along this direction of negative curvature as far as the constraint boundary. This variant was proposed by Steihaug (1983), while Toint (1981) suggests simply returning to the Cauchy point. As before, global convergence is ensured as either of these terminating points, as the objective function values there are no larger than at the Cauchy point. For consistency with the previous paragraph, we shall continue to refer to the terminating point in Steihaug's algorithm as the Steihaug-Toint point, although strictly Toint's point in this case may be different. The Steihaug-Toint method is basically unconcerned with the trust region until it blunders into its boundary and stops. This is rather unfortunate, particularly as considerable experience has shown that this frequently happens during the first few, and often the first, iteration(s) when negative curvature is present. The resulting step is then barely, if at all, better than the Cauchy direction, and this may lead to a slow but globally convergent algorithm in theory and a barely convergent method in practice. In this paper, we consider an alternative which aims to avoid this drawback by trying harder to solve the subproblem when the boundary is encountered, while maintaining the efficiencies of the conjugate gradient method so long as the iterates lie interior. The mechanism we use is the Lanczos method. The paper is organized as follows. In Section 2 we formally define the problem and any notation that we will use. The basis of our new method is given in Section 3, while in Section 4, we will review basic properties of the preconditioned conjugate-gradient and Lanczos methods. Our new method is given in detail in Section 5. Some numerical experiments demonstrating the effectiveness of the approach are given in Section 6, and a number of conclusions and perspectives are drawn in the final section. Solving the trust-region subproblem using the Lanczos method 3 2 The trust-region subproblem and its solution Let M be a symmetric positive-definite easily-invertible approximation to the symmetric matrix H. Furthermore, define the M-norm of a vector as where h\Delta; \Deltai is the usual Euclidean inner product. In this paper, we consider the M-norm trust-region problem minimize subject to ksk M - \Delta; (2.1) for some vector g and radius \Delta ? 0. A global solution to the problem is characterized by the following result. Theorem 2.1 (Gay, 1981, Sorensen, 1982) Any global minimizer s M of q(s) subject to satisfies the equation positive semi-definite, - M - 0 and - M (ks is positive definite, s M is unique. This result is the basis of a series of related methods for solving the problem which are appropriate when forming factorizations of H(-) j H +-M for a number of different values of - is realistic. For then, either the solution lies interior, and hence - g, or the solution lies on the boundary and - M satisfies the nonlinear equation denotes the pseudo-inverse of H. Equation (2.3) is straightforward to solve using a safeguarded Newton iteration, except in the so-called hard case for which g lies in the null-space of H(- M ). In this case, an additional vector in the range-space of H(- M ) may be required if a solution on the trust-region boundary is sought. Goldfeldt, Quandt and Trotter (1966), Hebden (1973) and Gay (1981) all proposed algorithms of this form. The most sophisticated algorithm to date, by Mor'e and Sorensen (1983), is available as subroutine GQTPAR in the MINPACK-2 package, and guarantees that a nearly optimal solution will be obtained after a finite number of factorizations. While such algorithms are appropriate for large problems with special Hessian structure - such as for band matrices - the demands of a factorization at each iteration limits their applicability for general large problems. It is for this reason that methods which do not require factorizations are of interest. Throughout the paper, we shall denote the k by k identity matrix by I k , and its j-th column by e j . A set of vectors fq i g are said to be M-orthonormal if hq the Kronecker delta, 4 N. I. M. Gould, S. Lucidi, M. Roma and Ph. L. Toint and the matrix Q formed from these vectors is an M-orthonormal matrix. The set of vectors fp i g are H-conjugate (or H-orthogonal) if hp 3 A new algorithm for large-scale trust-region subproblems To set the scene for this paper, we recall that the Cauchy point may be defined as the solution to the problem minimize subject to ksk M - \Delta; (3.1) that is as the minimizer of q within the trust region where s is restricted to the 1-dimensional subspace span \Phi . The dogleg methods (see, Powell, 1970, Dennis and Mei, 1979) aim to solve the same problem over a one-dimensional arc, while Byrd, Schnabel and Schultz (1985) do the same over a two-dimensional subspace. In each of these cases the solution is easy to find as the search space is small. The difficulty with the general problem (2.1) is that the search space R n is large. This leads immediately to the possibility of solving a compromise problem minimize subject to ksk M - \Delta; (3.2) where S is a specially chosen subspace of R n . Now consider the Steihaug-Toint algorithm at an iteration k before the trust-region boundary is encountered. In this case, the point s k+1 is the solution to (3.2) with the set span the Krylov space generated by the starting vector M \Gamma1 g and matrix M \Gamma1 H. That is, the Steihaug-Toint algorithm gradually widens the search space using the very efficient preconditioned conjugate gradient method. However, as soon as the Steihaug-Toint algorithm moves across the trust-region boundary, the terminating point s k+1 no longer necessarily solves the problem in (3.3), indeed it is very unlikely to do so when k ? 0. (As the iterates generated by the method increase in M-norm, once an iterate leaves the trust region, the solution to (3.3), and thus (2.1), must lie on the boundary. See, Steihaug, 1983, Theorem 2.1, for details). Can we do better? Yes, by recalling that the preconditioned conjugate gradient and Lanczos methods generate different bases for the same Krylov space. Solving the trust-region subproblem using the Lanczos method 5 4 The preconditioned conjugate-gradient and Lanczos methods The preconditioned conjugate-gradient and Lanczos methods may be viewed as efficient techniques for constructing different bases for the same Krylov space, K k . The conjugate gradient method aims for an H-conjugate basis, while the Lanczos method obtains an M-orthonormal basis. Algorithm 4.1: The preconditioned conjugate gradient method perform the iteration, Algorithm 4.2: Preconditioned Lanczos method perform the iteration, The conjugate gradient method generates the basis from Algorithm 4.1, while the Lanczos method generates the basis 6 N. I. M. Gould, S. Lucidi, M. Roma and Ph. L. Toint with Algorithm 4.2. The Lanczos iteration is often written in the more compact form k+1 and (4.14) is the matrix (q and the matrix @ A is tridiagonal. It then follows directly that The two methods are intimately related. In particular, so long as the conjugate-gradient iteration does not break down, the Lanczos vectors may be recovered from the conjugate-gradient iterates as while the Lanczos tridiagonal may be expressed as ff ff k\Gamma2 ff A The conjugate gradient iteration may breakdown if hp which can only occur if H is not positive definite, and will stop if hg On the other hand, the Lanczos iteration can only fail if K j is an invariant subspace for M \Gamma1 H. If q(s) is convex in the manifold K j+1 , the minimizer s j+1 of q in this manifold satisfies so long as the initial value s chosen. Thus this estimate is easy to recur from the conjugate-gradient iteration. The minimizers in successive manifolds may also be easily obtained using the Lanczos process, although the conjugate-gradient iteration is slightly less expensive, and thus to be preferred. Solving the trust-region subproblem using the Lanczos method 7 The vector g j+1 in the conjugate gradient method gives the gradient of q(s) at s j+1 . It is quite common to stop the method as soon as this gradient is sufficiently small, and the method naturally records the M \Gamma1 -norm of the gradient, kg This norm is also available in the Lanczos method as solves the tridiagonal linear system T k The last component, he k+1 of h k is available as a further by-product. 5 The truncated Lanczos approach Rather than use the preconditioned conjugate gradient basis fp for S, we shall use the equivalent Lanczos M-orthonormal basis fq g. The Lanczos basis has previously been used by Nash (1984) - to convexify the quadratic model - and Lucidi and Roma (1997) - to compute good directions of negative curvature - within linesearch based method for unconstrained minimization. We shall consider vectors of the form and seek solves the problem minimize s 2S subject to ksk M - \Delta: (5.2) It then follows directly from (4.15), (4.17) and (4.18) that h k solves the problem minimize subject to khk 2 - \Delta: (5.3) There are a number of crucial observations to be made here. Firstly, it is important to note that the resulting trust-region problem involves the two-norm rather than the M-norm. Secondly, as T k is tridiagonal, it is feasible to use the Mor'e-Sorensen algorithm to compute the model minimizer even when n is large. Thirdly, having found h k , the matrix Q k is needed to recover thus the Lanczos vectors will either need to be saved on backing store or regenerated. As we shall see, we only need Q once we are satisfied that continuing the Lanczos process will give little extra benefit. Fourthly, one would hope that as a sequence of such problems may be solved, and as T k only changes by the addition of an extra diagonal and superdiagonal entry, solution data from one subproblem may be useful for starting the next. We consider this issue in Section 5.2. The basic trust-region solution classification theorem, Theorem 2.1, shows that 8 N. I. M. Gould, S. Lucidi, M. Roma and Ph. L. Toint I k+1 is positive semi-definite, What does this tell us about s k ? Firstly, using (4.17), (4.18) and (5.4) we have and additionally that Comparing these with the trust-region classification theorem, we see that s k is the Galerkin approximation to s M from the space spanned by Q k . We may then ask how good the approximation is. In particular, what is the error (H g? The simplest way of measuring this error would be to calculate h k and - k by solving (5.3), then to recover s k as Q k h k and finally to substitute s k and - k into (H +-M)s+ g. However this is inconvenient as it requires that we have easy access to Q k . Fortunately there is a far better way. Theorem 5.1 and Proof. We have that iw k+1 from (4.14) iw k+1 from (5.4) This then directly gives (5.6), and (5.7) follows from the M \Gamma1 -orthonormality of w k+1 . 2 Therefore we can indirectly measure the error (in the M \Gamma1 -norm) knowing simply fl k+1 and the last component of h k , and we do not need s k or Q k at all. Observant readers will notice the strong similarity between this error estimate and the estimate (4.22) for the gradient of the model in the Lanczos method, but this is not at all surprising as the two methods are aiming for the same point if the trust-region radius is large enough. An interpretation of (5.7) is also identical to that of (4.22). The error will be small when either of fl k+1 or the last component of h k is small. We now consider the problem (5.3) in more detail. We say that a symmetric tridiagonal matrix is degenerate if one or more of its off-diagonal entries is zero; otherwise it is non-degenerate. We then have the following preliminary result. Solving the trust-region subproblem using the Lanczos method 9 Lemma 5.2 (See also, Parlett, 1980, Theorem 7.9.5) Suppose that the tridiagonal matrix T is non-degenerate, and that v is an eigenvector of T . Then the first component of v is nonzero. Proof. By definition for some eigenvector '. Suppose that the first component of v is zero. Considering the first component of (5.8), we have that the second component of v is zero as T is tridiagonal and non-degenerate. Repeating this argument for the i-th component of (5.8), we deduce that the the 1-st component of v is zero for all i, and hence that contradicts the assumption that v is an eigenvector, and so the first component of v cannot be zero. 2 This immediately yields the following useful result. Theorem 5.3 Suppose that T k is non-degenerate. Then the hard case cannot occur for the subproblem (5.3). Proof. Suppose the hard case occurs. Then, by definition, fl 0 e 1 is orthogonal to v k , the eigenvector corresponding to the leftmost eigenvalue, \Gamma' k , of T k . Thus, the first component of v k is zero, which, following Lemma 5.2, contradicts the assumption that v k is an eigenvector. Thus the hard case cannot occur. 2 This result is important as it suggests that the full power of the Mor'e and Sorensen (1983) algorithm is not needed to solve (5.3). We shall return to this in Section 5.2. We also have an immediate corollary. Corollary 5.4 Suppose that T n\Gamma1 is non-degenerate. Then the hard case cannot occur for the original problem (2.1). Proof. When the columns of Q forms a basis for R n , Thus the problems (2.1) and (5.2) are identical, and (5.2) and (5.3) are related through a nonsingular transformation. The result then follows directly from Theorem 5.3 in the case Thus, if the hard case occurs for (2:1), the Lanczos tridiagonal must degenerate at some stage. N. I. M. Gould, S. Lucidi, M. Roma and Ph. L. Toint Theorem 5.5 Suppose that T k is non-degenerate, that h k and - k satisfy (5.4) and that I k+1 is positive semi-definite. Then I k+1 is positive definite. Proof. Suppose that T k I k+1 is singular. Then there is a nonzero eigenvector v k for which Hence, combining this with (5.4) reveals that and hence that the first component of v k is zero. But this contradicts Lemma 5.2. Hence I k+1 is both positive semi-definite and nonsingular, and thus positive definite. 2 This result implies that (5.4) has a unique solution. We now consider this solution. Theorem 5.6 Suppose that he k+1 Proof. Suppose that T k is not degenerate. As the k 1-st component of h k is zero, then the non-degeneracy of T k and the k 1-st equation of (5.4), we deduce that the k-th component of h k is zero. Repeating this argument for the 1-st equation of (5.4), we deduce that the i-th component of h k is zero for 1 - i - k, and hence that h contradicts the first equation of (5.4), and thus T k must be degenerate. 2 Thus we see that of the two possibilities suggested by Theorem 5.1 for obtaining an s k for which will be the possibility fl that occurs before he k+1 Theorem 5.7 Suppose that the hard case does not occur for (2.1), and that fl Proof. If the Krylov space K k is an invariant subspace M \Gamma1 H, and by construction the first basis element of this space is M \Gamma1 g. As the hard case does not occur for (2.1), this space must also contain at least one eigenvector corresponding to the leftmost eigenvalue, \Gamma', of M \Gamma1 H. Thus one of the eigenvalues of T k must be \Gamma', and - k - ' as T k +- k I k+1 is positive semi-definite. But this implies that H positive semi-definite, which combines with (5.1), (5.5) and Theorem 5.1 with to show that s k satisfies the optimality conditions shown in Theorem 2.1. 2 Thus we see that in the easy case, the required solution will be obtained from the first non-degenerate block of the Lanczos tridiagonal. It remains for us to consider the hard case. In Solving the trust-region subproblem using the Lanczos method 11 view of Corollary 5.4, this case can only occur when T k is degenerate. Suppose therefore that T k degenerates into ' blocks of the form @ A where each of the T k i defines an invariant subspace for M \Gamma1 H and where the last block T k ' is the first to yield the leftmost eigenvalue, \Gamma', of M \Gamma1 H. Then there are two cases to consider. Theorem 5.8 Suppose that the hard case occurs for (2.1), that T k is as described by (5.9), and the last block T k ' is the first to yield the leftmost eigenvalue, \Gamma', of M \Gamma1 H. Then, 1. if ' - k1 , a solution to (2.1) is given by s k1 solves the positive-definite system 2. if ' ? - k1 , a solution to (2.1) is given by s @ A h is the solution of the nonsingular tridiagonal system u is an eigenvector of T k ' corresponding to \Gamma', and ff is chosen so that Proof. In case 1, H+- positive semi-definite as - k 1 - ', and the remaining optimality conditions are satisfied as k1 +1 is positive definite follows from Theorem 5.5. In case 2, H+'M is positive semi-definite. Furthermore, as is easy to show that khk 2 ! kh k1 k 2 - \Delta, and hence that there is a root ff for which ks \Delta. Finally, as each Q k i defines an invariant subspace, HQ k i Writing u, we therefore have N. I. M. Gould, S. Lucidi, M. Roma and Ph. L. Toint and all of the optimality conditions for (5.2). 2 Notice that to obtain s k as described in this theorem, we only require the Lanczos vectors corresponding to blocks one and, perhaps, ' of T k . We do not claim that to solve the problem as outlined in Theorem 5.8 is realistic, as it relies on our being sure that we have located the left-most eigenvalue of M \Gamma1 H. With Lanczos-type methods, one cannot normally guarantee that all eigenvalues, including the leftmost, will be found unless one ensures that all invariant subspaces have been investigated, and this may prove to be very expensive for large problems. In particular, the Lanczos algorithm, Algorithm 4.2, terminates each time an invariant subspace has been determined, and must be restarted using a vector q which is M-orthonormal to the previous Lanczos directions. Such a vector may be obtained from the Gram-Schmidt process by re-orthonormalizing a suitable vector - a vector with some component M-orthogonal to the existing invariant subspaces, perhaps a random vector - with respect to the previous Lanczos directions, which means that these directions will have to be regenerated or reread from backing store. Thus, while this form of the solution is of theoretical interest, it is unlikely to be of practical interest if a cheap approximation to the solution is all that is required. 5.1 The algorithm We may now outline our algorithm, Algorithm 5.1, the generalized Lanczos trust region (GLTR) method. We stress that, as our goal is merely to improve upon the value delivered by the Steihaug-Toint method, we do not use the full power of Theorem 5.8, and are content just to investigate the first invariant subspace produced by the Lanczos algorithm. In almost all cases, this subspace contains the global solution to the problem, and the complications and costs required to implement a method based on Theorem 5.8 are, we believe, prohibitive in our context. Solving the trust-region subproblem using the Lanczos method 13 Algorithm 5.1: The generalized Lanczos trust region method . Set the flag INTERIOR as true. For convergence, perform the iteration, using (4.20) If INTERIOR is true, but ff k - 0 or ks k reset INTERIOR to false. If INTERIOR is true else solve the tridiagonal trust-region subproblem (5.3) to obtain h k If INTERIOR is true test for convergence using the residual kg else test for convergence using the value fl k+1 jhe If INTERIOR is false, recover s by rerunning the recurrences or obtaining Q k from backing store. When recovering s by rerunning the recurrences, economies can be made by saving the during the first pass, and reusing them during the second. A potentially bigger saving may be made if one is prepared to accept a slightly inferior value of the objective function. The idea is simply to save the value of q at each iteration. On convergence, one looks back through this list to find an iteration, ' say, for which a required percentage of the best value was obtained, recompute h ' and then accept s as the required estimate of the solution. If the required percentage occurs at an iteration before the boundary is encountered, both the final point before the boundary and the Steihaug-Toint point are suitable and available without the need for the second pass. We note that we have used the conjugate-gradient method (Algorithm 4.1) to generate the Lanczos vectors. If the inner-product hp k ; Hp k i proves to be tiny, it is easy to continue using the Lanczos method (Algorithm 4.2) itself; the vectors 14 N. I. M. Gould, S. Lucidi, M. Roma and Ph. L. Toint required to continue the Lanczos recurrence (4.11) are directly calculable from conjugate-gradient method. At each stage of both the Steihaug-Toint algorithm and our GLTR method (Algorithm 5.1), we need to calculate ks k . This issue is not discussed by Steihaug as it is implicitly assumed that M is available. However, it may be the case that all that is actually available is a procedure which returns M \Gamma1 v for a given input v, and thus M is unavailable. Fortunately this is not a significant drawback as it is possible to calculate ks k +ffp k k M from available information. To see this, observe that ks ks and thus that we can find ks k+1 k 2 M from ks k k 2 M so long as we already know hs k ; Mp k i and M . But it is straightforward to show that these quantities may be calculated from the pair of recurrences and (5.12) where, of course, hg k ; v k i has already been calculated as part of the preconditioned conjugate-gradient method. 5.2 Solving the nondegenerate tridiagonal trust-region subproblem In view of Theorem 5.3, the nondegenerate tridiagonal trust-region subproblem (5.3) is, in theory, easier to solve than the general problem. This is so both because the Hessian is tridiagonal (and thus very inexpensive to factorize), and because the hard case cannot occur. We should be cautious here, because the so-called "almost" hard case - which occurs when g only has a tiny component in the range-space of H(- M ) - may still happen, and the trust-region problem in this case is naturally ill conditioned and thus likely to be difficult to solve. The Mor'e and Sorensen (1983) algorithm is based on being able to form factorizations of the model Hessian (which is certainly the case here as is tridiagonal), but does not try to calculate the leftmost eigenvalue of the pencil H + -M . In the tridiagonal case, computing the extreme eigenvalues is straightforward, particularly if a sequence of related problems are to be solved. Thus, rather than using the Mor'e and Sorensen algorithm, we prefer the following method. We restrict ourselves to the case where the solution lies on the trust-region boundary - we will only switch to this approach when the conjugate gradient iteration leaves the trust region. The basic iteration is identical to that proposed by Mor'e and Sorensen (1983), namely to apply Newton's method to where Solving the trust-region subproblem using the Lanczos method 15 Recalling that we denote the leftmost eigenvalue of T k by \Gamma' k , the main difference between our approach and Mor'e and Sorensen's is that we always start from some point in the interval this interval is characterized by both being positive definite and as then the resulting Newton iteration is globally linearly, and asymptotically quadratically, convergent without any further safeguards. The Newton iteration is performed using Algorithm 5.2. Algorithm 5.2: Newton's method to solve 1. Factorize are unit bidiagonal and diagonal matrices, respectively. 2. Solve BDB T 3. Solve 4. Replace - by - The Newton correction in Step 4 of this algorithm is given by while the exact form given is obtained by using the identity where w is as computed in Step 3. It is slightly more efficient to pick B to be unit upper- bidiagonal rather than unit lower-bidiagonal, as then the Step 2 simplifies to B T because of the structure of the right-hand side. To obtain a suitable starting value, two possibilities are considered. Firstly, we attempt to use the solution value - k\Gamma1 from the previous subproblem. Recall that T k is merely T with an appended row and column. As we already have a factorization of T trivial to obtain that of T and thus to determine if the latter is positive definite. If turns out to be positive definite, h k (- k\Gamma1 ) is computed from (5.15) and if is used to start the Newton iteration. Secondly, if - k\Gamma1 is unsuitable, we monitor T k to see if it is indefinite. This is trivial, as for instance, the matrix is positive definite so long as all of the ff i , generated by the conjugate-gradient method are positive. If T k is positive definite, we start the Newton iteration with the value which by assumption gives kh k (0)k 2 - \Delta as the unconstrained solution lies outside the trust region. Otherwise, we determine the leftmost eigenvalue, \Gamma' k , of T k , and start N. I. M. Gould, S. Lucidi, M. Roma and Ph. L. Toint with positive number chosen to make T k numerically "just" positive definite. By this we mean, that its BDB T factorization should exist, but that ffl should be as small as possible. We have found that a value (1 is the unit roundoff, is almost always suitable, but have added the precaution of multiplying this value by increasing powers of 2 so long as the factorization fails. If we need to compute the leftmost eigenvalue of T k , we use an iteration based upon the last-pivot function proposed by Parlett and Reid (1981). The last-pivot function, ffi k ('), is simply the value of the last diagonal entry of the BDB T factor D k (-) of T k \Gamma 'I k+1 . This value will be zero, and the other diagonal entries positive, when of uncertainty [' l ; ' u ] is placed around the required root. The initial interval is given by the Gersgorin bounds on the leftmost eigenvalue. When it is known, the leftmost eigenvalue, \Gamma' of T k\Gamma1 may be used to improve the lower bound, because of the Cauchy interlacing property of the eigenvalues of T k\Gamma1 and T k (see, for instance, Parlett, 1980, Theorem 10.1.2). Given an initial estimate of ' k , an improvement may be sought by applying Newton's method to ffi k ('); the derivative of ffi k is easy to obtain by recurrence. However, as Parlett and Reid point out, and thus has a pole at Hence it is better to choose the new point by fitting the model to the function and derivative value at the current ', and then to pick the new iterate as the larger root of ffi M ('). If the new iterate lies outside the interval of uncertainty, it is replaced by the midpoint of the interval. The interval is then contracted by computing ffi k at the new iterate, and replacing the appropriate endpoint by the iterate. The iteration is stopped if the length of the interval or the value of If ' k\Gamma1 is known, the initial iterate chosen as ' positive ffl - successive iterates generated from (5.16), the iterates convergence globally, and asymptotically superlinearly, from the left. If the Newton iteration is used, the required root is frequently obscured, and the scheme resorts to interval bisection. Thus the Parlett and Reid scheme is to be preferred. Other means of locating the required eigenvalue, based on using the determinant det(T were tried, but proved to be less reliable because of the huge numerical range (and thus potential overflow) of the determinant. Solving the trust-region subproblem using the Lanczos method 17 6 Numerical experiments The algorithm sketched in Sections 5.1 and 5.2 has been implemented as a Fortran 90 module, HSL VF05, within the Harwell Subroutine Library (1998). As our main interest is in using the methods described in this paper within a trust-region algorithm, we are particularly concerned with two issues. Firstly, can we obtain significantly better values of the model by finding better approximations to its solution than the Steihaug- Toint method? And secondly, do better approximations to the minimizer of the model necessarily translate into fewer iterations of the trust-region method? In this section, we address these outstanding questions. Throughout, we will consider the basic problem of minimizing an objective f(x) of n real variables x. We shall use the following standard trust-region method. Algorithm 6.1: Standard Trust-Region Algorithm An initial point x 0 and an initial trust-region radius \Delta 0 are given, as are constants ffl g , are required to satisfy the conditions 1. Stop if kr x f(x k )k 2 - ffl g . 2. Define a second-order Taylor series model q k and a positive-definite preconditioner . Compute a step s k to "sufficiently reduce the model" q k within the trust-region 3. Compute the ratio 4. Set Increment k by one and go to Step 1. We choose the specific values ffl set an upper limit of n iterations. The step s k in step 2 is computed using either Algorithm 5.1 or the Steihaug-Toint algorithm. Convergence in both algorithms for the subproblem occurs as soon as N. I. M. Gould, S. Lucidi, M. Roma and Ph. L. Toint or if more than n iterations have been performed. In addition, of course, the Steihaug-Toint algorithm terminates as soon as the boundary is crossed. All our tests were performed on an IBM RISC System/6000 3BT workstation with 64 Mega-bytes of RAM; the codes are all double precision Fortran 90, compiled under xlf90 with -O optimization, and IBM library BLAS are used. The test examples we consider are the larger examples from the CUTE test set (see Bongartz, Conn, Gould and Toint, 1995) for which negative curvature is frequently encountered. Tests were terminated if more than thirty CPU minutes elapsed. 6.1 Can we get much better model values than Steihaug-Toint? We first consider problems of the form (2.1). Our test examples are generated by running Algorithm 6.1 on the CUTE set for 10 iterations, and taking the trust-region subproblem at iteration as our example. The idea here is to simulate the kind of subproblems which occur in practice, not those which result at the starting point for the algorithm as such points frequently have special (favourable) properties. Our aim is to see whether there is any significant advantage in continuing the minimization of the trust-region subproblem once the boundary of the trust region has been encountered. We ran HSL VF05 to convergence, stopping when kg more than iterations had been performed. In all of the experiments reported here, the best value found was in fact the optimum value - a factorization of H was used to confirm that the matrix was positive semi-definite, while the algorithm ensured that the remaining optimality conditions hold - although, of course, there is no guarantee that this will always be the case. We measured the iteration (ST) and the percentage (ratio) of the optimal value obtained at the point at which the Steihaug-Toint method left the trust region, as well as the number of iterations taken to achieve 10%, 90% and 99% of the optimal reduction (10%, 90%, 99% respectively). The results of these experiments are summarized in Table 6.1. In this table we give the name of each example used, along with its dimension n, and the statistics "ratio"(expressed in the form x(y) as a shorthand for x \Theta 10 y ), "ST", "10%", "90%" and "99%" as just described. Some of the problems had interior solutions, in which case the "ratio" and "ST" statistics are absent (as indicated by a dash). We considered both the unpreconditioned method and a variety of standard preconditioners - a band preconditioner with semi-bandwidth of 5, and modified incomplete and sparse Cholesky factorizations, with the modifications as proposed by Schnabel and Eskow (1991) - used by the LANCELOT package (see, Conn, Gould and Toint, 1992, Chapter 3). The Cholesky factorization methods both failed for the problem MSQRTALS for which the Hessian matrix required too much storage. We make a number of observations. 1. On some problems, the Steihaug-Toint point gives a model value which is a good approximation to the optimal value. Solving the trust-region subproblem using the Lanczos method 19 no preconditioner 5-band example BRYBND 1000 3(-5) 23 24 28 CHAINWOO 1000 4(-5) 15 COSINE 1000 GENROSE 1000 MSQRTALS 1024 1(-1) 12 11 23 Incomplete Cholesky Modified Cholesky example COSINE 1000 GENROSE 1000 MSQRTALS 1024 factorization failure factorization failure Table 6.1: A comparison of the number of iterations required to achieve a given percentage of the optimal model value for a variety of preconditioners. See the text for a key to the data. N. I. M. Gould, S. Lucidi, M. Roma and Ph. L. Toint 2. On other problems, a few extra iterations beyond the Steihaug-Toint point pay handsome dividends. 3. Getting to within 90% or even 99% of the best value very rarely requires many more iterations than to find the Steihaug-Toint point. In conclusion, based on these numbers, we suggest that a good strategy would be to perform a few (say 5) iterations beyond the Steihaug-Toint point, and only accept the improved point if its model value is significantly better (as this will cost a second pass to compute the Lanczos vectors). We shall consider this further in the next section. 6.2 Do better values than Steihaug-Toint imply a better trust-region method? We now consider how the methods we have described for approximately solving the trust-region subproblem perform within a trust-region algorithm. Of particular interest is the question as to whether solving the subproblem more accurately reduces the number of trust-region iterations, or more particularly the cost of solving the problem - the number of iterations is of concern if the evaluation of the objective function and its derivatives is the dominant cost as then there is a direct correlation between the number of iterations and the overall cost of solving the problem. In Tables 6.2 and 6.3, we compare the Steihaug-Toint scheme with the GLTR algorithm (Algorithm 5.1) run to high accuracy. We exclude the problem HYDC20LS for our reported results as no method succeeded in solving the problem in fewer than our limit of n iterations, and the problems BROYDN7D and SPMSRTLS as a number of different local minima were found. In these tables, in addition to the name and dimension of each example, we give the number of objective function ("#f") and derivative ("#g") values computed, the total number of matrix-vector products ("#prod") required to solve the subproblems, and the total CPU time required in seconds. We compare the same preconditioners M as we used in the previous section. We indicate those cases where one or other method performs at least 10% better than its competitor by highlighting the relevant figure in bold. We observe the following. 1. The use of different M leads to radically different behaviour. Different preconditioners appear to be particularly suited to different problems. Surprisingly, perhaps, the unpreconditioned algorithm often performs the best overall. 2. In the unpreconditioned case, the model-optimum variant frequently requires significantly fewer function evaluations than the Steihaug-Toint method. However, the extra algebraic costs per iteration often outweigh the reduction in the numbers of iterations. The advantage in function calls for the other preconditioners is less pronounced. Ideally, one would like to retain the advantage in numbers of function calls, while reducing the cost per iteration. As we noted in Section 6.1, one normally gets a good approximation to the optimal model value after a modest number of iterations. Moreover, while the Steihaug-Toint point often gives a significantly sub-optimal value, a few extra iterations usually suffices to give Solving the trust-region subproblem using the Lanczos method 21 no preconditioner Steihaug-Toint model optimum example iterations 865 577 34419 145.02 DQRTIC 1000 43 43 83 0.3 43 43 91 0.3 FREUROTH 1000 17 17 34 0.4 17 17 34 0.4 GENROSE 1000 859 777 6092 28.8 773 642 24466 82.2 MSQRTALS 1024 44 34 7795 486.0 NONCVXUN 1000 492 466 177942 1017.9 ? 1800 seconds SENSORS 100 20 19 37 6.4 20 19 140 8.8 SINQUAD 5000 182 114 363 24.3 161 106 382 24.6 5-band Steihaug-Toint model optimum example CHAINWOO 1000 146 99 145 4.8 191 123 196 6.3 COSINE 1000 21 15 20 0.4 21 15 CRAGGLVY 1000 22 22 21 1.1 22 22 21 1.1 DQRTIC 1000 54 54 53 0.9 54 54 53 1.0 EIGENALS 930 56 43 171 75.2 53 42 222 75.8 FREUROTH 1000 20 iterations ? n iterations MANCINO 100 91 72 90 87.2 52 43 90 52.2 MSQRTALS 1024 88 62 9793 700.2 73 52 19416 1292.2 22 N. I. M. Gould, S. Lucidi, M. Roma and Ph. L. Toint Incomplete Cholesky Steihaug-Toint model optimum example CHAINWOO 1000 174 115 173 8.1 183 121 309 10.3 COSINE 1000 22 17 26 0.8 22 19 49 1.2 CRAGGLVY 1000 22 22 21 1.5 22 22 21 1.5 DQRTIC 1000 54 54 53 0.9 54 54 53 1.1 iterations ? n iterations GENROSE 1000 948 629 951 35.5 496 322 847 23.5 28 MSQRTALS 1024 factorization failure factorization failure 28 150 41.2 iterations ? n iterations iterations ? n iterations SINQUAD 5000 77 52 89 542.6 78 50 121 526.7 Modified Cholesky Steihaug-Toint model optimum example BRYBND 1000 15 15 14 2.2 59 37 61 7.7 CHAINWOO 1000 178 119 177 7.6 183 121 309 10.3 COSINE 1000 CRAGGLVY 1000 23 23 33 1.4 22 22 21 1.6 DQRTIC 1000 54 54 53 1.2 54 54 53 1.1 iterations ? n iterations GENROSE 1000 462 332 463 16.5 496 322 847 23.4 MANCINO 100 31 28 28 MSQRTALS 1024 factorization failure factorization failure Solving the trust-region subproblem using the Lanczos method 23 a large percentage of the optimum. Thus, we next investigate both of these issues in the context of an overall trust-region method. In Tables 6.4 and 6.5, we compare the number of function evaluations (#f), and the CPU time taken to solve the problem for the Steihaug-Toint ("ST") method with a number of variations on our basic GLTR method (Algorithm 5.1). The basic requirement is that we compute a model value which is at least 90% of the best value found during the first pass of the GLTR method. If this value is obtained by an iterate before that which gives the Steihaug-Toint point, the Steihaug- Toint point is accepted. Otherwise, a second pass is performed to recover the first point at which 90% of the best value was observed. The other ingredient is the choice of the stopping rule for the first pass. One possibility is to stop this pass as soon as the test (6.4) is satisfied. We denote this strategy by "90%best". The other possibility is to stop when either (6.4) is satisfied or at most a fixed number of iterations beyond the Steihaug-Toint point have occurred. We refer to this as "90%(ST+k)", where k gives the number of additional iterations allowed. We investigate the cases Once again, we compare the same preconditioners M as we used in the previous section. We highlight in bold those entries which are at least 10% better than the competition. The conclusions are broadly as before. Each method has its successes and failures, and there is no clear overall best method or preconditioner, although the unpreconditioned version performs surprisingly well. Restricting the number of iteration allowed after the Steihaug-Toint point has been found appears to curb the worst behaviour of the unrestricted method. N. I. M. Gould, S. Lucidi, M. Roma and Ph. L. Toint no preconditioner ST 90%(ST+1) 90%(ST+5) 90%(ST+10) 90%best example CRAGGLVY 1000 19 1.0 19 0.9 19 0.9 19 0.9 19 1.0 DQRTIC 1000 43 0.3 43 0.3 43 0.3 43 0.3 43 0.3 GENROSE 1000 859 28.8 748 38.9 721 48.1 738 57.3 728 60.0 MSQRTALS 1024 44 486.0 NONCVXUN 1000 492 1017.9 368 861.3 ? 1800 secs. ? 1800 secs. 433 1198.6 SENSORS 100 20 6.4 23 7.3 21 8.1 21 8.0 21 8.1 SINQUAD 5000 182 24.3 152 20.8 152 21.7 152 21.4 152 21.5 5-band ST 90%(ST+1) 90%(ST+5) 90%(ST+10) 90%best example CHAINWOO 1000 146 4.8 159 5.1 159 5.1 159 5.2 159 5.1 COSINE 1000 21 0.4 21 0.5 21 0.4 21 0.4 21 0.5 CRAGGLVY 1000 22 1.1 22 1.0 22 1.1 22 1.1 22 1.1 DQRTIC 1000 54 0.9 54 0.9 54 1.0 54 1.0 54 1.0 MANCINO 100 91 87.2 52 51.8 52 51.8 52 52.0 52 51.8 MSQRTALS 1024 88 700.2 97 756.7 73 704.9 74 844.7 79 981.5 Solving the trust-region subproblem using the Lanczos method 25 Incomplete Cholesky ST 90%(ST+1) 90%(ST+5) 90%(ST+10) 90%best example BRYBND 1000 55 3.9 56 4.2 56 4.3 56 4.3 56 5.0 CHAINWOO 1000 174 8.1 199 9.7 199 10.1 199 10.2 199 10.1 CRAGGLVY 1000 22 1.5 22 1.6 22 1.6 22 1.5 22 1.6 DQRTIC 1000 54 0.9 54 1.0 54 1.1 54 1.1 54 1.1 EIGENALS 930 76 94.6 77 97.2 74 97.2 74 97.3 74 96.8 GENROSE 1000 948 35.5 500 22.4 499 23.0 499 23.0 499 23.0 MSQRTALS 1024 fact. failure fact. failure fact. failure fact. failure fact. failure SINQUAD 5000 77 542.6 68 484.2 68 484.1 68 485.4 68 489.0 Modified Cholesky ST 90%(ST+1) 90%(ST+5) 90%(ST+10) 90%best example BRYBND 1000 15 2.2 15 2.2 15 2.3 15 2.2 15 2.2 CHAINWOO 1000 178 7.6 176 7.9 176 7.9 176 7.8 176 8.0 COSINE 1000 41 1.1 41 1.3 41 1.3 41 1.3 41 1.3 DQRTIC 1000 54 1.2 54 1.2 54 1.3 54 1.3 54 1.3 GENROSE 1000 462 16.5 434 18.8 434 19.3 434 19.1 434 19.1 MANCINO 100 31 129.3 64 232.3 77 275.9 77 275.5 77 275.6 MSQRTALS 1024 fact. failure fact. failure fact. failure fact. failure fact. failure 26 N. I. M. Gould, S. Lucidi, M. Roma and Ph. L. Toint 7 Perspectives and conclusions We have considered a number of methods which aim to find a better approximation to the solution of the trust-region subproblem than that delivered by the Steihaug-Toint scheme. These methods are based on solving the subproblem within a subspace defined by the Krylov space generated by the conjugate-gradient and Lanczos methods. The Krylov subproblem has a number of useful properties which lead to its efficient solution. The resulting algorithm is available as a Fortran 90 module, HSL VF05, within the Harwell Subroutine Library (1998). We must admit to being slightly disappointed that the new method did not perform uniformly better than the Steihaug-Toint scheme, and were genuinely surprised that a more accurate approximation does not appear to significantly reduce the number of function evaluations within a standard trust-region method. While this may limit the use of the methods developed here, it also calls into question a number of other recent eigensolution-based proposals for solving the trust-region subproblem (see Rendl, Vanderbei and Wolkowicz, 1995, Rendl and Wolkowicz, 1997, Sorensen, 1997, Santos and Sorensen, 1995). While these authors demonstrate that their methods provide an effective means of solving the subproblem, they make no effort to evaluate whether this is actually useful within a trust-region method. The results given in this paper suggest that this may not in fact be the case. This also leads to the interesting question as to whether it is possible to obtain useful low-accuracy solutions with these methods. We should not pretend that the formulae given in this paper are exact or even accurate in floating-point arithmetic. Indeed, it is well-known that the floating-point matrices Q k from the Lanczos method quickly loose M-orthonormality (see, for instance, Parlett, 1980, Section 13.3). Despite this, the method as given appears to be capable of producing usable approximate solutions to the trust-region subproblem. We are currently investigating why this should be so. One further possibility, which we have not considered so far, is to find an estimate - using the first pass of Algorithm 5.1, and then to compute the required s by minimizing the unconstrained model using the preconditioned conjugate gradient method. The advantage of doing this is that any instability in the first pass does not necessarily reappear in this auxiliary calculation. The disadvantages are that it may require more work than simply using (5.1), and that - must be computed sufficiently large to ensure that H + -M is positive semi-definite. Acknowledgement We would like to thank John Reid for his helpful advice on computing eigenvalues of tridiagonal matrices, and Jorge Mor'e for his useful comments on the Mor'e and Sorensen (1983) method. We are grateful to the British Council-MURST for a travel grant (ROM/889/95/53) which made some of this research possible. Solving the trust-region subproblem using the Lanczos method 27 --R CUTE: Constrained and unconstrained testing environment. A family of trust-region-based algorithms for unconstrained minimization with strong global convergence properties LANCELOT: a Fortran package for large-scale nonlinear optimization (Release Two new unconstrained optimization algorithms which use function and gradient values. Computing optimal locally constrained steps. Maximization by quadratic hill-climbing A catalogue of subroutines (release 13). An algorithm for minimization using exact second derivatives. Numerical experience with new truncated Newton methods in large scale unconstrained optimization. Computing a trust region step. The Symmetric Eigenvalue Problem. Tracking the progress of the Lanczos algorithm for large symmetric eigenproblems. Convergence properties of a class of minimization algorithms. A semidefinite framework for trust region subproblems with applications to large scale minimization. A new matrix-free algorithm for the large-scale trust-region subproblem A new modified Cholesky factorization. Newton's method with a model trust modification. Minimization of a large-scale quadratic function subject to a spherical constraint SIAM Journal on Optimization The conjugate gradient method and trust regions in large scale optimization. Towards an efficient sparsity exploiting Newton method for minimization. --TR --CTR Nicholas I. M. Gould , Philippe L. Toint, FILTRANE, a Fortran 95 filter-trust-region package for solving nonlinear least-squares and nonlinear feasibility problems, ACM Transactions on Mathematical Software (TOMS), v.33 n.1, p.3-es, March 2007 Giovanni Fasano , Massimo Roma, Iterative computation of negative curvature directions in large scale optimization, Computational Optimization and Applications, v.38 n.1, p.81-104, September 2007 Nicholas I. M. Gould , Dominique Orban , Philippe L. Toint, GALAHAD, a library of thread-safe Fortran 90 packages for large-scale nonlinear optimization, ACM Transactions on Mathematical Software (TOMS), v.29 n.4, p.353-372, December Nicholas I. M. Gould , Dominique Orban , Philippe L. Toint, CUTEr and SifDec: A constrained and unconstrained testing environment, revisited, ACM Transactions on Mathematical Software (TOMS), v.29 n.4, p.373-394, December Nicholas I. M. Gould , Philippe L. Toint, An iterative working-set method for large-scale nonconvex quadratic programming, Applied Numerical Mathematics, v.43 n.1-2, p.109-128, October 2002
trust-region subproblem;conjugate gradients;preconditioning;lanczos method
589272
Superlinearly Convergent Algorithms for Solving Singular Equations and Smooth Reformulations of Complementarity Problems.
We propose a new algorithm for solving smooth nonlinear equations in the case where their solutions can be singular. Compared to other techniques for computing singular solutions, a distinctive feature of our approach is that we do not employ second derivatives of the equation mapping in the algorithm and we do not assume their existence in the convergence analysis. Important examples of once but not twice differentiable equations whose solutions are inherently singular are smooth equation-based reformulations of the nonlinear complementarity problems. Reformulations of complementarity problems serve both as illustration of and motivation for our approach, and one of them we consider in detail. We show that the proposed method possesses local superlinear/quadratic convergence under reasonable assumptions. We further demonstrate that these assumptions are in general not weaker and not stronger than regularity conditions employed in the context of other superlinearly convergent Newton-type algorithms for solving complementarity problems, which are typically based on nonsmooth reformulations. Therefore our approach appears to be an interesting complement to the existing ones.
Introduction . In this paper we are interested in solving nonlinear equations in the case where their solutions can be singular, and smoothness requirements are weaker than those usually assumed in this context. Our development is partially motivated by the nonlinear complementarity problem, which we consider in detail, and for which our method takes a particularly simple and readily implementable form. be a given mapping, where V is a neighborhood of a point x in x being a solution of the system of equations In the sequel, F is assumed to be once (but not necessarily twice) dierentiable on V . In this setting x is referred to as singular solution if the linear operator F 0 (x) is singular, i.e., or, equivalently, In other cases x is referred to as regular solution. Computing Center of the Russian Academy of Sciences, Vavilova Str. 40, Moscow, GSP-1, Russia (izmaf@ccas.ru). Research of this author is supported by the Russian Foundation for Basic Research Grants 99-01-00472 and 01-01-00810. The rst author also thanks IMPA, where he was a visiting professor during the completion of this work. y Instituto de Matematica Pura e Aplicada, Estrada Dona Castorina 110, Jardim Bot^anico, Rio de Janeiro, RJ 22460-320, Brazil (solodov@impa.br). Research of this author is supported in part by CNPq Grant 300734/95-6, by PRONEX{Optimization, and by FAPERJ. A. F. IZMAILOV AND M. V. SOLODOV gives rise to numerous di-culties. It is well-known that for Newton-type methods at best one can guarantee linear convergence rate to a singular solution [6, 7, 9]. Moreover, it is not su-cient to choose a starting point only close enough to a solution (usually the set of appropriate starting points does not contain a full neighborhood of the solution, although this set is normally rather \dense" [18]). We refer the reader to the survey [19] and references therein. Another di-culty typical in this context is related to possible instability of a singular solution with respect to perturbations of F [27]. Certain special approaches to overcome those di-culties have been developed in the last two decades, but they employ second derivatives of F . Concerning methods for computing singular solutions, we cite [8, 20, 19, 43, 14, 1], and the more recent proposals in [26, 22, 21, 2, 27, 4] (of course, this list does not mention all contributions in this eld). One of the motivations for our new approach to solving singular equations lies in applications to the classical nonlinear complementarity problem (NCP) [37, 12, 13], which is to nd an x 2 R n such that smooth. One of the most useful approaches to numerical and theoretical treatment of the NCP consists in reformulating it as a system of smooth or nonsmooth equations [35, 29, 46]. One possible choice of a smooth reformulation is given by the following function (for other choices, see Section 5.1): It is easy to check that for this mapping the solution set of the system of equations coincides with the solution set of the NCP (1.2) [29, 47]. If x is a solution of the NCP, by direct computations (see Section 3), we obtain that denotes the standard basis in R n , and the index sets I 0 , I 1 and I 2 are dened by I 0 := I 1 := I 2 := It is immediately clear that F 0 (x) cannot be nonsingular, unless the index set I 0 is empty. The latter strict complementarity assumption is regarded rather restrictive. Therefore, smooth NCP reformulation provided by (1.3) gives rise to inherently singular solutions of the corresponding system of equations. In fact, it is known that any other smooth NCP reformulation has the same singularity properties [31] (see also Section 5.1). Furthermore, it is clear that F is once dierentiable with Lipschitz-continuous derivative (if g is twice continuously dierentiable), but F is not twice dierentiable when I 0 6= ;. This is also a common property shared by all useful smooth reformulations, e.g., see the collection [16]. Thus NCP reformulations provide an interesting example of once dierentiable nonlinear equations whose solutions are inherently singular. As discussed above, application of standard numerical techniques (e.g., Newton methods) in this context is prone to di-culties (and even failure) ALGORITHMS FOR SINGULAR EQUATIONS AND COMPLEMENTARITY 3 because of singularity. On the other hand, known special approaches to computing singular solutions are inapplicable, since these require second derivatives of F . This is the apparent reason why superlinearly convergent Newton-type algorithms for solving the NCP are typically based on nonsmooth equation reformulations and nonsmooth Newton methods (see [13] for a discussion and some references). In this paper we show that it is, in fact, possible to devise superlinearly convergent algorithms based on the smooth NCP reformulations. Specically, we propose an alternative approach based on computing singular solutions of the smooth reformulation stated above, and show that conditions needed for convergence of our method are principally dierent from those required for convergence of known nonsmooth algorithms. Thus the two can be considered as a complement to each other. We complete this section with some notation, which is fairly standard. We denote by L n the space of linear operators from R n to R n . For A 2 L n , let ker stand for its kernel (null space), and im stand for its image (range space). For a bilinear mapping and an element the linear operator B[p] 2 L n by Recall that symmetric bilinear mappings and linear operators of the form p are in isometrically isomorphic correspondence to each other, i.e., the correspondence is one-to-one, linear, and it preserves the norm. Therefore, in the sequel we shall not be making a formal distinction between those objects. Given a set S in a vector space, by conv S we denote its convex hull, and by span S its linear hull. Finally, by E we denote the identity operator in R n . 2. A General Approach to Solving Singular Equations. We start with describing an approach to computing singular solutions of twice dierentiable nonlinear equations, which was developed in [26, 22, 27]. We then extend it to the setting of once dierentiable mappings, and in the next section show how it applies to solving complementarity problems. A solution x of (1.1) being regular is equivalent to saying that im F 0 while singularity means that im F 0 (x) 6= R n . In this situation, one possibility to \regularize" a singular solution x is to add to the left-hand side of (1.1) another term, which vanishes at x (so that x remains a solution), and such that its Jacobian at x \compensates" for the singularity of F 0 (x) (so that to complement im F 0 (x) in R n ). It is natural to base this extra term on the information about the rst derivative of F . To this end, dene the mappings and consider the equation Suppose that P () is dened in such a way that for holds that Then, by the structure of , solution x of (1.1) is also a solution for (2.2). Furthermore, if F is su-ciently smooth (at least twice dierentiable at x), then under appropriate assumptions on the rst two derivatives of F at x, and on P () and h(), it is possible to ensure that is dierentiable at x, and x is a regular solution of (2.2). As these assumptions will not be used in this paper, we omit the details, referring the reader 4 A. F. IZMAILOV AND M. V. SOLODOV to [26, 27]. The regular solution x of (2.2) can be computed by means of eective special methods [26, 22, 27], or by conventional numerical techniques (the latter would typically require stronger assumptions, in order to ensure dierentiability of not only at x but also in its neighborhood). There exist certain general techniques to dene P () and h() with necessary properties (see [26, 27]). However, when one has additional information about the structure of singularity of F at x (e.g., recall (1.4) for the NCP reformulation), it can often be used to choose P () and h() in a particularly simple and constructive way. One such application is precisely the NCP, where the subspace im F 0 (x) can be identied (locally, but without knowing x), and so the two mappings can be chosen constant (see Section 3). In this paper, we shall focus exclusively on the case where it is possible to choose We emphasize that, of course, P should be determined without knowing the exact solution x. The simplest case when this is possible is when we know that corank F 0 when we are interested in determining a solution specically with this particular type of singularity. In that case, it is natural to take In Section 3, we show how appropriate P for the NCP reformulation can be determined using information available at any point close enough to a solution (but without knowing the solution itself). In general, if P () is dened as a constant satisfying (2.3), one also can usually take h() p, with p 2 R n n f0g being an arbitrary element. Indeed, with those choices the function dened by (2.1) takes the form and x is still a solution of (2.2), due to (2.3). If F is twice dierentiable at x, then it is clear that is dierentiable at this point, and Therefore, x is a regular solution of (2.2) if the linear operator in the right-hand side of (2.5) is nonsingular. This is possible under appropriate assumptions. Since the case of twice dierentiable F is not the subject of this paper, we shall not discuss here technical details. We only note that nonsingularity of (2.5) subsumes the condition Observe that the latter relation implies that (2.3) must hold as equality. Summarizing, we obtain the following assumptions on the choice of ker These assumptions clearly hold if, for example, P is the projector onto some complement of im F 0 (x) in R n parallel to im F 0 (x). With this choice, nonsingularity of (2.5) formally coincides with the notion of 2-regularity of at x with respect to p 2 R n , in the sense of [23, 3, 27]. We note, however, that this connection does not seem conceptually important, and in fact, appears to be in some sense a coincidence. Indeed, in the case of once dierentiable mappings considered below, nonsingularity condition that would be required no longer has any direct relation to 2-regularity for mappings with Lipschitzian derivatives, as dened in [24, 25]. As a nal note, we remark that it can be shown (by simple argument, see [26, 27]) that if there exists at least one element p 2 R n such that the operator (2.5) is nonsingular, then it will be so for almost every p 2 R n . ALGORITHMS FOR SINGULAR EQUATIONS AND COMPLEMENTARITY 5 We conclude the discussion of the twice dierentiable case by the following ex- ample, which is very simple but serves to illustrate the basic idea. Example 2.1. Let R be twice continuously dierentiable on V , where V is a neighbourhood of x 2 R which is a singular solution of (1.1). The latter means here that F Taking we obtain the following regularized equation: Obviously, which is distinct from zero for any p 2 R n f0g, This shows that in this example, if F 00 (x) 6= 0, singularity can be easily dealt with by using the second-order information. In the approach outlined above, F is assumed to be twice dierentiable. Suppose now that F is once (but not twice) dierentiable, and its rst derivative is Lipschitz-continuous on V . Then dened by (2.4) is also Lipschitz-continuous on V , and it is natural to try to apply to the corresponding equation (2.2) the generalized (non- smooth) Newton method [32, 33, 41, 42, 40, 28]. We emphasize that we shall use the nonsmooth Newton method to solve a (nonsmooth) regularization of a smooth equation. In the context of NCP, this should be compared to the more traditional approach of solving an inherently nonsmooth reformulation by the nonsmooth Newton method. As we shall show in Section 4, the two dierent approaches lead to two dierent regularity conditions, neither of which is weaker or stronger than the other. Let @(x) denote the Clarke's generalized Jacobian [5] of at x 2 V . That is where @B (x) stands for the B-subdierential [45] of at x, which is the set with D V being the set of points at which is dierentiable. With this notation, the nonsmooth Newton method is the following iterative procedure: It is well-known [42, 40, 28] that if (i) is semismooth [36] at x, and (ii) all the linear operators comprising @(x) are nonsingular, then the process (2.7) is locally well-dened and superlinearly convergent to x. More- over, if is strongly semismooth [36] then the rate of convergence is quadratic. The regularity condition (ii) can be relaxed if a more specic rule of determining employed. For example, if one chooses is enough to assume BD-regularity, i.e., that all elements in @B (x) are nonsingular [41]. In applications, usually has some special (tractable) structure, and at each iterate x k we are interested in obtaining just one, preferably easily computable, This would be precisely the case here. The choice of an element in @(x) that we suggest to use in the nonsmooth Newton method for solving (2.2) with given by (2.4), is the following: denotes the usual directional derivative of the mapping at with respect to a direction p 2 R n . In Section 3, we show that this 6 A. F. IZMAILOV AND M. V. SOLODOV H(x) is explicitly and easily computable for the NCP reformulations. The validity of the choice suggested in (2.8) for an element of @(x) is actually not so obvious. The possibility of choosing the directional derivative ( as an element in the generalized Jacobian of is based on the following fact. At a point x its derivative is in fact the second derivative of PF (). Due to can be considered as a symmetric bilinear mapping. This symmetry will be essential in the proof of Lemma 2.1 below. For a mapping x ! Q(x)p, where is an arbitrary Lipschitzian mapping, the inclusion can be in general invalid. Lemma 2.1. Suppose that F : V ! R n has a Lipschitzian derivative on V , where V is an open set in R n . Assume that for some the mapping is directionally dierentiable at a point x 2 V with respect to a direction p 2 R n . are dened in (2.8) and (2.4), respectively. Proof. Since clearly Lipschitz-continuous, using further the assumption that PF 0 is directionally dierentiable at x with respect to p, it follows that there exists a linear operator The above conclusion can be deduced from [42, Lemma 2.2(ii)] after identifying the space L n with the equivalent space R m , using the equivalence of the norms in nite-dimensional spaces. By the denition of the generalized Jacobian, B 2 @( means that there exist an integer m, sequences fx i;k g V and numbers i , with the following properties: i is dierentiable at each x i;k , and where the limits in the right-hand side of the second equality exist for each Note that dierentiability of at each x i;k means that the mapping twice dierentiable at these points. Taking into account the symmetry of the bilinear mapping representing the second derivative, we conclude that Therefore, where the second equality follows from (2.10), and the third from (2.9). Using the denition of the generalized Jacobian, we conclude that H(x) 2 @(x). Remark 2.1. There exists another way to construct the regularized equation which can have advantages in certain situations over the one described above. Specically, the mapping dened by (2.4) can be modied as follows: ALGORITHMS FOR SINGULAR EQUATIONS AND COMPLEMENTARITY 7 It is clear that with this denition, x is still a solution of Modifying H accordingly, we have Furthermore, it is clear that Lemma 2.1 is still valid with and H dened by (2.11) and (2.12). Finally, it is easy to see that since the possible limits of H(x) as x are the same, whether H is dened by (2.8) or (2.12). Hence, the regularity condition at x that would be needed for the superlinear convergence of our method is again the same, whether the method is applied to one regularized equation or the other. The possible advantage of the modied equation is the following. If singularity of F 0 (x) has a certain structure, then not all the components of F may need to be computed in (2.11). Furthermore, (2.12) can also take a simpler form in that case. For example, suppose that F 0 (x) is such that satisfying (2.3) can be chosen as the orthogonal projector onto the subspace span fe is the standard basis in R n and I P is the orthogonal projector onto span fe ng n Ig. It is easy to see that in this case, (2.11) would not require computing the function values F i I . Furthermore, the derivatives of would not appear in (2.12), and so this part would also be simplied. This feature would be further illustrated in the context of NCP in Section 3. In the sequel, we shall also consider the following modication of the Newton algorithm (2.7), which will be useful for solving the NCP reformulation in Section 3: This modication is essentially motivated by the idea of \truncating" elements of the (generalized) Jacobian by omitting the terms which vanish at the solution x. These terms typically involve some higher-order derivatives of the problem data (in the context of NCP (1.2), the second derivatives of g), and so it can be advantageous not to compute them, if possible. Note that the regularity condition which is typically employed in nonsmooth Newton methods consists of saying that every element in the generalized Jacobian @(x) (or the B-subdierential @B (x)) is nonsingular (recall condition (ii) stated above). This seems to be unnecessarily restrictive, because in most implementable algorithms some specic rule to choose H(x k used. We shall therefore replace the traditional condition by a weaker one. Specically, we shall assume that all the possible limits of H(x k ) as x k ! x are nonsingular, where H(x k ) is precisely the element given by (2.8) (or by (2.12)). To this end, we shall dene the set Hg: Our regularity assumption would be that elements in p (x) are nonsingular. We remind the reader that this set is the same for both choices of H , i.e., (2.8) and (2.12). We point out that unlike in the twice dierentiable case, this regularity condition cannot be related to the notion of 2-regularity [24, 25] of at x. Lemma 2.1 in hand, convergence of algorithms (2.7) and (2.13), with the data dened in (2.4) and (2.8) or (2.11) and (2.12), can be established similarly to [41], but taking into account the modied nonsingularity assumption. 8 A. F. IZMAILOV AND M. V. SOLODOV Theorem 2.2. Suppose has a Lipschitz-continuous derivative on is a neighborhood of a solution x of (1.1). Let Assume further that the mapping directionally dierentiable with respect to a direction p 2 R n at any point in V , and the mapping is semismooth at x. Let and H be dened by (2.4) and (2.8), or (2.11) and (2.12). Assume further that all linear operators comprising p (x) are nonsingular. Then the iterates given by (2.7) or (2.13) are locally well-dened, and converge to x superlinearly. If, in addition, the mapping strongly semismooth at x then the rate of convergence is quadratic. Proof. It is easy to see that under our regularity assumption, (H()) 1 is locally uniformly bounded. Indeed, assume the contrary, i.e., that there exists a sequence x, and the sequence f(H(x k is unbounded (this subsumes the possibility that some elements of the latter sequence are not even well-dened). Recall that the generalized Jacobian is locally bounded [5]. Since, by Lemma 2.1, H(x k ) 2 @(x k ) for every k, it follows that the sequence fH(x k )g is bounded. Hence, we can assume that fH(x k )g converges to some the inclusion is by the very denition of the set p (x). But then H is nonsingular, which is in contradiction with the earlier assumption that f(H(x k is unbounded. Consider rst algorithm (2.7), and suppose that the iterates are well-dened up to some index k 0. We have that where M > 0. Note that when semismooth, so is (). It is known [40, Proposition 1] that semismoothness of at x implies that sup (the latter property was introduced in the context of the nonsmooth Newton methods in [33]). Using Lemma 2.1 and combining the last two relations, well-denedness of the whole sequence fx k g and its superlinear convergence to x follow by a standard argument. In the strongly semismooth case, one has that sup and so convergence is quadratic. Consider now the iterates fx k g generated by (2.13). By our regularity assumption and the classical results of linear analysis, the condition implies that Hence, ALGORITHMS FOR SINGULAR EQUATIONS AND COMPLEMENTARITY 9 where the Lipschitz-continuity of was also used. It follows that the dierence between the original and modied steps is of the second order. By the obvious argument, it can now be easily seen that the modied algorithm has the same convergence rate as the original one. Note that in principle, our regularity condition depends not only on the structure of singularity of F at x, but also on the choice of p. Implementation of this approach presumes that there exists at least one p 2 R n for which this condition is satised. Furthermore, a way to choose such p should be available. Fortunately, a typical situation is the following. The existence of one suitable p can usually be established under some reasonable regularity assumption. Then, given the existence of one such it can further be proven that the set of appropriate elements is, in fact, open and dense in the whole space. Hence, p can be chosen arbitrary, with the understanding that almost any is suitable. We shall come back to this issue in Section 4, where regularity conditions for NCP are discussed. In the computational experience of [22, 26], where conceptually related methods for smooth operator equations are considered, a random choice of p does the job. Even though this choice certainly aects the rate and range of convergence, the dierences between dierent choices are usually not dramatic. Finally, we remark that the development presented above can be extended to the case when P () is not necessarily constant, but it is a Lipschitzian mapping satisfying with In that case, we would have to provide a technique to dene such P () in the general setting. Such techniques are possible, but they go beyond the scope of the present paper. Here we are mainly concerned with a specic application of our approach to the nonlinear complementarity problem, which we consider next. 3. Algorithm for the Nonlinear Complementarity Problem. Consider the nonlinear complementarity problem (1.2), and its reformulation as a system of smooth equations (1.1), given by (1.3). For convenience, we re-state the associated function F , which is We choose a specic reformulation for the clarity of presentation. In Section 5.1, we show that our analysis is intrinsic and extends to other smooth reformulations. Let x 2 R n be a solution of NCP. Suppose that g is twice continuously dier- entiable in some neighborhood V of x in R n . Then it is easy to see that F has a Lipschitz-continuous derivative on V , which is given by is the standard basis in R n . Recalling the three index sets I 0 := I 1 := I 2 := from (3.1) we immediately obtain that (3. A. F. IZMAILOV AND M. V. SOLODOV As already discussed in Section 1, the Jacobian F 0 (x) is necessarily singular whenever I 0 6= ;, the latter being the usual situation for complementarity problems of interest. Furthermore, F is not twice dierentiable. Hence, smooth NCP reformulations fall precisely within the framework of Section 2. Such equations cannot be eectively solved by previously available methods, and so our approach comes into play. We next show that in the setting of NCP, the general algorithm introduced in Section 2 takes a simple implementable form. Given the structure of F 0 (x), we have that Then the natural choice of condition (2.3)), is the operator with the matrix representation consisting of rows At the end of this section, we shall show how to dene knowing the solu- tion x (clearly, this task reduces to identifying the index set I 0 ). This is possible by means of error bound analysis. A su-cient condition for our error bound is weaker than b-regularity [39], which is currently the weakest assumption under which Newton methods for nonsmooth NCP reformulations are known to be (superlinearly) convergent [30, 34]. Once P is dened according to (3.3), we x p 2 R n n f0g arbitrarily. Then the function dened by (2.4) takes the form According to Section 2, x is a solution of which is our \regularized" equa- tion. We proceed to derive explicit forms for iterations of algorithms (2.7) and (2.13), and the regularity condition needed for their convergence. First, by (2.8) and (3.3), the matrix representation of H(x), which is the element of @(x) employed in algorithm (2.7), consists of rows Furthermore, the directional derivatives employed in (3.5) exist, and can be obtained explicitly from (3.1): where ALGORITHMS FOR SINGULAR EQUATIONS AND COMPLEMENTARITY 11 Note that according to (3.5), one has to compute Another useful observation which would suggest truncation of the Jacobian to be discussed later, is that for i 2 I 0 all the terms in (3.6) involving the second derivatives of g vanish at x. Furthermore, taking into account (3.5), (3.2), (3.6) and (3.7), we conclude that the matrix representation of an arbitrary limit point H of H(x) as x ! x consists of rows Hence, we can state the following su-cient condition for nonsingularity of every linear operator in p (x). Denote by J the collection of pairs of index sets (J that Our regularity condition consists of saying that for every pair of index sets (J holds that are linearly independent in R In Section 4, we shall discuss the relation between this condition and other regularity conditions for the NCP, as well as compare convergence results of our algorithm with convergence results of other locally superlinearly convergent equation-based methods for solving NCP. Under our assumptions, semismoothness of readily from (3.1) and standard calculus of semismooth mappings [36, Theorem 5]. Moreover, under the additional assumption of Lipschitz-continuity of g 00 () on V , strongly semismooth, which follows from results on the superposition of strongly semismooth mappings [15, Theorem 19]. Hence, () is (strongly) semismooth. Note that all the elements involved in the iteration scheme (2.7) are computed in this section by explicit formulas. In principle, computing H via (3.5)-(3.7) involves second derivatives of g. However, as already noted above, the terms containing second derivatives of g tend to zero as x ! x. This suggests the idea to modify the process by omitting these terms, which leads to the method represented by (2.13). We shall also take into account the structure of P , and make use of Remark 2.1. Note that for given by (3.3), we have that (E P ) is the orthogonal projector onto span fe g. According to (2.11), we can therefore re-dene Taking into account (2.12) and omitting further the terms that vanish at x, we can take ~ Comparing expressions (3.9) and (3.10) with (3.4) and (3.5), one can easily observe that the former are simpler and require less computations. A. F. IZMAILOV AND M. V. SOLODOV Furthermore, under our smoothness assumptions, it is easy to see that and so the modied Newton method given by (2.13) is applicable. We next give a formal statement of the convergence result for our methods applied to NCP, which is a corollary of Theorem 2.2. Theorem 3.1. Let be a twice continuously dierentiable mapping on being a neighborhood of a solution x of the NCP (1.2). Assume that for some condition (3.8) is satised for every pair of index sets (J Then the iterates given by (2.7) or (2.13) (with all the objects as dened in this section) converge to x locally superlinearly. If, in addition, the second derivative of g is Lipschitz-continuous on V , then the rate of convergence is quadratic. We next show how to construct knowing the solution x. Given the structure of P , see (3.3), it is clear that this task reduces to correct identication of the degenerate set I 0 . This can be done with the help of error bounds, as described next (our approach is in the spirit of the technique developed in [10] for identication of active constraints in nonlinear programming). To our knowledge, the weakest condition under which a local error bound for NCP is currently available is the 2- regularity of F given by (1.3) at the NCP solution x [25]. Specically, if F is 2-regular at x then there exist a neighborhood U of x in R n and a constant M 1 > 0 such that We shall not introduce the notion of 2-regularity formally here, as this would require an extensive discussion. We only emphasize that the bound (3.11) may hold when the so-called natural residual minfx; g(x)g does not provide an error bound, and always holds when it does (see [25], and in particular [25, Example 1]). Hence, 2-regularity of F is a weaker assumption than the R 0 type property or semistability, which in case of NCP are both equivalent to an error bound in terms of the natural residual [38]. And it is further weaker than b-regularity, see [25]. We note that in Lemma 3.2 below we could also use other error bounds for identifying I 0 . However, they would require either stronger local assumptions, or global assumptions. Lemma 3.2. Suppose that x is a solution of NCP, g is Lipschitz-continuous on is a neighborhood of x. Suppose nally that the local error bound (3.11) holds. Then for any 2 (0; 1) there exists a neighborhood U of x such that ng Proof. It is easy to observe that there exist some M 2 > 0 and some neighborhood U of x such that where the inequality follows from the Lipschitz-continuity of the functions involved. Therefore, by (3.11) (possibly adjusting the neighborhood U ), for an arbitrary xed we have that ALGORITHMS FOR SINGULAR EQUATIONS AND COMPLEMENTARITY 13 In particular, the quantity in the left-hand side of the inequality above tends to zero as x tends to x. On the other hand, it is clear that there exists " > 0 such that (U should be adjusted again, if necessary). Combining those facts, we obtain (3.12) for U su-ciently small. By Lemma 3.2, the index set I 0 , and hence the mapping P , are correctly identied by (3.12), provided one has a point close enough to the solution. We note that this requirement of closedness to solution is completely consistent with the setting of the paper, since the subject under consideration are superlinearly convergent Newton-like methods, which are local by nature. Finally, we mention other considerations that can also be useful for identifying I 0 . Sometimes the cardinality r of I 0 may be known from a priori analysis of the problem, or one can be interested in nding an NCP solution with a given cardinality of I 0 . Then for any x 2 R n su-ciently close to x, the set I 0 coincides with the set of indices corresponding to the r smallest values of j maxfg i (x); x i gj. In this case, no error bound is needed to identify I 0 . We note that in the present setting, cardinality of I 0 is closely related to corank of singularity. In the literature on numerical methods for solving singular equations, the assumption that corank of singularity is known is absolutely standard [20, 19, 43, 14, 1, 2]. In the complementarity literature, on the other hand, assumptions about cardinality of I 0 are not common, except possibly for I 4. Regularity Conditions. The weakest condition under which there exists a locally superlinearly convergent Newton-type algorithm for solving a (nonsmooth) equation reformulation of the NCP, is the b-regularity assumption, which can be stated as follows: for every pair of index sets (J holds that are linearly independent in R Under this assumption, the natural residual mapping x ! is BD-regular at x. Furthermore, it is also (strongly) semismooth under standard assumptions on g. Hence, the nonsmooth Newton method (2.7) based on it converges locally superlinearly [30, 34]. Note that Newton methods applied to another popular reformulation based on the Fischer-Burmeister function [17, 11], require for convergence the stronger R-regularity [44] assumption, see [34]. In what follows, we compare our regularity condition (3.8) with b-regularity, and show that they are essentially dierent. In general, neither is weaker or stronger than the other. This implies that our approach based on the smooth NCP reformulation is a complement to nonsmooth reformulations, and vice versa, as each approach can be successful in situations when the other is not. The next result is important to obtain an insight into the nature of our regularity condition (3.8). We start with the following denition. Definition 4.1. A solution x of the NCP (1.2) is referred to as quasi-regular, if for every pair of index sets (J there exists an element such that (3.8) is satised. Proposition 4.2. Suppose that the solution x of the NCP (1.2) is quasi-regular. Then there exists a universal p 2 R n which satises (3.8) for every pair (J J . Moreover, the set of such p is open and dense in R n . 14 A. F. IZMAILOV AND M. V. SOLODOV Proof. Fix a pair (J consider the determinant of the system of vectors in (3.8) as a function of p. This function is a polynomial on R n , and this polynomial is not everywhere zero, since it is not zero at But then the set where the polynomial is not zero is obviously open and dense in R n . Moreover, the intersection of such sets corresponding to pairs (J open and dense, since it is a nite intersection of open and dense sets. It follows that if x is a quasi-regular solution of NCP in the sense of Denition 4.1, then even picking a random p 2 R n one is extremely unlikely to pick a \wrong" (as the set of wrong elements is \thin"). Hence, under the assumption of quasi- regularity, for the implementation of the algorithm described in Section 3 we can choose arbitrarily, with the understanding that almost every p 2 R n is appropriate. In particular, for all practical purposes, we can think of quasi-regularity as the regularity condition needed for superlinear convergence of our algorithm. We next investigate the relationship between quasi-regularity and b-regularity. First, we show that if the cardinality of I 0 is equal to one, then quasi-regularity is in fact weaker than b-regularity. Proposition 4.3. Suppose that x is a b-regular solution of the NCP, and the cardinality of I 0 is equal to one. Then x is quasi-regular. Proof. Let I and denote g. In this setting, b-regularity clearly means that corresponding to the two possible choices of (J It follows that Assume for a contradiction that x is not quasi-regular. Then by Denition 4.1, there exists a pair (J such that for every p 2 R n condition (3.8) is violated. This means that is either or Taking any q 2 L ? n f0g, we deduce that for every p 2 R n either or Setting we then obtain that either or ALGORITHMS FOR SINGULAR EQUATIONS AND COMPLEMENTARITY 15 which contradicts b-regularity, because of (4.1). It is easy to see that in the setting of Proposition 4.3, the quasi-regularity condition can be satised without b-regularity. For example, let g 0 Then b-regularity is violated. On the other hand, quasi-regularity here is equivalent to saying that there exist elements (corresponding to the two possible choices of (J which is satised for almost any p 1 and p 2 , provided It is also quite clear that just choosing randomly should do the job. In general, i.e., in the cases of higher cardinality of I 0 , b-regularity and quasi-regularity become dierent, not directly related conditions. In particular, neither is stronger or weaker than the other, as illustrated by the following examples. Example 4.1. Let Then b-regularity is obvious, but This means that for J does not hold for any p, and so the quasi-regularity condition is not satised. Example 4.2. Let Then b-regularity does not hold (the linear independence condition is violated for but quasi-regularity is satised, which can be shown by straight-forward computations. We omit the details, as they do not provide any further insight We complete our discussion with a su-cient condition for quasi-regularity of x, which is meaningful when the cardinality of I 0 is not greater than n=2, half dimensionality of the space. Specically, suppose that are linearly independent in R n ; and are linearly independent in R It is clear that (4.3) is subsumed by b-regularity (where it must hold for all partitions of I 0 ). It is also not di-cult to see that (4.3) is necessary for quasi-regularity of x. Hence, this assumption does not introduce any additional restrictions with respect to regularity conditions under consideration. Furthermore, for non-pathological problems the cardinality of I 0 should not be too large compared to the dimensionality of the space, and so condition (4.2) should not be di-cult to satisfy. Therefore, (4.2) and (4.3) appear to be not restrictive. Proposition 4.4. Suppose that (4.2) and (4.3) hold. Then x is a quasi-regular solution of NCP. A. F. IZMAILOV AND M. V. SOLODOV Proof. Take any pair of index sets (J consider the system of (twice the cardinality of I 0 ) linear equations in the variable p 2 R n . Under the assumption (4.2), this system has a solution further that substituting this p into (3.8) reduces the system of vectors appearing in (3.8) precisely to the system of vectors appearing in (4.3), which is linearly independent by the hypothesis. Again, it is easy to see that the latter su-cient condition for quasi-regularity of x can hold without b-regularity. On the other hand, in general it is not implied by b-regularity. In particular, b-regularity need not imply (4.2). Summarizing the preceding discussion, we conclude that the regularity assumption required for the algorithm proposed in Section 3 for solving the NCP is dier- ent from b-regularity, which is the typical assumption in the context of nonsmooth Newton-type methods for solving nonsmooth NCP reformulations. In fact, the two assumptions are of a rather distinct nature. This is not surprising, considering that they result from approaches which are also quite dierent. 5. Some Further Applications. The general approach presented in Section 2 can be also useful in other problems where complementarity is present. Below we outline applications to another class of smooth reformulations of NCP (dierent from (1.3)), and to the mixed complementarity problems. We limit this discussion to exhibiting the structure of singularity associated with the smooth equation reformulations of those problems. Deriving the resulting regularity conditions and comparing them to known ones requires too much space. Without going into detail, we claim that regularity assumptions needed for our approach would again be dierent from assumptions of Newton methods for nonsmooth equations. 5.1. Other NCP Reformulations. The analysis presented in Sections 3 and 4 for NCP is intrinsic in the sense that it is also applicable to smooth reformulations other than the one given by (1.3). Indeed, following [35], consider the family of functions strictly increasing function such that It can be checked that the NCP solution set coincides with zeroes of F . As an aside, note that reformulation (1.3) cannot be written in the form stated above, so the two are really dierent. Suppose further that is dierentiable on R with 0 t > 0. For example, we could take Let x be some solution of NCP, and V be its neighborhood. If g is twice continuously dierentiable on V and 0 is Lipschitz-continuous, then the derivative of F is Lipschitz-continuous near x, and it is given by ALGORITHMS FOR SINGULAR EQUATIONS AND COMPLEMENTARITY 17 As is easy to see, Since 0 (t) > 0 for any t > 0, we conclude that the structure of singularity here is absolutely identical to that for F given by (1.3) (recall (1.4)). In particular, and all the objects and the analysis in Sections 3 and 4 can be derived in a similar fashion. 5.2. Mixed Complementarity Problems. The mixed complementarity problem (MCP) is a variational inequality on a (generalized) box ug, where l i 2 [1;+1) and u are such that l i < u Specically, the problem is to nd It can be seen that this is equivalent to x 2 R n satisfying the following conditions: for every if if if NCP is a special case of MCP corresponding to l We claim that solutions of MCP coincide with zeroes of the function F whose components are given by We omit the proof, which can be carried out by direct verication. Let x be some solution of MCP, and V be its neighborhood. If g is twice continuously dierentiable on V then the derivative of F is Lipschitz-continuous near x. Dening I 0 := I 1 := I 2 := it can be veried that A. F. IZMAILOV AND M. V. SOLODOV where In particular, i . Observing the structure of F 0 (x), further analysis can now follow the ideas of Sections 3 and 4. 6. Concluding Remarks. We have presented a new approach to solving smooth singular equations. Unlike previously available algorithms, our method is applicable when the equation mapping is not necessarily twice dierentiable. Important examples of once dierentiable singular equations are reformulations of the nonlinear complementarity problems, which we have studied in detail. In particular, we have demonstrated that in the case of NCP our method takes a readily implementable simple form. Furthermore, the structure of singularity can be completely identied by means of local error bound analysis, without knowing the solution itself. It was further shown that the regularity condition required for the superlinear convergence of the presented algorithm is dierent from conditions needed for the nonsmooth Newton methods applied to nonsmooth NCP reformulations. Thus the two approaches should be regarded as complementing each other. Finally, it was demonstrated that the main ideas of this paper should be also applicable to other problems where complementarity structures are present. --R Optimality Conditions: Abnormal and Degenerate Problems. An approach to Optimization and Nonsmooth Analysis. Newton's method at singular points. Newton's method at singular points. Convergence acceleration for Newton's method at singular points. Convergence rates for Newton's method at singular points. On the accurate identi A new merit function for nonlinear complementarity problems and a related algorithm. Pang (editors). Complementarity and variational problems: State of the Art. A geometric framework for the numerical study of singular points. Solution of monotone complementarity problems with locally Lipschitzian functions. On the resolution of monotone complementarity problems. Starlike domains of convergence for Newton's method at singularities. On solving nonlinear equations with simple singularities or nearly singular solutions. Characterization and computation of generalized turning points. Augmented systems for computation of singular points in Banach space problems. Stable methods for On certain generalizations of Morse's lemma. The theory of 2-regularity for mappings with Lipschitzian derivatives and its applications to optimality conditions bounds for 2-regular mappings with Lipschitzian derivatives and their applications Local regularization of certain classes of nonlinear operator equations. Semismoothness and superlinear convergence in non-smooth optimization and nonsmooth equations Some equation-based methods for the nonlinear complementarity problem Solving box constrained variational inequality problems by using the natural residual with D-gap function globalization A new class of semismooth Newton-type methods for nonlinear complementarity problems Newton's method for nondi Newton's method based on generalized derivatives for nonsmooth functions. A theoretical and numerical comparison of some semismooth algorithms for complementarity problems. Equivalence of the complementarity problem to a system of nonlinear equations. Pang Complementarity problems. Pang Private communication NE/SQP: A robust algorithm for the nonlinear complementarity Nonsmooth equations: Motivation and algorithms. Convergence analysis of some algorithms for solving nonsmooth equations. A nonsmooth version of Newton's method. Characterization and computation of singular points with maximum rank de Strongly regular generalized equations. Local structure of feasible sets in nonlinear programming Growth behavior of a class of merit functions for the nonlinear complementarity problem. --TR
singularity;superlinear convergence;complementarity;reformulation;nonlinear equations;regularity
589276
A Polynomial Time Algorithm for Shaped Partition Problems.
We consider the class of shaped partition problems of partitioning n given vectors in d-dimensional criteria space into p parts so as to maximize an arbitrary objective function which is convex on the sum of vectors in each part, subject to arbitrary constraints on the number of elements in each part. This class has broad expressive power and captures NP-hard problems even if either d or p is fixed. In contrast, we show that when both d and p are fixed, the problem can be solved in strongly polynomial time. Our solution method relies on studying the corresponding class of shaped partition polytopes. Such polytopes may have exponentially many vertices and facets even when one of d or p is fixed; however, we show that when both d and p are fixed, the number of vertices of any shaped partition polytope is $O(n^{d{p\choose 2}})$ and all vertices can be produced in strongly polynomial time.
Introduction The Partition Problem concerns the partitioning of vectors A in d-space into p parts so as to maximize an objective function which is convex on the sum of vectors in each part; see [3]. Each vector A i represents d numerical attributes associated with the ith element of the set ng to be partitioned. Each ordered partition - is then associated with the d \Theta p matrix A whose jth column represents the total attribute vector of the jth part. The problem is to find an admissible partition - which maximizes an objective function f given by a real convex functional on IR d\Thetap . Of particular interest is the Shaped Partition Problem, Department of Applied Mathematics, Chiaotung University, Hsinchu, 30045, Taiwan. Email address: fhwang@math.nctu.edu.tw. y William Davidson Faculty of Industrial Engineering and Management, Technion-Israel Institute of Technology, 32000 Haifa, Israel. Email address: onn@ie.technion.ac.il. Research supported in part by the Mathematical Sciences Research Institute at Berkeley California through NSF Grant DMS-9022140, by the N. Haar and R. Zinn Research Fund at the Technion, and by the Fund for the Promotion of Research at the Technion. z William Davidson Faculty of Industrial Engineering and Management, Technion-Israel Institute of Technology, 32000 Haifa, Israel. Email address: rothblum@ie.technion.ac.il. Research supported in part by the E. and J. Bishop Research Fund at the Technion and by ONR Grant N00014-92-J1142. Shaped partition problems where the admissible partitions are those - whose shape (j- 1 lies in a prescribed set of admissible shapes. In this article we concentrate on this later situation. The Shaped Partition Problem has applications in diverse fields that include circuit layout, clustering, inventory, splitting, ranking, scheduling and reliability, see [5, 9, 14, 15] and references therein. Further, as we demonstrate later on, the problem has expressive power that captures NP-hard problems such as the Max-Cut problem and the Traveling Salesman problem, even when the number p of parts or attribute dimension d are fixed. Our first goal in this article is to demonstrate, constructively, that a polynomial time algorithm for the Shaped Partition Problem does exist when both p and d are fixed. This result is valid when the set of admissible shapes and the function C are presented by oracles. So, our first result (formally stated and proved in Section 4) is: ffl Theorem 4.2: Any Shaped Partition Problem is solvable in polynomial oracle time using O(n dp 2 operations and queries. Our solution method is based on the observation that since C is convex, the Shaped Partition Problem can be embedded into the problem of maximizing C over the Shaped Partition A defined to be the convex hull of all matrices A - corresponding to partitions of admissible shapes. The class of Shaped Partition Polytopes is very broad and generalizes and unifies classical permutation polytopes such as Birkhoff's polytope and the Permuto- hedron (see e.g. [8, 19, 21]). Its subclass of bounded shaped partition polytopes with lower and upper bounds on the shapes was previously studied in [3], under the assumption that the vectors A are distinct. Therein a polynomial procedure for testing whether a given A - is a vertex of P A was obtained. This procedure is simplified and extended in [11]. A related but different generalization of classical permutation polytopes, arising when algebraic (representation-theoretical) constraints, rather than shape constraints, are imposed on the permutations involved, was studied in [19] and references therein. Since a Shaped Partition Polytope is defined as the convex hull of an implicitly presented set whose size is typically exponential in the input size even when both p and d are fixed, an efficient representation as the convex hull of vertices or as the intersection of half-spaces is not readily expected. Our second objective is to prove that, nevertheless, for fixed p and d, the number of vertices of Shaped Partition Polytopes is polynomially bounded in n, and that it is possible to explicitly enumerate all vertices in polynomial time. So, our second result (formally stated and proved in Section 4) is: ffl Theorem 4.3: Any Shaped Partition Polytope P A has O(n vertices which can be produced in polynomial oracle time using O(n d 2 p 3 operations and queries. An immediate corollary of Theorem 4.3 is that, for fixed d; p, the number of facets of P A is polynomially bounded as well and that all facets can be produced in polynomial oracle time (Corollary 4.4). Theorem 4.3 shows, in particular, that it is possible to compute the number of vertices efficiently. This might be extendible to the situation of variable d and counting vertices is generally a hard task (cf. [16]), as is counting partitions with various prescribed properties (see [4, 10]). The vertex counting problem for variable d and p will be addressed elsewhere. Frank K. Hwang, Shmuel Onn and Uriel G. Rothblum 3 A special role in our development is played by separable partitions, defined as partitions where vectors in distinct sets are (weakly) separable by hyperplanes. In the special case partitions had been studied quite extensively (see e.g. [2, 5, 7, 17]). The case has also been considered quite recently in [6]. Here we study such partitions for all d; p, as well as a class of generic partitions, and provide an upper bound on their number and an algorithm for producing them. In our recent related work [1], the precise extremal asymptotical behavior of such partitions is determined. The embedding of the partition problem into the problem of maximizing the convex function C over the partition polytope is useful due to the optimality of vertices in the latter problem. When consists of a single shape, the optimality of vertices holds for the more general class of Asymmetric Schur convex function, introduced in [13]; see [8]. All of our results apply with C as any asymmetric Schur convex function and consisting of a single shape. The article is organized as follows. In the next section we formally define the Shaped Partition Problem and Shaped Partition Polytope. We demonstrate the expressive power of this problem by giving four examples. For the first two examples, in which the parameters d; p are typically small and fixed, our Theorem 4.2 provides a polynomial time solution. The last two examples show that the Max-Cut problem and Traveling Salesman problem can be modeled as Shaped Partition Problems with fixed respectively, and that the corresponding polytopes have exponentially many vertices. In Section 3 we study separability properties of vertices of Shaped Partition Polytopes and discuss separable and generic partitions. In the final Section 4 we use our preparatory results of Section 3 to establish Theorems 4.2 and 4.3 and Corollary 4.4. Shaped Partition Problems and Polytopes A p-partition of [n] := ng is an ordered collection - sets (possibly empty) whose union is [n]. A p-shape of n is a tuple nonnegative integers n. The shape of a p-partition - is the p-shape of n given by j-j := (j- 1 j). If is a set of p-shapes of n then a -partition is any partition - whose shape j-j is a member of . Let A be a real d \Theta n matrix; for to denote the ith column of A. For each p-partition - of [n] we define the A-matrix of - to be the d \Theta p matrix with We consider the following algorithmic problem. Shaped Partition Problem. Given positive integers d; set of p-shapes of n, and objective function on -partitions given by C convex on IR d\Thetap , find a -partition - which maximizes f , namely satisfies 4 Shaped partition problems Of course, the complexity of this problem depends on the presentation of and C. But, we will construct algorithms that work in strongly polynomial time and can cope with minimal information on and C. Specifically, we assume that the set of admissible p-partitions can be represented by a membership oracle which on query - answers whether or not - 2 . The convex functional C on IR d\Thetap can be presented by an evaluation oracle that on query A - with - a -partition returns Since C is convex, the Shaped Partition Problem can be embedded into the problem of maximizing C over the convex hull of A-matrices of feasible partitions, defined as follows. Shaped Partition Polytope. For a matrix A 2 IR d\Thetan and nonempty set of p-shapes of n we define the Shaped Partition Polytope P A to be the convex hull of all A-matrices of -partitions, that is, A := conv We point out that for any A, the polytope P A is the image of the Shaped Partition Polytope I , with I the n \Theta n identity, under the projection X 7! AX . In [12] this is exploited, for the situation where the function C is linear and is a set of bounded shapes, to solve the corresponding Shaped Partition Problem for all n; d; p in polynomial time by linear programming over P I . We now demonstrate the expressive power of the Shaped Partition Problem. In par- ticular, we show that even if one of d or p is fixed, the Shaped Partition Problem may be NP-hard, and the number of vertices of the Shaped Partition Polytope may be exponen- tial. Therefore, polynomial time algorithms for optimization and vertex enumeration are expected to (and, as we show, do) exist only when both d and p are fixed. We start with two examples in which it is natural to have d; p small and fixed. Example 2.1 Splitting. The n assets of a company are to be split among its p owners as follows. For the jth owner prescribes a nonnegative vector A entries represent the relative values of the various assets to this owner. A partition - which splits the assets among the owners and maximizes the l q -norm ( q of the total value vector whose jth entry A j;i is the total relative value of the assets allocated to the jth owner by -. An alternative interpretation of the splitting problem concerns the division of an estate consisting of n assets among p inheritors having equal rights against the es- tate. With 2, the model captures a problem of a divorcing couple dividing their joint Frank K. Hwang, Shmuel Onn and Uriel G. Rothblum 5 For fixed p, Theorem 4.2 asserts that we can find an optimal partition in polynomial time O(n while the number p n of -partitions is exponential. We note that other (convex) functions C can be used within our framework. In particular, if C is linear on IR p\Thetap when our results of [12] apply and y ield a polynomial time solution even when p is variable. Example 2.2 Balanced Clustering. Given are objects represented by attribute vectors A . The objects are to be grouped in p clusters, each containing points, so as to to minimize the sum of cluster variance of a partition - given by , where jj \Delta jj denotes the l 2 -norm and - A j is the barycenter of the ith cluster. with d Here, we use the fact that . For fixed d; p, by Theorem 4.2 we can find an optimal balanced clustering in polynomial time O(n dp 2 while the number of -partitions is The next two examples show that unless both d; p are fixed, the Shaped Partition Problem may be NP-hard. The idea is simple: the formulation is such that every -partition - gives a distinct vertex A - of the Shaped Partition Polytope P A . Then, any function f on -partitions factors as f(-) := convex C on P A , say the one given by In the following examples, the membership oracle for and the evaluation oracle for f(-) := restricted to A-matrices, are easily polynomial time realizable from the natural data for the problem. Example 2.3 Max-Cut Problem and Unit Cube. Find a cut with maximum number of crossing edges in a given graph E). Here, the A-matrices of -partitions are precisely all (0; 1)-valued n \Theta 2 matrices with each row sum equals 1; in particular, each such matrix is determined by its first column. It follows that the Shaped Partition Polytope P A has 2 n vertices which stand in bijection with -partitions, and is affinely equivalent to the n-dimensional unit cube by projection of matrices onto their first column. So, each A - is a distinct vertex of P A and there is a convex C on IR d\Theta2 such that 6 Shaped partition problems Example 2.4 Travelling Salesman Problem and Permutohedron. Find a shortest Hamiltonian path on n sites under a given symmetric nonnegative matrix D, where D i;j represents the distance between sites i and j. where we regard a partition simply as the corresponding permutation. The matrices A - in this case are simply all permutations of A. The Shaped Partition Polytope P A has n! vertices which stand in bijection with -partitions, and is the so-called Permutohedron. Since each A - is a distinct vertex of P A , there is again a convex C on IR n such that 3 Vertices and generic partitions In this section we show that every vertex of any Shaped Partition Polytope P A equals the A-matrix A - of some A-generic partition, a notion that we introduce and develop below. The convex hull of a subset U in IR d will be denoted conv(U ). Two finite sets U; V of points in IR d are separable if there is a vector h 2 IR d such that h T and with u 6= v; in this case, we refer to h as a separating vector of U and V . The proof of the following characterization of separability is standard and is left to the reader. It implies, in particular, that if U and V are separable then Lemma 3.1 Let U and V be finite sets of IR d . Then U and V are separable if and only if their convex hulls are either disjoint or intersect in a single point which is a common vertex of both. Let A be a given d \Theta n matrix. For a subset S ' [n] let A the set of columns of A indexed by S (with multiple copies of columns identified). A p- partition A-separable if the sets A -r and A -s are separable for each pair that is, for each there is a vector h r;s 2 IR d such that r;s A j for all We have the following lemma which generalizes a result of [3] from matrices with no zero columns and no repeated columns. Lemma 3.2 Let A be a matrix in IR d\Thetan , let be a nonempty set of p-shapes of n, and let - be a -partition. If A - is a vertex of P A then - is an A-separable partition. Proof. The claim being obvious for suppose that p - 2. Let A - be a vertex of A . Then there is a matrix C 2 IR d\Thetap such that the linear functional on IR d\Thetap given by the inner product hC; i is uniquely maximized over P A at A - . Pick any pair 1 . Suppose there are with A i 6= A j (otherwise A -r and A -s are trivially separable). Let - 0 be the -partition obtained from - by switching i and j, i.e. taking - 0 Frank K. Hwang, Shmuel Onn and Uriel G. Rothblum 7 . By choice of C we have hC; A - 0 r;s This proves A -r ,A -s are separable for each pair 1 - r ! s - p, hence - is A-separable. We need some more terminology. Let A 2 IR d\Thetan . A p-partition - [n] is A-disjoint if are disjoint for each pair 1 - r ! s - p. As the convex hulls of finite sets are disjoint if and only if the sets can be strictly separated by a hyperplane, we have that - is A-disjoint if and only if for each there exists a vector h r;s 2 IR d such that (h r;s Of course, A-disjointness implies A-separability, and the two properties coincide when the columns of A are distinct. For the vector obtained by appending a first coordinate 1 to v. For a matrix A 2 IR d\Thetan and indices 1 - sign A (i A matrix A is generic if its columns are in affine general position, that is, if any set of d+ 1 vectors or less from among f - ng are linearly independent; in particular, if ? d this is the case if and only if all signs sign A (i are nonzero. Also, the columns of a generic matrix are distinct. We next provide a representation of the set of A-disjoint 2-partitions for generic matrices A. The case where n - d is simple. Lemma 3.3 Let A 2 IR d\Thetan be generic, p - 2 and n - d. Then every p-partition of [n] is A-disjoint. Proof. It suffices to consider the case 2. A standard result from linear algebra shows that as - A n are linearly independent, the range of [ - a 2-partition - of [n], there is a vector - 2 IR d+1 with - T A i ? 0 for each for each obtained from - by truncating its first coordinate - 1 we then have proving that - is A-disjoint. Let A 2 IR d\Thetan be generic with n - d. For any d-subset I of [n] with Of course, fI \Gamma A g is a 2-partition of [n] n I . Let I ' [n] be a d-set and 2-partition of I . The 2-partitions of [n] associated with A; I and are defined to be either of the two 2-partitions Lemma 3.4 Let A 2 IR d\Thetan be generic with n - d. Then the set of A-disjoint 2-partitions is the set of all 2-partitions associated with A, d-sets I ' [n] and 2-partitions I. 8 Shaped partition problems Proof. We will show that for each d-set I ' [n] and 2-partition I , the two 2-partitions associated with A; I and are A-disjoint and that each A-disjoint 2-partition is generated in this way. First, let I ' [n] have d-elements, say of I . Then H is a hyperplane that contains the columns of A indexed by I ; this hyperplane can be written as fx 2 Thus, h T A A . We next observe that assures that the 2-partition of [d] is B-disjoint. Thus, there exists a vector d 2 IR d with d T A i ? d T A j for all . For sufficiently small positive t we then have that (C proving that A-disjoint. It follows immediately that proving that the two 2-partitions of [n] associated with A; I and 2-partition of I are A-disjoint. Next assume that - is an A-disjoint 2-partition. Then there exists a hyperplane strictly separating A - 1 and A - 2 . Any such hyperplane can be perturbed to a hyperplane that is spanned by d columns of A and weakly separates A - 1 and A - 2 (the details of constructing such a perturbation are left to the reader). In particular, if A span the hyperplane A [ I or A [ I . In the former case we have I and in the latter case and 2. With each list [- 2-partitions of [n] associate a p-tuple - of subsets of [n] as follows: for of the p-tuple associated with the given list are pairwise disjoint. If [ p holds as well then - is a p-partition which will be called the partition associated with the given list. Lemma 3.5 For A 2 IR d\Thetan and p - 2, the set of A-disjoint p-partitions equals the set of p-partitions associated with lists of A-disjoint 2-partitions. Proof. First, consider a p-partition - associated with a list of Then for each so - is A-disjoint. Conversely, let - be an A-disjoint p-partition. Consider any pair 1 are disjoint, there is a hyperplane H r;s which contains no column of A and defines two corresponding half spaces r;s and H r;s that satisfy A -r ae H \Gamma r;s and A -s ae H r;s . Let - r;s := (- r;s 2 ) be the A-disjoint 2-partition defined by - r;s r;s g and - r;s r;s g. Let - 0 be the Frank K. Hwang, Shmuel Onn and Uriel G. Rothblum 9 p-tuple associated with the constructed - r;s 's. Then the sets of - 0 are pairwise disjoint and we have is the p-partition associated with the constructed list of A-disjoint 2-partitions. For each ffl ? 0 define the ffl-perturbation A(ffl) 2 IR d\Thetan of A as follows: for let the ith column of A(ffl) be A(ffl) i := A i is the image of i on the moment curve in IR d . Consider any 1 - n. Then the determinant d is a polynomial of degree d in ffl, with D d being the Van der Monde determinant d ] which is known to be nonzero. So for all sufficiently small ffl ? 0, sign A(ffl) (i the sign of the first nonzero coefficient among d and is either \Gamma1 or 1 and independent of ffl. We define the generic sign of A at (i as the common value of sign A(ffl) (i sufficiently small positive ffl. Lemma 3.6 Let A 2 IR d\Thetan and p - 1. For all sufficiently small ffl ? 0, A(ffl) is generic and the set of A(ffl)-disjoint p-partitions is the same. Further, for every d-set I 2 [n], the sets I \Gamma A(ffl) and I are independent of ffl. Proof. By Lemma 3.5, the set of A(ffl)-disjoint p-partitions is entirely determined by the set of A(ffl)-disjoint 2-partitions. Thus, it suffices to consider only 2. First assume that n ! d. In this case augment A with vectors to obtain a matrix A 0 2 IR d\Theta(d+1) . The above arguments show that for sufficiently small positive ffl, det - A 0 (ffl) is nonzero, implying that - A(ffl) n are linearly independent. From Lemma 3.3 it follows that for such ffl, the set of A(ffl)-disjoint 2-partitions of [n] is the set of all 2-partitions of [n]. Next assume that n ? d. As explained above, for all sufficiently small ffl ? 0, sign A(ffl) the nonzero generic sign -A (i It follows that for all sufficiently small ffl, the matrix A(ffl) is generic and for every d-set I , the sets I \Gamma A(ffl) and I are independent of ffl. By Lemma 3.4, the set of A(ffl)-disjoint 2-partitions is the set of all pairs of 2-partitions of [n] associated with A; d-sets I ' [n] and 2-partitions but each such pair depends only on I \Gamma . Hence is the same for all sufficiently small ffl ? 0. Let A 2 IR d\Thetan . A p-partition of [n] is A-generic if it is A(ffl)-disjoint for all sufficiently A the set of A-generic p-partitions. Lemma 3.6 shows that for all sufficiently small ffl ? 0, the set of A(ffl)-disjoint partitions is the same and equals \Pi p A . The final lemma of this section links vertices of Shaped Partition Polytopes with generic partitions. Shaped partition problems Lemma 3.7 Let A 2 IR d\Thetan and let be a nonempty set of p-shapes of [n]. Then every vertex of the polytope P A has a representation as the A-matrix A - of some A-generic - partition. Proof. Let d\Thetap be a vertex of P A and let C 2 IR d\Thetap be a matrix such that hC; \Deltai is uniquely maximized over P A at B. Let \Pi \Lambdag be the set of -partitions and let \Pi := f- Bg. Then there is a sufficiently small ffl ? 0 such that in addition, as guaranteed by Lemma 3.6, A(ffl) is generic and the set of A(ffl)-disjoint p-partitions equals \Pi p A . For such ffl hC; \Deltai is maximized over the perturbed polytope P A(ffl) at a vertex of the form A(ffl) - for some - 2 \Pi . By Lemma 3.2, - is A(ffl)-separable. Since A(ffl) is generic it has distinct columns and therefore - is also A(ffl)-disjoint. We conclude that - is A-generic, proving that - contains a generic partition. 4 Optimization and Vertex Enumeration We now use the facts established in the previous section to prove our main results. Our computational complexity terminology is fairly standard (cf. [20]). In all our algorithms, the positive integer n will be input in unary representation, whereas all other numerical data such as the matrix A will be input in binary representation. An algorithm is strongly polynomial time if it uses a number of arithmetic operations polynomially bounded in n, and runs in time polynomially bounded in n plus the bit size of all other numerical input. Lemma 4.1 Let d; p be fixed. For any A 2 IR d\Thetan , the set \Pi p A of A-generic p-partitions has Further, there is an algorithm that, given n 2 IN and A 2 l produces A in strongly polynomial time using O(n dp 2 Proof. If n - d, the set of A-generic p-partitions is the set of all partitions, of which there are p n - p d . Henceforth we assume that n ? d. If A := f([n])g consists of the single p-partition ([n]). Suppose now p - 2. For each choice 1 - compute the generic sign -A (i Evaluate the polynomial d at Each evaluation involves the computation of the determinant of a matrix of order d+1 and can be done, say by Gaussian elimination, using O(d 3 ) arithmetic operations and, for rational A, in strongly polynomial time. Then, solve the following linear system of equations d to obtain the indeterminates D . This can be done by inverting the nonsingular Van der Monde matrix of coefficients of this system, again by Gaussian elimination. The generic Frank K. Hwang, Shmuel Onn and Uriel G. Rothblum 11 sign -A (i is then the sign of the first nonzero D i . So, for fixed d the number of operations needed to compute all signs is O By Lemma 3.6, for sufficiently small positive ffl, for each d-set I ' [n], I \Gamma A(ffl) and I are independent of ffl. For a d-set I ' [n] and such ffl, I \Gamma A(ffl) and I are available from the above signs that determine det[ - permutation that puts - into the right location may be applied). Further, from Lemma 3.6 and 3.4, \Pi 2 A equals the common set of A(ffl)-disjoint partitions for sufficiently small positive ffl, and this set is the set of partitions of [n] of the form is a d-subset of [n] and 2-partition of I . For each d-set I ' [n], the common 2-partitions sufficiently small positive ffl has been determined; hence a list of the 2-partitions in \Pi 2 A is available (the construction may contain duplicates). As there are d d-subsets I and 2 d 2-partitions of each I , we have j\Pi 2 d all partitions in \Pi 2 A can be obtained from the generic signs using again O(n d+1 ) operations. For sufficiently small positive ffl, \Pi p A is the common set of A(ffl)-disjoint p-partitions and A is the common set of A(ffl)-disjoint 2-partitions. It follows from Lemma 3.5 that \Pi p A is the set of all p-partitions associated with lists of 2-partitions from \Pi 2 A . This shows that To construct \Pi p A , produce all such lists of 2-partitions from \Pi 2 A ; for each list form the associated p-tuple -, and test if it is a partition (i.e. if [ p [n]). As there are O(n lists, all this work can be easily done using O(n dp 2 operations which subsumes the work for computing the generic signs and constructing \Pi 2 A , and is the claimed bound. We can now provide the solution of the Shaped Partition Problem. The set of admissible p-partitions can be represented by a membership oracle which on query - answers whether or not - 2 . The convex functional C on IR d\Thetap can be presented by an evaluation oracle that on query A - with - a -partition returns The oracle for C will be called M-guaranteed if guaranteed to be a rational number whose absolute value is no larger than M for any -partition -. The algorithm is then strongly polynomial oracle time if it uses a number of arithmetic operations and oracle queries polynomially bounded in n, and runs in time polynomially bounded in n plus the bit size of A and M . Theorem 4.2 For every fixed d; p there is an algorithm that, given n; M 2 IN, A 2 l oracle presented nonempty set of p-shapes of n, and M-guaranteed oracle presented convex functional C on l solves the Shaped Partition Problem in strongly polynomial oracle time using O(n dp 2 operations and oracle queries. Proof. Use the algorithm of Lemma 4.1 to construct the set \Pi p A of A-generic p-partitions in strongly polynomial time using O(n dp 2 test shapes of the partitions in the list to obtain the subset \Pi := f- 2 \Pi p \Lambdag of A-generic - partitions by querying the -oracle on each of the j\Pi p partitions in \Pi p A . Since C is convex, it is maximized over the Shaped Partition Polytope P A at a vertex of P A . Shaped partition problems By Lemma 3.7, this vertex equals the A-matrix A - of some partition in \Pi . Therefore, any - 2 \Pi achieving is an optimal solution to the Shaped Partition Problem. To find such - compute for each - 2 \Pi the matrix , query the C-oracle for the value pick the best. The number of operations involved and queries to the C-oracle is again O(n dp 2 ). The bit size of the numbers manipulated throughout this process is polynomially bounded in the bit size of M and A hence the algorithm is strongly polynomial oracle time. Recall that the Shaped Partition Polytope is defined as P \Lambdag. The number of matrices in the set fA \Lambdag is typically exponential in n, even for fixed d; p. Therefore, although the dimension of P A is bounded by dp, this polytope can potentially have exponentially many vertices and facets as well. But, Lemmas 3.7 and 4.1 yield the following theorem which shows that, in fact, Shaped Partition Polytopes are exceptionally well behaved. Theorem 4.3 Let d; p be fixed. For any A 2 IR d\Thetan and nonempty set of p-shapes of n, the number of vertices of the Shaped Partition Polytope P A is O(n d( p) ). Further, there is an algorithm that, given n 2 IN, A 2 l presented , produces all vertices of A in strongly polynomial oracle time using O(n d 2 p 3 operations and queries. Proof. By Lemma 3.7, each vertex of P A equals the A-matrix A - of some partition in A . Therefore, the number of vertices of P A is bounded above by j\Pi p A j hence, by Lemma 4.1, is O(n To construct the set of vertices given a rational matrix A, proceed as fol- lows. Use the algorithm of Lemma 4.1 to construct the set \Pi p A of A-generic p-partitions in strongly polynomial time using O(n dp 2 operations. Test the shapes of the partitions in the list to obtain its subset \Pi := f- 2 \Pi p \Lambdag of A-generic -partitions by querying the -oracle on each of the j\Pi p partitions in \Pi p A . Construct the set of matrices U := fA - 2 \Pi g with multiple copies identified. This set U is contained in A and by Lemma 3.7 contains the set of vertices of P A . So u 2 U will be a vertex precisely when it is not a convex combination of other elements of U . This could be tested using any linear programming algorithm, but to obtain a strongly polynomial time procedure, we proceed as follows. By Carath'eodory's theorem, u will be a vertex if and only if it is not in the convex hull of any affine basis of U n fug. So, to test if u 2 U is a vertex of P A , compute the affine dimension a of U n fug. For each (a of U n fug, test if it is an affine basis of U n fug, and if it is compute the unique - P a P a in the convex hull of fu if and only a vertex of P A if and only if for each affine basis we get some Computing the affine dimension a, testing if an (a+1)-subset of U nfug is an affine basis and computing the - i , can all be done by Gaussian elimination in strongly polynomial time. Since we have to perform the entire procedure for each of the jU j - j\Pi elements for each such u the number of affine bases of U nfug is at most the number of arithmetic operations involved is O(jU j which absorbs the work for constructing \Pi and obeys the claimed bound. Frank K. Hwang, Shmuel Onn and Uriel G. Rothblum 13 As an immediate corollary of Theorem 4.3, we get the following polynomial bound on the number of facets of any Shaped Partition Polytope and a strongly polynomial oracle time procedure for producing all facets (by which we mean finding, for each facet F , a supporting P A at F ). Corollary 4.4 Let d; p be fixed. For any A 2 IR d\Thetan and nonempty set of p-shapes of n, the number of facets of the Shaped Partition Polytope P A is O(n d 2 p 3 Further, there is an algorithm that, given n 2 IN, A 2 l presented , produces all facets of P A in strongly polynomial oracle time using O(n d 2 p 3 operations and queries. Proof. By the well known Upper Bound Theorem [18], the number of facets of any k-dimensional polytope with m vertices is O(m k Applying this to P A with k - dp and get the bound on the number of facets of P A . To construct the facets, construct first the set V of vertices using the algorithm of Theorem 4.3. Compute the dimension a of aff(P (possibly empty) set S of dp \Gamma a points that together with V affinely span IR d\Thetap . For each affinely independent a-subset T of V , compute the hyperplane fX 2 spanned by S [ T . This hyperplane supports a facet of P A if and only if all points in V lie on one of its closed half-spaces. Clearly, all facets of P A are obtained that way, in strongly polynomial time and number of arithmetic operations and oracle queries bounded as claimed. Acknowledgment Shmuel Onn thanks the Mathematical Sciences Research Institute at Berkeley for its support while part of this research was done. --R Discrete Applied Mathematics A simple on-line randomized incremental algorithm for computing higher order (Voronoi) diagrams Optimal partitions having disjoint convex and conic hulls Mathematics of Operations Research Consecutive optimizers for a partitioning problem with applications to optimal inventory groupings for joint replenish- ment Cutting dense point sets in half The Art of Counting (edited by Joel Spencer) Partition polytopes over 1- dimensional points The Pareto set of the partition bargaining game Journal of Combinatorial Theory Ser. Representations and characterizations of the vertices of bounded-shape partition polytopes Linear programming over partitions Partitions: Clustering and Optimality SIAM Journal on Algebraic and Discrete Mathematics Discrete and Computational Geometry. On the number of halving lines The maximum numbers of faces of a convex polytope. Theory of Linear and Integer Programming Lectures in Polytopes --TR --CTR Uli Wagner, On the number of corner cuts, Advances in Applied Mathematics, v.29 n.2, p.152-161, August 2002 Sharon Aviran , Nissan Lev-Tov , Shmuel Onn , Uriel G. Rothblum, Vertex characterization of partition polytopes of bipartitions and of planar point sets, Discrete Applied Mathematics, v.124 n.1-3, p.1-15, 15 December 2002 Babson , Shmuel Onn , Rekha Thomas, The Hilbert zonotope and a polynomial time algorithm for universal Grbner bases, Advances in Applied Mathematics, v.30 n.3, p.529-544, April F. K. Hwang , J. S. Lee , Y. C. Liu , U. G. Rothblum, Sortability of vector partitions, Discrete Mathematics, v.263 n.1-3, p.129-142, 28 February
polynomial time;optimization;polytope;convex;programming;enumeration;partition;separation;cluster
589281
Cut Size Statistics of Graph Bisection Heuristics.
We investigate the statistical properties of cut sizes generated by heuristic algorithms which solve the graph bisection problem approximately. On an ensemble of sparse random graphs, we find empirically that the distribution of the cut sizes found by "local" algorithms becomes peaked as the number of vertices in the graphs becomes large. Evidence is given that this distribution tends toward a Gaussian whose mean and variance scales linearly with the number of vertices of the graphs. Given the distribution of cut sizes associated with each heuristic, we provide a ranking procedure that takes into account both the quality of the solutions and the speed of the algorithms. This procedure is demonstrated for a selection of local graph bisection heuristics.
Introduction . Algorithms for tackling combinatorial optimization problems [27] may be divided into two classes. Exact algorithms such as exhaustive search, branch-and-bound, or branch-and-cut, form the first class; they determine (exactly) the optimum of the cost function which is to be minimized. However, for NP-hard problems, they require large computation ressources, and in particular, large computation times. The second class consists of "heuristic" algorithms; these are not guaranteed to find the optimal (lowest cost) solution, nor even a solution very close to the optimum, but in practice they find good approximate solutions very fast. For problems in science, one's main interest is in the optimal solution, so an exact algorithm is required. However, for many engineering applications, the heuristic approach may be preferable. There are several reasons for this: (i) The computational ressources are simply insufficient to solve the instances of interest by exact methods; (ii) The cost function one wants to minimize is computationally very demanding, and limited resources force one to use an approximate cost function instead. This is the rule rather than the exception with very complex systems such as VLSI. If the true cost function cannot be used, there is little point in finding the true optimum for the wrong problem. (iii) Heuristic algorithms typically generate numerous "good enough" solutions, thus providing information about the statistical properties of low cost solutions. This information can in turn be used for generating better heuristics, or for finding new criteria for guiding the branching in exact algorithms such as branch-and-bound. For almost any combinatorial optimization problem, it is very easy to devise heuristic algorithms which perform quite well; this is probably why so many such algorithms have been proposed to date. Usually they fall into just a few families, the most popular of which are local search, simulated annealing, tabu search, and evolutionary computation. The practitioner is frequently confronted with the problem of choosing which method to use. Thus he would like to rank these algorithms and determine which one is best for his "instance" (the set of parameters which completely specify the cost function). A difficulty then arises because most heuristic algorithms are stochastic, so that they can give many different solutions for a single instance. In general, the distributions of solution costs generated by the different heuristics overlap, so that the winning algorithm varies from one trial to another. Furthermore, it is necessary to balance the quality of the solutions found against the time necessary to find them since in practice heuristics run at very different speeds. The final goal of this paper is to do just this kind of balancing: in Section 8 we shall introduce a generally applicable ranking method which is based on the possibility of performing multiple runs from random starts for each algorithm until an allotted amount of computer time is exhausted. Our ranking method then determines whether it is better to have a fast heuristic which gives not so good solutions or a slower heuristic which can give better solutions. Establishing a ranking on a single instance may be what is needed for a real world problem, but it is not a useful prediction tool. It is preferable to consider the effectiveness of a heuristic when it is applied to a family of instances. Since a detailed knowledge of the distribution of costs is necessary for our ranking procedure, the major part of this paper is an in depth study of the statistics of costs found by several classes of heuristics. The NP-hard [9] combinatorial optimization problem chosen for our study is the graph bisection problem, hereafter simply called the graph partitioning problem (GPP). This choice is justified by the wide range of practical applications of the GPP. These include host scheduling [3], memory paging and program segmentation [17], load balancing [21], and numerous aspects of VLSI-design such as logic partitioning [12] and placement [6, 19]. Because of these applications, the GPP has been used as a testing ground for many heuristics. For our work, a selection had to be made; in view of the previous studies by Johnson et al. [13], Lang and Rao [20], and Berry and Goldberg [4], we have restricted our study to iterative improvement heuristics based on local search and to simulated annealing. Having made a choice of optimization problem and algorithms, it remains to define the class of instances for the testbeds. Ideally, this family of instances should reflect the structure of the actual instances of interest to the practitioner. Since we do not have a particular application in mind, we shall follow the studies of [13, 20, 4], and consider an ensemble of sparse random graphs. From our numerical study, we have found that all of the heuristics tested share the following properties when the random graphs become large: (i) each algorithm can be characterized by a fixed percentage excess above the optimum cost; (ii) the partitions generated have a distribution of costs which becomes peaked, both within a given graph and across all graphs; (iii) these distributions tend towards Gaussians. Because of these properties, our ranking of heuristics on large graphs is largely determined by the mean and variance of the costs found, and thus a constant speed-up factor has only a very small effect on the ranking. We expect this property to hold for most problems and heuristics of practical interest, leading to a very robust ranking. The paper is organized as follows. In Section 2 we define the GPP as well as the ensemble of random graphs used for our testbed. Section 3 derives properties of random partitions, and shows that the Cut Size Statistics of Graph Bisection Heuristics 3 distribution of cut sizes has a relative width which goes to zero as the instance size grows. In Section 4 we argue why this property should hold also for the distribution of costs found by heuristic algorithms based on local iterative processes. In Section 5 we discuss the heuristic algorithms we have included in our tests. Section 6 gives the mean and standard deviation of the costs found as a function of graph size; the distribution for the costs is indeed found to be peaked. This leads to a first ranking which, however, does not take into account computation times. To implement our speed-dependent ranking, we must determine the distribution of cut sizes found by the different algorithms. This is the subject of Section 7, where evidence is given that the distribution on any typical graph tends towards a Gaussian in the limit of large graphs. In Section 8 we present our ranking method which takes into account both the quality of the solutions as well as the speed of the heuristics. In Section 9, finally, we discuss the results and conclude. 2. Minimum cuts. The graph partitioning (or graph "bisection") problem (GPP) can be defined as follows. Consider a graph E) which consists of a set of N vertices and a set of (non-oriented) edges connecting pairs of vertices. It is convenient to introduce the called the connectivity matrix, given by is connected to v j Since the edges are non-oriented, . (Some of what will be discussed applies to weighted graphs; represent the weight of the ij edge.) A partition of G is given by dividing the vertices of G into two disjoint subsets V 1 and V 2 such that . The number of edges connecting V 1 to V 2 is called the cut of the partition, and will be denoted by C. It is given by The GPP (or "Min-cut" problem) consists of finding the partition which the cost (2.1) is minimum subject to given constraints on the sizes of V 1 and V 2 . The GPP is NP-hard [9]. In the standard formulation to which we shall restrict ourselves in this work, V 1 and V 2 have equal sizes. For our study, it is necessary to fix an ensemble of graphs for the testbed. We have chosen G(N; p) the ensemble of random graphs of N vertices where each edge is present with probability p. The choice of G(N; p) is justified by its tractable mathematical properties and by the fact that many workers [13, 20, 4] have used graphs in this ensemble to test heuristics. The problem of finding the properties of the minimum cut size when the graphs belong to such an ensemble is sometimes called the stochastic GPP. Let us review some of the known results for this problem; this will serve to motivate our conjectures for the behavior of cuts obtained from heuristics. For each graph G i , call C 0 its minimum cut size. Taking G i from the ensemble G(N; p), C 0 is a random variable. Following derivations now standard in a number of other stochastic combinatorial optimization problems (COP), it is possible to show using Azuma's inequality [1] that the distribution of C 0 becomes peaked as N ! 1. This means that as N becomes large, , the relative fluctuations about the mean, tend to zero. This property, often referred to as "self-averaging", is typical of processes to which many terms contribute. For certain stochastic COP, it is possible to show further that the mean minimum cost satisfies a power scaling law in N , so that C 0 =N fl converges in probability to a limiting value as N ! 1. In the case of the stochastic GPP, there is no proof that such property hold. Nevertheless, it is believed that such a scaling holds: within the G(N; p) ensemble at p fixed, calculations show that C 0 =N 2 ! p=4 with probability one as [8]. As will be shown in the next section, this is also the limiting behavior of random cuts, and so the ensemble at p fixed is not a challenging one for heuristics. The reason for this "uninteresting" scaling is the high number of edges connecting to any vertex. Thus we consider in this work the ensemble G(N; p), ff is the mean connectivity (number of neighbors of a vertex) of the graphs. These graphs are sparse, in contrast to the dense graphs obtained by taking p to be independent of N . Consider the optimal partition. At a typical vertex in V 1 , some finite fraction of its edges will connect to vertices in V 2 . With each vertex contributing an O(1) amount to the cut size, C 0 is expected to grow linearly with N . Since C 0 =N is known to be peaked at large N , it is natural to conjecture the stronger property that C 0 =N tends towards a constant with probability one as N ! 1. A major motivation for this work is our expectation that an identical scaling law should hold if we replace C 0 by the cost found 4 G.R. Schreiber and O.C. Martin by a heuristic algorithm, albeit that the limiting constant depends on the heuristic. To motivate such a property, the next section analyzes the cut sizes of random partitions; then in Section 4 we consider the "statistical physics" of the GPP so as to interpolate between the case of minimum cuts and that of random cuts. 3. Cuts of random partitions. Here we show explicitly that a large N scaling law holds for the cut sizes of random partitions, and that asymptotically these random cuts have a Gaussian distribution with a relative variance proportional to 1=N . Consider any graph in G(N; p). One can always write the cut size of a random partition as where X is the mean (random) cut size for the graph under consideration, and hY is the average over the random partitions.) Averaging explicitly over all balanced partitions of the fixed graph, we find 1)]. The interpretation of this formula is very simple: any edge of weight E ij has a probability being cut. In the ensemble G(N; p) of random graphs, it is easy to calculate the first few moments of X . In particular, we find denotes the average over the ensemble G(N; p).) We also see that X is the sum of independent random this implies that the kth cumulant (connected moment) of the distribution of X statisfies c c At large N , we then have c in the constant p ensemble, and c - ffN in the p - ff=N ensemble. The random variable Y is more subtle as it is the sum of M correlated variables. Nevertheless, for any graph, it is possible to compute the moments of Y , and we have done this explicitly for the second and third moments. (The expressions are too long to be given here.) If we average Y 2 both over random partitions and over G(N; p), we obtain: The calculations get significantly more complicated for the higher moments. In order to keep to simple expressions, we limit ourselves to the ensemble with 1). Then we find: Furthermore, the graph to graph fluctuations of hY 2 i become negligible in relative magnitude, so that the ratio of a typical variance to the mean variance goes to 1 at large N . This however is not true for the higher moments; for instance, we find that the typical value of hY 3 i grows as N 1=2 , but taking in addition the mean over graphs leads to a N independent behavior. Finally, one can show that hY k i c =hY 2 with probability one. This shows that as N ! 1, Y has a Gaussian distribution, of zero mean, and of variance growing linearly with N , whose coefficient is graph independent. Coming back to the cut size of a random partition, we find that the normalized correlation coefficients between powers of X and Y tend to zero at large N , and thus X and Y become independent random variables in that limit. This, along with the results previously derived, shows that at large N , C itself has a Gaussian distribution. From these results, we deduce the large N behavior: so that relative deviations from the mean go to zero. Thus the distribution of C becomes peaked, and probability one as N ! 1. The convergence of the distribution of C=N to a "delta" function is referred to as the self-averaging of C. Cut Size Statistics of Graph Bisection Heuristics 5 The scaling of the variances can be summarized at large N by writing Y y where x and y are independent Gaussian random variables of zero mean and unit variance; oe is the standard deviation (rescaled by 1= N) of X , and oe ff=8 that of Y . Thus oe Y describes the fluctuations of the cut sizes within a graph, and oe X describes the fluctuations of the mean cut size from graph to graph. We have used these analytical results to test the validity of our computer programs. The first two moments of X allowed us to test our generation of random graphs in G(N; p). Similarly, a check on our random number generator was obtained by verifying on several graphs that the second moment of Y found by the numerics was in agreement with our formulae. Finally, we also checked that random cut sizes have a limiting Gaussian distribution, with a third moment which scales to zero at large N . (For this check, we performed random partitions on 100 000 graphs for 4. Statistical physics of the GPP. We saw that cut sizes of random partitions in G(N; p) have a self-averaging property; we conjectured that this property also holds for the minimum cut. It is possible to interpolate between these two kinds of partitions (random and min-cut) by following the formalizm of statistical physics. For any given graph, consider the "Boltzmann" probability distribution pB , defined for an arbitrary partition P of cut size C(P e \GammaC(P )=T Z Z is chosen so that pB is normalized (a probability distribution) and T is an arbitrary positive parameter called the temperature. When T !1, we recover the ensemble of random partitions where all partitions are equally probable, while when T ! 0, the ensemble reduces to the partitions of minimum cut size. For intermediate values of the temperature, the partitions are weighted according to an exponential of their cut size. In this "Boltzmann" ensemble, one can define the moments of the cut sizes just as was done in the case of random partitions. In most statistical physics problems, it is possible to show that the quantity in the exponential of Eq. (4.1) (here, the cut size) is self-averaging. For random graphs, however, the proofs are inapplicable; nevertheless, other evidence indicates that the cut size is self-averaging at any temperature [26]. This self-averaging can be understood qualitatively at low temperature as follows. The number N (C) of partitions of cut size C is a sharply increasing function of C, whereas the Boltzmann factor is a sharply decreasing function of C. Note that the probability distribution P (C) of C is given by the product of these two functions. Using naive but standard statistical physics arguments for N (C), one finds that P (C) has a peak at C (T ) which grows linearly with N and that the width of the distribution is O( N ), which gives the self-averaging property for C. In addition, this kind of argument says that becomes Gaussian at large N , a result which is usually correct in statistical physics systems. A number of statistical physics results have been obtained for the GPP in the ensemble of dense random graphs, i.e., for G(N; p) at p fixed. In particular, highly technical calculations [26, 8] indicate that the cut sizes are self-averaging at all temperatures, that is as N ! 1, relative fluctuations within a fixed graph become negligible, as well as those from graph to graph. The mean cut size is given by as N ! 1. (If the mean over graphs is not performed, the formula remains valid for "almost all" sequences of graphs with N !1.) In this equation, U(T ) is a function of temperature only, there is no dependence on p as long as p is independent of N . The limit T ! 0 gives the expected (and typical) value of the minimum cut, with 0:3816:. Although there is no proof yet that these calculations are exact, there is general agreement in the statistical physics community that the results are correct. The case of sparse random graphs (p - 1=N) has also been studied within the statistical physics approach [2, 5]. So far, however, the problem has proven to be intractable with no plausible solution in sight. Nevertheless, it is expected that the cut sizes are self-averaging at any temperature and that the mean of the distribution scales linearly with N at large N . 6 G.R. Schreiber and O.C. Martin The property of self-averaging seems quite generic. The reason it should hold in these systems is that the cut size of a partition is the sum of a large number of random variables which are not too correlated. It is very plausible that the cut size is self-averaging whenever partitions are generated by an iterative process involving just a few vertices at a time. All local search methods, and modifications thereof such as simulated annealing, fall into this category. Thus our claim is that any heuristic algorithm which generates partitions iteratively according to local (in vertex space) criteria will lead to cut sizes which are self-averaging. Thus the distribution of cut sizes found by any such heuristic should become peaked as N ! 1. Furthermore, in this limit, the distribution should converge towards a Gaussian in the way given by the central limit theorem. We will see in the sections to follow that this is indeed born out empirically for all of the heuristics which we have investigated. The arguments we have presented are not specific to the graph partitioning problem, so we expect them to apply to most stochastic COPs having many variables in their cost function. Surprisingly, there has been very little research on this topic. In the context of the "NK" model with binary variables, a study by Kauffman and Levin [16] found that the costs of local minima became peaked towards the value of a random cost as N grew. (This peculiar property is due to the structure of the energy landscape in that model.) However, concerning the behavior of heuristic solutions, research has almost exclusively focused on the case of the Euclidean traveling salesman problem where points are laid out on the plane. Most practitioners in that field know that local search heuristics give rise to costs whose relative variance decreases as the number of points increases. Furthermore, it was observed by Johnson and McGeoch [14] among others that the costs tend towards a fixed percentage excess above the optimum. Our purpose here is to show how this convergence occurs, albeit in a different combinatorial optimization problem, and to provide a theoretical framework for understanding where this behavior comes from. Also, we pay special attention to the distinction between fluctuations within an instance and from one instance to another. We believe our findings are quite general, and in particular that the ensemble of instances considered need not be based on points in a physical space. 5. Algorithms used in the testbed. In view of the previous arguments, we have restricted ourselves to local heuristics. Without trying to be complete nor representative, we have studied the statistics of cut sizes for three types of local search and four versions of simulated annealing algorithms. In this section we sketch the workings of these heuristics. In Sections 6 and 7, we show that the same self-averaging properties hold for all these algorithms in spite of their significant differences. There is thus no reason to believe that our claims are affected by the details of such algorithms; rather, the properties are most likely generic to dynamics which are local. 5.1. Kernighan-Lin (KL). In simple local search, one performs elementary transformations to a feasible solution of the COP as long as they decrease the cost, a procedure sometimes called -opting [22]. A more sophisticated version consists in using "variable depth" search: one builds a sequence of elementary transformations, usually according to a greedy criterion. p is not set ahead of time, and depends on the sequence of costs found. The elementary transformations are not imposed to decrease the cost, but the sequence of length p must do so if it is to be applied to the current solution. Such a procedure was first proposed by Kernighan and Lin [18], in fact in the framework of the GPP. Hereafter we will refer to their algorithm as "KL". The elementary transformation they use is the exchange of a pair of vertices, one vertex in V 1 being exchanged for one in V 2 . A sequence of such exchanges is built up in a greedy and tabu fashion by performing a "sweep" of all the vertices: at each step of the sweep, one finds the best (largest cost gain) pair to exchange among those vertices which have not yet been moved in the sweep (tabu condition). The sweep has length N=2. When the sweep is finished one finds the position along the sequence of exchanges generated where the cut size is minimum. If this minimum leads to an improved partition, the transformation of p exchanges is performed on the partition and another sweep is initiated; otherwise the search is stopped and the partition is "KL-opt", i.e., it is a local minimum under KL. The KL algorithm is deterministic although it is possible to introduce stochasticity to break degeneracies in selecting the best pair to exchange. Its computational complexity is not easy to estimate because the number of sweeps is not known in advance. (This is a generic difficulty in estimating the speed of iterative improvement heuristics.) However, in practice, one finds that KL finishes in a "small" number of sweeps. Thus the computational complexity is estimated to be a few times that of performing the last sweep, known as the check-out sweep. For our study, we have used our own implementation of KL [24], which uses heaps to find the best pair to exchange at each step. For sparse graphs, this leads to O(N operations per sweep. A nearly identical KL is provided in the Chaco software package, which Cut Size Statistics of Graph Bisection Heuristics 7 gives sensibly identical results. A faster implementation of the algorithm has been given by Fiduccia and Mattheyses [7] whenever the use of a radix sort is possible; then the time for each sweep is O(N ). In terms of quality of solutions found, KL is quite good. What is surprising is that although Kernighan and Lin proposed their method over 20 years ago, KL remains relatively unchallenged, at least as a general purpose method applicable to any kind of graph, regardless of its structure. Of course, for special kinds of graphs, such as meshes, other heuristics (e.g., spectral bisection) perform better [4, 11, 13, 15]. 5.2. A multilevel KL-algorithm: CHACO. The Chaco software package includes a number of heuristics for partitioning graphs. (For information about this package, see the Chaco user's guide [10].) For our purposes, we have used only its "multilevel" generalization of KL, hereafter referred to simply as CHACO. The CHACO algorithm is based on a coarse graining or "compactification" of the graph to be partitioned. At each level, vertices are paired using a matching algorithm, and paired vertices are then considered as the vertices of the next higher level of compactification. Because of this process, it is necessary to have weighted edges; the weights are also propagated to the higher level. The compactification is repeated until a sufficiently small graph is obtained to which spectral bisection is applied to get a first partition. Then this partition is used as the starting partition in KL for the graph at the level below it. This process is recursive, until one obtains a KL-opt partition of the original graph. (Note that this construction is deterministic, and does not require an initial "random" partition.) Such a multilevel strategy has been very successful for unstructured 2 and 3 dimensional meshes [11, 15], both in terms of solution quality (much better than for KL alone), and in terms of speed (much faster than KL because of the hierarchical nature). However, the usefulness of CHACO on random graphs is not a priori obvious, both in terms of speed and quality of solutions. 5.3. Simulated Annealing algorithms. We have choosen as a third comparative algorithm simulated annealing (SA). SA is based on a set of elementary moves, just like local search, but now moves which increase the cost are accepted with (low) probability. Because of this, it is sometimes appropriate to consider SA as a noisy local search method. Simulated annealing is really a family of algorithms. To include some of the different bells and whistles proposed for this algorithm, we have considered four variations. These are: (i) the SA as first introduced by Kirkpatrick et al. [19] (referred to as FSA) where the initial and final temperatures are fixed ahead of time by the user and where a predetermined number of trial moves are performed at each temperature; (ii) Kirkpatrick et al. also proposed to determine the initial and final temperatures of the schedule dynamically. They set the initial temperature at the beginning of the run using the criterion that about 80% of the trial moves are accepted at that tempera- ture. Similarly, they stop the cooling if for 5 cooling steps the energy does not decrease. We will refer to this method as KSA. (iii) Johnson et al. [13] improved the speed of this algorithm by allowing an early exit to the next temperature of the schedule; the condition they proposed for exiting is having accepted a minimum number of moves. Also they modified the termination criterion to having an acceptance rate less than a threshold value. We will refer to this version as JSA. All three of these SA methods use an exponential cooling schedule with a cooling factor of 0:95. (iv) The last SA variation consists in using an adaptive schedule whereby the next temperature value is determined on the fly according to the energy fluctuations at the current temperature. We have choosen for this variation the implementation of van Laarhoven and Aarts [28, 29]. To obtain good results one would have to spend a long time in the "freezing" phase of the cooling. Since this would increase the computation times significantly we have choosen not to use a fine-tuned adaptive schedule but one which provides a cooling factor of the same magnitude as in the other SA algorithms presented. This allows us to have similar computation times for all the simulated annealing algorithms investigated. In SA, one can use the same elementary moves as in local search, i.e., for the GPP, pair exchanges. However, once a low cost partition is obtained, it will take a long time (or a lot of luck) to find further good exchanges. Finding a good pair is best done by finding the first vertex to transfer and then the second, i.e., by using a sequential process. This suggests relaxing the constraint of having balanced partitions, and replacing it by a penalty function which keeps the sizes of V 1 and V 2 nearly equal (small off-balance). We have followed a slightly different approach where each move destroying the balance must be followed by a move restoring the balance. Then the Markov chain explores the partitions which are balanced and those with "off-balance" of \Sigma1. It is easy to see that this method is equivalent to having the cost of all the other partitions equal infinity; at fixed temperature and for long chains, one generates partitions with cut sizes given by the Boltzmann factor, within the constraint for the "off-balance". Indeed, the succession of accept/reject decisions makes the global probability distribution Boltzmannian in this enlarged space, so that we guarantee the same convergence properties as in the standard case. 8 G.R. Schreiber and O.C. Martin Some remarks concerning our implementations are in order. First, at fixed temperature, we perform a certain number of "sweeps". In each sweep, every vertex is sequentially considered as a candidate for changing sides of the partition; if the move were to violate our limit on the "off-balance", the move is rejected (in fact, it simply is not considered). A sweep thus requires O(N) operations. Our sweeps use random permutations rather than a fixed or random ordering of the vertices. The use of random permutations should - according to certain authors [13, 28, 29] - result in a enhancement of the quality of the solutions found. Second, the maximum number of sweeps at any temperature is set to ff-, with all of our implementations. For FSA and KSA, this is in fact the (actual) number of sweeps, so that their computational complexity is O(ff-N) times the number of temperature steps used. The cases of JSA and ASA are more difficult to evaluate. In practice we find that JSA is faster than KSA, but not by more than a constant factor. ASA on the other hand spends quite a lot of time at intermediate temperatures, all the more so that N increases; empirically, we have found an O(N 3=2 ) complexity. In terms of quality, we are aware of no systematic study on sparse random graphs. In a previous SA work on the GPP, Van Laarhoven and Aarts used an adaptive decrement rule [28, 29] and claim a gain of about 13% over simpler non-adaptive algorithms. They also compared their results to those from the algorithm used by Johnson et al. for the GPP, who claimed an enhancement of about 5% for JSA over the Kernighan-Lin algorithm. The small gain found by Johnson [13] is, according to van Laarhoven and Aarts [28, 29], due to the use of a non-adaptive choice of the temperature decrement rule. However, we have found for sparse random graphs that the different variants of simulated annealing are nearly indistinguishable in terms of quality of solutions. This may be due to our not using a penalty term or to the different nature of the graphs used in the present study. 5.4. Chained-Local-Optimization (CLO). The chained-local-optimization (CLO) strategy is a synthesis of local search and of simulated annealing [25]. The essential idea is to have simulated annealing sample not all solutions, but only locally optimal solutions. This strategy is guaranteed to be at least as good as local search, and has been successfully applied to the traveling salesman problem [23] and to the partitioning of unstructured meshes [24]. In this work, we use KL as the local search engine. Given any initial KL-opt partition P i , the simplest implementation of CLO will: (i) apply a perturbation or "kick" to modify significantly the partition (in practice this means exchanging clusters of vertices); (ii) run KL on the modified partition so as to reach a new KL-opt partitionP f ; (iii) apply the accept/reject procedure for going from the initial partition to the final one (P f ). This defines the analogue of one move of a simulated annealing algorithm, except that many modifications to the partition have occured in this single step. The temperature may be modified according to a schedule if desired, but for simplicity, we have set the temperature to zero in all of our runs. As was discussed in the context of simulated annealing, it is inefficient to exchange vertices or clusters simultaneously, it is better to do it sequentially. Our present CLO algorithm thus proceeds as follows. Given P i an initial balanced KL-opt partition, choose a (connected) cluster of p vertices in V 1 (or and move them into V 2 (respectively V 1 ). KL-optimize this partition to generate an intermediate (off- balanced) partition. Now choose a cluster of p vertices in V 2 this modified partition to generate P f , the final (and balanced) partition. This whole procedure is our "simulated annealing" step, and we apply the accept/reject criterion for going from P i to P f . When runing CLO on irregular meshes [24], it was possible to perform large kicks, exchanging many vertices at once. Unfortunately, for sparse random graphs, we find that the acceptance when doing so becomes low. We have thus used "small" kicks, creating clusters of sizes varying randomly between 3 and 13. Given such small kicks, KL usually terminates in just 2 sweeps, and the speed of CLO per kick is about half that of KL. Consider now the limit of large N . Using the analogy with simulated annealing, if a fixed (N - independent) number of small kicks are used, it can be expected that CLO will perform no better than KL itself. We have thus chosen to use a number of kicks which scales linearly in N , namely -N with This choice of course influences the quality of the solutions generated, a larger value of - giving a priori better results. The computational complexity of this algorithm is then of order N 2 log(N ). 6. Self-averaging of the cut size. In the rest of this paper, we study the statistical properties of the cut sizes generated by the algorithms described in Section 5 when applied to random initial partitions. The ensemble of graphs used is that of random graphs with mean connectivity Section 2). This value was chosen because at much larger connectivities, the ratio between the best and worst cut size approaches 1, and at lower connectivities, algorithms taking explicit advantage of Cut Size Statistics of Graph Bisection Heuristics 9 disconnected parts of the graph will outperform general purpose heuristics. In order to minimize effects associated with our finite sample of graphs in the ensemble, we have benchmarked all the algorithms on the same graphs. The number of graphs used during the production runs was 10 000 with values of N ranging between 50 and 200; however, because the CHACO algorithm was fast, we have also performed runs on 100 000 graphs for that heuristic. The purpose of this section is to give numerical evidence that the distribution of cut sizes becomes peaked in the limit of large graphs, for each of the heuristics considered. (Further properties of the distribution will be given in Section 7.) We find that each algorithm generates cut sizes for which both the mean and variance scale linearly in N . From this behavior, it is clear that the distribution of cut sizes becomes peaked at large N , i.e., that the cut sizes are self-averaging. Also, assuming (cf. Section 2) that the minimum (i.e., optimum) cut size scales linearly with N at large N , we then see that each heuristic algorithm leads to a fixed percentage excess above the true optimum. (Note that the worst cut size also has a linear scaling in N .) This percentage excess provides a first ranking of the algorithms, which, however, does not take into account the speed of execution. If C(i; m) is the cut obtained by a heuristic for the graph G i and an initial partition m, define the mean cut per vertex by: m) where the averages are over initial partitions and over the ensemble of graphs studied (cf. Section 3 for the notation). We compute these ensemble averages numerically using the standard estimator (hereafter, overlines refer to numerical averages): The approximation is due to a statistical error e associated with fluctuations of C(i; m) both with m and i. It is not difficult to see that for our problem, one does not need to perform an average over m; using any finite number R of partitions for each graph G i provides an unbiassed estimator of . Furthermore, the statistical error e is not very sensitive to R, making it numerically inefficient to take a large value for R. Because of this, we have performed the numerical averages with this leads to a simple expression for e, the statistical error on c: ci Figure 6.1 shows the dependence of c on 1=N . (The error bars are too small to be visible. Also, in order to avoid cluttering the figure, we have included among the simulated annealing algorithms only KSA; the other implementations of simulated annealing give nearly identical results.) For all algorithms, the figure suggests that there is a limiting large N value for c and that the convergence to this limit is linear in 1=N . We have thus fitted the data to a linear function: The values of the A and B coefficients obtained from the fits are given in Table 6.1, and the - 2 values show that the fits are good. An identical analysis can be performed on the variance of the cuts found by the different algorithms. Figure 6.2 shows the dependence on N for the rescaled quantity The scaling in N is apparent, just as it was for c. In summary, our data lead us to conclude that the mean and variance of C scale linearly with N at large N . Then the relative width of the distribution of C is proportional to 1= showing that the distribution for the cut sizes becomes peaked for all the algorithms investigated. (One can also say that the distribution of C(i; m)=N tends towards a delta function as N !1, which is what we mean by self- averaging.) Since the fluctuations of C(i; m) include both graph to graph fluctuations and fluctuations G.R. Schreiber and O.C. Martin CHACO KSA KL CLO Fig. 6.1. Scaled mean cut sizes for the different algorithms. algorithm A B % excess KSA 0.4485 4.95 0.00 ASA 0.4499 4.96 0.32 JSA 0.4513 4.88 0.63 CLO 0.4568 4.85 1.8 CHACO 0.4802 5.81 7.1 KL Table Estimates for the large N value and slope of the mean cut size per vertex and percentage excess relative to the KSA heuristic. within a graph, we can conclude that the relative fluctuations within a fixed typical graph necessarily also go to zero. (N.B.: although for our runs we use our observable is an unbiassed estimator for ci which includes both types of fluctuations.) Thus in the large N limit, each algorithm will give a fixed percentage excess above the minimum for almost all graphs and almost all random initial partitions. A speed independent ranking. Since each algorithm is characterized by a percentage excess, we can introduce a ranking of the different heuristics according to their excess in the large N limit. (Of course, this ranking does not take into account the speed of the algorithms!) For our graphs and our implementation of the different heuristics, the winners are in the class of simulated annealing. The best is KSA; using this as the reference rather than the true min cut size (which is unknown), JSA has an excess of 0:63%, ASA an excess of 0:32%, and FSA an excess of 0:08%. The next best heuristic is the CLO-algorithm, followed by CHACO, and finally KL. (The results for the excesses are given in Table 6.1.) We have also included for general interest the excess obtained by a zero temperature "simulated annealing": 18:21%; note that it gives much less good results than KL, while true simulated annealing gives much better results than KL. As a comment, let us remark that the relative solution quality of the algorithms is determined to higher precision than the absolute quality. Simply put, the cut sizes we obtain for the different algorithms are correlated because they are performed on the same graphs, so that the statistical error on for instance is 3:2 times smaller than the statistical error on alone. This is why it is possible to give reliable values for the excesses of the different simulated annealing algorithms even though their solution quality is very similar. Nevertheless the ranking for the simulated annealing algorithms is not without ambiguity. The FSA algorithm is, for larger N , within the statistical error of the KSA algorithm, and hence we have no strong evidence that one is better than the other. The other algorithms are easily ranked. KL and CHACO are 9:6% and 7:1% worse than KSA, but CLO is only 1:8% worse. The comparison with KL is qualitatively (though not quantitatively) similar Cut Size Statistics of Graph Bisection Heuristics 11 KSA KL CLO CHACO Fig. 6.2. Scaled variance of the cut sizes for the different algorithms. to that given by Johnson et al. [13] and by van Laarhoven and Aarts [29]. Both claimed a gain of the SA-algorithm over the KL-algorithm of about 5% and 13%, respectively. The differences with our results have several origins. First, we have performed an average over an ensemble of graphs. Second, our graphs have slightly different characteristics from the ones they use. Third we have not introduced a penalty term in our implementation of simulated annealing; this probably affects the quality of the solutions found. 7. Distribution of cut sizes. In this section we deepen our statistical study of C. As shown in the previous section, the distribution of C=N tends towards a delta function; it is natural to ask how this limit is reached, and to understand the nature of intra- and inter-graph fluctuations. It is convenient to use the framework introduced in Section 3 but where random partitions are replaced by the partitions found by applying one of our heuristics to a random start. For each graph G i , and each initial partition m, we define where hY (i; so that X(i) is the average cut size found on graph G i , and Y (i; m) gives the fluctuation of the cut size about its mean for that graph. For each of our heuristics, our study indicates that for a large random graph G i , Y has a nearly Gaussian distribution, and that the width of this distribution is essentially independent of i. We study this distribution at large N and show that its width is self-averaging and that its relative asymmetry goes to zero. Finally, we have evidence that X and Y become independent variables at large N . These properties will lead to a fast and robust ranking of the heuristics in Section 8. Figure 7.1 shows the distribution of cut sizes found by KL on one graph chosen at random from G(N; p) with 1). Superposed is a Gaussian with the same mean and variance. The figure gives good evidence that the distribution of Y for that graph is very close to a Gaussian. Then an obvious question is whether the distribution of Y is similar across different graphs. For each of our heuristics, we find that the answer is yes, as indicated by the following study of the moments of Y . (Note that for the CHACO algorithm, the default parameter setting generates the initial starting partition deterministically by application of the coarse graining strategy, then a spectral method is applied. Since there is no "random" initial partition, there are no fluctuations in the cut size as a function of m and so little in this section applies to CHACO with these parameter settings.) To quantify how oe 2 m)i varies from graph to graph, we measured its mean and variance over i. First, we measured the ensemble averages Y (i) =N . For each heuristic, the data extrapolates to a limiting value as N becomes large. Comparing with the results for the mean cut size, we find that the algorithms which lead to the best cut sizes also have the smallest widths for the Y distribution. Second, we studied the variance of oe 2 Y (i), i.e., oe 2 Y (i) \Delta . This study requires high statistics, and so was performed to high accuracy only for KL, the fastest of our algorithms; however the other algorithms show qualitatively the same behavior. Figure 7.2 displays for KL the 1=N dependence of the relative variance of oe 2 Y (i), i.e., the inter-graph variance of oe 2 Y (i) divided by the square of its mean. As can be seen from the figure, the ratio goes to zero at large N , showing that oe 2 Y (i) is self-averaging. Simply put, G.R. Schreiber and O.C. Martin frequency of cut sizes Fig. 7.1. Histogram of KL cut sizes for one graph with overlaid Gaussian. this means that the width (over m) of the Y distribution has relative fluctuations from graph to graph which dissapear as N ! 1. (Our lower statistics data for the other heuristics are consistent with this Y Fig. 7.2. Relative variance of the intra-graph cut size variance oe 2 Y Following the statistical physics analogy given in Section 4, there is reason to believe that the distribution of Y tends towards a Gaussian as in the case of random partitions. To test this conjecture, we have measured the asymmetry of the distribution of Y on numerous graphs for KL. First, we find that the typical asymmetry is small, and that the mean of the third moment of Y satisfies Y (i) as N !1. Second, we have checked that the average of the squared asymmetry is also small, i.e., Y (i) 0: These properties give strong evidence that the distribution of Y for any graph tends towards a Gaussian of zero mean and of variance AN as N ! 1, where A depends on the heurisitic but not on the actual graph. The distribution of X(i) can be studied similarly. The previous section gave its mean as a function of N and also showed that it is self-averaging. It is of interest to quantify the decrease with N of its relative variance. We have found that the distribution of X is roughly compatible with a Gaussian distribution of width proportional to N for each of the algorithms. (Unfortunately, a quantitive test of this requires very high statistics.) However, the distribution of X(i) is not essential for our ranking procedure as will be clear in the next section, so we have not studied it in greater depth. Cut Size Statistics of Graph Bisection Heuristics 13 Finally, to completely specify the statistics of C(i; m), it is necessary to describe the correlations between X(i) and Y (i; m). We have found numerically that these variables are nearly uncorrelated, with in particular the correlation between X(i) and oe 2 Y (i) tending towards zero as N ! 1. Assuming that this holds and that X has a Gaussian distribution, then the distribution of C(i; m) is also Gaussian. Our measurement of the asymetry (jointly over i and m) of C(i; m) is compatible with this property at large N . (The total variance is then given by the sum of the variances of X and Y .) This can be summarized mathematically by introducing two Gaussian random variables x and y of zero mean and unit variance, and modeling the rescaled cut size as the following sum: m) - oe oe Y This equation is then the exact analogue of what was derived for the cut sizes of random partitions (see Eq. 3.5). 8. A speed dependent ranking of heuristics. In this section we come back to the initial motivation for this work, namely the necessity of comparing heuristics of very different speeds. The possibility of doing so is very relevant, as for most combinatorial optimization problems local search is quite fast and simulated annealing notoriously slow. Any meaningful ranking must determine whether it is better to have a fast heuristic which gives not so good solutions, or a slower heuristic giving better solutions. We now show how to introduce such a ranking when considering first just one graph, and then generalize to an ensemble of graphs. Finally, we illustrate what this ranking gives in the case of the heuristics in our testbed when applied to sparse random graphs. The case of one graph. Consider a single graph G on which one is to provide a ranking of a number of heuristics which give various cut sizes and run at different speeds. To take into account both the speed of the algorithms and the quality of the solutions they generate, we fix the amount of computation time allotted per algorithm. Call this time - (measured for instance in CPU seconds on a given machine). Each heuristic then generates (non-optimal) solutions during that time using multiple random initial starts. Suppose that the speed of the algorithm of interest is such that k independent starts can be performed in the allotted time - . (We shall assume that the execution time is insensitive to the random initial start, as this is the case in practice with our heuristics. Knowledge of the speed of the algorithm then gives the value of k which can be used.) For each start, there is an output or "best-found" cost. The output at the end of the k starts is the best of these k costs, hereafter called "best-of-k". The different algorithms are then ranked on the basis of the ensemble mean of their "best-of-k" (the value of k depending on - and on the algorithm). This ensemble average is the average over the random numbers used both for the random initial starts and for running the algorithms (if any). This establishes a ranking for a particular graph and for a given amount of computation time - . It is inefficient to perform the average just mentioned in a "direct" way, i.e., by extracting values of "best-of-k" over many multiple runs; it is far better to compute the average starting with the distribution of the "best-found" cut sizes associated with single random starts. Call P (C) the probability of finding a "best-found" cut size of value C, and Q(C) the associated cumulative distribution, i.e., the probability of finding a cut size (strictly) smaller than C. Since the cut sizes are integer valued, we then have Q(C). Introducing the analogous probabilities ~ for the "best-of-k" values, one has: The distribution for "best-of-k" can thus be generated from that of "best-found", and then C , the mean of "best-of-k", is easily extracted. (This construction explains why we studied the distribution of single cut sizes in Section 7.) Note also that it is possible to extract C for a whole range of - values with essentially no extra work since - affects only k and the determination of the mean of "best-of-k" represents a negligible amount of work once the distribution of "best-found" is known. The quantity C is in effect a quantitative measure of the effectiveness of the algorithm. Of course, C depends on the amount of computation ressources allotted, i.e. As - increases, k increases (in jumps of unity), and C decreases. The broader the distribution of "best-found", the faster the decrease of C and the more useful it is to perform multiple runs. To establish the ranking, simply order the algorithms according to their C . In general, this ranking may depend on - , and clearly it is sensitive to the lower tail of the distribution of "best-found". Let us 14 G.R. Schreiber and O.C. Martin illustrate this by considering for instance two heuristics H 1 and H 2 having two overlapping distributions for "best-found", with averages satisfying hC H1 In the mean, H 1 seems better than H 2 , but if H 2 is significantly faster, and if the tail of its distribution extends well into the domain of CH1 , then one can have C H1 . H 2 may then be the more effective algorithm, assuming of course that - is large enough so that indeed H 2 can be run multiple times. Some general properties may be derived assuming for instance that CH1 and CH2 are described by the same distribution but are shifted with respect to one another. Then if the tail of the distribution falls off as an exponential or faster, H 2 will not become more effective than H 1 as - !1. Ranking on an ensemble of graphs. The extension of this ranking to an ensemble of graphs is straight-forward. Assume that C is known for each graph G and for each heuristic. C is a (real number) measure of the effectiveness of the heuristic on that graph, given an amount of computation time - . We can then generalize this measure from one graph to an ensemble of graphs by considering , the mean of C over the relevant ensemble. The final ranking is then simply given by the ordering of the algorithms according to their mean effectiveness. Our expectation is that in a relatively homogeneous ensemble, the effectiveness (and thus the ranking) will be nearly the same for essentially all sufficiently large graphs and so the average behavior is also the typical behavior. We can expect this to happen whenever the distribution of cut sizes associated with the different heuristics do not overlap too much and have the same pattern regardless of the graph. This is what occurs in the case of our ensemble of random graphs: indeed, we saw that each algorithm leads to a fixed percentage excess cost at large N and that the distribution of costs is peaked. Then two algorithms have non overlapping distributions as give rise to the same percentage excess). It is then clear that at large N , the mean ranking is the same as the typical ranking. It is also clear that increasing the amount of computer resources (- and thus speeding up an algorithm while keeping the quality of its solutions the same does very little to improve its ranking. Illustration. For each value of N and - , we can follow the procedure just given to obtain C for the different heuristics of interest for any given graph G, and repeat this for many graphs in G(N; p). There are, however, a number of possible speed-ups in our case because of the statistical properties derived in the previous sections. First, although in principle the "best-of-k" construction has to be repeated for each graph, the results of Section 7 provide a short-cut. Since the distribution for "best-found" is (to high accuracy) Gaussian, it is possible to map the mean of "best-found" to that of "best-of-k" once and for all: the mapping is just a shift by a k-dependent number of standard deviations. Second, noting that at fixed N , the variance of this Gaussian as well as the speed of the algorithm is essentially constant from graph to graph, we can calculate (the average over graphs) in terms of: (i) the CPU time necessary to find one "best-found"; (ii) the mean cut size, (iii) the variance of the intra-graph cut sizes, m)i, which is graph independent at large N . These quantities were measured for a number of values of N , and then fits were performed to interpolate to arbitrary values of N . From these fits, it is possible to compute analytically the values of for any values of N and - , and in particular the "winning" algorithm (the first in our ranking). From this, define regions in (N ,- ) space where a given heuristic is the winner, leading to a "diagram" as in Figure 8.1. In our construction of this diagram, we have included JSA in our ranking but not FSA, KSA, nor ASA. This is because for our choice of parameters, all of the simulated annealing algorithms tested give very similar quality solutions, but JSA is slightly faster. Although the effectiveness of all these SA algorithms are nearly identical, their ranking depends on N and - because of the discrete jumps in k. (Whenever one algorithm increases its k before the others, it may change its ranking.) In the diagram of Figure 8.1, we have labeled the different regions according to the associated "winner", and have indicated the boundaries separating them. (Again, because of the discrete nature of k, we have smoothed these curves.) The labeling "SA" in fact corresponds to JSA. The CPU time is expressed in multiples of CPU- cycles. To give these units a machine independent and less technical meaning, it is enough to say that the lower boundary of the CHACO region corresponds to the time CHACO needs to run once. From this diagram, we see that at large N , given enough CPU time, the best algorithm is simulated annealing, simply because its mean excess cost is lower than that of the other algorithms. In this limit, the distributions for the cut sizes overlap very little, so the ranking is relatively insensitive to the algorithm's speed: using multiple random starts does very little to improve the quality of the solutions found as fluctuations about the mean become negligible. At smaller values of N , the fluctuations arising from Cut Size Statistics of Graph Bisection Heuristics 150 CLO SACHACO KL0 CPU TIME5000Fig. 8.1. Ranking diagram different random starts are not negligible, so faster algorithms can outperform simulated annealing by using the best of k runs. If we compare KL, CHACO, and CLO, we see that CLO is a bit slower but leads to substantially better solutions, and so is the winner if the amount of CPU time is enough for it to run. The other algorithms are competitive only if neither CLO nor simulated annealing can terminate a run. This explains why the KL region is nearly invisible, squeezed under the CHACO region, itself below the CLO and SA region. (Note: (i) on our random graphs, CHACO is slower than KL; (ii) the initial partition is set deterministically within the default settings of CHACO, so that its "best-found" and "best-of-k" values are identical.) 9. Discussion and Conclusions. We have studied the statistics of cut sizes generated by graph partitioning heuristics, both within a given graph and over an ensemble of graphs. Motivated by a statistical physics analogy and by what happens for random partitions (Section 3), we obtained strong numerical evidence that the cut sizes generated on sparse random graphs are self-averaging, i.e., that their distribution becomes peaked as the number of vertices N becomes large. (Quantitatively, this simply means that the relative fluctuations about the mean tend tend to zero as N ! 1.) For the mean cut size, we found a linear dependence on N , indicating that each heuristic leads to a fixed percentage excess cut size above the true minimum. We expect analogous properties to hold for all local heuristics applied to any combinatorial optimization problem in which each variable is coupled to just a few others. We also investigated how the distribution of cut sizes approaches its limiting large N behavior, and gave evidence that on typical graphs the distribution of cut sizes generated becomes Gaussian as N !1. In that limit, each heuristic is then characterized by a mean cut size (over all graphs) and a variance describing the fluctuations in the cut sizes on any typical graph. This variance seems to scale linearly with N in the large N limit and to be self-averaging also. The principal motivation for this work was to introduce a method to rank heuristics while taking into account both the quality of the solutions found and the speed of the algorithms. Knowledge of the distribution of cut sizes allows one to establish a meaningful ranking of the heuristics by assuming that the algorithms may be applied to k different random starts, with the best of the k runs giving the final cost. Although this ranking can be done by brute force, we have used the properties just described to demonstrate it on the heuristics in our testbed. At "large" values of N (N ? 700), the winner is almost always simulated annealing. In fact, at large N , the distributions associated with the algorithms we have tested do not overlap significantly, so that the use of multiple runs to explore the tail of the distributions is not effective. For smaller values of N , the faster algorithms are more competitive, and we find that the winner is CLO except when the allotted time is too short for running even one run of CLO. Since the graph to graph fluctuations in the variance of the cut sizes found are small, this ranking "in the mean" is also in almost all cases the ranking on individual graphs; it is thus very robust. G.R. Schreiber and O.C. Martin A number of questions remain open. How can one characterize the distribution of X(i), the mean cut size on graph i? To what extent do similar properties hold for heuristics which are manifestly not local? Can the information found help generate better heuristics? Concerning this last question, it is worth pointing out that although simulated annealing is a general purpose method, it outperforms the other heuristics which were specifically developped for the graph partitioning problem. This suggests that some improvements in these methods might be obtainable by suitable modifications. 10. Acknowledgement . We are indebted to Bruce Hendrickson and Robert Leland for providing us with their software package Chaco 2.0. We also thanks S. W. Otto and N. Sourlas for stimulating discussions. G.R.S. acknowledges support from an Individual EC research grant under contract number ERBCHBICT941665, and O.C.M. acknowledges support from the Institut Universitaire de France. Fur- thermore, G.R.S. would like to express his gratitude to Professor J.M. G'omez G'omez for his generous hospitality at the Department of Theoretical Physics of the Universidad Complutense de Madrid, where part of this work was accomplished. --R Weighted sums of certain dependent random variables Graph bipartitioning and statistical mechanics A partitioning strategy for non-uniform problems on multiprocessors Path optimization for graph partitioning problems. Replica symmetry breaking in finite connectivity systems: a large connectivity expansion at finite and zero temperature A procedure for placement of standard-cell VLSI circuits A linear-time heuristic for improving network partitions Application of statistical mechanics to NP-complete problems in combinatorial optimization Computers and Intractability: A Guide to the Theory of NP-Completeness The Chaco user's guide: Version 2.0 Partitioning of vlsi circuits and systems Optimization by simulated annealing: An experimental evaluation The traveling salesman problem: A case study in local optimization A fast and high quality multilevel scheme for partitioning irregular graphs Towards a general theory of adaptive walks on rugged landscapes Some Graph Partitioning Problems Related to Program Segmentation An efficient heuristic procedure for partitioning graphs Optimization by Simulated Annealing an empirical evaluation An empirical study of static load balancing algorithms Computer solutions of the traveling salesman problem Partitioning of unstructured meshes for load balancing Glass Theory and Beyond Algorithms and Complexity A general approach to combinatorial optimization problems --TR --CTR Andrew E. Caldwell , Igor L. Markov, Toward CAD-IP Reuse: A Web Bookshelf of Fundamental Algorithms, IEEE Design & Test, v.19 n.3, p.72-81, May 2002 Angel , Vassilis Zissimopoulos, On the Hardness of the Quadratic Assignment Problem with Metaheuristics, Journal of Heuristics, v.8 n.4, p.399-414, July 2002 Andrew E. Caldwell , Andrew B. Kahng , Andrew A. Kennings , Igor L. Markov, Hypergraph partitioning for VLSI CAD: methodology for heuristic development, experimentation and reporting, Proceedings of the 36th ACM/IEEE conference on Design automation, p.349-354, June 21-25, 1999, New Orleans, Louisiana, United States Andrew E. Caldwell , Andrew B. Kahng , Igor L. Markov, Design and implementation of move-based heuristics for VLSI hypergraph partitioning, Journal of Experimental Algorithmics (JEA), 5, p.5-es, 2000 Olivier C. Martin , Rmi Monasson , Riccardo Zecchina, Statistical mechanics methods and phase transitions in optimizationproblems, Theoretical Computer Science, v.265 n.1_2, p.3-67, 08/28/2001
graph partitioning;heuristics;ranking;self-averaging
589287
On the Local Convergence of a Predictor-Corrector Method for Semidefinite Programming.
We study the local convergence of a predictor-corrector algorithm for semidefinite programming problems based on the Monteiro--Zhang unified direction whose polynomial convergence was recently established by Monteiro. Under strict complementarity and nondegeneracy assumptions superlinear convergence with Q-order 1.5 is proved if the scaling matrices in the corrector step have bounded condition number. A version of the predictor-corrector algorithm enjoys quadratic convergence if the scaling matrices in both predictor and corrector steps have bounded condition numbers. The latter results apply in particular to algorithms using the Alizadeh--Haeberly--Overton (AHO) direction since there the scaling matrix is the identity matrix.
Introduction The study of superlinear convergence of interior-point methods for linear programming (LP) was initiated in the early 90s in an effort to explain the fact that interior point methods tend to perform significantly better in practice than indicated by the polynomial complexity bounds. This discrepancy is due to the limitation of the worst case analysis used in deriving polynomial complexity bounds and reflects the inherent conflict between the requirements of global convergence and fast local convergence. Superlinear convergence is especially important for semidefinite programming (SDP) since no finite termination schemes exist for such problems. As predicted by theory and confirmed by numerical experiments the condition number of the linear systems defining the search directions increases as 1=-, where - is the normalized duality gap, so that the respective systems become very ill conditioned as we approach the solution. Therefore an interior point method that is not superlinearly convergent is unlikely to obtain high accuracy in practice in spite of its theoretical "polyno- mial complexity". On the other hand a superlinearly convergent interior point method will achieve good accuracy (e.g. 10 \Gamma10 or better) in substantially fewer iterations than indicated by its worse case global linear convergence rate that is related to polynomial complexity. The local convergence analysis for interior point algorithms for SDP is much more challenging than those for LP as shown by a relatively smaller number of papers addressing this subject. The first two papers investigating superlinear convergence of interior point algorithms were written independently by Kojima, Shida and Shindoh [4] and by Potra and Sheng [13]. The algorithm investigated in these papers is an extension of Mizuno-Todd- Ye predictor-corrector algorithm for LP and uses the KSH/HRVW/M search direction (see the next section for a definition of this search direction). Kojima, Shida and Shindoh [4] established the superlinear convergence under the following three assumptions: (A) SDP has a strictly complementary solution; nondegenerate in the sense that the Jacobian matrix of its KKT system is (C) the iterates converge tangentially to the central path in the sense that the size of the neighborhood containing the iterates must approach zero, namely, lim Here k:k F denotes the Frobenius norm of a matrix and "ffl" denotes the corresponding scalar product (see the next section for precise definitions). In [13] we have not used assumptions (B) and (C). Instead we proposed a sufficient condition for superlinear convergence that is implied by the above assumptions. In [14] we improved this result and obtained superlinear convergence under assumption (A) and the following condition: (D) lim which is clearly weaker than (C). Of course both (C) and (D) can be enforced by the al- gorithm, but the practical efficiency of such an approach is questionable. However, from a theoretical point of view it is proved in [14] that the modified algorithm in [4] that uses several corrector steps in order to enforce (C) has polynomial complexity and is superlinearly convergent under assumption (A) only. It is well known that assumption (A) is necessary for superlinear convergence of standard interior point methods even in the QP case (see [10]). Kojima, Shida and Shindoh [4] also gave an example suggesting that interior point algorithms for SDP based on the KSH/HRVW/M search direction are unlikely to be superlinearly convergent without imposing a condition like (C). In [5] the same authors showed that a predictor-corrector algorithm using the AHO direction is quadratically convergent under assumptions (A) and (B) (see the next section for a definition of the AHO search direction). They also proved that the algorithm is globally convergent but no polynomial complexity bounds have been found for this algorithm. It is shown that condition (C) is automatically satisfied by the iteration sequence generated by the algorithm. It appears that the use of the AHO direction in the corrector step has a strong effect on centering. We exploited this property in [15] where we showed that a direct extension of Mizuno-Todd-Ye algorithm, based on the KSH/HRVW/M direction in the predictor step and the AHO direction in the corrector step, has polynomial complexity and is superlinearly convergent with Q-order 1:5 under assumptions (A) and (B). An interesting superlinearly convergent predictor-corrector algorithm based on the NT search direction was proposed by Luo, Sturm and Zhang [7]. The algorithm depends on a parameter ffl ? 0. It produces points (X is defined in (2.7), ffl=4. The algorithm starts from a feasible point (X and for any given ~ ffl - ffl=4 finds a feasible point (X in at most O( iterations. However this bound on the number of iterations is not proved to hold for hence the algorithm is not polynomial in the usual sense. The algorithm is superlinearly convergent under assumption (A). It turns out that (C) is enforced by the algorithm since it is proved in [7] that for sufficiently large k It is also proved that if one uses one predictor and r correctors per iteration, then - k converges to zero with Q-order 2=(1 In this paper we investigate the local behavior of the predictor-corrector algorithm considered by Monteiro [9] for SDP using the MZ-family of search directions. We show that the sufficient condition of Potra and Sheng [13] for superlinear convergence applies for this algorithm. The sufficient condition is independent of scaling matrices. In particular we show that the algorithm is superlinearly convergent if (A) and (D) are satisfied. More specifically, we show that under the assumptions (A) and (B), superlinear convergence with Q-order 1.5 is obtained if the scaling matrices in the corrector step have bounded condition num- ber. Finally, we propose a new version of the predictor-corrector algorithm which enjoys quadratic convergence if the scaling matrices in both predictor and corrector steps have bounded condition numbers and (A) and (B) are satisfied. The following notation and terminology are used throughout the paper: the p-dimensional Euclidean space; nonnegative orthant of IR the positive orthant of IR the set of all p \Theta q matrices with real entries; the set of all p \Theta p symmetric matrices; : the set of all p \Theta p symmetric positive semidefinite matrices; : the set of all p \Theta p symmetric positive matrices; the (i; j)-th entry of a matrix M; Tr(M the trace of a p \Theta p matrix, equals 0: M is positive semidefinite; 0: M is positive definite; n: the eigenvalues of M 2 S the largest, smallest, eigenvalue of M 2 S Euclidean norm of a vector and the corresponding norm of a matrix, i.e., Frobenius norm of a matrix; k(G; G; 2 The predictor-corrector algorithm for SDP We consider the semidefinite programming (SDP) problem: and its associated dual problem: are given data, and are the primal and dual variables, respectively. By G ffl H we denote the trace of (G T H). Also, for simplicity we assume that A i are linearly independent. Throughout this paper we assume that both (2.1) and (2.2) have finite solutions and their optimal values are equal. Under this assumption, X and (y ; S ) are solutions of (2.1) and (2.2) if and only if they are solutions of the following nonlinear system: We denote the feasible set of the problem (2.3) by and its solution set by F , i.e., We consider the symmetrization operator [17] Since, as observed by Zhang [17], for any nonsingular matrix P , any matrix M with real spectrum, and any - 2 IR, it follows that for any given nonsingular matrix P , (2.3) is equivalent to A perturbed Newton method applied to the system (2.4) leads to the following linear system: m \Theta S n is the unknown search direction, - 2 [0; 1] is the centering parameter, and is the normalized duality gap corresponding to (X; The search direction obtained through (2.5) is called the Monteiro-Zhang (MZ) unified direction [17, 11]. The matrix P used in (2.5) is called the scaling matrix for the search direction. It is well known that taking I results in the Alizadeh-Haeberly-Overton (AHO) search direction [1], corresponds to the Kojima-Shindoh-Hara/Helmberg- Rendl-Vanderbei-Wolkowicz/Monteiro (KSH/HRVW/M) search direction [6, 3, 8], and the case of P T coincides with the Nesterov-Todd (NT) search direction [12]. Monteiro and Zhang [11] established the polynomiality of a long-step path-following method based on search directions defined by scaling matrices belonging to the class such that Following [11], Sheng et al. [16] proved the polynomiality of a Mizuno-Todd-Ye type predictor-corrector algorithm for SDP by imposing the scaling matrices to be chosen from the class n\Thetan is nonsingular and PXSP Moreover, its superlinear convergence was proved under an addtional simple condition. The primal-dual algorithms considered by Monteiro [9] are based on the centrality measure \Theta S n 1), we denote by N (fl) the following neighborhood of the central path: Monteiro's generalized predictor-corrector algorithm for semidefinite programming based on the MZ family of directions consists of a predictor step and a corrector step at each iteration. Starting from a strictly feasible pair (X generates a sequence of iterates in N (ff). An iteration of Monteiro's generalized predictor-corrector algorithm can be described as follows. Algorithm Given choose nonsingular n \Theta n matrices P k and P k ffl Predictor Step. Solve the system (2.5) with (X; Denote the solution (U; m \Theta S n , and set Compute the step length ffl Corrector Step. Solve the system (2.5) with (X; and be the solution, and set End of iteration. Using an elegant analysis, Monteiro [9] proved that the predictor-corrector algorithm defined above with properly chosen parameters ff and fi (0 well defined and that it needs at most O( iterations for producing a pair (X is the initial gap. More precisely, Monteiro showed that for all k - 0. 3 Technical results In analyzing the local behavior of the predictor-corrector algorithm of Monteiro, we need the following technical result proved in [8, Lemma 2.6] and [9, Lemma 2.1(b)]. Lemma 3.1 Suppose that M 2 IR p\Thetap is a nonsingular matrix and E 2 IR p\Thetap has at least one real eigenvalue. Then, The following lemma is part of Lemma 3.5 of Monteiro [9]. Lemma 3.2 Let W 2 IR n\Thetan be such that GWG \Gamma1 is skew-symmetric for some nonsingular n\Thetan . Then, The following technical result will play an important role in our analysis. Lemma 3.3 Let (X; 1). Suppose that (D x ; \Deltay; D s n\Thetan \Theta IR m \Theta S n\Thetan is a solution of the linear system: \Deltay for some K 2 IR n\Thetan . Then we have F , where Proof. By denoting we can write and It is easily seen that - and Using the notation it follows that Using (3.8) and Lemma 3.2 with On the other hand, using (3.8) again, we obtain \GammakX \Gamma1=2 D x X \Gamma1=2 (X 1=2 SX s s s which implies (i). Then (ii) follows from (i), (3.9), and the fact that It is interesting to note that the inequalities in the above lemma are independent of the scaling matrix P . In the next lemma we establish a lower bound for the stepsize ' k , which together with Lemma 3.3 enables us to analyze the asymptotic behavior of the predictor-corrector algorithm. be generated by the predictor-corrector algorithm. Then where Proof. For simplicity, let us omit the index k. By (2.8), we have which together with the linearity of H P (\Delta), the fact that T r[H P (M and (2.5a) with imply that Using the fact that U ffl Therefore, and Hence, we have X(') - 0 and S(') - 0 for all ' 2 [0; - ']. Otherwise, there exists a ' 0 2 [0; - such that X(' 0 )S(' 0 ) is singular, which means On the other hand, (3.2) with implies that which contradicts (3.10). Using (3.4) with Therefore, '. The result follows from the definition of '. 4 A sufficient condition for superlinear convergence In this section we will investigate the asymptotic behavior of the predictor-corrector algorithm and obtain a sufficient condition for superlinear convergence. Definition 4.1 A triple (X ; y is called a strictly complementary solution of Throughout the paper we assume that the following condition holds. Assumption 1. The SDP problem has a strictly complementary solution (X be an orthogonal matrix such that q are eigenvectors of X and S , and define It is easily seen that IB [ ng. For simplicity, let us assume that where B and N are diagonal matrices. Here and in the sequel, if we write a matrix M in the block form then we assume that the dimensions of M 11 and M 22 are jIBj \Theta jIBj and jINj \Theta jINj, respectively. In the next lemma we use the following notation: Lemma 4.2 (Potra-Sheng [13, Lemma 4.4]) Under Assumption 1 we have ks ks Using Lemma 4.3, we can write O( O( O( Using the same techniques, we obtain a similar result for the predicted pair (X k Lemma 4.3 Let X Assumption 1 is satisfied, then we have O( O( O( As in [13], let us define a linear manifold: It is easily seen that if (X Lemma 4.4 (Potra-Sheng [13, Lemma 4.5]) Under Assumption 1, F ae M. Lemma 4.5 (Potra-Sheng [13, Lemma 4.6]) Under Assumption 1, every accumulation point of strictly complementary solution of (2.3). Let us define is the solution of the following minimization problem: and \Gamma is a constant such that k(X k ; S k )k F - \Gamma; 8k. Note that every accumulation point of belongs to the feasible set of the above minimization problem and the feasible set is bounded. Therefore ( - exists for each k. Theorem 4.6 Under Assumption 1, if then the predictor-corrector algorithm is superlinearly convergent. Moreover, if there exists a constant oe ? 0 such that then the convergence has Q-order at least 1+oe in the sense that - Proof. For simplicity, let us omit the index k. It is easily seen that (U Here we have used the relation - clearly satisfies the equation Denoting and applying (i) of Lemma 3.3, we obtain which implies Similarly, By Lemma 4.3 and the fact that ( - In a similar manner we obtain Let us observe that Then from (4.6), (4.7), (4.8), (4.9) and (4.10), we get Hence, Applying (ii) of Lemma 3.3, we obtain Noting that we deduce Finally, if k ) for some constant oe ? 0, then we have k ). From Lemma 3.4, it follows that Therefore, Lemma 4.6 was originally obtained by Potra and Sheng [14]. Based on Lemma 4.6, we establish the following generalization of the result of Potra and Sheng [14, Theorem 6.1]. Theorem 4.7 Under Assumption 1, if X k S !1, then the predictor-corrector algorithm is superlinearly convergent. Moreover, if X k S constant oe ? 0, then the convergence has Q-order at least 1 0:5g. 5 Superlinear convergence under strict complementarity and nondegeneracy Throughout this section, we will assume that Assumption 1 (strict complementarity) holds. Let (X ; y ; S ) be a strictly complementary solution of (2.1) and (2.2). We will also assume the following nondegeneracy condition introduced by Kojima, Shida and Shindoh [4, 5]. First, let us define an affine space G 0 by Assumption 2. (Nondegeneracy) If X As remarked in Section 5 of Kojima, Shida and Shindoh [5], under the strict complementarity assumption, the above nondegeneracy condition is equivalent to the combination of primal and dual nondegeneracy conditions given by Alizadeh, Haeberly and Overton [2]. Under Assumptions 1 and 2, the solution (X ; S ) is unique. Therefore the iteration sequence converges to (X ; S ) and so does the sequence of predicted pairs Lemma 5.1 (Kojima-Shida-Shindoh [5], Lemma 5.3) Assume that H I (US +X V Let R be a nonsingular matrix and ~ (R It is easily seen that the R-scaled SDP ~ ~ also satisfies the strict complementarity and nondegeneracy conditions. Its unique solution is (RX R T (R Using Lemma 5.1 and considering the new SDP (5.1), we can easily obtain the following lemma. Lemma 5.2 Assume that for some nonsingular matrix R, In the next lemma cond F denotes the condition number of a matrix B. Lemma 5.3 If cond F (P k Proof. Let R the corrector step of the algorithm, we have U it is easily seen that Suppose (5.2) is not true, i.e., the sequence is unbounded. Then we can choose a subsequence such that (R and Obviously, (U . The fact that the matrices A are linearly indepen- dent, together with (U implies that (U Dividing both sides of (5.3) by letting k !1 along a subsequence, we obtain which contradicts Lemma 5.2. Theorem 5.4 Under the strict complementarity and nondegeneracy assumptions, if cond F (P k O(1), then the algorithm is superlinearly convergent with Q-order at least 1.5. Proof. At the predictor step, we have Thus, Then, by Lemma 5.3, we obtain Note that F F Therefore, which ends the proof by invoking Theorem 4.7. The above result says that the superlinear convergence of the predictor-corrector algorithm is independent of the choice of the scaling matrix P k in the predictor step of the algorithm, while the scaling matrices used in the corrector step need to be "well-conditioned" for superlinear convergence. Clearly, the family of scaling matrices admissible in the corrector step for superlinear convergence includes the identity matrix defining the AHO as a special case. By imposing the same assumption on the scaling matrices used in the predictor step and a new strategy for the step size, we can improve the order of convergence stated in Theorem 5.4. In order to achieve quadratic convergence we need to slightly modify the choice of the step size. Instead of ' k given by (2.9), we will use: The predictor-corrector algorithm with this new strategy will be called the modified predictor-corrector algorithm. It is easily seen that the modified predictor-corrector algorithm still has polynomial complexity. In what follows we will show that it is also quadratically convergent. Theorem 5.5 Under the hypothesis of Theorem 5.4, if cond F (P k then the modified predictor-corrector algorithm is quadratically convergent. Proof. From the proof of Theorem 5.4 (cf. (5.5)), we have Using (5.7) and an argument similar to that employed in the proof of Lemma 5.3 we get Then we can write g. As in (5.4)-(5.5), we can prove that Using (5.9) and the same argument as in Lemma 5.3, we get Observing that we have where C 2 is a positive constant. Without loss of generality, we may assume which, together with (5.6) and (5.13), implies that and Let Evidently, k ], we have This means Therefore and - 6 Remarks In this paper we only consider the feasible version of the predictor-corrector method to keep the presentation simple. However, the analysis used here can be easily extended to the infeasible predictor-corrector algorithms based on the unified direction proposed by Monteiro and Zhang. Under the strict complementarity and nondegeneracy assumptions we have established the superlinear convergence with Q-order 1.5 of the "pure" predictor-corrector algorithm if the scaling matrices for the corrector step satisfy cond F (P k superlinear convergence can be obtained under a weaker condition is an interesting topic for future research. Finally, we mention that quadratic convergence is established for the predictor-corrector algorithm with a slight modification of the step size selection. It would be interesting to find out whether quadratic convergence can be proved for the "original" predictor-corrector algorithm. --R Complementarity and nondegeneracy in semidefinite programming. An interior-point method for semidefinite programming Local convergence of predictor-corrector infeasiblee-interior-point algorithms for semidefinite programs A predictor-corrector interior-point algorithm for the semidefinite linear complementarity problem using the Alizadeh-Haeberly-Overton search direction Superlinear convergence of a symmetric primal-dual path following algorithm for semidefinite programming Polynomial convergence of primal-dual algorithms for semidefinite programming based on Monteiro and Zhang family of directions Local convergence of interior-point algorithms for degenerate monotone LCP A unified analysis for a class of path-following primal-dual interior-point algorithms for semidefinite programming A superlinearly convergent primal-dual infeasible-interior- point algorithm for semidefinite programming Superlinear convergence of interior-point algorithms for semidefinite programming Superlinear convergence of a predictor-corrector method for semidefinite programming without shrinking central path neighborhood On a general class of interior-point algorithms for semidefinite programming with polynomial complexity and superlinear convergence On extending primal-dual interior-point algorithms from linear programming to semidefinite programming --TR --CTR Y. B. Zhao, Enlarging neighborhoods of interior-point algorithms for linear programming via least values of proximity measure functions, Applied Numerical Mathematics, v.57 n.9, p.1033-1049, September, 2007
interior point method;superlinear convergence;semidefinite programming
589294
A Spectral Bundle Method for Semidefinite Programming.
A central drawback of primal-dual interior point methods for semidefinite programs is their lack of ability to exploit problem structure in cost and coefficient matrices. This restricts applicability to problems of small dimension. Typically, semidefinite relaxations arising in combinatorial applications have sparse and well-structured cost and coefficient matrices of huge order. We present a method that allows us to compute acceptable approximations to the optimal solution of large problems within reasonable time.Semidefinite programming problems with constant trace on the primal feasible set are equivalent to eigenvalue optimization problems. These are convex nonsmooth programming problems and can be solved by bundle methods. We propose replacing the traditional polyhedral cutting plane model constructed from subgradient information by a semidefinite model that is tailored to eigenvalue problems. Convergence follows from the traditional approach but a proof is included for completeness. We present numerical examples demonstrating the efficiency of the approach on combinatorial examples.
Introduction . The development of interior point methods for semidefinite programming [19, 31, 1, 46] has increased interest in semidefinite modeling techniques in several fields such as control theory, eigenvalue optimization, and combinatorial optimization. In fact, interior point methods proved to be very useful and reliable solution methods for semidefinite programs of moderate size. However, if the problem is defined over large matrix variables or a huge number of constraints interior point methods grow terribly slow and consume huge amounts of memory. The most efficient methods of today [15, 23, 2, 32, 45, 29] are primal-dual methods that require, in each iteration of the interior point method, the factorization of a dense matrix of order equal to the number of constraints and one to three factorizations of the positive semidefinite matrix variables within the line search. For a typical workstation this restricts the number of constraints to 2000 and the size of the matrix variables to 500 if reasonable performance is required. For larger problems time and memory requirements are prohibitive. It is important to realize that either the primal or the dual matrix is generically dense even if cost and coefficient matrices are very sparse. Very recently, a pure dual approach was proposed in [4] which offers some possibilities to exploit sparsity. It is too early to judge the potential of this method. In combinatorial optimization semidefinite relaxations where introduced in [27]. At that time they were mainly considered a theoretical tool for obtaining strong bounds [11, 28, 40]. With the development of interior point methods hopes soared high that these relaxations could be of practical value. Within short time several approximation algorithms relying on semidefinite programming were published, most of them based on the approach by Goemans and Williamson [8]. On the implementational side [14, 16, 20] cutting plane approaches for semidefinite relaxations of Konrad-Zuse-Zentrum fur Informationstechnik Berlin, Takustrae 7, 14195 Berlin, Germany. helmberg@zib.de, http://www.zib.de/helmberg y Universitat Klagenfurt, Institut f. Mathematik, Universitatsstr. 65-67, A 9020 Klagenfurt, Aus- tria. franz.rendl@uni-klu.ac.at. Financial support through the Austrian FWF Project P12660- MAT is greatfully acknowledged. constrained quadratic 0-1 programming problems proved to yield solutions of high quality. However, as mentioned above, they were very expensive to compute even for problems of small size (a few hundred 0-1 variables). Problems arising in practical applications (starting with a few thousand 0-1 variables) were out of reach. We believe that the method proposed in this paper will open the door to problems of this size. Although combinatorial applications are our primary concern we stress that the method is not restricted to this kind of problems. In fact it will be a useful alternative to interior point methods whenever the number of constraints or the order of the matrices is quite large. We transform a standard dual semidefinite program into an eigenvalue optimization problem by reformulating the semidefinite constraint as a non-negativity constraint on the minimal eigenvalue of the slack matrix variable and lifting this constraint into the cost function by means of a Lagrange multiplier. The correct value of the Lagrange multiplier is known in advance if the primal feasible matrices have constant trace. (This is the case for the combinatorial applications we have in mind.) In this paper we develop a bundle method for solving the problem of minimizing the maximal eigenvalue of an affine matrix function with an additional linear objective term. These functions are well known to be convex and non-smooth. A very general method for optimizing non-smooth convex functions is the bundle method, see e.g. [21, 42, 17, 18]. In each step the function value and a subgradient of the function is computed for some specific point. By means of the collected subgradients a cutting plane model of the function is formed. The minimizer of the cutting plane model augmented by a regularization term yields the new point. In the case of eigenvalue optimization the subgradient is formed by means of an eigenvector to the maximal eigenvalue. Extremal eigenvalues and associated eigenvectors of large symmetric matrices can be computed efficiently by Lanczos methods (see e.g. [9]). Lanczos methods need a subroutine that computes the product of the matrix with a vector. This allows to exploit any kind of structure present in the matrix. The polyhedral cutting plane model used in traditional bundle algorithms is up-dated by new subgradient information such as to approximate well the subdifferential, and thus the function itself, in the vicinity of the current point. For eigenvalue optimization problems the subdifferential is generated by a semidefinite set, in particular by the intersection of a simple affine constraint and a face of the semidefinite cone. This suggests to use, instead of the traditional polyhedral cutting plane model, a semidefinite cutting plane model that works with an approximation of this face of the semidefinite cone. This specialization of the cutting plane model is the main contribution of the paper. The semidefinite bundle approach allows for an intriguing interpretation in terms of the original semidefinite program. The cutting plane model requires that the dual slack matrix of the semidefinite program is positive semidefinite only with respect to a subspace of vectors, thus it may be interpreted as a relaxation of the dual semidefinite program. In general the optimal solution of this relaxed semidefinite problem will produce an indefinite dual slack matrix. One or more of the negative eigenvalues and corresponding eigenvectors of the slack matrix are used to update the subspace in order to improve the relaxation and the process is iterated. This process trivially provides the optimal solution if the subspace grows to the full space. However, we show that during the algorithm generically the dimension of the subspace is bounded by (roughly) the square root of the number of constraints. If this A SPECTRAL BUNDLE METHOD 3 is still considered too large the introduction of an aggregate subgradient guarantees convergence for restricted bundle sizes. In the extreme the bundle may consist of one new eigenvector to the maximal eigenvalue only. In contrast, the 'classical' algorithms of Cullum, Donath, and Wolfe [6] and Polak and Wardi [38] require in each iteration the computation of all eigenvectors to eigenvalues within an "-distance of the maximal eigenvalue, thus close to the optimal solution this number is at least as large as the multiplicity of the maximal eigenvalue in the optimal solution. In the quadratically convergent algorithm of Overton [35] each step is computed from a complete spectral decomposition of the matrix and a guess of the exact multiplicity of the maximal eigenvalue in the optimal solution. In recent work [33, 34] Oustry reinterprets the algorithm of Overton within the frame-work of the U-Lagrangian introduced in [26] and embeds it in a first order method to ensure global convergence. Again, for global convergence the approach relies on the spectrum of all eigenvalues within "-distance of the maximal eigenvalue and makes use of the entire spectral information to obtain local quadratic convergence. Because of the restricted bundle size quadratic convergence is out of reach for our algorithm, it is a first order method only. In principle convergence follows from the traditional approach (see e.g. [21]) but we include a proof for completeness. We also present a primal-dual interior point code for solving the quadratic semidefinite programming problems associated with the semidefinite cutting plane models and discuss efficiency aspects. The properties of the algorithm are illustrated on several combinatorial examples. In x2 some basic properties of semidefinite programs are stated. Then we transform semidefinite programs into eigenvalue optimization problems. Section 3 introduces the bundle method. The algorithm and the proof of convergence is given in x4. The quadratic semidefinite subproblems arising in the bundle method can be solved by interior point methods as explained in x5. Section 6 gives an outline of the implementation and briefly discusses the computation of the maximal eigenvalue and an associated eigenvector. Numerical examples on combinatorial problems are presented in x7. We conclude the paper with a summary and possible extensions and improvements in x8. For the convenience of the reader an appendix explaining the notation and the symmetric Kronecker product is included at the end of the paper. 2. Semidefinite programs and eigenvalue optimization. We denote the set of symmetric matrices of order n by Sn which we regard as a space isomorphic to R ( n+1 As scalar product of A; B 2 Sn (or more general, use the trace is the sum of the diagonal elements of a square matrix. We will often use the same symbol for the canonical scalar product of vectors a; b a, the appropriate space will be clear from the context. The subset of positive semidefinite matrices S n is a full-dimensional, non- polyhedral convex cone in Sn and defines a partial order on the symmetric matrices by A B n . Positive definite matrices are denoted by S ++ n or A 0. Consider the standard primal-dual pair of semidefinite programs, (D) Z 0: linear operator and A T its adjoint operator, defined by hAX; yi ff for all X 2 Sn and y They are of the form with A i 2 Sn , is the cost matrix, b 2 R m the right-hand-side vector. We assume some constraint qualification to hold, so that these problems satisfy strong duality in the sense that for any optimal solution X of (P) and any optimal solution (y ; Z ) of (D) we have The following assumption allows a simple reformulation of the dual (D) as an eigenvalue optimization problem. We assume that for some constant a ? 0. In this case we can add tr a as a redundant constraint to the primal problem and obtain the following dual equivalent to (D) Now a ? 0 implies X 6= 0 at the optimum, hence any optimal Z of this dual is singular. Therefore all dual optimal solutions Z satisfy leading to Thus we have shown that (D) is equivalent to min y amax convenience we assume deal with the following problem. The eigenvalue problem (E) is a convex, non-smooth optimization problem. It is well studied in the literature. Here we only recall some basic facts. The function is differentiable if and only if the maximal eigenvalue has multiplicity one. When optimizing eigenvalue functions, the optimum is generically attained at matrices whose maximal eigenvalue has multiplicity larger than one. In this case one has to consider the subdifferential of max at X , (see e.g. [35]). In particular, for any v 2 R n belonging to the eigenspace of the maximal eigenvalue of X , contained in the subdifferential of max at X . For the function of interest, A SPECTRAL BUNDLE METHOD 5 the subdifferential of f at y can be derived by standard rules (see [17]), Observe that the set of all subgradients is bounded. Remark 2.1. Even though our assumption (2.2) might look artificial, it does hold for SDP arising from quadratic 0-1 optimization. It also holds for many other SDP derived as relaxations of combinatorial optimization problems, see for instance [1, 12, 24]. 3. The bundle method. In this section we develop a new method for minimizing f . We use two classical ingredients, the proximal point idea, and the bundle concept. The new contribution lies in the way that we derive the new iterate from the 'bundle' of subgradient information collected from previous iterates. Since our approach builds on several subtle ideas, we proceed in small steps and explain first, how we derive a minorant of f from local information. 3.1. Minorizing f by " f. Our first goal is to obtain a minorant " f of f which approximates f in the neighborhood of the current iterates reasonably well, and which is easier to handle than f . Introducing the function we can express f(y) as This formulation shows that lower approximations of f can be obtained by constraining W to a subset of all semidefinite matrices with tr We propose the following choice for this subset. Let P be n \Theta r with P T n with tr be two matrices. We restrict W to be contained in the set c The f , defined through P and W , now reads Wg: By definition, we have " f(y) f(y) 8y: If, for some " W for some eigenvector v to f("y). This is e.g. the case if v is a column of P or v is contained in the range space of P . The intuitive idea behind our specific choice of c W is as follows: the matrix P contains subgradient information from the current point " y, and perhaps from previous iterates. We explain below in detail, how we propose to select and update the matrix P . For computational efficiency, we would like to keep the number r of columns of P small, independent of the multiplicity of the largest eigenvalue. Therefore we collect indispensable subgradient information, that has to be removed from P , in an aggregate subgradient. This aggregation is the final ingredient of our local model of f . The matrix W plays the role of an aggregate subgradient. Again, we will discuss be- low, how W is updated during the algorithm. The main point here is that instead of optimizing over all semidefinite matrices W , we constrain ourselves to a small subset. Remark 3.1. If we set use for the matrix P a set of eigenvectors to the r largest eigenvalues at " y, we would end up with a model closely related to the approach from [6]. In this case it would be important to select r at least as large as the multiplicity of the largest eigenvalue. In our present approach this is not necessary. 6 C. HELMBERG AND F. RENDL 3.2. Proximal point idea. The next goal is to minimize " f instead of f . Since f is built from local information from a few previous iterates, this model function is unlikely to be reliable for points far from the current iterate. Therefore we use the proximal point idea and add a penalty term for the displacement from the current point. Thus we determine a new candidate y from the current iterate " y by solving the following convex problem, referred to as the augmented model. (Here u ? 0 is some fixed real weight.) min y We note that this minimization problem corresponds to the Lagrangian relaxation of Thus we replace the original function f by its minorant " f and minimize locally around " y. The weight u controls (indirectly) the radius s of the sphere around " y, over which we minimize. Substituting the definition of " f , this problem is the same as min y This problem can be simplified, because y is unconstrained. Note that Therefore we obtain min y W2c W; b\GammaAW +u(y\Gamma"y)=0 W2c The first equality follows from interchanging min and max (see Corollary 37.3.2 of [41]) and using first order optimality for the inner minimization with respect to y, The final problem is a semidefinite program with (concave) quadratic cost function. We will discuss in x5 how problems of this kind can be solved efficiently. Its optimal solution W k+1 gives the new trial point y by (3.3). Remark 3.2. The choice of the weight u is somewhat of an art. There are several clever update strategies published in the literature, see for instance [21, 42]. 3.3. One iteration of the algorithm. The main ingredients of our approach have now been explained, so we can give a formal description of a general iteration k of the algorithm. To be consistent with the notation of the algorithm given in x4, let us denote by x k what was called " y in x3.2. The algorithm may have to compute several trial points y k+1 , y keeping the same x progress is not considered satisfactory (null step). For each y k+1 the function is evaluated and a subgradient (eigenvector) is computed. This information is added to c W k to form an improved model c W k+1 . Therefore, we assume that the current 'bundle' A SPECTRAL BUNDLE METHOD 7 contains an eigenvector of its span (y k may or may not be equal to x k ). Other than that, P need only The minorant of f in iteration k is denoted by " Here c represents the current approximation to the set of all semidefinite matrices of trace one, see (3.1). It will be convenient to introduce also the regularized version of " The new trial point y k+1 is obtained by minimizing f k (y) with respect to y. As described above, this can be done as follows. First, solve by interior point methods (see x5) yielding a (not necessarily unique) maximizer W use (3.3) to compute To finish an iteration, we have to decide whether enough progress is made to perform a serious step or not, i.e. whether we are going to set x how to update P k and W k If P k does not yet use the maximum number of columns allowed then the update process is simple: orthogonalize the new eigenvector with respect to P k , add it as a new column to form P k+1 and continue. In general, however, P k will already use the maximum number of columns and so we have to make room for the new subgradient information. Instead of simply eliminating some columns of P k we can do better by exploiting the information available in ff and V . Let Q\LambdaQ T be an eigenvalue decomposition of V . Then the 'important' part of the spectrum of W k+1 (the important subspace within the space spanned by P k ) is spanned by the eigenvectors associated with the 'large' eigenvalues of V . Thus we split the eigenvectors of Q into two parts (with corresponding spectra 1 and 2 containing as columns the eigenvectors associated to 'large' eigenvalues of V and Q 2 containing the remaining columns, Now the next P k+1 is computed to contain P k Q 1 and at least one eigenvector v k+1 to the maximal eigenvalue of (The operator orth(.) indicates that we take an orthonormal basis of [P k The next aggregate matrix is built in such a way that W k+1 2 c contains only the important part of P k , given by P k Q 1 , we include the remaining part of P k , given by P k Q 2 in W k+1 (ff W k Note that W k+1 is scaled to have trace equal one. Proposition 3.3. Update rules (3.7) and (3.8) ensure that W k+1 2 c W k+1 . Proof. Let W k+1 be of the form (3.6). By (3.7) there is an orthonormal matrix such that P k+1 W k+1 . We summarize some easy facts, which will be used in the convergence analysis of the algorithm. since y k+1 is minimizer of f k . Because f k Next let Using the definition of y k+1 from (3.5) it follows easily that (y the augmented model of the next iteration will satisfy (y) 8y: Remark 3.4. While the choice for the update of P k is fairly natural, we could use other update formulas, such as W . The main properties guiding the update are that W k+1 2 c ensuring (3.11) and that in y k+1 the model is now supported by a subgradient of f pushing the model towards f in the vicinity of the last minimizer. 4. Algorithm and convergence analysis. In the previous section we focused on the question of doing one iteration of the bundle method. Now we provide a formal description of the method and point out that except for the choice of the bundle, the nature of the subproblem, and some minor changes in parameters the algorithm and its proof are identical to the algorithm of Kiwiel as presented in [21]. To keep the paper self-contained we present and analyze a simplified variant for fixed u. We refer the reader to [21] for an algorithm with variable choice of u. Algorithm 4.1. Input: An initial point y to the maximal eigenvalue of C \Gamma A T y 0 , an " ? 0 for termination, an improvement parameter mL 2 an upper bound R 1 on the number of columns of P . 1. 2. (Direction finding) Solve (3.4) to get y k+1 from (3.5). Decompose V into using (3.8). 3. (Evaluation) Compute and an eigenvector v k+1 . Compute P k+1 by (3.7). 4. (Termination) If f(x k A SPECTRAL BUNDLE METHOD 9 5. (Serious step) If then set x continue with Step 7. Otherwise continue with Step 6. 6. (Null step) Set x 7. Increase k by 1 and go to Step 2. We prove convergence of the algorithm for If the algorithm stops after a finite number of iterations then by and thus by (3.5) 0 2 @f(x k ), so x k is optimal. Assume in the following that the algorithm does not stop. First consider the case that only null steps occur after some iteration K. Lemma 4.2. If there is a K 0 such that (4.1) is violated for all k K, then Proof. For convenience we set Using the relations (3.10), and (3.9), we obtain for all k K (y Therefore the f k (y k+1 ) converge to some f f(x) and the computed gradient of f in y k+1 and observe that the linearization f of f in y k+1 ff W k+1 . Thus \Gamma\Omega ff The convergence of the f k (y k+1 ), the boundedness of the gradients and the fact that imply that the last term goes to zero for k !1. So for all there is an M 2 N such that for all k ? M where '!' follows from (4.1) being violated for all k ? K. Thus the sequences f(y k+1 ) both converge to f(x). y k+1 is the minimizer of the regularized function f k . On the one hand this implies that y k+1 ! x. On the other hand 0 must be contained in the subgradient @f k (y k+1 Therefore there is a sequence h k 2 @ " subgradients converging to zero. The converge to f(x) and the y k+1 converge to x, hence zero must be contained in @f(x). We may concentrate on serious steps in the following. In order to simplify notation we will speak of x k as the sequence generated by serious steps with all duplicates eliminated. By f k (and the corresponding " refer to the function whose minimization gives rise to x k+1 . The next lemma investigates the case that the f(x k ) remain above some value f(~x) for some fixed ~ x. Lemma 4.3. If fixed ~ m and all k then the x k converge to a minimizer of f . Proof. First we prove the boundedness of the x k . To this end denote by g k+1 2 subgradient arising from the optimal solution of the minimization problem observe that by (3.3) Therefore the distance of x k+1 to ~ x can be bounded by 2\Omega ~ ff 2\Omega ~ ff For any k ? K, a recursive application of the bound above yields uX By (4.1) the progress of the algorithm in each serious step is at least mL (f(x k together with (4.2) we obtainX Therefore the sequence of the x k remains bounded and has an accumulation point x. By replacing ~ x by x in (4.4) and choosing K sufficiently large, the remaining sum can be made smaller than an arbitrary small ffi ? 0, thus proving the convergence of the x k to x. As the x k+1 converge to x the g k+1 converge to zero by (4.3), and since the sequence (f(x k has to converge to zero as well, we conclude that x is a minimizer of f . The lemma also implies that f(x k there are no minimizers. We summarize the discussion in the following theorem. Theorem 4.4. [21] If the set of minimizers of f is not empty then the x k converge to a minimizer of f . In any case f(x k A SPECTRAL BUNDLE METHOD 11 Remark 4.5. We have just seen that the bundle algorithm works correctly even if P contains only one column. In this case the use of the aggregate subgradient is crucial. To achieve correctness of the bundle algorithm without aggregate subgradients, it suffices to store in P only the subspace spanning the eigenvectors corresponding to non-zero eigenvalues of an optimal solution W k+1 of (3.2). Using the bound of [36] it is not too difficult to show that in this case the maximal number of columns one has to provide is the largest plus the number of eigenvectors to be added in each iteration (this is at least one). In our computational experiments we found that this upper bound is hardly ever reached. In fact, typical values for the maximal rank are around half this upper bound. 5. Solving the subproblem. In this section we concentrate on how the minimizer of f k can be computed efficiently. We have already seen in x3 that this task is equivalent to solving the quadratic semidefinite program (3.4). Problems of this kind can be solved by interior point methods, see e.g. [7, 23]. Dropping the iteration index k and the constants in (3.4) we obtain for ff ff 0; V 0: Expanding into the cost function yields ff ff ff 0; V 0: Using the svec-operator (see the appendix for a definition and important properties of svec and the symmetric Kronecker product\Omega s ) to expand symmetric matrices from S r into column vectors of length we obtain the quadratic program (recall that, for I svec(V where (after some technical linear algebra) ff ff +\Omega C; W At this point it is advisable to spend some thought on W . The algorithm is designed for very large and sparse cost matrices C. W is of the same size as C. Initially it might be possible to exploit the low rank structure of W for efficient representations, but as the algorithm proceeds, the rank of W grows inevitably. Thus it is impossible to store all the information of W . However, as we can see in (5.2) to (5.6), it suffices to have available the vector AW 2 R m and the scalar\Omega C; W ff to construct the quadratic program. Furthermore, by the linearity of A(\Delta) and hC; \Deltai, these values are easily updated whenever W is changed. To solve (5.1) we employ a primal-dual interior point strategy. To formulate the defining equations for the central path we introduce a Lagrange multiplier t for the equality constraint, a dual slack matrix U 0 as complementary variable to V , a dual slack scalar fi 0 as complementary variable to ff and a barrier parameter ? 0. The system reads ts I \Gamma I svec(V The step direction (\Deltaff; \Deltafi; \DeltaU; \DeltaV; \Deltat) is determined via the linearized system I svec(\DeltaV In the current context we prefer the linearization because it makes the system easy to solve for \DeltaV with relatively little computational work per iteration. The final system for \DeltaV reads ff It is not too difficult to see that the system matrix is positive definite (because suffices to show that Q using 0). The main work per iteration is the factorization of this matrix (with v 2 S r this is it is not possible to do much better since Q 11 has to be inverted at some point. Because of the strong dominance of the factorization it pays to employ a predictor corrector approach, but we will not delve into this here. For strictly feasible primal starting point is a strictly feasible dual starting point can be constructed by choosing t 0 sufficiently negative such that A SPECTRAL BUNDLE METHOD 13 Starting from this strictly feasible primal-dual pair we compute the first by compute the step direction (\Deltaff; \Deltafi; \DeltaU; \DeltaV; \Deltat) as indicated above , perform a line search with line search parameter strictly feasible, move to this new point, compute a new by ae oe with and iterate. We stop if (hU; 6. Implementation. In our implementation of the algorithm we largely follow the rules outlined in [21]. In particular u is adapted during the algorithm. The first guess for u is equal to the norm of the first subgradient determined by v 0 . The scheme for adapting u is the same as in [21] except for a few changes in parameters. For example the parameter mL for accepting a step as serious is set to the parameter mR indicating that the model is so good (progress by the serious step is larger than mR [f(x k that u can be decreased is set to The stopping criterion is formulated in relative precision, in the implementation. The choice of the upper bound R on the number of columns r of P and the selection of the subspace merits some additional remarks. Observe that by Remark 4.5 it is highly unlikely the r violates the bound even if the number of columns of P is not restricted. is also the order of the system matrix in (5.7) and is usually considerably smaller than the size of the system matrix in traditional interior point codes for semidefinite programming which is of order m. Furthermore the order of the matrix variables is r as compared to n for traditional interior point codes. Thus if the number of constraints m is roughly of the same size as n and a matrix of order m is still considered factorizable then running the algorithm without bounding the number of columns of P may turn out to be considerably faster than running an interior point method. This can be observed in practice, see x7. For huge n and m primal-dual interior point methods are not applicable any more, because X , Z \Gamma1 , and the system matrix are dense. In this case the proposed bundle approach allows to apply the powerful interior point approach at least on an important subspace of the problem. The correct identification of the relevant subspace in V is facilitated by the availability of the complementary variable U . U helps to discern between the small eigenvalues of V (because of the interior point approach we have V 0!). Eigenvectors v of V that are of no importance for the optimal solution of the subproblem will have a large value v T U v, whereas eigenvectors, that are ambiguous, will have both, a small eigenvalue v T V v and a small value v T U v. In practice we restrict the number of columns of P to 25 and provide room for at least five new vectors in each iteration (see below). Eigenvectors v that correspond to small but important eigenvalues of V are added to W ; important eigenvectors are added to W only if more room is needed for new vectors. For large m the computation of (5.2) to (5.6) is quite involved. A central object appearing in all constants is the projection of the constraint A i on the space spanned 14 C. HELMBERG AND F. RENDL by P , P T A i P . Since the A i are of the same size as X which we assume to be huge, it is important to exploit whatever structure is present in A i to compute this projection efficiently. In combinatorial applications the A i are of the form vv T with v sparse and the projection can be computed efficiently. In the projection step and in particular in forming Q 11 the size of r is again of strong influence. If we neglect the computation of the computation of Q 11 still requires 2m flops. Indeed, if m is large then for small r the construction of Q 11 takes longer than solving the associated quadratic semidefinite program. The large computational costs involved in the construction and solution of the semidefinite subproblems may lead to the conviction that this model may not be worth the trouble. However, the evaluation of the eigenvalue-function is in fact much more expensive. There has been considerable work on computing eigenvalues of huge, sparse matrices, see e.g. [9] and the references therein. For extremal eigenvalues of symmetric matrices there seems to be a general consensus, that Lanczos type methods work best. Iterative methods run into difficulties if the eigenvalues are not well separated. In our context it is to be expected that in the course of the algorithm the largest eigenvalues will get closer and closer till all of them are identical in the optimum. For reasonable convergence block Lanczos algorithms with blocksize corresponding to the largest multiplicity of the eigenvalues have to be employed. During the first ten iterations the largest eigenvalue is usually well separated and the algorithm is fast. But soon the eigenvalues start to cluster, larger and larger blocksizes have to be used, and the eigenvalue problem gets more and more difficult to solve. In order to reduce the number of evaluations it seems worth to employ powerful methods in the cutting plane model. The increase in computation time required to solve the subproblem goes hand in hand with the difficulty of the eigenvalue problem because of the correspondence of the rank of P and the number of clustered eigenvalues. Iterative methods for computing maximal eigenvectors generically offer approximate eigenvectors to several other large eigenvalues, as well. The space spanned by these approximate eigenvectors is likely to be a good approximation of the true eigenspace. If the maximal number of columns for P is not yet attained it may be worth to include several of these approximate eigenvectors as well. In our algorithm we use a block Lanczos code of our own that is based on a Fortran code of Hua (we guess that this is Hua Dai of [47]). It works with complete orthogonalization and employs Chebyshev iterations for acceleration. The choice of the blocksize is based on the approximate eigenvalues produced by previous evaluations but is at most 30. Four block Lanczos steps are followed by twenty Chebyshev iterations. This scheme is repeated till the maximal eigenvalue is found to the required relative precision. The relative precision depends on the distance of the maximal to the second largest eigenvalue but is bounded by 10 \Gamma6 . As starting vectors we use the complete block of eigenvectors and Lanczos-vectors from the previous evaluation. 7. Combinatorial applications. The combinatorial problem we investigate is quadratic programming in f\Gamma1; 1g variables, In the case that C is the Laplace matrix of a (possible weighted) graph the problem is known to be equivalent to the max-cut problem. The standard semidefinite relaxation is based on the identity x T Cx ff . For all f\Gamma1; 1g n vectors, xx T is a positive semidefinite matrix with all diagonal elements equal to one. We relax xx T to X 0 and and obtain the following A SPECTRAL BUNDLE METHOD 15 primal-dual pair of semidefinite programs, Z 0: For non-negatively weighted graphs a celebrated result of Goemans and Williamson [8] says, that there is always a cut within :878 of the optimal value of the relaxation. One of the first attempts to approximate (DMC) using eigenvalue optimization is contained in [39]. The authors use the Bundle code of Schramm and Zowe [42] with a limited number of bundle iterations, and so do not solve (DMC) exactly. So far the only practical algorithms for computing the optimal value were primal-dual interior point algorithms. However these are not able to exploit the sparsity of the cost function and have to cope with dense matrices X and Z \Gamma1 . An alternative approach based on a combination of the power method with a generic optimization scheme of Plotkin, Shmoys, and Tardos [37] was proposed in [22] but seems to be purely theoretical. In Table 7.1 we compare the proposed bundle method to our semidefinite primal-dual interior point code of [14] (called PDIP in the sequel) for graphs on nodes that were generated by rudy, a machine independent graph generator written by G. Rinaldi. Table 7.7 contains the command lines specifying the graphs. Graphs G 1 to G 5 are unweighted random graphs with a density of 6% (approx. 19000 edges). G 6 to G 10 are the same graphs with random edge weights from f\Gamma1; 1g. G 11 to G 13 are toroidal grids with random edge weights from f\Gamma1; 1g (1600 edges). G 14 to G 17 are unweighted 'almost' planar graphs having as edge set the union of two (almost maximal) planar graphs (approx. 4500 edges). G to G 21 are the same almost planar graphs with random edge weights from f\Gamma1; 1g. In all cases the cost matrix C is the Laplace matrix of the graph divided by 4, i.e., let A denote the (weighted) adjacency matrix of G, then For a description of the code PDIP see [14], the termination criterion requires the gap between primal and dual optimal solution to be closed to a relative accuracy of For the bundle algorithm, (DMC) is transformed into an eigenvalue optimization problem as described in x2. In addition the diagonal of C is removed so that, in fact, the algorithm works on the problem min with This does not change problem (PMC) because the diagonal elements of X are fixed to one. The offset 1e T (Ae \Gamma diag(A)) is added to the output only and has no influence on the algorithm whatsoever, in particular it has no influence on the stopping criterion. As starting vector y 0 we choose the zero vector. All other parameters are as described in x6. All computation times, for the interior point code PDIP as well as for the bundle code, refer to the same machine, a Sun sparc Ultra 1 with a Model 140 UltraSPARC CPU and 64 MB RAM. The time measured is the user time and it is given in the leading zeros are dropped. The first column of Table 7.1 identifies the graphs. The second and third refer to PDIP and contain the optimal objective value produced (these can be regarded as highly accurate solutions) and the computation time. The fourth and fifth column give the same numbers for the bundle code. On these examples the bundle code is superior to PDIP. Although the examples do belong to the favorable class of instances having small m and relatively large n, the difference in computation time is astonishing. Note that the termination criterion used in the bundle code is quite accurate, except for G 11 which seems to be a difficult problem for the bundle method. This deviation in accuracy is not caused by cancellations in connection with the offset. The difficulty of an example does not seem to depend on the number of nonzeros but rather on the shape of the objective function. For toroidal grid graphs the maximum cut is likely to be not unique, thus the objective function will be rather flat. This flatness has its effect on the distribution of the eigenvalues in the optimal solution. Indeed, for G 11 more eigenvalues cluster around the maximal eigenvalue than for the other problems. We illustrate this in Table 7.2, which gives the largest eigenvalues of the solution at termination for problems G 1 , G 6 , G 11 , G 14 , and G Table Comparison of the interior point (PDIP) and the bundle (B) approach. sol gives the computed solution value and time gives the computation time. PDIP-sol PDIP-time B-sol B-time G2 12089.43 1:19:14 12089.45 5:19 G6 2656.16 1:24:53 2656.18 3:57 G 11 629.16 1:28:41 629.21 45:26 G G 17 G G 19 Table 7.3 provides additional information on the performance of the bundle algorithm on the examples of Table 7.1. The second column gives the accumulated time spent in the eigenvalue computation, it accounts for roughly 90% of the computation time. serious displays the number of serious steps, iter gives the total number of iterations including both, serious and null steps. kgk is the norm of the subgradient arising from the last optimal W k+1 before termination. For G 11 the norm is considerably higher than for all other examples. Since the desired accuracy was not achieved for G 11 by the standard stopping criterion it may be worth to consider an alternative stopping criterion taking into account the norm of the subgradient as well. Column max-r gives the maximal rank of P attained over all iterations. The rank of P would A SPECTRAL BUNDLE METHOD 17 Table The maximal eigenvalues after termination of examples G 1 , G6 , G11 , G14 , and G18 . 13 3.1190 3.2239 0.7651 1.0557 1.4047 14 3.1135 3.2181 0.7650 1.0515 1.4007 19 2.7214 2.7716 0.7647 1.0398 1.3725 22 2.6834 2.6756 0.7644 1.0341 1.3583 26 2.6274 1.8722 0.7636 1.0239 1.3480 28 2.5137 1.7974 0.7633 1.0211 1.3397 29 2.4840 1.4859 0.7630 1.0180 1.3345 have been bounded by 25, but this bound never came into effect for any of these examples. Aggregation was not necessary. Observe that the theoretic bound allows for r up to 39, yet the maximal rank is only half this number. The last column gives the time when the objective value was first within 10 \Gamma3 of the optimum. For combinatorial applications high accuracy of the optimal solution is of minor importance. An algorithm should deliver a reasonable bound fast and its solution should provide some hint on how a good feasible solution can be constructed. The bundle algorithm offers both. With respect to computation time the bundle algorithm displays the usual behavior of subgradient algorithms. Initially progress is very fast, but as the bound approaches the optimum there is a strong tailing off effect. We illustrate this by giving the objective values and computation times for the serious steps of example G 6 (the diagonal offset is +77 in this example) in Table 7.4. After one minute the bound is within 0:1% of the optimum. For the other examples see the last column of Table 7.3. With respect to a primal feasible solution observe that P k V k successively better and better approximation to the primal optimal solution X . In case too much information is stored in the aggregate vector AW k (remember that it is not advisable to store W k itself), P k may be enriched with additional Lanczos-vectors from the eigenvalue computation. The solution of this enlarged quadratic semidefinite subproblem will be an acceptable approximation of X . It is not necessary to Table Additional information about the bundle algorithm for the examples of Table 7.1. -time gives the total amount of time spent for computing the eigenvalues and eigenvectors, serious gives the number of serious steps, iter the total number of iterations including null steps. kgk refers to the norm of the gradient resulting from the optimal solution of the last semidefinite subproblem. max-r is the maximum number of columns used in P (the limit would have been 25). 0.1%-time gives the time when the bound is within 10 \Gamma3 of the optimum in relative precision. -time serious iter kgk max-r 0.1%-time G1 3:12 22 33 0.1639 G4 2:38 19 27 0.08235 19 54 G 9 2:59 G13 17:24 43 78 0.218 15 6:17 G G 19 11:24 41 71 0.1571 15 3:34 construct the whole matrix X . In fact, the factorized form (P k much more convenient to work with. For example the approximation algorithm of Goemans and Williamson [8] requires precisely this factorization. A particular x ij element of X is easily computed by the inner product of row i and j of the n \Theta r . In principle this opens the door for branch and cut approaches to improve the initial relaxation. This will be the subject of further work. Table 7.5 gives a similar set of examples for A last set of examples is devoted to the Lov'asz #-function [27] which yields an upper bound on the cardinality of a maximal independent (or stable) set of a graph. For implementational convenience we use its formulation within the quadratic f\Gamma1; 1g programming setting, see [24]. For a graph with k nodes and h edges we obtain a semidefinite program with matrix variables of order constraints. The examples we are going to consider have more than one thousand nodes and more than six thousand edges. For these examples interior point methods are not applicable any more because of memory requirements. It should be clear from the examples of Table 7.5 that there is also little hope for the bundle method to terminate within reasonable time. However, the most significant progress is achieved in the beginning and for the bundle method memory consumption is not a problem. We run these examples with a time limit of five hours. More precisely, the algorithm is terminated after the first serious step that occurs after five hours of computation time. The graph instances are of the same type as above. The computational results are displayed in Table 7.6. The new columns n and m give the order of the matrix variable and the number of constraints, respectively. Observe that the toroidal grid graphs A SPECTRAL BUNDLE METHOD 19 Table Detailed account of the serious steps of example G 6 . iter value time kgk max-r 8 2670.10 28 5.992 15 9 2666.07 34 4.173 14 2656.31 1:57 0.431 19 2656.19 3:33 0.07243 Table Examples for B-sol B-time -time serious iter kgk max-r G22 14135.98 38:11 28:00 26 52 0.0781 23 G26 14132.93 34:45 26:37 31 48 0.3066 23 G 28 4100.81 29:41 21:08 23 G G G 36 8006.04 2:56:10 2:31:09 62 115 0.2634 24 G 37 G 38 8015.01 4:03:53 3:39:24 58 155 0.1937 22 G G 48 and G 49 are perfect with independence number 1500; the independence number of G 50 is 1440 but G 50 is not perfect. We do not know the independence number of the other graphs. Except for G 48 and G 49 , which have '(G 48 perfectness, it is hard to judge the quality of the solutions. Tracing the development of the bounds the last serious steps of examples G 43 to G 47 and G 51 to G 54 still produced improvements of 0.5% to 1%. This and the rather large norm of the subgradient of Table Upper bound on the #-function after five hours of computation time. serious iter kgk max-r G 43 G 44 1001 10991 310.13 5:06:31 3:14:25 G48 3001 9001 1526.53 5:11:31 4:59:57 54 94 0.4062 15 G G 50 3001 9001 1536.12 5:17:51 5:01:25 50 124 0.4728 15 G 53 1001 6915 463.86 5:08:36 4:36:34 41 104 2.593 25 G 51 and G 54 indicate that the values cannot be expected to be 'good' approximations of the #-function. Also note, that the size of the subspace required for G 48 to G 50 is still well below 25. In examples G 51 to G 54 the value of ff is almost negligible, but for G 43 to G 47 the value of ff is roughly 1=3 at termination. Thus for these examples the restriction to 25 columns became relevant. The computational results of Table 7.6 demonstrate that the algorithm has its limits. Nonetheless the bounds obtained are still useful and the primal approximation corresponding to the subgradient is a reasonable starting point for primal heuristics. 8. Conclusions and extensions. We have proposed a proximal bundle method for solving semidefinite programs with large sparse or strongly structured coefficient matrices. The semidefinite constraint is lifted into the objective function by means of a Lagrange multiplier a whose correct value is not known in general, except for problems with fixed primal trace. In the latter case a is precisely the value of the trace. The approach differs from previous bundle methods in that the subproblem is tailored for semidefinite programming. In fact the whole approach can be interpreted as semidefinite programming over subspaces where the subspace is successively corrected and improved till the optimal subspace is identified. The set of subgradients modeled by the semidefinite subproblem is a superset of the subgradients used in the traditional polyhedral cutting plane model. Therefore convergence of the new method is a direct consequence of previous proofs for traditional bundle methods. It is not yet clear whether the specialized model admits stronger convergence results. The choice of u is still very much an open problem of high practical importance. For (constrained) quadratic f\Gamma1; 1g-programming the method offers a good bound within reasonable time and allows to construct an approximate primal optimal solution (of the relaxation) in compact representation. To improve the bound by a cutting plane approach the algorithm must be able to deal with sign constraints on the y-variables. In principle it is not difficult to model the sign constraints in the semidefinite subproblem. However, as a consequence the influence of the sign constrained y variables on the cost coefficients of the quadratic subproblem cannot be eliminated any longer, rendering the method impractical even for a moderate number of cutting planes. Alternatively one might consider active set methods but these entail the danger of destroying convergence. Together with K.C. Kiwiel we are currently working on alternative methods for incorporating sign constraints on y [13]. The backbone of the method is an efficient routine for computing the maximal eigenvalue of huge structured symmetric matrices. Although our own implementation A SPECTRAL BUNDLE METHOD 21 Table Arguments for generating the graphs by the graph generator rudy. G1 -rnd graph 800 6 8001 G2 -rnd graph 800 6 8002 G3 -rnd graph 800 6 8003 G4 -rnd graph 800 6 8004 G5 -rnd graph 800 6 8005 G6 -rnd graph 800 6 8001 -random G7 -rnd graph 800 6 8002 -random G8 -rnd graph 800 6 8003 -random G9 -rnd graph 800 6 8004 -random G12 -toroidal grid 2D 50 G13 -toroidal grid 2D 25 G14 -planar 800 99 8001 -planar 800 G17 -planar 800 99 8007 -planar 800 G22 -rnd graph 2000 1 20001 G23 -rnd graph 2000 1 20002 G24 -rnd graph 2000 1 G25 -rnd graph 2000 1 20004 G26 -rnd graph 2000 1 20005 G27 -rnd graph 2000 1 20001 -random G28 -rnd graph 2000 1 20002 -random G29 -rnd graph 2000 1 -times G31 -rnd graph 2000 1 20005 -random G32 -toroidal grid 2D 100 20 -random 0 1 -times G35 -planar 2000 99 20001 -planar 2000 G36 -planar 2000 99 -planar 2000 G37 -planar 2000 99 20005 -planar 2000 G38 -planar 2000 99 20007 -planar 2000 -planar 2000 99 20001 -planar 2000 G40 -planar 2000 99 -planar 2000 -planar 2000 99 20005 -planar 2000 -times 2 -plus G42 -planar 2000 99 20007 -planar 2000 G43 -rnd graph 1000 2 10001 G44 -rnd graph 1000 2 10002 G45 -rnd graph 1000 2 10003 G46 -rnd graph 1000 2 10004 G47 -rnd graph 1000 2 10005 G48 -toroidal grid 2D 50 -toroidal grid 2D G50 -toroidal grid 2D 25 120 G52 -planar 1000 100 10003 -planar 1000 100 10004 G53 -planar 1000 100 10005 -planar 1000 100 10006 G54 -planar 1000 100 10007 -planar 1000 100 10008 (based on the code of Hua) seems to work sufficiently stable there is certainly much room for improvement. A straight forward approach to achieve serious speed-ups is to implement the algorithm on parallel machines, see for instance [43]. Rather recently interest in the Lanczos method has risen again, see [25, 3, 5, 10, 30] and references therein. Most of these papers are based on the concept of an implicit restart proposed in [44] which is a polynomial acceleration approach that does not require additional matrix vector multiplications. It will be interesting to test these new ideas within the bundle framework. We thank K.C. Kiwiel for fruitful discussions and C. Lemar'echal and an anony- 22 C. HELMBERG AND F. RENDL mous referee for their constructive critisism that helped to improve the presentation. Appendix . Notation. R real column vector of dimension n real matrices real matrices positive definite matrices positive semidefinite matrices A 0 A is positive definite A 0 A is positive semidefinite I , I n identity of appropriate size or of size n e vector of all ones of appropriate dimension maximal eigenvalue of A tr A trace of A 2 M n;n , tr product in Mm;n , dimensional vector representation of A 2 Sn Kronecker product of A 2 M m;n diag(A) the diagonal of A 2 Mn as a column vector diagonal matrix with v on its main diagonal Sn is isomorphic to R via the map svec(A) defined by stacking the columns of the lower triangle of A on top of each other and multiplying the offdiagonal elements with 2, a The factor for offdiagonal elements ensures that, for The symmetric Kronecker product\Omega s is defined for arbitrary square matrices M n;n by its action on a vector svec(C) for a symmetric matrix C 2 Sn , Both concepts were first introduced in [2]. Here we use the notation introduced in [45]. From the latter paper we also cite some properties of the symmetric Kronecker product for the convenience of the reader. 1. B\Omega s A 2. (A\Omega s B) 3. A\Omega s I is symmetric if and only if A is. 4. 5. (A\Omega s 6. If A 0 and B 0 then (A\Omega s B) 0 7. --R Interior point methods in semidefinite programming with applications to combinatorial optimization Iterative methods for the computation of a few eigenvalues of a large symmetric matrix Solving large-scale sparse semidefinite programs for combinatorial optimization An implicitly restarted Lanczos method for large symmetric eigenvalue problems The minimization of certain nondifferentiable sums of eigenvalues of symmetric matrices Improved approximation algorithms for maximum cut and satisfiability problems using semidefinite programming Matrix Computations A shifted block Lanczos algorithm for solving sparse symmetric generalized eigenproblems Geometric Algorithms and Combinatorial Optimization Fixing variables in semidefinite relaxations Incorporating inequality constraints in the spectral bundle method An interior-point method for semidefinite programming Quadratic knapsack relaxations using cutting planes and semidefinite programming Convex Analysis and Minimization Algorithms I An interior-point method for minimizing the maximum eigenvalue of a linear combination of matrices Solving graph bisection problems with semidefinite programming Proximity control in bundle methods for convex nondifferentiable minimization Efficient approximation algorithms for semidefinite programs arising from MAXCUT and COLORING Connections between semidefinite relaxations of the max-cut and stable set problems --TR --CTR Jiahai Wang, Letters: An improved discrete Hopfield neural network for Max-Cut problems, Neurocomputing, v.69 n.13-15, p.1665-1669, August, 2006 Samuel Burer , Renato D. C. Monteiro , Yin Zhang, Interior-Point Algorithms for Semidefinite Programming Based on a Nonlinear Formulation, Computational Optimization and Applications, v.22 n.1, p.49-79, April 2002 Abraham Duarte , ngel Snchez , Felipe Fernndez , Ral Cabido, A low-level hybridization between memetic algorithm and VNS for the max-cut problem, Proceedings of the 2005 conference on Genetic and evolutionary computation, June 25-29, 2005, Washington DC, USA Gerald Gruber , Franz Rendl, The bundle method for hard combinatorial optimization problems, Combinatorial optimization - eureka, you shrink!, Springer-Verlag New York, Inc., New York, NY, Tijl De Bie , Nello Cristianini, Fast SDP Relaxations of Graph Cut Clustering, Transduction, and Other Combinatorial Problems, The Journal of Machine Learning Research, 7, p.1409-1436, 12/1/2006 Kazuhide Nakata , Makoto Yamashita , Katsuki Fujisawa , Masakazu Kojima, A parallel primal-dual interior-point method for semidefinite programs using positive definite matrix completion, Parallel Computing, v.32 n.1, p.24-43, January 2006 Hernn Alperin , Ivo Nowak, Lagrangian Smoothing Heuristics for Max-Cut, Journal of Heuristics, v.11 n.5-6, p.447-463, December 2005 Alper Yildirim , Xiaofei Fan-Orzechowski, On Extracting Maximum Stable Sets in Perfect Graphs Using Lovsz's Theta Function, Computational Optimization and Applications, v.33 n.2-3, p.229-247, March 2006 Stephen Braun , John E. Mitchell, A Semidefinite Programming Heuristic for Quadratic Programming Problems with Complementarity Constraints, Computational Optimization and Applications, v.31 n.1, p.5-29, May 2005 Henry Wolkowicz , Miguel F. Anjos, Semidefinite programming for discrete optimization and matrix completion problems, Discrete Applied Mathematics, v.123 n.1-3, p.513-577, 15 November 2002
proximal bundle method;eigenvalue optimization;semidefinite programming;convex optimization;large-scale problems
589295
Cones of Matrices and Successive Convex Relaxations of Nonconvex Sets.
Let F be a compact subset of the n-dimensional Euclidean space Rn represented by (finitely or infinitely many) quadratic inequalities. We propose two methods, one based on successive semidefinite programming (SDP) relaxations and the other on successive linear programming (LP) relaxations. Each of our methods generates a sequence of compact convex subsets Ck . .) of Rn such that (a) the convex hull of $F \subseteq C_{k+1} \subseteq C_k$ (monotonicity), (b) $\cap_{k=1}^{\infty} C_k = \text{the convex hull of F (asymptotic convergence). Our methods are extensions of the corresponding Lovsz--Schrijver lift-and-project procedures with the use of SDP or LP relaxation applied to general quadratic optimization problems (QOPs) with infinitely many quadratic inequality constraints. Utilizing descriptions of sets based on cones of matrices and their duals, we establish the exact equivalence of the SDP relaxation and the semi-infinite convex QOP relaxation proposed originally by Fujie and Kojima. Using this equivalence, we investigate some fundamental features of the two methods including (a) and (b) above.
Introduction . Consider a maximization problem with a linear objective function c T x: maximize c T x subject to x # F, where c denotes a constant vector in the n-dimensional Euclidean space R n and F a subset of R n . We can reduce a more general maximization problem with a nonlinear objective function f(x) to a maximization problem having a linear objective function represented by a new variable, x n+1 , if we replace f(x) by x n+1 and then add the inequality f(x) # x n+1 to the constraint. Thus (1.1) covers such a general optimization problem. Throughout the paper we assume that F is compact. Then the problem (1.1) has a global maximizer whenever the feasible region F is nonempty. For any compact convex set C containing F , the maximization problem maximize c T x subject to x # C serves as a convex relaxation problem, which satisfies the properties that (i) the maximum objective value # of the problem (1.2) gives an upper bound for the maximum objective value # of the problem (1.1), i.e., # , and # Received by the editors March 31, 1998; accepted for publication (in revised form) July 19, 1999; published electronically March 21, 2000. http://www.siam.org/journals/siopt/10-3/33645.html Department of Mathematical and Computing Sciences, Tokyo Institute of Technology, 2-12-1 Oh-Okayama, Meguro-ku, Tokyo 152-8552, Japan (kojima@is.titech.ac.jp). # Department of Combinatorics and Optimization, Faculty of Mathematics, University of Water- loo, Waterloo, Ontario N2L 3G1, Canada (ltuncel@math.uwaterloo.ca). This work was completed while this author was visiting Tokyo Institute of Technology, Department of Mathematical and Computing Sciences, on a sabbatical leave from University of Waterloo. The research of this author was supported in part by Tokyo Institute of Technology and by a research grant from NSERC of Canada. (ii) if a maximizer - lies in F , it is a maximizer of (1.1). Since the objective function of (1.1) is linear, we know that if we take the convex hull (defined as the intersection of all the convex sets containing F ) for C in (1.2), then (ii) # the set of the maximizers of (1.2) forms a compact convex set whose extreme points are maximizers of (1.1). Therefore, if we solve the relaxation problem (1.2) with a convex feasible region C which closely approximates c.hull(F ), we can expect to get not only a good upper bound # for the maximum objective value # but also an approximate maximizer of the problem (1.1). We can further prove that for almost every c # R n (in the sense of any maximizer x # of (1.2) is an extreme point of c.hull(F ), which also lies in F ; hence x # is a maximizer of (1.1). This follows from a result due to Ewald, Larman, and Rogers [5] for consequences of related results; see also [17]. Furthermore, for many representations of various convex sets C, given - x # C, we can very e#ciently find x # , an extreme point of C, such that c T x # c T - x. Indeed, the relaxation technique mentioned above has been playing an essential role in practical computational methods for solving various problems in the fields of combinatorial optimization and global optimization. It is often used in hybrid schemes with the branch-and-bound and branch-and-cut techniques in those fields. See, for instance, [2]. The aim of this paper is to present a basic idea on how we can approximate the convex hull of F . This is a quite di#cult problem, and also too general. Before making further discussions, we at least need to provide an appropriate (algebraic) representation for the compact feasible region F of the problem (1.1) and the compact convex feasible region C of the relaxation problem (1.2). We employ quadratic inequalities for this purpose. denote the set of n - n symmetric matrices and the set of positive semidefinite matrices, respectively. Given Q and # R, we write a quadratic function on R n with the quadratic term x T Qx, the linear and the constant term # as p(-; #, q, Q): Then the set Q of quadratic functions on R n and the set Q+ of convex quadratic functions are defined as and respectively. We also write p(-) # Q (or Q+ ) instead of p(-; #, q, Q) # Q (or Q+ ) are irrelevant. Throughout the paper, we assume that the feasible region F of the problem (1.1) is represented by a set of quadratic inequalities such that where PF denotes a set of quadratic functions, i.e., PF # Q, and we will derive convex relaxations, C, represented by convex quadratic inequalities such that where PC denotes a set of convex quadratic functions, i.e., PC # Q+ . We allow cases where PF and/or PC involve infinitely many quadratic functions. Thus (1.1) or (1.2) (or both) can be a semi-infinite quadratic optimization problem (QOP). Here we use the word "semi-infinite" for optimization problems having a finite number of scalar variables and possibly an infinite number of inequality constraints. There are some reasons why we have chosen quadratic inequalities for the representation of both problems, the maximization problem (1.1) that we want to solve and its convex relaxation problem (1.2). First, quadratic inequalities form a class of relatively easily manageable nonlinear inequalities, yet they have enough power to describe any compact feasible region F in R n . Indeed, if F is closed, then its complement R n \F is open so that it can be represented as the union of the open balls {x over all x # G for some G # R n We also know that any single polynomial inequality can be converted into a system of quadratic inequalities; for example, can be converted into Second, we know that we can solve some classes of maximization problems having linear objective functions and a convex-quadratic-inequality constrained feasible region C e#ciently. Among others, we can apply interior-point methods [1, 16] to the problem (1.2) when either PC is finite or PC is infinite, but its feasible region C is described as the projection of a set characterized by linear matrix inequalities in the space S n of n - n symmetric matrices onto the n-dimensional Euclidean space R n . Third, and also most importantly, we can apply the semidefinite programming (SDP) relaxation, which was originally developed for 0-1 integer programming problems by Lov-asz and Schrijver [12] and later extended to nonconvex quadratic optimization problems [6, 18, 19], to the entire class of maximization problems having a linear objective function and finitely or infinitely many quadratic inequality constraints. See also [1, 8, 9, 13, 15, 23, 24, 29]. In addition to the reasons above, we should mention that the maximization problem with a linear objective function and quadratic inequality constraints involves various optimization problems such as 0-1 integer linear (or quadratic) programming problems which, in principle, include all combinatorial optimization problems [1, 9, 18]. Linear complementarity problems [4], bimatrix games, and bilinear matrix inequalities [14, 20] are also included as special cases. For some optimization problems, some of the semidefinite programming (SDP) relaxations we provide may be solved in polynomially many iterations (of an interior-point method or an ellipsoid algorithm) approximately. Such conclusion requires, in the case of the ellipsoid method, the existence of a certain polynomial-time separation oracle for the underlying convex cone constraint (see [9]). In the case of interior-point algorithms (whose e#ciency in the theory and practice of SDP has been well established), we need to have an e#ciently computable self-concordant barrier for the feasible solutions set or at least for the underlying cone constraints (see [16]). Some of the most exciting activities in combinatorial optimization are currently centered around the applications of SDP to combinatorial optimization problems (see [7]). Such activity in theory and practice is fueled by theoretical results establishing that certain simple SDP relaxations of a combinatorial optimization problem can be e#ectively utilized in developing polynomial-time approximation algorithms with worst-case approximation-ratio guarantees much better than those previously proven using linear programming or other techniques. (See Goemans [7], Goemans and Williamson [8], Nesterov [15], and Ye [29].) Also outstanding are the results on the stable set problem establishing the fact that SDP techniques can be used in optimizing over a relaxation of the stable set polytope which is contained in the polytope defined by the clique inequalities. (Note that it is NP-hard to optimize over the latter-mentioned polytope, whereas Gr-otschel, Lov-asz, and Schrijver [9] and Lov-asz, and Schrijver [12] were able to utilize polynomial-time methods to achieve a better goal, as far as the proof of approximate optimality of some feasible solutions of the stable set problem is concerned.) Given an initial approximation C 0 of F , i.e., a compact convex set C 0 containing F , both of the methods, proposed in this paper, generate a sequence of compact convex subsets (a) It should be noted that the compactness of each C k and property (b) imply that (c) if (detecting infeasibility). To generate C k+1 at each iteration, the SDP relaxation and the linear programming relaxation play an essential role, and the entire method may be regarded as an extension of the Lov-asz-Schrijver lift-and-project procedure for 0-1 integer programming problems to semi-infinite nonconvex quadratic optimization problems, with the use of the SDP relaxation in the first method and the LP relaxation in the second method. The LP relaxation, referred to above, is essentially the same as the reformulation-linearization technique developed for nonconvex quadratic optimization problems by Sherali and Alameddine [21]; see also [2, 22]. However, we should caution the reader that the methods presented here are mostly conceptual in the general settings, because we need to solve a semi-infinite SDP (or a semi-infinite LP) at each iteration. For such a task, an e#cient practical algorithm may not be currently available. In their paper [6], Fujie and Kojima proposed the semi-infinite convex QOP relaxation for nonconvex quadratic optimization problems and showed that the semi-infinite convex QOP relaxation is not stronger than the SDP relaxation in general, but the two relaxations are essentially equivalent under Slater's constraint qualifica- tion. We establish the exact equivalence between the two relaxations for semi-infinite nonconvex quadratic optimization problems without any constraint qualification. Using this equivalence, we derive some fundamental features of our methods including (a) and (b) above. One of the common themes in this paper is the usage of cones of matrices (and duality) in our constructions. This was also one of the themes of [12]. The other themes of this paper are the successive applications of SDP relaxations and LP relaxations. We call the related procedures the successive SDP relaxation method and the successive semi-infinite LP relaxation method, respectively. Section 2 is devoted to preliminaries, where we provide some basic definitions and properties on quadratic inequality representations for closed subsets of R n , the homogeneous form of quadratic functions, the SDP relaxation, etc. In section 3, we present our first method in detail as well as the main results, including the features (a) and (b). After we present some fundamental characterizations of the SDP relaxation in section 4, we give proofs of the main results in section 5. In section 6, we apply our method to 0-1 semi-infinite nonconvex quadratic optimization problems. Incorporating the basic results on the lift-and-project procedure given by Lov-asz and Schrijver [12] for 0-1 integer convex optimization problems, we show that our method terminates in at most (n iterations either to generate the convex hull of the feasible region or to detect the emptiness of the feasible region, where n denotes the number of 0-1 variables of the problem. Section 7 contains our second method, which is based on semi-infinite LP relaxations. We establish the same theoretical properties as we do for the successive SDP relaxation method. In section 8, we present two numerical examples showing the worst-case behavior of some of our procedures. In particular, we know from the second example that the best of our procedures requires infinitely many iterations to generate the convex hull of F in the worst case. 2. Preliminaries. 2.1. Semi-infinite quadratic inequality representation. In this subsection, we discuss some representations of a closed subset F of R n in terms of (possibly infinitely many) quadratic inequalities. If p(-; #, q, Q) # Q, and p(x; #, q, Q) # 0 holds for all x # F , we say that p(x; #, q, Q) # 0 is a quadratic valid inequality for F and that p(-; #, q, Q) induces a quadratic valid inequality for F . A quadratic valid inequality p(x; #, q, Q) # 0 for F is linear and # R such that # a T x #x # F , rank-2 quadratic if and # R such that a T x # and b T x #x # F , spherical if ellipsoidal if convex quadratic if Q # S n respectively. It should be noted that if a quadratic valid inequality p(x; #, q, Q) # 0 for F is rank-2, then the rank of the matrix Q is at most 2 but that the converse is not necessarily true. We say that F has a (semi-infinite) quadratic inequality representation P # Q if holds. To designate the underlying representation P of F , we often write F (P) instead of F . Whenever F is a closed proper subset of R n , F has infinitely many represen- tations. We allow the cases where P consists of infinitely many quadratic functions. Hence can be a semi-infinite system of quadratic inequalities. If inequality representation of F and if p(-) # c.cone(P), then is a quadratic valid inequality, where c.cone(P) denotes the closed convex cone generated by P. Hence if P # P # c.cone(P), then P # is a quadratic inequality representation of F ; F inequality representation P of F is finite if it consists of a finite number of quadratic functions, and infinite otherwise. If F is a compact convex subset of R n , it has a quadratic inequality representation; in fact, the set of all the linear (rank-2 quadratic or spherical) valid inequalities for F forms an inequality representation of F . If, in addition, F is polyhedral, we can take a finite linear inequality representation. Let C be a compact subset of R n . We use the following symbols: the set of p(-)'s that induce linear valid inequalities for C, the set of p(-)'s that induce rank-1 quadratic valid inequalities for C, the set of p(-)'s that induce rank-2 quadratic valid inequalities for C, the set of p(-)'s that induce spherical valid inequalities for C, the set of p(-)'s that induce ellipsoidal valid inequalities for C, the set of p(-)'s that induce convex quadratic valid inequalities for C, the set of p(-)'s that induce all quadratic valid inequalities for C. By definition, we see that Note that if C is convex, then the equality holds with each these, P # (C) is the strongest quadratic inequality representation of C. 2.2. Homogeneous form of quadratic functions-lifting to the space of symmetric matrices. We introduce a di#erent description of quadratic functions, which we call the homogeneous form. This form leads us to a lifting of a quadratic function defined on the Euclidean space to the space of symmetric matrices and to the SDP relaxation (or to the semi-infinite LP relaxation in section 4.2). For every quadratic function p(-; #, q, Q) # Q, we connect the variable vector x # R n to the positive semidefinite matrix x and the triplet of the constant # R, q # R n , and Q # S n to the (1 . Then we have the identity p(x; #, q, x Thus, if P # Q is a quadratic inequality representation of F , then provides an equivalent representation of F ; Now we have two kinds of description for a quadratic function on R n : the usual form p(-; #, q, and the homogeneous form introduced above. The former is used in section 5, where we prove our main results, while the latter is suitable for the compact description of the SDP relaxation in section 2.3 and the proof of its equivalence to the semi-infinite convex QOP relaxation in section 4. We will use both forms in parallel, choosing whichever is convenient to us in a given situation. It should be noted that the correspondence is not only one-to-one but also linear. To save notation, we identify the set Q of quadratic functions with the set S 1+n of (1 any subset of Q with the corresponding subset of S 1+n . Specifically, we write identify the set of (1 symmetric matrices with the set Q of quadratic functions from R n to R. 2.3. SDP relaxation. Let P be a semi-infinite quadratic inequality representation The SDP relaxation - F (P) of F (P) with the quadratic inequality representation P is given by and and x and P . # 1 x T This implies that x # - F (P) and F (P) # - F (P). We also see that - F (P) is convex. Hence F (P). The SDP relaxation was originally proposed for combinatorial optimization problems and 0-1 integer programming problems [12], and later extended to quadratic optimization problems. See [1, 6, 8, 9, 15, 19, 18, 23, 24, 29]. 3. Main results. Now we are ready to describe our method for approximating a quadratic-inequality-constrained compact feasible region F of the minimization problem (1.1). Before running the method, we need to fix a semi-infinite quadratic inequality representation PF of F , and choose an initial approximation C 0 of the convex hull of F , i.e., a compact convex set which contains c.hull(F ). Starting from C 0 , the method generates a sequence of compact convex sets we expect to converge to c.hull(F ). At each iteration, we choose a semi-infinite quadratic inequality representation P k of the kth approximation C k of c.hull(F ). Since the union (PF # P k ) forms a semi-infinite quadratic inequality representation of F . We then apply the SDP relaxation to (PF # P k ) to generate the next iterate C It should be emphasized that during none of the iterations do we modify or strengthen the representation PF directly. We only utilize the semi-infinite quadratic inequality representation of the compact convex set C k that has been computed in the previous iteration. Successive SDP Relaxation Method. Step 0: Let Step 2: Choose a semi-infinite quadratic inequality representation P k for C k . Step 3: Let and Step 4: Let to Step 1. We state two convergence theorems below. We choose the spherical inequality representation at Step 2 of each iteration in the first theorem, while we choose the rank-2 quadratic inequality representation P 2 (C k ) for C k at Step 2 of each iteration in the second theorem. Their proofs will be given in section 5. Theorem 3.1. Assume that PF is a semi-infinite quadratic inequality representation of a compact subset F of R n , and that C 0 # F is a compact convex subset of R n . If we choose P of each iteration in the successive SDP relaxation method, then the monotonicity property (a) and the asymptotic convergence property (b) stated in the introduction hold. Theorem 3.2. Under the same assumptions as in Theorem 3.1, if we choose of each iteration in the successive SDP relaxation method, then (a) and (b) remain valid. We know that if P # Q and P # Q are semi-infinite quadratic inequality representations of C k and if F (P). Hence, even if we replace )" in Theorem 3.1 by "P k # P S (C k )" (or "P )" in Theorem 3.2 by "P k # P 2 (C k )"), the properties (a) and (b) remain valid. In particular, (a) and (b) remain valid when we choose any of P If we take the linear representation P L (C k ) of C k at every iteration, then we can prove that (See Lemma 4.1.) Hence (b) does not follow in general. In section 8, we will give two numerical examples. The first example shows that the rank-1 quadratic inequality representation strong enough to ensure (b). The second example shows that even when we choose the strongest quadratic inequality representation P # (C k ) of C k for P k at every iteration, not only does the convergence "C k # c.hull(F )" require infinitely many iterations, but its speed also becomes extremely slow in the worst case. 4. Fundamental characterization of successive convex relaxation. 4.1. Semi-infinite convex QOP relaxation and its equivalence to SDP relaxation. The semi-infinite convex QOP relaxation of F (P) with the semi-infinite quadratic inequality representation P is defined as We observe that F (P) and that the set - F (P) is a closed convex set. Hence F (P) # c.hull(F F (P). The semi-infinite convex QOP relaxation was introduced by Fujie and Kojima [6]. It was called the relaxation using convex-quadratic valid inequalities for F (P) in their paper [6]. The following basic properties of the relaxation are essentially due to them. Lemma 4.1. Let PF be a semi-infinite quadratic inequality representation of a closed set F # R n . Let P be a set of convex quadratic valid inequalities for F , i.e., Then Let P be a set of linear valid inequalities for F , i.e., (iii) Let x # c.hull(F ). Suppose that p(x #, q, Q) # 0 for some p(-; #, q, Q) # PF with a positive definite Q. Then x # - F (PF ). Proof. Part (i) follows directly from the definition of the semi-infinite convex QOP relaxation. Now we show (ii). Let we see that Hence it su#ces to show that - F (PF #P). Let p(-) # c.cone(P F #P)#Q+ . Then there exist p(-) i # PF positive m) such that are linear functions, we see that F (PF ). Moreover, Therefore, This proves (ii). Finally we will show (iii). Since x # F , there is a p # PF such that su#ciently small, we obtain that This implies x # - F (PF ), and proves (iii). When P is finite and F (P) satisfies Slater's constraint qualification, Fujie and Kojima [6] showed that the semi-infinite convex QOP relaxation is essentially equivalent to the SDP relaxation in the sense that - F (P) coincides with the closure of - F (P). The theorem below shows the exact equivalence between them, without any constraint qualification, for more general semi-infinite quadratic inequality representation cases. F (P) is closed, one of the consequences of the next theorem is that - F (P) is always closed. Note that we can assume without loss of generality that P is a closed convex cone, since every closed set F admits such a representation. Theorem 4.2. Let P be a closed convex cone, giving a semi-infinite quadratic inequality representation of a closed subset F of R n its SDP relaxation and its semi-infinite convex QOP relaxation coincide with each other; - F (P). Proof. Using the dual cone of P, we can express the sets - F (P) and - F (P) as follows: and # . For the last identity above, we have used the fact that for any pair of closed convex cones K 1 and K 2 in R m , we have (K 1 # K 2 First let x # - F (P). Then there exists an X # S n such that 760 MASAKAZU KOJIMA AND LEVENT TUNC-EL Consider the identity -x -X The first matrix on the right-hand side is in P # and in the second matrix of the right-hand side, we have X - xx T since it is the Schur complement of 1 in the symmetric, positive semidefinite matrix # 1 x T We have proved x # - F (P) and hence - F (P). For the converse, let x # - F (P); that is, there exists some H # S n such that The matrix is positive semidefinite if and only if (H is. But the latter was already established. So, . Therefore x # - F (P), and - F (P) is proved. 4.2. Semi-infinite LP relaxation. In section 7, we will also need an analog of the above theorem for our successive semi-infinite LP relaxation method. For every semi-infinite quadratic inequality representation P of a compact subset F of R n , let us define and of Sherali and Alameddine [21]. Here, L denotes the set of linear functions on The next result can be obtained by following the steps of the proof of Theorem 4.2. Corollary 4.3. Let P be a closed convex cone, giving a semi-infinite quadratic inequality representation of a closed subset F of R n F L (P). Proof. We observe that and Since it is easy to see that #X # S n such that # 1 x T only if the proof is complete. 4.3. Invariance under one-to-one a#ne transformation. Let b be an arbitrary one-to-one a#ne transformation on R n , where A is an n - n non-singular matrix and b # R n . Then of f(F (P)). This means that the semi-infinite SDP and LP relaxations are invariant under the one-to-one a#ne transformation We also see that holds, where U # {L, 1, 2, E, C, #}. Therefore, P L (C), are invariant under one-to-one a#ne transformations on R n . If in addition A is a scalar multiple of an orthogonal matrix, then the above identity also holds for is invariant under such a one-to-one a#ne transformation on R n . At each iteration of the successive SDP relaxation method, we observe that forms a semi-infinite quadratic inequality representation of f(F ) and P # inequality representation of f(C k ). Furthermore, if we choose one of the invariant semi-infinite quadratic inequality representations P L (C k ), under any one-to-one a#ne transformation for P k , we see that P U hence the identity above turns out to be Here U # {L, 1, 2, E, C, #}. Therefore the successive SDP relaxation method is invariant under any one-to-one a#ne transformation. The same comment applies to the successive semi-infinite LP relaxation method, which we will present in section 7. 762 MASAKAZU KOJIMA AND LEVENT TUNC-EL 5. Proofs of Theorems 3.1 and 3.2. We present three lemmas, Lemma 5.1 in section 5.1, Lemma 5.2 in section 5.2, and Lemma 5.3 in section 5.4. Lemma 5.1 proves the monotonicity property (a) in Theorems 3.1 and 3.2 simultaneously. Lemma 5.2 is used to prove Theorem 3.1 in section 5.3, and Lemma 5.3 to prove Theorem 3.2 in section 5.5. 5.1. Monotonicity. We first establish the monotonicity in general. Lemma 5.1. Let C 0 be a compact convex set containing F . Fix a closed convex cone S 1+n # K and Assume that # K and Proof. Since K # S 1+n contains all symmetric rank-1 matrices of the form Now, as in the arguments in section 2.3, it follows that c.hull(F We will show by induction that C k+1 # C k for all the construction of C 1 and the assumption imposed on C 0 , we first observe that Now assume that C k # C k-1 for some k # 1. Then P U (C k-1 which implies that PF # P U (C k-1 desired. 5.2. Separating hypersphere. The following lemma easily follows from the separating hyperplane theorem, and the proof is omitted here. Lemma 5.2. Let C be a compact convex subset of R n and x # C. Then there exists a hypersphere S # {x # R n : #x-d#} which strictly separates the point x # and C such that where d # R n and # > 0. 5.3. Proof of Theorem 3.1. The monotonicity property (a) follows from Lemma 5.1 by letting K # S 1+n and U # S. Let C # k=0 C k . We know by (a) that that all the sets c.hull(F ), C, and C k are compact sets. To prove (b), we have the following left to show: C # c.hull(F ). Assume on the contrary that there exists some x # C such that x # c.hull(F ). Then, by Lemma 5.2, there exists a hypersphere S # {x # R strictly separates the point x # C from c.hull(F ) such that there is a quadratic function, cuts o# x Q). Note that if p 1 (-; #, q, Q) is such a quadratic function, then so is #p 1 (-; #, q, Q) for any # > 0. Hence we may assume that the minimum eigenvalue of the matrix Q # S n is at least (-1). Now consider a quadratic function p 2 (-) defined by By the definition of # , we see that This means that the open ball B+ # {x # R with the center d and the a neighborhood of the compact set C. On the other hand, the sequence {C k } of compact subsets of R n satisfies So, we can find a finite positive number # such that the open ball B+ contains C # . Hence, We also see that Thus we have shown that Therefore, x # C . This is a contradic- tion. The theorem is proved. 5.4. A family of inequalities of the convex cone of rank-2 quadratic valid inequalities for the unit ball. Let B denote the unit ball {x # R 1}. Let Q be an arbitrary n - n symmetric matrix, and let u # R n be an arbitrary vector on the boundary of B; #u# = 1. We will construct a family of quadratic valid inequalities, which lie in the convex cone of rank-2 quadratic valid inequalities, with a parameter # (0, #/8) for the unit ball B satisfying the properties (i), (ii), and (iii) listed in Lemma 5.3. We first apply the eigenvalue decomposition to the matrix Q # S n . We may assume that the first m eigenvalues are nonnegative and the last are nonpositive for some nonnegative integer m # n. Then we can write the matrix denote eigenvectors of Q, which are orthogonal to each other, and - j (j = 1, 2, . , m) and - j (j = m+ 1, . , n) denote the eigenvalues corresponding to them. For each # (0, #/8), we define a a are nonzero vectors, and a are linear valid inequalities for the unit ball B. For all # (0, #/8), define In particular, p # (u) # 0 # (0, #/8). Lemma 5.3. tends to 0. (iii) The Hessian matrix of p # (-) coincides with -Q. Proof. Part (i) was already shown. (ii) Let j be fixed. It su#ces to show that converge to zero as # (0, #/8) tends to 0. First, we derive that # j (#) converges to zero as # (0, #/8) tends to 0. We see from (5.2) that sin #) sin # (b T sin # (b T Since both the numerator and the denominator above converge to zero as # (0, #/8) tends to 0, we calculate their derivatives at The derivative of the numerator turns out to be (b T which vanishes at On the other hand, the derivative "2 cos #" of the denominator "2 sin #" in (5.5) does not vanish at converges to 0 as # (0, #/8) tends to 0. Similarly, we can prove that - # j (#) converges to 0 as # (0, #/8) tends to 0. (iii) It follows from the definitions (5.2) and (5.4) that the Hessian matrix of the quadratic function p # (-) a a j (#) T= - From the lemma above, we see that the cone rich enough to contain rank-2 quadratic functions with any prescribed Hessian, leading to valid inequalities that are tight at any given point on the boundary of B. 5.5. Proof of Theorem 3.2. The monotonicity property (a) follows from Lemma 5.1 by letting K # S 1+n 2. To derive (b), it su#ces to show that C # k=0 C k # c.hull(F ) as in the proof of Theorem 3.1. Assume on the contrary that x # c.hull(F ) for some x # C. By Lemma 5.2, there exists a hypersphere strictly separates the point x # C and c.hull(F ) such that the successive SDP relaxation method using the rank-2 quadratic representation for C k at each iteration is invariant under the a#ne transformation maps d to the origin and the hypersphere onto the unit hypersphere {x # R may assume that 1. Thus, we have obtained that Since u # F , there is a quadratic function p 1 (-; #, q, Q) # PF that cuts o# u; be the quadratic function introduced in section 5.4. See (5.2) and (5.4). By Lemma 5.3, we can choose a # (0, #/8) for which p # (u) # -p 1 (u; #, q, Q)/3 holds. Now we define 766 MASAKAZU KOJIMA AND LEVENT TUNC-EL By construction, we know that p # k (-) # c.cone(P 2 (C k )). Since both quadratic functions p # (-) and p # k (-) have the common Hessian matrix -Q, We will show that for every su#ciently large k. Then the above two relations imply u # C k+1 for such a large k. This contradicts the fact Since the sequence of compact convex subsets we see that as k # (j = 2, 3, . , n). By continuity, we see then that for every su#ciently large Thus we have shown that (5.6) holds for every su#ciently large k. This completes the proof of Theorem 3.2. 6. Application to 0-1 semi-infinite, nonconvex quadratic optimization problems. We briefly recall two of the Lov-asz-Schrijver procedures for 0-1 integer programming problems, and relate them to our successive SDP relaxation method. Let F be a subset of {0, 1} n whose convex hull is to be approximated. In the Lov-asz- Schrijver procedures, we assume that a compact convex subset C 0 of R n satisfying is given in advance. We define Let K I denote the convex cone spanned by the 0-1 vectors in K Here the 0th coordinate is special. It is used in homogenizing the sets of interest in R n . Clearly The closed convex cone K 0 serves as an initial relaxation of K I . Given the current relaxation K k of K I , first a convex cone M+ (K k , K k ) in the space of (1 symmetric matrices is defined (the lifting operation). Then a projection of this cone gives the next relaxation N+ (K k ) of K I . Now, we define the lifting operation in general. Let K and T be closed convex cones in R 1+n . A (1 real entries is in (This condition is equivalent to Y K # T .) Here, e 0 denotes the unit vector with 0th coordinate 1. Item (ii) above serves an important role in Lov-asz-Schrijver procedures as well as in some of the SDP relaxations used by Goemans and Williamson [8], Nesterov [15], and Ye [29]. This equation is valid simply because for each j for which x j # {0, 1}, the equation x 2 is valid. Indeed, our general framework applies to any compact set in R n , and the equation Y e not utilized in earlier sections (as it is not valid). In this section, however, the equation is valid and we utilize it. As will be noted in the proof of Theorem 6.3, the inclusion of this equation will be guaranteed by our choice of the initial formulation. The third condition of Lov-asz-Schrijver procedures is very interesting. They present a couple of possibilities for the choice of cone T in 0-1 integer programming. Among them is the cone spanned by all 0-1 vectors with the first component x This choice, since the cone T # has a very simple set of generators, allows for the development of polynomial-time algorithms for approximately solving the successive SDP relaxations as long as the number of iterations of the successive procedure is O(1). Their result only assumes that a polynomial-time weak separation oracle is available for K. The key is that since T # has only O(n) extreme rays, it becomes trivial to check condition (iii) in polynomial time. On the other hand, Lov-asz and Schrijver [12] note that the choice T # K is also possible and leads to at least as good relaxations as the former choice for T . (In many cases the successive relaxations for are significantly tighter than the successive relaxations with the simpler choice of T .) In the case of the latter choice, the possibility of polynomial-time solvability of the first few successive relaxations depends on the availability of polynomial-time algorithms to check Y K # K. Our procedure uses T # K. Now, we describe the projection step. We also define the iterated operators N k use the notation N+ (K), whereas N+ (K, K) is used in [12].) Another procedure studied in [12] uses a weaker relaxation by removing the condition (i) in the lifting procedure. Let M(K,K) and N(K) denote the related sets for this procedure. We will refer to the first procedure using the lifting M+ (K, K) (and the projection N+ ) as the N+ procedure. We will call the other (using M(K,K), and N) the N procedure. Lov-asz and Schrijver prove the following. Theorem 6.1. and 768 MASAKAZU KOJIMA AND LEVENT TUNC-EL Let us see how our successive SDP relaxation method applies to 0-1 nonconvex quadratic optimization problems. Consider a 0-1 nonconvex quadratic program: subject to x # F # {x # {0, 1} We may assume that the set P # contains the quadratic functions x 1, 2, . , n. Then we can replace the 0-1 constraint imposed on the variable x i by the inequality -x i by adding the quadratic functions -x 1, 2, . , n, to P # , we obtain a quadratic inequality representation PF of the feasible region F . Let C 0 # [0, 1] n . Note that F #= C 0 #{0, 1} our general setting here. However, has been assumed for some compact convex subset C 0 of R n in the Lov-asz-Schrijver procedures discussed above. Lemma 6.2. Suppose that we take C and Proof. Let C # 1 be the semi-infinite convex QOP relaxation of the set F with the quadratic inequality representation PF # In view of Theorem 4.2 and Lemma 5.1, we know that Hence it su#ces to show that If F contains all the 0-1 vectors, the inclusion relation above obviously holds. Now assume that x # F is a 0-1 vector. Then there is a quadratic function p 1 (-, #, q, Q) # PF such that On the other hand, we know that the quadratic function with the identity matrix as its Hessian matrix, is a member of c.cone(P 0 ), and that Hence if # > 0 is su#ciently small, then This implies that x # C # As a consequence of the lemma above, we see that the 0-1 nonconvex quadratic optimization problem (6.1) is equivalent to the 0-1 convex quadratic optimization problem subject to x # Using this observation, we can prove that in the case of 0-1 nonconvex quadratic optimization problem (6.1), our successive SDP relaxation method converges in (1+n) iterations. Theorem 6.3. The successive SDP relaxation method, applied to the 0-1 non-convex quadratic optimization problem (6.1), using C as the initial approximation of c.hull(F ) and in each iteration, terminates in at most (1 +n) iterations with Proof. We note that by Lemma 6.2, after one iteration of the successive SDP relaxation method, we obtain the 0-1 convex quadratic optimization problem (6.2) that can be used with the original Lov-asz-Schrijver procedure. We only have to note that the successive SDP relaxation method becomes the Lov-asz-Schrijver procedure after the first iteration. For this purpose, we compare conditions (i), (ii), and (iii) of the Lov-asz-Schrijver procedure for to the conditions used to construct in the successive SDP relaxation method. Here First, we observe that #X # S n such that Y # 1 x T if and only if # 0, . Hence (i) is satisfied. For (ii), note that implies the constraint Y e implies Y e 0 # Diag(Y ). Finally, for (iii), note that a linear inequality a T x # is valid for C k if and only if (#, -a T Therefore, we see that Step (3.1) of the successive SDP relaxation method implies that . Thus, we conclude by noting that Theorem 6.1 implies that n more steps of the procedure is su#cient. The above discussion and the results show that our successive SDP relaxation method generalizes the Lov-asz-Schrijver N+ procedure by ignoring condition (ii), which is no longer valid. Our results in the previous sections already showed that in this full generality, we still have the asymptotic convergence of the method. It is therefore interesting to investigate the same questions about the weaker procedure . What is the generalization of procedure N? . Does the generalization of procedure N satisfy the same theoretical properties as the successive SDP relaxation method? We answer both of these questions in the next section. As is shown in [12], in some cases the procedure N+ is significantly better than N . Procedure N is weaker, but the relaxations given by it are always polyhedral sets (so LP techniques can be employed) and N+ requires more general techniques. Hence, sometimes procedure N might be more manageable even if the procedure N+ is not. We should expect that the generalization of procedure N should be only using condition (iii), Y K # K, in the definition of the lifting. We would also expect that the generalization should lead to semi-infinite LP (rather than SDP) relaxations. We show in the next section that the above-mentioned generalization of procedure N leads to successive semi-infinite LP relaxations and all the analogs of the theoretical properties established for our successive SDP relaxations can also be established for the successive semi-infinite LP relaxations. 7. Successive semi-infinite LP relaxation. Successive Semi-Infinite LP Relaxation Method. Step 0: Let Step 2: Choose a quadratic inequality representation P k for C k . Step 3: Let #X # S n such that (The equalities above follow from Corollary 4.3.) Step 4: Let to Step 1. Theorem 7.1. Assume that PF is a semi-infinite quadratic inequality representation of a compact subset F of R n , and that C 0 # F is a compact convex subset of R n . If we choose P of each iteration in the successive semi-infinite LP relaxation method, then the monotonicity property (a) and the asymptotic convergence property (b) stated in the introduction hold. Proof. We can apply the same proof as the one given for Theorem 3.2 in section 5.5 to the theorem. Note that we can define another semi-infinite LP relaxation based on the semi-infinite convex QOP relaxation. Clearly, if Q # S n So, we can define a semi-infinite LP relaxation based on the above observation: F L q O and F L q O In this case, the equivalence - F L F L is evident. The convergence of the successive semi-infinite LP relaxation method using - F L can be established by following the proofs of Theorems 3.1 and 3.2. Instead, we note - F L F L . Therefore, Theorem 7.1 also implies that this particular semi-infinite LP relaxation method has the properties (a) and (b) mentioned in the theorem. 8. Further discussions on successive convex relaxations. 8.1. Conic quadratic inequality representation. The conic quadratic inequality presented below is a generalization of the linear matrix inequality [3, 28] and the bilinear matrix inequality [14, 20]. It will be shown that any conic quadratic inequality can be reduced to a semi-infinite system of standard quadratic inequalities and vice versa. Let K and K K} be a closed convex cone in R m and its dual. Here u - v denotes an inner product of u . For all lies in K. Now we introduce a conic quadratic vectors in R m . We may assume without loss of generality that . The inequality (8.1) turns out to be a system of m usual quadratic inequalities on R n if we take the nonnegative orthant R m for the cone K. The inequality (8.1) turns out to be a quadratic matrix inequality, which is a generalization of linear and bilinear matrix inequalities [3, 28] if we identify the space of # symmetric matrices with R m and we take the positive semidefinite of matrices for the cone K, where We can rewrite the conic quadratic inequality (8.1) as a semi-infinite system of standard quadratic inequalities in the homogeneous form. for some P # . This means that we can easily include any conic quadratic inequality in the semi-infinite quadratic inequality representation of the feasible region F of the maximization problem (1.1). To see the equivalence between (8.1) and (8.2) for some P # we observe that (8.1) can be rewritten as Therefore, if we define 772 MASAKAZU KOJIMA AND LEVENT TUNC-EL we obtain the desired semi-infinite system (8.2) of standard quadratic inequalities, which is equivalent to (8.1). Let F (P) denote the solution set of (8.2) with its quadratic inequality representation Applying the SDP relaxation to F (P), we obtain that and and and The set in the last line corresponds to the SDP relaxation to the solution set of (8.1). This implies that we can apply the SDP relaxation directly to the conic quadratic inequality converting it into the semi-infinite system (8.2) of standard quadratic inequalities. Conversely, we can reduce any semi-infinite system of standard quadratic inequalities to a conic quadratic inequality. To show this, consider a semi-infinite system (8.2) of standard quadratic inequalities in the homogeneous form. We may assume without loss of generality that P # S 1+n is a closed convex cone. We can rewrite (8.2) as x which is a conic quadratic inequality. Let F denote the solution set of the conic quadratic inequality (8.3) that we have derived from (8.2) above. Applying the SDP relaxation to F , we obtain that and # . Note that the set in the last line corresponds to the SDP relaxation of the solution set of the semi-infinite system (8.2) of standard quadratic inequalities. In view of the discussions above, we know that the conic quadratic inequality representation is as general as the semi-infinite quadratic inequality representation and that the SDP relaxations to both representations are equivalent. When we deal with the semi-infinite convex QOP relaxation, however, the semi-infinite quadratic inequality representation seems more convenient than the conic quadratic inequality representation. 8.2. A counterexample to the convergence for the rank-1 quadratic inequality representation case. The example below shows that the rank-1 quadratic inequality representation is not strong enough to ensure the convergence of the successive SDP relaxation method. Let where denotes the rank-1 quadratic inequality representation of the unit ball, which consists of all quadratic functions such that (a T x - 1)(a see that Theorem 8.1. Suppose that we take P representation of C k ) in the successive SDP relaxation method applied to the example above. Then C Proof. By definition, C which su#ces to establish the theorem. First observe that C 1 # B. Hence it su#ces to show equivalently for all p(-) # c.cone(P F ) # Q+ , fixed. Then we can choose # i # 0 #) such that Now assume that # 0 > 0. In this case, we may further assume without loss of generality that # It follows from p(-) # Q+ that the Hessian matrix # I # is positive semidefinite. Hence if we denote the largest and the smallest eigenvalues of the matrix We also see that Hence 8.3. A counterexample to the finite termination for the strongest quadratic inequality representation case. The example below shows that in the worst case, even when we take the strongest quadratic inequality representation P # (C k ) for C k at every iteration, . the successive SDP relaxation method requires infinitely many iterations, and . the convergence is extremely slow. For every 5. Then Theorem 8.2. Suppose that we take P strongest quadratic inequality representation of C k ) in the successive SDP relaxation method applied to the example above. (i) C k is symmetric with respect to the x 2 axis: only if (-x 1 , x 2 (ii) Let Then Proof. We will prove (i) and (ii) by induction. (i) Obviously the assertion is true for Assume that C k is symmetric with respect to the x 2 axis. Then we know that This ensures that C k+1 is symmetric with respect to the x 2 axis. (ii) By definition, we know that # 2. Hence (8.4) holds for Assuming that (8.4) holds, we prove that (8.5) holds. We first observe that It su#ces to show that (0, - # C k+1 or equivalently Assume on the contrary that such that Here we remark that p 1 (-) can be incorporated into p # (-) since p 1 (-) # P k . By the symmetry with respect to the x 2 axis, we see that Thus, defining - we obtain that It follows from p # and the third inclusion relation of (8.6) that p # (0, - We may further assume without loss of generality that redefine all the relations above remain valid. Since - we see that Q 11 # 1 and Q 22 # 1. By (8.6) and hence 776 MASAKAZU KOJIMA AND LEVENT TUNC-EL Therefore, by the convexity of the quadratic function - p(x), we obtain that This contradicts (8.7). The above example is simple, yet it illustrates great di#culties for the successive relaxation method. For example, # k+1 /# k # 1. Therefore, the convergence is slower than linear. Note that, in any dimension, if we take a pair of ball constraints, one convex (inclusion), the other nonconvex (exclusion), then both of the successive SDP and semi-infinite LP relaxation methods stop in one iteration, returning the convex hull of the intersection. Also, in the above example, if we knew that p 2 (-) a#ects only the definition of F in the region x 1 # 0 and that p 3 (-) is only e#ective in the region we could do elementary modifications to the method to speed up convergence tremendously. This is a good elementary example to illustrate the fact that for such methods to become more e#cient in practice, hybrid approaches including branch- and-bound and branch-and-cut seem necessary. We make further remarks in the next section. 9. Concluding remarks. We propose extensions of two fundamental lift-and- project procedures N and N+ of Lov-asz and Schrijver [12]. The original procedures were proposed for 0-1 integer programming problems to compute the convex hull of feasible (integer) solutions. Our procedure applies to any nonconvex region and as a result we do not use the key equations, Y e used in N and N+ procedures. Therefore, our relaxations are based either on two conditions: Y is positive semidefinite and Y K # K (successive SDP relaxation method), or on only one condition: Y K # K (successive semi-infinite LP relaxation method). In both cases we established the properties (a) monotonicity and (b) asymptotic convergence. The weakest version of our procedures satisfying the properties (a) and (b) uses only rank-2 quadratic valid inequalities. We showed in section 6 that such inequalities ensure the condition Y K # K. Finally, in section 8 we showed that even the strongest of such relaxation procedures (using all quadratic valid inequalities) uses infinitely many iterations to converge. In the above sense, the strongest positive result is given in section 7 by the successive semi-infinite LP relaxation method based on rank-2 valid inequalities. On the one hand, theoretically speaking, the best results are given in section 7: the weakest algorithm achieving the strongest results. Moreover, the successive semi-infinite LP relaxation method is more likely to be practical for a given general problem. On the other hand, the relative value of SDP relaxations has been quite impressive so far on some very special problems (e.g., the stable set problem [12]) and less impressive on others (e.g., the matching problem [25]). Therefore, one interesting research direction is to search for interesting classes of nonconvex sets for which the successive SDP relaxation method is significantly better than the successive semi-infinite LP relaxation method. For the same reason, (partial) characterizations of nonconvex sets on which both methods perform comparably are also important. Our convergence proofs are by contradiction, but the main argument is about cutting o# a point using valid inequalities induced by the underlying construction. The strongest convergence result (for the weakest algorithm) uses separating hyper- spheres. In the other proofs, for the bad points, the separating hyperspheres may have huge radii and converge to hyperplanes. However, for certain points and shapes, the advantage of using more general convex quadratic inequalities is clear. This discussion motivates us to suggest another avenue for research. It would be interesting to find certain invariants and measures of the input of our procedures that lead to nontriv- ial, descriptive convergence rates for our methods, perhaps only for some interesting subclass of problems. Recently, Kojima and Takeda [11] discussed the computational complexity of the successive SDP and semi-infinite LP relaxation methods. They gave an upper bound on the number of iterations which the methods require to attain a convex relaxation of a quadratically constrained compact set F with a given accuracy # > 0, in terms of #, the diameter of the initial relaxation C 0 , the diameter of F , and some other quantities characterizing the Lipschitz continuity and the nonconvexity and nonlinearity of the quadratic inequality representation PF of F . The major di#culty in implementing the idea of the successive SDP (or semi-infinite relaxation method in practice is the solution of a continuum of semi-infinite SDPs (or semi-infinite LPs) to generate a new approximation C k+1 of the convex hull of the feasible region F of a nonconvex quadratic program at each itera- tion. In their succeeding paper [10], the authors propose implementable variants by introducing two new techniques, a discretization technique for approximating continuum of semi-infinite SDPs (or semi-infinite LPs) by a finite number of standard SDPs (or LPs) with a finite number of linear inequality constraints, and a localization technique for generating a convex relaxation of F that is accurate only in certain directions in a neighborhood of the objective direction c. They established that, Given any positive number #, there is an implementable discretized-localized variant of the successive SDP (or semi-infinite LP) relaxation method which generates an upper bound of the objective values within # of their maximum in a finite number of iterations. See also [27] for a practical implementation of this variant and some numerical results. --R Interior point methods in semidefinite programming with applications to combinatorial optimization Linear Matrix Inequalities in System and Control Theory The Linear Complementarity Problem The directions of the line segments and of the r-dimensional balls on the boundary of a convex body in Euclidean space Improved approximation algorithms for maximum cut and satisfiability problems using semidefinite programming Discretization and Localization in Successive Convex Relaxation Methods for Nonconvex Quadratic Optimization Problems Complexity Analysis of Conceptual Successive Convex Relaxation of Nonconvex Sets 2nd rev A cone programming approach to the bilinear matrix inequality problem and its geometry On the Generic Properties of Convex Optimization Problems in Conic Form A recipe for semidefinite relaxation for (0 An Algorithmic Analysis of Multiquadratic and Semidefinite Programming Problems Control system synthesis via bilinear matrix inequalities A new reformulation-linearization technique for bilinear programming problems A reformulation-convexification approach for solving nonconvex quadratic programming problems Systems Sci. Dual quadratic estimates in polynomial and boolean programming On a representation of the matching polytope via semidefinite liftings Convexity and Optimization in Finite Dimensions I Towards the implementation of successive convex relaxation method for Approximating quadratic programming with bound and quadratic constraints --TR --CTR Masakazu Kojima , Levent Tunel, Some Fundamental Properties of Successive Convex Relaxation Methods on LCP and Related Problems, Journal of Global Optimization, v.24 n.3, p.333-348, November 2002 Akiko Takeda , Katsuki Fujisawa , Yusuke Fukaya , Masakazu Kojima, Parallel Implementation of Successive Convex Relaxation Methods for Quadratic Optimization Problems, Journal of Global Optimization, v.24 n.2, p.237-260, October 2002 Mituhiro Fukuda , Masakazu Kojima, Branch-and-Cut Algorithms for the Bilinear Matrix Inequality Eigenvalue Problem, Computational Optimization and Applications, v.19 n.1, p.79-105, April 2001 Henry Wolkowicz , Miguel F. Anjos, Semidefinite programming for discrete optimization and matrix completion problems, Discrete Applied Mathematics, v.123 n.1-3, p.513-577, 15 November 2002
nonconvex quadratic optimization problem;global optimization;semi-infinite programming;linear matrix inequality;bilinear matrix inequality;semidefinite programming
589300
Global Error Bounds for Convex Conic Problems.
This paper aims at deriving and proving some Lipschitzian-type error bounds for convex conic problems in a simple way. First, it is shown that if the recession directions satisfy Slater's condition, then a global Lipschitzian-type error bound holds. Alternatively, if the feasible region is bounded, then the ordinary Slater condition guarantees a global Lipschitzian-type error bound. These can be considered as generalizations of previously known results for inequality systems, which also follow from general results by Bauschke and Borwein in [ SIAM Review, 38 (1996), pp. 367--426] and Bauschke, Borwein, and Li in a 1997 report. However, the proofs in the current paper are considerably simpler. Some of the results are generalized to the intersection of multiple shifted cones (with different shifts). Under Slater's condition alone, a global Lipschitzian-type error bound does not hold. It is shown, however, that such an error bound holds for a specific conic region. For linear systems we establish that the sharp constant involved in Hoffman's error bound is nothing but the condition number for linear programming as used by Vavasis and Ye in [ Math. Programming, 74 (1996), pp. 79--120].
Introduction In optimization theory it is often desirable to measure the distance to the solution set from a certain given point. In general, this distance can be difficult to assess, since one may not have a complete knowledge about the solution set. However, if the form of the solution set is explicitly given, then in some cases it is possible to estimate the distance to the solution set by the so-called constraint violation which is computable. This kind of estimation is termed error bound relation. The first such result was obtained by Hoffman [7] for systems of linear equalities and inequalities. We shall discuss Hoffman's error bound in the paper too. A recent extensive survey on various types of error bound results can be found in Pang [15]. Most papers discussing error bound results assume that the solution set is given by equations and inequalities, e.g. For a given point x the amount of constraint violation can be measured as the following quantity with the notation (y) A measure for constraint violation is similar to a penalty function in the sense that it takes positive value for points outside the set, and zero otherwise. Note that a measure for constraint violation should be easy computable, such as the case for the above defined function v(x). Hoffman's lemma [7] states that if S 6= ;, and f i and g j are all affine linear functions, then there is a positive dist (x; S) -v(x) (1.1) for all x 2 ! n . This means that the distance to S is of the same magnitude as v(x). Such a relation is known as a Lipschitzian type error bound. In the case that f i and g j are not linear, the above inequality (1.1) does not hold in general. Early results concerning nonlinear functions are due to Robinson [17] and Mangasarian [13]. Robinson [17] showed that for inequality systems if all functions are convex and differentiable, S is bounded and the Slater condition holds, i.e. there is a - x such that g j (-x) ! 0 for all j, then relation (1.1) holds. Mangasarian [13] removed the assumption that S is bounded by assuming an additional asymptotic constraint qualification condition, which however can be difficult to verify in general. In this paper we consider the following convex conic set: is a subspace of ! n and K ' ! n is a closed convex cone. Polynomial-time interior-point algorithms for solving convex optimization problems with convex conic feasible set were introduced in a systematic manner by Nesterov and Nemirovskii [14]. It turns out that many important classes of optimization problems, such as linear programming and semidefinite programming, can be cast in this form. The focus of this paper is to discuss how error bound type relation can be established for such problems. Throughout this paper we make the following assumption: Assumption 1 F 6= ;. The organization of the paper is as follows. In the next section we prove that with a proper definition of constraint violation a Lipschitzian type error bound (1.1) can be established for general convex conic problems, under various conditions on the relations between L and K, including Slater type conditions. In Section 3 we discuss a link between the constant in Hoffman's error bound and the so-called condition number for linear programming. Finally, we conclude the paper in Section 4. We use the following notation in this paper. Matrices are denoted by capital letters, e.g. X. For indicates the maximum eigenvalue of X, and - min (X) the minimum eigenvalue of X. We denote n-dimensional Euclidean space by ! n and its nonnegative quadrant by . The space of all symmetric n by n matrices is denoted by S n\Thetan and the cone of all symmetric positive semidefinite n by n matrices by S n\Thetan . Vector e represents a vector of all ones with appropriate dimension. For a vector we use the capitalized letter V to denote the diagonal matrix which takes v as its diagonal elements. For two vectors as the component-wise Hadamard product. We use the Euclidean norm for vectors and the spectral norm for matrices. A vector a - 0 means that each component of a is nonnegative, and X - 0 indicates that X is positive semidefinite. systems Consider the convex conic set (1.2). For convenience we further assume that K is a pointed and solid cone, i.e. K " The dual of K is Since K is pointed and solid, K too is a closed, convex, pointed and solid cone. An immediate next question is: How can we define a constraint violation function for F? For this purpose we note the following lemma due to Moreau (see Theorem 31.5 in [18]). Lemma 2.1 For any x 2 ! n there is a unique x p 2 K and x d 2 K such that In fact, x p is simply the projection of x onto K and kx d k measures the distance from x to K. A natural definition for the constraint violation for F is now in order: Definition 2.1 For any x 2 ! n define dist as the constraint violation function for F . It is readily seen that v(x; It is, however, not immediately clear how the function v(x; F) can be computed. Below we shall see some examples in which this function is explicitly derived. First we consider the case , the nonnegative quadrant of ! n . Clearly, which is exactly the usual definition of the violation for nonnegativity constraints. Another example is , the cone of n by n symmetric positive semidefinite matrices. Consider a given n by n symmetric matrix X. Following Lemma 2.1 we know that there is unique positive semidefinite matrices X p and X d such that and X d can be computed as follows. Let an orthonormal matrix and is a diagonal matrix with eigenvalues of X as its components. Splitting \Gamma denote the nonnegative and nonpositive parts of respectively, it follows that X denotes the minimum eigenvalue of X. Finally, we consider another popular convex cone: the second order cone K defined as It can be shown that in this case 2: In general, Definition 2.1 is only related to the geometry of the object under consideration. Consider now an arbitrary point z Assume that z 62 F . The following problem yields a unique point in F with the shortest Euclidean distance to z: subject to x Let this optimal solution be - x. The Karush-Kuhn-Tucker optimality condition for (Proj) is given as follows: Hence, where the first inequality follows from the fact that z p 2 K and - 2 K . Let the projection of z onto the affine subspace b + L be z l . Then, dist Substitute this relation into (2.2) we obtain (dist (z; dist In Section 3 we shall discuss how to further bound the errors when K is a polyhedral cone, which is the situation when the original Hoffman lemma applies. In the rest of this section we assume that K is a general convex cone. In addition to this we assume that the Slater condition is satisfied, i.e. Assumption The following lemma is well-known; see e.g., Duffin [5], Borwein and Wolkowicz [2], Luo, Sturm and Zhang [11], Nesterov and Nemirovskii [14] and Sturm [21]. For completeness we provide here a short proof. Lemma 2.2 Suppose that Assumption 2 holds. Then, for any y 2 must follow that b T y ? 0. Proof. Suppose, for the sake of deriving a contradiction, that there is y 6= 0 such that y 2 Consider the hyperplane For any x while for any x 2 K, since y 2 K we have y T x - 0. This means that H y separates b +L and K, yielding a contradiction to the fact that b +L intersects with the interior of K.For fixed - x we consider again the system (KKT) in terms of - and -. After some re-arrangements this yields 8 which is a closed convex cone as well. Note that - case and is omitted in our proof. In many applications, 0 62 L and so We shall mention another easy case, i.e. - x lies in the interior of K then - f0g. In this case x, and therefore dist (z; F) - dist due to (2.3). In what remains we shall only concentrate on the situation when - x 62 int K. Remark that for - K is known as a face of K . The condition (2.4) is equivalent to Lemma 2.3 If Assumption 2 holds then Proof. Suppose for contradiction that there is y K . This means that However, - which is impossible due to Lemma 2.2. 2 Now we define the minimum angle between L ? and - K as Note that both L ? and - K are closed cones. According to Lemma 2.3, it follows that for any - In order to pursue our analysis further, one of the following two mutually exclusive cases will be considered. Assumption 3 The set Assumption Let us first consider the situation when Assumption 3 holds. In that case we know that there exists such that for any - we always have Now take - K . Let the projection of 0 onto - -s s s s Figure and the cone - K . Let the angle between - and - \Gamma p be '. Clearly, ' -=2. Moreover, Denote We are now in a position to prove the following error bound result. Theorem 2.1 If Assumption 2 and Assumption 3 hold then dist (z; F) -v(z; F) for all z Proof. By (2.5) we have dist (z; F)= sin ' -dist (z; F): Using the first equation in (2.4) we also have ')dist (z; -dist (z; F): Recall relation (2.3). By the above estimations on k-k and k-k, it follows from (2.3) that (dist (z; F)) 2 -dist (z; F)(kz d k dist and consequently dist (z; F) -v(z; F):In the other situation, namely if Assumption 4 holds, then a similar result can be shown. Theorem 2.2 If Assumption 4 holds, then for any b 2 ! n we must have (b Moreover, there is a constant - ? 0, independent of b, such that dist (z; F) -v(z; F) Proof. First we show that (b Suppose otherwise that there is b with (b Then, there will be a hyperplane separating b +L and K, say with such that Since K is a closed cone, the above separation implies that y T x - 0 for all x 2 K and Moreover, we also have y T This is in contradiction with the condition Compared with Lemma 2.3, we have now a stronger relation: f0g. This means that the proof of Theorem 2.1 can remain exactly the same, except that now ' ? 0 can be taken as the minimum angle between L ? and K which is independent of b. 2 We remark that both Theorem 2.1 and Theorem 2.2 easily extend to the case when L is a closed cone. Theorem 2.3 Suppose that K 1 is a closed convex cone and K 2 is a closed, convex, solid and pointed cone. Furthermore, suppose that (b compact. Then there is a constant - ? 0 such that dist (z; F) -(dist (z; b dist (z; K 2 for all z Proof. We follow similar lines as in the proof of Theorem 2.1. Consider subject to x 2 b +K 1 Let the optimal solution be - x. The Karush-Kuhn-Tucker optimality condition yields: Let and Both - are closed convex cones. Now we claim that Suppose such is not the case. Then, one should be able to find - 6= 0 satisfying Hence, b T This implies that contradicting the Slater condition. are closed convex cones and, moreover, - 2 is a solid pointed cone, we derive from (2.6) that - 2 can be strictly separated from \Gamma - 1 . Due to compactness of F we may let ' be a positive lower bound on the minimum angle between this separating hyperplane and - 2 . Then we have and consequently dist (z; b +K 1 )k- dist (z; K 2 dist (z; K 2 The desired result thus follows. 2 Similarly, we have the following result, the proof of which is pretty much the same as that of Theorems 2.2 and 2.3 and is omitted here. Theorem 2.4 Suppose that K 1 is a closed convex cone and K 2 is a closed, convex, solid and pointed cone. Furthermore, suppose that here is a constant - ? 0, independent of b, such that dist (z; F) -(dist (z; b dist (z; K 2 for all z When more than two cones are concerned, a similar result holds under Slater's condition. First we note the following lemma, see e.g. [11]. Lemma 2.4 Let K be a convex cone and int K 6= ;. Then, x 2 int K if and only if for any holds that Theorem 2.5 Let K i be convex cones, m. Suppose that Then, there is - ? 0 such that dist dist (z; K i for any z 2 ! n . Proof. Consider subject to x 2 K Hence, for the optimal solution - x the KKT condition yields with Let By Lemma 2.4 there exists m. Let with z ip 2 K i , z id 2 K ip z due to Lemma 2.1. Moreover, kz id dist (z; K i m. Therefore, z z T kz id kk- i k: On the other hand, since dist (z; K i and so by letting it follows that dist dist (z; K i ):Theorem 2.1 can be viewed as an analogue to Robinson's result for convex inequality systems. In the form of convex inequality systems, Theorem 2.2 can be found in Hu and Wang [9] and Deng and Hu [3]. In particular, Deng and Hu [3] investigated the case when K is the cone of positive semidefinite matrices. This case is known as linear matrix inequalities (LMIs for short). In its optimization version it is also called semidefinite programming and has received intensive research attention recently. Sturm [20] mainly investigated error bounds for LMIs in the absence of Slater's condition. In fact, in the context of LMIs, both Theorem 2.1 and Theorem 2.2 also follow from the analysis in [20]. Moreover, an example was given in Sturm [20] showing that Assumption 2 alone cannot guarantee a global Lipschitzian type error bound even for LMIs. Such an error bound is only possible when an additional scaling factor is present. Below we shall discuss how to derive some conditioned error bound relation for the convex conic problem (1.2) under Assumption 2, without assuming Assumption 3 and Assumption 4. In this situation the recession cone L " K must be non-empty and it is not contained in the interior of K. For a fixed positive angle consider the following cone the projection of x onto L and the cone L " K has an angle at least -=2 Theorem 2.6 Suppose that Assumption 2 holds. There exists a constant - ? 0 such that dist (z; F) -v(z; F): for all z 2 C. Proof. Observe that if - x is the projection of z on F , then it must also be the projection of z on F for any y 2 L ? . This can be seen as follows. The fact that - x is the projection of z is equivalent to the existence of - 2 L ? and - 2 K such that (See also (2.1)). Now if z is changed to z y, then we need only to change - to - satisfy the same set of KKT conditions. Remark also that to prove the theorem it is sufficient to show that, for any z 2 C, its projection onto F is contained in a compact set. Suppose that the theorem is false and that there is a sequence fz :::g, such that the corresponding projection on F , f-x unbounded. Due to the above remarks we have made, we need only to consider the projection of z (k) onto the subspace L. Without loss of generality, assume that z For sufficiently large k we have where the first inequality is because - x (k) must be pointing towards the cone of recession directions and the last inequality is due to the fact that k-x (k) k ! 1. This contradicts to - x being the closest point in F to z (k) .For any given point z C. The following relation is immediate. Lemma 2.5 dist (z; F) - dist (z Proof. Let the projection of z 2 onto F be - x 2 . Then, dist (z; F) - where we used the fact that z 1 2 L " K and so - 2.5 and Theorem 2.6 we have Theorem 2.7 Suppose that Assumption 2 holds. Then dist (z; F) -v(z for all z 3 Hoffman's error bound and the condition number In this section we shall discuss error bounds for the linear system fx j A T x - bg with A 2 ! m\Thetan and m. This is the setting for which Hoffman's error bound result applies ([7]). Our purpose is to see how the constant in Hoffman's bound is related to other known quantities for the linear system. Previous results on the constant of Hoffman's bound can be found, e.g., in [12, 1, 6, 10]. By introducing a slack confine ourselves to the range space of A T , i.e. Accordingly, . For a given z . Let which minimizes the distance to s(z). Let ng n K: Then, for this given s(-x) - 0 we can rewrite (2.1) as A J - As (3.1) is a necessary condition for optimality, it is certain that (3.1) is feasible. What remains to be analyzed is the size of the solution. A key ingredient in our analysis is the following lemma. Lemma 3.1 Suppose that A has full row rank. Then, diagonal and D - Lemma 3.1 was first shown by Dikin [4] and was used in his convergence analysis for affine scaling methods. Among others, Stewart [19] and Todd [23] rediscovered this result later. The meaning of Lemma 3.1 can be interpreted as follows. It is well known that 0g and are orthocomplements to each other. Obviously, for a given positive diagonal matrix D, Null(A) can only intersect with at the origin, hence there must be a positive angle between them. Lemma 3.1 further states that the minimum angle between Null(A) and uniformly bounded from below by a positive constant which is independent of D. To understand this fact we may consider the following example. Let simply the line x 1 For a given positive diagonal matrix D, contained in the first and the third quadrants. The angle between these two subspaces never exceeds -=4. An important property of the constant -(A) is that it reflects an intrinsic, geometric relationship of the spaces. Vavasis and Ye [24] used this constant -(A) as a measure of complexity for solving the related linear programming problem. Their results showed that, in a real-number computation model, linear program is solvable in polynomial-time, in terms of total number of basic operations, with respect to the dimension n and the complexity measure log -(A). For problems with integral input data, this result yields the usual polynomiality complexity result for linear programs in terms of the input-length. Holder, Sturm and Zhang [8] showed that -(A) plays an important role in sensitivity analysis for linear programming. Furthermore, Sturm and Zhang [22] extended some of the results in [8] to semidefinite programming. It is known however, that Lemma 3.1 cannot extend to general semidefinite programming for arbitrary invariant scaling of the cone S n\Thetan Fortunately, in analyzing (3.1) we need only to deal with a polyhedral cone. To see how condition number -(A) can play a role in error bound analysis, we need to introduce a number of technical lemmas. First we note the following equivalent definition of -(A) for arbitrary matrix A due to Vavasis and Ye [24]. Lemma 3.2 It holds that kck positive diagonalg: For our analysis it is important to know the size of a solution for a linear system. To this end, we note the following two lemmas. Remark that Renegar [16] studied similar problems in a quite general framework using a quantity called distance to ill-posedness. Lemma 3.3 Suppose that A has full row rank. Further assume that fx j Then, there is a solution - x in 0g such that k-xk -(A)kbk: Proof. Consider a linear program subject to and its dual (D) maximize b T y subject to A T y Both (P) and (D) satisfy Slater's condition. Therefore their respective analytic central paths satisfying the following relation: x(-)s(-e: Multiplying the second equation in (3.2) with X(-), the diagonal matrix with x(-) as its diagonal components, and applying the first equation in (3.2) we obtain e: Substituting this into the second equation and finally using the third relation in (3.2) we have Now we can apply Lemma 3.1 to obtain The lemma is proven. 2 Next we shall extend this result to the case when Slater's condition is no longer assumed. Lemma 3.4 Suppose that A has full row rank. Further assume that fx j Then, there is a solution - x in 0g such that k-xk -(A)kbk: Proof. Let Consider a perturbed set Clearly, F ffi contains an interior point and therefore Lemma 3.3 can be invoked. Let x The set fx be a cluster point of x ffi as and shall compare the condition number of A and that of its submatrices. Proof. By Lemma 3.2, kck positive diagonalg: partitioned in accordance with For fixed c 1 6= 0 and fixed positive diagonal matrix D 1 . Let c positive diagonal and D 2 ! 0. Clearly, the set of solutions minimizing kD 1=2 converges to the set of solutions minimizing kD 1=2 )k. For given c and D let y(c; D) be a maximum norm solution among solutions which minimize kD 1=2 similarly. It follows that lim sup As a consequence, and so the lemma is proven. 2 Applying Lemmas 3.4 and 3.5 to (3.1) we have Finally we shall give a bound on the constant in Hoffman's error bound for linear systems. Theorem 3.1 Suppose that and A has full row rank. It holds that dist (z; F) -(A)(cond(AA T for any z 2 Proof. Using (3.1) and (3.3), By (2.3), on one hand we have On the other hand, Combining these two inequalities, the desired result follows. 2 Conclusions In this paper we discuss error bounds for sets in convex conic form. The notion of constraint violation is extended to this class of problems. For a number of applications the measure of constraint violation is easy computable. We show that under Slater's condition, and additionally, if either the feasible set is bounded or the recession directions satisfy the Slater's condition, then there is a global Lipschitzian type error bound for general convex conic problems. These results can be generalized to the intersection of multiple convex cones, or intersection of two shifted convex cones, one of them being pointed and solid. If only Slater's condition is satisfied without additional assumptions on the feasible region, then a global error bound is impossible as shown by Sturm [20]. In this case, one may still identify a region in which Lipschitzian type error bound holds. Finally, we discuss the bounds in Hoffman's lemma for linear systems. It is shown that such a bound is linked closely with the condition number for linear programming as investigated by Vavasis and Ye [24]. Acknowledgement I would like to thank Jos Sturm for pointing out an error in an earlier version of the paper. --R The distance to a polyhedron Characterizations of optimality for the abstract convex program with finite dimensional range Computable error bounds for semidefinite programming Iterative solution of problems of linear and quadratic programming Linear Inequalities and Related Systems Approximations to solutions to systems of linear inequalities On approximate solutions of systems of linear inequalities sensitivity analysis and parametric programming On approximate solutions of infinite systems of linear inequalities The sharp Lipschitz constants for feasible and optimal solutions of a perturbed linear program Duality results for conic convex programming A condition number of linear inequalities and equalities A condition number for differentiable convex inequalities bounds in mathematical programming Some perturbation theory for linear programming An application of error bounds for convex programming in a linear space Convex Analysis On scaled projections and pseudoinverses bounds for linear matrix inequalities On sensitivity of central solutions in semidefinite programming A Dantzig-Wolfe-like variant of Karmarkar's interior-point linear programming algorithm A primal-dual interior point method whose running time depends only on the constraint matrix --TR
convex conic problems;error bound;condition number
589314
Minimizing a Quadratic Over a Sphere.
A new method, the sequential subspace method (SSM), is developed for the problem of minimizing a quadratic over a sphere. In our scheme, the quadratic is minimized over a subspace which is adjusted in successive iterations to ensure convergence to an optimum. When a sequential quadratic programming iterate is included in the subspace, convergence is locally quadratic. Numerical comparisons with other recent methods are given.
Introduction In this paper we consider the problem of minimizing a quadratic over a sphere: subject to kxk where A is a symmetric n n matrix, b 2 R n , T denotes transpose, and k k is the Euclidean norm. This minimization problem is often called the trust region subproblem since it must be solved in each step of a trust region algorithm [1, 2, 3, 15, 19]. Problems of this form arise in many other applications including regularization methods for ill-posed problems [14, 26], and graph partitioning problems [10]. Although the solution to (1) can be expressed in terms of a diagonalization of A, this representation is practical only when n is small. In this paper, we focus on the large-scale case. One approach to the large-scale case, developed by Golub and von Matt in [5] (also see [4]), is to (partially) tridiagonalize A using the Lanczos process and then solve tridiagonal problems to obtain an approximate solution to (1). For further developments of this approach, including preconditioning and a Fortran 90 implementation HSL VF05 in the Harwell Subroutine Library, see Gould et al. [7]. For the method developed in this paper, we use an approach in the spirit of the Golub/von Matt/Gould et al. scheme to obtain a starting guess. Parametric eigenvalue approaches to the sphere constrained problem (1) are developed by Sorensen [24] and by Rendl and Wolkowicz [20]. The relationship between these two approaches is discussed in detail in [20]. Roughly, Sorensen's approach involves constructing an approximation to the solution of (1) from the solution to a related eigenvalue problem. Since this approximation may not satisfy the bound on the norm of the solution, a series of eigenvalue problems are solved, and in the limit, the bound on the norm of the solution is fullled. In the approach of Rendl and Wolkowicz, the same eigenvalue problem is solved in each iteration, however, the bound on the norm of the solution is satised by maximizing a related dual func- tion. The eigenvalue problems arising in either approach can be solved using Arnoldi techniques such as those developed in [13]. In the \hard case" (see [16]) where b is orthogonal to the eigenvectors associated with the smallest eigenvalue of A, Sorensen's approach needs to be modied. An e-cient algorithm for the hard case is developed by Rojas in her thesis [21]. She also uses this algorithm to solve some di-cult ill-posed problems of Hansen [11, 12]. The approach of Rendl and Wolkowicz does not need modication in the hard case, however, the convergence of algorithms for the eigenvalue problem may be slower when the computed eigevalue is not simple. The approach in this paper, which we call the sequential subspace method (SSM), involves solving (1) with the additional constraint that x is contained in a subspace. We show that convergence is locally quadratic (locally cubic when if the subspace contains the iterate generated by one step of the sequential quadratic programming (SQP) algorithm applied to (1). The convergence is quadratic even when the original problem is degenerate with multiple solutions, and with a singular Jacobian for the rst-order optimality system. Descent of the cost at a nonoptimal point can be ensured by including in the subspace either the cost gradient or an eigenvector associated with the smallest eigenvalue of A. We observe in numerical experiments that appropriate small dimensional subspaces are generated by preconditioned Krylov space and minimum residual techniques. Comparisons with the algorithms of Sorensen [24], of Rendl and Wolkowicz [20], and of Gould, Lucidi, Roma, and Toint are given in Section 5. A solution of the problem subject to is any eigenvector associated with the smallest eigenvalue of A. In comparing the SSM approach to algorithms for solving the eigenproblem, it follows from the discussion of Sleijpen and Van der Vorst in [22] that an SQP iterate for (2) is closely connected with the Rayleigh quotient iteration [18, p. 70], which is cubically convergent [18, p. 73]. In [22] approximate solutions to the SQP system are used to build up subspaces containing the approximation to the eigenvector. In this paper, we solve the SQP system relatively precisely, and we form a small dimensional subspace containing the SQP iterate. After computing the new approximation in the subspace, the previous information is discarded; hence, the computer memory requirements are relatively small. Complete diagonalization If there exists a solution y of (1) with kyk < r, then A is positive semidenite and y is the global minimizer of the quadratic x T Ax 2b T x. Thus, when a minimizer of (1) lies in the interior of the constraining sphere, the constraint can be ignored and the optimization problem can be approached using techniques for unconstrained optimization. Consequently, we restrict our attention to the following equality constrained problem: subject to The solutions to (3) are characterized by the following result (see [23, Lem. 2.4, Lem. Lemma 1. The vector x is a solution of (3) if and only if r and there exists such that A + I is positive semidenite, and The solution to (3) can be expressed in terms of the eigenpairs of A. Let T be a diagonalization of A where is a diagonal matrix with diagonal elements and is the matrix whose columns 1 , are orthonormal eigenvectors of A. Dening Lemma 2. The vector is a solution of (3) if and only if c is chosen in the following way: (a) Degenerate case: If then are arbitrary scalars satisfying the condition (b) Nondegenerate case: If (a) does not hold, then c where > 1 is chosen so that Proof. Simply check that the su-cient optimality conditions of Lemma 1 are satised. The degenerate case, where the Jacobian of the rst-order optimality system may be singular, coincides with the \hard case" of More and Sorensen [16] where b is orthogonal to the eigenspace associated with the smallest eigenvalue of A and the multiplier is equal to 1 . In the nondegenerate case, the multiplier is chosen so that A+ I is positive denite and the solution the constraint x T In the nondegenerate case, equation (5) leads to upper and lower bounds for the multiplier . Since i kbk r To obtain a lower bound, observe that which yields the relation Utilizing the upper and lower bound u and l and the strict convexity of the left side of (5) on the interval ( l ; u it is easy to devise e-cient algorithms to compute a solution of (5). 3 Incomplete diagonalization, local convergence At iteration k in the sequential subspace method (SSM) for (3), we impose the additional constraint that x lies in a subspace S k of R n . Hence, the new iterate x k+1 is a solution of the problem subject to We show that the convergence is locally quadratic, even when the original problem (3) is degenerate, if we include an SQP iterate associated with x k in S k If V is an n l matrix with orthonormal columns that span S k , then (8) is equivalent to the problem subject to After substituting for x, (9) reduces to the following problem in R l : subject to l is small, then (10) can be solved by complete diagonalization as in Section 2. Or if B has a sparse factorization, then (10) can be solved quickly using the Newton approach developed in [16]. In theory, a tridiagonal B is generated using the Lanczos process [6]. In particular, if v 1 is a unit vector and v i is the i-th column of V, then the Lanczos process can be expressed as follows: Algorithm 1 (Lanczos) ksk Here d is the diagonal and u is the superdiagonal of the tridiagonal matrix B. If then the Lanczos process is terminated and the column space of V and AV coincide. It is well-known that the columns of V generated by this process may deviate signicantly from orthogonality due to the propagation of rounding errors. And when this happens, (9) is no longer equivalent to (10). Nonetheless, Gould et al. observe in [7] that the solution to (10) often provides a good approximation to the solution of despite the loss of orthogonality. The Lanczos process can be repaired, in order to restore orthogonality, by using a Householder process to generate the columns of V. This process, however, requires products between a vector and each of the previously computed columns of V. Thus the overhead needed to maintain orthogonality grows like nl 2 in the number of ops and like nl in storage. This overhead can be signicant when n or l is large. On the other hand, to compute a high accuracy solution, we need to maintain orthogonality in order to obtain an equivalent problem (10). This leads us to focus on approaches that involve subspaces where l is much smaller then n. In particular, for an implementation (Algorithm 4) of the SSM proposed later, l is either 4 or 5. Since sequential quadratic programming (SQP) techniques often converge rapidly, with a good starting guess, we always include the SQP approximation in the subspace . If x k is the current iterate, which we assume satises the constraint and k is the current approximation to the multiplier associated with the constraint, then the SQP iterate can be expressed in the following way: and z and are solutions of the following linear system: When the coe-cient matrix in (11){(12) is singular, we let (z; ) be the minimum residual/minimum norm solution; that is, (z; ) is gotten (in theory) by multiplying the right side by the pseudoinverse of the coe-cient matrix (see [8]). The SQP method is equivalent to Newton's method applied to the nonlinear system A solution x k+1 to the subspace problem (8) is an approximation to the solution of (3). To obtain an estimate for the multiplier of Lemma 1, we minimize the Euclidean norm of the residual b Ax k+1 x k+1 over the scalar . This works out to give (b Ax) T x We now examine the local convergence of a solution x of (8) and the multiplier estimate (14) under the assumption that S k contains is a solution to (11). Let S denote the set of minimizers of (3), and let be the multiplier given by Lemma 1. In the nondegenerate setting where A+ I is positive denite, we show that the iteration is locally, quadratically convergent to the unique solution of (3). In the degenerate case = 1 where S has more than one element, we obtain local quadratic convergence to S , where distance is measured in the usual way: In the nondegenerate degenerate-case where contains a single element, we obtain local quadratic convergence for a \safe-guarded" choice of k . Our convergence result in the special nondegenerate degenerate-case is given later in Lemma 5, while our local convergence result in either the nondegenerate case or the degenerate case with multiple solutions is the following: Theorem 1. Let be the multiplier of Lemma 1 associated with the set of solutions S of (3), and suppose that either A+ I is positive denite, or with (4) a strict inequality. Then there exist positive constants and C with the property that for any (x k ; k ) such that and for any subspace S k that contains the SQP iterate xSQP associated with (11){ (12), any solution x k+1 of (8) and associated multiplier k+1 given by (14) satisfy the following estimate: The eigenvalue problem (2), corresponding to b = 0, is always degenerate (with multiple solution) and the error has the following special form: When the multiplier is estimated using (14), it can be shown, when that the error in the multiplier is bounded by a constant times the error in the solution vector squared (see the remark at the end of Section 3.1). It follows that for some constant the same as the convergence result for the Rayleigh quotient iteration. 3.1 Nondegenerate problems We begin the derivation of Theorem 1 with the nondegenerate case: Lemma 3. If (3) has a solution x and an associated multiplier with > 1 , then there exist a neighborhood N of (x ; ) and a constant C with the property that for any and for any subspace S k that contains the SQP iterate xSQP associated with (11){(12), any solution x k+1 of (8) and associated multiplier k+1 given by (14) satisfy the following estimate: Proof. Since > 1 , the matrix A+ I is positive denite, and the Jacobian of the nonlinear system (13) is nonsingular at By the standard convergence theorem for Newton's method applied to a smooth system of equations, there exist a neighborhood N of ) and a constant c such that whenever Let and be positive scalars chosen so that for all x 2 R n , let f be the cost function in (3), let L be the Lagrangian dened by A Taylor expansion around x yields the following relation: for any x 2 B r rg. Combining this with (16) gives for any x 2 B r If p is the projection of xSQP onto B r , then Hence, we have xSQP for some , it follows that p 2 S k and f(x k+1 ) f(p). Combining this inequality with (17) and (18) gives which implies that Making this substitution gives r Combining (21) with (19), the proof is complete. Remark. For the eigenvalue problem (2), we have x In this case, (20) yields and (21) becomes 3.2 Degenerate problems Now consider local convergence in the degenerate case where Referring to Lemma 2, the degenerate case can only happen when i Any solution to (3) in the degenerate case can be expressed x and 1 is any linear combination of the vectors i satisfying the relation Initially, we suppose that k 1 0, in which case the projection of S on the eigenspace associated with contains a sphere of radius -. Our convergence result is the following: Lemma 4. Suppose that the multiplier of Lemma 1 associated with the set of solutions S of (3) is given by is the component of an element of S in the eigenspace associated with E 1 . Then there exist positive constants and C with the property that for any (x k ; k ) such that and for any subspace S k that contains the SQP iterate xSQP associated with (11){ (12), any solution x k+1 of (8) and associated multiplier k+1 given by (14) satisfy the following estimate: Proof. Initially, let us assume that k is near 1 , but k 6= 1 . In this case, the linear system (11){(12) is nonsingular, and there exists a unique solution (z; ). We expand z and x k in terms of the eigenvectors of A writing and x . Utilizing (11), we obtain Substituting this in (12) gives Let us dene I)x k and . For If x 2 S , then since we have Let k be the error at step k dened by By (26), we have since is near - 2 > 0 when x k is near S . From (23), we have Let x be the closest element of S to x k and dene . Then we have By (30) the i component of in error by O( k ) since i , the i component of x k , is in error by O( k by (31). implies that . Combining this with (28) and Hence, for i 2 E+ the i component of xSQP is in error by O( 2 be the seminorm associated with projection into the eigenspace associated with Then we have for all x ng. Proceeding as we did earlier, but replacing norms by seminorms, where p is the projection of xSQP onto the ball B r , and xSQP for some by (30) and (32), and z is perpendicular to x k by (12), we have This implies that Consequently, kp x which combines with (34) to give By the triangle inequality, Let k k 1 be the seminorm dened by and recall that kx k . By the Pythagorean theorem and the fact that x k+1 has length r, we have which implies that The distance from x k+1 to S is given by where x is any element of S . Relations (35){(38) yield dist(x Combining these estimates, we have k+1 This analysis was given under the assumption that k 6= 1 . In the special case k show how the analysis should be modied. With the change of variables z = and the substitution x , the SQP system (11){(12) is equivalent, by orthogonal transformation, to D# where D is a diagonal matrix with diagonal elements then the rst s diagonal elements of D and the rst s components of D vanish. Hence, the rst s equations in (39) imply that The next n s equations give while the last equation in (39) gives The minimum norm solution to this last equation is By (40), Combining these bounds, we have these relations, all the analysis from (33) onward can be applied, leading us to the estimate k+1 Lemmas 3 and 4 yield Theorem 1. 3.3 Nondegenerate degenerate-problems Finally, let us consider the nondegenerate degenerate-case where and the 1 component of x in the eigenspace associated with the smallest eigenvalue of A vanishes. Our convergence result is the following: Lemma 5. If (3) has a solution x is given by (22), then there exist a neighborhood N of (x ; 1 ) and a constant C with the property that for any and for any subspace S k that contains the SQP iterate xSQP associated with (11){ (12), the solution x k+1 of (8) and associated multiplier k+1 given by (14) satisfy the following estimate: In the case that k k, C can be chosen so that Proof. Focusing on the numerator in (24), and substituting , we With this substitution for the numerator of in (24), we obtain the denominator terms in (43) have the following lower bound: Another lower bound is gotten by neglecting terms corresponding to indices i where the seminorm k k 1 is dened in (36). Combining (43){(45) yields: Returning to our previous analysis of the degenerate case, it follows from (29) and (46) that for Here we exploit the fact that for . In order to analyze (47), we consider two separate cases: (i) kx k x k+ and (ii) kx k x k+ , where is any xed constant satisfying In case (i), x k+ x k+ We now derive a similar bound for the left side of (49) in case (ii). In this case, it follows from (42) that x k+ for any x 2 R n , where a + subscript on a vector is used to denote its projection on the eigenspace associated with E+ . After substituting for using (20), we obtain for any x 2 B r . Assuming x k is a unit vector (note that when kx k only if x We will establish a uniform bound for the expression (51) when x k is near x , To facilitate this analysis, we rst consider whether the equation has a solution of the form y of this form, the Schwarz inequality gives jy T Since the unit vector y is orthogonal to the eigenspace associated with 1 , Multiplying (52) by y T and using both (53) and (54) gives For any x 2 B r , we have which implies that since yields the relation: Referring to (55), we have a contradiction when kx x k+ . In summary, the equation (52) has no solution over the set Y consisting of those y that satisfy the following conditions: lies in the closure of Y, then by (57), since any solution of (52) satises (55), y cannot be a solution of (52). Since (52) has no solution over the closure of Y, the following constant - is strictly positive: Since lim min (51) is bounded uniformly over all x k near x with x k+ . Thus in either case (i) or (ii), the left side of (49) is bounded, and by (47), we have the same as relation (30) in the degenerate case. To establish the analogue of (32) for indices i 2 E+ , we need a dierent bound for the next to last term in (43). From the identity Hence, we have Since implies that It follows that This estimate along with the lower bound (44) for the denominator in (43) yields the relation The reminder of the analysis is identical to that given for the degenerate case (Lemma 4), starting with (32). Since S it follows from the analysis of Lemma 4 that In the special case k Hence, the in (60) can be absorbed in the kx k completes the proof. Implementation In our experimentation with the SSM, we put the following four vectors in S k in each iteration: xSQP , x k , and an estimate for an eigenvector of A associated with the smallest eigenvalue. By including x k in S k , the value of the cost function can only decrease in consecutive iterations. The multiple b Ax k of the cost function gradient ensures descent if the current iterate does not satisfy the rst-order optimality conditions. The eigenvector associated with the smallest eigenvalue will dislodge the iterates from a nonoptimal stationary point. We also use this vector in a \safe-guard" strategy designed to keep A positive denite. 4.1 The SQP system Now consider the SQP system (11){(12). According to (12), z is orthogonal to the prior iterate x k . Let P be the matrix that projects a vector into the space perpendicular to Multiplying (11) by P yields according to (12), we have We haved found that preconditioned Krylov space methods, such as the Gauss-Seidel scheme in [9], converge very quickly when applied to (61). As a small illustra- tion, let us consider the second test problem from [24] with where is a 1000 1000 diagonal matrix with diagonal elements selected randomly from a uniform distribution on ( :5; :5) and I 2qq T where q is gotten by rst generating random numbers on ( :5; :5) and then scaling the resulting vector to have unit length. The vector b is generated in the same way as q. The solid curve in Figure 1 gives the convergence when a Lanczos type process (Algorithm 1, with starting vector v is used to generate the matrix V used in (9). The Lanczos process was modied to ensure orthogonality of the columns of V. For each value of l in Algorithm 1, we solve the l l tridiagonal problem (10) to obtain an approximate solution x and associated multiplier for the original problem (3). In the solid curve of Figure 1, we plot the base 10 logarithm of the norm of the residual kb (A+I)xk. According to Lemma 1, the residual vanishes at an optimal solution. The dashed curve of Figure 1, based on the SSM approach, is gotten in the following way: Taking Algorithm 1, we generate a V with 40 orthonormal columns. Solving (10), we obtain starting guess x 0 . In iteration k of the SSM phase, we start with the vector v we use the Gauss-Seidel/Krylov space approach of [9] to generate a matrix V, with orthonormal columns, that approximately contains a solution of (61) in its range. Using the V generated in this way, we solve (9) to obtain the next iterate x k+1 . The associated multiplier is estimated using (14). Each kink in the dashed curve of Figure 1 corresponds to the number of iterations needed to obtain an approximate solution of (61). In this example, roughly 15 multiplications by the elements of the matrix A are used to solve (61). The quadratic convergence of SSM is re ected in the rapid decay of the residual norm. This approach for generating V, using a nonsymmetric Gauss-Seidel matrix, Krylov spaces, and orthogonalization, can become expensive when n is really large Matrix-vector products Figure 1: Convergence of the tridiagonalization approach (solid) and SSM (dashed) for the second test problem from [24]. since each of the columns of V should be stored in memory. Hence, in the remainder of this paper, we focus on low-storage symmetric techniques for solving (61), which we compare to other approaches. We solve (61) using a preconditioned version of Paige and Saunders' MINRES algorithm [17]. More precisely, we use Algorithms 3 and 3a in [9] and three dierent choices for the symmetrizing preconditioner W in that paper: (i) corresponding to unconditioned iterations, (ii) is the diagonal matrix whose diagonal matches that of L is the strictly lower triangular matrix whose lower triangle matches that of C. The implementations of SSM associated with the latter two preconditioners are denoted and SSM l respectively. Typically, the L matrix associated with I)P is dense, even when A is sparse, since P is often dense. Nonetheless, linear systems of the form can be solved in time proportional to the number of nonzero elements in the lower triangle of A due to the special structure of C. In terms of the vectors w, q and p dened by I)w and the diagonal d of C can be expressed while the o-diagonal elements of C are Exploiting this structure, it can be shown that the solution to (L can be computed in the following way: Algorithm The statement y a i+1:n;i of Algorithm 2 only requires the nonzero elements in column i of A beneath the diagonal. Hence, the number of oating point operations for Algorithm 2 is O(n) plus the number of nonzero elements in the lower triangle of A. The analogous procedure for the transposed system is the following: Algorithm 3 4.2 Positive deniteness In theory, the MINRES algorithm we use to solve (61) can be applied to any symmetric matrix. In practice convergence can be extremely slow when C is indenite. For this reason, we try to choose k so that A+ k I is positive denite. If e is an eigenvector of the matrix B in (10) associated with the smallest eigenvalue , then the pair (v; ), approximates an eigenpair of A corresponding to the smallest eigenvalue. The error in can be estimated in the following way: If is closer to 1 than the other eigenvalues of A, then after substituting in the residual r = Av v, we have since 1. Thus j 1 j krk, which implies that With this insight, we replace the least squares estimate (14) by the following safeguarded estimate: When the approximate eigenpair (v; ) is not very accurate, then the safe-guarded step (62) is a safe, but poor approximation to . Hence, whenever we apply one iteration of SSM to the quadratic eigenvalue problem (2) in order to compute a more accurate eigenpair. Due to the third and sixth order estimates in (15), simply one iteration of SSM for the eigenproblem often yields a highly accurate eigenpair. 4.3 The algorithm We now collect our observations and present the algorithm that was used to generate the numerical results of the next section. To simplify the presentation, we introduce the following subroutines: This routine applies Algorithm 1 to the matrix A, starting from the vector v 1 , to generate a matrix V with columns l This routine solves the problem (8) generating a solution denoted x, and an associated multiplier is a matrix whose columns are an orthonormal basis for S k , then an estimate (v; ) for the smallest eigenvalue of A and an associated eigenvector is gotten by computing the smallest eigenvalue and an associated eigenvector e for setting This routine computes a (minimum residual, minimum solution (z; ) of the following linear system: Our implementation of the sequential subspace method combines these three routines and the safe-guarded step (62): Algorithm 4 (Safe-guarded SSM with Lanczos startup) while ( == & it while ( kb Algorithm 4 For the computational results reported in the next section, we took :01ng. The \rand" function appearing at the start of Algorithm 4 generates a vector with components uniformly distributed on [0; 1]. 5 Computational results In this section we compare the performance of SSM to the performance of the algorithms in [7], [20], and [24], denoted GLRT, RW, and S respectively, using the three test problems presented in [24]. The results that we report for S were extracted from [24], while the results reported for GLRT and RW were obtained using codes provided by the authors. We thank the authors for providing access to their codes. Each of these codes used dierent stopping criteria. GLRT stopped when kb (A+I)xk=kbk was bounded by a given tolerance, while RW stopped when the gap between the value of the primal and dual problem, and hence the error in the primal cost function, was smaller than a given tolerance. In order to ensure that each code computed a solution with the same accuracy, we adjusted the error tolerance parameter of each code until the value of kb for the computed solution was smaller than a given tolerance (specied below). In the rst test problem of [24], is the standard 2-D discrete Laplacian on the unit square based on a 5-point stencil with equally-spaced mesh points. Taking a series of 20 problems were generated where b was a vector with elements uniformly distributed on [0; 1]. Each of these problems was solved using three dierent tolerances, In Table 1 we give the average number of matrix-vector products involving A for each algorithm. Each iteration of the preconditioned MINRES algorithm with lower trian- Tolerance S RW GLRT SSM SSM d SSM l Table 1: Problem 1, average number of matrix-vector products versus tolerance. gular preconditioner involves roughly twice as many ops as an iteration of either the identity or the diagonal preconditioned schemes. Hence, in doing the bookkeeping, we charged for two matrix-vector products in each iteration of the triangular preconditioned scheme. As seen in Table 1, SSM l converges more than twice as fast as the identity and diagonal preconditioned schemes, and overall, SSM l uses the smallest number of matrix-vector products for this test problem. Since the parametric eigenvalue algorithms S and RW compute an extreme eigenvalue for a series of matrices, we also list in parentheses in Table 1 the number of these eigenproblems that are solved. Hence, RW is very economical in terms of the number of these eigenproblems that are solved. The second suite of test problems in [24] utilizes the matrix described earlier in Section 3. In these problems, the radius of the sphere is varied and the number of matrix-vector products is tabulated. For radii of one or smaller, solutions can be computed extremely quickly, so we focused on and an error tolerance of 10 7 . In Table 2 we see that for had the fewest matrix- Radius S RW GLRT SSM SSM d SSM l Table 2: Problem 2, average number of matrix-vector products versus radius. vector products, while GLRT had the fewest for The nal problem of [24] again employed the discrete Laplacian matrix, but with 100. The vector b was designed to make the problem degenerate; rst a random b was generated, then its 1 component was removed. Table 3 gives the results for the various algorithms. SSM l Table 3: Problem 3, average number of matrix-vector products. We placed an asterisk by the result in Table 3 for GLRT since this routine reduced the error to 10 4 , not the 10 7 tolerance used by the other routines. Among the routines that achieved the error tolerance, SSM l performed the best relative to the number of matrix-vector products. Note that the number of matrix-vector products given in Table 3 for S was taken from [24] while Rojas, in her recent thesis [21], developed a more e-cient implementation of Sorensen's approach for degenerate problems. In summary, a Lanczos type process seems to be very eective when the problem is very nondegenerate ( >> 1 ). As the problem becomes more degenerate, preconditioned schemes such as SSM d or SSM l appear more eective. The number of times that RW computes an extreme eigenpair is often around 5. For the numerical experiments reported in this paper, Matlab's eig routine was used to compute this extreme eigenpair. If this routine for computing an extreme eigenpair could be sped possibly using the Jacobi type methods of Sleijpen and Van der Vorst [22] or the truncated RQ iteration of Sorensen and Yang [25], the number of matrix-vector operations used in the parametric eigenvalue approach would be reduced. --R A trust region algorithm for nonlinearly constrained optimization A trust region strategy for non-linear equality constrained optimization Least squares with a quadratic constraint Quadratically constrained least squares and quadratic problems Matrix Computations Solving the trust-region subproblem using the Lanczos Method Applied Numerical Linear Algebra Iterative methods for nearly singular linear systems Graph partitioning and continuous quadratic programming a MATLAB package for analysis and solution of discrete ill-posed problem Geophysical Data Analysis: Discrete Inverse Theory Solution of sparse inde The Symmetric Eigenvalue Problem A trust region algorithm for equality constrained optimization A semide A Large-scale Trust-region Approach to the Regularization of Discrete Ill-posed Problems A Jacobi-Davidson iteration method for linear eigenvalue problems Newton's method with a model trust region modi Minimization of a large-scale quadratic function subject to a spherical constraint A truncated RQ iteration for large scale eigenvalue calculations Inverse Problem Theory --TR --CTR Peter A. Graf , Wesley B. Jones, A projection based multiscale optimization method for eigenvalue problems, Journal of Global Optimization, v.39 n.2, p.235-245, October 2007 Stanislav Busygin, A new trust region technique for the maximum weight clique problem, Discrete Applied Mathematics, v.154 n.15, p.2080-2096, 1 October 2006
symmetric successive overrelaxation;sparse optimization;arnoldi m orthogonalization;gauss-seidel;minimal residual;large-scale optimization;trust region subproblem;quadratic programming;krylov space;preconditioning;quadratic optimization
589318
Rescaling and Stepsize Selection in Proximal Methods Using Separable Generalized Distances.
This paper presents a convergence proof technique for a broad class of proximal algorithms in which the perturbation term is separable and may contain barriers enforcing interval constraints. There are two key ingredients in the analysis: a mild regularity condition on the differential behavior of the barrier as one approaches an interval boundary and a lower stepsize limit that takes into account the curvature of the proximal term. We give two applications of our approach. First, we prove subsequential convergence of a very broad class of proximal minimization algorithms for convex optimization, where different stepsizes can be used for each coordinate. Applying these methods to the dual of a convex program, we obtain a wide class of multiplier methods with subsequential convergence of both primal and dual iterates and independent adjustment of the penalty parameter for each constraint. The adjustment rules for the penalty parameters generalize a well-established scheme for the exponential method of multipliers. The results may also be viewed as a generalization of recent work by Ben-Tal and Zibulevsky [SIAM J. Optim, 7 (1997), pp. 347--366] and Auslender, Teboulle, and Ben-Tiba [ Comput. Optim. Appl., 12 (1999), pp. 31--40; Math. Oper. Res., 24 (1999), pp. 645--668] on methods derived from $\varphi$-divergences. The second application established full convergence, under a novel stepsize condition, of Bregman-function-based proximal methods for general monotone operator problems over a box. Prior results in this area required strong restrictive assumptions on the monotone operator.
Introduction denote the possibly unbounded n-dimensional "box" ([a This paper considers two closely-related problems, the minimization problem min f(x) (1) is a closed proper convex function, and the variational inequality where T is a (possibly set-valued) maximal monotone operator, and NB (x) denotes the cone of vectors normal to the set B at x. It is well known that, under mild regularity conditions, (1) is the special case of (2) for which the subgradient mapping of f . The last decade has seen considerable progress in the theory of proximal point methods based on generalized distances [11, 13, 19, 5, 21, 31, 14, 2, 3, 17]. Such methods use a scalar-valued regularization function to derive better-behaved versions of problems (1) and (2). In this article, we consider separable regularization terms of the form are scalar functions conforming to very general assumptions (see Assumption 2.1 below). In particular, we assume that as x 2 int B approaches the boundary of B, kr 1 D(x; y)k !1, where r 1 denotes the gradient with respect to the first vector argu- ment. The distance-like measure D can be, for example, the squared Euclidean distance, a Bregman distance [8], or a '-divergence [19] (see Section 2.2 below). Using these regularization terms, proximal methods for (1) take the form: x where ff k is a positive n-dimensional vector whose elements are called stepsizes. Note that we allow different stepsizes for each coordinate, as suggested by a variety of computational and theoretical studies [32, 5, 2, 3]. Moreover, since kr 1 D(x; x k )k !1 as x approaches the boundary of B, the regularization acts not only as a stabilizing proximal term but also as a kind of barrier function keeping the iterates within int B. In the case of the variational inequality (2), (3) generalizes to finding x k+1 satisfying the recursion RRR We derive some general results for these types of algorithms in Section 2, assuming that the stepsizes conform to a special rule that takes into account the curvature of the proximal term. This rule, although restrictive, appears to cover cases of the greatest practical interest; as we shall see, it covers the stepsize/penalty selection rules proposed in [32, 5, 2, 3]. Section 3 uses the results of Section 2 to obtain subsequential convergence results for the generalized proximal minimization algorithm (3). A critical application of (3), considered in Section 3.2, is when f is minus the dual function of a convex program such as s.t. where are differentiable convex functions. 1 We also assume that this problem is feasible, i.e., there is - y Choosing B to be any box containing the nonnegative orthant and f to be the negative of the dual function of (5), we may implement (3) via a multiplier method in which a sequence of unconstrained penalized versions of (5) must be solved. This construction leads to a class of multiplier methods that is extremely broad, subsuming both the classical quadratic augmented Lagrangian and the exponential method of multipliers [32, 6]. For these multiplier methods, our stepsize choice ensures that for indices i with x k the corresponding penalty term is augmented so it does not become so "flat" as to permit infeasibility of primal limit points. Empirically, the technique speeds convergence, and it also appears in a convergence rate analysis in [32] for the exponential method of multipliers case. Ben-Tal and Zibulevsky [5] have proved the optimality of the accumulation points of the exponential method, together with a class of proximal terms closely related to '- divergences, and their results are extended in [3]. Section 3 places such results in a broader context that includes Bregman distances. In Section 4, we restrict our attention to Bregman distances. It has been known for the better part of a decade that, when D(\Delta; \Delta) is any Bregman distance and the stepsizes do not vary by coordinate, the recursion (4) converges to a solution of the variational inequality (2) in various special cases: when the subdifferential of a closed proper convex function f , or when domT ' int B, meaning that all constraints must already be embedded in the operator T . In [9], these results were extended to "paramonotone" operators T , a category which includes as a special case. Unfortunately, many interesting practical cases, such as the subdifferential maps of saddle functions, are not paramonotone. More recently, Auslender et al. [2] have obtained strong results for general maximal monotone T , but only for a specific '-divergence choice of D(\Delta; \Delta). As noted in [4], these results can be extended to the (generally non-Bregman) case where D(\Delta; \Delta) is obtained by adding a quadratic to any member of the class \Phi 2 of [3]. 1 Actually, the results of Section 3.2 continue to hold [28] if one only supposes that are closed proper convex and assumes appropriate conditions on the effective domains of the objective and constraints, as in [24, Chapter 28]. However, this further generality makes the proofs more convoluted and is dropped for the sake of simplicity in the exposition. Page 4 RRR 35-99 Section 4 shows convergence, for general maximal monotone T , of the proximal method (4), where D(\Delta; \Delta) is a Bregman distance, to a solution of (2). We do impose some additional assumptions, derived from those of Section 2. First, we assume that the Bregman function used to construct the distance is twice-differentiable, which is not part of the standard Bregman function setup. Second, in addition to our general stepsize rule, we also require that the stepsizes do not vary by coordinate, that is, ff n for all k. The resulting condition is stronger than the usual requirement that the stepsize is simply bounded away from zero, but is crucial to the analysis, which blends the techniques of Section 2 with traditional Fej'er monotonicity arguments. Still, we have managed to substitute conditions on D(\Delta; \Delta) and ff k , which are parts of the algorithm, for conditions on T , which is part of the problem to be solved. Finally, we allow the calculations required for the recursions (3) and (4) to be performed approximately, as is likely to be necessary in practice. For the rescaling minimization case of Section 3, we adopt a constructive approximation criterion inspired by [17] and [29]. However, our criterion, which is tailored to the proximal minimization case, appears to be new. In the variational inequality analysis of Section 4, we use the simple, verifiable criterion of [14], although extension to the more sophisticated criterion of [29] may well be possible. In summary, the primary contributions of this paper are: ffl A novel convergence proof framework for a broad class of proximal algorithms. ffl Using this framework to establish subsequential convergence of a wide range of proximal minimization algorithms (3) with differing stepsize parameters for each coordinate; this result in turn leads to subsequential convergence of a broad class of multiplier methods with differing penalty parameters for each constraint. ffl Using the framework to show convergence of "interior" Bregman proximal point algorithms for maximal monotone operators, with a novel stepsize condition, but without the usual restrictive assumptions on the operator T . The new proximal minimization approximation criterion of Section 3 constitutes an additional contribution. Fundamental Analysis This section develops the fundamental analysis necessary for our results. We concentrate our attention on the variational problem (2), since it subsumes the minimization problem (1) under mild assumptions. In order to simplify the notation, we denote, for d 00 RRR We are now able to present the necessary assumptions on the functions d Assumption 2.1 For has the following properties: 2.1.1. For all y closed and strictly convex, with its minimum at y i . Moreover, int dom d i (\Delta; y 2.1.2. d i is continuously differentiable over (a exists and is strictly positive. 2.1.3. For all y essentially smooth [24, Chapter 26]. 2.1.4. There exist ae; ffl ? 0 such that if either The assumption of strict convexity is standard in generalized proximal methods. The assumption of twice differentiability is also quite common, although many existing results require only a once-differentiable d i . The essential smoothness assumption makes the distance act like a barrier function, forcing the iterates defined by the recursion (4), and hence its approximate version (6) below, to remain in the interior of the box B. In Section 2.2, we specialize these assumptions to the case of Bregman distances and '-divergences, where similar comments can be made. Finally, the fourth part of the assumption is new to the theory of generalized proximal methods, but is not very restrictive in practice. In particular, we show in Section 2.2 that, for Bregman distances and '-divergences, this condition can be written in terms of the kernels used to obtain the regularizations, and that it holds for most of the examples we are aware of. In addition, we make the following standard regularity assumption which, in view of the barrier function properties of d i , is required for any sensible application of (4): Assumption 2.2 domT " int B 6= ;. We are now able to present the proximal minimization algorithm: Rescaling Proximal Method for Variational Inequality (RPMVI) 1. Initialization: Let Choose a scalar c ? 0, and an initial iterate x 0 2 int B. 2. Iteration: (a) Choose ff k 2 R n such that ff k \Phi (b) Find x k+1 and e k+1 such that Page 6 RRR 35-99 (c) Let repeat the iteration. To guarantee the convergence of the RPMVI, we need additional assumptions on the stepsizes fff k and the error sequence fe k g; see Assumption 2.3 below. We define whence it is clear from (6) that Assumption 2.3 Let ffi k g be a real sequence converging to zero. The error sequence, fe k g, the regularization functions d and the stepsizes, fff k must be chosen in order to guarantee that: 2.3.1. ff 2.3.2. If - x is an accumulation point of fx k g, i.e., there is an infinite set K ' N such that x, then, for each or there is an infinite set K 0 ' K such that x Assumption 2.3.2 may seem artificial at this point, but Sections 3 and 4 will describe settings where it is easily verifiable. 2.1 Convergence Analysis We assume throughout this section that Assumptions 2.1 and 2.2 hold, and that sequences conforming to the recursions of the RPMVI algorithm and Assumption 2.3 exist. In Sections 3 and 4 we will present conditions which, in more specific settings, guarantee the existence of such sequences. Lemma 2.4 Let - x 2 R n be a limit point of fx k g, i.e., x k !K - x for some infinite set K ' N. Then for lim lim inf lim sup Proof. For each i, we consider the three possible cases: First, suppose i is such that - For the sake of a contradiction, assume that using Assumption 2.3.2, there is an infinite set K 0 ' K and a i ? 0 such RRR that for all k 2 K 0 , jfl k x i . Therefore -ff ff [Assumption 2.3.1] [Choice of ff k This result contradicts Next, consider the case - x suppose that lim inf k!K1 fl k using Assumption 2.3.2, there must be a i ? 0 and an infinite set K 0 ' K such that for all k 2 K 0 , ff -ff -cd 00 Let ffl be as in Assumption 2.1.4. If there is an infinite set K 00 ' K 0 such that x for all k 2 K 00 , we can conclude from the assumption that: 2d 00 aecd 00 since x for sufficiently large k 2 K 0 . Page 8 RRR 35-99 As d i (\Delta; x its minimum at x k\Gamma1 implies that d 0 Hence ff for sufficiently large k 2 K 0 , a contradiction with Finally, the case of - is analogous to the case - Lemma 2.5 Let - x be a limit point of fx k g, i.e., x k !K - x for some infinite set K ' N. Proof. By Assumption 2.2, there must exist some e (ex). The monotonicity of T implies that, for all k - 0, We will show that unboundedness of ffl k gK would contradict this inequality for some sufficiently large k. is unbounded, there must exist an infinite K 0 ' K such that ffl k gK 0 converges in , with at least one ffl k implies that for each unbounded coordinate i, either or Therefore, for each unbounded coordinate of ffl k gK 0 , we have or On the other hand, for coordinates such that ffl k also bounded. Thus, for sufficiently large k 2 K must be negative, contradicting (9). 2 Finally, the main convergence theorem for the RPMVI follows: RRR Theorem 2.6 If fx k g is a sequence generated by the RPMVI algorithm with Assumptions 2.1, 2.2, and 2.3 holding, then all the limit points of fx k g are solutions to the variational inequality problem (2). Proof. Let - x be any limit point of fx k g, i.e., x k !K - x, for some infinite set K ' N. From Lemma 2.5, we know that the corresponding sequence must exist some K 0 ' K with must be outer semicontinuous [27, 12.8(b)], it follows that - fl 2 T (-x). Lemma 2.4 implies that and these conditions are equivalent to Incidentally, it is possible to eliminate the requirement of twice-differentiability of d i (\Delta; y i ), at the cost of some additional complexity in the description of the method. Specifically, consider replacing Assumption 2.1.4 with the condition that there exist functions If the stepsizes are now selected so that for some scalar c ? 0, we have for all and k - 0 that ff k the conclusions of Theorem 2.6 continue to hold. We may examine this variation of the analysis in subsequent research. The present approach is equivalent to taking L i (y choice since d 00 (y rate of change of d 0 (\Delta; y i ) around y i . 2.2 Some examples of d i functions We present some example of d i functions that conform with Assumption 2.1. In particu- lar, we show that two classes of regularizations widely studied in the literature, Bregman distances [11, 13] and '-divergences [19], conform to the assumption under very mild restrictions 2.2.1 Bregman distances Bregman distances were introduced in [8] and have been studied in the context of proximal methods in [11, 12, 13], as well as many subsequent works. To construct each regularization one uses an auxiliary convex function h i and defines d i Nonseparable distances can also be constructed in a similar way, but the separable case is the most common. The following properties guarantee that Assumption 2.1 holds for such Assumption 2.7 For has the following properties: 2.7.1. h i is closed, int continuously differentiable, with a strictly positive second derivative throughout (a 2.7.2. h i is essentially smooth. 2.7.3. There exist ae ? 0 and ffl ? 0 such that if either Note that Assumption 2.7.1 implies that each h i is strictly convex. Assumption 2.7.3 corresponds to Assumption 2.1.4, since d 00 Fortunately, it is not very restrictive. Consider the case of finite a i . Since lim x we know that h 00 must be unbounded above as x i & a i . To violate the assumption, h 00 would have to oscillate unboundedly as x i & a i . As far as we are aware, every separable Bregman function proposed so far conforms not only to Assumption 2.7.3, but to a more stringent, easier-to-verify condition, as follows: Lemma 2.8 If there is an ffl ? 0 such that for all x i is non-increasing, and for all x 2 (b i is non-decreasing, then Assumption 2.7.3 holds. Proof. Suppose that a Therefore, Assumption 2.7.3 holds with 1. The case b i ! 1 is analogous. 2 Examples of functions h i where all these assumptions hold are: log x, with a with with a Finally, we note that for finite a i we do not yet assume that h i must approach a finite limit as x i & a i , nor similarly for x i Such an assumption is quite common in the theory of Bregman distances [11, 13, 9, 29], but, similarly to [21], it is not needed for the results of Section 3 below. We will use it, however, in the variational inequality analysis of Section 4. RRR 35-99 Page 11 2.2.2 '-divergences The '-divergence regularizations have been studied in the context of proximal methods, for example, in [19], and more recently in [5, 3]. In these works, the box considered is the positive orthant, i.e., . An auxiliary strictly convex scalar function ' is used to define the distance d i , but this time by: The following hypotheses can be used to guarantee Assumption 2.1 when Assumption 2.9 The function ' : R! (\Gamma1; +1] is such that: 2.9.1. ' is closed and convex, with int 2.9.2. ' is twice differentiable on (0; +1), with ' 00 (t) ? 0 for all t ? 0; 2.9.3. 2.9.4. ' is essentially smooth; 2.9.5. There exists a ae ? 0 such that ae' 0 (t) - ' 00 Slight variations on these assumptions appear, for example, in [5, 3], together with the following examples: The next lemma states that Assumption 2.9.5 above implies Assumption 2.1.4: Lemma 2.10 Let (a be defined as in (10). Then Assumption 2.1.4 is equivalent to the existence of a ae ? 0 such that ae' 0 (t) - ' 00 Proof. First we observe that: d 00 and so d 00 Page 12 RRR 35-99 Therefore, Assumption 2.1.4 reduces to Taking letting y i range over (0; x i ], and setting Conversely, if (12) is true, (11) holds for an arbitrary choice of ffl ? We note that in [5], one assumes that the iterations are of the form: where each ff k i is greater than c=x k being a positive constant. In [2, 3], this property is guaranteed by redefining the distance measure to be ~ ~ and assuming stepsizes bounded away from zero. In this case, the iteration is ~ with lim inf k!1 e rewriting the iteration with respect to D, instead of ~ D, we recover the rule from [5]. It turns out that these techniques are a special case of our stepsize choice rule, which gives in the case of a '-divergence that which is identical if one redefines the constant factor c. Thus, the reader should note that the class of '-divergences described by Assumption 2.9 encompasses the regularizations studied in [5, 2, 3]. In particular, it includes the classes \Phi 1 and \Phi 2 described in [3]. However, the stepsize rule in the RPMVI is more stringent than the one in [5, 2, 3], as it also assumes that the stepsize is bounded away from zero. To overcome this slight restriction, we point out that the assumption ff k used here only in the first part of the proof of Lemma 2.4, and it can be replaced by the assumption that d 00 continuous and strictly positive over (a This condition holds for '-divergences, since d 00 In this sense, the results here can be seen as extensions of those in [5, 2, 3]. RRR 3 Proximal Minimization Methods with Rescaling This section applies the analysis of the RPMVI method to the minimization problem (1). We leave Assumption 2.1 as a standing assumption; we also make the following standard regularity assumption, which in view of the barrier function properties of D, is required for any sensible application of (3): Assumption 3.1 dom f " int B 6= ;. Note that, since int B is open, this assumption implies that ri dom f " int B 6= ;, which implies that dom @f " int B 6= ;. Then, using [24, Theorem 23.8], one can show that the minimization problem (1) is equivalent to the variational inequality problem (2) with Moreover, Assumption 2.2 holds. Then, we specialize the RPMVI to: Rescaling Proximal Minimization Method (RPMM) 1. Initialization: Choose c ? 0 and oe 2 [0; 1]. Choose nonnegative scalar sequences fs k g and fz k g with 2. Iteration: (a) Choose ff k 2 R n such that ff k \Phi (b) Find x oe ae s k+1 oe with the standing convention that min \Psi is z k+1 whenever (c) Let repeat the iteration. Note that if one chooses s k ; z reduces to the "constructive" criterion reminiscent of [29]. Page 14 RRR 35-99 3.1 Convergence analysis We start by showing that the iteration step is well defined if f is bounded below on B: Lemma 3.2 If f is bounded below on B, then there is a unique point that solves the iteration step of the RPMM with e a solution to (13)-(14) exists if f is bounded below on B. Proof. Let ' be a lower bound of f on B. Given i 2 R, the level set This last set is a level set of i ) on B, which must be bounded, since by Assumption 2.1.1 this function attains its minimum at the unique point x k [24, Corollary 8.7.1]. Therefore, f(\Delta) attains a minimum on B. The uniqueness of the minimum follows from the strict convexity of D(\Delta; x k ). 2 To apply the convergence analysis of the previous section to the sequence fx k g computed by the RPMM, it suffices to show that Assumption 2.3 holds. Verification of Assumption 2.3.1 is straightforward: Lemma 3.3 With the definition ae s k oe for all k - 1, Assumption 2.3.1 holds for the RPMM. Proof. From the nonnegativity of fs k g and fz k g, it follows that ffi k g is also nonnegative. one also has fi k ! 0. Moreover, since oe 2 [0; 1], oe ff for all k, so Assumption 2.3.1 holds. 2 As in (7), we define for all k - 0 and and let fl k 2 R n be the vector with elements Lemma 3.4 RRR 35-99 Page 15 Proof. The claim that fl k 2 @f(x k ) follows from the definition of fl k . For the second claim, we have, using the convexity of d i (\Delta; x ff Using (14), it then follows that ff ae s k ae s k oe fi fi x ae s k oe \Gammas k :Before proving the next result, we state a helpful technical lemma: Lemma 3.5 [22, Section 2.2] Suppose fa k g, ffl k g ae R are sequences such that fa k g is bounded below, exists and is finite, and the recursion a k+1 - a k holds for all k. is convergent. It is now possible to establish that Assumption 2.3.2 also holds: Lemma 3.6 If f is bounded below on B, then ff(x k )g is convergent and Hence Assumption 2.3.2 holds for the RPMM. Proof. Using Lemma 3.4, ns Then, recalling that fs k g is summable, Lemma 3.5 implies that ff(x k )g is a convergent sequence. For Page Using Lemma 3.4 once again, it follows that Taking limits, we conclude that fl k Thus, Theorem 2.6 implies the optimality of all accumulation points of the sequence g. We strengthen this observation below: Theorem 3.7 Suppose that Assumptions 2.1 and 3.1 hold, and that f is bounded below on B. If fx k g has a limit point, then ff(x k )g converges to the infimum of f on B and all limit points of fx k g will be minimizers of f on B. A condition that guarantees the existence of limit points of fx k g is the boundedness of the solution set, or any other level set of f . Proof. As just noted, Lemma 3.6 implies that Assumption 2.3.2 holds, and so Assumption 2.3 holds in its entirety. Assumption 2.1 holds by hypothesis, and, setting Assumption 3.1 implies Assumption 2.2. Thus, the conclusions of Theorem 2.6 apply. Let x be a limit point of fx k g, i.e. x k !K - x, for some infinite set K ' N. Theorem 2.6 asserts that Assumption 3.1, - x is a minimizer of f on B. Moreover, since Lemma 2.5 states that ffl k gK is bounded, and since ff(x k )g is convergent by Lemma 3.6, min lim Therefore, lim k!1 f(x k Finally, the boundedness of any level set of a proper closed convex function implies boundedness of all level sets [24, Corollary 8.7.1], and Lemma 3.6 states that ff(x k )g is convergent, consequently it is bounded. So, fx k g is also bounded and has limit points. 2 3.2 Multiplier Methods We now discuss applying the RPMM to the dual of the convex program (5) to obtain multiplier methods. The use of proximal methods to derive multiplier methods for constrained convex optimization is a now-classical subject and may be traced to the seminal paper [26]. In the context of generalized proximal methods, applications can be found, for example, in [30, 13, 19, 21, 31, 3, 17]. In this section, we consider only the case in which the proximal step is done exactly, i.e., we will let e as in [30, 13, 19, 17]. Unfortunately, our approximate-step acceptance rule for the RPMM does not translate directly to an easily verifiable acceptance criterion for an approximate solution of the penalized problem (17) be- low. However, partial results in this direction may be obtained under stringent assumptions on the original problem (5); see Appendix B. A criterion in the spirit of (14) that does not depend on such assumptions is the subject of ongoing research [15]. We further observe that the approximation criteria of [17, 29] also do not translate readily to a multiplier method RRR 35-99 Page 17 setting. On the other hand, under the assumption that the primal objective function g 0 is strongly convex, [26, 21, 3] present some inexact multiplier methods based on a rather different acceptance rule involving optimizing the augmented Lagrangian function to within some tolerance ffl of its minimum value. Consider the convex problem (5), and let ffi C denote the indicator function of a convex set C. Then we define f to be minus the dual function associated with (5), plus The dual problem to (5) is then equivalent to the minimization of f . Furthermore, we assume Assumption 3.83.8.1. The primal problem (5) has a finite optimal value, and it conforms to the Slater condition. 3.8.2. For all conform to Assumption 2.1 for a i - 3.8.3. There is an - x ? 0 such that - x 2 dom f , where f is as defined in (15). This assumption has the following consequences: Assumption 3.8.1 implies that the dual solution set is non-empty and bounded [16] and that there is no duality gap. Assumption 3.8.3 implies that Assumption 3.1 holds for f as defined by (15). Under Assumption 3.8, if we fix e each iterate x k+1 of the RPMM applied to the negative dual functional f may be calculated by the following multiplier method whenever the unconstrained problems (17) have solutions: \Phi d \Phi denotes the monotone conjugate [24, p. 111] with respect to the first argument, that is, d \Phi )g. 3 Theorem 3.10 below gives conditions guaranteeing that a y k+1 satisfying (17) exists. We relegate the technical aspects of the proof of the equivalence of (16)-(18) to the RPMM applied to the f defined in (15) to Appendix A, since they are very similar to earlier 2 The case a is of interest because it includes the classical method of multipliers for problems with inequality constraints [26], along with various extensions described in [13, 20]. 3 The classical conjugate / of a function / is defined [24, Chapter 12] via / for any / : R n ! (1; +1]. The monotone conjugate of / is then the classical conjugate of , that is, Page proofs for various special cases of (17)-(18), for example in [30, 13, 19, 21, 17]. In particular, Corollary A.4 establishes the equivalence of the two calculations. Given this equivalence, Theorem 3.7 asserts the subsequential convergence of the sequence to a dual solution of (5). For the primal sequence, however, it has historically been harder to prove good behavior. For example, in the case of Bregman distances, a guarantee of feasibility of primal accumulation points has relied on stringent assumptions like R n ae int B, as in [13], or strict complementarity [18]. In the case of the RPMM, with its strong stepsize restrictions, the feasibility, and therefore optimality, of accumulation points of fy k g is easily demonstrated. Theorem 3.9 Suppose that Assumption 3.8 holds. Pick a scalar c ? 0, let x 0 2 R n and suppose that it is possible to obtain a sequence f(ff k that obeys the recursions (16)- (18). Then, fx k g is bounded and all its accumulation points are solutions of the dual of (5). Moreover, lim sup lim and fg 0 (y k )g converges to the optimal value of the primal problem (5). Therefore, any accumulation point of fy k g solves the primal problem. Proof. As shown in Corollary A.4, the sequence fx k g is the same as would be computed by using the RPMM to solve the dual problem, that is, to minimize f . In particular, fx k g and all its limit points must be nonnegative. Moreover, the Slater condition implies that the dual function has bounded level sets. Then, the boundedness of fx k g and the optimality of its limit points follow from Theorem 3.7. Let us analyze the primal sequence. For each the same role as in (7), with e k Let fx k gK be any convergent subsequence of fx k g, and - x the respective accumulation point, x k !K - x. Lemma 2.4 implies that As fx k g is bounded, the above relations imply that RRR 35-99 Page 19 Now, suppose for the purposes of contradiction that (20) does not hold. Then, for some must be an infinite set K ae N and an ffl ? 0 such that is bounded, there exists a refined subsequence K 0 ' K such that fx k g K 0 is convergent, with limit - imply that g i (y k ) !K 0 \Gamma1. Since Lemma 2.5 asserts that fi k is bounded, we can conclude that i k \Gamma1. However, this divergence would imply that x k i should be 0 for infinitely many k 2 K 0 ' K, once again a contradiction of (23). Therefore, lim and (20) holds. Finally, we prove that fg 0 (y k )g converges to the optimal value. We may use (17), (18), and the chain rule to see that y k minimizes the Lagrangian corresponding to the primal problem with the fixed multiplier x k . Hence, Let \Gammaf denote the dual optimal value, which is equal to the primal optimal value since there is no duality gap. Theorem 3.7 states that f(x k Taking limits in (25) and using (24), it follows that lim The feasibility and optimality of the accumulation points of fy k g are then consequences of the continuity of g i , Finally, it is natural to seek conditions under which the penalized subproblems (17) must have solutions, and the primal sequence fy k g is bounded. The following result addresses these questions under the standard assumption of a bounded solution set: Theorem 3.10 Suppose that the primal solution set is bounded. Given any ff k ? 0 and exist satisfying the recursions (17)-(18). Moreover, the primal sequence fy k g is bounded. Proof. For the first assertion, it suffices to show that the penalized problems (17) have solutions. Given any closed proper convex function /, we define its recession function /1 via dom/ may be chosen arbitrarily [24, Theorem 8.5]. The boundedness of the primal solution set is equivalent [7, Section 5.3] to: Page 20 RRR 35-99 Thus, the existence of a solution to (17) is a corollary of Lemma A.5 in the appendix, along with the sum rule for recession functions [24, Theorem 9.3]. We now prove that fy k g is bounded. Theorem 3.9 shows that the sequences fg i (y k )g, above. From (27), unboundedness of fy k g would imply that . But such unboundedness would contradict g 0 (y k )'s convergence to the optimal value. 2 We remark that the penalty parameter adjustment rule (16), as discussed in Section 2.2.2, essentially subsumes, in a context broader than '-divergences, the corresponding rules described in [32] for the exponential method of multipliers and in [5, 3, 4] for a general '-divergence setting. We end this section giving some examples of d \Phi functions that may be derived from separable Bregman distances (see Section 2.2.1). Further examples may be obtained from [21, 28]. For a Bregman-derived distance, we have d i whence d \Phi where h \Phi denotes the standard monotone conjugate of h. Note that when such a d \Phi used in the minimization operation in (17), the additive terms h i (w are constant and may be discarded. The following examples may now be easily verified: may be disregarded; this choice gives the classical quadratic method of multipliers for inequality constraints. where the \Gammaw i term may be disregarded, yielding the exponentional method of multipliers. 4 Bregman Interior Point Proximal Methods for Variational Inequalities We now turn our attention to the box-constrained variational inequality problem (2), where (possibly set-valued) maximal monotone operator. In this section, we confine ourselves to Bregman distances, as defined in Section 2.2. We augment Assumption 2.2 as follows: RRR 35-99 Page 21 Assumption 4.1 T is maximal monotone, the solution set of (2) is non-empty, and there exists some e Our goal is to show convergence of an approximate version of the iteration (4), without further conditions on T . We modify and extend Assumption 2.7 as follows: Assumption 4.2 For have the same properties specified in Assumption 2.7, and furthermore, h i is continuous on [a defining 4.2.1. For all x 2 B and ff 2 R, the level set fy 2 int B j D h (x; y) - ff g is bounded. 4.2.2. If fx k g ae int B converges to x 2 R n , then lim k!1 D h (x; x k 4.2.3. rge h Note that at finite a i 's and b i 's, the corresponding h i is now required to take a finite value. The algorithm can now be stated: Box Interior Proximal Point Algorithm (BIPPA) 1. Initialization: Let 2. Iteration: Choose ff k such that ff k - c maxf1; h 00(x )g. Find vectors repeat the iteration. 4.1 Convergence analysis First, we cite a result showing that the iteration step of BIPPA is well defined: Lemma 4.3 [13, Theorem 4(i)] Under Assumption 4.2, there is a unique point x k+1 that solves the iteration step (28) of the BIPPA with e We note that it is shown in the unpublished dissertation [28] that (28) has a unique exact solution even if Assumption 4.2.3 does not hold. This result permits one to dispense completely with Assumption 4.2.3. However, the proof, while essentially a minor modificiation of that of [1, Theorem A.1], is quite involved, so we do not include it here. To guarantee the convergence of the BIPPA, we must assume some vanishing behavior for fe k g; we will use the assumptions of [14]. Although not as general as the criterion used in RPMM, these conditions are better suited to our analysis, since they will permit us to use properties associated with Fej'er monotonicity, and are still feasible to enforce computationally. Page 22 RRR 35-99 Assumption 4.4 [14] The error sequence fe k g conforms to:X exists and is finite. Note that this assumption implies that Assumption 2.3.1 holds with k1 . We now state some necessary lemmas: Lemma 4.6 If Assumption 4.4 holds, then the sequence fx k g is bounded and D h Proof. The result will follow from [14, Lemma 3] once we show that, for z 2 (T +NB E(z) exists and is finite. But,X and Assumption 4.4 implies that the right hand side of this relation is finite. Hence, exists and is finite. Using Assumption 4.4 once more, we conclude that E(z) exists and is finite. 2 We also use a key result from Solodov and Svaiter [29]: Theorem 4.7 [29, Theorem 2.4] Let h i satisfy Assumption 4.2. Given two sequences fx k g ae B and fy k g ae int B, either one of which is convergent, with lim k!1 D h the other sequence also converges to the same limit. This theorem implies that Bregman function in the classical sense [8, 10]. Using Theorem 4.7 and Lemma 4.6, we derive: Corollary 4.8 Under Assumptions 4.1, 4.2, and 4.4, fx k g has at least one limit point. Moreover, if for some infinite set K ' N, we have x k !K - x, then x x. Therefore, Assumption 2.3.2 holds. RRR Before presenting the main convergence theorem for the BIPPA, we present a final technical lemma that will help us to prove the uniqueness of the accumulations points of fx k g. Lemma 4.9 Under Assumption 4.4, for all z converges to a value in [0; +1) which we will denote by d(z). Proof. Consider any z 2 (T implies that (29) holds. Using Assumption 4.4 and D h the hypotheses of Lemma 3.5 are satisfied with a converges, necessarily to a nonnegative value. 2 Now, the main convergence theorem follows: Theorem 4.10 Under Assumptions 4.1, 4.2, and 4.4, fx k g converges to a solution of Proof. Let - x be an accumulation point of fx k g, i.e. x k !K - x, for some infinite set K ' N. Such a point exists by Lemma 4.6. From Theorem 2.6, We now prove the uniqueness of the limit point: from Assumption 4.2.2, we know that as defined in Lemma 4.9, is zero. Suppose that fx k g has another accumulation point x k !K 0 x 0 for some infinite set K 0 ' N. We then have that it follows from Theorem 4.7 that x Another possible application of our fundamental analysis is to try to generalize the idea of adding the square of the Euclidean norm and an arbitrary generalized distance to obtain Fej'er monotonicity to solutions of (2), as in [2, 3] for the special case of '-divergences. The difficulty here is to generalize the condition that defines the class \Phi 2 in [3]. This topic is the subject of ongoing research. --R An interior-proximal methods for convex linearly constrained problems and its extension to variational problems A logarithmic-quadratic proximal method for variational inequalities Interior proximal and multiplier methods based on second order homogeneous kernels. Modified Lagrangian Methods for Variational Inequality Problems. Penalty/barrier multiplier methods for convex programming problems. Nonlinear Programming Constrained Optimization and Lagrange Multiplier Methods. The relaxation method of finding the common point of convex sets and its application to the solution of problems in convex programming. An interior-point method with Bregman functions for the variational inequality problem with paramonotone operators An iterative row-action method for interval convex program- ming The proximal minimization algorithms with D-functions A convergence analysis of proximal-like minimization algorithms using Bregman functions Nonlinear proximal point algorithms using Bregman functions Approximate iterations in Bregman-function-based proximal algorithms A Practical General Approximation Criterion for Methods of Multipliers Based on Bregman Distances. A necessary and sufficient condition to have bounded multipliers in nonconvex programming. Strict convex regularizations Augmented Lagrangian methods and proximal points methods for convex optimization. On the twice differentiable cubic augmented Lagrangian. Proximal minimization methods with generalized Bregman functions. Introduction to Optimization. Extension of Fenchel's duality theorem for convex functions. Convex Analysis. Conjugate Duality and Optimization. Augmented Lagrangians and applications of the proximal point algorithm in convex programming. Variational Analysis Springer-Verlag T'opicos em M'etodos de Ponto Proximal. An inexact hybrid generalized proximal point algorithm and some new results on the theory of Bregman functions. Entropic proximal mappings with applications to nonlinear programming. Convergence of proximal-like algorithms On the convergence of the exponential multiplier method for convex programming. --TR --CTR Alfred Auslender , Paulo J. Silva , Marc Teboulle, Nonmonotone projected gradient methods based on barrier and Euclidean distances, Computational Optimization and Applications, v.38 n.3, p.305-327, December 2007 Paulo J. Silva , Jonathan Eckstein, Double-Regularization Proximal Methods, with Complementarity Applications, Computational Optimization and Applications, v.33 n.2-3, p.115-156, March 2006
convex programming;varphi-divergence;proximal algorithms;variational inequalities;bregman distances
589323
On the Complexity of a Practical Interior-Point Method.
The theory of self-concordance in convex optimization has been used to analyze the complexity of interior-point methods based on Newton's method. For large problems, it may be impractical to use Newton's method; here we analyze a truncated-Newton method, in which an approximation to the Newton search direction is used. In addition, practical interior-point methods often include enhancements such as extrapolation that are absent from the theoretical algorithms analyzed previously. We derive theoretical results that apply to such an algorithm, one similar to a sophisticated computer implementation of a barrier method. The results for a single barrier subproblem are a satisfying extension of the results for Newton's method. When extrapolation is used in the overall barrier method, however, our results are more limited. We indicate (by both theoretical arguments and examples) why more elaborate results may be difficult to obtain.
Introduction . In their 1993 book [16], Nesterov and Nemirovsky derive complexity results for convex optimization problems. Their basic algorithm is an interior-point method where each subproblem is solved using a damped Newton method. If a nonlinear optimization problem is large (and hence complexity is an important issue) then Newton's method is not normally used because of its computational costs, so these results might be considered primarily of theoretical interest. Our goal in this paper is to derive comparable complexity results for algorithms that more closely resemble practical interior-point algorithms for large-scale optimization (see, e.g., [1, 9, 10, 11, 13, 19, 23]). The interior-point method we analyze is strongly related to the barrier method in [13]. The essential features of this algorithm are that each barrier subproblem is solved approximately using a truncated-Newton method; then the solutions to the subproblems are extrapolated to obtain an initial guess for a new subproblem. Many of the enhancements discussed in [13]-such as preconditioning, a specialized matrix-vector product, and a numerically stable formula for the search direction-fit into the theoretical framework used here. The major exception is the line search (see below). We derive a bound on the number of truncated-Newton iterations required to solve a barrier subproblem to within some tolerance. Each truncated-Newton iteration involves the approximate solution of the Newton equations via (say) the conjugate-gradient method, requiring at most O(n 3 ) computations in exact arithmetic, although typically the number of computations would be O(n) or O(n 2 ), in problems where the Hessian matrix is sparse. In the algorithm analyzed here, a prescribed step length is used, so there is no line search. (Here is where the theoretical and practical algorithms *Received by the editors Month? Date?, 199?; accepted for publication (in revised form) Month? 1997. yOperations Research and Engineering Department, George Mason University, Fairfax, VA 22030. The work of this author was supported by National Science Foundation grant DMI-9414355. zOperations Research and Engineering Department, George Mason University, Fairfax, VA 22030. The work of this author was supported by National Science Foundation grant DMI-9414355. G. NASH AND ARIELA SOFER differ, since a practical method would likely use an adaptive line search based on minimizing a one-dimensional approximation to the barrier function.) Ignoring the computations for evaluating the gradient and Hessian, the algorithm determines the solution to within a tolerance in a number of operations that is polynomial in M and the problem dimensions. If polynomial algorithms exist to evaluate the gradient and Hessian, the overall algorithm for the barrier subproblem is polynomial in the dimensions of the optimization problem. The theoretical result that we obtain reduces to the result for Newton's method if the inner convergence tolerance for the truncated-Newton method is set to zero. For this reason, we consider this result to be a satisfying extension of the theory for Newton's method. In the second major part of the paper we analyze how a simple linear extrapolation scheme can accelerate the algorithm by providing improved initial guesses for each subproblem. We show that improved performance can be achieved by an algorithm based on linear extrapolation, when barrier subproblems are solved exactly. We also indicate, via an example, that it may be difficult to derive complexity results either when subproblems are solved inexactly, or when higher-order extrapolation is used. Our analysis is based on the framework established in Nesterov and Nemirovsky (1993) and Nemirovsky (1994) as adapted by us in our book [14]. In the rest of the paper we cite theoretical results for an algorithm based on Newton's method. These results are due to Nesterov and Nemirovsky, although we frequently cite [14] because our discussion here more closely follows the organization and notation of that book. (Related discussions can be found in [3, 8, 22].) In the barrier method, we assume that the barrier function is "self concordant," a property that we define below. Self-concordant barrier functions were introduced in [16]; recent work on this topic includes [4, 6, 7, 8, 17]. For linear programs and convex quadratic programs, the ordinary logarithmic barrier function can be used. Related barrier functions can be used for semi-definite programming [18, 20, 21, 24]. It is possible to prove that, for any convex feasible region with the properties we specify below, there exists an appropriate self-concordant barrier function, although in general it may not be practical for computation. Thus the results we describe here provide a general theoretical approach for solving convex programming problems. Two major results are required to prove that the overall algorithm is a polynomial algorithm. The first states that if the truncated-Newton method is applied to a single barrier subproblem, and the initial guess is "close" to the solution, then the number of iterations required to find an approximate solution of this subproblem is bounded. This is the topic of Section 3. The second states that the linear extrapolation can improve the initial guess for the next subproblem. This topic is addressed in Section 4. These two results are less obvious than they might at first seem. If any constraint in the convex programming problem is binding at the solution, then the solution will be on the boundary of the feasible region where the barrier function has a singular- ity. Since standard convergence results for Newton-type methods assume that the Hessian at the solution has a bounded condition number, a traditional analysis is not appropriate. To analyze the behavior of Newton-type methods in this case, we must in some manner take this singularity into account. To do this, we define a norm k\Deltak x in terms of the Hessian of the barrier function evaluated at a point x. We will measure "closeness" in terms of this norm. This norm depends on the Hessian, and changes as the variables change. For this COMPLEXITY OF A PRACTICAL INTERIOR-POINT METHOD 3 norm to be useful, the rate of change of the Hessian matrix must not be "too great." This reasoning leads to the imposition of a bound on the third derivatives of the barrier function in terms of the Hessian (see Section 2). This bound is all that is required to prove the first major result corresponding to the behavior of the truncated-Newton method on a single barrier subproblem. To prove that the approximate solution of one subproblem will not be too far from the solution of the next subproblem, it is necessary that the values of the barrier functions not change "too quickly" as the barrier parameter changes. To guarantee this, we impose a bound on the first derivatives of the barrier functions in terms of the Hessian (see Section 4). By measuring all quantities in terms of the Hessian, we are able to circumvent the difficulties associated with the singularity of the barrier function at the solution. If the barrier function has these properties, then an interior-point method can be designed so that the optimal solution of a convex programming problem can be found (to within some tolerance) using a polynomial number of truncated-Newton iterations. Practical experience suggests that an improved barrier method can be obtained if the approximate solutions to the subproblems are extrapolated to produce an initial guess for the next subproblem. We analyze this idea in Section 4. We are able to derive some theory to support this idea in the case when linear extrapolation is used, and when the subproblems are solved exactly. Through an example, we suggest that it may be difficult to obtain a comparable result for either a more elaborate algorithm (with higher-order extrapolation) or a more realistic algorithm (with the subproblems solved inexactly). 2. Basics. In this section, we define self concordance and establish some basic lemmas. An extensive discussion of the theory of self concordance is given in the book by Nesterov and Nemirovski [16]. The presentation here parallels our book [14]. Let S be a bounded, closed, convex subset of ! n with nonempty interior int S. (The assumption that S is bounded is not that important, since we could modify the optimization problem by adding artificial, very large bounds on the variables.) Let F be a convex function defined on the set S, and assume that F has three continuous derivatives. Then F is self concordant on S if: (i) (barrier property) F along every sequence f x i g ae int S converging to a boundary point of S. (ii) (differential inequality) F satisfies for all x 2 int S and all In this definition, that is, it is a third-order directional derivative of F . As an example, the logarithmic barrier function log(a T is self concordant on the set \Psi . 4 STEPHEN G. NASH AND ARIELA SOFER The constant 2 in the definition is arbitrary. If instead for some constant C, then the scaled function - concordant. The number 2 is used in the definition so that the function F log x is self concordant without any scaling. We make several assumptions to simplify our discussion. They are not essential; in fact, almost identical results can be proved without these assumptions. We assume that r 2 F (x) is nonsingular for all x 2 int S. (See Theorem 2.1.1 in [16] for an approach that avoids this assumption.) This allows us to define a norm as follows: We also assume that F has a minimizer x 2 int S. Because F is convex, these assumptions guarantee that x is the unique minimizer of F in S. The following lemmas indicate some basic properties of self-concordant functions. The first shows that the third-order directional derivative can be bounded using this norm. Lemma 1. If F is self-concordant on S then for all x 2 int S and for all Proof. See [8] or [16]. The next lemma bounds how rapidly a self-concordant function F (x) and its Hessian can change if a step is taken whose norm is less than one. The first result is an analog of a Taylor series expansion for a self-concordant function. The second is a bound on how rapidly the norm can change when x changes. Lemma 2. Let F be self concordant on S. Let x 2 int S and suppose that khk x ! 1. Then x where The lower bound in (1) is satisfied even if khk x - 1. Furthermore, for any g 2 ! n , Proof. See [16]. Our convergence results for a barrier subproblem are phrased in terms of a quantity called the "Newton decrement." It is defined below. The Newton decrement measures the norm of the Newton direction, but indirectly it can be interpreted as a "proximity measure" for the distance to the barrier trajectory. We use the Newton decrement in place of more traditional measures of convergence, such as COMPLEXITY OF A PRACTICAL INTERIOR-POINT METHOD 5 If x 2 int S and pN is the Newton direction for F at x, then the Newton decrement of F at x is Consider the Taylor series approximation to F The Newton direction pN minimizes this approximation and is the solution to The optimal value of the Taylor series approximation is indicating why ffi (F; x) is called the Newton decrement. We have the following lemma. Lemma 3. The Newton decrement satisfies Proof. See [16]. We will obtain bounds on F in terms of the Newton decrement, and we will also measure the progress at each iteration of the truncated- Newton method in terms of the Newton decrement. Thus, statements about the convergence of the method in terms of the Newton decrement will indirectly provide us with information about convergence as measured in the more traditional ways. 3. Convergence of a Truncated-Newton Method. We now study the consequences of using a truncated-Newton method, rather than Newton's method, to minimize a self-concordant function: minimize where S and F are as in the previous section. In the truncated-Newton method, a search direction p will be computed that satisfies the acceptance criterion where x is the current estimate of the solution to the barrier subproblem, pN is the Newton direction, and ffl is some tolerance. For simplicity, we assume that the tolerance ffl is fixed, although similar results could be derived in the case where ffl varied from iteration to iteration. It is not that important how the search direction is computed, as long as the number of arithmetic operations is polynomial in the dimensions of the problem. Practical truncated-Newton methods often use the conjugate-gradient method which, in exact arithmetic, is guaranteed to converge to the Newton direction in a finite number of operations. Thus, in exact arithmetic, this would be an appropriate procedure. 6 STEPHEN G. NASH AND ARIELA SOFER The acceptance criterion (2) is impractical since it involves the Newton direction pN . It is, however, closely related to practical rules for terminating the inner iteration of a truncated-Newton method. If we define to be the value of the quadratic model for F (x) at p, then x x In [12] it is recommended that the inner iteration of a truncated-Newton method be terminated based on the value of the quadratic model, and the barrier method in [13] uses a related rule. Other practical acceptance rules are based on the value of the relative residual, and have the form for some tolerance j [2]. It is straightforward to derive thatp cond 2 (r 2 F (x)) where is the condition number of r 2 F (x) in the 2-norm. These inequalities provide a further demonstration of the relationship of (2) with practical termination rules for the inner iteration of the truncated-Newton method. Some useful consequences of the acceptance criterion (2) are stated in the following lemma. Lemma 4. Suppose that the acceptance criterion (2) is satisfied at x. then and Proof. The first result is a straightforward consequence of (2). The second is obtained by squaring the acceptance criterion: COMPLEXITY OF A PRACTICAL INTERIOR-POINT METHOD 7 The function F will be minimized using a "damped" truncated-Newton method, that is, a step is taken along the truncated-Newton direction but with a specified step length that is less than one. If we denote the search direction at x by p, then the method is defined by p: The reason for including this step length is that the resulting displacement will always have norm less than one, so that Lemma 2 applies. It also guarantees that the damped truncated-Newton step is well-defined, in the sense that the iterates remain in int S. As the method converges and the truncated-Newton direction approaches zero, the step length approaches one, so that (asymptotically) rapid rates of convergence can be attained. The rest of this section develops the properties of the damped truncated-Newton method. The next lemma gives a lower bound on how much the function F will be decreased by a step of the damped truncated-Newton method. Lemma 5. If x+ is the result of the damped truncated-Newton iteration, then Proof. Let be the step length. Using Lemma 2 and (4) we obtain x The desired result is just a re-arrangement of this last inequality. The lemma provides a lower bound for F The lower bound is zero when kpk and is positive and strictly increasing for kpk x ? 0. (The derivative of the right-hand side with respect to kpk x is positive.) The result gives a lower bound on how much progress is made at each truncated-Newton iteration. Figure 1 illustrates this lower bound for various values of ffl. If kpk x remains large, then the truncated-Newton method must decrease the value of F (x) by a nontrivial amount. Since the function is bounded below on S, this cannot go on indefinitely, and kpk x must ultimately become small. The next theorem analyzes the convergence of the method in the case where are related by (3), these two results provide a bound on the number of truncated-Newton iterations required to solve the optimization problem to within some tolerance. (This argument is made precise in Theorem 7.) The theorem also determines a bound on kx \Gamma x k x in terms of ffi (F; x), and thus shows that if ffi (F; x) is small, then the norm of the error is small as well. Theorem 6. If x 2 int S then Let x be the minimizer of F in S, and assume that ffi 8 STEPHEN G. NASH AND ARIELA SOFER Figure Bound on F various values of ffl. bound bound bound bound Proof. The proof is in two parts, proving each of the results in turn. Part 1: We will derive the bound on ffi (F; x+ ). Let p be the truncated-Newton direction at x, pN be the Newton direction at x, and ffp. For any h 2 ! n we define This function is twice continuously differentiable for By Lemma 1 and Lemma 2 we have x+tp x Note that Z ffh Z x d- x Then x d- dt x x khk x COMPLEXITY OF A PRACTICAL INTERIOR-POINT METHOD 9 x khk x x x khk x x Since this is true for all h we have This completes Part 1 of the proof. Part 2: The proof of can be found in [14] or [15]. We conclude with a summary theorem. It provides a bound on the number of truncated-Newton iterations required to minimize F to within a tolerance. Theorem 7. Let S be a bounded, closed, convex subset of ! n with non-empty in- terior, and let F (x) be a convex function that is self concordant on S. Given an initial guess x0 2 int S, the damped truncated-Newton method is defined by the recurrence where p is an approximation to the Newton direction pN at for some tolerance 0 F At some iteration i we must have and then for every j - i we have G. NASH AND ARIELA SOFER The number of truncated-Newton steps required to find a point x bounded by for some constant C that depends on ffl and that decreases as ffl ! 0. If then the number of steps is bounded by for some constant - C. Proof. The bound on F derived in [16]. The remaining conclusions are consequences of the earler results. The formulas for the constants C and - C can be derived as follows. Let be the first iteration for which Such an index i must exist, since F is bounded below on S. If then by (3) It follows from Lemma 5 that Thus the number of initial iterations i is at most The progress of the later iterations is described by Theorem 6. If j - i, then Hence the number of later iterations is at most Summing these two bounds determines the constant C. then at the later iterations and a similar analysis determines - C. and the truncated-Newton method computes the Newton direction, then the theorem is the same as that reported for Newton's method in [14, 15]. It establishes the polynomiality of the truncated-Newton method applied to a barrier subproblem (see Section 1 for details). 4. Extrapolation. In Section 3 we analyzed the behavior of a truncated-Newton method when applied to a single barrier subproblem. We now consider the overall interior-point method based on solving a sequence of subproblems. Our main concern is with the effects of extrapolating the (approximate) solutions of several subproblems to obtain an improved initial guess for the next subproblem. COMPLEXITY OF A PRACTICAL INTERIOR-POINT METHOD 11 The complexity results for the overall method depend on an additional assump- tion, that is, a bound on the first-derivative of the barrier function. Although we do not make much direct use of this assumption in this paper, it underlies many of our comments, and so we state it here. Let S be a set with the same properties as in Section 3. Following [16], a self- concordant function F on S is a self-concordant barrier function for S if, for some for all x 2 int S and all h 2 ! n . We may assume, without loss of generality, that - 1. The function F barrier function with the function log(a T is a self-concordant barrier function with for the set \Psi . A self-concordant barrier function exists for any closed convex set S [16]; evaluating such a barrier function may not be computationally practical, however. We also assume that the convex program is written in the following standard subject to x 2 S where c 6= 0. An optimization problem with a general nonlinear objective function can be converted to this form by adding an additional variable and constraint. The problem (P) will be solved using a path-following method of the following form. For ae ? 0 we define where F is a self-concordant barrier function for the set S with parameter - 1, and where the Hessian of F is nonsingular for all x 2 int S. (The nonsingularity assumption is not essential; see [16].) Let x (ae) be the minimizer of F ae (x) for x 2 int S. Our method will generate x A complexity result for a particular algorithm based on this approach and Theorem 7 can be proved as in [14, 15]. In this algorithm, x i 2 int S is accepted as an approximate minimizer of F ae i if for some 2 , and then x i is used as an initial guess for minimizing F ae i+1 . Theorem 8. Suppose that we solve the problem (P) on a bounded, closed, convex domain S using the path-following method described above, where F is a self- concordant barrier function with parameter - 1. Let 1be the parameter in the acceptance test, and assume that the penalty parameters are updated via G. NASH AND ARIELA SOFER Assume that the method is initialized with ae 0 and x0 2 int S where x0 satisfies the acceptance test for F ae 0 . Then ae 0 If a truncated-Newton method based on (2) is used to minimize F ae i to within proximity -, then the number of truncated-Newton iterations required to find an approximate solution to subproblem i does not exceed a constant N -;';ffl depending only on -, ', and ffl. In particular, the total number of truncated-Newton iterations required to find an x satisfying c T bounded above by log with constant C -;';ffl depending only on -, ', and ffl. It is possible that a better initial guess for a subproblem (and hence a better algorithm) can be obtained by extrapolation of previous solutions. The technique of extrapolation-initially proposed by Fiacco and McCormick [5]-approximates the barrier trajectory x(ae) by a polynomial of degree q. The coefficients of the polynomial are computed from the solutions of q subproblems, and are then used to predict the solution of the barrier subproblem for the new value of ae. The results in [5] indicate that extrapolation is a powerful computational tool. Our own computational experiments [13] also indicate that better initial guesses (and better overall perfor- mance) can be obtained by extrapolating the solutions of a sequence of subproblems. An example illustrating the usefulness of extrapolation is summarized in Table 1; for details, see [13]. In this case, the use of cubic extrapolation reduces the number of truncated-Newton iterations by a factor of 2.1, and the number of gradient evaluations by a factor of 2.7. (Complexity results for a predictor-corrector method can be found in [15]; this predictor-corrector method is a form of extrapolation, but appears to be less practical for large nonlinear programs than the technique used here. It is not clear that this predictor-corrector approach can be extended effectively to a truncated-Newton method.) The lemma below gives theoretical support to these computational results. The technique of extrapolation may have limitations, however, as we shall discuss in the latter part of this section. We first examine linear extrapolation and assume that x linear extrapolation predicts that where More generally, \Delta i can be considered as a "search direction" along the barrier tra- jectory, as in the lemma below. The result shows that linear extrapolation can be used to produce initial guesses that are at least as good as when extrapolation is not used. For simplicity, we choose fl i j fl for all i, although it would be easy to extend the result to the case where the fl i 's are not constant. (The lemma uses our general COMPLEXITY OF A PRACTICAL INTERIOR-POINT METHOD 13 Table Effect of Cubic Extrapolation (iter = number of truncated-Newton iterations, ng = number of gradient evaluations). Based on [13] with elaborate barrier algorithm, test problem 51, 1000 variables. Extrapolation No Extrapolation ae iter ng iter ng 28 277 Totals 51 367 107 982 assumption that r 2 F is positive definite. If r 2 F is only positive semi-definite, then the lemma is still true under the assumption that c T \Delta i 6= 0.) Lemma 9. Let ae assume that x Define the linear extrapolation direction If then that is, \Delta i is a descent direction for F i+1 at x i . We define with a more specific upper bound provided by (1). Proof. We first prove that c T x series expansion gives Because x i 6= x positive definite, the last term is positive. Since 14 STEPHEN G. NASH AND ARIELA SOFER which implies that F Similarly, by switching the roles of x i and x i\Gamma1 , we obtain If we multiply the first of these inequalities by ae i , the second by ae i\Gamma1 , and rearrange, then Combining this with the above results We now use this result to prove that \Delta i is a descent direction. Since x Using these results, we obtain This completes the first part of the proof. The second part of the proof relies on (1). Using the formulas above, the upper bound in (1) becomes Note that by the assumptions in the lemma. If we solve OE 0 we obtain the solution It is straightforward to verify that OE 00 (ff so that this is a local minimizer of OE. we obtain that and so x Since the upper bound on F i+1 is decreasing at is a local minimizer of the upper bound, this completes the proof. COMPLEXITY OF A PRACTICAL INTERIOR-POINT METHOD 15 This result could be extended to the case where the subproblems are not solved exactly, as long as the magnitudes of were sufficiently small so as not to interfere with the inequalities in the lemma. It appears to be difficult to generalize the above result greatly, however, as the discussion below indicates. Lemma 9 shows that an appropriate step along the extrapolation direction can produce a decrease in the objective value of the barrier function for the updated barrier parameter. However taking the "full" extrapolation step (fl i =fl guaranteed to be beneficial, no matter how slight the change in the barrier parameter. This is true even when the subproblems are solved exactly. To see this, let and consider the extrapolated point The effect of extrapolation is described by Now d d 0: Thus, for a small change in the barrier parameter (i.e., as to first order there is no improvement in the objective value of the new barrier function at the extrapolated point. Somewhat more insight can be obtained by analyzing / 00 (0): The second term is positive by convexity, but the first term is less than or equal to zero (see Lemma 9). If we define then we can write From (5) and our comments above we obtain G. NASH AND ARIELA SOFER If kvk x i - then the lower bound on / 00 (0) is positive, and hence no improvement in the objective value is obtained. Thus the full extrapolation step is not guaranteed to be useful. So far, we have only considered linear extrapolation. We now indicate via an example that higher-order extrapolation may lead to predicted solutions that are far worse than those obtained without extrapolation. In this example, we consider both exact and approximate solutions to subproblems. As above, an approximate solution x i to a barrier subproblem will be accepted if x i 2 int S and if for some - 0. Consider the one-variable problem subject to x - 0: This problem is already in standard form, and log(x) can be used as a barrier function with It is easy to analyze this example. The solution to the optimization problem is The solution to the barrier subproblem is x 1=ae. The norm has the value jh=xj. The Newton decrement is xaej. The solution to a subproblem is accepted if The penalty parameter is updated via ae Here we use four values of 0:5. The choice corresponds to solving the subproblems exactly. We initialize the penalty parameter with ae In each case, the approximate solution to the subproblem is chosen at the upper bound of the acceptable range. This choice is consistent with the class of theoretical algorithms that we study in this paper, and is plausible for a practical algorithm. In Figure 2 we show the results of applying quadratic extrapolation on this prob- lem. The exact solutions of the first seven barrier subproblems are marked with \Theta. The vertical bars indicate the range of x values that satisfy the acceptance criterion for a subproblem. The dotted lines show the path of extrapolated approximate solutions, with a * used to indicate the extrapolated initial guess. In each of the four cases, the extrapolated values are exceedingly poor initial guesses for the next subproblem. In fact, the extrapolation paths move away from, rather than toward, the solution of the next subproblem. In Figure 3 we show the results of applying linear extrapolation. In this case, the approximate solutions are chosen as the lower bound of acceptable values for the first subproblem, and the upper bound for the second subproblem. This is allowed by the theory, but we think it unlikely that a practical algorithm could produce such approximate solutions. In this case, even with linear extrapolation, the extrapolated points can (when - is sufficiently large) point away from the solution of the next subproblem. Hence extrapolation is worse than doing nothing. As mentioned, we do not expect these circumstances to arise for a practical algorithm. Nevertheless, any complexity theory COMPLEXITY OF A PRACTICAL INTERIOR-POINT METHOD 17 Figure Quadratic Extrapolation of Approximate Solutions. rho x Quadratic: kappa=0 rho x Quadratic: kappa=.1 rho x Quadratic: kappa=.25 rho x Quadratic: kappa=.5 Figure Linear Extrapolation of Approximate Solutions. rho x rho x rho x rho x for such an algorithm would have to rule out this possibility, and hence would have to be based on a more elaborate theoretical framework than that used here. This example suggests that an algorithm that uses extrapolation must monitor the effectiveness of the extrapolation scheme, and must not use it blindly. It also suggests that the algorithm would have to be carefully designed so that inaccuracies in the solutions of barrier subproblems did not interfere with the performance of the extrapolation scheme. We have not been able to derive complexity results for a more elaborate algorithm of this type. This example leads us to think that this would be a difficult enterprise. G. NASH AND ARIELA SOFER --R Computational experience with penalty/barrier methods for nonlinear programming Interior Point Approach to Linear A sufficient condition for self-concord- ance Sequential Unconstrained Minimization Techniques Osman G- uler Two interior-point algorithms for a class of convex programming problems Interior point methods via self-concordance or relative Lipschitz condition A practical interior-point method for convex pro- gramming An unconstrained optimization technique for large-scale linearly constrained minimization problems Assessing a search direction within a truncated-Newton method A barrier method for large-scale constrained optimiza- tion Linear and Interior point polynomial time methods in convex programming Homogeneous interior-point algorithms for semidefinite program- ming An infinitely summable series implementation of interior-point methods On the long step path-following method for semidefinite programming Complexity Issues Interior methods for constrained optimization Extending primal-dual interior point algorithms from linear programming to semidefinite programming --TR
convex programming;interior-point method;truncated-Newton method;self-concordance;large-scale optimization;complexity
589351
An Unsymmetrized Multifrontal LU Factorization.
A well-known approach to computing the LU factorization of a general unsymmetric matrix A is to build the elimination tree associated with the pattern of the symmetric matrix A + AT and use it as a computational graph to drive the numerical factorization. This approach, although very efficient on a large range of unsymmetric matrices, does not capture the unsymmetric structure of the matrices. We introduce a new algorithm which detects and exploits the structural asymmetry of the submatrices involved during the processing of the elimination tree. We show that with the new algorithm, significant gains, both in memory and in time, to perform the factorization can be obtained.
Introduction We consider the direct solution of sparse linear equations based on a multifrontal approach. The systems are of the form A is an n n unsymmetric sparse matrix. The multifrontal method has been developed by Du and Reid [11, 12] for computing the solution of indenite sparse symmetric linear equations using Gaussian elimination and then has been extended to solve more general unsymmetric matrices by Du and Reid [13]. The multifrontal method belongs to the class of methods that separate the factorization into an analysis phase and a numerical factorization. The analysis phase involves a reordering step, which will reduce the ll-in during numerical factorization and a symbolic phase that This work was supported by the Director, Oce of Science, Division of Mathematical, Information, and Computational Sciences of the U.S. Department of Energy under contract number DE-AC03-76SF00098. y amestoy@enseeiht.fr, ENSEEIHT-IRIT, 2 rue Camichel 31071 Toulouse (France) and PRamestoy@lbl.gov, NERSC, Lawrence Berkeley National Laboratory, 1 Cyclotron Rd. Berkeley CA 94720 z NERSC, Lawrence Berkeley National Laboratory, 1 Cyclotron Rd., Berkeley CA 94720 builds the computational tree, so called elimination tree [9, 18, 20], whose structure gives the dependency graph of the multifrontal approach. The analysis phase is generally not concerned with numerical values and is only based on the sparsity pattern of the matrix. As far as the analysis phase is concerned, the approaches introduced by Du and Reid for both symmetric and unsymmetric matrices are almost identical. When the matrix is unsymmetric, the structurally symmetric matrix the summation is performed symbolically, is used in place of the original matrix A. The elimination tree of the unsymmetric LU factorization is thus identical to that of the Cholesky factorization of the symmetrized matrix M. To control the growth of the factors during LU factorization, partial threshold pivoting is used during the numerical factorization phase. The pivot order, used during the analysis to build the elimination tree might not be respected. Numerical pivoting can then result in an increase in the estimated size of the factors and in the number of operations. To improve the numerical behaviour of the multifrontal approach it is common to involve a step of preprocessing based on the numerical values. In fact if the matrix is not well-scaled, which means that the entries in the original matrix do not have the same order of magnitude, a good prescaling of the matrix can have a signicant impact on the accuracy and performance of the sparse solver. In some cases it is also very benecial to precede the ordering by performing an unsymmetric permutation to place large entries on the diagonal. Du and Koster [10] have designed algorithms to permute large entries onto the diagonal and have shown that it can very signicantly improve the behaviour of multifrontal solvers. The multifrontal approach by Du and Reid [13] is used in the Harwell Subroutine Library code 3] and in the distributed memory code MUMPS developed in the context of the PARASOL project (EU ESPRIT IV LTR project 20160) [4, 5]. Another way to represent the symbolic LU factorization of a structurally unsymmetric matrix is to use directed acyclic graphs (see for example [14, 15]). These structures more costly and complicated to handle than a tree, capture better the asymmetry of the matrix. Davis and Du [6] implicitly use this structure to drive their unsymmetric-pattern multifrontal approach. We explain, in this article, how to use the simple elimination tree structure of the symmetric matrix M to detect, during the numerical factorization phase, structural asymmetry in the factors. We show that, with the new factorization phase, we very signicantly reduce the computational time, the size of the LU factors and the total memory requirement with respect to the standard multifrontal approach [13]. In Section 2, we rst recall the main properties of the elimination tree and describe the standard multifrontal factorization algorithm. We then introduce the new algorithm and use a simple example to show the benets that can be expected from the new approach. In Section 3, our set of test matrices is introduced. We analyse the performance gains (in terms of size of the factors, memory requirement and factorization time) of the new approach with respect to the standard multifrontal code on our set of test matrices. We add some concluding remarks in Section 4 2 Description of the multifrontal factorization algorithms Let A be an unsymmetric matrix and let M denote the structurally symmetric matrix A+A T . The elimination tree is dened using the structure of the Cholesky factors of M. If the matrix M is reducible then the tree will be a forest. Liu [18] denes the elimination tree as the transitive reduction of the directed graph of the Cholesky factors of M. The characterization of the elimination tree and the description of its properties are beyond the scope of this article. In our context, we are interested in the elimination tree only as the computational graph for the multifrontal factorization. For a complete description of the elimination tree the reader can In the multifrontal approaches, we actually use an amalgamated elimination tree, referred to as the assembly tree [12] which can be obtained from the classical elimination tree. Each node of the assembly tree corresponds to Gaussian elimination operations on a full submatrix, called a frontal matrix. The frontal matrix can be partitioned as shown in Figure 1. fully summed rows - partly summed rows - fully summed columns partly summed columns Figure 1: Partitioning of a frontal matrix. Each frontal matrix factorization involves the computation of a block of columns of L, termed fully summed columns of the frontal matrix, a block of rows of U, termed fully summed rows, and the computation of a Schur complement matrix F 22 F 21 F 1 called a contribution block. The rows (columns) of the F 22 block are referred to as partly summed rows (columns). original nonzero new entry in M Figure 2: Example of matrix A and The unsymmetric matrix A, on the left-hand side of Figure 2, will be used to illustrate the main properties of the assembly tree and to introduce the new algorithm. In Figures 2 and 3 an \X" denotes a nonzero position from the original matrix A and a \ i " corresponds to a new entry introduced during symmetrization. In Figure 3, we indicate the structure of the lled matrix is the matrix of the Cholesky factor of M. Entries with an \F" corresponds to ll-in entries in the L factor. F F F original nonzero Fill-in (w.r.t. M) new entry in M Figure 3: Structure of the Cholesky factors of the matrix M. The matrix M F is used to dene the assembly tree (see Figure 4) associated with the multifrontal LU factorization of the matrix A. From the fact that the factorization is based on the assembly tree associated with the Cholesky factorization of M F , it results that where Struct() denotes the matrix pattern. Let us denote by structural zero a numerical zero that does not result from numerical cancellation. Typically, due to the symmetrization, the matrix M might contain many structural zeros that will propagate during the numerical factorization phase. What has motivated our work is the following question. Is it possible, during the processing of the assembly tree to eciently detect and remove structural zeros that appear in matrix M F and that are direct or indirect consequence of the symmetrization of Although it is not so clear from the structure of the matrix M F , we will show that blocks of structural zeros can be identied during the processing of the assembly tree. In the following, we rst describe how the assembly tree is exploited during the standard multifrontal algorithm. We then report and analyse the sparsity structure of the frontal matrices involved in the processing of the assembly tree associated with our example matrix. Based on these observations, we will introduce the new factorization algorithm. The assembly tree is rooted (a node of the tree called the root is chosen to give an orientation to the tree) and is processed from the leaf nodes to the root node. If two nodes are adjacent in the tree, then the one nearer the root is the parent node, and the other is termed its child. Each edge of the assembly tree indicates a data dependency between parent and child. It involves sending a contribution block from the child to the parent. A parent node process will start when the processes associated with all of its children are completed. U U U U original nonzero arrowhead of var. 3 contribution block Root (2) (1) Figure 4: Assembly tree associated with our test matrix. For example, in Figure 4, node (3) must wait for the completion of nodes (1) and (2) before starting its computations. The subset of variables which can be used as pivots (boldface variables in Figure 4) are the fully summed variables of node (k). The contribution blocks of the children and the entries from the original matrix corresponding to the fully summed variables of node (k) are used to build the frontal matrix of the node. This will be referred to as the assembly process. During the assembly process of a frontal matrix, we need for each fully summed variable j, to access the nonzero elements in the original matrix that are in rows/columns of indices greater than j. A way to eciently access the original matrix is to store it in arrowheads according to the reordered matrix. For example during the assembly process of node (3) the arrowheads of variables 3 and 4 from matrix A together with the contribution blocks of nodes (1) and (2) are used to assemble the frontal matrix of node (3). One should note that, by construction, the list of indices in the partly summed rows is identical to that of the partly summed columns (row and column indices of block F 22 in Figure 1). Therefore, during the assembly process, only the list of row indices of the partly summed rows is built. This list is obtained by merging all the row and column indices of the arrowheads of the matrix A with the row indices of the contribution blocks of all the sons. Once the structure of the frontal matrix is built, the numerical values from both the arrowheads and the contribution blocks can be assembled at the right place in the frontal matrix. The oating point operations involved during the assembly process will be referred to as assembly operations (only additions) whereas oating-point operations involved during the factorization of the frontal matrices will be referred to as elimination operations. Partial threshold pivoting is used to control the element growth in the factors. Note that pivots can be chosen only from within the block F 11 of the frontal matrix. The LU factors corresponding to the fully summed variables are computed and a new contribution block is produced. When a fully summed variable of node (k) cannot be eliminated during the node process because of numerical considerations, then the corresponding arrowhead in the frontal matrix is added to the contribution block and the fully summed variable will be included in the fully summed variables at the parent of node (k). This process creates additional ll-in in the LU factors. In a multifrontal algorithm, we have to provide space for the frontal matrices and the contribution blocks, and to reserve space for storing the factors. We need working space to store both real and integer information. This will be referred to as the total working space of the factorization phase. The same integer array can be used to describe a frontal matrix, its corresponding LU factors and its contribution block. The management of the integer working array can thus be done in a simple and ecient way. In a uniprocessor environment, it is possible to determine the order in which the assembly tree will be processed. Furthermore, if we process the assembly tree with a depth rst search order, we can use a stack to manage the storage of the factors and the contribution blocks. This mechanism is ecient both in terms of total memory requirement and amount of data movement (see [12]). A stack mechanism, starting from the beginning of the real working array, is used to store the LU factors. Another stack mechanism starting from the end of the real working array is used to store the contribution blocks. After the assembly phase of a node the working space used by the contribution blocks of its children can be freed and, because the assembly tree is processed with a depth rst search order, the contribution blocks will always be at the top of the stack. In the remainder of this paper, the maximum stack size of the contribution blocks will be referred to as the maximum stack size. The standard and new algorithms for multifrontal factorization During a multifrontal factorization, each frontal matrix can be viewed as the minimum structure to perform the elimination of the fully summed variables and to carry the contribution blocks from all of its sons. In Figure 5, we have a closer look at the frontal matrices involved in the processing of the assembly tree of Figure 4 to identify the structural zeros. We report, beside each node, the structure of the factorized frontal matrix assuming that the pivots are chosen down the diagonal of the fully summed block and in order (i.e. no numerical pivoting is required). An \X" corresponds to a nonzero entry and a \O" corresponds to a structural zero. One can see that, for our example, the frontal matrices have many structural zeros. There are two kinds of structural zeros: those forming a complete zero column (or row), and more isolated zero entries in a nonzero column or row (for example entries (4,3) and (4,7) in the e standard algorithm. frontal matrix of node (3)). If one knows how to detect a partly summed row (or column) with only structural zeros then the corresponding row (or column) can be suppressed from the frontal matrix because this row (or column) will not add any contribution to the father node. Structural zero rows (or columns) can be detected during the assembly process of a frontal matrix because of the following property: if a row (or column) index does not appear in the row (or column) indices both of the arrowheads of the original matrix and of the contribution blocks of the sons, then this index will correspond to a row (or column) with only structural zeros. This property is used to deduce the assembly process of the new algorithm. Note that if the matrix is not structurally decient then each fully summed row (or column) must have at least one nonzero entry. Therefore, we can restrict our search for zero rows (columns) to the partly summed rows (columns). In the new assembly algorithm, the list of indices of the partly summed rows of a frontal matrix is dened as the merge of the row indices in the arrowheads of the fully summed variables of the node with the row indices of the contribution blocks of its sons. The column indices are dened similarly. As it is illustrated in Figure 6, the new assembly process can result in signicant modications in the processing of the assembly tree. For example, on node (1), row 3 and column 5 are suppressed from the frontal matrix; on node (2), all the partly summed rows are suppressed; on node (3), row 7 and column 5 are suppressed. As it can be noticed in Figure 6, frontal matrices naturally become unsymmetric in structure. We nally indicate in Figure 7 the structure of the LU factors obtained with the new algorithm. Figure Processing the assembly tree associated with the matrix A in Figure 2 using the new algorithm. This should be compared to the matrix M F in Figure 3 showing the structure of the factors obtained with the standard algorithm. It can be seen that nonzero entries corresponding to original nonzero Fill-in (w.r.t. M) new entry in M Figure 7: Structure of the LU factors obtained with the new algorithm. ll-in (for example (7,4) in M F ) or introduced during the symmetrization of M (for example in M F ) might be suppressed by the new algorithm. On the other hand, the new algorithm will never suppress structural zeros in a block of fully summed variables (for example (4,3) in node (3) of Figure 5). On our small example, the total number of entries in the factors reduces from 31 to 23. Comparing Figures 5 and 6, one can notice that the new algorithm might also lead to a signicant reduction in both the number of operations involved during the assembly process and the maximum stack size. The latest combined with a reduction in the size of the factors will result in a reduction in the total working space. On our example the number of assembly operations drops from 30 to 20 (18 entries from A plus 2 from the contribution blocks). The maximum stack size reduces from 8 to 1 (obtained in both cases after stacking the contribution blocks of nodes (1) and (2)). 3 Results and performance analysis We describe in Table 1 the set of test matrices (order, number of nonzero entries, structural symmetry and origin). We dene the structural symmetry as the percentage of the number of nonzeros matched by nonzeros in symmetric locations over the total number of entries. A symmetric matrix has a value of 100. Although, our performance analysis will focus on matrices with a relatively small structural symmetry, all classes of unsymmetric matrices are represented in this set. The selected matrices come from the forthcoming Rutherford-Boeing Sparse Matrix Collection [8] 1 , Tim Davis collection 2 , and SPARSEKIT2 3 . The Harwell Subroutine Library [16] code ma41 has been used to obtain the results for the standard multifrontal method. The factorization phase of ma41 has then been modied with the new algorithm. The ma41 code has a set of parameters to control its eciency. We have used the default values for our target computer. Approximate minimum degree ordering [1] has been used to reorder the matrix. As we have mentioned in the Introduction, it is often quite benecial for very unsymmetric matrices to precede the ordering by performing an unsymmetric permutation to place large entries on the diagonal and then scaling the matrix so that the diagonal entries are all of modulus one and the o-diagonals have modulus less than or equal to one. We use the Harwell Subroutine Library code mc64 [10] to perform this preordering and scaling on all matrices of structural symmetry smaller than 55. When mc64 is not used, our matrices are always row and column scaled (each row/column is divided by its maximum value). All results presented in this section, have been obtained on one processor (R10000 MIMPS RISC 64-bit processor) of the SGI Cray Origin 2000 from Parallab (University of Bergen, Norway). The processor runs at a frequency of 195 Mhertz and has a peak performance of 400 M ops per second. Web page http://www.cse.clrc.ac.uk/Activity/SparseMatrices/ Web page http://www.cise.ufl.edu/ davis/sparse/ 3 Web page http://iftp.cs.umn.edu/pub/sparse/ Matrix name Order No. entries StrSym Origin (Discipline) av4408 4408 95752 0 Vavasis (Partial di. eqn.) [21] bbmat 38744 1771722 54 Rutherford-Boeing (CFD) cavity26 4562 138187 95 SPARSEKIT2 (CFD) ex11 16614 1096948 100 SPARSEKIT2 (CFD) goodwin 7320 324784 64 Davis (CFD) lhr14c 14270 307858 1 Davis (Chemical engineering) lhr17c 17576 381975 0 Davis (Chemical engineering) lhr34c 35152 764014 0 Davis (Chemical engineering) lhr71c 70304 1528092 0 Davis (Chemical engineering) lns 3937 3937 25407 87 Rutherford-Boeing (CFD) Rutherford-Boeing (Economics) Rutherford-Boeing (Demography) raefsky6 3402 137845 2 Davis (Structural engineering) Rutherford-Boeing (Chemical engineering) rim 22560 1014951 sherman5 3312 20793 78 Rutherford-Boeing (Oil reservoir simul.) shyy161 76480 329762 77 Davis (CFD) twotone 120750 1224224 28 Davis (Circuit simulation) wang4 26068 177196 100 Rutherford-Boeing (Semiconductor) Table 1: Test matrices. StrSym denotes the structural symmetry. In the following graphs, we report the performance ratios of the new factorization algorithm over the standard algorithm. Matrices are sorted by increasing structural symmetry of the matrix to be factored, i.e. after application of the column permutation when mc64 is used. We use the same matrix order in the graphs and in the complete set of results provided in Tables 2 and 3. In this way, one can easily nd, given a point in the graph, its corresponding entry in the tables. On the complete set of test matrices, we rst analyse in Figure 8 what is probably of main concern for the user of a sparse solver, i.e. the time to factor the matrix and the total working space (as dened in the previous section). In Figure 8, we divide the matrices into three categories: matrices of structural symmetry smaller than 50 for which the time reduction is between 20% and 80%, matrices whose structural symmetry is between 50 and 80 for which the time reduction is between 3% and 20% and nearly structurally symmetric matrices for which there is almost no dierence between the standard and new version. It is interesting to notice that even on symmetric matrices the added work to detect asymmetry does not aect much the performance of the factorization. In the remainder of this paper, we will not report results on almost structurally symmetric matrices (symmetry greater than 80). 43 43 48 new over standard version total working space Figure 8: Study of the factorization time and the total working space. sy (in abscissa) corresponds to structurally symmetric matrices. In Figure 9, we relate the gain in the factorization time with the reduction in the number of elimination operations and in the number of assembly operations. Although the number of operations due to the assembly is always much smaller that the number of operations involved during factorization (see Tables 2 and 3), the assembly process can still represent an important part of the time spent in the factorization phase (see for example [2]). This is illustrated in Figure 9 where we see that the high reduction in the number of assembly operations (more than 50%) signicantly contributes to reducing the factorization time. Note that on a relatively large matrix (twotone) of symmetry 57 still signicant gains in time and in number of assembly operations (more than 40%) can be obtained. In Figure 10, we relate the total working space reduction to the size of the factors and to the maximum stack size. Although a reduction in the maximum size of the stack might not always introduce a reduction in the total working space, we see that in practice it is often the case. An extreme example of this reduction is matrix orani678 of symmetry 9 (see Table 2) for which maximum stack size is reduced by more than one order of magnitude (5482454 to 457312). Finally, we notice that a large reduction in the maximum stack size (Figure 10) will generally correspond to a large reduction in the number of assembly operations (Figure 9). Structural Symmetry new over standard version oper for factors oper for assem. Figure 9: Impact of the reduction in the number of oating point operations on the time. 43 43 48 new over standard version maximum size of stack space for LU factors total working space Figure 10: Correlation between the factor size, the maximum stack size, and the total working space. Matrix Total Operations in Facto. (Str.Sym.) Version LU stack space Elimin. Assemb. Time raefsky6 Stnd 1509016 606458 2017106 4.795E+08 4.348E+06 2.17 (2) New 998064 145575 1119135 2.313E+08 1.049E+06 1.14 raefsky5 Stnd 1757680 378792 2095082 3.746E+08 3.874E+06 1.83 av4408 Stnd 551354 227787 677254 6.872E+07 1.451E+06 0.46 lhr14c Stnd 2167304 415090 2439502 2.092E+08 7.944E+06 1.77 lhr34c Stnd 5613656 755710 6249187 6.282E+08 2.064E+07 5.24 lhr17c Stnd 2813418 639172 3204801 3.089E+08 1.085E+07 2.65 lhr71c Stnd 11615170 729857 12711920 1.402E+09 4.304E+07 13.16 twotone Stnd 22086166 15899616 34489449 2.933E+10 2.171E+08 183.84 onetone1 Stnd 4713485 3348215 6212037 2.282E+09 2.675E+07 14.54 rdist1 Stnd 258096 53767 279999 8.150E+06 5.054E+05 0.13 Table 2: Comparison of the standard (Stnd) and the new algorithms on matrices of structural symmetry < 50. Matrix Total Operations in Facto. (Str.Sym.) Version LU stack space Elimin. Assemb. Time bbmat Stnd 44111480 8351266 48035816 3.676E+10 2.283E+08 185.75 utm3060 Stnd 324640 78679 806970 2.683E+07 6.973E+05 0.22 utm5940 Stnd 701496 131839 2799224 6.640E+07 1.529E+06 0.51 onetone2 Stnd 2253553 898540 385816 5.085E+08 7.628E+06 3.88 goodwin Stnd 1264140 307777 1385604 1.612E+08 2.841E+06 1.05 rim Stnd 4127204 833290 4371615 5.648E+08 9.194E+06 3.37 shyy161 Stnd 7437816 377535 290589 9.945E+08 1.178E+07 6.56 shyy41 Stnd 251336 28523 8137032 1.036E+07 6.337E+05 0.14 sherman5 Stnd 167412 61227 217769 1.284E+07 4.414E+05 0.13 lns 3937 Stnd 285517 89578 335285 1.920E+07 5.482E+05 0.20 cavity15 Stnd 202629 33004 230111 1.033E+07 3.453E+05 0.10 cavity26 Stnd 394164 58589 447293 2.433E+07 6.877E+05 0.22 ex11 Stnd 11981558 3960507 13614400 6.678E+09 3.836E+07 27.76 fidapm11 Stnd 15997220 4863371 19069150 9.599E+09 4.705E+07 39.78 olaf1 Stnd 5880174 1506068 6794919 1.965E+09 1.685E+07 8.91 wang4 Stnd 11561486 5063375 15900237 1.048E+10 4.087E+07 46.19 Table 3: Comparison of the standard (Stnd) and the new algorithms on matrices of structural symmetry >= 50 Concluding remarks We have described a modication of the standard multifrontal LUfactorization algorithm that can lead to a signicant reduction in both the computational time and the total working space. The standard multifrontal algorithm [13] for unsymmetric matrices is based on the assembly tree of a symmetrized matrix and involves frontal matrices symmetric in structure. Therefore, it produces LU factors such that the matrix F = L+U is symmetric in structure. This approach is currently used in the context of two publically available packages (ma41 [2, 3] and MUMPS [5, 4]) and has the advantage, with respect to other unsymmetric factorization algorithms [6, 7, 17], of having the LU factorization based on the processing of an assembly tree, while the other approaches explicitly or implicitly use a more complex to handle graph structure. We have demonstrated that, based on the same assembly tree, one can derive a new multifrontal algorithm that will introduce asymmetry in the frontal matrices and in the matrix of the factors F. The detection of the asymmetry is only based on structural information and is not costly to compute as it has been illustrated on structurally symmetric matrices, for which both algorithms behave similarly. On a set of unsymmetric matrices, we have shown that the new algorithm will reduce both the factor size and the number of operations by a signicant factor. We have also observed that the reduction in the number of indirect memory access operations involved during the assembly process is generally much higher than the reduction in the number of elimination operations. Finally, we have noticed that the gain in the maximum stack size is also relatively high and is comparable to the gain in the number of assembly operations. Space for Operations LU Stack Total Elim Assemb. Time mean 0.79 0.35 0.69 0.67 0.41 0.59 median 0.80 0.35 0.76 0.68 0.40 0.61 50 Structural symmetry < 80 mean 0.93 0.85 0.93 0.89 0.82 0.86 median 0.95 0.91 0.95 0.93 0.89 0.90 Table 4: Performance ratios of the new algorithm over the standard algorithm. To conclude, we report in Table 4 a summary of the results (mean and median) obtained on the test matrices with structural symmetry smaller than 80. For very unsymmetric matrices (structural symmetry smaller that 50), we obtain an average reduction of 31% in the total working space and of 41% in the factorization time. The maximum stack size and the number of assembly operations are reduced by respectively 65% and 59%. Finally, it is interesting to observe that, even on fairly symmetric matrices (50 structural symmetry < 80), it can still be worth trying to identify and exploit asymmetry during the processing of the assembly tree. Acknowledgements . We want to thank Horst Simon and Esmond Ng who gave us the opportunity to work at NERSC (LBNL) for one year. We are grateful to Sherry Li for helpful comments on an early version of this paper. --R Exploiting structural symmetry in unsymmetric sparse symbolic factorization. Elimination structures for unsymmetric sparse lu factors. A scalable sparse direct solver using static pivoting. The role of elimination trees in sparse factorization. The multifrontal method for sparse matrix solution: Theory and Practice. A new implementation of sparse Gaussian elimination. --TR --CTR Gianmarco Manzini , Mario Putti, Mesh locking effects in the finite volume solution of 2-D anisotropic diffusion equations, Journal of Computational Physics, v.220 n.2, p.751-771, January, 2007 Abdou Guermouche , Jean-Yves L'excellent, Constructing memory-minimizing schedules for multifrontal methods, ACM Transactions on Mathematical Software (TOMS), v.32 n.1, p.17-32, March 2006 Xiaoye S. Li , James W. Demmel, SuperLU_DIST: A scalable distributed-memory sparse direct solver for unsymmetric linear systems, ACM Transactions on Mathematical Software (TOMS), v.29 n.2, p.110-140, June Timothy A. Davis, A column pre-ordering strategy for the unsymmetric-pattern multifrontal method, ACM Transactions on Mathematical Software (TOMS), v.30 n.2, p.165-195, June 2004 Wook Ryol Hwang , Martien A. Hulsen , Han E. H. Meijer, Direct simulation of particle suspensions in sliding bi-periodic frames, Journal of Computational Physics, v.194 n.2, p.742-772, March 2004 See Jo Kim , Wook Ryol Hwang, Direct numerical simulations of droplet emulsions in sliding bi-periodic frames using the level-set method, Journal of Computational Physics, v.225 n.1, p.615-634, July, 2007
elimination tree;gaussian elimination;unsymmetric matrices;multifrontal methods;sparse linear equations
589352
Improved Symbolic and Numerical Factorization Algorithms for Unsymmetric Sparse Matrices.
We present algorithms for the symbolic and numerical factorization phases in the direct solution of sparse unsymmetric systems of linear equations. We have modified a classical symbolic factorization algorithm for unsymmetric matrices to inexpensively compute minimal elimination structures. We give an efficient algorithm to compute a near-minimal data-dependency graph for unsymmetric multifrontal factorization that is valid irrespective of the amount of dynamic pivoting performed during factorization. Finally, we describe an unsymmetric-pattern multifrontal algorithm for Gaussian elimination with partial pivoting that uses the task- and data-dependency graphs computed during the symbolic phase. These algorithms have been implemented in WSMP---an industrial strength sparse solver package---and have enabled WSMP to significantly outperform other similar solvers. We present experimental results to demonstrate the merits of the new algorithms.
Introduction . Typical dire ct solve rs for ge ne ral sparse syste ms of line ar e quations of the form have four distinct phase s: analysis, comprising orde ring for fill-in re duction and symbolic factorization; nume rical factorization of the sparse coe #cie nt matrix A into triangular factors L and U using Gaussiane limina- tion with partial pivoting; forward and backwarde limination to solve for x using the triangular factors L and U and the right-hand-side ve ctor b; and ite rative re fine me nt of the compute d solution. In this pape r, we de scribe some of the algorithms that are use d in the unsymme tric symbolic and nume rical factorization phase s of the Watson Sparse Matrix Package (WSMP)-a high-pe rformance and robust software for solving ge ne ral sparse line ar syste ms. The se algorithms are crucial to WSMP's pe rformance , which has be e n shown to be significantly be tte r than that of othe r similar solve rs [18]. An important contribution of this pape r is to show that, contrary to conve ntional wisdom, it is possible to symbolically de te rmine a static communication patte rn for paralle l unsymme tric sparse LU ve n in the pre se nce of partial pivoting. The proce ss of factoring a sparse matrix can be e xpre sse d by a dire cte d acyclic task-de pe nde ncy graph (task-DAG). The ve rtice s of this DAG corre spond to the tasks of factoring rows or columns, or groups of rows and columns, of the sparse matrix, and the e dge s corre spond to the de pe nde ncie s be twe e n the tasks. A task is re ady fore xe cution if and only if all tasks with incominge dge s to it have comple te d. In addition to a task-DAG, the re is a data-de pe nde ncy graph (data-DAG) associate d with sparse matrix factorization. The ve rte x se t of the data-DAG is the same as that of the task-DAG for a give n sparse matrix. Ane dge from a ve rte x i to a ve rte x j in the data-DAG de note s that at le ast some of the output data of task i is re quire d as input by task j. In this pape r, we de fine task i as the task of computing column i of L and row i of U . Once the tasks are de fine d, the task-DAG is unique to a sparse # Received by the editors October 3, 2001; accepted for publication (in revised form) by E. G. Ng October 18, 2002; published electronically December 3, 2002. http://www.siam.org/journals/simax/24-2/39604.html T.J.W atson Research Center, P.O. Box 218, Yorktown Heights, NY 10598 (anshul@watson. ibm.com). matrix for a give n pe rmutation of rows and columns; howe ve r, the data-DAG is a function of the sparse factorization algorithm. Multifrontal algorithms [9, 14, 23] for sparse factorization can work with a minimal data-DAG (i.e ., a data-DAG with the smalle st possible dge s) for a give n matrix. In the case of symme tric sparse matrice s, the minimal task- and data-DAGs for the factorization proce ss are a tre e calle d the e limination tre e [22]. Howe ve r, for un- symme tric sparse matrice s, the task- and data-DAGs are ge ne ral DAGs. More ove r, the e dge -se t of the minimal data-DAG for unsymme tric sparse factorization can be a rse t of the e dge -se t of a task-DAG. Gilbe rt and Liu [16] de scribe e limination structure s for unsymme tric sparse LU factors and give an algorithm for sparse unsym- me tric symbolic factorization. The se e limination structure s are two DAGs that are transitive re ductions of the graphs of the factor matrice s L and U , re spe ctive ly, and can be use d to de rive a task-DAG for sparse LU factorization. Some re se arche rs have argue d that computing ane xact transitive re duction can be tooe xpe nsive [9, 15] and have propose d using subminimal DAGs with more e dge s than ne ce ssary. Howe ve r, trave rsing unne ce ssary dge s during nume rical factorization can be a source of ove rhe ad. More ove r, in a paralle l imple me ntation,e xtra dge s can be pote ntial source s of unne ce ssary synchronization or communication. In this pape r, we show how a re lative ly straightforward modification to Gilbe rt and Liu's symbolic factorization algorithme nable s ane #cie nt computation of the minimale limination DAGs. We also de fine a se t dge s that must be adde d to the task-DAG in orde r to ge ne rate a minimal data-DAG that is valid as long as partial pivoting with dynamic row and columne xchange s is not pe rforme d during factorization. Finally, we de scribe how supple me nting this data-DAG furthe r with a small se t ofe xtrae dge s can yie ld a ne ar-minimal data-DAG that is su#cie nt to handle an arbitrary numbe r of pivot failure s and the re sulting row and columne xchange s during nume rical factorization. A pivot failure occurs the pivot pre dicte d by the analysis phase must be alte re d during nume rical factorization be cause the nume rical value of the pivot is too small. By me ans ofe xpe rime nts on a suite of unsymme tric sparse matrice s from re al applications, we show that computing the final data-DAG xtre me ly fast. Furthe rmore , for the matrice s in our te st suite , this data-DAG has only a slightly highe r ofe dge s than the task-DAG constructe d using comple te transitive re duction. The multifrontal me thod [9, 14, 23] for sparse matrix factorization usually o#e rs a significant pe rformance advantage ove r conve ntional factorization sche me s by pe r- mittinge #cie nt utilization of paralle lism and me mory hie rarchy. Du# and Re id [14] de scribe d a symme tric-patte rn multifrontal algorithm for unsymme tric matrice s that ge ne rate s ane limination tre e base d on the symme tric structure of the union of the structure s of A and the transpose of A to guide the nume rical factorization. This algorithm works on square frontal matrice s (se e se ction 4.1) and can incur a substantia ove rhe ad for ve ry unsymme tric matrice s due to unne ce ssary data de pe nde ncie s in the e limination tre e and due toe xtra ze ros in the artificially symme trize d frontal matrice s. Davis and Du# [9] and Hadfie ld [20] introduce d an unsymme tric-patte rn multifrontal algorithm that ove rcome s the de ficie ncie s of a symme tric-patte rn algo- rithm. Our powe rful symbolic phase e nable s us to use a much more simplifie d and e #cie nt ve rsion of the unsymme tric-patte rn multifrontal algorithm with partial pivoting We de scribe the unsymme tric-patte rn multifrontal algorithm that is use d in WSMP ande xpe rime ntally compare it with othe r state -of-the -art sparse unsymme tric factorization code s. UNSYMMETRIC SPARSE MATRIX FACTORIZATION 531 Table Test ith their order (N), number of nonzeros (NNZ), and the application area of origin. Number Matrix N NNZ Application Finite element analysis 3 bayer01 57735 277774 Chemistry 4 bbmat 38744 1771722 Fluid dynamics programming 6 e40r0000 17281 553956 Fluid dynamics 7 e40r5000 17281 553956 Fluid dynamics simulation 9 epb3 84617 463625 Thermodynamics engineering 14 mixtank 29957 1995041 Fluid dynamics engineering simulation simulation pre2 659033 5959282 Circuit simulation 19 raefsky3 21200 1488768 Fluid dynamics 22 tib 18510 145149 Circuit simulation twotone 120750 1224224 Circuit simulation wang3old 26064 177168 Circuit simulation simulation In Table 1.1, we introduce the suite of randomly chose n te st matrice s that we will use ine xpe rime nts throughout this pape r. The table shows the orde r ofe ach matrix, the numbe r of nonze ros in it, and the application are a of the origin of the matrix. All matrice s in our te st suite arise in re al-life proble ms and are in the public domain. The e xpe rime nts re porte d in this pape r we re conducte d on an IBM RS6000 WH-2 with a Gbyte s of RAM, 8 Mbyte s of le ve l-2 cache , and 64 Kbyte s of le ve l-1 cache . The organization of this pape r is as follows. ction 2 introduce s the te rms, conve ntions, and notations use d in the pape r. A symbolic factorization algorithm that compute s the structure of the triangular factors and minimale limination structure s is de scribe d in se ction 3. In se ction 4, we de scribe how to compute ne ar-minimal data-DAGs for unsymme tric multifrontal factorization. The nume rical factorization algorithm is discusse d in de tail in se ction 5. We finish with concluding re marks in se ction 6. The last subse ction ofe ach major se ction containse xpe rime ntal re sults pe rtaining to the algorithms in that se ction. 2. Terminology and conventions. We assume that the original n - n sparse unsymme tric coe #cie nt matrix is irre ducible and cannot be pe rmute d into a block-triangular form. This is not a se rious re striction, be cause a ge ne ral matrix can first be re duce d to a block-triangular form and the n only the irre ducible diagonal blocks ne e d to be factore d [12]. We assume that the coe #cie nt matrix A is factore d into a lowe r triangular matrix L and an uppe r triangular matrix U . Multiple row and column pe rmutations may be applie d to A during various stage s of the solution proce ss. Howe ve r, for the sake of clarity, we will always de note the coe #cie nt matrix by A and the factors by L and U . The state of pe rmutation of A, L, and U will usually be cle ar from the conte xt. We de note the dire cte d graph corre sponding to an n - n matrix M by re graph may not always be associate d with ane xplicitly de fine d matrix. Howe ve r, the n ane dge i#j # EM if and only if m ij is a structural nonze roe ntry in the sparse matrix M . The transpose of a matrix M is re pre se nte d by M # . If i#j # EM , the n j#i # EM # , and vice -ve rsa. the se t of indice s of the columns in M that have a structural nonze roe ntry in row i. This is also the se t of all ve rtice s to which i has an outbound e dge inG M . Similarly, Struct(M #,i ) is the se t of indice s of the rows in M that have a structural nonze roe ntry in column i and is also the se t of all ve rtice s from which i has an inbounde dge inG M . A dire cte d path from node i to node j in the dire cte d graphG M is de note d by i#j. The transitive re ductionG M O(VM , EM O ) of a graph the graph with the smalle st dge s that has a dire cte d path i#j if and only ifG M has a dire cte d path i#j. Since we are primarily de aling with the nonze ro structure of matrice s rathe r than the actual value s, we may also loose ly re fe r to M O as the transitive re duction of M O is a transitive re duction ofG M . The le ading i - i submatrix of M is de note d by M i and the corre sponding graph and its transitive re duction andG M O re spe ctive ly. The e dge s and paths in some of the graphs use d in this pape r are labe le d. An e dge in a labe le d graph can have one of the thre e labe ls-L, U, or LU. De pe nding on its labe l, ane dge can be an L-e dge , a U-e dge , or an LU-e dge . L-, U-, and LU-e dge s from ve rte x i to j are de note d by i L #j, iU #j, and i LU #j, re spe ctive ly. An L-path from i to j, de note d by i L #j, is a dire cte d path containing only L- and LU-e dge s. Similarly, a U-path from i to j, de note d by iU #j, is a dire cte d path containing only U- and LU-e dge s. If an L-e dge i L #je xists in the graph, the L-pare nt(i). Similarly, if iU #je xists, the U-pare nt(i), and if i LU #je xists, the LU-pare nt(i). We de fine 1 a supe rnode [q : r] as a maximal se t of conse cutive indice s {q, q 1, . , r} such that for all and matrice s L and U , we de fine supe rnodal matrice s L and U such thate ach supe rnode [q : r] in L and U is re pre se nte d by a single row and column in L and U . re m # n is the total numbe r of supe rnode s. Furthe rmore , if and r < s, the n g < h; that is, the column and row indice s in L and U maintain the re lative orde r of supe rnode s in L and U . 3. Computing a task-DAG and the structures of L and U . Gilbe rt and pre se nt an unsymme tric symbolic factorization algorithm to compute the structure s of the factors L and U and the ir transitive re ductions L O and U O . Figuresummarize s Gilbe rt and Liu's algorithm. The algorithm compute s the structure of L, U , and L O row by row and compute s the structure of U O by columns. The total time that the algorithm shown in Figure spe nds in ste p 1 is bounde d by flops(LU O ) [16], which is the numbe r of ope rations re quire d to multiply the sparse matrice s L and U O . Similarly, the time spe nt in ste p 3 is bounde d by flops(UL O ). The total computational cost of ste ps 2 and 4 is O| )). This is be cause transitive re duction is pe rforme d on n rows of U and columns of L, and the ith ste p could pote ntially trave rse alle dge s inG L O andG U O Ste ps 2 and 4 of Gilbe rt and Liu's algorithm are much more costly than ste ps 1 and 3. The cost of the se ste ps Other definitions of supernodes in the context of unsymmetric sparse factorization have been used in the literature [11]. UNSYMMETRIC SPARSE MATRIX FACTORIZATION 533 to n do 1. Compute trave rsingG U O and using the fact that # j < i, j # Struct(L i,# ) if and only if # k # j such that k # the re is a path k#j in U O i-1 . 2. Transitive ly re duce Struct(L i,# ) usingG L O ande xte nd it toG L O 3. Compute 4. Transitive ly re duce Struct(U #,i ) usingG U #O ande xte nd it toG U #O end for Fig . 3.1. Gilbert and Liu's unsymmetric symbolic factorization algorithm [16]. to n do 1. Transitive ly re duce Struct(L i,# ) usingG L O ande xte nd it toG L O 2. Compute 3. Transitive ly re duce Struct(U #,i ) usingG U #O ande xte nd it toG U #O 4. Compute end for Fig . 3.2. A modified symbolic factorization algorithm. has prompte d re se arche rs to se e k alte rnative s, such as computing fast but incomple te transitive re duction [9, 15]. The use of such alte rnative s toG L O andG U O with more e dge s thanG L O andG U O , re spe ctive ly, can incre ase the cost of ste ps 1 and 3, as we ll as that of nume rical factorization. 3.1. A modification to Gilbert and Liu's algorithm. We now de scribe a re lative ly simple modification to the algorithm shown in Figure 3.1. We start by splitting the original coe #cie nt matrix into a lowe r triangular part store d by columns and an uppe r triangular part store d by rows. In our modifie d symbolic factorization algorithm, we compute the structure of L by the columns (i.e ., L # by rows) and that of U by the rows. This is achie ve d by simply re formulating the algorithm shown in Figure 3.1 to pe rform only ste ps 2 and 3, but twice fore ach i on two se ts of ide ntical data structure s-one corre sponding to L # and the othe r corre sponding to U . The modifie d algorithm is shown in Figure 3.2. Note that in the algorithm of Figure 3.2, ste ps 3 and 4 are ide ntical to ste ps 1 and 2, re spe ctive ly. The first two ste ps compute the ith rows of L O and U and the last two ste ps compute the ith columns of U O and L. An actual code of this algorithm can use the same pair of routine s with di#e re nt argume nts to imple me nt all four ste ps. The re duction in the size of the code by half, howe ve r, is a se condary be ne fit of the modifie d algorithm. The primary advantage of this sche me is that it allows imme diate de te ction of supe rnode s during symbolic factorization. This, as we shalle xplain in se ction 3.2, allows us to avoid computing and storingG L O andG U #O e xplicitly. Inste ad, we can work only with the ir supe rnodal counte rpartsG L O and G U #O . 3.2. Use of supernodes to speed up transitive reduction. Most mode rn sparse factorization code s re ly he avily on toe #cie ntly utilize me mory hie rarchie s and paralle lism in the hardware . are so crucial to high pe r- formance in sparse matrix factorization that the crite rion for the inclusion of rows and columns in the same supe rnode is ofte n re laxe d [7] to incre ase the size of the supe rnode s. Conse cutive rows and columns with ne arly the same but not ide ntical structure s are ofte n include d in the same supe rnode , and artificial nonze roe ntrie s with a nume rical value of 0 are adde d to maintain ide ntical row and column structure s for all me mbe rs of a supe rnode . The rationale is that the slight incre ase in the numbe r of nonze ros and floating-point ope rations involve d in the factorization is more than compe nsate d for by a highe r factorization spe e d. WSMP's LU factorization algorithm also works on the re laxe d ge ne rate d by its symbolic factorization. In the symbolic factorization algorithm, as soon as are compute d in the ith ite ration of the oute r loop, the y can be compare d with Struct(L #,i-1 ) and Struct(U i-1,# ) to de te rmine if the y be long to the curre nt supe rnode . A ne w row-column pair is adde d to the curre nt supe rnode if its structure ide ntical or ne arly ide ntical to the pre vious row-column pair. If the ith row-column pair fails to me e t the crite rion for me mbe rship into the curre nt supe rnode , the n a ne w supe rnode is starte d at i. The use of supe rnode s allows us to significantly re duce the cost of computing the transitive re ductions. In ste p 1 of the algorithm shown in Figure 3.2, inste ad of transitive ly re ducing the e ntire Struct(L i,# ), we re duce only the se t {h re Ste p 3 is tre ate d similarly. As a re sult of working only with supe rnode s, the uppe r bound on the cost of computing the transitive re duction de cre ase s from O| )) to O| )). This is be cause only the supe rnodal DAGsG L O andG U O are se arche d duringe ach of the n transitive re duction ste ps. Strict supe rnodal andG U O would have at le ast fe we re dge s thanG L O andG U O , re m is the numbe r of supe rnode s. The re ason is that U O and L #O do not contain anye dge s i#j, re supe rnode . The use of re laxe d re duce s the numbe r ofe dge s e ve n furthe r be cause some pote ntiale dge s of the form i#j, re be e liminate d from the task-DAG node s i and are artificially me rge d. 3.3. Task-DAGs for LU factorization. In this pape r, we will re fe r to two type s of task-DAGs: a conve ntional DAG de note d by and a supe rnodal DAG de note d by T S . Each ve rte x of the conve ntional task-DAG re fe rs to the task of computing a single row of U and the corre sponding column of L. On the othe r hand, a ve rte x of the supe rnodal task-DAG corre sponds to a se t of row-column pairs that constitute a supe rnode . Although, in a practical imple me ntation, we always work with supe rnodal DAGs, we will ofte n use conve ntional task- and data-DAGs in the re mainde r of the pape r to the e xposition simple . All re sults and de scriptions pre se nte d in te rms of the conve ntional DAGs map naturally to the supe rnodal case . We first show how to compute TC in te rms of the conve ntional structure s L #O and U O . The transpose matrix L # is use d to indicate that for all i#j Theorem 3.1. is a task-DAG for LU factorization if its ve rte x se {1, 2, . , n} and itse dge -se Proof. To prove that is a task-DAG, we show that E T C is su#cie nt to re p- re se nt a prope r orde ring of the ne limination tasks de note d by V T C . Struct(L #,i ) can contribute to Struct(L #,j ) only if i # Struct(U #,j ), and if this is the case , the n the symbolic factorization algorithm of Figure 3.2e nsure s that U O containse ithe r i#j UNSYMMETRIC SPARSE MATRIX FACTORIZATION 535 Table Comparison of conventional symbolic factorization (due to Gilbert and Liu [16])w ith supernodal | is the size of the largest diagonal block in the matrix onw hich symbolic factorization is performed; Nsup is the number of supernodes; t C and t S are the times in seconds of the tw are the number of edges in the task-DAGs produced by the tw algorithms. Matrix |V | Nsup Conventional Supernodal |t Ct S bbmat 38744 4877 10. 41260 1.7 6077 7.9 6.8 5.9 e40r5000 17281 2755 .60 19891 .16 3182 6.3 6.3 3.8 fidap011 16614 1262 2.3 16613 .42 1261 13. 13. 5.5 mil053 530238 166155 15. 530237 4.5 166154 3.2 3.2 3.3 mixtank 29957 2984 7.8 30949 1.2 3203 10. 9.7 6.5 nasasrb 54870 3808 4.9 54869 .97 3807 14. 14. 5.1 pre2 629628 243693 30. 765210 6.4 317216 2.6 2.4 4.7 raefsky3 21200 1282 2.1 21199 .41 1281 17. 17. 5.1 raefsky4 19779 1359 2.9 19778 .50 1358 15. 15. 5.8 tib 17583 7823 .11 22904 .07 10060 2.2 2.3 1.6 twotone 105740 34304 2.6 126656 .91 44856 3.1 2.8 2.9 wang3old 26064 8451 3.1 26063 .54 8450 3.1 3.1 5.7 wang4 26068 8254 3.0 26067 .53 8253 3.2 3.2 5.7 or i#j. The same is true for Struct(U i,# ), Struct(U j,# ), and L #O . The re fore ,e ve ry row-column pair i that update s row and column j must be e liminate d be fore j. The ore m 3.1 can be e asilye xte nde d to the supe rnodal case . The supe rnodal task-DAG T S is de fine d by a ve rte x se ane dge se re m is the numbe r of supe rnode s. 3.4. Experimental results. In Table 3.1, we compare Gilbe rt and Liu's symbolic factorization algorithm [16] with the supe rnodal symbolic factorization algorithm de scribe d in se ction 3.2. We re port the ir CPU time s tC and t S , re spe ctive ly, and the numbe re dge s in task DAGs and T S ge ne rate d by the m. The last column of Table 3.1 shows the factor by which the supe rnodal symbolic factorization is faste r than the conve ntional algorithm. The table also shows ave rage supe rnode size (n/m) and the ratio dge s in and T S fore ach matrix. The se two ratios are close ly re late d. The ratio of tC and t S be ars some corre lation to the ratio dge s in and T S , but the actual ratio is matrix de pe nde nt. Note that only the time of transitive re duction ste ps 1 and 3 of the algorithm in Figure 3.2 is re duce d by the use of supe rnode s; the time of computing the structure s of L and U in ste ps 2 andre mains mostly unchange d (othe r than some re duction in the numbe r of structure s me rge d due to supe rnode re laxation). The re fore , the actual re duction achie ve d in the symbolic factorization time de pe nds on the re lative amounts of time spe nt in 536 ANSHUL transitive re duction and computing L and U structure s. More ove r, Table re ports the numbe r dge s in the task-DAGs, not the numbe r dge s in the actual lowe r and uppe r triangular transitive ly re duce d graphs that are trave rse d during symbolic factorization. Re call that the e dge -se t of a task-DAG is the union of the e dge -se ts of the corre sponding lowe r and uppe r triangular transitive ly re duce d graphs. The amount of structural symme try in the matrix a#e cts the numbe r of commone dge s be twe e n the uppe r and lowe r transitive ly re duce d graphs, which in turn de te rmine s the actual dge s in the task-DAG. Eise nstat and Liu [15] pre se nt an alte rnative to comple te transitive re duction to re duce the cost of this ste p in sparse unsymme tric symbolic factorization. The y propose e xploiting structural symme try in the matrix to compute partial transitive re ductions. Although the y pre se nte xpe rime ntal re sults on a di#e re nt se t of much smalle r matrice s, it appe ars that the use of supe rnode s as propose d in se ction 3.2 can achie ve much highe r spe e dups in symbolic factorization while computinge xact transitive re ductions than the partial transitive re duction sche me propose d in [15]. Howe ve r, Eise nstat and Liu's algorithm too can be spe d up by the use of supe rnode s. A supe rnodal ve rsion of this algorithm has be e n imple me nte d in the Supe rLU dist [21] sparse solve r package . We compare d our symbolic factorization time with that of Supe rLU dist and found the latte r to be slowe r by about 25% ove rall on our te st suite . This could be partly due to imple me ntation di#e re nce s and partly due to the fact that while Eise nstat and Liu's algorithm save s time in the transitive re duction computation, it spe ndse xtra time in me rging structure s due to re dundante dge s in the DAG. It appe ars that the use of supe rnode s in Gilbe rt and Liu's algorithm can spe e d up its transitive re ductione nough for it to match or outpe rforme ve n a supe rnodal ve rsion of Eise nstat and Liu's algorithm ine xe cution time . 4. Data-DAGs for unsymmetric multifrontal LU factorization. The original multifrontal algorithm [14, 23] was de scribe d in the conte xt of a symme tric- patte rn coe #cie nt matrix but has be e n applie d to matrice s with unsymme tric patte rns by introducing ze ro-value de ntrie s at appropriate locations to conve rt the original matrix into one with the patte rn of A+A # [14, 2, 4]. This can cause a substantial ove rhe ad for ve ry unsymme tric matrice s due to the e xtra computation pe rforme d on the introduce de ntrie s and the re sulting fill-in. Davis and Du# [9] and Hadfie ld [20] introduce d an unsymme tric-patte rn multifrontal algorithm to ove rcome this short- coming. In this se ction, we de ve lop ne ar-minimal data-DAGs for the unsymme tric multifrontal algorithm-an aspe ct of unsymme tric multifrontal factorization that has not be e n we ll inve stigate d in pre vious works. As we shall show in se ction 5, the availability of a ne ar-minimal data-DAG aids in the e #cie nt imple me ntation of the nume rical factorization phase . It would also he lp minimize the synchronization and communication ove rhe ads in a paralle l imple me ntation. 4.1. Outline of the symmetric multifrontal algorithm. The symme tric- patte rn multifrontal algorithm is guide d by an asse mbly ore limination tre e [22, 23, 19], which se rve s as both the task- and data-de pe nde ncy graphs for the factorization proce ss. The data associate d withe ach supe rnode of the e limination tre e is a square frontal matrix. A frontal matrix F g associate d with a supe rnode de nse matrix whose dime nsions are e qual )| or| Struct(U q,# )| . The contiguous local row and column indice s in the de nse frontal matrix corre spond to noncontiguous global indice s of the matrix L Eache ntry in a frontal matrix corre sponds to a structural nonze roe ntry in the global matrix. Afte r a frontal matrix F g is fully asse mble d or populate d, the le ading r - q columns corre sponding to UNSYMMETRIC SPARSE MATRIX FACTORIZATION 537 the supe rnode (also known as the pivot block) are factore d and be come parts of the factors U and L, re spe ctive ly. The re maining trailing part of the frontal matrix is now calle d the update or the contribution matrix, de note d by C g . The contribution matrix corre sponding to a supe rnode is asse mble d comple te ly into the frontal matrix of its only pare nt supe rnode and is ne ve r acce sse d again. This is be cause if the pare nt of supe rnode the e limination tre e , the n Struct(L #,r The same is true for columns of U due to symme try. In a re cursive formulation of the symme tric-patte rn multifrontal algorithm, the task corre sponding to a supe rnode first comple te s ide ntical subtasks fore ach of its childre n in the e limination tre e , the n asse mble s the ir contribution matrice s into its frontal matrix, and finally pe rforms the partial factorization on the frontal matrix. Calling a re cursive proce dure to pe rform the task de scribe d above on the root su- pe rnode of the e limination tre e comple te s the factorization of a sparse matrix with a symme tric structure . 4.2. Outline of the unsymmetric multifrontal algorithm. The ove rall structure of an unsymme tric-patte rn multifrontal algorithm is similar to its symme t- ric counte rpart and can be e xpre sse d in the form of a re cursive proce dure starting at the root (the supe rnode with no outgoinge dge s) of the task-DAG. Howe ve r, the re are two major di#e re nce s. The first di#e re nce is in the control-flow. In the unsymme tric multifrontal algorithm, be fore starting a subtask for a child, the task corre sponding to the pare nt supe rnode must che ck to se e if the child supe rnode has alre ady be e n proce sse d by anothe r pare nt. Only the first pare nt to re ach a child actually pe rforms the re cursive computation starting at that child. The se cond di#e re nce is in the data- flow, or the way contribution matrice s are asse mble d into frontal matrice s. This is e xplaine d be low in gre ate r de tail. Re call that the e dge -se t E T C of the task-DAG is the union of the e dge -se ts O of the transitive re ductions of L # and U , re spe ctive ly. We now assign labe ls to the e dge s in . The e dge s contribute d to E T C sole ly by E L #O are labe le d as L-e dge s. dge s contribute d to E T C sole ly by E U O are labe le d as U-e dge s. The third type of labe l, the LU-labe l, is assigne d to the e dge s that be long to the inte rse ction E L #O and E U O . Finally, an L-e dge i L #j is conve rte d to an LU-e dge i LU #j if the re is a U-path iU #j in , and a U-e dge iU #j is conve rte d to i LU #j if the re is an L-path i L #j in . The e dge s of the supe rnodal task-DAG T S are de fine d similarly. Unlike the symme tric multifrontal algorithm, the frontal and contribution matrice s in the unsymme tric multifrontal algorithm are , in ge ne ral, re ctangular rathe r than square . Furthe rmore , a contribution matrix in the unsymme tric multifrontal algorithm can pote ntially be asse mble d into more than one frontal matrix be cause a supe rnode in the data-DAG can have more than one pare nt. As de scribe d in [20], the asse mbly of contribution matrice s into the pare nt frontal matrice s in the unsymme tric multifrontal algorithm proce e ds as follows. #h be an L-e dge in the data-DAG, re have an inde x i in common, the n alle le me nts of row i of U in C g can pote ntially be asse mble d into F h . Similarly, if gU #h is a U-e dge and have an inde x i in common, the n alle le me nts of column i of L in C g can pote ntially be asse mble d into F h . Finally, if g LU #h is an LU-e dge , the n the e ntire trailing submatrix of C g with global row and column indice s gre ate r than ore qual to s can be asse mble d into F h . Ce rtaine ntrie s of C g may have pote ntial de stinations in the frontal matrice s of Multifrontal factorization guided by the task-DAG Matrix LU U LU LU LU U intended destination task-DAG U Fig . 4.1. An example of the inability of a task-DAG to guide complete assembly of all contribution matrices in the unsymmetric multifrontal algorithm. An 'X' denotes a nonzero in the coe#cient matrix and a '+' denotes a nonzero created due to fill-in. more than one pare nt of ge ve n if the data-DAG contains no unne ce ssarye dge s. This is be cause C g can have common rows (columns) with the frontal matrice s of more than one among g's LU- and L-pare nts (U-pare nts). The unsymme tric multifrontal algorithm muste nsure that anye ntry of a contribution matrix is not use d to update more than one frontal matrix. Additionally, a corre ct data-DAG must have su#cie nt outgoinge dge s from all supe rnode s so thate ache ntry of a contribution matrix has a pote ntial de stination in at le ast one frontal matrix. 4.3. Inadequacy of task-DAG for unsymmetric multifrontal algorithm. By me ans of a smalle xample in Figure 4.1, we show that if the task-DAG de fine d in se ction 3.3 is use d as a data-DAG, the n all contribution matrice s may not be fully absorbe d into the ir pare nt frontal matrice s. The figure shows a sparse matrix with factorization fill-in, the transitive ly re duce d DAGs L #O and U O , and the task- DAG with itse dge s labe le d as de scribe d in se ction 4.2. For the sake of clarity,e ach supe rnode is chose n to be of size 1. The figure shows all frontal and contribution (shade d portions) matrice s and the flow of data from the contribution to frontal matrice s along the e dge s of the task-DAG. Note that alle dge s may not le ad to a data LU #5. It asily se e n that the U-e dge 1U #4, which is abse nt from the task-DAG (be cause it is re move d while transitive ly re ducing U to U O ), is ne ce ssary for the comple te asse mbly of C 1 . 4.4. A data-DAG for a predefined pivot sequence. Having shown that the minimal task-DAG cannot se rve as a data-DAG for unsymme tric multifrontal we now de fine a data-DAG that is su#cie nt for the prope r asse mbly of all contribution matrice s, as long as rows and columns are note xchange d among di#e re nt supe rnode s for pivoting. We will use D N to de note such a DAG, whe re the supe rscript N stands for "no pivoting." A data-DAG D P that can accommodate UNSYMMETRIC SPARSE MATRIX FACTORIZATION 539 pivoting will be de scribe d in se ction 4.5. Theorem 4.1. If a column s all of the following conditions, the n a U-e dge i U #j is ne ce ssary for C i to be comple te ly asse mble d into its pare nts' frontal matrice s: 1. The LU-pare nt of i, if ite xists, is gre ate r than j. 2. None of i's U-pare nts are in Struct(U #,j ). 3. The re e xists a k # Struct(L #,i ) such that k > j. The transpose of this the ore m can be state d similarly. Proof. The contribution matrix C i has a column that contribute s to L #,j , be cause , at the le ast, the re is ane le me nt corre sponding to L k,j in C i . At the same time , none of i's U-pare nts' frontal matrice s have column j, so the y cannot absorb L #,j from C i . Since the LU-pare nt of i is gre ate r than j, it too cannot absorb L #,j from C i . The addition of iU #j make s it possible for C i to contribute L #,j to F j . The transpose case can be prove n similarly. The ore m 4.1 capture s the situation illustrate d in Figure 4.1 for 4, and pre scribe s the addition of toe nsure comple te asse mbly of C 1 . Theorem 4.2. If D N is a DAG forme d by adding all possible e dge s to according to The ore m 4.1, provide d that the se e dge s don't alre adye xist, the n D N is a data-de pe nde ncy DAG for the unsymme tric multifrontal algorithm without pivoting. Proof. To show that D N is a data-DAG, we must show that itse dge -se t is su#cie nt for the comple te absorption of all contribution matrice s into the ir pare nt frontal matrice s. We prove this by contradiction. Without loss of ge ne rality, assume that ane le me nt corre sponding to L k,j in C i is not asse mble d. Note that i < j < k. If L k,j is in C i , the ithe r the re is a U-path iU #j in . If the n alle ntrie s with row indice s gre ate r than ore qual to j in column of C i will be absorbe d by F j , and the se e ntrie s include the one corre sponding to L k,j . If the n a U-path iU #je xists in the re are two possibilitie s: e ithe r LU-pare LU-pare nt(i) > j. LU-pare nt(i). If l # j, the n the e ntire trailing submatrix of C i with row and column indice s gre ate r than l, including be asse mble d into F l . If l > j, the n conside r two furthe r possibilitie s:e ithe r one of i's U-pare nts is in Struct(U #,j ) or is not. If one is, the n its frontal matrix will absorb column j from C i . If none of i's U-pare nts is in Struct(U #,j ), the n all conditions for the applicability of The ore m 4.1 are satisfie d. The re fore , iU #j would have be e n adde d to D N and would have cause d the e ntry corre sponding to L k,j in C i to be absorbe d into F j . Thus, it is not possible for the e ntry corre sponding to L k,j to be le ft unasse mble d in any C i . Similarly, it can be shown that the e ntry corre sponding to any U j,k cannot be le ft unasse mble d in any C i . Having shown that the e dge -se t of D N is su#cie nt for unsymme tric multifrontal factorization without pivoting, we now show that not alle dge s that D N inhe rits from may be ne ce ssary if pivoting is not pe rforme d during factorization. Theorem 4.3. For LU factorization without pivoting, ane dge i U #j (i L in is re dundant if the maximum inde x in Struct(L #,i ) (Struct(U i,# )) is smalle r than j. Proof. Re call that Struct(L #,j the maximum inde x in Struct(L #,i ) is smalle r than j, the n doe s not contribute to Struct(L #,j ). The proof for L #O and Struct(U i,# ) is similar. Note that The ore m 4.3 is valid only if row and columne xchange s are not pe r- forme d during LU factorization. Othe rwise , additional fill-in cause d by pivoting could cre ate an inde x gre ate r than ore qual to j in Struct(L #,i ) or Struct(U i,# ),e ve n if it is not pre dicte d by the symbolic factorization on the original pe rmutation of the matrix. The re fore , alle dge s in could pote ntially be use d. Supe rnodal ve rsions of The ore ms 4.1-4.3 for T S can be prove n similarly. To summarize the re sults of this subse ction, we have shown how to construct a data- DAG for unsymme tric multifrontal factorization without pivoting from a task-DAG and we have shown that although the task-DAG is de rive d from the strict transitive re ductions of L # and U (or L # and U), it may still pass one dge s to the data-DAG that are re dundant if pivoting is not pe rforme d during factorization. The re fore , the data-DAG is not minimal. Howe ve r, if pivoting is pe rforme d, the n pote ntially all the e dge s could ge t use d. 4.5. Supplementing the data-DAG for dynamic pivoting. We will now show that the e dge -se t of data-DAG D N constructe d in se ction 4.4 may not be su#- cie nt if pivoting is pe rforme d during factorization. We also discuss how to supple me nt to ge ne rate a data-DAG D P whose e dge -se t is su#cie nt to handle any amount of pivoting. We start with an ove rvie w of the pivoting me thodology in the unsymme tric multifrontal algorithm, which has be e n de scribe d in de tail in [20]. If a diagonale le me nt A i,i (q # i # r) in a supe rnode [q : r] fails to me e t the pivoting crite rion, the n first an atte mpt is made toe xchange row and column i with a row j and a column k such that i < the pivoting crite rion. Such intrasupe rnode pivoting has noe #e ct on the structure of the factors and factorization can continue as usual. Howe ve r, it may not always be possible to find a suitable row-column pair within a supe rnode 's pivot block to satisfy the pivoting crite rion. In this situation, inte rsupe rnode pivoting is ne ce ssary. If the LU-pare nt of the data-DAG and a suitable ith pivot cannot be found within the pivot block of F g , the n all row-column pairs from i to r are symme trically pe rmute d to ne w locations from s - (r Thus,e #e ctive ly, shrinks to [q the supe rnode [s : t]e xpands to [s - (r As a side e #e ct of this pivoting, the re is additional fill-in in all the ance stors of g in the data-DAG that are smalle r than h. In particular, the columns of L of all of g's U-ance stors smalle r than h ge te xtra row indice the rows of U of all of g's L-ance stors smalle r than h ge te xtra column indice failure in supe rnode h is handle d similarly in a re cursive manne r. In D N , whose construction is de scribe d in se ction 4.4, all supe rnode s may not have an LU-pare nt to support the symme tric pivoting me thod de scribe d above . The re fore , as the first ste p towards de riving D P from D N , we alte r the e dge -se t of the latte r as follows. Fore ach g from 1 to m re m is the total numbe r of supe rnode s), the smalle st supe rnode h to which both g L #h and gU #he xist is de signate d as the LU-pare nt of g; that is, if ane dge g#h doe s note xist, the n an LU-e dge g LU #h is adde d to the data-DAG, or if an L- or a U-e dge g#he xists, the n it is conve rte d to an LU-e dge . The n, alle dge s g#k such that k > h are de le te d. If the original matrix is not re ducible to a block-triangular form, the n afte r this modification,e ach supe rnode othe r than the root supe rnode has an LU-pare nt to accommodate row-column pairs that fail to satisfy the pivoting crite rion in the ir original locations [20]. It asily se e n that this modification has noe #e ct on The ore ms 4.1-4.3 be cause g L #h (gU #h) is in the modifie d D N only if g L #h (gU #h) is in the original D N as de fine d in se ction 4.4. UNSYMMETRIC SPARSE MATRIX FACTORIZATION 541 old indices Original matrix Multifrontal factorization without any pivot failure FLU Factorization with failure of pivot 1 Handling failure of pivot 1 new unassembled intended destination data-DAG13LU U LU LU Fig . 4.2. An example factorization to show how the failure of pivot 1 is handled by a symmetric permutation of row and column 1 to merge themw ith their LU-parent supernode, 4. An 'X' denotes a nonzero in the coe#cient matrix and a '+' denotes a fill-in. The circled `X' and '+' are created due to pivoting. A '0' denotes a fill-in predicted by the original symbolic factorization that has a value of zero due to pivoting-related movement of row s and columns. The figure also show s that the absence of 2 L #4 leaves the entry U 1,5 unassembled from C 2 . Figure 4.2 shows how the failure of pivot row and column 1 is atte mpte d in the unsymme tric multifrontal factorization of a small 5 - 5e xample matrix. Row-column 1 is symme trically pe rmute d to a ne w location adjace nt to 1's LU-pare nt 4 in the data- DAG. This re sults in an addition of row inde x 1 to 1's U-pare nt 2 and an addition of column inde x 1 to 1's L-pare nt 3. Additionally, afte r moving to the ir ne w locations, row 1 in U and column 1 in L ge t fill-in in column and row positions re row 4 in U and column 4 in L have nonze ros (i.e ., U 1,5 , L 4,1 , and L 5,1 ). Figure 4.2 also shows that afte r pivoting, the ne w row 1 of C 2 cannot be fully asse mble d in the abse nce of an L-e dge 2 L #4. arly, in addition to adding dge s as de scribe de arlie r, D N re quire s furthe r modifications in orde r to se rve as a data-DAG for unsymme tric multifrontal algorithm with dynamic pivoting. Figure give s anothe re xample of a factorization whe re D N is unable to guide a comple te asse mbly in the e ve nt of a pivot failure . Note that thise xample satisfie s the first two conditions of The ore m 4.1. Howe ve r, since it doe s not satisfy condition 3, no dge s are adde d and the abse nce of pre clude s a comple te asse mbly of C 2 into its pare nts' frontal matrice s We now state and prove a the ore m that pre scribe s a modification of D N to pre ve nt the situation illustrate d in Figure 4.3. Theorem 4.4. If a column s all of the following conditions, the n a U-e dge i U #j is ne ce ssary for C i to be comple te ly asse mble d into its pare nts' frontal matrice s in the e ve nt of failure of pivot k. 1. The LU-pare nt of i is gre ate r than j. 2. None of i's U-pare nts are in Struct(U #,j ). 3. A xists such that the re is a U-path k U #i in LU-pare nt(k) > j. The transpose of this the ore m can be state d similarly. Proof. Note that The ore m 4.4 is ve ry similar to The ore m 4.1. The only di#e re nce is condition 3. If pivot k fails, the n it will add a row in Struct(L #,i ) that corre sponds to LU-pare nt(k) -1, which is the ne w location of k and is gre ate r than j - 1, the ne w inde x for j. Thus, the failure of pivot k transforms condition 3 of The ore m 4.4 into condition 3 for the applicability of The ore m 4.1, which has alre ady be e n prove d. The ore m 4.4 state s thate ve n if Struct(L #,i ) doe s not have any inde x gre ate r than j but all othe r conditions for the applicability of The ore m 4.1 are satisfie d and iU #j is not pre se nt in the DAG, the n pivoting may re sult in incomple te asse mbly unle ss thise dge is adde d. This is be cause pivoting can cre ate a nonze roe ntry L k,i such that j. This is what happe ns in the e xample shown in Figure 4.3 for the original indice s. Pivoting change s i, j, and k to 1, 7, and 8, re spe ctive ly. In light of The ore m 4.4, we introduce anothe r modification to D N . Inste ad of using The ore m 4.1 strictly to de rive D N from we omit che cking for condition 3 and de rive D N by adding all those e dge s to practice ) that satisfy conditions 1 and 2. Now, by me ans of The ore m 4.5, we will show that the data-DAG D N ,e ve n afte r the modifications de scribe d above , is not su#cie nt toe nsure comple te asse mbly of all contribution matrice s in the e ve nt of inte rsupe rnode pivoting. The re ade r can ve rify that Figure the transpose case of The ore m 4.5 for 5. Finally, The ore m 4.6 will show that supple me nting the data- DAG with additionale dge s pre scribe d by The ore m 4.5 make s it su#cie nt to handle all contribution matrice s in the face of inte rsupe rnode pivoting. As we dide arlie r in this pape r, for the sake of clarity and simplicity, we will state and prove The ore ms 4.5 and 4.6 in the conte xt of conve ntional DAGs with single -node supe rnode s. The re sults naturallye xte nd to supe rnodal DAGs. Theorem 4.5. If h is the LU-pare nt of j and all of the following conditions hold, the n a U-e dge i U #h is ne ce ssary for C i to be comple te ly asse mble d into its pare nts' frontal matrice s in the e ve nt that j fails to me e t the pivot crite rion in its original location. 1. The re e xists an L-path j L #i such that i < h and LU-pare nt(i) > h. 2. None of i's U-pare nts are in Struct(L #,j ). 3. Eithe r # k # Struct(L #,i ) such that k > h, or the re is a U-path k U LU-pare The transpose case can be state d similarly. UNSYMMETRIC SPARSE MATRIX FACTORIZATION 543 indices indices new F 9 F 486423748 9 LU UU U U U U U Original matrix data-DAG9 X unassembled5 Handling failure of pivot 1 intended destination LU LU LU LU LU LU Fig . 4.3. An example factorization to show that the edges in D N are not su#cient to assemble its parents' frontal matrices in the event of the failure of pivot 1. The convention for representing di#erent types of structural nonzeros is the same as in Figure 4.2. Proof. If pivot j fails, the n, along with the othe r faile d LU-childre n of h, it occupie s a ne w position just be fore h. Since the re is an L-path j L #i, column j is adde d to C i afte r the failure of pivot j; that is, in the ne w matrix afte r pivoting, j # We know that the LU-pare nt of i is gre ate r than the ne w j, be cause LU-pare nt(i) > h. Since none of i's U-pare nts we re in the old Struct(L #,j ), the y are not in the ne w Struct(U #,j )e ithe r. Thus the first two conditions for the applicability of The ore ms 4.1 and 4.4 are satisfie d. Condition 3 of The ore m 4.5 quivale nt to condition 3 of The ore ms 4.1 and 4.4. The re fore , a U-e dge iU #j is ne e de d for prope r multifrontal factorization of the ne w matrix afte r pe rmuting j to its ne w location. Since , in its ne w location, j is me rge d with h into a common supe rnode , a U-e dge iU #h in the original matrix would have su#ce d. The proof of the transpose case is similar. Theorem 4.6. If D P is a DAG forme d by adding all possible e dge s according to The ore m 4.5 to D N , the n D P is an ade quate data-DAG for unsymme tric multifrontal factorization with pote ntially unlimite d inte rsupe rnode pivoting. Proof. We prove this by showing that with D P , it is not possible for anye le me nt of a contribution matrix C i to re main unasse mble d. Without loss of ge ne rality, conside r ane le me nt corre sponding to L k,j in C i . If L k,j is in C i , the ne ithe r k # Struct(L #,i ) the original L and U pre dicte d by symbolic factorization, or row k or column j or both we re adde d to C i due to pivoting. If row k and column j are parts of the original structure of C i , the n The ore m 4.2 has alre ady shown that the e dge -se t of D N , which is a subse t of the e dge -se t of D P , is su#cie nt to asse mble L k,j . We now show that L k,j will be absorbe d from C i by one of i's pare nts in D P adde d to C i due to pivoting, irre spe ctive of whe the r row k be longe d to the original Struct(L #,i ) or if it too was adde d to C i due to pivoting. LU-pare nt(i) and LU-pare nt(j). We conside r two case s: (1) g # h and (2) g > h. If g # h, F g will have both row k and column j and will absorb the e le me nt corre sponding to L k,j from C i . If g > h, the n the first condition for the applicability of The ore m 4.5 has be e n d. Now we conside r two furthe r sce narios: (2a) At le ast one of i's U-pare nts is in the original Struct(L #,j ), or (2b) none of i's U-pare nts is in the original Struct(L #,j ). In case of (2a), afte r pivoting, at le ast one of i's U-pare nts is in the ne w Struct(U #,j ) and the frontal matrix of this U-pare nt will absorb column j from C i , including the e ntry corre sponding to L k,j . In case (2b), the se cond condition for the applicability of The ore m 4.5 has be e n d. Finally, whe the r row k was in the original Struct(L #,i ) or was adde d to C i due to the failure of a U-de sce nde nt k, in its final location, k must be gre ate r than h. The re ason is that if j # k # h (i.e ., k's ne w location is in the e xte nde d supe rnode h), the n h must be an LU-ance stor of i be cause implie s that the re are both i L #h and iU #h in the data-DAG. But that is not possible be cause we are alre ady working unde r the assumption that the LU-pare nt g of i is gre ate r than h. The re fore , k > h and the third condition of The ore m 4.5 has also be e n d. As a re sult, The ore m 4.5 would have e nsure d that a U-e dge iU #h is pre se nt in D P to asse mble column j from Similarly, we can prove that noe ntry corre sponding to any U j,k will be le ft unasse mble d in C i . 4.6. Experimental results. In se ctions 4.4 and 4.5, we showe d how to supple me nt the e dge -se t of the task-DAG to construct a data-DAG for the unsym- me tric multifrontal algorithm. Table showse xpe rime ntal re sults of WSMP's UNSYMMETRIC SPARSE MATRIX FACTORIZATION 545 Table Time required for and the number of edges in each DAG. Matrix Symbolic Supplement-1 Supplement-2 Total Time |ED P | bbmat 1.7 6077 .06 6142 .00 6181 1.76 1.02 mil053 4.5 166154 .51 166154 .07 166154 5.08 1.00 mixtank 1.2 3203 .05 3203 .00 3203 1.25 1.00 pre2 6.4 317216 .84 320063 .16 320942 7.40 1.01 twotone .91 44856 .12 45866 .01 45918 1.04 1.02 wang3old .54 8450 .03 8450 .00 8450 0.57 1.00 imple me ntation of the proce dure s to ge ne rate the various DAGs. Thre e DAGs are conside re d in Table 4.1: the supe rnodal task-DAG T S , the supe rnodal data-DAG D N for unsymme tric multifrontal factorization without pivoting, and the supe rnodal data-DAG D P for unsymme tric multifrontal factorization with pivoting. The table shows the time to compute e ach of the DAGs and the numbe r dge s in the m for the matrice s in our te st suite . T S is compute d by the basic symbolic factorization algorithm de scribe d in se ction the re fore , t S is the basic symbolic factorization time . We re fe r to the proce ss of computing D N from T S as Supple me nt-1. Supple me nt-1 che cks for the first two conditions of The ore m 4.1 to find the e dge s to be adde d to E T S and the n adds outgoing dge s from supe rnode s without LU-pare nts to ld E D N . Supple me nt-2 is the proce ss that addse dge s base d on the first two conditions of The ore m 4.5 to to ld The e xe cution time of Supple me nt-1 and Supple me nt-2 is de note d by t 1 and t 2 , re spe ctive ly. Note that not all the e dge s in D N and D P are ne ce ssary. For the sake of com- putational spe e d, Supple me nt-1 and Supple me nt-2 in WSMP do not che ck for all the conditions of The ore ms 4.1, 4.4, and 4.5 while addinge dge s. The last conditions of all thre e the ore ms are skippe d. Eve n if all conditions of the se the ore ms we re che cke d, not all the e dge s in the re sulting data-DAGs may be ne ce ssary. The re fore , D N and D P are not minimal data-DAGs for unsymme tric multifrontal factorization. Howe ve r, as Table 4.1 shows, the se DAGs do not have many more e dge s than T S for most re al-life matrice s. The ave rage fore xce sse dge s in supe rnodal D P ove r T S is only about 4% for our te st suite . We have shown that the e dge s in the task-DAGs or T S are insu#cie nt to dire ct the data flow in unsymme tric multifrontal factorization. On the othe r hand, the e dge s in D P are su#cie nt,e ve n with pivoting. The re fore , the numbe r dge s in a truly minimal supe rnodal data-DAG is some whe re be twe e n the numbe r dge s in T S and in the supe rnodal D P . The e xpe rime ntal re sults in Table 4.1 show that the se two rs are usually fairly close . The table also shows that the time re quire d to construct D N and D P is also small compare d to the basic symbolic torization time . Thus, the me thodology de scribe d in this se ction for the construction of data-DAGs for unsymme tric multifrontal factorization nt in both time and the numbe r of DAGe dge s. A comparison of the Table 4.1 with WSMP's LU factorization time give n in Table 5.1 shows that the total symbolic time is usually significantly le ss than the nume rical factorization time . 5. Implementation details of unsymmetric factorization. A brie f outline of the unsymme tric multifrontal algorithm base d on the work of Hadfie ld [20] and Davis and Du# [9] is found in se ction 4.2. We now add more de tails to it and pre se nt a comple te algorithm that is imple me nte d in WSMP. WSMP is ge are d towards multiple factorizations of matrice s with the same sparsity patte rn but di#e re nt nonze ro value s. The re fore , symbolic phase is pe rforme d only once and its output is use d in all subse que nt nume rical ve n with di#e re nt pivot se que nce s re sulting from di#e re nt nume rical value s. A fundame ntal data-structure in our unsymme tric multifrontal algorithm is the frontal matrix. A frontal matrix is associate d withe ach supe rnode . Figure 5.1 shows the organization of a typical frontal matrix for a supe rnode The core of this frontal matrix is a| Struct(L #,q )| -| Struct(U q,# )| portion, re are pre dicte d by the symbolic factorization. In the abse nce of pivoting, the first r - q columns of this matrix would be factore d and would be save d as parts of U and L, re spe ctive ly. The re maining trailing submatrix would constitute the contribution matrix whose conte nts would be absorbe d into the frontal matrice s of the pare nts of g in D P . Extra rows F [q:r] s [q:r] Extra columns r pivot block pivot block Fig . 5.1. Organization of a typical frontal matrix for a supernode r]). The p failed pivots from the LU-children of the supernode are appended at the beginning of the frontal matrix and the extra row s and columns inherited from U- and L-descendents, respectively, are appended at the end. UNSYMMETRIC SPARSE MATRIX FACTORIZATION 547 In the pre se nce of nume rical pivoting,e xtra pivots as we ll as othe r rows and columns may be adde d to the frontal matrix de pe nding on the labe ls and pivot failure s of the childre n of g in D P . Extra pivots (row-column pairs with the same indice s) are adde d to F g if some of the pivots of g's LU-childre n fail to satisfy the pivoting crite rion. The LU-childre n of g the mse lve s may have inhe rite d some or all of the se faile d pivots from the ir own LU-childre n. The re fore , faile d pivots from any of the LU-de sce nde nts of g cane nd up in its frontal matrix. If p such pivots are adde d, the n the size of the pivot block incre ase s from r - q The frontal matrix F g can similarly inhe rite xtra rows corre sponding to faile d pivots in its U-de sce nde nts whose LU-pare nts are gre ate r than g ande xtra columns corre sponding to faile d pivots in its L-de sce nde nts whose LU-pare nts are gre ate r than g. Irre spe ctive of the ir ne w indice s, the se e xtra rows and columns are always ap- pe nde d at the e nd of the original rows and columns of F g and a d list of the ir indice s is maintaine d ate ach supe rnode . Eve ntually, the se are asse mble d into the e xtra pivots of the frontal matrice s of the LU-pare nts of the supe rnode s whe re the se pivots faile d. The row and column structure s pre dicte d by symbolic factorization are intact for future factorizations of matrice s with the same nonze ro patte rn. The additions to the se structure s due to pivoting, which de pe nd on the nonze ro value s in the matrix be ing factore d, are maintaine d se parate ly and are discarde d be fore e ach ne w factorization. The availability of a static data-DAG D P that is su#cie nt for handling an arbitrary amount of dynamic pivoting is critical to our imple me ntation of the unsymme tric multifrontal algorithm. Figure give s a high-le ve l pse udocode of our factorization algorithm. The algorithm starts with the root supe rnode of task- and data-DAGs. At any supe rnode , first, it re cursive ly factors all the unfactore d childre n of that supe r- node . The n it looks at the faile d pivots (if any) of its childre n to figure out the numbe r and indice s of the e xtra rows, columns, and pivots, if any, and accordingly allocate s a frontal matrix of the appropriate size . In the ne xt ste p, the contribution from the original coe #cie nt matrix and the contribution matrice s of the curre nt supe rnode 's childre n are accumulate d in the appropriate locations inside the frontal matrix. Finally, the algorithm proce e ds to factor the pivot block of the frontal matrix and update s the re mainde r of the frontal matrix. The le ading succe ssfully factore d rows and columns are save d as portions of U and L for use during triangular solve s. The re maining contribution matrix ve ntually asse mble d into the frontal matrice s of its pare nts and is re le ase d by the last pare nt to pick up its contribution. The frontal matrix of the LU-pare nt of a supe rnode picks up all its faile d pivot row-column pairs as we ll as the e ntire trailing submatrix of its contribution matrix with row and column indice s gre ate r than ore qual to the first inde x of the pare nt supe rnode . The re maining rows and columns of a supe rnode 's contribution matrix are asse mble d into the frontal matrice s of its L- and U-pare nts in D P . It is possible for more than one L- or U-pare nts' frontal matrice s to have the same row or column indice s in common with the child's contribution matrix. Howe ve r,e ache le me nt of a contribution matrix must be adde d intoe xactly one frontal matrix. Some simple bookke e ping to track of rows and columns that have be e n asse mble d su#ce s to e nsure this condition for the re lative ly fe w rows and columns that have the pote ntial to be copie d into the frontal matrice s of multiple L- and U-pare nts, re spe ctive ly. Figure 5.2 and the de scription in this se ction show that WSMP's unsymme tric multifrontal algorithm is fairly straightforward to imple me nt. The static task- and data-DAGs compute d during the symbolic phase and the use of re cursion make the 548 ANSHUL function uns mf (root) { /* 1. Re cursive calls to root's childre n */ fore ach child k of root in T S do if not alre ady proce sse d k then Call uns mf (k); Flag supe rnode k as alre ady proce sse d; end for /* 2. Colle ct pivoting info to de te rmine size of F root */ fore ach child k of root in D P do if k is an L-child then if k has faile d pivots then Add the m to the sorte d list of F root 'se xtra columns; hase xtra columns then Add those whose LU-pare nt is gre ate r than root to the sorte d list of F root 'se xtra columns while che cking for duplicate s; else if k is a U-child then if k has faile d pivots then Add the m to the sorte d list of F root 'se xtra rows; hase xtra rows then Add those whose LU-pare nt is gre ate r than root to the sorte d list of F root 'se xtra rows while che cking for duplicate s; else if k is an LU-child then if k has faile d pivots then Add the m to the sorte d list of F root 'se xtra pivots; hase xtra columns then Add those whose LU-pare nt is gre ate r than root to the sorte d list of F root 'se xtra columns while che cking for duplicate s; hase xtra rows then Add those whose LU-pare nt is gre ate r than root to the sorte d list of F root 'se xtra rows while che cking for duplicate s; end for /* 3. Initialize root's frontal matrix */ Allocate F root of appropriate size and fill it with ze ros; Populate F root withe ntrie s from A corre sponding to supe rnode root; /* 4. Asse mbly from childre n's contribution matrice s into F root */ fore ach child k of root in D P do Copy appropriate contribution from C k into F root ; if root is the last pare nt of k to pick up C k 's contribution then Fre e the space occupie d by C end for 5. Nume rical factorization */ Factor the pivot block of F root and update the trailing part to ld C root ; function uns mf. Fig . 5.2. A simple and e#cient unsymmetric multifrontal algorithm. UNSYMMETRIC SPARSE MATRIX FACTORIZATION 549 nume rical factorization algorithm much simple r to de scribe and imple me nt than the e arlie r de scriptions of the unsymme tric patte rn multifrontal algorithm in [20] and [9]. Othe r than UMFPACK [8], WSMP is the only sparse solve r available that is base d on an unsymme tric patte rn multifrontal algorithm. It is also the first such paralle l solve r available for ge ne ral use . Although Hadfie ld [20] provide de xpe rime ntal re sults from a paralle l imple me ntation on the nCUBE, a practical paralle l solve r did not re sult from thate #ort. The algorithm of Figure 5.2 is not only re lative ly simple in de scription but is also computationally le an be cause it the none sse ntial non-floating-point ope rations and can handle pivot failure s fairlye #cie ntly. It is also note worthy that for structurally symme tric matrice s, the algorithm in Figure 5.2 naturally re duce s to a symme tric-patte rn multifrontal algorithm guide d by ane limination tre e , which re place s both T S and D P . Othe r than a state me nts fore ach supe rnode , the re is no ove rhe ad in using this algorithm for structurally symme tric matrice s. 5.1. Experimental results. We now compare the unsymme tric LU factoriza- tion time of WSMP with that of thre e state -of-the -art multifrontal sparse solve rs, name ly, MUMPS ve rsion 4.1.6 [4, 5], MA41 [2, 3], and UMFPACK ve rsion 3.2 [8]. A de taile d comparative study that include s more solve rs can be found in [18]. The software compare d in this se ctione mploy di#e re nt variants of the multifrontal me thod. MUMPS contains a symme tric-patte rn multifrontal factorization code base d on the classical multifrontal algorithm [14]. MA41, in some se nse , is a hybrid be twe e n sym- me tric and unsymme tric patte rn multifrontal solve rs. It use s ane limination tre e to guide factorization, but the frontal matrice s are prune d of all-ze ro rows and columns. UMFPACK 3.2 contains a variation of the unsymme tric-patte rn multifrontal algorithm [9] that use s ane limination tre e de rive d from the structure of A # A. Apart from the factorization algorithm, the re are othe r significant di#e re nce s among the four software package s that a#e ct the ir pe rformance . First, the y use diffe re nt sche me s for fill-re ducing ring. By de fault, WSMP use s a symme tric pe rmu- tation base d on a ne ste d-disse ction ring [17] compute d on the structure of A+A # . MUMPS and MA41 use a symme tric pe rmutation base d on the approximate mini- mum de gre e (AMD) algorithm [1] applie d to the structure of A+A # . UMFPACK use s a column AMD algorithm [10] to pre pe rmute only the columns of A and compute s a row pe rmutation base d on nume rical and sparsity crite ria during factorization. The se cond di#e re nce is the use of a maximal matching algorithm [13] to pe rmute the rows of the coe #cie nt matrix to maximize the product of the magnitude s of its diagonal e ntrie s. As shown in [6, 18], this can a#e ct factorization time s be cause it change s the amount of structural symme try and the amount of nume rical pivoting during factor- ization. WSMP use s this pre proce ssing on all matrice s, MUMPS and MA41 use it only if the structural symme try in the original matrix is le ss than 50%, and UMF- doe s not use it at all. The third di#e re nce is that WSMP re duce s the coe #cie nt matrix into a block-triangular form, while MUMPS, MA41, and UMFPACK 3.2 do not. Table 5.1 shows nume rical factorization time s and ope ration counts of MUMPS, MA41, UMFPACK, and WSMP run with the options in MUMPS, MA41, and WSMP change d to minimize the di#e re nce s be twe e n the code s othe r than the factorization algorithm. We switche d o# the pe rmutation to a he avy-diagonal form and the associate d scaling in MUMPS, MA41, and WSMP. For WSMP, inste ad of its de fault ne ste d-disse ction ring, we use d an approximate minimum fill orde ring, which is ve ry similar to AMD. Eve n with the se change s, di#e re nce s re main be twe e n the four Table LU factorization times and operation counts of MUMPS, MA41, UMFPACK 3.2, and WSMP w ith similar permutation options. The best time is in boldface and the second best time is underlined. 3.2W SMP Matrix time ops time ops time ops time ops af23560 3.89 2.56 3.58 2.54 8.59 3.46 4.06 3.22 bbmat 54.3 41.6 56.3 41.1 78.7 39.1 27.6 21.5 e40r0000 4.93 2.53 3.63 1.58 6.23 2.17 0.80 .419 ecl32 64.2 64.6 67.1 64.4 191. 112. 139. 77.6 fidap011 8.58 7.01 8.79 6.96 17.0 8.51 6.51 5.74 mil053 43.5 31.8 40.0 31.8 107. 46.2 28.2 20.8 mixtank 151. 141. 152. 64.1 363. 243. 76.3 64.6 nasasrb 12.8 9.45 11.9 9.43 55.9 28.2 10.4 8.78 pre2 fail fail fail fail fail fail 346. 301. raefsky3 4.44 2.90 3.88 2.90 16.0 7.87 4.88 4.17 raefsky4 107. 74.4 92.9 44.7 25.0 12.9 43.4 22.5 twotone 56.5 38.3 37.6 31.8 30.1 10.8 2.87 1.49 wang3old 72.9 57.8 57.7 51.0 40.6 24.2 45.8 32.3 wang4 11.8 10.5 12.2 10.5 53.4 30.7 8.84 7.94 code s. For instance , MUMPS, MA41, and WSMP first pe rmute the matrix such that it has a diagonal of structural nonze ros. This initial pe rmutation is the same for MUMPS and MA41 be cause both use the same code to compute it. Howe ve r, it can be di#e re nt for WSMP. The pivoting strate gy of UMFPACK base d on row is inhe re ntly di#e re nt from the symme tric inte rsupe rnode pivoting strate gy use d in MUMPS, MA41, and WSMP. WSMP's algorithms work only with a pe rmutation to the block-triangular form, which is not imple me nte d in MUMPS, MA41, and UMF- PACK. Howe ve r, with the e xce ption of comp2c, the e #e ct of block-triangularization on the ope ration count for factorization is insignificant, if any. As a re sult of the se di#e re nce s and due to the fact that MUMPS may pe rform more ope rations than ne c- e ssary on structurally unsymme tric matrice s, the factorization ope ration counts for the four code s in Table are di#e re nte ve n with a similar ring algorithm for fill-re duction. In Table 5.1, the faste st factorization time fore ach matrix is in boldface and the se cond faste st time is unde rline d. Although di#e re nce s othe r than the factorization algorithm itse lf a#e ct the pe rformance of the se code s, it ise asy to se e the broad picture thate me rge s from Table 5.1. Most of the boldface e ntrie s are in the WSMP column and most of the unde rline de ntrie s are in the MA41 column. For many matrice s, the e #e ct of the algorithmic choice s of the software ise vide nt in the factorization statistics in Table 5.1. MUMPS usually re quire s more floating-point ope rations for factorization than MA41 and WSMP be cause it use s artificially symme trize d frontal matrice s padde d with ze ros. For the same re ason, UMFPACK is faste r than MUMPS UNSYMMETRIC SPARSE MATRIX FACTORIZATION 551 for ve ry unsymme tric matrice s (such as baye r01, one tone 2, and twotone howe ve r, it is slowe r for matrice s with more structural symme try (such as fidap011, mil053, and wang4) partly be cause it use s a fill-re ducing pe rmutation on the columns of the coe #cie nt matrix be fore starting LU factorization. MA41 o#e rs a significant improve me nt ove r MUMPS for matrice s with a ve ry unsymme tric structure , such as comp2c, one tone 1, and twotone . It se e ms that MA41's me chanism for finding all- ze ro rows and columns incurs a slight ove rhe ad that it cannot o#se t for matrice s with a ne arly symme tric structure (such ase cl32, fidap011, and wang4), for which it is some what slowe r than MUMPS. It is cle ar from Table 5.1 that WSMP has the smalle st ove rall factorization time se ve n its de fault options are modifie d to compare it with the othe r solve rs. With its de fault options, WSMP's factorization time s are usually much smalle r [18] than those shown in Table 5.1. 6. Concluding remarks. This pape r de scribe s sparse unsymme tric symbolic and nume rical factorization algorithms that improve pre vious similar algorithms. Our phase , in particular, is more powe rful than othe rs de scribe d in the lite rature . It ine xpe nsive ly compute s minimale limination structure s that are transitive re ductions of the uppe r and lowe r triangular factors of the original coe f- ficie nt matrix. In addition, it compute s ne ar-minimal data-de pe nde ncy DAGs for unsymme tric multifrontal factorization with and without pivoting. A data-DAG that has only a slightly highe r ofe dge s than a minimal task-DAG and that is capable ofe xpre ssing all possible data-de pe nde ncie s in the face of dynamic pivoting is a fe ature of our symbolic phase . We show how this data-DAG aids in a high-pe rformance imple me ntation of the unsymme tric multifrontal LU factorization algorithm. This factorization algorithm is not only faste r than othe r sparse LU factorization algorithms but is also simple r than the unsymme tric multifrontal algorithms de scribe d pre viously in the lite rature . Our algorithms do not introduce additional ove rhe ads while factoring matrice s with a symme tric nonze ro patte rn. Whe n pre se nte d with a sparse matrix with a symme tric structure , both the symbolic and the nume rical factorization algorithms and the data-structure s ge ne rate d by the m grace fully transform into the ir symme tric counte rparts without re quiring any significant amount ofe xtra proce ssing or storage . In a distribute d-me mory paralle l imple me ntation of unsymme tric sparse LU the e dge s of the data-DAG conne cting tasks mappe d onto di#e re nt proce sse s de te rmine the inte rproce ss communication patte rn. The static and ne ar- minimal nature of the data-DAG use d in our algorithms would be e xtre me ly use ful for pote ntial paralle l imple me ntations of unsymme tric multifrontal factorization, re changing the data-DAG dynamically could be cumbe rsome and ine #cie nt and the unne ce ssary dge s could incre ase synchronization and communication ove rhe ads. Acknowledgments . The author wishe s to thank Andre w Conn, Fre d Gustavson Jose ph Liu, Sivan Tole do, and the anonymous re fe re e s fore xtre me ly use ful comme nts one arlie r drafts of this pape r. --R An approximate minimum degree ordering algorithm Vectorization of a multiprocessor multifrontal code Memory management issues in sparse multifrontal methods on multiprocessors A fully asynchronous multifrontal solver using distributed dynamic scheduling Multifrontal parallel distributed symmetric and unsymmetric solvers Analysis and comparison of tw sparse solvers for distributed memory computers The influence of relaxed supernode partitions on the multifrontal method An unsymmetric-pattern multifrontal method for sparse LU factorization A Column Approximate Minimum Degree Ordering Algorithm A supernodal approach to sparse partial pivoting Direct Methods for Sparse Matrices On algorithms for permuting large entries to the diagonal of a sparse matrix The multifrontal solution of unsymmetric sets of linear equations Exploiting structural symmetry in unsymmetric sparse symbolic factorization Elimination structures for unsymmetric sparse LU factors Fast and e Recent advances in direct methods for solving unsymmetric sparse systems of linear equations Highly scalable parallel algorithms for sparse matrix factorization On the LU Factorization of Sequences of Identically Structured Sparse Ma- tricesw ithin a Distributed Memory Environment A scalable sparse direct solver using static pivoting The role of elimination trees in sparse factorization The multifrontal method for sparse matrix solution: Theory and practice --TR --CTR Timothy A. Davis, A column pre-ordering strategy for the unsymmetric-pattern multifrontal method, ACM Transactions on Mathematical Software (TOMS), v.30 n.2, p.165-195, June 2004
sparse matrix factorization;parallel sparse solvers;multifrontal methods
589360
Numerical Approximation of an SQP-Type Method for Parameter
This paper deals with the numerical approximation of the Levenberg--Marquardt SQP (LMSQP) method for parameter identification problems, which has been presented and analyzed in [M. Burger and W. Mhlhuber, Inverse Problems, pp. 943--969]. It is shown that a Galerkin-type discretization leads to a convergent approximation and that the indefinite system arising from the Karush--Kuhn--Tucker (KKT) system is well-posed.In addition, we present a multilevel version of the Levenberg--Marquardt method and discuss the simultaneous solution of the discretized KKT system by preconditioned iteration methods for indefinite problems. From a discussion of the numerical effort we conclude that these approaches may lead to a considerable speed-up with respect to standard iterative regularization methods that eliminate the underlying state equation. The numerical efficiency of the LMSQP method is confirmed by numerical examples.
Introduction Parameter identication denotes the procedure of determining unknown parameters appearing in an underlying state equation (usually a partial dierential equation), from indirect measurements related to the solution of this equation. Such problems frequently appear in many applications, where mathematical models of physical, chemical, biological or economical processes are used (cf. e.g. [1, 12, 16] and the references there). Since such problems are ill-posed in general, i.e., the parameter to be reconstructed does not depend on the observation in a stable way, regularization methods have to be used in order to compute a stable approximation of the parameter in presence of data noise. Due to the ill-posedness of the identication problem, the numerical approximation of such problems is not a simple task. The standard approach that can be found in literature is based on an a-priori elimination of the state equation, and an application of a discretized regularization method to the resulting operator equation involving the parameter-to-output map, which is the operator mapping the parameter to the corresponding observation. The main part in the This work has been supported by the Austrian National Science Foundation FWF under project grants F 13/08 and F 13/09. INTRODUCTION evaluation of this map is the solution of the underlying state equation for given parameter, which is numerically realized by standard discretizations such as nite elements. This approach, in particular combined with iterative regularization methods (cf. [17] for an overview), has been applied with success even to rather complicated parameter identication problems (cf. e.g. [9, 24, 25]). However, since this methods need a high number of direct solves (i.e., solutions of the state equation), ne discretizations of the parameter yield a considerable computational eort, which results in high CPU-times or even in the impossibility to use ne discretizations. Another drawback of this approach is that the discretizations of state and parameter are rather independent, which makes the numerical analysis extremely di-cult. Therefore, fundamentally dierent methods for the solution of parameter identication problems have been investigated recently, whose common idea is to avoid the a-priori elimination of the state equation (cf. [10, 20, 26]). The aim of this paper is to discuss the numerical approximation of an iterative regularization method based on the idea of sequential quadratic programming (cf. [10]). We investigate Galerkin-type discretizations in the product space for parameter, state variable and a corresponding Lagrangian variables, which leads to a sequence of well-posed indenite systems. With this approach we are able to show convergence of the numerical approximation both for the quadratic programming problem arising in each iteration step and for the overall minimization procedure. The general setup in this paper is as follows: we assume that we are given a noisy measurement z - satisfying where the exact data satisfy z := E^u; (1.2) Our aim is to identify the parameter q 2 Q ad Q (where Q ad is a closed subset of Q with non empty interior) in the underlying equation is a continuous nonlinear operator with In this setup X, X , Q and Z are Hilbert spaces, and X can be identied with the dual of X. Finally, we assume that e is continuously Frechet-dierentiable on X Q and that the partial derivative e u 2 L(X; X ) is self-adjoint and satises the coercivity condition he u (u; q)v; vi e kvk 2 for some e The above setup is typical for a partial dierential equation of elliptic type, which is also the main type of application we have in mind. We want to mention that the innite- dimensional analysis carried out in the preceding paper [10] was not restricted to elliptic problems, but only assumed well-posedness of the state equation for given parameter. How- ever, since the numerical approximation techniques for elliptic problems dier from the ones for parabolic or hyperbolic problems (cf. e.g. [32] for an overview), one cannot expect a successful unied approach to corresponding parameter identication problems. For this reason we start with an investigation of the elliptic case in this paper, but we want to mention that the numerical identication of parameters in transient equations or even mixed systems of equations is an important and challenging problem for future research. In [10], it has been mentioned that the parameter identication problem in the above setup is an ill-posed inverse problem and we have proposed the following iterative regularization method based on the idea of sequential quadratic programming: Programming Method). Let be a given initial value and let ( k ) k2N be a bounded sequence of positive real numbers. The Levenberg-Marquardt sequential quadratic programming (LMSQP) method consists of the iteration procedure where is the minimizer of the quadratic programming problem2 subject to the linear constraint The iteration procedure is stopped as soon as with appropriately chosen > 1. Due to the results of [10], the LMSQP-method is a convergent regularization method, in particular the quadratic programming problems of the form (1.7), (1.8), which have to be solved in each iteration step, are well-posed. Our aim in this paper is to investigate the numerical approximation of the LMSQP-method by a Galerkin-type approach. We shall show below that this leads to an indenite system in each iteration step, whose solution is an approximation of optimal order to the solution of (1.7), (1.8). Moreover, we show that the reconstructions obtained with the discretized LMSQP-method converge to a solution of the parameter identication problem as the noise level and the discretization size tend to zero, if an appropriate stopping rule is used, which relates the residual to the noise level and some measures for the discretization. Moreover, we shall discuss the solution of the discretized Karush-Kuhn Tucker system, which is an indenite linear system to be solved for the discretized equivalents of state, parameter and Lagrangian variable. The standard approaches to the solution of such discretized problems arising from partial dierential equations are reduced SQP-methods, where state and Lagrangian variable are eliminated a-priorily. We recall the basic properties of the reduced SQP-approach, but we mainly focus on the iterative solution of the whole system with appropriate preconditioning. This promising approach has been employed recently for parameter identication (cf. [20, 26]) and optimal control problems (cf. [2, 3, 4, 5]) with good numerical results, in particular with respect to e-ciency. The paper is organized as follows: in Section 2 we investigate the numerical approximation of the LMSQP-method by a Galerkin-type approach and discuss the well-posedness, stability and approximation properties of the discretized Karush-Kuhn-Tucker (KKT) system; the convergence of the discretized solutions is shown in Section 3. Some further numerical methods and the implementation of the outer iteration, i.e., the SQP-iteration under the assumption that we are able to solve the quadratic optimization problems arising in each step of the LMSQP method, are examined in Section 4. We discuss the correct scaling of variables, globalization strategies as well as a multi-level approach, which leads to a further speed-up of the method. Section 5 is devoted to the inner iteration, i.e., the numerical solution of the discretized KKT-system. Some basic properties of this symmetric indenite problem are studied, as well as its iterative solution with appropriate preconditioning. As a rst application we investigate the identication of a potential in an elliptic boundary-value problem, where we can give quantitative error estimates in terms of the discretization sizes. Some numerical experiments related to this identication problems are presented in Section 7, before we nally conclude and give an outlook to further interesting problems related to this topic in Section 8. Discretization Techniques In the following we investigate the discretization of the LMSQP-method by a Galerkin ap- proach. First of all, we assume that we have discretized data z -; 2 Z Z of the form z is the orthogonal projector onto the nite-dimensional subspace Z . Note that we can give an error estimate for z -; using (1.1) and kR Now let X h X, Q h Q be nite-dimensional subspaces of X and Q, with the corresponding orthogonal projectors . Then we can discretize the LMSQP-Method as follows: be as above and let be a given initial value. Moreover, let ( k ) k2N be a bounded sequence of positive real numbers. The Galerkin Levenberg-Marquardt sequential quadratic programming (GLMSQP) method consists of the iteration procedure where is the minimizer of the quadratic programming problem2 kR (Eu z - )k 2 subject to the linear constraint 2.1 The Discretized Karush-Kuhn-Tucker System 5 Note that the constraint (2.5) can be rewritten in operator form as ~ to be solved for (u; with the notation and P is the adjoint of P h . Under the assumption (1.5), we obtain that for all v 2 X h , i.e., the discrete bilinear form associated with the operator P coercive on X h . This implies by the Lax-Milgram theorem, that (2.6) is uniquely solvable with respect to u for given q 2 Q h . Consequently, in an analogous way to the proof of Proposition 2.1 in [10] we may show the following result on the well-posedness of the quadratic programming problem that has to be solved in each step of Method 2: Proposition 2.1. Let e be continuously Frechet-dierentiable, let (1.5) hold and let k > 0. Then the quadratic programming problem (2.4), (2.5) has a unique solution which is also the only local minimum. 2.1 The Discretized Karush-Kuhn-Tucker System In [10], the Karush-Kuhn-Tucker system for the innite-dimensional version of the LMSQP- method has been derived and analyzed in the framework of linear saddle point problems. Now we will discuss the discretized analogue of this system, namely the rst-order optimality conditions for the quadratic programming problem (2.4), (2.5). The Lagrangian of (2.4), (2.5) is given by for are equal to the identity on X h and Q h , respectively, we can rewrite the Lagrangian as ~ with the operators K k and L k dened by (2.7), (2.8). The KKT-system can now be deduced by computing the partial derivatives of the Lagrangian with respect to u, q and , i.e., solves the linear saddle-point problem@ P ~ ~ ~ ~ As in [10], we dene the symmetric bilinear form a a and the bilinear form b Moreover, we use the right-hand sides Then the KKT-system (2.12) can be interpreted as the Galerkin approximation of an indenite variational problem, i.e., (u; q; is the solution of a In an analogous way to the proof of Theorem 2.3 in [10] we can show that the bilinear form a satises the kernel-ellipticity condition on X h Q h , i.e., there exists a constant a > 0 such that a and that b satises the LBB-condition sup for some b > 0. This implies the following well-posedness result (cf. [7, 8]) for the discretized problem (2.17), (2.18): Theorem 2.2. Let e be continuously Frechet-dierentiable, let (1.5) hold and let k > 0. Then the indenite system (2.17), (2.18) has a unique solution (u; q; which depends continuously on the right-hand sides f k and g k . Since the constants a and b are the same as in the corresponding innite-dimensional conditions in X Q, they are in particular independent of the discrete subspaces X h and This allows to deduce an approximation result for the solutions of (2.17), (2.18) to the solution of the innite-dimensional KKT-System, given in variational form as a k (u; q; with a k given by as above and g k dened by 2.1 The Discretized Karush-Kuhn-Tucker System 7 Theorem 2.3. Suppose that the assumptions of Theorem 2.2 are satised and let denote the unique solution of (2.17), (2.18). Then there exists a constant c > 0 independent of X h and Q h such that where (u; q; ) denotes the unique solution of (2.19), (2.20) and Proof. First, let (~u h ; ~ the solution of (2.17), (2.18) with a k replaced by a k , k . Then Theorem 2.1 in [8] implies the existence of a constant c 1 > 0 (independent of X h and Q h ) such that Moreover, the stable dependence of the solutions of (2.17), (2.18) on the right-hand side implies the existence of c 2 > 0 with sup hg ja sup (R (R and with the triangle inequality we may conclude (2.23). Theorem 2.3 provides an error estimate for the solutions of the discretized saddle-point problem (2.17), (2.18), consisting of two parts corresponding to the numerical approximation in the image space Z and in the pre-image spaces X and Q. An obvious estimate for the rst term is ;h inf ~ which possibly does not lead to a quantitative estimate, since there is no additional information on the smoothness of the noisy data. An alternative estimate is ~ The inmum of can usually be estimated more easily, since the exact data z are smoother due to the fact that ^ u is the solution of the state equation for some parameter ^ q. E.g., if the state equation is of elliptic type with solution ^ is the embedding operator, and R results from a standard nite element discretization on a grid with neness , then we have at least Another important observation is that the last term vanishes if the discrete spaces Z and are equal, which can be achieved in some applications. The second term in (2.23) shows that the Galerkin approximation of the KKT-system is of optimal order in X h Q h X h ; it can be estimated by standard methods for nite element discretizations; quantitative estimates can be obtained using the regularity of the iterates. This part depends of course strongly on the specic application. 3 Convergence Analysis In this section we will analyze the Galerkin LMSQP method with respect to convergence, i.e., the convergence of the reconstruction obtained with an appropriate stopping rule as the noise level and the measure for the discretization neness tend to zero. With identify the innite-dimensional case, i.e., We assume that the discrete subspaces satisfy If we denote by e k and f k the error terms we can rewrite the Karush-Kuhn-Tucker system (2.12) as@ P ~ ~ ~ ~ ~ ~ r kA (3.2) where the r k denotes the remainder As in [10], we require a condition on the nonlinearity, which is summarized in the following: Assumption 1. Let (1.5) be satised and dene the remainder r(u; q) by Then we assume that there exists a constant kEe and that there exists a solution (^u; of the parameter identication problem. If we dene the discretization measures h , h by where kEe and by Then for all (u; holds, where Remark 1. If X h , Z and Q h are standard nite-element spaces on some triangulations, then h , and h can be estimated by the approximation error of these elements. In particular, if the discretization parameter (i.e., the maximal size of a triangle) tends to zero and if the triangulation is regular, one can guarantee that h , and h tend to zero (cf. [32] for further details). For the choice of the stopping index we use a numerical version of (1.9), which involves the discretization measures dened above: For an appropriate choice of , this allows us to prove the following monotonicity property of the iterates: Lemma 3.1. Let Assumption 1 be fullled, let the noise be bounded by (1.1), and assume that In addition, k is chosen such that k k 1 for all k 2 N and that s and the stopping index k is chosen according to the generalized discrepancy principle (3.11) with then and the estimates and hold for all k < k . Proof. Assume that q we deduce the identity The noise bound (1.1) implies that and using the Cauchy-Schwarz inequality together with (3.9) we obtain the estimate (3.16) follows from dividing (3.15) by k and the fact that s By induction we can now show that q k 2 B (q 0 ) for k < k and satisfying (3.14). In an analogous way to the proof of Lemma 3.2 in [10] we can prove the following statement on the niteness of the stopping index k if - > 0: Lemma 3.2. Under the assumptions of Lemma 3.1, the discrepancy principle (3.11) yields a nite stopping index k if and is chosen according to (3.14). One observes that in the above estimates, the term - ;h now plays the same role as the noise level - in the innite-dimensional setup. Therefore it is also possible to prove convergence as in the same way as convergence in the innite-dimensional case for - ! 0 (cf. [10, Theorem 3.5]). Consequently, we do not give the detailed convergence proof, but refer to [10] for further details on the technique of the proof. We only recall the basic assumptions on e and give the nal convergence result, where we use the notation (u -;;h k ) for the iteration according to (2.12) with initial value (P h discretization parameter h and . Assumption 2. In addition to Assumption 1, assume that e is of the form with continuously Frechet-dierentiable (nonlinear) operators A , such that Moreover, we assume that A and N satisfy the nonlinearity conditions kEe and kEe for some positive constants 2 and 3 . Theorem 3.3 (Convergence). Let Assumption 2 and (3.12) be fullled with , suciently small, and let the noise be bounded by (1.1). Moreover, let k be chosen such that and that (3.13) is satised. If the perturbed iteration is stopped with according to the generalized discrepancy principle (3.11) with (uniformly bounded in h and ) satisfying (3.14), then (q -;;h where (u; q) is a solution of (1.3) with Proof. Analogous to the proof of Theorem 3.5 in [10]. 4 Numerical Realization of the SQP-Iteration In the following we want to discuss some numerical methods and variants for the 'outer iteration', i.e., the Galerkin LMSQP algorithm under the assumption that we are able to solve the discretized KKT-system numerically. The 'inner iteration', namely the numerical solution of the indenite system (2.12) will be investigated in Section 5. 4.1 Scaling of State Variable, Parameter and Lagrangian Variable The performance of an iteration algorithm often depends crucially on the way the problem is formulated. Scaling is a well-known technique for reformulating an optimization problem whose main objective is twofold: on the one hand all the variables should be of similar magnitude, on the other hand also the value of the derivatives should all be of similar size. In unconstrained optimization, a problem should be rescaled in such a way, that changes of the iterate in one direction do not result in by far larger changes of the value of the objective than changes in another direction. In constrained optimization the above statements are also true for each constraint. Additionally the set of constraints should be well balanced with respect to each other such that each constraint has equal weight. Furthermore the set of constraints should be balanced with respect to the objective. As scaling is of high practical importance for any optimization problem, many aspects can be found in monographs on optimization (cf. e.g. [19, 30]). We want to consider only the last aspect in this context, i.e., the scaling of the state constraint with respect to the objective which is also of high importance for achieving fast convergence of the outer iteration. For the inner iteration, the aspect of scaling can be included in the construction of a good preconditioner. The outer iteration of an SQP method tries to attain two goals at the same time: feasibility of the iterate with respect to the state constraint and optimality of the iterate with respect to the objective. One aspect dominating the other results usually in bad convergence properties: If the feasibility aspect dominates, only very small changes of the iterate are possible in order to ensure "almost" feasibility. If the optimality aspect dominates, any violation of the state constraint is reduced too slowly. For the LMSQP method in the form of (2.4),(2.5) it turned out that in many situations the feasibility aspect is strongly dominating. Using line search methods for globalization (see also Subsection 4.2) this results usually in step lengths much smaller than one. Replacing the state constraint by a preconditioned state constraint leads to a better balanced formulation and to much faster convergence. Furthermore a step length parameter equal to one is accepted in almost all steps. Another aspect of this kind of rescaling is treated in Subsection 6.2. 4.2 Globalization Strategies The LMSQP method is a variant of Newton's method and therefore only locally convergent (see also the analysis in Section 3). For this reason, globalization strategies, such as trust region methods or line search strategies (which are the two most popular classes of globalization techniques in optimization), are needed. The basic idea of trust region methods is to add an additional constraint on the maximal increment to the quadratic optimization problem for the correction step of the current iterate, i.e. instead of (2.4), (2.5) one would solve (2.4), (2.5) and k(u u k ; q q k )k k with k chosen appropriately. Trust region methods have been successfully applied to PDE constrained optimization problems (see e.g. [14, 38]), often using a reduced SQP approach. We want to mention that a similar eect as with trust-region methods could be reached in principle by controlling the penalty parameter k , which also restricts the step size and produced good numerical results (see Example 7.1). For a comprehensive overview of trust region methods we refer to Conn et. al. [13]. In the code used for two-dimensional problems (cf. [10] and Example 7.2), we use a line search algorithm for globalization. In contrast to trust region methods, the calculation of the increment is split into two phases: rst of all, a search direction is determined, and secondly the estimation of a step length parameter indicating how far into the search direction one should go. For the computation of the search direction we solve the optimization problem and (2.5). In order to determine the step length we cannot use the objective itself as a criterion (as in unconstrained optimization), but have to use a merit function which balances the minimization of the objective with the feasibility with respect to the state constraint. Applied to a discretized optimization problem of the form ~ subject to an equation constraint of the form possible choices are the l 1 -merit function 4.3 Nested Multi-Level Optimization Techniques 13 and its variants, and the augmented Lagrangian where is an estimate of the Lagrangian variable corresponding to the discretized equation constraint. Both merit functions are exact in the sense that for su-ciently large, minimizers of the original constrained optimization problem also minimize the merit function. A crucial property in the design of a merit function is that it should accept step length one close to a solution in order to preserve the quadratic convergence of the SQP method. The augmented Lagrangian works well, as long as the estimate for the Lagrangian multiplier is accurate enough, whereas the l 1 -merit function sometimes suers from the so-called "Marathos-eect", i.e. it does not accept unit step length and therefore causes a slow-down of the convergence. A strategy to overcome this di-culty using a second order correction can be found in [30]; nevertheless, it performed very well in our numerical experiments (see Example 7.2). 4.3 Nested Multi-Level Optimization Techniques Important tools for the e-cient numerical approximation of innite-dimensional optimization problems are multi-level optimization methods. In the nested multi-level setup, one starts the optimization procedure at a coarse level X , where the iteration procedure can be carried out e-ciently. If an appropriate stopping rule is satised, one interpolates the state and parameter obtained in this way to a ner level X h 2 (for serving now as a starting value on this level. This procedure is repeated until the nest level is reached. Usually, nested space are used in this approach, i.e., X h 1 (for which leads to simple interpolation operators. Since one cannot choose the discretization of the data arbitrarily in general , we consider only the case of xed here, but a multi-level approach in can be realized in an analogous way, if necessary. Nested multi-level methods outperform standard discretization techniques in many cases (cf. e.g. [21, 22, 29]); usually a considerable number of iterations is needed on the coarse level only , where the numerical eort per iteration is very low. On the nest levels, the stopping rule is often satised already after one iteration step and so the overall eort is less than for a direct discretization on the nest level. For the Galerkin LMSQP method, we can formulate a multi-level algorithm as follows: Algorithm 4.1 (Nested Multi-Level Galerkin LMSQP). Given a decreasing sequence '=1;:::;L with nested spaces X h ' X h '+1 , Q h ' Q h '+1 (e.g. h non-increasing sequence ' satisfying (3.14), the nested multi-level Galerkin LMSQP method consists of the following iterative procedure: 1. 2. Perform the Galerkin LMSQP method until the stopping criterion (3.11) is satised with stopping index k ('). 3. If stop the iteration, else prolongate the iterate (u ' k ) to the ner level which results in a new starting value (u '+1 and go to step 2. The analysis in Section 3 shows that for ' , the estimate holds, where ' is the error corresponding to the interpolation of the iterates from level ' 1 to level ', i.e., kR Ee ' Z kf ' This monotonicity estimate corresponds very well to the intuition that only few iterations are needed on the ne levels, in particular if ' k is decreasing, which leads to kR Ee ' Z ' For a ne level with small , we can expect that and the second term ' can be expected to be negligible. I.e., the stopping rule at level ' is probably satised with k Under typical conditions, where X h ' and Q h ' correspond to standard nite-element spaces on dierent renement levels of an initial triangulation of a domain one can show that at least consequently ch ch 1 for some constant c 2 R+ , where Together with the above estimate one can show with a standard proof technique that the converges to a solution (u; q) of the parameter identication problem for 5 Numerical Solution of the KKT-System In the following we will discuss the numerical solution of the discretized KKT-system (2.12) for xed iteration number k. We have seen above that the Galerkin-type approximation (2.12) of the original KKT-system is stable and convergent, now we discuss some of its structural properties, which are important for the application of iterative solution methods and for the construction of preconditioners. 5.1 The System Matrix M 15 Choosing bases of the nite-dimensional subspaces X h and Q h , we may represent via with coordinate vectors V; In order to transform (2.12) into a linear system for V , S and , we dene the matrices and the vectors This allows us to rewrite the discretized KKT-system (with penalty parameter respectively as where M is the matrix in (5.6) and The structural properties of M and its sub-matrices will be examined in the following section. 5.1 The System Matrix M Due to the well-posedness result on the discretized KKT-system (2.12) (cf. Theorem 2.2), we may conclude that the system matrix M is regular. In order to obtain further insight into the structure of M , we investigate the properties of the sub-matrices G, H, K and L: Proposition 5.1. The matrices K 2 R mm and H 2 R nn are symmetric positive denite, and the matrix G 2 R mm is symmetric positive semi-denite. If, in addition, the operator is injective on X h , then G is regular, too. Proof. Let u and q be as in (5.2), then there exist constants c 1 (h) and c 2 (h) such that where j:j denotes the euclidean norm in R n and R m , respectively. Thus, we have and Moreover, the identity implies that G is positive semi-denite and regular under the assumption that E is injective on X h . The symmetry of the matrices G, H and K can be veried in a similar way, using the symmetry of scalar products and the self-adjointness of the operator K k . The matrix L 2 R mn is di-cult to analyze, it is neither symmetric nor regular in general (in particular if n 6= m). However, some fundamental properties of M (such as its regularity) rely rather on G, H and K than on L. Moreover, the classical splitting of a symmetric saddle-point problem as@ G 0 K T I I where H := H and C is the Schur-complement is possible if we only know that G and H are regular. In particular, we may conclude that M has n +m positive and m negative eigenvalues. 5.2 Reduced SQP Approaches The basic idea of reduced SQP-methods is the a-priori elimination of the equality constraint, which can be written in matrix form as which is equivalent to an elimination of V and in (5.6). Due to Proposition 5.1, K is a regular, symmetric matrix and thus, we may compute which yields after some calculations the n n-system with The reduced SQP-approach seems of particular interest if n m, which is a frequently used discretization strategy for parameter identication and optimal control problems (cf. e.g. [35, 36, 37]). The original matrix M is an indenite matrix of size (2m+n) (2m+n), while 5.3 Simultaneous Solution of the KKT-System 17 the reduced system matrix M r in (5.12) is of size n n. However, M r is not a sparse matrix even if all the sub-matrices of M are sparse, since it involves the inverse of K. Moreover, the evaluation of M r is more expensive than the evaluation of the original system matrix M , since it involves the solution of two systems of the form with dierent right-hand sides g, while for the evaluation of M only direct evaluations of K are needed, which are very cheap for typical nite element discretization of the state constraint. In practice, one usually tries to compensate this disadvantage of reduced SQP-methods by using a Broyden-type update for the reduced system matrix instead of the exact matrix M r , which leads to e-cient optimization algorithms for small n. 5.3 Simultaneous Solution of the KKT-System Recently, the simultaneous solution of KKT-systems by iterative methods has been investi- gated, in particular in connection with optimal control problems (cf. [2, 4, 5, 20]). Compared to the reduced SQP-approach, a simultaneous solution strategy has the obvious advantage that the allocation and evaluation of the system matrix M is much cheaper than of M r . The pay-o is that M is indenite and larger than M r , which might cause additional eort. How- ever, the main eort in the reduced SQP-approach is related to the evaluation or assembly of the system matrix M r , respectively, and therefore a simultaneous solution of the KKT-system can result in a tremendous speed-up of the SQP-method, in particular for ne discretizations. At a rst glance, it seems rather straight-forward to solve (5.7) by a standard iterative method for indenite systems such as inexact Uzawa methods (cf. [6, 15]) or Krylov-subspace methods such as GMRES (cf. [34]), MINRES (cf. [31]) and QMR (cf. [18]). However, in the case of large-scale problems, we have to expect a large condition number (note that is usually small and that M is singular for and a complicated eigenvalue pattern of the matrix M , which might cause iterative methods to diverge or to need a high number of iterations. Therefore, an appropriate preconditioning technique seems necessary for any of the methods. We do not go into details here, but refer to the forthcoming paper [11] for a discussion of preconditioners. In the following we distinguish two types of solvers that seem appropriate for the solution of the indenite system (5.7) and discuss their basic properties with respect to the special structure of M . Inexact Uzawa Iterations Inexact Uzawa methods and similar iteration procedures have been developed for the solution of the classical Stokes system and similar problems (cf. [32] for an overview). The classical Uzawa method is just a gradient method for the dual of the corresponding Lagrange functional, the inexact Uzawa method can be interpreted as a preconditioned version (cf. [32]). Following the exposition by Zulehner [39], we can write an inexact Uzawa method for a system of the form (5.6) as A followed by A is a preconditioner for the diagonal matrix A := C is a preconditioner for the Schur-complement C dened by (5.8). In terms of (5.7) we can write the inexact Uzawa iteration as M is a preconditioner for the system matrix, given by . A convergence analysis of this method is available only in the case when A is a regular matrix (cf. [6, 39]), which means that we have to assume that G is regular. The latter is true e.g. if the data z represent distributed data for the state, i.e., E is an embedding operator. In this case, the structure of A is rather simple and it is not a di-cult task to construct a preconditioner, even exact preconditioning seems possible (note that G is just a mass matrix for a typical nite element discretization). Since the matrices G and H do not change during the SQP-iteration we may even compute decompositions in a preprocessing step. The construction of a preconditioner for the Schur-complement C is more di-cult and must take into account the specic nature of the underlying state equation. Krylov-Subspace Methods The Krylov-subspace methods GMRES and QMR are variants of the CG-algorithm that are applicable to indenite problems, too. The basic idea of such methods is a defect minimization in the Krylov-subspace generated by X 1 , in the k-th iteration step. Since preconditioned CG-methods are probably the most successful class of iteration methods for positive denite systems, such methods seem very attractive also in the indenite case, although additional di-culties may arise (cf. e.g. [34]). The convergence analysis in [34] and [18] shows that the error bounds obtained for both methods are essentially the same, and mainly dependent on the eigenvalue distribution and the condition number of the system matrix M . Therefore, appropriate preconditioning is again of high importance, in this case also with the possibility that G is singular. We refer to [11] for a detailed discussion of this problem. As a rst application we investigate the identication of the potential q in the elliptic boundary value problem in @ from a state observation in L 2 which is a well-studied problem in literature (cf. e.g. [33]). In [10], it has been shown that in the setup (d denotes the space dimension) the operators satisfy all assumptions needed for the convergence analysis of the LMSQP-method. Now we shall study a concrete nite-element discretization of the KKT-system and the derivation of estimates for the numerical errors , h and h . 6.1 Error Estimates for the Discretized KKT-System In this case we can write the whole KKT-system in classical form as in in in again with homogenous Dirichlet boundary conditions upon v and on @ where L d is a dimension-dependent dierential operator of order 2d corresponding the norm in H 7 e.g., we have supplemented by homogenous boundary conditions up to order d 1. If f 2 L and 3 a standard elliptic regularity argument shows that ^ 0(for all k 2 N. In the same way we can show that k 2 H and s k 2 H 7 This additional regularity can be employed to derive standard error estimates for nite-element discretizations of the KKT-system (2.12). If we use piecewise linear nite elements on regular triangulations T and T h for the discretization spaces Z and X h , where and h represent the neness of the grids, then a classical approximation result for nite elements (cf. [32, p.96]) implies that Of course, one could also use piecewise constant elements on T , which would yield However, in practical applications a higher-order approximation in is often desirable, since can be signicantly larger than a reasonable choice of h. A canonical approximation of the parameter q is a nite element space of order greater or equal d on a regular triangulation T ~ h assumptions on the exact solution ^ q one can obtain quantitative estimates for h in terms ~ h. At a rst glance it seems surprising that one needs a-priori assumptions on the parameter, but not on the state in order to derive error estimates. However, due to the ill-posedness of the identication problem with respect to the parameter q, such a-priori knowledge seems to be necessary. The approximation of the state corresponds rather to the approximation of the underlying elliptic state equation, which is well-posed with respect to u and yields further regularity. We nally want to mention that according to the theory developed above, one could choose T ~ h independent of T h , but this would cause unnecessary complications in the implementation of the method. We note that alternatively one can use the space for d 3, which yields An appropriate discretization strategy is e.g. to choose Q h as the space of piecewise constant elements on an underlying grid T ~ h . The advantage of this approach is that elements of order greater than one, which are necessary for (d 2), can be avoided. 6.2 Structure of the System Matrix For the potential identication problem, some parts of the system matrix M in (5.6) are well- understood. First of all, G is an L 2 -mass matrix and it is positive denite if the triangulations T and T h coincide, which we assume in the following. The eigenvalues of G are then all of order h d . The matrix H is the stiness matrix for the dierential operator L d , with minimal eigenvalue of order h d and maximal eigenvalue of order h d . The matrix K is the sum of a stiness matrix for the Laplacian and a weighted mass matrix (with weight q k in the L 2 -scalar product), where one can expect the rst part in this sum to be dominating. Thus, the stiness matrix ^ K for the Laplacian will be a good preconditioner for K. The maximal and minimal eigenvalues of K and K are of order h d 2 and h d , respectively. The remaining part in the system matrix, namely the matrix L, is di-cult to understand, since its elements are weighted L 2 -scalar products of basis functions of dierent nite element spaces. However, the spectral norm of L can be estimated, it is of order ~ h d . The construction of preconditioners for G and H is well-investigated, even exact preconditioning seems to be applicable. For K it seems reasonable to use a preconditioner ^ K for the Laplacian, e.g. a multi-grid preconditioner. With preconditioning for K, the system matrix can be transformed to ~ with the corresponding Schur-complement ~ K is an appropriate preconditioner for K, then we can estimate the minimal eigenvalue by min ( ~ C) min and the maximal eigenvalue by O Hence, the condition number of ~ C is independent of h, but only depends on and ~ h h . One observes that the condition number is decreasing as ~ h tends to h from above (note that usually ~ h h). For the Uzawa iteration, one can choose the preconditioner ^ C in this case as a multiple of ^ K 1 or even of G 1 . If ~ h h, the Uzawa iteration seems not to be optimal, in this case one can apply either a reduced SQP-approach or use Krylov-subspace methods with dierent preconditioning strategies. For the details on the latter we refer to [11]. 7 Numerical Experiments In order to test our theoretical results, we numerically solve some model problems, which have already been investigated with respect to the convergence behavior of the LMSQP-method in [10]. Example 7.1. Our rst example is the identication of the potential q in (6.1), (6.2) from a state observation u 2 L 2(with The exact potential is given by which is an element of 3 This problem was implemented in the software-system MATLAB. The data are generated by solving the state equation on a ne grid and subsequent interpolation to a coarser grid; the noise is an additive high-frequency perturbation. We used uniform grids with m nodes for the discretization of the state u and the Lagrange-parameter and n nodes for the parameter q, i.e., . The parameters k are chosen according to which lead to convergence of the method even for starting value q 0. The KKT-system (5.6) is solved by the QMR method, using an Uzawa-type preconditioner as described in Section 6.2, with ^ K a preconditioner for the Laplacian and ^ . The convergence results for the overall LMSQP-method have been shown in [10] and compared to a Levenberg-Marquardt method following the feasible path. It turned out that both methods lead to almost the same iteration sequence q k . In particular, the number of iterations needed until the stopping rule is satised, is the same for both methods. Now we compare the numerical e-ciency of the LMSQP-method with feasible path approaches, namely the Levenberg-Marquardt method (LM) on the feasible path (with the same Galerkin discretization as for LMSQP and solution of the Gauss-Newton system by a preconditioned CG-method) and a Broyden-type variant of the Levenberg-Marquardt method (cf. [23] for further details). 22 7 NUMERICAL EXPERIMENTS Table 1: CPU-time (in seconds) needed for the LMSQP-method, the LM-method and a Broyden-type variant of the LM-method. For this sake we choose dierent discretization levels (xed during the iteration) and measure the CPU-time needed for the LMSQP-method, until the stopping rule is satised (for xed noise level -). From the results shown in Table 1 one observes that the LMSQP-method with simultaneous solution of the KKT-system outperforms the feasible-path approaches for all dierent discretizations. Since the LMSQP and the LM-method need the same number of outer iterations, the dierence in the numerical eort is caused by the fact that the eort for the evaluation of the system matrix in the LM-method is signicantly higher than evaluation and preconditioning of the system matrix in the simultaneous LMSQP-method. Obviously, the gain in the numerical eort for the evaluation of the system matrix increases with the number of discretization points, which explains the extremely large CPU-time for the LM- method at the nest discretization level is much faster than the LM-method, which is again caused by the fact that the evaluation of the system matrix can be carried out e-ciently. However, the number of iterations needed for the Broyden-type variant is much larger than for the other two methods, which use the full information about the derivatives. Finally, we investigate the spectral condition of the system matrix M as well as of the matrix ~ M dened by (6.13), where we use a preconditioner for the Laplacian as ^ K. >From the left picture in Figure 1, which shows the condition number as a function of the discretization size h (in logarithmic scale) for xed one observes that the condition number of M grows quadratically with h 1 , while the condition number of ~ M is much smaller and almost independent of h. The second part of Figure 1 shows a plot of the condition numbers vs. the parameter in doubly logarithmic scale, from which it seems that the growth of the condition number as ! 0 is slower for ~ M than for the original matrix M . In both cases, the condition number seems to be a convex function of , which has a unique minimum at some . However, this value is rather large and values of that are signicantly larger than are not of interest for our purpose, since they would cause a tremendous slow-down of the outer iteration. Therefore we can focus our attention to the case < , where the condition number increases in a monotonically with 1 . Example 7.2. Our second numerical example is the identication of the conductivity q 2 log(cond(M)) Condition Number vs. Discretization Size Original Preconditioned State log b log(cond(M)) Condition Number vs. b Original Preconditioned State Figure 1: Plot of the spectral condition of the matrix M vs. the discretization size h (in logarithmic scale, left) and vs. the parameter (in doubly logarithmic scale, right). The solid line shows the condition number of the original matrix M , the dashed line of the matrix ~ M with preconditioned state equation. in in on @ from a state observation u 2 L 2 The domain is a ball in R 2 with missing rst quadrant, i.e., in radial coordinates The exact parameter to be reconstructed is ^ q 1, the right-hand side in (7.1) is given by r with The corresponding solution of the state equation is ^ 3r). The data are generated using the exact solution ^ u perturbed by uniformly distributed random noise. For the discretization we used triangular nite elements with piecewise quadratic shape functions for the state u and the Lagrange parameter and piecewise constant shape functions for the parameter q. The results were calculated using the nite element code FEPP [27], developed at the Department for Analysis and Computational Mathematics of the University of Linz. We want to mention that this identication problem is quite challenging not only due to the complicated geometry, but also due to the fact that q is not identiable along a level line in the interior, where u attains an extremum. This does not destroy the theoretical identiability results, because it is a set of Lebesgue-measure zero, but it can be expected to create numerical di-culties. Results for exact data can be found in Table 2. The good performance of the method with respect to both, CPU time and number of outer iterations can be observed clearly. Especially for problems with ne discretizations of the parameter q, this method can still be realized e-ciently, while classical approaches do not yield results in reasonable time. A plot Level dim q dim u avg QMR it SQP it time Table 2: CPU-time and number of inner (QMR) and outer (SQP) iterations for exact data Figure 2: Parameter distribution for exact data at level 4 of the parameter q can be found in Figure 2, from which one observes that the parameter is reconstructed very well except in a neighborhood of the level curve 0g. Additional speed-up can be gained using a multi-level approach as described in Subsection 4.3. We used nested spaces for q and u by subdividing each triangular element into four smaller elements, when rening the mesh. Table 3 presents results for this approach. It can be seen that on ne discretization levels one SQP step is su-cient for fullling the stopping criterion, which corresponds very well to the theoretical predictions made in Section 4.3. A comparison of the results to the ones in Table 2 shows that for xed discretization level, the solution of the identication problem on level 5 is only slightly faster than the identication of q on level 6 (with about the fourfold number of parameters) using a multi-level approach. A plot of Level dim q dim u avg QMR it SQP it time acc. time Table 3: CPU-time per level, accumulated time and number of inner (QMR) and outer (SQP) iterations for exact data using a nested multi-level approach Figure 3: Parameter distribution for exact data at level 4 using a nested multi-level approach the parameter can be found in Figure 3. Here the approximation of the parameter in the area where it can not identied is by far better than in the classical approach using only one discretization level (compare Figure 2). 8 Conclusions and Outlook We have developed a framework for Galerkin-type approximations of the LMSQP-method for parameter identication problems in elliptic partial dierential equations and we have discussed the implementation of the Galerkin LMSQP-method with iterative solution of the KKT-system. The numerical results show that the resulting iteration method clearly outperforms state-of-the-art methods for iterative regularization and provides a tool for the e-cient solution of identication problems with ne discretizations. Moreover, we have developed a multi-level version of the Galerkin-LMSQP method, which yields a further speed-up. The crucial point for the possibility to obtain an e-cient implementation of the LMSQP- method is the preconditioning of the KKT-system, which is then solved iteratively as an indenite problem in the product space for state, parameter and Lagrangian variable. The construction of such preconditioners is not a simple task and has not been discussed in detail in the present paper, but will be investigated in [11], where dierent preconditioning techniques will be compared. Other numerical aspects to be investigated in future research are adaptive discretization strategies and fast parallel solvers based on domain-decomposition techniques. The adaptive discretization of optimal control problems, which is a closely related subject, has been discussed by Becker et al. [3]; possibly the ideas of this work can be carried over to identi- cation problems, too. The parallel solution of optimal control problems has been investigated by Lions and Pironneau [28] in the case of quadratic problems; recently Biros and Ghattas [4, 5] performed a numerical study of a parallel solver with an SQP-method for the outer and preconditioned Krylov-subspace methods for the inner iteration. Many of their ideas seem to be applicable also for parameter identication problems that are solved with the LMSQP- method, which rises the hope that e-cient parallel versions of the LMSQP-method can be designed also for large-scale identication problems such as impedance tomography. Finally, we want to recall that the framework of this problem does not apply to transient problems of parabolic or hyperbolic type. Since numerical methods for dierent types of partial dierential equations have many type-specic features in general, it is not surprising that also the numerical treatment of parameter identication problems should depend on the type of the underlying state equation. However, it seems possible to construct e-cient and convergent discretized methods at least in the case of parabolic equations, which is an important task for future research. Acknowledgments The authors thank Dr. Walter Zulehner (University of Linz) and Dr. Joachim Schoberl (cur- rently Texas A & M University) for useful and stimulating discussions on the preconditioning of the indenite system (5.6). --R Estimation Techniques for Distributed Parameter Systems Preconditioners for Karush-Kuhn-Tucker matrices arising in the optimal control of distributed systems Parallel Lagrange-Newton-Krylov-Schur methods for PDE-constrai- ned optimization problems Parallel Lagrange-Newton-Krylov-Schur methods for PDE-constrai- ned optimization problems Analysis of the inexact Uzawa algorithm for saddle point problems On the existence Mixed and Hybrid Finite Element Methods Iterative regularization of a parameter identi Inverse Problems in Partial Di Inexact and preconditioned Uzawa algorithms for saddle point problems Inverse Problems in Di Convergence rate results for iterative methods for solving non-linear ill-posed problems QMR: a quasi-minimal residual method for non-Hermitian linear systems Preconditioned all-at-once methods for large The numerical solution of a control problem governed by a phase On Broyden's method for the regularization of nonlinear ill-posed prob- lems A projection-regularized Newton method for nonlinear ill-posed problems with application to parameter identi cation problems with nite element discretization Sur le controle parallele des systemes distribues Solution of sparse inde Numerical approximation of partial di Determination of a source term in the linear di GMRES: a generalized minimal residual algorithm for solving non-symmetric linear systems Control applications of reduced SQP methods Partially reduced SQP methods for large-scale nonlinear optimization problems Solving discretized optimization problems by partially reduced SQP methods Global convergence of trust-region interior-point algorithms for in nite-dimensional nonconvex minimization subject to pointwise bounds Analysis of iterative methods for saddle point problems: a uni --TR
parameter identification;iterative regularization;galerkin methods;sequential quadratic programming;indefinite systems
589719
Mappings for conflict-free access of paths in bidimensional arrays, circular lists, and complete trees.
Since the divergence between the processor speed and the memory access rate is progressively increasing, an efficient partition of the main memory into multibanks is useful to improve the overall system performance. The effectiveness of the multibank partition can be degraded by memory conflicts, that occur when there are many references to the same memory bank while accessing the same memory pattern. Therefore, mapping schemes are needed to distribute data in such a way that data can be retrieved via regular patterns without conflicts. In this paper, the problem of conflict-free access of arbitrary paths in bidimensional arrays, circular lists and complete trees is considered for the first time and reduced to variants of graph-coloring problems. Balanced and fast mappings are proposed which require an optimal number of colors (i.e., memory banks). The solution for bidimensional arrays is based on a particular Latin Square. The functions that map an array node or a circular list node to a memory bank can be calculated in constant time. As for complete trees, the mapping of a tree node to a memory bank takes time that grows logarithmically with the number of nodes of the tree. The problem solved here has further application in minimizing the number of frequencies assigned to the stations of a wireless network so as to avoid interference.
Introduction In recent years, the traditional divergence between the processor speed and the memory access rate is progressively increasing. Thus, an efficient organization of the main memory is important to achieve high-speed computations. For this purpose, the main memory can be equipped with cache memories which have about the same cycle time as the processors - or can be partitioned into multibanks. Since the cost of the cache memory is high and its size is limited, the multibank partition has mostly been adopted, especially in shared-memory multiprocessors [3]. However, the effectiveness of such a memory partition can be limited by memory conflicts, that occur when there are many references to the same memory bank while accessing the same memory pattern. To exploit to the fullest extent the performance of the multibank partition, mapping schemes can be employed that avoid or minimize the memory conflicts [15]. Since it is hard to find universal mappings - mappings that minimize conflicts for arbitrary memory access patterns - several specialized mappings, designed for accessing regular patterns in specific data structures, have been proposed in the literature (see [12, 2] for a complete list of references). In particular, for bidimensional arrays, Budnik and Kuck [7], Balakrishnan et al. [4], Kim and Prasanna [12], and Das and Sarkar [8] studied mappings that provide conflict-free access to rows, columns, positive and negative diagonals, subarrays, and distributed subarrays. The techniques used range from Latin squares to Perfect Latin squares, from linear mappings to quasi-groups [11]. Subse- quently, mappings for other data structures like complete trees and binomial trees have been devised. In particular, mappings that provide conflict-free access to complete subtrees, root-to-leaves paths, sub- levels, and composite patterns obtained by their combination, have been investigated in [8, 9, 1, 10, 14]. The mapping schemes proposed in those papers are optimal, i.e., they use as few memory modules as possible; balanced, i.e., the nodes of data structures are distributed as evenly as possible among the banks; fast, i.e., the bank address to which a node is assigned is computed quickly with no knowledge of the entire structure mapping; and flexible, i.e., they can be used for templates of different size. In the present paper, optimal, balanced and fast mappings are designed for conflict-free access of paths in bidimensional arrays, circular lists, and complete trees. With respect to the above mentioned papers, paths in bidimensional arrays and circular lists are dealt with for the first time. Moreover, access to any (not only to root-to-leaves) paths in complete trees is provided. The remainder of this paper is organized as follows. In Section 2, the conflict-free access problem is formally stated. In Section 3, the problem of accessing paths in bidimensional arrays is solved. The proposed solution is a variant of a graph-coloring, which requires an optimal number of colors and is achieved using a combinatorial object similar to a Latin Square. As a byproduct, the memory bank to which an array node is assigned is computed in constant time. In Section 4, the problem of accessing paths in circular lists is optimally solved and the function that maps a circular list node to a memory bank can be calculated in constant time. In Section 5, the same problem on complete trees is also optimally solved via a variant of a graph-coloring problem. The time needed to assign a tree node to a memory bank grows logarithmically with the number of nodes of the tree. Conclusions are offered in Section 6. Conflict-Free Access When storing a data structure D, represented in general by a graph, on a memory system consisting of N memory banks, a desirable issue is to map any subset of N arbitrary nodes of D to all the N different banks. This problem can be viewed as a coloring problem where the distribution of nodes of D among the banks is done by coloring the nodes with a color from the set f0; it is hard to solve the problem in general, access of regular patterns, called templates, in special data structures - like bidimensional arrays, circular lists, and complete trees - are considered hereafter. A template T is a connected subgraph of D. The occurrences fT of T in D are the template instances. For example, if D is a complete binary tree, then a path of length k can be a template, and all the paths of length k in D are the template instances. After coloring D, a conflict occurs if two nodes of a template instance are assigned to the same memory bank, i.e., they get the same color. An access to a template instance T i results in c conflicts if belong to the same memory bank. Given a memory system with N banks and a template T , the goal is to find a memory mapping that colors the nodes of D in such a way that the number of conflicts for accessing any instance of T is minimal. In fact, the cost for T i colored according to U , CostU (D; defined as the number of conflicts for accessing T i . The template instance of T with the highest cost determines the overall cost of the mapping U . That is, A mapping U is conflict-free for T if Among desirable properties for a conflict-free mapping, a mapping should be balanced, fast, and optimal. A mapping U is termed balanced if it evenly distributes the nodes of the data structure among the N memory banks. For a balanced mapping, the memory load is almost the same in all the banks. A mapping U will be called fast if the color of each node can be computed quickly (possibly in constant time) without knowledge of the coloring of the entire data structure. Among all possible conflict-free mappings for a given template of a data structure, the more interesting ones are those that use the minimum possible number of memory banks. These mappings are called optimal. It is worth to note that not only the template size but also the overlapping of template instances in the data structure determine a lower bound on the number of memory banks necessary to guarantee a conflict-free access scheme. This fact will be more convincing by the argument below for accessing paths in D. E) be the graph representing the data structure D. The template P k is a path of length k in D. The template instance P k [x; y] is the path of length k between two vertices x and y in V , that is, the sequence of vertices such that (v h ; v h+1 The conflicts can be eliminated on P k [x; y] if are assigned to all different memory banks. The conflict-free access to P k can be reduced to a classical coloring problem on the associated graph GDP k obtained as follows. The vertex set of GDP k is the same as the vertex set of GD , while the edge (r; s) belongs to the edge set of GDP k iff the distance d rs between the vertices r and s in GD satisfies d rs - k, where the distance is the length of the shortest path between r and s. Now, colors must be assigned to the vertices of GDP k so that every pair of vertices connected by an edge is assigned a couple of different colors and the minimum number of colors is used. Hence, the role of maximum clique in GDP k is apparent for deriving lower bounds on the conflict-free access on paths. A clique K for GDP k is a subset of the vertices of GDP k such that for each pair of vertices in K there is an edge. By well-known graph theoretical results, a clique of size n in the associated graph GDP k implies that at least n different colors are needed to color GDP k . In other words, the size of the largest clique in GDP k is a lower bound for the number of memory banks required to access paths of length k in D without conflicts. On the other hand, the conflict-free access to P k on GD is equivalent to color the nodes of GD in such a way that any two nodes which are at distance k or less apart have assigned different colors. Unfortunately, this latter coloring problem is NP-complete [13] for general graphs. This justifies the investigation either for good heuristics for general graphs or optimal algorithms for special classes of graphs. In the next three sections, optimal mappings for bidimensional arrays, circular lists and complete binary trees will be derived for conflict-free accessing P k . Accessing Paths in Bidimensional Arrays Let a bidimensional array A be the data structure D to be mapped into the multibank memory system. An array r \Theta c has r rows and c columns, indexed respectively from 0 to r \Gamma 1 (from top to bottom) and from 0 to c \Gamma 1 (from left to right), with r and c both greater than 1. The graph E) representing A is a mesh, whose vertices correspond to the elements of A and whose arcs correspond to any pair of adjacent elements of A on the same row or on the same column. For the sake of simplicity, A will be used instead of GA since there is no ambiguity. Thus, a generic node x of A will be denoted by its row index and j is its column index. least l (k+1) 2m memory banks are required for conflict-free accessing P k in A. Proof Consider a generic node of A, and its opposite node at distance k on the same column, i.e., All the nodes of A at distance k or less from both x and y are mutually at distance k or less, as shown in Figure 1. Therefore, in the associated graph GAP k , they form a clique, and they must be assigned to different colors. In details, such a clique, denoted as KA (x; k), is defined as follows: KA ni l km l km l km l km l kmo Summing up over t, the size of the clique results to be ''- k- Hence, at least l (k+1) 2m colors are required. 2 (b) x x y y Figure 1: A subset KA (x; of nodes of A that forms a clique in GAP k Below, a conflict-free mapping is given to color all the nodes of an array A using as few colors as in Lemma 1. Therefore, the mapping is optimal. From now on, the color assigned to node x is denoted by fl(x). Algorithm Array-Coloring (A; k); l (k+1) 2m and even if k is odd ffl Assign to each node x = (i; A the color Intuitively, the above algorithm first covers A with a tessellation of basic sub-arrays of size M \Theta M . Each basic sub-array S is colored in a Latin Square fashion as follows: ffl the colors in the first row of S appear from left-to-right in the sequence 0; ffl the color sequence for a generic row is obtained from the sequence at the previous row by a \Delta left-cyclic shift. For the coloring of A, decomposed into 6 basic sub-arrays of size M \Theta M , is illustrated in Figure 2. Theorem 1 The Array-Coloring mapping is optimal, fast, and balanced. Proof To prove optimality, it must be shown that the mapping is conflict-free and that the minimum number of colors is used. Figure 2: An array A of size 16 \Theta 24 with a tessellation of 6 sub-arrays of size 8 \Theta 8 colored by the Array-Coloring algorithm to conflict-free access P 3 . Consider a generic node x = (g; f) of A and the associated clique KA (x; k), defined in Lemma 1. In order to prove that the mapping is conflict-free, one only needs to show that all the nodes of KA (x; k), which are mutually at distance no more than k, are assigned by the Array-Coloring algorithm to different colors. Formally, consider an arbitrary pair of nodes belonging to KA (x; k), such that 0, the roles of w and z could be swapped). Then the mapping is conflict-free if the Array-Coloring algorithm guarantees that the colors fl(w) and fl(z) are different. Moreover, let oe(w; z) be the difference between the two colors assigned to w and z. Then, the mapping is conflict-free if the following two conditions simultaneously hold: (1) In order to show that the conditions in (1) hold for any pair of nodes of KA (x; k), the two cases k even and k odd must be distinguished. When k is even, one has that l (k+1) 2m that oe(w; z) the congruence oe(w; z) 6j 0 mod M is equivalent to oe(w; z) 6= 0 and oe(w; z) 6= M . Clearly, oe(w; which is verified only if either z = w or j' \Gamma jj is a multiple of k + 1. But, since j' \Gamma jj - k implies oe(w; z) 6= 0, no two distinct nodes of KA (x; can have the same color. Thus, it remains to prove that oe(w; z) 6= M . Assume by contradiction that oe(w; Therefore, three cases may occur: l M In case (i), oe(w; which contradicts the fact that jj \Gamma 'j - k. In case (ii), oe(w; z) can be equal to M if and only if is, . Thus, in case (ii), for any pair of nodes z and w of KA (x; which do not satisfy the first condition in (1), it results that is equal to a positive integer and precisely, But this violates the second condition in (1) because (i Finally, in case (iii), oe(w; if and only if jj \Gamma 1). That is, for any pair of nodes z and w of KA (x; not satisfying the first condition in (1), it yields precisely, But again this violates the second condition in (1) because the distance between w and z is (i In conclusion, for k even, any two nodes whose colors differ exactly by M are k relative positions are depicted in Figure 3(a). When k is odd, it follows that l (k+1) 2m that oe(w; z) equivalent to oe(w; z) 6= 0 and oe(w; z) 6= M . Clearly, oe(w; which is verified only if either or is a multiple of k. Hence, two distinct nodes of KA (x; which have the same color are at distance (i It remains to prove that oe(w; z) 6= M . As before, three cases may occur: l M Note that 2 and l M Repeating the same reasoning done for k even, one can show again that any two nodes whose colors differ by M are k apart. Their relative positions are illustrated in Figure 3(b). So, the Array-Coloring Algorithm is conflict-free. Moreover, since it uses the minimum number of colors, the proposed mapping is optimal. (b) (a) Figure 3: Relative positions in A of two nodes which are assigned to the same color: (a) k even, (b) k odd. It is easy to see that the time required to color all the nodes of an array is O(n). Moreover, to color only a single node x = (i; j) of the tree requires only O(1) time, since and hence the mapping is fast. In order to prove that the mapping is balanced, observe that each color appears once in each sub-row of size M . Hence, the number m of nodes with the same color verifies rb c e: 2 Observe that the Array-Coloring Algorithm guarantees conflict-free access to some paths longer than k. Specifically, it is possible to access without conflicts any horizontal path of length M and any vertical path of length g.c.d.(M;\Delta) because L is the minimum integer such that Finally, since the distance between two consecutive nodes on the same diagonal of A is 2, any b k consecutive elements on a diagonal can be accessed with no conflicts. Accessing Paths in Circular Lists Let a circular list C be the data structure D to be mapped into the multibank memory system. A circular list of n nodes, indexed consecutively from 0 to n \Gamma 1, is a sequence of n nodes such that node i is connected to both nodes (i \Gamma 1) mod n and (i The graph E) representing C is a ring, whose vertices correspond to the elements of C and whose arcs correspond to any pair of adjacent elements of C. For the sake of simplicity, C will be used instead of GC since there is no ambiguity. At least M memory banks are required for conflict-free accessing P k in C. Proof For conflict-free accessing P k in C two nodes with the same color must be at distance at least k + 1. When all the nodes are mutually at distance less than k and must all be colored with different colors. When each color may appear at most times. Therefore, Figure 4: Conflict-free access to P 4 in a circular list C of 13 nodes colored by the Circular-List-Coloring algorithm with 7. at least \Upsilon colors are needed. Observed that follows that at least memory banks are required. 2 Below, an optimal conflict-free mapping is provided to color all the nodes of a circular list C using as few colors as in Lemma 2. As before, the color assigned to node x is denoted by fl(x). Algorithm Circular-List-Coloring (C; k); ffl Assign to node x 2 C, the color Note that a linear (that is, non circular) list L can be optimally colored to conflict-free access P k with which matches the trivial lower bound given by the number of nodes in P k . In fact, L can be optimally colored by a naive algorithm which assigns to node x the color Such a naive algorithm does not work for circular lists. For example, consider the circular list C of nodes, shown in Figure 4, to be colored to access P 4 . Applying the naive algorithm with M only the first 10 nodes can be feasibly colored with 5 colors, but 3 additional colors are then required for feasibly coloring the last 3 nodes, for a total of 8 colors. In contrast, the optimal Circular-List- (b) (a) Figure 5: A circular list C of 17 nodes colored to conflict-free access P 3 according to: (a) the Circular- Coloring algorithm requires 7 colors only. Moreover, it is worth to point out that the naive algorithm does not always work for circular lists even when applied with M . For instance, for 5. Applying the naive algorithm with M to this instance, 15 nodes can be colored using 5 colors, but 2 additional colors are needed for feasibly coloring the last 2 nodes for a total of 7 colors (as shown in Figure 5(b)). Instead, the optimal coloring provided by the Circular-List-Coloring algorithm uses only 5 colors, as shown in Figure 5(a). Indeed, the naive algorithm always produces a feasible (although not necessarily optimal) coloring if applied using Theorem 2 The Circular-List-Coloring mapping is optimal, fast, and balanced. Proof To prove optimality, two cases may be distinguished. If and the Circular-List-Coloring algorithm reuses the same color at distance M . Hence, no conflict arises. If n 6j 2. Two nodes get the same color only if they are at distances M or M \Gamma 1, which are both greater than or equal to k+ 1. Hence, as before, no conflict arises. Since the algorithm uses as few colors as possible, the mapping is optimal. It is also fast since each node is colored in constant time. Finally, each color is assigned to exactly n nodes when n is a multiple of M , and no more than l min(n;') l max(n\Gamma';0) nodes are colored with the same color in all the other cases. 2 It is interesting to note at this point that, given a circular list of n nodes, the minimum number of colors required to conflict-free access P k satisfies the following properties (see Figure 6): ffl Up to results, i.e. all the nodes must have different colors. Indeed, all of them are mutually at distance no more than k and, therefore, they form a clique on the graph depends on both n and k, and, for a fixed k, is not a monotone M(n, Figure The number of colors M(n; 6) required to conflict-free access P 6 when n ranges between 1 and 58. function of n. In contrast, for arrays and trees (as will be proved in the next section), M depends only on k and is monotone. Accessing Paths in Complete Trees Let a rooted complete binary tree B be the data structure to be mapped into the multibank memory system. The level of node x 2 B is defined as the number of edges on the path from x to the root, which is at level 0. The maximum level of the nodes of B is the height of B. Let LevB (i) be the set of all nodes of B at level i - 0. A complete binary tree of height H is a rooted tree B in which all the leaves are at the same level and each internal node has exactly 2 children. Thus, LevB (i) contains 2 i nodes. The h-th ancestor of the node (i; j) is the node (i its children are the nodes (i in the left-to-right order. From now on, the generic node x, which is the j-th node of LevB (i), with counting from left to right, will be denoted by Therefore, the generic path instance P k [x; y] will be denoted by Lemma 3 At least memory banks are required to conflict-free access P k in B. Proof Consider a generic node x = (i; j). All the 2 b k nodes in the subtree S of height b k rooted at the b k c-th ancestor of x are mutually at distance not greater than k. In addition, consider the d k , ancestors of x, on the path I of length d k 2 e from the b k c-th ancestor of x up to the k-th ancestor of x. All these nodes are at distance not greater than k from node x, and together with the nodes of S they are at mutual distance not greater than k. x (a) (b) Figure 7: A subset KB (k) of nodes of B that forms a clique in Moreover, for nodes in the complete subtree of height rooted at the - j 's child which does not belong to I. Such nodes are at distance not greater than k from x. Furthermore, these nodes, along with the nodes of S and I, are all together at mutual distance not greater than k. Hence, in the associated graph GDP k there is at least a clique of size From that, the claim easily follows. Figure 7 shows a subset KB (k) of nodes of B which are at pairwise distance not greater than k, for 4, and hence forms a clique in the associated graph GBP kAn optimal conflict-free mapping to color a complete binary tree B acts as follows. A basic subtree KB (k) defined as in the proof of Lemma 3 is identified and colored. Such a tree is then overlaid to B in such a way that the uppermost levels of B coincide with the lowermost levels of KB (k). Then, the complete coloring of B is produced level by level by assigning to each node the same color as an already colored node. Formally, for a given k, define the binary tree KB (k) as follows: ffl KB (k) has a leftmost path of k nodes. ffl the root of KB (k) has only the left child; ffl a complete subtree of height is rooted at the right child of the node at level i on the leftmost path of KB (k). (a) (b)6 7 8 934 Figure 8: Coloring of B for conflict-free accessing: (a) P 3 , (b) P 4 . (Both KB (3) and KB (4) are depicted by dash splines.) The nodes of KB (k) must be colored with 2 b k different colors. Thus, the uppermost levels of B are already colored. For the sake of simplicity, to color the remaining part of B, the levels are counted starting from the root of KB (k). That is, the level of the root of B will be renumbered as level 1. Now, fixed the algorithm to color B acts as follows. Algorithm Binary-Tree-Coloring (B; k); ffl Color KB (k) with M colors; ffl Visit the tree B in breadth first search, and for each node x of B, with mod 2; - Assign to x the same color as that of the node y and Examples of colorings to conflict-free access P 3 and P 4 are illustrated in Figure 8. x y Figure 9: For inherits the same color as node y Theorem 3 The Binary-Tree-Coloring mapping is optimal, fast and balanced. Proof. To prove that the mapping is optimal, it must be shown that it is conflict-free and it uses as few colors as those given by Lemma 3. First, observe that the 2 b k c leaves of a subtree of height are at mutual distance not greater than k, and therefore they must be colored with all different colors. Thus, let each level of B be partitioned (starting from the leftmost node) into consecutive blocks of size 2 b kc . The block b(i; w), with w - 0, at level i of B consists of the 2 b kc consecutive nodes (i; w2 b k which must all be assigned to a different color. Consider the node x = (i; j) to be colored. The node x = (i; belongs to the block b and it appears in the (- position inside the block. Consider the leftmost node z of b x , where . Then, a generalization KB (z; of KB (k) can be defined depending on z. KB (z; includes the following nodes of B: ffl the nodes on the path \Gamma of length k from the father of z up to the (k 1)-th ancestor of z; ffl for the nodes of the complete binary tree S q of height k \Gamma q rooted at the child, which does not belong to \Gamma, of the q-th ancestor of z; ffl the nodes of the complete binary tree S of height rooted at the -th ancestor of z. It is crucial to note that all the following nodes are at distance k + 1 from all the nodes in b x : (i) the root of KB (z; k), (ii) the leaves of S q , with (iii) the leaves of S, which are not parents of any node in b x . The nodes of b are colored from left to right copying the same colors used in the nodes of KB (z; specified in (i), (ii), and (iii) above, and considered by increasing level and from left to right, as illustrated in Figures 10 and 11 for k even and odd, respectively. In particular, is assigned to the same color as the root of KB (z; k), which is the ancestor of x; 2, the 2 k\Gammaq nodes of b x , (i; b j are assigned to the same colors as the leaves of the tree S q . Observe that the number of nodes colored with the two steps above is 1 When k is odd, this is enough to color the entire block since 2 d k c . In fact, the set of nodes of KB (z; specified in (iii) above is empty for k odd. In contrast, when k is even, only the first half of the block has been colored since 2 d k to color the second half of the block, one further step is required, which uses the colors of the nodes of KB (z; specified in (iii) above: ffl The rightmost 2 b k nodes of b x are assigned to the same colors as the rightmost (resp., leftmost) leaves of the complete binary tree rooted at ( 1)-th ancestor of z, depending on the fact that the -th ancestor of z is a left (resp., right) child of its father. In order to prove that the mapping is conflict-free, an inductive reasoning on the level i of the tree is followed. The basis for the induction is when the tree coincides with KB (k) and it is colored, by definition, with all different colors. For i ? k, consider a generic node x = (i; j), its block b x and its leftmost node z. By inductive hypothesis, all the nodes in the tree up to level are colored in a conflict-free manner, but with color repetitions. In particular, the subtree KB (z; k) is conflict-free and since its nodes are mutually at distance at most k they must have been assigned to all different colors. The algorithm colors b x copying the colors of some nodes in KB (z; k), specified in (i), (ii), and (iii), which are exactly at distance k + 1 from the nodes of b x . Therefore, there are no color repetitions in b x and no conflict can arise. Note that nodes in different blocks at level i may inherit the same color, but since any two nodes in different blocks are at distance at least k conflict can arise. Therefore, all the nodes in the tree up to level i are colored in a conflict-free manner. Finally, since the tree is colored with the colors of KB (k), whose number equals the lower bound of Lemma 3, the tree-coloring mapping is optimal. It is easy to see that the time required to color all the n nodes of a tree is O(n). However, to color only a single node x of the tree requires only O(log n) time since, in the worst case, all the nodes along a path from x up to the root must have been colored. One can readily see that, if the height H of the tree B is a multiple of k, then the nodes of B can be partitioned into subsets, each of which induces a copy of KB (k). Therefore, each color is used m times, and the mapping is balanced. 2 z G Figure 10: The generalization KB (z; 6) of KB (6) for the node z. The root of KB (z; 6), the leaves of the and the rightmost leaves of S are used to color the nodes in the block b z . G z Figure 11: The generalization KB (z; 5) of KB (5) for the node z. The root of KB (z; 5) and the leaves of the subtrees S 4 and S 5 are used to color the nodes in the block b z . The results shown for binary trees can be extended to a q-ary tree Q, with q - 2. Corollary 1 At least memory modules are required to conflict-free access P k in a q-ary tree Q. 2 Similarly to the binary case, for a given k, define a q-ary tree K q ffl K q Q (k) has a leftmost path of k ffl the root of K q Q has only the leftmost child; ffl a complete subtree of height is rooted at the q \Gamma 1 rightmost children of the node at level i on the leftmost path of K q Such a K q Q (k) is then overlaid to Q in such a way that the uppermost levels of Q coincide with the lowermost levels of K q Then, the complete coloring of Q is produced level by level by assigning to each node the same color as an already colored node. For the sake of simplicity, to color the remaining part of Q, the levels are again counted starting from the root of K q That is, the level of the root of Q will be renumbered as level 1. Now, the algorithm to color Q is the following: Algorithm q-ary-Tree-Coloring (Q; k); ffl Color K q ffl Visit the tree Q in breadth first search, and for each node x of Q, with do: mod q; - Assign to x the same color as that of the node y and By a reasoning similar to that employed for complete binary trees, the optimality of the q-ary-Tree- Coloring Algorithm easily follows. 6 Conclusions In this paper, the problem of conflict-free accessing arbitrary paths P k in particular data structures, such as bidimensional arrays, circular lists and complete trees, has been considered for the first time and reduced to variants of graph-coloring problems. Optimal, fast and balanced mappings have been proposed. Indeed, the memory bank to which a node is assigned is computed in constant time for arrays and circular lists, while it is computed in logarithmic time for complete trees. However, it remains as an open question whether a tree node can be assigned to a memory bank in constant time. On the other hand, the conflict-free access to P k on an arbitrary data structure D is NP-complete [13], and this justifies the investigation of good heuristics. This problem is equivalent to the classical node coloring problem in the associated graph GDP k . Therefore, it can be solved by the most effective coloring heuristic known so far, that is, the saturation-degree heuristic [6], which works as follows. Let N(x) be the neighborhood of node x in the associated graph GDP k . At each iteration, the saturation- degree heuristic selects the node x to be colored as one with the largest number of different colors already assigned in N(x). Ties between nodes are broken by preferring the node x with the largest number of colored nodes in N(x). Once selected, node x is assigned the lowest color not yet assigned in N(x). As experimentally proved in [5], the saturation-degree heuristic is especially effective when the minimum number of colors is given by the size of the largest clique K of GDP k . Therefore, it should work efficiently also for the conflict-free access problem, and, in particular, for d-dimensional arrays as well as for generic, i.e. not necessarily complete, trees. Indeed, it is expected in such cases that the minimum number of required memory banks be equal to the lower bound given by the size of the largest clique K of GDP k , as happened for bidimensional arrays and complete trees. Unfortunately, the resulting coloring is not guaranteed to be optimal, fast or balanced. Moreover, it is still an open question to determine whether the problem of conflict-free accessing paths on d-dimensional arrays and generic trees is NP-complete. Finally, in a more practical perspective, the number of memory banks available could be fixed to a constant -, depending on the memory configuration. Then, if the number of memory modules M(k) required for a given P k is larger than -, no conflict-free access is possible. However, assume that P k 0 is the longest path that can be accessed without conflicts using - memory banks, i.e. M(k 0 ) -. Then, accessing P k , no more than d k conflicts may arise. Hence, the proposed mappings are scalable. Acknowledgement The authors are grateful to Richard Tan for his helpful comments, and to Thomas McCormick for having provided the reference [13]. --R "Toward a Universal Mapping Algorithm for Accessing Trees in Parallel Memory Systems" "Multiple Template Access of Trees in Parallel Memory Systems" "Accounting for Memory Bank Contention and Delay in High-Bandwidth Multiprocessors" "On Array Storage for Conflict-Free Memory Access for Parallel Processors" "Assigning Codes in Wireless Networks: Bounds and Scaling Properties" "The Organization and Use of Parallel Memories" "Conflict-Free Data Access of Arrays and Trees in Parallel Memory Systems" "Parallel Priority Queues in Distributed Memory Hypercubes" "Load Balanced Mapping of Data Structures in Parallel Memory Modules for Fast and Conflict-Free Templates Access" New York "Optimal Approximation of Sparse Hessians and its Equivalence to a Graph Coloring Problem" "Conflict-Free Template Access in k-ary and Binomial Trees" "Theoretical Limitations on the Efficient Use of Parallel Memories" --TR Conflict-free template access in <italic>k</italic>-ary and binomial trees Accounting for Memory Bank Contention and Delay in High-Bandwidth Multiprocessors Multiple templates access of trees in parallel memory systems Assigning codes in wireless networks New methods to color the vertices of a graph Latin Squares for Parallel Array Access Optimal and Load Balanced Mapping of Parallel Priority Queues in Hypercubes Load Balanced Mapping of Data Structures in Parallel Memory Modules for Fast and Conflict-Free Templates Access Toward a Universal Mapping Algorithm for Accessing Trees in Parallel Memory Systems --CTR Alan A. Bertossi , Cristina M. Pinotti , Richard B. Tan, Channel Assignment with Separation for Interference Avoidance in Wireless Networks, IEEE Transactions on Parallel and Distributed Systems, v.14 n.3, p.222-235, March Sajal K. Das , Irene Finocchi , Rossella Petreschi, Conflict-free star-access in parallel memory systems, Journal of Parallel and Distributed Computing, v.66 n.11, p.1431-1441, November 2006
bidimensional array;conflict-free access;complete tree;path template;multibank memory system;mapping scheme;frequency assignment;circular list
589754
Compiler-optimized simulation of large-scale applications on high performance architectures.
In this paper, we propose and evaluate practical, automatic techniques that exploit compiler analysis to facilitate simulation of very large message-passing systems. We use compiler techniques and a compiler-synthesized static task graph model to identify the subset of the computations whose values have no significant effect on the performance of the program, and to generate symbolic estimates of the execution times of these computations. For programs with regular computation and communication patterns, this information allows us to avoid executing or simulating large portions of the computational code during the simulation. It also allows us to avoid performing some of the message data transfers, while still simulating the message performance in detail. We have used these techniques to integrate the MPI-Sim parallel simulator at UCLA with the Rice dHPF compiler infrastructure. We evaluate the accuracy and benefits of these techniques for three standard message-passing benchmarks on a wide range of problem and system sizes. The optimized simulator has errors of less than 16% compared with direct program measurement in all the cases we studied, and typically much smaller errors. Furthermore, it requires factors of 5 to 2000 less memory and up to a factor of 10 less time to execute than the original simulator. These dramatic savings allow us to simulate regular message-passing programs on systems and problem sizes 10 to 100 times larger than is possible with the original simulator, or other current state-of-the-art simulators.
Introduction Predicting parallel application performance is an essential step in developing large applications on highly scalable parallel architectures, in sizing the system configurations necessary for large problem sizes, or in analyzing alternative architectures for such systems. Considerable research is being done on both analytical and simulation models for performance prediction of complex, scalable systems. Analytical methods typically require custom solutions for each problem and may not be tractable for complex interconnection networks or detailed modeling scenarios; simulation models are likely to be the primary choices for general-purpose performance prediction. As is well known, however, detailed simulations of large systems can be very computation-intensive and their long execution times can be a significant deterrent to their widespread use. The current generation of parallel program simulators use two techniques to reduce model execution times: direct execution and parallel simulation. In direct execution, the simulator uses the available system resources to directly execute portions of the program. Parallel simulation distributes the computational workload among multiple processors, while using appropriate synchronization algorithms to ensure that execution of the model produces the same result as if all events in the model were executed in their causal order. However, the current state of the art is such that even using direct execution and parallel simulations, the simulation of large applications designed for architectures with thousands of processors can run many orders of magnitude slower than their physical counterparts. In this paper, we propose, implement, and evaluate practical, automatic optimizations that exploit compiler support to enable efficient simulation of very large message-passing parallel programs. Our goal is to enable the simulation of target systems with thousands of processors, and realistic problem sizes expected on such large platforms. The key idea underlying our work is to use compiler analysis to isolate fragments of local computations and message data whose values do not affect the performance of the program. For example, computations that determine loop bounds, control flow, or message patterns and volumes all have an effect on performance, whereas computations of many array values have no significant effect on performance. These computations can be abstracted away while simulating the rest of the program in detail to predict the performance characteristics of the application. Similarly, it is also possible to avoid performing data transfers for many messages whose values do not affect performance, while simulating the performance of the messages in detail. There are two major aspects to the compiler analysis required to accomplish this optimization: identifying the values within the program that could affect program performance, and isolating the computations and communications that determine these values. To perform the first step, we use a compiler-synthesized static task graph model [4, 5], an abstract representation that identifies the sequential computations (tasks), the parallel structure of the program (task scheduling, precedences, and explicit communication), and the control-flow that determines the parallel structure. The symbolic expressions in the task graph for control flow conditions, communication patterns and volumes, and scaling expressions for sequential task execution times directly capture all the values (i.e., the references within those expressions) that impact program performance. The second step uses a compiler technique called program slicing [21] to identify the portions of the computation that determine these values. The compiler can then emit simplified MPI code that contains exactly the computations that must be actually executed during the simulation (in addition to the communication), while the remaining code fragments are abstracted away. The compiler also needs to estimate the execution time of the abstracted code by using parameterized by direct measurement. In addition to reducing simulation times, these optimizations can dramatically reduce the memory requirements for the simulation (if major program arrays are only referenced in redundant computations, they do not have to be allocated at all during the simulation). The memory savings can potentially allow much larger problem sizes and architectures to be studied than would otherwise be feasible. In order to demonstrate the impact of these optimizations, we have combined the MPI-Sim parallel simulator [6, 25-27] with the dHPF compiler infrastructure [2], to develop a program simulation framework that incorporates the new techniques described above. The original MPI-Sim simulator used both direct execution and parallel simulation to achieve substantial reductions in the simulation time of parallel programs. dHPF, in normal usage, compiles an HPF program to MPI (or to a variety of shared memory systems), and provides extensive parallel program analysis capabilities. The integrated tool can allow us to evaluate the impact of the preceding optimizations with existing MPI and HPF programs without requiring any changes to the source code. In previous work, we modified the dHPF compiler to automatically synthesize the static task graph model and symbolic task time estimates for MPI programs compiled from HPF source programs. 1 In this work, we use the static task graph plus program slicing to perform the simulation optimizations described above. We have also extended MPI-Sim to exploit the information from the compiler, and avoid executing significant portions of the computational code. The hypothesis is that this will significantly reduce the memory and time requirements of the simulation and therefore enable us to simulate much larger systems and problem sizes than were previously possible. We use a number of widely used benchmarks to evaluate the utility of the integrated framework: Sweep3D [1], a benchmark; SP from the NPB benchmark suite [8] and Tomcatv, a SPEC92 benchmark. The simulation models of each application were validated against measurements over a range of problem sizes and numbers of processors. The errors in the predicted execution times, compared with direct measurement, were at 1 In the future, we plan to synthesize this information for existing MPI codes as well. The dHPF infrastructure supports very general computation partitioning, communication analysis, and symbolic analysis capabilities that make this feasible for a wide class of MPI programs. most 16% in all cases we studied, and often were substantially less. The validation has been done for the distributed memory IBM SP architecture, as well as the shared memory SGI Origin 2000 (note, that MPI-Sim simulates the MPI communication, not the communications via shared memory). The optimizations had a significant impact on the performance of the simulators: the total memory usage of the simulator using the compiler synthesized model was a factor of 5 to 2000 less than the original simulator, and the simulation time was typically lower by a factor of 5-10. These dramatic savings allow us to simulate systems or problem sizes that are 10-100 times larger than is possible with the original simulator, without significant reductions in the accuracy of the simulator. For example, we were successful in simulating the execution of a configuration of Sweep3D for a target system with 10,000 processors! In many cases, the simulation time was faster than the original program. The remainder of the paper proceeds as follows. Section 2 first describes the state of the art of parallel program simulation, to set the stage for our work. Section 3 provides a brief overview of MPI-Sim and the static task graph model. Section 4 describes the optimization strategy and the compiler and simulator extensions required to implement the strategy. Section 5 describes our experimental results, and Section 6 presents our main conclusions. Related Work Because analytical performance prediction can be intractable for complex applications, program simulations are commonly used for such studies. It is well known that simulations of large systems tend to be slow. To improve the simulators, direct-execution has been used [20, 26, 28]. Direct execution simulators make use of available system resources to directly execute portions of the application code and simulate architectural features that are of specific interest, or are unavailable. For example, simulators can be used to study various architectural components such as the memory subsystem or the interconnection network. Specifically, if one is interested in determining if a faster communication fabric for a network of workstations is of value for a given set of applications, one can run the application on the currently available machines and only simulate the projected network's behavior. The benefits of this direct-execution simulation are obvious: first, one can estimate the value of the new hardware without the expense of purchasing it; second, one can do the simulation fast-there is no need to simulate the workstation's behavior (for example down to the level of memory references) since that part of the hardware is readily available. Many of the early simulators were designed for sequential execution [9, 13, 14]. However, even with the use of abstract models and direct execution, sequential program simulators tended to be slow with slowdown factors ranging from 2 to 35 for each process in the simulated program [9]. Several recent efforts have been exploring the use of parallel execution [10, 16, 17, 23, 24, 27, 28] to reduce the model execution times, with varying degrees of success. In order to have multiple simulation processes and maintain accuracy, simulations use protocols to synchronize the processes. One of the widely used protocols is the Quantum protocol, which lets the processes compute for a given quantum before synchronizing them. In general, synchronous simulators that use the quantum protocol must trade-off simulation accuracy with speed-frequent synchronizations slowdown the simulation, but synchronizing less frequently introduces errors, by possibly executing statements out-of-order. Both LAPSE [16, 17] and Parallel Proteus use some form of program analysis to increase the simulation window beyond a fixed quantum. MPI-Sim uses parallel discrete event simulation with the conservative protocol [24, 27]. Supported protocols include the Null Message Protocol (NMP) [11], the Conditional Event Protocol (CEP) [12], and a new protocol, which is a combination of the two [22]. As discussed in the next section, MPI-Sim exploits the determinism present in the communication pattern of the application to reduce, and in many cases, completely eliminate synchronization overheads. Although simulation protocol optimizations have reduced simulation times, the resulting improvements are still inadequate to simulate the very large problems that are of interest to high-end users. For instance, Sweep3D is a kernel application of the ASCI benchmark suite released by the US Department of Energy. In its largest configuration, it requires computations on a grid with one billion elements. The memory requirements and execution time of such a configuration makes it impractical to simulate, even when running the simulations on high performance computers with hundreds of processors. To overcome this computational intractability, researchers have used abstract simulations, which avoid execution of the computational code entirely [18, 19]. However, this leads to major limitations that make the approach inapplicable to many real world applications. The main problem with abstracting away all of the code is that the model is essentially independent of program control flow, even though the control flow may affect both the communication pattern as well as the sequential task times. Also, the preceding solution requires significant user modifications to the source program (in the form of a special input language) in order to express required information about abstracted sequential tasks and communication patterns. This makes it difficult to apply such a tool to existing programs written with widely used standards such as Message Passing Interface (MPI) or High Performance Fortran (HPF). 3 Background and Goals 3.1 MPI-SIM: Parallel Simulation of MPI programs using Direct Execution The starting point for our work is MPI-Sim, a direct-execution parallel simulator for performance prediction of MPI programs. MPI-Sim simulates an MPI application running on a parallel system (referred to as the target program and system respectively). The machine on which the simulator is executed (the host machine) may be either a sequential or a parallel machine. In general, the number of processors in the host machine will be less than the number of processors in the target architecture being simulated, so the simulator must support multi-threading. The simulation kernel on each processor schedules the threads and ensures that events on host processors are executed in their correct timestamp order. A target thread is simulated as follows. The local code is simulated by directly executing it on the host processor. Communication commands are trapped by the simulator, which uses an appropriate model to predict the execution time for the corresponding communication activity on the target architecture. supports most of the commonly used MPI communication routines, such as point-to-point and collective communications. In the simulator, all collective communication functions are implemented in terms of point-to- point communication functions, and all point-to-point communication functions are implemented using a set of core non-blocking MPI functions. In general, the host architecture will have fewer processors than the target machine (for sequential simulation, the host machine has only one processor); this requires that the simulator provide the capability for multithreaded execution. Since MPI programs execute as a collection of single threaded processes, it was necessary to provide a capability for multithreaded execution of MPI programs in MPI-Sim. Further, memory and execution time constraints of sequential simulation led to the development of parallel implementations of MPI-Sim. MPI-Sim has been ported to multiple parallel architectures including a distributed memory IBM SP2 as well as a shared-memory SGI Origin 2000. The simulation kernel provides support for sequential and parallel execution of the simulator. Parallel execution is supported via a set of conservative parallel simulation protocols [26], which typically work as follows: Each 2 In the future, we plan to synthesize this information for existing MPI codes as well. The dHPF infrastructure supports very general computation partitioning, communication analysis, and symbolic analysis capabilities that make this feasible for a wide class of MPI programs. application process in the simulation is modeled by a Logical Processes (LP) 3 . Each LP can execute independently, without synchronizing with other LPs, until it executes a wait operation (such as an MPI-Recv, MPI-Barrier etc.); a synchronization protocol is used to decide when such an LP can proceed. We briefly describe the default protocol used by MPI-SIM. Each LP in the model computes local quantities called Earliest Output Time (EOT) and Earliest Input Time (EIT) [7]. The EOT represents the earliest future time at which the LP will send a message to any other LP in the model; similarly the EIT represents a lower bound on the receive timestamp of future messages that the LP may receive. Upon executing a wait statement, an LP can safely select a matching message (if any) from its input buffer, that has a receive timestamp less than its EIT. Different asynchronous protocols differ only in their method for computing EIT. Our implementation supports a variety of such protocols as mentioned previously. The primary overhead in implementing parallel conservative protocols is due to the communications to compute EIT and the blocking suffered by an LP that has not been able to advance its EIT. We have suggested and implemented a number of optimizations to significantly reduce the frequency and strength of synchronization in the parallel simulator thus reducing unnecessary blocking in its execution [26, 27]. Our optimizations were geared towards exploiting determinism in applications. For instance, consider an LP that is blocked at a receive statement and its input buffer contains a single message. In general, the LP cannot proceed by removing that message from the buffer as it might be possible that another message destined for this LP is in transit, and that message has a lower timestamp. However, if the receive statement is known by the process to be deterministic, it follows that there must exist a unique message that matches the receive statement. As soon as the LP receives this message, it can proceed without the need for any synchronizations with other LPs in the model. In the best case, if every receive statement in the model is known to be deterministic, no synchronization messages will be generated in the model and the parallel simulation can be extremely efficient. The preceding optimizations have two limitations: first, it works only with communications statements that are a priori known to be deterministic. Second, the use of direct execution in the simulator implies that the memory and computation requirements of the simulator are at least as large as that of the target application, which restricts the target systems and application problem sizes that can be studied even using parallel host machines. The compiler-directed optimizations discussed in the next section are primarily aimed at alleviating these restrictions. 3.2 The Static Task Graph Representation As will be seen in the next section, the compiler analysis to be performed can be greatly facilitated by exploiting an appropriate abstract representation for the parallel behavior of the program. As part of the POEMS project [3, 15], we have developed an abstract program representation called the static task graph (STG) that captures extensive static information about a parallel program [5]. The STG is designed to be computed automatically by a parallelizing compiler. It is a compact, symbolic representation of the parallel structure of a program, independent of specific program input values or the number of processors. Each node of the STG represents a set of possible parallel tasks, typically one per process, identified by a symbolic set of integer process identifiers. To illustrate, the STG for the example MPI program is shown in Figure 1. The compute node for the loop nest represents a set of tasks, one per process, denoted by the symbolic set of process ids }0 p . Each node also includes markers describing the corresponding region of source code of the original program (for now, each node must represent a contiguous region of code). Each edge of the graph represents a set of edges connecting pairs of parallel tasks described by a symbolic integer mapping. For example, the communication edge in the figure is labeled with a mapping indicating that each process p ( 1 sends to process nodes fall into one of three categories: control-flow, computation and communication. Each computational node includes a symbolic scaling function that captures how the number of loop iterations (if any) in the task 3 In general, an LP can be used to simulate multiple application processes. scales as a function of arbitrary program variables. Each communication node includes additional symbolic information describing the pattern and volume of communication. Overall, the STG serves as a general, language- and architecture-independent representation of message-passing programs. In previous work, we extended the dHPF compiler to synthesize static (and dynamic) task graphs for MPI programs generated by the dHPF compiler from HPF source programs [4]. In the future, we will extract task graphs directly from existing MPI codes. This compiler support is extremely valuable because it enables the techniques developed in this paper to be applied fully automatically, i.e., without user intervention, for efficient simulation of parallel programs. Compiler-Supported Techniques for Efficient Large-Scale Simulation This section begins by motivating the overall strategy we use to address the key restriction on simulation scalability identified above, namely, the time and cost required for simulating the detailed computations of the target program. We then describe more specifically how this strategy is accomplished. 4.1 Optimization Strategy and Challenges Parallel program simulators used for performance evaluation execute or simulate the actual computations of the target program for two purposes: (a) to determine the execution time of the computations, and (b) to determine the impact of computational results on the performance of the program, due to artifacts like communication patterns, loop bounds, and control-flow. For many parallel programs, however, a sophisticated compiler can extract extensive information from the target program statically. In particular, we identify two types of relevant information often available at compile-time: 1. The parallel structure of the program, including the sequential portions of the computation (tasks), the mapping of tasks to threads, and the communication and synchronization patterns between threads. 2. Symbolic estimates for the execution time of isolated sequential portions of the computation. If this information can be provided to the simulator directly, it may be possible to avoid executing substantial portions of the computational code during simulation, and therefore reduce the execution time and memory requirements of the simulation. To illustrate this goal, consider the simple example MPI code fragment in Figure 1. The code performs a "shift" communication operation on the array D, where every processor sends its boundary values to its left neighbor, and then the code executes a simple computational loop nest. In this simple example, the communication pattern and the number of iterations of the loop nest depend on the values of the block size per processor (b), the array size (N), the number of processors (P), and the local processor identifier (myid). Therefore, the computation of these values must be executed or simulated during the simulation. However, the communication pattern and loop iteration counts do not depend on the values stored in the arrays A and D, which are computed and used in the computational loop nest (or earlier). We refer to these latter values as redundant computations (from the point of view of performance estimation). If we can estimate the performance of the computational loop nest analytically, we could avoid simulating the code of this loop nest, while still simulating the communication behavior in detail. We could achieve this optimization by using the compiler to generate the simplified code shown on the right in the figure. In this code, we have replaced the loop nest with a call to a special simulator-provided delay function. We have extended MPI-Sim to provide such a function, which simply forwards the simulation clock on the double precision A(NMAX, 1 double precision call mpi_comm_size(MPI_COMM_WORLD, P, ierr) call mpi_comm_rank(MPI_COMM_WORLD, myid, ierr) read(*, N) if (myid .gt. <SEND D(2:N-1, myid*b+1) to processor myid-1> endif if (myid .lt. P) then <RECV D(2:N-1, (myid+1)*b+1) from processor myid+1> endif do do endif integer, allocatable :: dummy_buf call mpi_comm_size(MPI_COMM_WORLD, P, ierr) call mpi_comm_rank(MPI_COMM_WORLD, myid, ierr) call read_and_broadcast(w_1) read(*, N) allocate dummy_buf((N-2)*2) if (myid .gt. <SEND dummy_buf(:) to processor myid-1> endif if (myid .lt. P) then <RECV dummy_buf(:) from processor myid+1> endif call delay((N-2) * (min(N,myid*b+b) - Figure 1: Example to illustrate (a) a simple MPI program, (b) task-graph for MPI program, and (c) simplified MPI program for efficient simulation. (a) Original MPI Code (c) Simplified MPI Code (b) Task Graph for Original MPI Code Task Pairs: {[p] - [q]: DO Compute DO I Control-flow edge C Communication edge Compute Tasks: simulation thread by a specified amount. The compiler estimates the cost of the loop nest in the form of a simple scaling function shown as the argument to the delay call. This function describes how the computational cost varies with the retained variables (b, N, P and myid), plus a parameter w 1 representing the cost of a single loop iteration. We currently obtain the value of w 1 by direct measurement for one or a few selected problem sizes and number of processors, and use the scaling function to compute the required delay value for other problem sizes and number of processors. Note in the example that the compiler has avoided allocating the arrays A and D, which significantly reduces the memory required to simulate the program. As an additional optimization, if the compiler can prove that the data transferred in the message is also "redundant", the simulator can also avoid performing an actual data transfer, although it will simulate the message operation in detail. It can also avoid allocating any memory for the message buffer. This message optimization can lead to further savings in simulation time and memory usage. This paper develops automatic compiler-based techniques to perform the optimizations described above, and evaluates the potential benefits of these techniques. In particular, our goal is to use the compiler-generated static task graph (plus additional compiler analysis) to avoid simulating or executing substantial portions of the computational code of the target program and sending unnecessary data. We use the task graph to identify the computational tasks that are candidates for elimination, to compute the scaling expressions for those delay functions, and (most importantly) to identify which values computed in the program have an impact on performance. We use additional compiler analysis to distinguish the computations that compute such values, i.e., those that are not redundant as defined above. More specifically, there are four major challenges we must address in achieving the above goals, of which the first three have not been addressed in any previous system known to us: a) We must transform the original parallel program into a simplified but legal MPI program that can be simulated by MPI-Sim. The simplified program must include only the computation and communication code that needs to be executed by the simulator. It must yield the same performance estimates as the original program for total execution time (for each individual process), total communication and computation times, as well as more detailed metrics of the communication behavior. b) We must be able to abstract away as much of the local computation within each task as feasible and eliminate as many data structures of the original program as possible, by isolating the redundant computations in the program. c) We must identify the messages whose contents do not directly affect the computation at the receiver, and exploit this information to reduce simulation time and memory usage. d) We must estimate the execution times of the abstracted computational tasks for a given program size and number of processors. Accurate performance prediction for sequential code is a challenging problem that has been widely studied in the literature. We use a fairly straightforward approach described in Section 4.5. Refining this approach is part of our ongoing work in the POEMS project. The following subsections describe the techniques we use to address these challenges, and their implementation in dHPF and MPI-Sim. We first describe the basic process of using the task graph to generate the simplified MPI program, then describe the compiler analysis needed to identify redundant computations, and finally discuss the approach we use to estimate the performance of the eliminated code. 4.2 Translating the static task graph into a simplified MPI program The STG directly identifies the local (sequential) computational tasks, control flow, and communication tasks and patterns of the parallel program. By using the compiler-generated STG as the basis for our analysis, we can avoid having to perform a complex, ad hoc analysis to identify these components. Given this information, the first step is to identify contiguous regions of computational tasks and/or control-flow in the STG that can be collapsed into a single condensed (or collapsed) task, such as the loop nest of Figure 1. Note that this is simply a transformation of the STG for simplifying further analysis and does not directly imply any changes to the parallel program itself. We refer to the task graph resulting from this transformation as the condensed task graph. In later analysis, we can consider only a single computational task or a single collapsed task at a time for deciding how to simplify the code (we refer to either as a single sequential task). The criteria for collapsing tasks depend on the goals of the performance study. First, as a general rule, a collapsed region must not include any branches that exit the region, i.e., there should be only a single exit at the end of the region. Second, for the current work, a collapsed region must contain no communication tasks because we aim to simulate communication precisely. Finally, deciding whether to collapse conditional branches involves a difficult tradeoff: it is important to eliminate control-flow that references large arrays in order to achieve the savings in memory and time we desire, but it is difficult to estimate the performance of code containing such control-flow. We have found, however, that there are typically few branches that involve large arrays that do have a significant impact on program performance. For example, one minor conditional branch in a loop nest of Sweep3D depends on intermediate values of large 3D arrays. The impact of this branch on execution time is relatively negligible, but detecting this fact, in general, can be difficult within the compiler because it may depend on expected problem sizes and computation times. Therefore, there are two possible approaches we can take. The more precise approach is to allow the user to specify through directives that specific branches can be eliminated and treated analytically for program simulation. A simpler but more approximate approach is to eliminate any conditional branches inside a collapsible loop nest, and rely on the statistical average execution time of each iteration to provide a good basis for estimating total execution time of the loop nest. With either approach, we can use profiling to estimate the branching probabilities of eliminated branches We have currently taken the second approach, but the first one is not difficult to implement and could provide more precise performance estimates. While condensing the task graph, we also compute a scaling expression for each collapsed task that describes how the number of computational operations scales as a function of program variables. We introduce time variables that represent the execution time of a sequence of statements in a single loop iteration (denoted w i for task i). The approach we use to estimate the overall execution time of each sequential task is described in Section 4.5. Based on the condensed task graph (and assuming for now that the compiler analysis of Section 4.3 is not needed), we generate the simplified MPI program as follows. We retain any control-flow (loops and branches) of the original MPI code that is retained in the condensed task graph, i.e., the control-flow that is not collapsed. Second, we retain the communication code of the original program, in particular only the calls to the underlying message-passing library. If a program array that is otherwise unused is referenced in any communication call, we replace that array reference with a reference to a single dummy buffer used for all the communication. (Note that without the message optimization described later in this section, the simulator must still perform the actual data transfer between processes when simulating the message. The message optimization attempts to eliminate this data transfer itself.) We use a buffer size that is the maximum of the message sizes of all communication calls in the program and allocate the buffer statically or dynamically (potentially multiple times), depending on when the required message sizes are known. Third, we replace the code sequence for each sequential task of the task graph by a call to the MPI-Sim delay function, and pass in an argument describing the estimated execution time of the task. We insert a sequence of calls to a runtime function, one per w i parameter, at the start of the program to read in the value of the parameter from a file and broadcast it to all processors. Finally, we eliminate all the data variables not referenced in the simplified program. 4.3 Program slicing for identifying redundant computations and data The major challenge in performing the transformations mentioned earlier correctly and effectively is to identify the redundant computations, i.e., the ones that can be safely eliminated. The solution we propose is to use program slicing to retain those parts of the computational code (and the associated data structures) that affect the program execution time. Given a variable referenced in some statement, program slicing finds and isolates a subset of the program computation and data that can affect the value of that variable [21]. The subset has to be conservative, limited by the precision of static program analysis, and therefore may not be minimal. The key requirement in applying program slicing is to identify the variable values that affect the execution time of the program. Once again, the compiler-generated static task graph captures this information directly and precisely, allowing us to avoid a complicated and ad hoc analysis of the entire source code. In particular, the values that affect performance are exactly the variable references that appear in the retained control-flow of the condensed graph, in the scaling functions of the sequential tasks and communication events, and in the source and destination expressions of the communication descriptors (or the communication calls themselves). Once these values are identified, program slicing can be used to isolate the computations and data that affect those variable values. Program slicing is essentially a reachability analysis on the dependence graph of the program, including both data and control dependences. In particular, given a particular target reference, we use a reachability analysis to identify the statements in the program that can affect the value of that reference through some chain of dependences (i.e., through some feasible path in the dependence graph). Because this is a well-known compiler technique, we omit the details here. A state-of-the-art algorithm for program slicing is described in [21] and was used as the basis for our implementation. Applying this technique, however, requires that the target reference be part of the program so that it appears in the program dependence graph computed by the compiler. Some of the expressions of the static task graph are not directly derived from corresponding expressions in the program, and therefore cannot be used as starting points for program slicing. For such expressions, we introduce dummy procedure call statements at appropriate points in the target program, passing those expressions as arguments, and then rebuild the program dependence graph. Now, these expressions can be used as starting points for slicing. The dummy procedure calls can later be eliminated. Obtaining the memory and time savings we desire requires full interprocedural program slicing, so that we completely eliminate the uses of as many large arrays as possible. General interprocedural slicing is a challenging but feasible compiler technique that is not currently available in the dHPF infrastructure. For now, we take limited interprocedural side effects into account, in order to correctly handle calls to runtime library routines (including communication calls and runtime routines of the dHPF compiler's runtime library). In particular, we assume that these routines can modify any arguments passed by reference but cannot modify any global (i.e., common block) variables of the MPI program. This is necessary and sufficient to support single-procedure benchmarks. We expect to incorporate full interprocedural slicing in the future, to support continuing work in POEMS. The final output of the slicing analysis is the set of computations that must be retained in the simplified MPI code, while the remaining computations of the program (except for I/O statements and communication calls) can be considered redundant. The code generation for the simplified MPI program (described in the previous section) is modified slightly to use this information. For each sequential task, the non-redundant computations are retained in the generated program, while the rest of the task is replaced with a single call to the simulator delay function. For precise performance prediction, the simulator delay calls should not include the time for the retained computations since those will be simulated (and their time accounted for) explicitly. The execution time estimates computed above, however, apply to the entire task. In practice, we have found that the amount of non-redundant code is very small for most tasks and therefore we do not adjust the execution time estimates to account for this retained code. 4.4 Message optimization for simulating redundant messages As noted in the previous Section, the data transferred in some of the messages may also be "redundant" from the point of view of performance. If such cases can be identified, we can avoid performing the data transfers during the simulation, potentially leading to additional time and memory savings. Although this is conceptually similar to redundant computations, we discuss this "message optimization" separately because the mechanism for achieving this optimization is somewhat different, as explained below. First, the compiler can identify redundant messages as a direct result of the program slicing analysis described above. In particular, the technique described above to account for interprocedural side-effects during slicing directly identifies those message receive calls that receive redundant values. The corresponding message send calls are already known to the compiler. The compiler provides this information to the simulator by flagging the MPI calls that are redundant. The buffers used by these messages are not allocated in the resulting simplified MPI program. The actual message optimization is as follows. If a call is not flagged, MPI-Sim simulates the call in detail (by sending the necessary protocol messages and predicting the end-to-end latency for the messages) and sends the data to the receiving simulation thread, so that the actual data is available to the simulated application. However, if the call is flagged by the compiler as "redundant", then MPI-Sim still simulates the call in the detail with respect to the MPI communication protocol, but sends only an empty message to the receiving simulation thread. Since "redundant" receives are also flagged, the receiver does not copy the data in the buffer. The messages need to be present in the simulated application because they provide information about the synchronization in the program. Although this optimization does not reduce the number of messages sent, the size of the messages is reduced, and the memory used by the messages does not need to be allocated. This results in lower latencies incurred by the messages that are sent between processors as well as smaller communication overheads due to copying the data enclosed in the messages into/from the communication buffers. It also results in lower memory usage by the simulator. 4.5 Estimating task execution times The main approximation in our approach is to estimate sequential task execution times without direct execution. Analytical prediction of sequential execution times is an extremely challenging problem, particularly with modern superscalar processors and cache hierarchies. There are a variety of possible approaches with different tradeoffs between cost, complexity, and accuracy. The simplest approach, and the one we use in this paper, is to measure task times (specifically, the w i ) for one or a few selected problem sizes and number of processors, and then use the symbolic scaling functions derived by the compiler to estimate the delay values for other problem sizes and number of processors. Our current scaling functions are symbolic functions of the number of loop iterations, and do not incorporate any dependence of cache working sets on problem sizes. We believe extensions to the scaling function approach that capture the non-linear behavior caused by the memory hierarchy are possible. Performance Estimates Measured task times Simplified MPI code MPI code with timers Parallel Program dHPF Parallel System Figure 2: Compilation, parameter measurement and simulation for a parallel program. Two alternatives to direct measurement of the task time parameters are (a) to use compiler support for estimating sequential task execution times analytically, and (b) to use separate offline simulation of sequential task execution times [15]. In both cases, the need for scaling functions remains, including the issues mentioned above, because it is important to amortize the cost estimating these parameters over many prediction experiments. The scaling functions for the tasks can depend on intermediate computational results, in addition to program inputs. Even if this is not the case, they may appear to do so to the compiler. For example, in the NAS benchmark SP, the grid sizes for each processor are computed and stored in an array, which is then used in most loop bounds. The use of an array makes forward propagation of the symbolic expressions infeasible, and therefore completely obscures the relationship between the loop bounds and program input variables. We simply retain the executable scaling expressions, including references to such arrays, in the simplified code and evaluate them at execution time. We have been able to automate fully the modeling process for a given HPF application compiled to MPI. The modified dHPF compiler automatically generates two versions of the MPI program. One is the simplified MPI code with delays calls described previously. The second is the full MPI code with timer calls inserted to perform the measurements of the w parameters. The output of the timer version can be directly provided as input to the delay version of the code. This complete process is illustrated in Figure 2. We performed a detailed experimental evaluation of the compiler-based simulation approach. We studied three issues in these experiments: 1. The accuracy of the optimized simulator that uses the compiler-generated information, compared with both the original simulator and direct measurements of the target program. 2. The reduction in memory usage achieved by the optimized simulator compared with the original and the resulting improvements in the overall scalability of the simulator in terms of system sizes and problem sizes that can be simulated. 3. The performance of the optimized simulator compared with the original, in terms of both absolute simulation times and in terms of relative speedup as compared to sequential model execution, when simulating a large number of target processors. Results in each of the above categories are presented for both types of the optimizations considered in this paper: elimination of local computations and elimination of data contents from large messages. We begin with a description of our experimental methodology and then describe the results for each of these issues in turn. 5.1 Experimental Methodology We used three real-world benchmarks (Tomcatv, Sweep3D and NAS SP) and one synthetic communication kernel (SAMPLE) in this study. Tomcatv is a SPEC92 floating-point benchmark, and we studied an HPF version of this benchmark compiled to MPI by the dHPF compiler. Sweep3D, a Department of Energy ASCI benchmark [1], and SP, a NAS Parallel Benchmark from the NPB2.3b2 benchmark suite [8], are MPI benchmarks written in Fortran 77. Finally, we designed the synthetic kernel benchmark, SAMPLE, to evaluate the impact of the compiler-directed optimizations on programs with varying computation granularity and message communication patterns that are commonly used in parallel applications. For Tomcatv, the dHPF compiler automatically generates three versions of the output MPI code: (a) the normal MPI code generated by dHPF for this benchmark, where the key arrays of the HPF code are distributed across the processors in contiguous blocks in the second dimension (i.e., using the HPF distribution (*,BLOCK)); (b) the simplified MPI code with the calls to the MPI-Sim delay function, making full use of the techniques described in Section 4; and (c) the normal MPI code with timer calls inserted to measure the task time parameters, as described in Section 4.5. Since dHPF only parses and emits Fortran and MPI-Sim only supports C, we use f2C to translate each version of the generated code to C and run it on MPI-Sim. For the other two benchmarks, Sweep3D and NAS SP, we manually modified the existing MPI code to generate the simplified MPI and the MPI code with timers for each case (since the task graph synthesis for MPI codes is not implemented yet). These codes serve to show that the compiler techniques we developed can be applied to a large range of codes with good results. For each application, we measured the task times (values of w i ) on 16 processors. These measured values were then used in experiments with the same problem size on different numbers of processors. The only exception was NAS SP, where we measured the task only for a single problem size (on 16 processors), and used the same task times for other problem sizes as well. Recall that the scaling functions we use currently do not account for cache working sets and cache performance. Changing either the problem size or the number of processors affects the working set size per process and, therefore, the cache performance of the application. Nevertheless, the above measurement approach provided very accurate predictions from the optimized simulator, as shown in the next subsection. All benchmarks, except SAMPLE, were evaluated for the distributed memory IBM SP (with up to 128 processors); the SAMPLE experiments were conducted on the shared memory SGI Origin 2000 (with up to 8 processors). 5.2 Validation The original MPI-Sim was successfully validated on a number of benchmarks and architectures [6, 26, 27]. The new techniques described in Section 4, however, introduce additional approximations in the modeling process. The key new approximation is in estimating the sequential execution times of portions of the computational code (tasks) that have been abstracted away. Our aim in this section is to evaluate the accuracy of MPI-Sim when applying these techniques. For each application, the optimized simulator (henceforth denoted as MPI-SIM-TG) was validated against direct measurements of the application execution time and also compared with the predictions from the original simulator. We studied multiple configurations (problem size and number of processors) for each application. In all cases MPI-SIM-TG is validated against the measured system. 4 We begin with Tomcatv, which is handled fully automatically through the steps of compilation, task measurements, and simulation shown in Figure 2. The size of Tomcatv used for the validation was 2048-2048. Figure 3 shows the results from 4 to 64 processors. Even though MPI-Sim with the analytical model (MPI-SIM- TG) is not as accurate as MPI-Sim with direct execution (MPI-SIM-DE), the error in the performance predicted by MPI-SIM-TG was below 16% with an average error of 11.3% against the measured system. 4 The message optimizations further introduced do not modify the underlying communication model and thus do not affect validation. Validation of MPI-SIM for Tomcatv2060100140 number of processors runtime (in sec) measured Figure 3: Validation of MPI-Sim for (2048-2048) Tomcatv (on the IBM SP). Figure 4 shows the execution time of the model for Sweep3D with a total problem size of 150-150-150 grid cells as predicted using MPI-SIM-TG, MPI-SIM-DE, as well as the measured values, all for up to 64 processors. The predicted and measured values are again very close and differ by at most 9.8%. On average, MPI_SIM_DE differed from the measured value by 3.7% and MPI-SIM-TG by 7.2%. number of processors runtime (in sec) measured Figure 4: Validation of Sweep3D on the IBM SP, Fixed total Problem Size. Finally, we validated MPI-SIM-TG on the NAS SP benchmark. The task times were obtained from the 16 processor run of the class A, the smallest of the three built-in sizes (A, B and C) of the benchmark, and used for experiments with all problem sizes. Figures 5 and 6 show the validation for class A and the largest size, class C. The validation for class A is good (the errors are less than 7%). The validation for class C is also good with an average error of 4%, even though the task times were obtained from class A. This result is particularly interesting because, for programs of the same size, class C on average runs 16.6 times longer than class A. This demonstrates that the compiler-optimized simulator is capable of accurate projections across a wide range of scaling factors. Furthermore, cache effects do not appear to play a great role in this code or the other two applications we have examined. This is illustrated by the fact that the errors do not increase noticeably when the task times obtained on a small number of processors were used for a larger number of processors. Validation for SP Class A100300500700 number of processors runtime (in sec) measured Figure 5: Validation for NAS SP, class A on the IBM SP. Validation for SP class C5001500250016 36 64 100 number of processors runtime in seconds measured Figure Validation for NAS SP, class C on the IBM SP. Figure 7 summarizes the errors that MPI-SIM-TG incurred when simulating the three applications. All the errors are within 16%. The figure emphasizes that the compiler-supported approach combining analytical model and simulation is very accurate for a range of benchmarks, system sizes, and problem sizes. It is hard to explore these errors further without detailed analysis of each application. Therefore, to better quantify what errors can be expected from the optimized simulator, we used our SAMPLE benchmark, which allows us to vary the computation to communication ratio as well as the communication patterns. %Error between MPI-SIM-TG Predictions and the Measured number of processors Tomcatv Sweep3D(150cubed) Figure 7: Percent Error Incurred by MPI-SIM-TG when Predicting Application Performance. was validated on the Origin 2000. Two common communication patterns were selected: wavefront and nearest neighbor. For each pattern, the communication to computation ratio was varied from 1 to 100 to a ratio of 1 to 1. Figure 8 plots the total execution time for the program and MPI-SIM-TG prediction. In order to demonstrate better the impact of computation granularity on the validation, Figure 8 plots the percentage variation in the predicted time as compared with the measured values. As can be seen from the figure, the predictions are very accurate when the ratio of computation to communication is large, which is typical of many real-world applications. As the amount of computation granularity in the program decreases, the simulator incurs larger errors. This can expected because both measurement errors and task time estimation errors can become relatively more significant. Nevertheless, the graph shows that the predicted values differ by at most 15% from the measured values, even for small communication to computation ratios. Validation of SAMPLE Measured vs. Predicted with Optimization on Origin 2K50015000.01 0.0125 0.0167 0.025 Communication to Computation Ratio Time in seconds Wvfrnt-Measured NN-Measured Figure 8: Validation of SAMPLE on the Origin 2000. Percent variation of measured time from predicted time5150.01 0.03 0.10 0.30 0.50 0.70 0.90 Communication to Computation ratio Difference wvfrnt nn Figure 9: Effect of Communication to Computation Ratio on Predictions. The accuracy of MPI-SIM-TG for large computation to communication ratio (below 5% error) indicates that the slightly higher errors we observed for Tomcatv, Sweep3D and NAS SP must be due to the presence of small computation to communication ratios. 5.3 Expanding the simulator to larger systems and problem sizes. The main benefit of using the compiler-generated code is that we can decrease the memory requirements of the simplified application code. Since the simulator uses at least as much memory as the application, decreasing the amount of memory for the application decreases the simulator's memory requirements, thus allowing us to simulate large problem sizes and systems. Number of processors total memory use total memory use Memory Reduction Factor Sweep 3D, 4-4-255 per Proc. Problem Size 4900 2884MB 30MB 96 Sweep 3D, 6-6-1000 per Proc. Problem Size 6400 215GB 122MB 1762 Tomcatv, 2048-2048 4 236MB 118.4KB 1993 Table 1: Memory Usage in MPI-SIM-DE and MPI-SIM-TG for the benchmarks. Table 1 shows the total amount of memory needed by MPI-Sim when using the analytical (MPI-SIM-TG) and direct execution (MPI-SIM-DE) models. For Sweep3D, with 4900 target processors, the analytical models reduce memory requirements by two orders of magnitude for the 4-4-255 per processor problem size. Similarly, for the 6-6-1000 problem size, the memory requirements for the target configuration with 6400 processors are reduced by three orders of magnitude! Three orders of magnitude reduction is also achieved for Tomcatv, while smaller reductions are achieved for SP. This dramatic reduction in the memory requirements of the model allows us to (a) simulate much larger target architectures, and (b) show significant improvements in execution time of the simulator. To illustrate the improved scalability achieved in the simulator with the compiler-derived analytical models, we consider Sweep3D. In this paper, we study a small subset of problems that are of interest to application developers. They are represented by the 20 million cell total problem size, which can be divided into 4-4-255, 7-7-255, and 28-28-255 per processor problem sizes which need to run on 4,900, 1,600 and 100 processors, respectively. The scalability of the simulator for the 4-4-255 problem size can be seen in Figure 10. The memory requirements of the direct execution model restricted the largest target architecture that could be simulated to 2500 processors. With the analytical model, it was possible to simulate a target architecture with 10,000 processors. Since the application's predicted runtime for 10,000 processors is 11.0955 seconds and the runtime of the simulator for that configuration is 148.118 seconds, the simulator's slowdown is only 13.35! Note that instead of scaling the system size, we could scale the problem size instead (for the same increase in memory requirements per process), in order to simulate much larger problems. Validation and Scalability of Sweep3D (4x4x255/proc)26101 10 100 1000 10000 number of processors runtime (in sec) Measured Figure 10: Scalability of Sweep3D for the 4-4-255 per Processor Size (IBM SP). 5.4 Performance of MPI-Sim The benefits of compiler-optimized simulation are not only evident in memory reduction but also in improved performance. We characterize the performance of the simulator in four ways: 1. performance gains when using the message optimization (MPI-SIM-TGMO) and MPI-SIM-TG as compared to MPI-SIM-DE, 2. absolute performance (i.e., total simulation time) of MPI-SIM-TG vs. MPI-SIM-DE and vs. the application, 3. parallel performance of MPI-SIM-TG, in terms of both absolute and relative speedups, and 4. performance of MPI-SIM-TG when simulating large systems on a given parallel host system. Effect of Optimizations on Simulator's Performance To illustrate the performance improvements between MPI-SIM-DE, MPI-SIM-TG, which takes advantage of only the local optimizations and MPI-SIM-TGMO, which additionally optimizes the messages being sent, we conducted experiments on the three benchmarks. In case of Sweep3D we compared the performance of the three versions of the simulator when each had a given number of host processors available. The problem size per processor was fixed, and the number of target processors in the experiment was increased. This study demonstrates the ability of each simulator to efficiently simulate large problem sizes. For NAS SP, since the problem size of the application is given (here class C), we fixed the number of target processors and varied the number of host processors available to the simulator. This study illustrates not only the relative performance of the simulators, but also their ability to use computational resources. Figures 11, 12 and 13 show the performance of MPI-SIM-TGMO, MPI-SIM-TG and MPI-SIM-DE when simulating Sweep3D for three sizes per processor sizes: 7-7-255, 14-14-255 and 28-28-255. All simulators use host processors to simulate up to 4,900 target processors. The improvements in performance between MPI- SIM-DE and MPI-SIM-TG for the above sizes are on the average 39.7%, 67.28% and 88.07% respectively. As the problem size per processor grows larger, the amount of computation per processor increases thus the amount of computation abstracted away increases resulting in runtime savings. 7x7x255 Per Processor Size, 64 Host Processors20060010 100 1000 10000 target processors runtime in sec. MPI-SIM-TGMO Figure 11: Sweep3D, 7x7x255 Per Processor Size, (MPI-SIM-TGMO is MPI-SIM-TG+ the message optimization). 14x14x255 Per Processor Size, 64 Hosts2006001000 target processors runtime in sec MPI-SIM-TGMO Figure 12: Sweep3D, 14-14-255 Per Processor Size. 28x28x255 Per Processor Size, 64 host procs200600100010 100 1000 10000 target procs runtime in sec MPI-SIM-TGMO Figure 13: Sweep3D, 28-28-255 Per Processor Size. Although the biggest performance gain is in the computation optimization, reducing the size of the messages sent, where possible, is beneficial. The simulation, MPI-SIM-TGMO, runs faster than the simulation, which just optimizes the computation (MPI-SIM-TG). The improvements for the sizes 7-7-255, 14-14-255 and 28-28-255 are 28.04%, 31.23% and 13.9% respectively. The benefits of the message optimizations are limited for the Sweep3D application, because it uses a large number of barrier synchronizations as well as collective operations such as (MPI_Allreduce). These operations either take no data or only single data items. We also observed great performance improvements for the NAS SP benchmark, class C, the largest size available in the suite. Figures 14 and 15 show the performance of MPI-SIM-TG and MPI-SIM-TGMO for two target processor configurations: 16 and 64. The simulations were run on a variety of host processors from 1 to 64. First, both MPI-SIM-TG and MPI-SIM-TGMO ran faster than the actual application. The measured runtime of the application executing on 16 processors is 2623.38 seconds, whereas running on 64 processors it is 790.67 seconds. Additionally, Figures 14 and 15 illustrate that the simulation can run an order of magnitude faster than MPI-SIM- when the message optimization is used. In Figure 14, the jump in runtime for MPI-SIM-TG (from 1 to 2 host processors) is due to the large communication costs. The size of the messages sent between processors is 605,161 doubles. Therefore the cost of sending these messages increases considerably when more than one processor is used. When only 2 host processors are used this increased cost is not compensated by the increased computational power. However, as the number of host processors increases, better performance is achieved. Since the size of these large messages can be reduced to 0 in the MPI-SIM-TGMO simulation, this communication overhead is significantly reduced and the simulator performs substantially better than MPI-SIM-TG. As the number of target processors increases (to 64 in Figure 15), the size of the messages in the simulation is reduced (to 370,441 for the target processor code.) Still, using the message optimization results in an order of magnitude decrease in the simulator's runtime. Absolute Performance, Local Code Optimization Only To compare the absolute performance of MPI-Sim, we gave the simulator as many processors as were available to the application (#host processors = # target processors). class C2006001000 host processors runtime in sec. MP I-S IM-TG MP I-S IM-TGMO Figure 14: A 16 Target Processor Simulation of NAS SP, Class C Running on Various Number of Host Processors. Processors , NAS SP C lass C100300500700 host Processors runtime in seconds MP I-S IM-TG MP I-S IM-TGMO Figure 15: A 64 Target Processor Simulation of NAS SP, Class C Running on Various Number of Host Processors. Figure shows the absolute performance for Sweep3D with a total problem size of 150 3 . MPI-SIM-DE is on the average 2.8 times slower than the actual application (Measured in the Figure). However, MPI-SIM-TG is initially faster then the measured application starting at 13 times faster when running on 4 processors, gradually becoming only 2.2 times faster for processors and finally being twice as slow as the application running on 64 processors. Message optimizations present in MPI-SIM-TGMO further decrease the simulators' runtime by on the average 18% as compared to MPI-SIM-TG. Both MPI-SIM-TG and MPI-SIM-TGMO are always faster (on the average and 18.5 times faster respectively) than MPI-SIM-DE, showing the clear benefits of compiler optimizations. However, as the number of processors increases the amount of communication relative to the computation increases thus exposing the overhead of simulating the communications and making MPI-SIM-TG and MPI-SIM- TGMO slower than the application. cubed Sweep3D, Total Problem Size1010000 number of processors Runtime in seconds Measured MPI-SIM-TGMO Figure Absolute Performance of MPI-Sim for Fixed Total Problem Size Sweep3D. (Vertical Scale is Logarithmic) Figure 17 shows the runtime of the application and the measured runtime of the two versions of the simulator running NAS SP class A. We observe that MPI-SIM-DE is running about twice slower than the application it is predicting. However, MPI-SIM-TG is able to run much faster than the application, even though detailed simulation of the communication is still performed. In the best case (for 36 processors), it runs 2.5 times faster. For 100 processors, it runs 1.5 times faster. The relative performance of MPI-SIM-TG decreases as the number of processors increases because the amount of computation in the application decreases with increased number of processors and thus the savings from abstracting the computation are decreased. Absolute Performance of MPI-Sim for NAS SP206010014030 50 70 90 Number of processors Runtime in Seconds Measured Figure 17: Absolute Performance of MPI-Sim for the NAS SP Benchmark, class A. Even more dramatic results were obtained with Tomcatv, where the runtime of MPI-SIM-TG does not exceed 2 seconds for all processor configurations as compared to the runtime of the application which ranges from 130 to seconds (Figure 18). This is due to the ability of the compiler to abstract away most of the computation. All that the simulator needs to directly execute is the skeleton code that controls the flow of the computation and communication patterns. Absolute Performance of MPI-Sim for Tomcatv20601001400 number of processors runtime (in seconds) application Figure Absolute Performance of MPI-Sim for Tomcatv (2048x2048). Parallel Performance To evaluate the parallel performance of the simulator, we study how well can it take advantage of increasing system resources (her processors) to solve a given problem (fixed total problem size). Figures 14 and 15 indirectly demonstrate the performance of the simulator; to illustrate the performance better, the speedup achieved for the 16 target configuration is depicted in Figure 19. Although MPI-SIM-TGMO, has a smaller runtime than MPI-SIM-TG, it scales well for only up to 8 host processors. This is because, as the number of host processors increases, the communication overhead between the host begins to dominate the runtime. On the other hand, MPI- SIM-TG, which had to send large messages, suffers most when more than one host is used, but then is able to distribute that overhead among more processors. 16Target NAS SP, Class C0.51.52.53.5 number of host processors MPI-SIM-TGMO Figure 19: Speedup of MPI-Sim for NAS SP. Clearly, the performance of the simulator is better when larger systems are simulated. For the 64-target processor case ( Figure 15), the runtime decreases steadily as the number of processors is increased. However, using more than host processors actually increases the simulator's runtime. (64 Target Class C could not be run on a single processor due to memory constraints, so direct speedup comparisons are not possible.) Better scalability is seen for the Sweep3D application. Figure 20 shows the performance of MPI-SIM-TG and MPI-SIM-DE simulating the 150 3 Sweep3D running on 64 target processors when the number of host processors is varied from 1 to 64. The data for the single processor MPI-SIM-DE simulation is not available because the simulation exceeds the available memory. Clearly, both MPI-SIM-DE and MPI-SIM-TG scale well. The speedup of MPI-SIM-TG is also shown in Figure 21. The steep slope of the curve for up to 8 processors indicates good parallel efficiency. For more than 8 processors the speedup is not as impressive, reaching about 15 for 64 processors. This is due to the decreased computation to communication ratio in the application. Still, the runtime of MPI-SIM-TG is on the average 5.4 times faster than that of MPI-SIM-DE. Runtime of S imu lator Vs. Application (150x150x150 Sweep3d , 64Target proc)100300500700 number host processors runtime (in sec) MP I-SIM -DE MP I-SIM -TG Measured Figure 20: Parallel Performance of MPI-Sim. Speedup of MPI-SIM-TG (150cubed Sweep3D, 64 Target Processors)515 berofprocessors speedup MPI-SIM-TG Figure 21: Speedup of MPI-SIM-TG for Sweep3D. Performance for Large Systems To quantify further the performance improvement for MPI-SIM-TG, we have compared the running time of the simulators when predicting the performance of a large system; in this case we want to simulate a billion-cell problem for Sweep3D. This application's developers envision this problem to utilize 20,000 processors, which corresponds to a 6-6-1000 per processor problem size. Figure 22 shows the running time of the simulators as a function of the number of target processors, when 64 host processors are used. The problem size is fixed per processor, so the problem size increases with the increased number of processors. The figure clearly shows the benefits of the optimizations. In the best case, when the performance of 1,600 processors is simulated (corresponding to the 57.6 million problem size) the runtime of the optimized simulator is nearly half the runtime of the original simulator. However, even with the optimizations, the memory requirements are still too large to be able to simulate the desired target system. MPI-SIM runtime for the 6x6x1000 per processor size host processors)20060010000 500 1000 1500 2000 2500 3000 number of target host processors runtime in seconds Figure 22: Performance of MPI-SIM when Simulating Sweep3D on Large Systems. 6 Conclusions This work has developed a scalable approach to detailed performance evaluation of communication behavior in Message Passing Interface (MPI) and High Performance Fortran (HPF) programs. Our approach is based on using compiler analysis to identify portions of the computation whose results do not have a significant impact on program performance, and therefore do not have to be simulated in detail. The compiler builds an intermediate static task graph representation of the program which enables it to identify program values that have an impact on performance, and also enables it to derive scaling functions for computational tasks. The compiler then uses program slicing to determine what portions of the computations are not needed in determining performance. Finally, the compiler abstracts away those parts of the computational code (and corresponding data structures), replacing them with simple, analytical performance estimates. It also flags messages for which the data transfer does not have to be performed within the simulation. All of the communication code is retained by the compiler, and is simulated in detail by MPI-Sim. Our experimental evaluation shows that this approach introduces relatively small errors into the prediction of program execution times. The benefit we achieve is significantly reduced simulation times (typically more than a factor of 2) and greatly reduced memory usage (by two to three orders of magnitude). This gives us the ability to accurately simulate detailed performance behavior of systems and problem sizes that are 10-100 times larger than is possible with current state-of-the-art simulation techniques. In our current work, we are also exploring a number of alternative combinations of modeling techniques. For example, we can use detailed simulation for the sequential tasks, instead of analytical modeling and measurement. This will not only allow to get accurate estimates of task execution times, but also enable us to study the application's performance on a processor and memory architecture different from the currently available platforms. Within POEMS, we aim to support any combination of analytical modeling, simulation modeling and measurement for the sequential tasks and the communication code. The static task graph provides a convenient program representation to support such a flexible modeling environment [5]. One potential limitation of our work is that the benefits would not be as large for applications where the parallelism and communication patterns depend extensively on intermediate results of the computations. In particular, so-called irregular applications may have this property. Evaluating the benefits for such applications requires further research, and perhaps a refinement of the techniques developed here. Another interesting direction is whether the techniques described here can be extended to other types of distributed applications (i.e., non-scientific applications) that use network communication intensively. If very fast simulation techniques could be developed for such applications, they could prove extremely valuable in controlling runtime optimization decisions such as object migration, load balancing, or adaptation for quality-of- service requirements, which are critical decisions for many distributed applications. Acknowledgements This work was supported by DARPA/ITO under Contract N66001-97-C-8533, "End-to-End Performance Modeling of Large Heterogeneous Adaptive Parallel/Distributed Computer/Communication Systems," (http://www.cs.utexas.edu/users/poems/). The work was also supported in part by the ASCI ASAP program under DOE/LLNL Subcontract B347884, and by DARPA and Rome Laboratory, Air Force Materiel Command, USAF, under agreement number F30602-96-1-0159. We wish to thank all the members of the POEMS project for their valuable contributions. We would also like to thank the Lawrence Livermore National Laboratory for the use of their IBM SP. This work was performed while Adve and Sakellariou were with the Computer Science Department at Rice University. --R "The ASCI Sweep3D Benchmark Code," "Using integer sets for data-parallel program analysis and optimization.," "POEMS: End-to-end Performance Design of Large Parallel Adaptive Computational Systems," "Compiler Synthesis of Task Graphs for a Parallel System Performance Modeling Environment.," "Application Representations for a Multi-Paradigm Performance Modeling Environment for Parallel Systems," "Performance Prediction of Large Parallel Applications using Parallel Simulations," "Parsec: a parallel simulation environment for complex systems," "The NAS Parallel Benchmarks 2.0," "PROTEUS: a high-performance parallel-architecture simulator," "Optimistic simulation of parallel architectures using program executables," "Distributed simulation: a case study in design and verification of distributed programs," "The Conditional Event Approach to Distributed Simulation," "The Rice parallel processing testbed," "Multiprocessor Simulation and Tracing using Tango.," "POEMS: End-to-end Performance Design of Large Parallel Adaptive Computational Systems.," "A Distributed Memory LAPSE: Parallel Simulation of Message-Passing Programs," "Parallelized direct execution simulation of message-passing parallel programs," "FAST: a functional algorithm simulation testbed," "Functional Algorithm Simulation of the Fast Multipole Method: Architectural Implications," "Improving the Accuracy vs. Speed Tradeoff for Simulating Shared-Memory Multiprocessors with ILP Processors," "Interprocedural slicing using dependence graphs," "Transparent implementation of conservative algorithms in parallel simulation languages," "Reducing Synchronization Overhead in Parallel Simulation," "An adaptive synchronization method for unpredictable communication patterns in dataparallel programs," "Parallel Simulation of Data Parallel Programs," "MPI-SIM: using parallel simulation to evaluate MPI programs," "Asynchronous Parallel Simulation of Parallel Programs," "The Wisconsin Wind Tunnel: VIrtual Prototyping of Parallel Computers," --TR The rice parallel processing testbed Interprocedural slicing using dependence graphs PROTEUS: a high-performance parallel-architecture simulator The Wisconsin Wind Tunnel A distributed memory LAPSE Reducing synchronization overhead in parallel simulation Optimistic simulation of parallel architectures using program executables Parallelized Direct Execution Simulation of Message-Passing Parallel Programs Transparent implementation of conservative algorithms in parallel simulation languages Using integer sets for data-parallel program analysis and optimization Poems MPI-SIM Performance prediction of large parallel applications using parallel simulations Asynchronous Parallel Simulation of Parallel Programs Improving lookahead in parallel discrete event simulations of large-scale applications using compiler analysis Parsec POEMS An adaptive synchronization method for unpredictable communication patterns in dataparallel programs Compiler Synthesis of Task Graphs for Parallel Program Performance Prediction Parallel Simulation of Data Parallel Programs FAST Improving the Accuracy vs. Speed Tradeoff for Simulating Shared-Memory Multiprocessors with ILP Processors --CTR Yasuharu Mizutani , Fumihiko Ino , Kenichi Hagihara, Fast performance prediction of master-slave programs by partial task execution, Proceedings of the 4th WSEAS International Conference on Software Engineering, Parallel & Distributed Systems, p.1-7, February 13-15, 2005, Salzburg, Austria
performance modeling;parallel simulation;parallelizing compilers
589795
Local behavior of the Newton method on two equivalent systems from linear programming.
Newton's method is a fundamental technique underlying many numerical methods for solving systems of nonlinear equations and optimization problems. However, it is often not fully appreciated that Newton's method can produce significantly different behavior when applied to equivalent systems, i.e., problems with the same solution but different mathematical formulations. In this paper, we investigate differences in the local behavior of Newton's method when applied to two different but equivalent systems from linear programming: the optimality conditions of the logarithmic barrier function formulation and the equations in the so-called perturbed optimality conditions. Through theoretical analysis and numerical results, we provide an explanation of why Newton's method performs more effectively on the latter system.
Introduction Newton's method is generally accepted as an effective tool for solving a system of nonlinear It is a locally and quadratically convergent method under reasonable assumptions (see e.g. Dennis and Schnabel (Ref. 1)). It is often not fully appreciated, however, that Newton's method can exhibit significantly different local and global behavior on two equivalent systems. By equivalent systems, we refer to two systems of nonlinear equations that can be derived from one another and essentially share the same set of solutions (though some auxiliary variables/equations may be present in one but not in another). In this paper, we compare the behavior of Newton's method applied to two well-known equivalent systems of nonlinear equations associated with linear programming. The first of these equivalent systems consists of the first-order optimality conditions of the log-barrier formulation of the linear program. The second system consists of equations in the perturbed first-order optimality conditions for the linear program. Though the two nonlinear systems have essentially the same set of solutions, El-Bakry, Tapia, Tsuchiya, and Zhang (Ref. 2) show that Newton's method necessarily generates different iterates for the two systems. In this paper, we show that Newton's method applied to the perturbed optimality conditions for the linear program has a larger sphere of convergence than Newton's method applied to the optimality conditions of the log-barrier formulation of the linear program. Of these two equivalent systems, the perturbed first-order optimality conditions are widely used in interior-point methods for linear programming. However, the reasons for favoring this system have not been fully analyzed. In this paper, we provide an explanation on why the system associated with the perturbed optimality conditions is the system of choice. The paper is organized as follows. In Section 2, we present the two equivalent nonlinear systems under consideration. In Section 3, we introduce the notion of the sphere of convergence of Newton's method and provide theoretical results on the radius of the sphere of convergence of Newton's method applied to the two equivalent systems. In Section 4 we present numerical results supporting the theory we developed in the previous section. Finally, we make some concluding remarks in Section 5. 2. Two Equivalent Formulations In this section, we introduce the linear programming problem and the two equivalent nonlinear systems under consideration. We consider the linear programming problem in the standard form m. The Lagrangian function associated with problem (1) is where y are, respectively, the vectors of Lagrange multipliers associated with the equality and the inequality constraints. The first-order optimality conditions for problem (1) are 2.1. Two Equivalent Systems We derive one of the equivalent systems by formulating problem (1) in the logarithmic barrier framework. This framework, which was first introduced by Frisch (Ref. 3), consists of solving a sequence of equality constrained minimization problems with decreasing values of the barrier parameter - ? 0. For problem (1) and a given value of - ? 0, the log-barrier subproblem has the following form log x i Assume that the feasible set fx : for every value of - ? 0, there exists a unique solution x - of the log-barrier subproblem. Under mild assumptions (see e.g. Fiacco and McCormick (Ref. 4)), as - ! 0 the sequence of iterates fx converges to a solution x of problem (1), i.e. lim -!0 x The optimality conditions for the log-barrier subproblem are derived by differentiating the Lagrangian function, log where y is the vector of Lagrange multipliers associated with the equality constraints, and setting the gradient of the Lagrangian equal to zero. Then the optimality conditions are Observe that the Jacobian of FB is given by \Gamma-X \Gamma2 A T If In applications of Newton's method near the solution, the Jacobian necessarily becomes ill-conditioned as - approaches zero (see, (Ref. 5, 6)). Now we derive a nonlinear system equivalent to system (3). Consider the introduction of an auxiliary variable, z 2 R n , and define which is written equivalently as Substituting z into system (3) and adding the equation that relates x; z and - yields the system The Jacobian of F P is given by The Jacobian is nonsingular if solution where x and z contain zero components, the Jacobian may or may not be nonsingular, depending on the degeneracy of the solution. Kojima, Mizuno and Yoshise (Ref. 7) first proposed to use system (5) to solve the linear program in a primal-dual interior-point method. El-Bakry, Tapia, Tsuchiya, and Zhang (Ref. 2) show that although systems (3) and (5) are equivalent, Newton's method necessarily generates different iterates for the two systems. Two things are worth noting in comparing the two systems (3) and (5). First, while FB and are undefined for are defined. Second, while F 0 B is dependent on -, F 0 P is not. These differences will greatly affect the behavior of Newton's method when applied to the two systems as - approaches zero. We remark that although the Jacobian F 0 does not depend on -, we will nevertheless use the present notation for the Jacobian to stress its association with system (5) for a given value of - 0. System (5) can also be obtained by considering the first-order optimality conditions (2) of the linear program and perturbing the complementarity equation, 2.2. Central Paths Assume that the strictly feasible set f(x; nonempty. Let the solution to system (3) for a particular value of - ? 0, and similarly let the solution to system (5). Then by the central path for system (3), we mean the set The set of points in CB forms a continuous path such that lim -!0 (x (Ref. 4, 8)). We remark that systems (3) and (5) are equivalent, in the sense that for - ? 0, and (x for z -(X system (5), we have lim -!0 (x system (5) the central path is defined as the set 2.3. Assumption and Notation Throughout the paper, we make use of the following assumption and notation. Nondegeneracy Assumption. Let the matrix A be of full rank m, and let (x ; y ; z ) be a primal and dual nondegenerate solution of system (2). Without loss of generality, we assume that the first m components of x are positive and the remaining (n \Gamma m) components are zero. The nondegeneracy assumption guarantees that (x ; y ; z ) is an isolated solution point in the primal-dual space. It is also well known that the pair (x ; z ) satisfies strict complementarity: x 0g. Then by the nondegeneracy assumption, ng. The matrix A will be partitioned into where A B denotes the matrix consisting of the columns of A indexed by B and similarly for AN . Note that If u is a vector, then its uppercase counterpart U will denote the diagonal matrix whose diagonal consists of the elements of u. For a vector u 2 R n , u B is the vector of the first m components of u and uN is the vector of the remaining (n \Gamma m) components of u. The quantity u 2 represents the vector u whose components are individually squared. All norms k \Delta k are assumed to be the Euclidean norm unless otherwise noted. 3. Sphere of Convergence: Analysis Standard local theory of Newton's method applied to a nonlinear system (see e.g. (Ref. 1)) provides the existence of a neighborhood about a solution in which Newton's method is well- defined. More importantly, starting from any point in the neighborhood, Newton's method guarantees convergence to the solution. For systems (3) and (5), such a neighborhood also exists about the solution for any given - ? 0 under our nondegeneracy assumption. In this section, we introduce the notion of the sphere of convergence for Newton's method. We analyze the behavior of the radius of the sphere of convergence associated with systems (3) and (5) by considering Newton's method applied to these equivalent systems as - ! 0. Under the nondegeneracy assumption, our analysis shows that the radius of the sphere of convergence of Newton's method on system (3) decreases to zero in the same order as - ! 0. However, we show the radius of the sphere of convergence of Newton's method applied to system (5) has a lower-bound estimate independent of -. These results provide a theoretical explanation on why Newton's method is more efficient on system (5) than on system (3) at least for small values of 3.1. Preliminaries We introduce the notion of the sphere of convergence for Newton's method. Then, we present lemmas to be used in our analysis for the radius of the sphere of convergence of Newton's method on system (3). We remark that the notion of the sphere of convergence is not new. Several references can be found in the literature where this notion or similar concept is used, see (Ref. 1, p. 91), for example. To conduct a rigorous study on the radius of convergence for Newton's method, we give a formal definition for the sphere of convergence below. Definition 3.1. We define the closed ball with radius r centered at v as B(v rg. Definition 3.2. For a given nonlinear system, F (v) = 0, and a solution v , the sphere of convergence of Newton's method at v is defined as the largest closed ball centered at v such that starting from any interior point in the sphere, excluding v , Newton's method (with unit steplength) is well-defined and generates a sequence that converges to v . Lemma 3.1. Consider - ? 0 and (x contained in CB . Then under the nondegeneracy assumption, there exists - so that for - - there is a ball B(x such that for any for constants C Proof. Since x strictly positive, and x exist - constants such that for - -, we have G 1 - (x -. - is an interior point of R n Such a point must also satisfy First we show that x i for are bounded away from zero. From (11) for are bounded above and below and ffi - G 1 =2, from (12) we obtain Thus Now, we show the second part of the proof. By the nondegeneracy assumption, strict complementarity (8) holds at the solution, which, together with the definition (4), implies lim Hence, for sufficiently small - and for some constants G 3 Consequently, By (12) and (13) for i 2 N we obtain Therefore, where We note that fi can be chosen so that C 3 ? Lemma 3.2. Define where under the nondegeneracy assumption, there exists ~ -, and for any is such that Lemma 3.1 holds, for constants C Proof. Consider - is such that Lemma 3.1 holds. Without loss of generality, consists of the columns A i of A with i 2 B. Similarly, we can define Substituting in the definition of P we obtain Now, introduce the m \Theta (n \Gamma m) matrix R where Then P can be partitioned as follows Applying the bounds in (9) to (16), we obtain kRk - C 5 - for a constant C 5 ? 0. Since - such that for all - ~ -, we obtain kRR T k ! 1. Then using the Neumann series on (I m +RR from (17) that for - and constants C 3.2. Sphere of Convergence for System (3) We provide a tight result showing that the radius of the sphere of convergence of Newton's method on system (3) decreases to zero in the same order that - ! 0. Our result follows from showing that a lower-bound and an upper-bound of order - exist for the radius of the sphere of convergence. Lemma 3.3. Under the nondegeneracy assumption, there exist ~ such that for any - ~ -, the radius of the sphere of convergence, r B (-), of Newton's method satisfies Proof. We will prove the above result by showing that the sequence of Newton iterates converges to the solution (x the initial point x 0 satisfies Consider ~ - given in Lemma 3.2. Assume Newton's method is applied to system (3) for a particular value of - ~ -. Denote (x; y) as the current Newton iterate where x 2 B(x and x satisfies the conditions given in (9). Now, consider the next Newton iteration x y Using the fact that FB (x x y -(X By Taylor's Theorem, for some - Substituting (20) into (19) we obtain where Making the above substitution for and multiplying the right-hand-side of (21) we obtain Using the definition of P in (14), we rewrite We now consider first the vector - ) in (22). If we partition its basic and nonbasic components and use the notation for P in (14), then which leads to Applying the bounds given in (9) and (15) to the above, we obtain for some constant C ? 0. Recall that It follows from (23) that if the initial iterate satisfies ae fi;C oe then the x-component of the Newton iteration sequence will converge to x - . Now, consider the remaining m components of (22). Taking the norm and partitioning matrices, we obtain Applying (9), we have Then it follows that for some constant - Thus, the y-component of the Newton iteration sequence converges to y holds. In view of (24) and (25), we conclude that the Newton iteration sequence converges to (x the initial iterate x 0 satisfies (18) for all - ~ - and for K 1 defined as the constant in the right-hand side of (24). 2 The above lemma shows that the radius of the sphere of convergence of Newton's method It establishes only a lower-bound result for the radius of the sphere of convergence of system (3). To establish that the radius of the sphere of convergence decreases to zero at exactly the same order as - ! 0, we need an upper-bound of the same order. The following lemma establishes such an upper-bound. Lemma 3.4. Consider Newton's method applied to system (3). There exist constants - and such that for any given - -, the radius of the sphere of convergence, r B (-), corresponding to this - satisfies Proof. It suffices to show the existence of a point x - 0 with Newton's method does not converge or is not defined. From Lemma 3.1 there exist - constant such that for - - and for i 2 N , (x Consider an i 2 N , and let where e i is the ith canonical vector. Obviously, Newton's method is not defined at x. Therefore, r B (- K 2 -. 2 Now we are ready to give the main result for system (3). Theorem 3.1. There exist constants ~ - ? 0 and K 1 such that for - ~ -, the radius for the sphere of convergence, r B (-), of Newton's method applied to system (3) satisfies Proof. Application of Lemma 3.3 and Lemma 3.4 produces the result. 2 Since system (3) is not well-defined in a neighborhood of the solution for it is not surprising that as - ! 0, the sphere of convergence would decrease to zero. However, it was previously not known that the radius of the sphere of convergence would decrease to zero at exactly the same rate as - goes to zero. For the log-barrier formulation of the nonlinear program with inequality constraints, S. Wright (Ref. lower-bound result for the radius of the sphere of convergence. In (Ref. 9), it is shown that there exists a - such that for - convergence to the solution x - can be obtained from any point x 0 that satisfies C- ff (26) for constant - In the case of linear programming, our result for system (3) is tight and shows that the radius of the sphere of convergence decreases in the same order as our results. 3.3. Sphere of Convergence for System (5) We now give a lower-bound estimate for the radius of the sphere of convergence of Newton's method on system (5), which is independent of the value of -. This result shows that the sphere of convergence is bounded away from zero as - ! 0. Proposition 3.1. Under the nondegeneracy assumption, there exist constants R ? 0 and ~ such that for any - ~ -, the radius of the sphere of convergence, r P (-), of Newton's method satisfies Proof. We will show that Newton's method applied to system (5) generates iterates that converge to the solution (x the initial point which then implies that r P (- R ? 0. At a given value of -, let (x; respectively the current iterate and the solution of Newton's method applied to system (5). Since F 0 by continuity there exist positive constants j and D such that is independent of -. choose ~ - such that for all - ~ -, x y z Now let (x; -. Then x y z Hence, for such chosen - and (x; y; z), the Jacobian F 0 is nonsingular and satisfies in view of (28). The Newton iterates are of the formB B @ z +C C A =B B @ x y Hence, It follows from (29) that if the initial iterate for any value of - 2 (0; ~ -), Newton's method converges to the solution (x a lower bound estimate for the radius of the sphere of convergence of Newton's method is which is independent of -. 2 Our analysis shows that the radius of the sphere of convergence is independent of - and thus stays bounded away from zero as - ! 0. This result indicates that the sphere of convergence associated with system (5) would eventually be larger than the sphere of convergence associated with system (3); that is, at least for small - values, r B (- r P (-). In the next section, we show numerically that this is indeed the case. 4. Sphere of Convergence: Numerical Results In Section 3, we provided bounds on the radii of the spheres of convergence of Newton's method on systems (3) and (5) under the nondegeneracy assumption. Our analysis shows that at least for small values of -, the sphere of convergence for system (5) is larger than that for system (3). In this section, we try to compute numerical upper-bound estimates on the radii of the spheres of convergence for Newton's method on systems (3) and (5). The purpose of these computations is not only to confirm our theory for nondegenerate problems for small values of -, but also to obtain empirical information on degenerate problems and for relatively large values of -. 4.1. Description of the Numerical Experiments We note that (a) the variables are (x; y) for system (3), and (x; (z (c) the variable y appears linearly in both systems (3) and (5). For the sake of comparison, we will only estimate the radii of the spheres of convergence for both systems in the x-space, using a fixed initial point for y. More specifically, for any given - ? 0 and any chosen initial point x 0 , we set y In the rest of the section, the term "sphere of convergence" is always restricted to the x-space only. Our upper-bound estimates are based on the following simple idea. Let x ff 2 R n be an arbitrary unit vector and - ? 0 be a scalar. Consider applying Newton's method to systems (3) and (5) starting from initial points of the form and with z (5). If for - ff ? 0, Newton's method does not converge to (or to (x -(X system (5)), then obviously - ff is an upper bound for the radius of the sphere of convergence of Newton's method at x - . This upper bound is the tightest possible in this particular direction if Newton's method converges to x - for any - 2 (0; - ff ). Numerically, this upper bound - ff can be approximated by gradually increasing - from zero by a small increment until Newton's method fails to converge. We can generate a tighter upper bound by calculating - ff for a set of random unit vectors fx ff g, and then taking min ff f- ff g as an upper bound. Under the nondegeneracy assumption, for system (5) Newton's method is well-defined in a neighborhood of the solution to the linear program, which includes negative values for x and z. Therefore, we can choose x ff to be any unit random vector. In our experiments on system (5), ten unit random vectors x ff are selected using the Matlab function randn followed by a normalization. As we mentioned earlier, because of the presence of the term X \Gamma1 , system (3) is not well-defined nor is Newton's method in any neighborhood of the solution to the linear program. This fact implies that the sphere of convergence of Newton's method shrinks to zero However, it is not clear at all that the largest half-sphere inside the positive orthant where Newton's method is well-defined and convergent should also shrink to zero as - ! 0. To be fair to system (3), we use only positive unit random vectors x ff . In this way, we actually estimate an upper bound for the radius of the "half-sphere" of convergence instead of the sphere of convergence. In our experiments on system (3), ten positive unit random vectors x ff are selected using the Matlab function rand followed by a normalization. To observe the behavior of the radii of the half-sphere of convergence for system (3) and of the sphere of convergence for system (5) as - ! 0, the numerical procedure described above was performed for a set of values of - ? 0: We include large values of in order to see the behavior of the radius of the sphere of convergence of Newton's method when far from the solution at The parameter - in (30) was given an initial value of 10 \Gamma10 and was incremented when the convergence criteria was satisfied at some iteration k, where v Such residual definitions were designed to prevent the stopping criterion from being in favor of one system or another. Nonconvergence was recorded for a particular run with a given - value and initial point of the form given in (30) if the maximum number of iterations, which we set to 50, was reached. The convergence tolerance was set to . The numerical solution v was obtained by solving system (5) with a given value of - in (31) and with a stopping tolerance of 10 \Gamma8 . In particular, system (5) was solved using an interior-point primal-dual method. For the given set of - values, the estimates for the radii of the half-sphere or sphere of convergence were recorded as min ff f- ff g, where by our construction - - k, and We emphasize that in these experiments, we always used the pure Newton's method with the unit step-length. In our implementation, we used a plain partial-pivoting Gaussian elimination (Matlab back- slash) to solve all linear equations in computing Newton directions for both systems (3) and (5). This should minimize the effect of ill-conditioning caused by different elimination schemes that exploit sparsity. 4.2. Test Problems Test problems consisted of six randomly generated problems r1-r6 which are all nondegenerate, the Netlib nondegenerate problems: scagr7, sc50b, share1b and the Netlib degenerate prob- lems: adlittle, afiro, blend, sc50a, and share2b. For the random problems, the data were generated from a uniform distribution on the interval (0; 1) using the Matlab function rand. For a given problem, the same ten unit random vectors x ff were used for all values of - in (31). The problems were run on a Sun Ultra Sparc workstation using Matlab version 5.1. Test problem dimensions can be found in Table I, where the first nine problems are nondegenerate and the last five are degenerate. We mention that some Netlib problems are not in the standard form and have inequality constraints, and the numbers of variables shown are before the addition of slack variables. 4.3. Results for Nondegenerate Problems We present numerical results for only four nondegenerate problems since we obtained similar results for the remaining problems. Figures 1-2 show the radii of the half-sphere of convergence associated with system (3) and the sphere of convergence associated with system (5) graphed against the values of - given in (31). Figure 1 contains the graph for a random problem, and the remaining graphs show results for the Netlib problems. The results show that the radius of the sphere of convergence of Newton's method on system (5) is bounded away from zero even for - sufficiently small, but the radius of the half-sphere of convergence of Newton's method on system (3) appears to decrease to zero as - ! 0 in a linear fashion. Furthermore, our tests show a larger radius of the sphere of convergence of Newton's method on system (5) than on system (3) even before - becomes small. In the case of problem r2, the radius for system (3) is noticeably larger than that for system (5) only when - is very large. 4.4. Results for Degenerate Problems For degenerate problems, we do not have a theory for the radius of the sphere of convergence of Newton's method on either of the two systems. We hope that numerical results would provide some empirical information on the behavior of Newton's method applied to these problems. We present numerical results on four of the five degenerate problems, as shown in Figures 3-4, omitting results for the problem adlittle because they are similar to the presented results. The results show that the radius of the half-sphere of convergence of Newton's method on system (3) appears to decrease to zero as - approaches zero, as in the case with the nondegenerate problems. We observe that unlike the case of nondegenerate problems, the radius of the sphere of convergence on system (5) also appears to decrease to zero with -. In these tests, we observe that the radius of the sphere of convergence of system (5) is always larger than or equal to the radius of the sphere of convergence of system (3) for all the - values given in (31). In particular, the radius associated with system (5) stays well above that for system (3), by at least an order of magnitude, as - ! 0. 5. Conclusions In this paper, we studied the local behavior of Newton's method on two equivalent systems from linear programming: the optimality system (3) for the log-barrier formulation of the linear program and the perturbed optimality system (5) for the linear program itself. For nondegenerate problems, we have shown that the radius of the sphere of convergence of Newton's method on system (3) decreases to zero at exactly the same order as - ! 0, while the radius of the sphere of convergence associated with system (5) stays bounded away from zero as - ! 0. These theoretical results are established for exact arithmetics and hence are independent of the numerical conditioning of the Jacobian matrices for systems (3) and (5). The numerical experiments have confirmed our theoretical results. Interestingly, on the majority of our test problems the estimated radius of the sphere of convergence of Newton's method was consistently larger on system (5) than on system (3); not only for small values of -, but also for medium and large values of - for which numerical ill-conditioning does not play a critical role. There are multiple reasons why Newton's method performs more favorably on system (5) than on system (3) (see (Ref. 10) for a recent work on this subject). Contrary to previous belief, M. Wright (Ref. 11, 12) has shown that numerical ill-conditioning is not a determining factor. We believe that the results in this paper provide another fundamental reason why system (5) should be the system of choice to be used in an interior-point path-following framework. Similar results have been extended to the nonlinear program and will be reported in a subsequent paper. --R Numerical Methods for Unconstrained Optimization and Nonlinear Equations The Logarithmic Potential Method of Convex Programming Sequential Unconstrained Minimization Techniques Hessian Matrices of Penalty Functions for Solving Constrained Optimization Problems. Analytic Expressions for the Eigenvalues and Eigenvectors of Hessian Matrices of Barrier and Penalty Functions. A Primal-Dual Interior Point Algorithm for Linear Pro- gramming An Analogue of Moreau's Proximation Theorem On the Convergence of the Newton/Log-Barrier Method Why a Pure Primal Newton Barrier Step May be Infeasible. Some Properties of the Hessian in the Logarithmic Barrier Function --TR A primal-dual interior point algorithm for linear programming Some properties of the Hessian of the logarithmic barrier function On the formulation and theory of the Newton interior-point method for nonlinear programming Ill-Conditioning and Computational Error in Interior Methods for Nonlinear Programming --CTR D. C. Jamrog , R. A. Tapia , Y. Zhang, Comparison of two sets of first-order conditions as bases of interior-point Newton methods for optimization with simple bounds, Journal of Optimization Theory and Applications, v.113 n.1, p.21-40, April 2002
equivalent systems;linear programming;newton's method;sphere of convergence
589926
The dyadic stream merging algorithm.
We study the stream merging problem for media-on-demand servers. Clients requesting media from the server arrive by a Poisson process, and delivery to the clients starts immediately. Clients are prepared to receive up to two streams at any time, one or both being fed into a buffer cache. We present an on-line algorithm, the dyadic stream merging algorithm, whose recursive structure allows us to derive a tight asymptotic bound on stream merging performance. In particular, let be the Poisson request arrival rate, and let L be the fixed media length. Then the long-time ratio of the expected total stream length under the dyadic algorithm to that under an algorithm with no merging is asymptotically equal to 3/log(L)2L;. Furthermore, we establish the near-optimality of the dyadic algorithm by comparisons with experimental results obtained for an optimal algorithm constructed as a dynamic program. The dyadic algorithm and the best on-line algorithm of those recently proposed differ by less than a percent in their comparison with an off-line optimal algorithm. Finally, the worst-case performance of our algorithm is shown to be no worse than that of earlier algorithms. Thus, the dyadic algorithm appears to be the first near optimal algorithm that admits a rigorous average-case analysis.
INTRODUCTION At a sequence of random times, clients request content streaming from a given media server, e.g., videos from a video-on-demand server, with delivery for each client to begin immediately. To reduce the potentially heavy tra-c burden created by these media streams, it is clearly desirable to combine streams of the same content; this can be implemented in practice by using multicast protocols (e.g., see [28]). With a multicast protocol This work is supported by the NSF Grant No. 0092113. Journal of Algorithms, 2002, to appear. CILOVI in place, a stream sent to a client can be received by all other clients at a minimal possible usage of network resources. To see how this can be done and still preserve immediate-start delivery, we need the following as- sumptions: clients can receive two streams in parallel and each has a cache for buering stream content. Although multimedia streaming embraces video, audio, and data streaming, we will stay with video terminology for simplicity. The basic idea of stream merging can be explained on the following example. Consider a situation in which (i) client C 1 arrives at t 1 and requests a video of duration L and (ii) client C 0 is currently playing the same video from a stream S 0 that began at time t 0 < t 1 . Client C 1 missed the rst := t 1 t 0 time units of the video from S 0 and that part of the video needs to be sent to C 1 by the server in stream S 1 . However, C 1 can make use of stream S 0 by buering its content for later playback. In that way the stream S 1 can be terminated after time units. This process is called stream merging; in the present case, S 1 was discontinued after being "merged" at time with the earlier starting Note that the total streaming time has been reduced from 2L; with no merging, to a minimum achievable value of streaming time is a simple and eective measure of bandwidth consumption that we will retain throughout the paper. merging becomes much more involved as we increase the number of streams that are candidates for merging, because then the number of ways in which merging can be done also increases. For example, consider the case of three clients C 0 , C 1 , and C 2 arriving at times t 0 < and initiating streams S 0 , S 1 , and S 2 for a video of duration L. Let be the interarrival times. Figure 1 illustrates an example in which the t i 's are given by 0, 3 and 4, and Consider the ways in which we can merge the streams for all three clients. For the given setup, the two possible merging patterns are shown in Fig. 1. In Fig. 1(a), S 1 and S 2 are merged independently with S 0 as described earlier: C 1 caches S 0 during during at the end of the respective intervals S 1 and S 2 are merged with The second possibility is rst to merge S 2 with S 1 and then S 1 with S 0 . This scenario is illustrated in Fig. 1 (b). Figure 2 breaks down Fig. 1 (b) into the individual schedules for C 1 and C 2 . Client C 1 plays S 1 and caches during its buer which is only fed by S 0 during the last L 2 1 time units of the video. Client C 2 caches and plays from S 2 during at which point S 2 is discontinued, and play proceeds from C 2 's buer. Client C 2 continues to cache S 1 ; but in addition, it caches the remainder of S 0 (in a suitably chosen region of the cache where the two buering operations can not overlap). This continues 2 at which point S 1 is shut down and S 0 becomes the only (a) (b) FIG. 1. Stream merging examples. The position of the video runs diagonally. The x-axis represents time. By following the zig-zag lines one obtains which part of the video is being played from which stream. The dashed lines show where the play of the video changes from one stream to another. schedule C schedule FIG. 2. Individual schedules for the clients C 1 and C 2 in Fig. 1 (b). active stream while it is supplying the last L units of the video to the buer of C 2 . In this process, C 2 has played the rst 2 time units of the video directly from S 2 , the next time units from a cached segment of S 1 and the last L time units from a cached segment of S 0 . At any given time, a vertical line in Fig. 2 crosses each of the streams currently being received. Accordingly, in the schedules for C 1 and C 2 the bold lines incident to the vertical lines at time t indicate that the buer content at time t consists of the corresponding segments of S 0 and S 1 . CILOVI Note that, although the streaming at C 1 is the same as in the rst merging example, does not terminate at time no longer needed by the media server must still send S 1 to C 2 until C 2 can switch to which occurs at in order to facilitate the requirements of C 2 . Without such an extension C 2 would not be in a position to receive all parts of the movie. Note also that the cost (sum of stream durations) of the second merging pattern is 16 as compared to the cost 17 of the rst pattern. In general, the best merge pattern for an arrival at time t depends not only on arrival times before t; but also on the arrival times after t. As it will become clear in the next section, for this example the solution, that our dyadic tree algorithm yields, corresponds to Figure (b). The technique of stream merging originated with Eager, Vernon, and Zahorjan [11, 12] as a model of the pyramid broadcasting scheme introduced by Viswanathan and Imielinski [36,37]. This paradigm provides the multi-cast basis for sharing streams and is built upon the assumption that clients can receive more bandwidth than they need for play-out. The skyscraper broadcasting scheme [15,22,31] is another example of these new techniques. A number of related techniques go under the names of batching [1, 9, 10], patching [6, 16, 21], tapping [7, 8], and piggy-backing [2, 18, 19, 29] and the general problem has several parameters and useful performance metrics. Other parameters include delay guarantees, receiving bandwidth, server bandwidth, and buer size [5, 13{15, 17, 20, 23{27, 30{35]. The maximum number of streams is another metric that is of greater interest in certain circumstances. In this setting, the algorithm of this paper has the properties It is on-line, i.e., the media server does not know arrival times in advance It gives a zero-delay guarantee, i.e., all video requests are satised immediately. It restricts the number of streams being received by a client at any one time to at most two { the receive-two model. The buer size can accommodate up to half of the video. The last two assumptions are justied in the papers by Bar-Noy and Ladner [3, 4], which supply the primary motivation for the work here. In particular, most of the improvement of merging streams is already present in the receive-two model. The L=2 buer size limit comes about because our algorithm does not attempt merging with an existing stream that is already at least half over. As Bar-Noy and Ladner argue, this convenience does not lead to increased average cost even for only moderately large arrival rates. For further discussion of the literature on stream merging, we refer the reader to the mini-survey of [3]. Many excellent numerical/experimental studies have appeared in the stream-merging literature, but the absence of mathematical foundations has stood out, at least until the work in [3, 4], which focuses on compet- itive, or worst-case, analysis. Here, we give what appears to be the rst rigorous average-case analysis of a near-optimal algorithm. The paper is organized as follows. In Section 2 we present the dyadic tree algorithm and state our main results. Section 3 contains numerical experiments that verify the algorithm's performance and conclusions. The proofs of the main results can be found in Section 4. 2. ALGORITHM AND RESULTS The problem of stream merging can be posed as a problem on trees (see [3, 4]). A merge tree is a representation of a stream merging diagram, such as those shown in Figure 1. Each stream of the merging diagram corresponds to a node in the associated merge tree. Thus, the number of nodes in the merge tree is equal to the number of requests placed with the server, i.e., the number of clients. If stream S j is merged directly to an earlier starting stream S i ; then the node associated with S j is a child of the node associated with S i . It is convenient to label the nodes with the arrival times of the corresponding streams. A root stream is merged with no other stream, i.e., it is the root in a merge tree. The length of the root stream is always the full length of the video, L. The start rule below provides a simple way to determine which streams are roots, i.e. it denes a sequence of merge trees. Let t the stream starting times. Start rule: Node t 0 is a root. If t i is a root, then t L=2g is a root. In other words, the start rule says that a node will be in a given tree only if the root stream of that tree started less the L=2 time units ago. As noted earlier, this constraint simplies the algorithmics; there is a sacrice in e-ciency, but only when tra-c is low. For example, suppose we have a root stream starting at time t 0 and an arrival at time t 1 with made a descendant of t 0 ; then no other node arriving after can be merged with t 1 without extending its length to L. Hence, some gain is achieved only if there are no arrivals in the interval However, this is an unlikely scenario under high tra-c load. When the arrivals are Poisson, the sequence of merge trees becomes a renewal process. This fact allows us to focus our analysis on a single merge tree rooted at t 0 . Let ft n g 1 n=0 be a sample path of a Poisson process with rate on the non-negative reals, and assume for convenience that t 6 COFFMAN, JELENKOVI CILOVI I I I 14 I 2 I 5FIG. 3. Dyadic partition of the interval. The total length of all streams in a merge tree is dened as l n 1ft n < L=2g; (1) where l n denotes the length of the stream initiated by the arrival at time the indicator function 1fAg is equal to 1 if A is true and 0 otherwise. By denition l L. The quantity T will measure the eectiveness of stream merging algorithms. Our new stream merging algorithm is implicit in the following algorithm for constructing merge trees from a given root. The Dyadic Tree Algorithm: The input is a sequence of n > 0 arrival times and the output is a tree of n nodes. The arrival at time 0 determines the root. To nd the children of the root, rst divide the interval [0; L=2) into dyadic subintervals I with lengths shown in Figure 3. If I i contains at least one arrival time, then t (i) denotes the earliest such time; otherwise, Each t (i) > 0 is made a child of the root. Then for each t (i) > 0; the algorithm is applied recursively to the interval [t (i) to determine the subtree rooted at t (i) . It is not di-cult to verify that this can be formulated as an on-line algorithm, as we show at the end of this section. In particular, the decision as to where a node t i should be attached to an existing tree is unaected by arrivals after time t i . The following theorem gives our rst result, a uniform bound on total stream length. We postpone the proof until Section 4. Throughout the paper we use log to denote log 2 . Theorem 2.1. The total cost of the dyadic tree algorithm satises4 L log(L)4 L ET (L; Furthermore, it can be shown that the upper bound of the preceding theorem is asymptotically tight for large values of L. A detailed proof of the next theorem is given in Section 4. Theorem 2.2. The total cost of the dyadic tree algorithm satises lim Observe that, by Theorem 2.2, the long-time ratio of the expected total stream length under the dyadic algorithm to that under an algorithm with no merging is asymptotically equal to 3 log(L)=(2L). Here we point out that, by Lemma 1 of [3], the length l of the non-root stream initiated at time t > 0 is given by where t p is its parent and t l is the last stream that merges with it. If t is a leaf then t l = t, i.e., In order to consider the worst-case performance we examine a slightly dierent model. This modication is necessary owing to the fact that in the original model the number of requests in [0; L=2) is unbounded, so that the worst case performance is meaningless. Let time be slotted and let the video have a length of 2n time slots. We assume that in each of the slots at most one stream can be initiated. According to the start rule a merge tree is being built on n slots. The total stream length achieves its maximum when a stream is initiated in every time slot. In [4] it is proved that the worst-case performance of the optimal algorithm is (n log n). Let T (2n) denote the total stream length for the worst-case merge tree built on n slots, 1. It is easy to show by induction that T (n) is monotonic in n; hence, one can assume that n is a power of 2. Next, consider two merge trees built on n=2 slots each, i.e., 1. The key fact is that in these two cases only the lengths of streams initiated in the 0th and n=2-th slot dier. This follows from the fact that the length of the stream initiated at t depends only on t and starting times of the parent stream and the last stream that merges with it (see (2)). In the rst case the lengths of streams initiated at are 2n and 3n=2, respectively. In the second case the lengths are equal to n. Thus, the dierence is 3n=2 and one obtains T The solution to this recurrence has the form T n). Thus, the dyadic algorithm is within a constant factor of optimal in the worst-case. A more detailed numeric comparison of the dyadic algorithm and the optimal algorithm is made in the next section. We conclude this section with a straightforward on-line implementation of the algorithm. 8 COFFMAN, JELENKOVI CILOVI On-line Dyadic Stream Merging: Let S be a stack with push and pop operations dened for triples of numbers (t a Each triple corresponds to a stream: t a is the time at which the stream was initiated, t r is the time after which newly arrived streams will not be allowed to merge with it and t e is the time when the stream terminates. push the root triple (0; L=2; L) onto S. At time t < L=2 of a new request: 1. pop the triples (t a t; at this point let be the top of the stack, 2. for all but the root triple in S increase the last component by 2t is the arrival time of the parent of 3. add the new stream to the stack by performing push (t; t a g; the stream started at t is the child of the stream started at ^ t a . This procedure uniquely and explicitly denes the merge tree as well as the stream termination times. 3. NUMERICAL RESULTS AND CONCLUSIONS This section provides a numerical validation of the asymptotic approxi- mation The rst example investigates the dependency of the total cost on the length of the stream for xed values of the arrival rate . The parameter values are set within the regions that are plausible for real-life systems. In particular, we set plot the ratio ET=T 0 in Figure 4, where ET is obtained by simulating 10,000 trees for each set of values. Points marked with "o", "+" and "x" correspond to 1 equal to 5, 20, and 60 sec, respectively. Note that for the merge tree consists of only 11 nodes on average. In the second example we x L and look at ET (; L) as a function of the rst argument. The simulation results of ET=T 0 are plotted in Figure 5. As in the previous case we simulated 10,000 trees for each point. Values of L are set to 120, 60, and 30 min and denoted respectively by the symbols "o", "+", and "x". Using approximation T 0 with the appropriate multiplicative factor yields excellent engineering estimates for all reasonable values of L and . Finally, we compare the performance of the dyadic tree algorithm to the performance of the optimal o-line algorithm. The cost of the latter can be 2.5 30.10.30.50.70.9Length of the stream, hours FIG. 4. ET=T 0 as a function of the stream length for three values of the arrival rate. Expected interarrival times are 5 sec ("o"), 20 sec ("+") and 60 sec ("x"). FIG. 5. ET=T 0 as a function of the arrival rate for three values of the stream length. The stream length is set to 120 min ("o"), ("x"). determined by a dynamic program (see [2]). Let T opt (i; j) be the optimal cost of the merge tree for streams initiated at 0 t i < < t j < L=2. The optimal merge tree satises the preorder traversal property [4] and, hence, 1kn CILOVI Increase in cost, FIG. 6. Performance of the algorithm in comparison with the optimal o-line algorithm. The length of the stream is equal to 2 hours. with T opt (i; L. The last term represents the gain from a merge of optimal trees rooted at t 0 and t k . We used the fact that the length of the stream t is given by (2). For numerical comparison, let the length of the video be 2 hours and let the value of the expected interarrival time vary from 5 sec to 60 sec in steps of 5 sec. For every pair (; L) we simulated 1,000 trees and based on that computed the average cost for two algorithms. The increase in expected cost when using the dyadic tree algorithm instead of the optimal o-line algorithm is rather small as shown in Figure 6. For all parameter values the increase did not exceed 8%. In summary, we have been able to prove the tight asymptotic average-case behavior 3 4 L log(L) for the dyadic stream merging algorithm, and to show in addition that its average-case and worst-case performance are comparable to those of the best on-line algorithms known to date. 4. PROOFS We start by introducing a recursive procedure for labeling the arrival times in [0; L=2). For the purposes of the proof these labels replace the t i labels. The procedure can be thought of as a function EL : T 7! ! that maps a set T of arrival times to the space of indices !. Each index ! consists of a number of digits equal to the depth of the node in the merge tree that corresponds to the given arrival. In general, 2:::, and the parent of the node labeled ! is a FIG. 7. An illustration of the labeling algorithm. In this example there are seven points that need to be labeled. On the rst call of the procedure three points are assigned labels (1,2 and 4). The recursive algorithm is applied until all points are labeled. node labeled with the prex . The algorithm labels the arrivals as follows. The interval [0; L=2) is divided into dyadic intervals in increasing order from the root as shown in Figure 3. If a point t is the rst point in the subinterval I i then its label is i. Label the rest of the points in [t; 2 i L) recursively by using the parent's label as a prex for childrens' labels. An example of how the points are labeled is shown in Figure 7. 4.1. Proof of Theorem 2.1 Lower bound: By applying the above labeling procedure, it is not hard to verify that (1) becomes l n 1ft n < l where l !1 :::! n is the length of the stream starting at the point labeled . If for a particular realization of the Poisson process there is no point with label Next, we estimate the expected values of l !1 :::! n . Let ; fn g 1 n=1 be a set of i.i.d. exponential random variables with mean 1 , and consider rst the streams that are children of the root, i.e., the streams whose indices consist of a single digit. Given that, for a particular realization of the Poisson process, there exists a stream with label ! 1 , its length must be at least 2 !1 L=2 according to (2). Therefore, l !1 L=2 L=2 CILOVI taking into account the memoryless property of the Poisson process, we conclude that L=2 A node with label of form a child of the node with label ! 1 . Considering the preceding inequality, the recursive nature of the merging algorithm and the size of the problem in which node ! 1 is the root one obtains L=2 L=2 The recursive structure of the merging algorithm shows that for a stream with an arbitrary index with the understanding that the sum in the above expression is equal to zero if and, hence, the expected value of an individual stream length is further lower bounded by L=2 Now observe that the number of indices with a digit sum equal to k is since the above sum is equal to the number of ways one can partition a set of cardinality k. Rearrange the sum in (3), use the bound (4) and apply (5) to nd L +X L=2 L +X L=2 where the last step follows from Jensen's inequality. Finally, simple manipulations of the preceding bound yield blog L Llog(L) L from which we conclude that the lower bound holds. Upper bound: Consider the streams that are children of the root. For such streams we have by (2) l !1 3 L=2 The inequality is tight when there is an arrival right after time 2 !1 L=2 and an arrival just before time 2 !1 L. Next we examine the streams that can be reached from the root in exactly two steps. An upper bound on their length is l !1!2 3 whereupon the memoryless property of the Poisson process gives 14 COFFMAN, JELENKOVI CILOVI Note that (6) and (7) are of the same form. In the rst inequality the size of the problem is L=2 while in the second the size is (2 !1 L=2 infft n . Since the merging algorithm is recursive, for streams that have depth n 2 in the merge tree one can conclude that 3E L=2 Recall (5) in order to verify that the number of indices with the digit sum k and last digit i is equal to 2 k The length of the root stream is always L so (3), (6), (8) and (9) yield L=2 A simple computation shows that E (1 therefore, by changing the order of summation and setting one obtains Finally, straightforward but tedious calculations show which in conjunction with bound (10) and the monotonicity of the function log(L)je This concludes the proof. 4.2. Proof of Theorem 2.2 The upper bound is a direct consequence of Theorem 2.1. Below we provide the proof of the lower bound. Let P P (; the probability of having at least one Poisson arrival in an interval of length . By conditioning on an arrival in both (2 and one obtains from (2) 3L=2 Extending the above reasoning to the streams with two-digit labels yields a lower bound on their expected lengths 3L=2 In the above inequality we conditioned on the position of the stream its parent and the last stream that will merge to it. Due to the recursive structure of the algorithm, for a stream with an arbitrary label the lower bound has the following form 3L=2 CILOVI Next, the preceding inequality, (1) and (5) result in 3L=2 6 3L Lblog(L)c 6L: Finally, setting log log(L) and using log e > 1 yield lim log(L) and, therefore, blog(L)c log(L)log log(L) log(L) as L !1: This concludes our proof. ACKNOWLEDGMENT The authors would like to thank the anonymous reviewer for helpful comments. --R On optimal batching policies for video- on-demand storage servers On optimal piggyback merging policies for video-on-demand systems Competitive on-line stream merging algorithms for media-on-demand Optimizing patching performance. Improving video-on-demand server e-ciency through stream tapping Improving bandwidth e-ciency of video-on-demand servers A periodic broadcasting approach to video-on-demand service Dynamic batching policies for an on-demand video server Minimizing bandwidth requirements for on-demand data delivery Optimal and e-cient mergind schedules for video-on-demand servers Optimized regional caching for on-demand data delivery Dynamic skyscraper broadcasts for video-on-demand Supplying instantaneous video-on-demand services using controlled multicast Reducing I/O demand in video-on- demand storage servers Adaptive piggybacking: A novel technique for data sharing in video-on-demand storage servers Exploiting client bandwidth for more e-cient video broadcast Patching: a multicast technique for true video- on-demand services Skyscraper broadcasting: A new broadcasting scheme for metropolitan video-on-demand systems Fast broadcasting for hot video access. Harmonic broadcasting for video-on-demand service Staircase data broadcasting and receiving scheme for hot video service. Enhancing harmonic data broadcasting and receiving scheme fo popular video service. Fast data broadcasting and receiving scheme for popular video service. Computer Networking: A Top-Down Approach Featuring the Internet Merging video streams in a multimedia storage server: complexity and heuristics. Data broadcasting and seamless channel transition for highly-demanded videos Pyramid broadcasting for video-on-demand ser- vice Metropolitan area for video-on-demand service using pyramid broadcasting --TR Reducing I/O demand in video-on-demand storage servers Dynamic batching policies for an on-demand video server Adaptive piggybacking On optimal piggyback merging policies for video-on-demand systems Metropolitan area video-on-demand service using pyramid broadcasting Skyscraper broadcasting Merging video streams in a multimedia storage server <italic>Patching</italic> Improving bandwidth efficiency of video-on-demand servers Zero-delay broadcasting protocols for video-on-demand Optimal and efficient merging schedules for video-on-demand servers Catching and selective catching An efficient bandwidth-sharing technique for true video on demand systems Competitive on-line stream merging algorithms for media-on-demand Computer Networking Dynamic Skyscraper Broadcasts for Video-on-Demand Fast broadcasting for hot video access Supplying Instantaneous Video-on-Demand Services Using Controlled Multicast A Low Bandwidth Broadcasting Protocol for Video on Demand Exploiting Client Bandwidth for More Efficient Video Broadcast On Optimal Batching Policies for Video-on-Demand Storage Servers Video-on-Demand Server Efficiency through Stream Tapping --CTR Marcus Rocha , Marcelo Maia , talo Cunha , Jussara Almeida , Srgio Campos, Scalable media streaming to interactive users, Proceedings of the 13th annual ACM international conference on Multimedia, November 06-11, 2005, Hilton, Singapore Marcelo Maia , Marcus Rocha , talo Cunha , Jussara Almeida , Srgio Campos, Network bandwidth requirements for optimized streaming media transmission to interactive users, Proceedings of the 12th Brazilian symposium on Multimedia and the web, November 19-22, 2006, Natal, Rio Grande do Norte, Brazil Wun-Tat Chan , Tak-Wah Lam , Hing-Fung Ting , Prudence W. H. Wong, On-line stream merging in a general setting, Theoretical Computer Science, v.296 n.1, p.27-46, 4 March Amotz Bar-Noy , Richard E. Ladner, Competitive on-line stream merging algorithms for media-on-demand, Journal of Algorithms, v.48 n.1, p.59-90, August Raj Kumar Rajendran , Dan Rubenstein, Optimizing the quality of scalable video streams on P2P networks, Computer Networks: The International Journal of Computer and Telecommunications Networking, v.50 n.15, p.2641-2658, October 2006
video-on-demand;average-case analysis;stream merging
590354
Symbolic representation of user-defined time granularities.
In the recent literature on time representation, an effort has been made to characterize the notion of time granularity and the relationships between granularities. The main goals are having a common framework for their specification, and allowing the interoperability of systems adopting different time granularities. This paper considers the mathematical characterization of finite and periodic time granularities, and investigates the requirements for a user-friendly symbolic formalism that could be used for their specification. Instead of proposing yet another formalism, the paper analyzes the expressiveness of known symbolic formalisms for the representation of granularities, using the mathematical characterization as a reference model. Based on this analysis, a significant extension to the collection formalism defined in [15] is proposed, in order to capture a practically interesting class of periodic granularities.
Introduction There is a wide agreement in the AI and database community on the requirement for a data/knowledge representation system of supporting standard as well as user-defined time granularities. Examples of standard time granularities are days, weeks, months, while user defined granularities may include business- weeks, trading-days, working-shifts, school-terms, with these granularities having different definitions in different application contexts. The work in [3, 4] represents an effort to formally characterize the notion of time granularity and the relationships between granu- larities, in order to have a common framework for their specification and to allow the interoperability of systems adopting different time granularities. The formal definition, however, is essentially a mathematical characterization of the granules, and it is not suitable for presentation and manipulation by the common user. The goal of this paper is identifying an intuitive formalism which can capture a significant class of granularities within the formal framework and which is closed for this class with respect to the operations it allows. This class can be intuitively described as containing all finite granularities, as well as all periodical ones. Instead of inventing yet another symbolic formalism for this purpose, in this work we consider some existing proposals, analyzing their expressiveness with respect to our goal. A symbolic formalism, based on collections of temporal intervals, was proposed in [11] to represent temporal expressions occurring in natural language and used in several application domains like appointment scheduling and time management. This formalism has been adopted with some extensions by many researchers in the AI [9, 15, 6] and Database area [8, 5]. From the deductive database community, a second influential proposal is the slice formalism introduced in [14], and adopted, among others, in [2]. None of these formalisms and extensions seems to have the expressive power we are seeking, despite some of the proposals include features that go beyond what is needed in our framework. For example, [6] provides a powerful formalism to represent calendars and time repetition, including existential and universal quantification, which supports the representation of uncertainty, a feature not considered in our frame- work. Moreover, some calendar expressions in [6] go beyond the specification of granularities, as defined in [4, 3] and in this paper, allowing the representation of overlapping granules of time. The formalism can represent recurring events in the form of non-convex in- tervals, but it does not seem to be able to represent what in the following we call gap-granularities, where gaps may not only occur between one granule and the next, but also within granules. A business-month seen as an indivisible time granule defined as the union of all business-days within a month is an example. Relevant work on non-convex intervals and repetition includes [10, 13], but the emphasis in these works is more on reasoning with qualitative relations than on calendar expression representation. In addition to the research cited above, significant work on time granularity includes [16, 12, 7]. The contribution of this paper is twofold: on one side we give results on the expressiveness of the formalisms proposed in [11] and [14] which we identify as the two basic approaches to symbolic representa- tion, while, on the other side, we propose an extension to one of these formalisms that allows to capture exactly the class of finite and infinite periodical granularities we defined in [3]. In the next section we introduce the formal notion of time granularity. In Section 3 we briefly describe the collection and slice symbolic representation formalisms, and we evaluate their expressiveness and formal properties. In Section 4, we propose an extension to the collection formalism to capture gap- granularities, and we conclude the paper in Section 5. Appendix A summarizes the syntax of the symbolic formalism, and Appendix B contains the proofs of the results in the paper. 2. Characterization of time granularities In this section we introduce the mathematical characterization of time granularities as proposed in [4] and further refined and summarized in [3]. Granularities are defined with respect to an underlying time domain, which can be formally characterized simply as a set whose elements are ordered by a relation- ship. For example, integers (Z; ), natural numbers rational (Q; ), and real numbers (R; ) are all possible choices for the time domain. granularity is a mapping G from the integers (the index set) to subsets of the time domain such and G(j) are non- empty, then each element of G(i) is less than all elements of G(j), and (2) if are non-empty, then G(k) is non-empty. The first condition states that granules in a granularity do not overlap and that their index order is the same as their time domain order. The second condition states that the subset of the index set that maps to non-empty subsets of the time domain is contiguous. While the time domain can be discrete, dense, or con- tinuous, a granularity defines a countable set of gran- ules, each one identified by an integer. The index set can thereby provide an "encoding" of the granularity in a computer. The definition covers standard granularities like Days, Months, Weeks and Years, bounded granularities like Years-since-2000, granularities with non-contiguous granules like Business-Days, and gap-granularities, i.e., granularities with non-convex intervals as granules like Business-Months. As an example of the encoding, Years-since-2000 can be defined as a mapping G, with G(1) mapped to the subset of the time domain corresponding to the year 2000, G(i + 1) to the one corresponding to the year 2001, and so on, with Independently from the integer encoding, there may be a "textual representation" of each non-empty granule, termed its label, that is used for input and output. This representation is generally a string that is more descriptive than the granule's index (e.g.,"August 1997", "1/2/2000", etc. Among the many relationships between time granularities (see [4]), the following defines an essential concept for this paper. periodical with respect to a granularity G if 1. For each i 2 Z there exists a (possibly infi- nite) subset S of the integers such that 2. There exist R; P than the number of non-empty granules of H , such that for j2S G(j) and H(i+R) 6= The first condition states that any non-empty granule H(i) is the union of some granules of G; for instance, assume H(i) is the union of the granules The periodicity property (condition 2 in the definition) ensures that the R th granule after H(i), i.e., H(i non-empty, is the union of G(a This results in a periodic "pattern" of the composition of R granules of H in terms of granules of G. The pattern repeats along the time domain by "shifting" each granule of H by P granules of G. P is also called the "period" of H . The condition on R enforces that at least one granule of H is a periodic repetition of another granule. A granularity H which is periodical with respect to G is specified by: (i) the R sets of indexes of G describing the non-empty granules of H within one period; (ii) the value of P ; (iii) the indexes of first and last non-empty granules in H , if their value is not infinite. Then, if S are the sets of indexes of G describing spectively, then the description of an arbitrary granule H(j) is given by 1 S Many common granularities are in this kind of relationship, for example, Years is periodical with respect to both Days and Months. Business-Months is periodical with respect to Business-Days, which in turn is periodical with respect to Days. Most practical problems seem to require only a granularity system containing a set of time granularities which are all periodical with respect to a basic granularity. Usually Days, Hours, Seconds or Microseconds take this role, depending on the accuracy required in each application context. In this paper, for simplicity, we assume there is a fixed basic granularity covering the whole time domain. Definition 3 We say that a granularity G is periodical if it is periodical with respect to the basic granularity. In Figure we represent the whole set of granu- larities, according to Definition 1, partitioned in two main subsets: those having all granules with contiguous values (NO-GAP) and those admitting granules with non-contiguous values (GAP). The inner circle identifies finite and periodical granularities: finite granularities are divided (dash line) into finite irregular and finite periodical 2 while infinite periodical granularities are divided into those having a first non-empty granule and no last granule (INFINITE-R), those having a last non-empty granule and no first granule (INFINITE-L), and those infinite on both sides (INFINITE). This classification will be useful when considering the expressive power of symbolic formalisms 3. Two approaches to symbolic representa- tion In this section we first remind the syntax and semantics of collection and slice formalisms, and then analyze their expressiveness with respect to the class of periodical granularities. 3.1. Collections and slices The temporal intervals collection formalism was proposed in [11]. A collection is a structured set of intervals where the order of the collection gives a measure of the structure depth: an order 1 collection is This formula is correct provided that no granule of H is empty, but it can be easily adapted to the case with finite index for first and last non-empty granules. 2 Despite this formal distinction, finite granularities will be treated uniformly in the results. GAP NO-GAP INFINITE-R INFINITE-L Figure 1. A classification of time granularities an ordered list of intervals, and an order n (n ? 1) collection is an ordered list of collections having order Each interval denotes a set of contiguous moments of time. For example, the collection of Months, where each month is represented as the collection of days in that month, is a collection of order 2. In order to provide a user-friendly representation of collections, the authors introduce two classes of operators on collections and the notion of calen- dar, as a primitive collection. A calendar is defined as an order 1 collection formed by an infinite number of meeting 3 intervals which may start from a specific one. The two classes of operators are called dicing and slic- ing. A dicing operator allows to further divide each interval within a collection into another collection. For example, Weeks:during:January1998 divides the interval corresponding to January1998 into the intervals corresponding to the weeks that are fully contained in that month. Other dicing operators are allowed, adopting a subset of Allen's interval relations [1]. Slicing operators provide means of selecting intervals from collections. For example, selects the first and last week from those identified by the dicing operator above. In general, slicing can be done using a list of integers, as well as with the keyword the, which identifies the single interval of the collection (if it is single), and the keyword any, which gives non-deterministically one of the intervals. Collection expressions can be arbitrarily composed using these two classes of operators starting from calendars, which are explicitly specified either by a periodic set of intervals, or as a grouping of intervals from previously defined meets interval I 2 if I 2 starts when I 1 finishes. calendars. The slice formalism was introduced in [14] as an alternative to the collection formalism in order to have an underlying evaluation procedure for the symbolic expressions. It is based on the notions of calendar and slice. Similarly to the collection formalism, calendars are periodic infinite sets of consecutive inter- vals, but there is no first nor last interval. Intervals in a calendar are indexed by consecutive integers. Once a basic calendar is given in terms of the time domain, other calendars can be defined dynamically from existing ones by the construct Generate(sp; C; l which generates a new calendar with m intervals in each period, the first one obtained grouping l 1 granules of calendar C, starting from C(sp), the second grouping the successive l 2 granules, and so on, with treated as a circular list. A calendar C 1 is a subcalendar of C each interval of C 2 is exactly covered by finite number of intervals of C 1 . Weeks, Days, Months are calendars with DaysvMonths, DaysvWeeks, Weeks6vMonths. A slice is a symbolic expression built from calendars and denoting a (finite or infinite) set of not necessarily consecutive intervals. It has the form where the sum identifies the starting points of the intervals and D their duration. Each C i is a symbol denoting a calendar and O i is either a set of natural numbers or the keyword all. If the sum is simply O 1 :C 1 , it denotes the starting points of the intervals of C 1 whose index belongs to O 1 , or the starting points of all intervals all. If the sum is On :Cn with On = fon g it denotes the starting points of the on -th interval of Cn following each point in . For example, the sum all.Years denotes the set of points corresponding to the beginning of the first day of February and April of each year. The duration D has the form h:C d where C d is a symbol denoting a calendar such that C d v Cn , and h is the number of successive intervals of C d specifying the duration. Hence, the slice all.Years f1g.Days . 2.Days denotes a set of intervals corresponding to the first 2 days of February and April of each year. 3.2. Expressiveness and relationships Both collections and slices essentially characterize periodic sets. Similarly to granularities, even in these formalisms there is the notion of a basic cal- endar, which defines the finest time units in the do- main. Without loss of generality, in the following of the paper we assume that this basic calendar (denoted by C) is the basic granularity we mentioned in Section 2. A period, in terms of C can be associated with each slice expression S as well as with any collection expression E. Intuitively, the period indicates the number of instants of C after which the same pattern of intervals denoted by the expression is repeat- ing; each interval in a period can be obtained by a constant shift of the corresponding interval in another pe- riod. If C are the calendars appearing in the expression, then the period is the least common multiple of P eriod(C i =C). Technically, P eriod(C i =C), is defined as is a list of integers, each one denoting the duration of an interval of C i in terms of returns the j th element of the list, and length(list) returns the number of elements in the list. For example, and, hence, P We now consider the expressiveness of slice expressions with respect to the formal notion of granularity introduced in Section 2. If all the intervals denoted by a slice S are disjoint, we call S a disjoint slice. We also say that a granularity G is equivalent to a slice S, if each granule of G is formed by the union of a set of granules of the basic granularity (C) and this set is represented by one of the intervals denoted by the slice; moreover, each of the intervals must describe one of these sets. Theorem 1 Given a disjoint slice S, there exists a no- gap finite granularity, or a no-gap infinite periodical granularity G equivalent to S. Technically, if is an infinite slice we have an algorithm to derive the intervals f[r is the length in terms of the basic calendar C corresponding to h granules of C d , starting at r i . These intervals are the ones denoted by S within a slice period. Then, a periodical granularity G can be defined by taking P eriod is the slice period in terms of C, and C(x) for each It is shown that G is equivalent to S. When S is finite, the same algorithm can be easily adapted to derive all the intervals S denotes. Then, the equivalent granularity is simply defined explicitly mapping each granule to one of these intervals. Disjointness ensures that the result of this mapping is indeed a granularity. Ignoring exceptions to leap years. Example 1 Let S=all.Weeks . 12.Hours be an infinite slice and Hours be the basic calendar. The slice P eriod is 168 hours (the number of hours in a week) and in the period containing Hours(1) the slice denotes the set of intervals f[25; 36]; [49; 60]g. The periodical G, equivalent to S, is defined by taking (the number of intervals in a period), x=25 Hours(x) and The following example shows that if a slice is non-disjoint, then there is no equivalent granularity. Example 2 Let S=all.Weeks . 3.Days. According to the slice semantics, this expression denotes all intervals spanning from Tuesday to Thursday and all intervals from Wednesday through Friday. By Definition 1, no pair of granules of the same granularity can overlap. Hence, no granularity can be found which is equivalent to S. 2 To understand the expressiveness of the slice formalism with respect to granularities, we still need to check if any granularity in the identified classes is representable by a disjoint slice. Theorem 2 Given a no-gap finite granularity or a no-gap infinite periodical granularity, there exists an equivalent slice. The theorem states that any finite (periodical or not) granularity can be represented by a slice, and that the same holds for periodical granularities which are unbounded on both sides. INFINITE-R and INFINITE- granularities cannot be represented by a slice, since the only way to denote an infinite set of intervals with a slice is to have O all, and there is no way within the slice formalism to impose a minimum or a maximum on that set. 5 From the above results we can conclude that disjoint slices can represent exactly the set of granularities identified in Figure 2, while non-disjoint ones do not represent granularities at all. Unfortunately, it seems that there is no way to enforce disjointness by simple syntax restrictions. We now consider the collection formalism. Proposition 1 Any collection E resulting from the application of a dicing or slicing operator is such that 5 Note however, that the addition of a reference interval (bound) to each slice, as used in [2], provides an easy extension to capture all no-gap periodical granularities.0000000000000000001111111111111111111111110000000000001111111111111111110000000000000000001111111111111111111111110000000000111111111111111000000111111 GAP NO-GAP INFINITE-R INFINITE-L Disjoint slice expressions Figure 2. The subset of the granularities captured by the slice formalism any two intervals t and u contained in E are either equal or disjoint. Proposition 1 follows from the semantics of the operators, and from the fact that each calendar contains only disjoint intervals. Similarly to slices, we say that a granularity G is equivalent to a collection E, if each granule of G is formed by the union of the granules of C represented by one of the intervals in the collection; moreover, each interval in the collection describes the composition of one of the granules of G. Theorem 3 Given a collection expression, there exists an equivalent no-gap periodical or finite non periodical granularity. Similarly to Theorem 1, we developed an algorithm to parse the expression, to derive its period, the intervals it denotes within the period 6 , and lower/upper bounds if present. Once the intervals are derived, we have all the data that is needed to define the granularity G, since it will have the same period, the intervals within the period define the corresponding granules, and the lower/upper bounds are used to impose a start- ing/ending non-empty granule. Example 3 Consider E=f1/Mondays:during:Years.2000g. This collection expression identifies an order 1 6 The intervals may be structured in a collection of order higher than 1, but this is irrelevant with respect to the time granules that the expression denotes. collection that contains all first Mondays of each year starting since Monday, January 1st 2001. We assume Days is the basic calendar with Saturday, Jan 1st 2000. We first have to compute the expression period. Since Mondays is defined as 1/Days:during:Weeks with the periods of Days and Weeks equal to 1 and 7 respec- tively, is the period computed for Mondays. Similarly, since the period for Years with respect to the basic calendar is 1461 (4 years in Days), the whole expression period is computed as lcm(7; years in Days). Then, G is defined as having period 28 (the number of granules in each period), (7/1/2002), . , and To obtain these intervals the algorithm first restricts years to those after 2000, then it represents all Mondays within those years, and in the end it extracts the intervals corresponding to the first Monday. 2 We also have the counterpart of Theorem 3. Theorem 4 Given a no-gap periodical or finite non- periodical granularity, there exists an equivalent collection GAP NO-GAP INFINITE-R INFINITE-L collection expressions Figure 3. The subset of the granularities captured by the collection formalism Note that in this case, all granularities in the right side of the inner circle of Figure 3 are captured. We can conclude that slices and collections have incomparable expressiveness, since slices can represent sets of overlapping intervals, and collections can represent INFINITE-R and INFINITE-L periodical granularities. From the above results, it is clearly possible to translate from one formalism to the other, when considering expressions denoting FINITE or INFINITE granular- ities, but it seems to be difficult to devise general rules to translate at the symbolic level, preserving the intuitiveness of the expression. Indeed, despite the in slices may be intuitively interpreted as equivalent to :during: in collections, they actually have a different semantics. The collection formalism has been extended with some additional operators in [8]. In particular, control statements if-then-else and while are introduced to facilitate the representation of certain sets of intervals, as for example, the fourth Saturday of April if not an holiday, and the previous business-day otherwise. Unfortunately, the syntax allows the user to define collections which contain overlapping intervals 7 . This implies that there are collection expressions in the extended formalism for which there does not exist an equivalent granularity. 4. An extension proposal Both the collection and slice formalisms as well as their known extensions cannot represent gap gran- ularities. Indeed, this requires a non-convex interval representation for each granule which is formed by non-contiguous instants. For example, they cannot represent Business-Months, where each granule is defined as the set of Business-Days within a month, and it is perceived as an indivisible unit. We propose an extension to the collection formalism in order to capture the whole set of periodical granularities. We introduce the notion of primitive collection, which includes calendars as defined in the collection formalism as well as order 1 collections of non-convex intervals, where each of the intervals represents a gran- ule. A primitive collection PC can be specified by sp is a synchronization point with respect to an existing calendar is the period expressed in terms of C 0 , and X is the set of non-convex intervals 8 identifying the position of granules of PC within a period. The synchronization point sp says that PC(1) will start at the same instant as C 0 (sp). 7 For example, consider an expression representing a semester following the last day of the month, if it is a Sunday, otherwise the week following that day. Considering 31/5/1998 and 30/6/1998, both the semester starting 1/6/1998 and the week starting will be denoted, with the first properly containing the second. 8 Each x is the non-convex interval representing the i-th granule. Example 4 Suppose a company has 2 weekly working shifts for its employees: shift1=fMonday, Wednesday, Saturdayg and shift2=fTuesday, Thursday, Fridayg. It may be useful to consider these as two periodic granularities, where each shift is treated as a single time granule within a week. If Thursday 1/1/1998 is taken as Days(1), shift1 = Generate(5; Days; 7; fh[1; 1]; [3; 3]; [6; 6]ig). In- deed, the synchronization point is 5, since the first granule of shift1 following Days(1) starts on Monday January 5th 1998 which is 5 days later. C 0 is Days, the period P is 7 days and X is composed by x which identifies the single granule within the period, formed by the first, third, and sixth day, starting from 5/1/1998, and repeating every 7 days. Similarly, Generate(6; Days; 7; fh[1; 1]; [3; 3]; [4; 4]ig) denotes the first, third and fourth day, starting from 6/1/1998, and repeating every 7 days. 2 The user can specify collection expressions by arbitrarily applying dicing and slicing operators starting from primitive collections. Since operators now apply to non-convex intervals, we need to revise their definition. Let t and u be non convex in- tervals, with an x b n g and S be the sets of values represented by t and u respectively. Dicing operators are based on the following binary relations on non-convex intervals: 9 during u iff S ' S 0 intersects u starts u iff (a A dicing operator :rel: takes an order 1 collection as its left operand and an interval u as its right operand, and it returns an order 1 collection rel ug. If the strict form : rel : is used, then rel ug, i.e., only the portion of t which is contained in u is part of the resulting 9 This set of relations is similar to the one chosen in [11] for convex intervals. We consider it only as a good basic set which allows the representation of most common granularities while having a simple implementation. It can be extended to a richer set considering, for example, the taxonomy of relations given in [10]. collection. When the right operand is a collection, instead of a single interval, the same procedure is applied for each of its intervals, resulting in a collection of one order higher. A slicing operator k=E replaces each collection contained in E with the k-th non-convex interval in that collection, while replaces it with the collection made of the subset of intervals whose position in the collection is specified by g. Example 5 Consider the collection expression Weeks:?:2/shift1:during:1998/Years where shift1 was defined in Example 4. This expression denotes all weeks following the end of the second work-shift of 1998. Years is the order ity, we assume the interval [1::365] corresponds to year 1998. Then, the slicing 1998/Years returns the interval h[1::365]i, and the dicing shift1:during:1998/Years returns the finite collection of order 1 composed by all the work-shifts during 1998: fh[5; 5], [7; 7], [10; 10]i, . , h[355; 355], 360]ig. The selection of the second of those work-shifts returns the non-convex interval Finally, the dicing Weeks:>:h[12; 12]; [14; 14]; [17; 17]i generates the collection of all the weeks that start after January 17-th, i.e., fh[19; 23]i; h[26; We state a formal property of the proposed extension Theorem 5 The extended collection formalism can represent all and only the granularities which are either periodical or finite non-periodical. To support this result, the algorithm used in the proof of Theorem 3 has been extended to consider non-convex intervals. The granularities captured by the proposed extension are shown in Figure 4. 5. Conclusions In this paper we have considered a recently proposed theoretical framework for time granularities and we have analyzed two of the most influential proposals for calendar symbolic representation. On one side, we have shown that the theoretical framework is general enough to capture all the sets of disjoint intervals representable by those formalisms. On the other side we have shown exactly which subclass of granularities can be represented by each formalism. From this GAP NO-GAP INFINITE-L INFINITE-R extended collection expressions Figure 4. The subset of granularities captured by the proposed extension analysis, we have proposed an extension of the collection formalism which captures a well-defined and large class of granularities, providing a good coverage of granularities that may be found in database and temporal reasoning applications. We are currently working at the definition and implementation of set operations, performed at the symbolic level, among extended collection expressions. This problem has interesting applications (see e.g., [2]) but it is not addressed in [11] and derivative work for collections, and only briefly investigated in [14] for slices. --R Maintaining Knowledge about Temporal Intervals An Access Control Model Supporting Periodicity Constraints and Temporal Reasoning A General Framework for Time Granularity and its Application to Temporal Reasoning in book Database support for workflow management: the WIDE project Expressing Time Intervals and Repetition within a Formalization of Calendars Temporal Granularity for Unanchored Temporal Data Implementing calendars and temporal rules in next generation databases a taxonomy of interval relations A representation for collections of temporal intervals Metric and Layered Temporal Logic for Time Granularity An efficient symbolic representation of periodic time Reasoning About Periodic Events The TSQL2 Temporal Query Language for any integer s and 1 is one of those denoted by S. is among the intervals denoted by the slice. "!" --TR --CTR Claudio Bettini , X. Sean Wang , Sushil Jajodia, Temporal Reasoning in Workflow Systems, Distributed and Parallel Databases, v.11 n.3, p.269-306, May 2002 flexible approach to user-defined symbolic granularities in temporal databases, Proceedings of the 2005 ACM symposium on Applied computing, March 13-17, 2005, Santa Fe, New Mexico Lavinia Egidi , Paolo Terenziani, A mathematical framework for the semantics of symbolic languages representing periodic time, Annals of Mathematics and Artificial Intelligence, v.46 n.3, p.317-347, March 2006
time granularity;knowledge representation;temporal reasoning;time representation
590507
Probabilistic Default Reasoning with Conditional Constraints.
We present an approach to reasoning from statistical and subjective knowledge, which is based on a combination of probabilistic reasoning from conditional constraints with approaches to default reasoning from conditional knowledge bases. More precisely, we introduce the notions of i>z-, lexicographic, and conditional entailment for conditional constraints, which are probabilistic generalizations of Pearl's entailment in system i>Z, Lehmann's lexicographic entailment, and Geffner's conditional entailment, respectively. We show that the new formalisms have nice properties. In particular, they show a similar behavior as reference-class reasoning in a number of uncontroversial examples. The new formalisms, however, also avoid many drawbacks of reference-class reasoning. More precisely, they can handle complex scenarios and even purely probabilistic subjective knowledge as input. Moreover, conclusions are drawn in a global way from all the available knowledge as a whole. We then show that the new formalisms also have nice general nonmonotonic properties. In detail, the new notions of i>z-, lexicographic, and conditional entailment have similar properties as their classical counterparts. In particular, they all satisfy the rationality postulates proposed by Kraus, Lehmann, and Magidor, and they have some general irrelevance and direct inference properties. Moreover, the new notions of i>z- and lexicographic entailment satisfy the property of rational monotonicity. Furthermore, the new notions of i>z-, lexicographic, and conditional entailment are proper generalizations of both their classical counterparts and the classical notion of logical entailment for conditional constraints. Finally, we provide algorithms for reasoning under the new formalisms, and we analyze its computational complexity.
Introduction In this paper, we elaborate a combination of probabilistic reasoning from conditional constraints with approaches to default reasoning from conditional knowledge bases. As a main result, this combination provides new notions of entailment for conditional constraints, which respect the ideas of classical default reasoning from conditional knowledge bases, and which are generally much stronger than the classical notion of logical entailment based on conditioning. Moreover, the results of this paper can also be applied for handling inconsistencies in probabilistic knowledge bases. Informally, the ideas behind this paper can be described as follows. Assume that we have the following knowledge at hand: "all penguins are birds" (G1), "between 90 and 95% of all birds fly" (G2), and "at most 5% of all penguins fly" (G3). Moreover, assume a first scenario in which "Tweety is a bird" (E1) and second one in which "Tweety is a penguin" (E2). What do we conclude about Tweety's ability to fly? A closer look at this example shows that the statements G1-G3 describe statistical knowledge (or objective knowl- while E1 and E2 express degrees of belief (or subjective knowledge). One way of handling such combinations of statistical knowledge and degrees of belief is reference class reasoning, which goes back to Reichenbach (1949) and was further refined by Kyburg (1974; 1983) and Pollock (1990). Another related field is default reasoning from conditional knowledge bases, where we have generic statements of the form "all penguins are birds", "generally, all birds fly", and "generally, no penguin flies" in addition to some concrete evidence as E1 and E2. The literature contains several different approaches to default reasoning and extensive work on the desired properties. The core of these properties are the rationality postulates proposed by Kraus et al. (1990). These rationality postulates constitute a sound and complete axiom system for several classical model-theoretic entailment relations under uncertainty measures on worlds. In detail, they characterize classical model-theoretic entailment under preferential structures (Shoham 1987; Kraus et al. 1990), infinitesimal probabilities (Adams 1975; Pearl 1989), possibility measures (Dubois & Prade 1991), and world rankings (Spohn 1988; Goldszmidt & Pearl 1992). They also characterize an entailment relation based on conditional objects (Dubois & Prade 1994). A survey of all these relationships is given in (Benferhat et al. 1997). Recently, Friedman and Halpern (2000) showed that many approaches describe to the same notion of inference, since they are all expressible as plausibility measures. Mainly to solve problems with irrelevant information, the notion of rational closure as a more adventurous notion of entailment has been introduced by Lehmann (Lehmann 1989; Lehmann & Magidor 1992). This notion of entailment is equivalent to entailment in system Z by Pearl (1990), to the least specific possibility entailment by Benferhat et al. (1992), and to a conditional (modal) logic-based entailment by Lamarre (1992). Finally, mainly in order to solve problems with property inheritance from classes to exceptional subclasses, the maximum entropy approach to default entailment was proposed by Goldszmidt et al. (1993); the notion of lexicographic entailment was introduced by Lehmann (1995) and Benferhat et al. (1993); the notion of conditional entailment was proposed by Geffner (Geffner 1992; Geffner Pearl 1992); and an infinitesimal belief function approach was suggested by Benferhat et al. (1995). Coming back to our introductory example, we realize that G1-G3 and E1-E2 represent interval restrictions for conditional probabilities, also called conditional constraints (Lukasiewicz 1999b). The literature contains extensive work on reasoning about conditional constraints (Dubois & Prade 1988; Dubois et al. 1990; 1993; Amarger et al. 1991; Jaumard et al. 1991; Th-one et al. 1992; Frisch & Haddawy 1994; Heinsohn 1994; Luo et al. 1996; Lukasiewicz 1999a; 1999b) and their generalizations, for example, to probabilistic logic programs (Lukasiewicz 1998). Now, the main idea of this paper is to use techniques for default reasoning from conditional knowledge bases in order to perform probabilistic reasoning from statistical knowledge and degrees of beliefs. More precisely, we extend the notions of entailment in system Z, Lehmann's lexicographic entailment, and Geffner's conditional entailment to the framework of conditional constraints. Informally, in our introductory example, the statements G2 and G3 are interpreted as "generally, a bird flies with a probability between 0.9 and 0.95" (G2 ? ) and "generally, a penguin flies with a probability of at most 0.05" (G3 ? ), respectively. In the first scenario, we then simply use the whole probabilistic knowledge to conclude under classical logical entailment that "Tweety flies with a probability between 0.9 and 0.95". In the second scenario, it turns out that the whole probabilistic knowledge precisely, is inconsistent in the context of a pen- guin. In fact, the main problem is that G2 ? should not be applied anymore to penguins. That is, we can easily re-solve the inconsistency by removing G2 ? , and then conclude from classical logical entailment that "Tweety flies with a probability of at most 0.05". Hence, the results of this paper can also be used for handling inconsistencies in probabilistic knowledge bases. More precisely, the new notions of nonmonotonic entailment coincide with the classical notion of logical entailment as far as satisfiable sets of conditional constraints are con- cerned. Furthermore, they allow desirable conclusions from certain kinds of unsatisfiable sets of conditional constraints. We remark that this inconsistency handling is guided by the principles of default reasoning from conditional knowledge bases. It is thus based on a natural preference relation on conditional constraints, and not on the assumption that all conditional constraints are equally weighted (as, for ex- ample, in the work by Jaumard et al. (1991)). The work closest in spirit to this paper is perhaps the one by Bacchus et al. (1996), which suggests to use the random worlds method (Grove et al. 1994) to induce degrees of beliefs from quite rich statistical knowledge bases. How- ever, differently from (Bacchus et al. 1996), we do not make use of a strong principle such as the random worlds method (which is closely related to probabilistic reasoning under maximum entropy). Moreover, we restrict our considerations to the propositional setting. The main contributions of this paper are as follows: We illustrate that the classical notion of logical entailment for conditional constraints is not very well-suited for default reasoning with conditional constraints. We introduce the notions of z-entailment, lexicographic entailment, and conditional entailment for conditional constraints, which are a combination of the classical notions of entailment in system Z (Pearl 1990), Lehmann's lexicographic entailment (Lehmann 1995), and Geffner's conditional entailment (Geffner 1992; Geffner & Pearl 1992), respectively, with the classical notion of logical entailment for conditional constraints. We give some examples that analyze the nonmonotonic properties of the new notions of entailment for default reasoning with conditional constraints. It turns out that the new notions of z-entailment, lexicographic entailment, and conditional entailment have similar properties like their classical counterparts. We show that the new notions of z-entailment, lexicographic entailment, and conditional entailment for conditional constraints properly extend the classical notions of entailment in system Z, lexicographic entailment, and conditional entailment, respectively. We show that the new notions of z-entailment, lexicographic entailment, and conditional entailment for conditional constraints properly extend the classical notion of logical entailment for conditional constraints. Note that all proofs are given in (Lukasiewicz 2000). Preliminaries We now introduce some necessary technical background. We assume a finite nonempty set of basic propositions (or atoms) . We use ? and > to denote the propositional constants false and true, respectively. The set of classical formulas is the closure of [f?;>g under the Boolean operations : and ^. A strict conditional constraint is an expression real numbers l; u2 [0; 1] and classical formulas and . A defeasible conditional constraint (or default) is an expression ( k)[l; u] with real numbers classical formulas and . A conditional constraint is a strict or defeasible conditional constraint. The set of strict probabilistic formulas (resp., probabilistic formulas) is the closure of the set of all strict conditional constraints (resp., conditional constraints) under the Boolean operations : and ^. We use and to abbreviate :(:F ^:G), :(F ^:G), and (:(:F ^G))^ (:(F ^:G)), respectively, and adopt the usual conventions to eliminate parentheses. A probabilistic default theory is a pair P is a finite set of strict conditional constraints and D is a finite set of defeasible conditional constraints. A probabilistic knowledge base KB is a strict probabilistic formula. Informally, default theories represent strict and defeasible generic knowledge, while probabilistic knowledge bases express some concrete evidence. A possible world is a truth assignment I : ! ftrue, falseg, which is extended to classical formulas as usual. We use I to denote the set of all possible worlds for . A possible world I satisfies a classical formula , or I is a model of , denoted I A probabilistic interpretation Pr is a probability function on I (that is, a mapping Pr : I ! [0; 1] such that all Pr(I) with I 2 I sum up to 1). The probability of a classical formula in the probabilistic interpretation Pr , denoted Pr(), is defined as follows: For classical formulas and with Pr () > 0, we use Pr( to abbreviate Pr(). The truth of probabilistic formulas F in a probabilistic interpretation Pr , denoted Pr inductively defined as follows: Pr Pr Pr Pr G. We remark that there is no difference between strict and defeasible conditional constraints as far as the notion of truth in probabilistic interpretations is concerned. A probabilistic interpretation Pr satisfies a probabilistic formula F , or Pr is a model of F , iff Pr a set of probabilistic formulas F , or Pr is a model of F , denoted Pr is a model of all F 2F . We say F is satisfiable iff a model of F exists. We next define the notion of logical entailment as fol- lows. A strict probabilistic formula F is a logical consequence of a set of probabilistic formulas F , denoted F iff each model of F is also a model of F . A strict conditional constraint ( j)[l; u] is a tight logical consequence of F , denoted F is the infimum (resp., supremum) of Pr( j) subject to all models Pr of F with Pr() > 0 (note that we canonically define We remark that every notion of entailment for conditional constraints is associated with a notion of consequence and a notion of tight consequence. Informally, the notion of consequence describes entailed intervals, while the notion of tight consequence characterizes the tightest entailed inter- val. That is, if ( j)[l; u] is a tight consequence of F , then Motivating Examples What should a probabilistic knowledge base entail under a probabilistic default theory? To get a rough idea on the reply to this question, we now introduce two natural notions of entailment and analyze their properties. It will turn out that neither of these two notions is fully adequate for probabilistic default reasoning with conditional constraints. In the sequel, let D) be a probabilistic default theory. We first define the notion of 0-entailment, which applies to probabilistic knowledge bases of the In detail, a strict conditional constraint ( j)[l; u] is a 0-consequence of KB , denoted It is a tight 0-consequence of KB , denoted KB k 0 tight ( j)[l; u], iff Informally, we use the concrete evidence in KB to fix our "point of interest" and the generic knowledge in T to draw the requested conclusion. That is, we perform classical conditioning. We next define the notion of 1-entailment, which applies to all probabilistic knowledge bases KB . A strict probabilistic formula F is a 1-consequence of KB , denoted strict conditional constraint ( j)[l; u] is a tight 1-consequence of KB , denoted tight ( j)[l; u], iff P[D[fKBg Informally, we draw our conclusion from the union of the concrete evidence in KB and the generic knowledge in T . We now analyze the properties of these two notions of entailment. Our first example concentrates on the aspects of ignoring irrelevant information and property inheritance. Example 1 The knowledge "all penguins are birds" and "at least 95% of all birds have legs" can be expressed by the following probabilistic default theory should entail that "generally, birds have legs with a probability of at least 0.95" (that is, e.g., if we know that Tweety is a bird, and we do not have any other knowl- edge, then we should conclude that the probability of Tweety having legs is at least 0.95). Indeed, this conclusion is drawn under both 0- and 1-entailment (see item (1) in Table 1). should entail that "generally, yellow birds have legs with a probability of at least 0.95" (as the property "yellow" is not mentioned at all in T 1 and thus irrelevant), and that "generally, penguins have legs with a probability of at least 0.95" (as the set of all penguins is a nonexceptional subclass of the set of all birds, and thus penguins should inherit all properties of birds). However, while 1-entailment still allows the desired conclusions, 0-entailment just yields the interval [0; 1] (see items (2)-(3) in Table 1). 2 We next concentrate on the principle of specificity and the problem of inheritance blocking. Example 2 Let us consider the following probabilistic default theory (fly k bird)[:9; :95]; (fly k penguin)[0; This default theory should entail that "generally, penguins fly with a probability of at most 0.05" (as properties of more specific classes should override inherited properties of less specific classes). Indeed, 0-entailment yields the desired conclusion, while 1-entailment reports an unsatisfiability (see item (7) in Table 1). Moreover, should entail that "generally, penguins have legs with a probability of at least 0.95", since penguins are exceptional birds w.r.t. to the ability of being able to fly, but not w.r.t. the property of having legs. However, 0-entailment provides only the interval [0; 1], and 1-entailment reports even an unsatisfiability (see item (5) in Table 1). 2 The following example deals with the drowning problem (Benferhat et al. 1993). Example 3 Let us consider the following probabilistic default theory f(fly k bird)[:9; :95]; (fly k penguin)[0; :05]; (easy to see k This default theory should entail that "generally, yellow penguins are easy to see", as the set of all yellow penguins Table 1: Examples of 0- and 1-entailed tight intervals. tight (easy to seej>) [0; 1] [1; 0] undefined [:86; :91] undefined [1; 0] is a nonexceptional subclass of the set of all yellow objects. But, 0-entailment gives only the interval [0; 1], and 1-entail- ment reports an unsatisfiability (see item (8) in Table 1). 2 The next example is taken from (Bacchus et al. 1996). Example 4 Let us consider the following probabilistic default theory This default theory should entail "generally, the probability that magpies chirp is between 0.7 and 0.8", since we know more about birds w.r.t. the property of being able to chirp than about magpies. Indeed, both 0- and 1-entailment yield the desired conclusion (see item (9) in Table 1). 2 The following example concerns ambiguity preservation (Benferhat et al. 1995). Example 5 Let us consider the following probabilistic default theory f(fly k metal wings)[:95; 1]; (fly k bird)[:95; 1]; (fly k penguin)[0; Assume now that Oscar is a penguin with metal wings. As Oscar is a penguin, we should conclude that the probability that Oscar flies is at most 0.05. However, as Oscar has also metal wings, we should conclude that the probability that Oscar flies is at least 0.95. As argued in the literature on default reasoning (Benferhat et al. 1995), such ambiguities should be preserved. Indeed, 0-entailment yields the desired interval [0; 1], while 1-entailment reports an unsatisfiability (see item (10) in Table 1). 2 What about handling purely probabilistic evidence? Example 6 Let us consider again the probabilistic default theory T 2 of Example 2. Assume a first scenario in which our belief is "the probability that Tweety is a bird is at least 0.9" and "the probability that Tweety is a penguin is at least 0.1" and a second scenario in which our belief is "the probability that Tweety is a bird is at least 0.9" and "the probability that Tweety is a penguin is at least 0.9". What do we conclude about Tweety's ability to fly in these scenarios? The notion of 0-entailment is undefined for such purely probabilistic evidence, whereas the notion of 1-entailment yields the probability interval [:86; :91] in the first scenario, and reports an unsatisfiability in the second scenario (see items (11)-(12) in Table 1). 2 Summarizing the results, 0-entailment is too weak, while 1-entailment is too strong. In detail, 0-entailment often yields the trivial interval [0; 1] and is even undefined for purely probabilistic evidence, while 1-entailment often reports unsatisfiabilities (in fact, in the most interesting sce- narios, as 1-entailment is actually monotonic). Roughly speaking, our ideal notion of entailment for probabilistic knowledge bases under probabilistic default theories should lie somewhere between 0- and 1-entailment. One idea to obtain such a notion could be to strengthen 0- entailment by adding some inheritance mechanism. Another idea is to weaken 1-entailment by handling unsatisfiabilities. In the rest of this paper, we will focus on the second idea. Probabilistic Reasoning In this section, we extend the classical notions of entailment in system Z (Pearl 1990), Lehmann's lexicographic entailment (1995), and Geffner's conditional entailment (Geffner 1992; Geffner & Pearl 1992) to conditional constraints. The main idea behind these extensions is to use the following two interpretations of defaults. As far as default rankings and priority orderings are concerned, we interpret a "generally, if is true, then the probability of is between l and u". Whereas, as far as notions of entailment are concerned, we interpret ( k)[l; u] as "the conditional probability of given is between l and u". Preliminaries A probabilistic interpretation Pr verifies a default ( k)[l;u] It falsifies a default set of defaults D tolerates a default d under a set of strict conditional constraints has a model that verifies d. A set of defaults D is under P in conflict with d iff no model of verifies d. A default ranking on D maps each d 2D to a nonnegative integer. It is admissible with iff each set of defaults D 0 D that is under P in conflict with some default d 2D contains a default d 0 such that (d 0 ) <(d). A probabilistic is -consistent iff there exists a default ranking on D that is admissible with T . It is -inconsistent iff no such default ranking exists. A probability ranking maps each probabilistic interpretation on I to a member of f0; for at least one interpretation Pr . It is extended to all strict probabilistic formulas F as follows. If F is sat- isfiable, then We say is admissible with F iff (:F It is admissible with a default ( k)[l; u] iff Roughly speaking, the intuition behind this definition is to "generally, if is true, then the probability of is between l and u". A probability ranking is admissible with a probabilistic default theory is admissible with all F 2P and all d 2D. System Z We now extend the notion of entailment in system Z (Pearl 1990; Goldszmidt & Pearl 1996) to conditional constraints. In the sequel, let D) be a -consistent probabilistic default theory. The notion of z-entailment is linked to an ordered partition of D, a default ranking z, and a probability ranking z . We first define the z-partition of D. Let (D the unique ordered partition of D such that, for each D i is the set of all defaults in D that are tolerated under P by D that we define D call this (D the z-partition of D. Example 7 The z-partition for the probabilistic default theory is given as follows: (f(legs k bird)[:95; 1]; (fly k bird)[:9; :95]g; f(fly k penguin)[0; We now define the default ranking z. For each d 2D j is assigned the value j under z. The probability ranking z on all probabilistic interpretations Pr is then defined as follows: z(d) otherwise. The following result shows that, in fact, z is a default ranking that is admissible with T , and z is a probability ranking that is admissible with T . Lemma 8 a) z is a default ranking admissible with T . b) z is a probability ranking admissible with T . We next define a preference relation on probabilistic in- terpretations. For probabilistic interpretations Pr and Pr 0 , we say Pr is z-preferable to Pr 0 iff z (Pr) < z (Pr 0 A model Pr of a set of probabilistic formulas F is a z- minimal model of F iff no model of F is z-preferable to Pr . We are now ready to define the notion of z-entailment as follows. A strict probabilistic formula F is a z-con- sequence of KB , denoted KB k z F , iff each z-minimal model of P [ fKBg satisfies F . A strict conditional constraint ( j)[l; u] is a tight z-consequence of KB , denoted tight ( j)[l; u], iff l (resp., u) is the infimum (resp., supremum) of Pr( j) subject to all z-minimal models Pr of P [ fKBg with Pr() > 0. Coming back to Examples 1-6, it turns out that the non-monotonic properties of z-entailment differ from the ones of 0- and 1-entailment (see Table 2). In detail, in the given examples, z-entailment ignores irrelevant information, shows property inheritance to globally nonexceptional subclasses, and respects the principle of specificity. Moreover, it may also handle purely probabilistic evidence. However, properties are still not inherited to more specific classes that are exceptional with respect to some other properties. Moreover, z-entailment still has the drowning problem and does not preserve ambiguities. The following examples illustrate how z-entailed tight intervals are determined. Example 9 Given T 2 of Example 2, we get: tight (legs j >)[0; 1] Here, the interval "[0; 1]" comes from the tight logical consequence Given T 5 of Example 5, we get: tight (fly j Here, the interval "[0; :05]" comes from the tight logical consequence metal wingsj>)[1; 1]g Lexicographic Entailment We now extend Lehmann's lexicographic entailment (Lehmann 1995) to conditional constraints. In the sequel, let D) be a -consistent probabilistic default theory. We now use the z-partition (D of D to define a lexicographic preference relation on probabilistic interpretations. Table 2: Examples of z-, lexicographically, and conditionally entailed tight intervals. tight k lex tight k ce tight (fly j>) [:9; :95] [:9; :95] [:9; :95] (fly j>) [0; :05] [0; :05] [0; :05] (easy to see j>) [0; 1] [:95; 1] [:95; 1] (fly j>) [0; :05] [0; :05] [0; 1] (fly j>) [:86; :91] [:86; :91] [:86; :91] (fly j>) [0; :15] [0; :15] [0; :15] For probabilistic interpretations Pr and Pr 0 , we say Pr is lexicographically preferable to Pr 0 iff there exists some for all i < j k. A model Pr of a set of probabilistic formulas F is a lexicographically minimal model of F iff no model of F is lexicographically preferable to Pr . We now define the notion of lexicographic entailment as follows. A strict probabilistic formula F is a lexicographic consequence of KB , denoted KB k lex F , iff each lexicographically minimal model of P [fKBg satisfies F . A strict conditional constraint ( j)[l; u] is a tight lexicographic consequence of KB , denoted KB k lex tight ( j)[l; u], iff l (resp., u) is the infimum (resp., supremum) of Pr( subject to all lexicographically minimal models Pr of P[fKBg with Pr () > 0. Coming back to Examples 1-6, it turns out that lexicographic entailment has nicer nonmonotonic features than z-entailment (see Table 2). In detail, in the given examples, lexicographic entailment ignores irrelevant information, shows property inheritance to nonexceptional subclasses, and respects the principle of specificity. Moreover, it does not block property inheritance, it does not have the drowning problem, and it may also handle purely probabilistic evidence. However, lexicographic entailment still does not preserve ambiguities. The following examples illustrate how lexicographically entailed tight intervals are determined. Example 11 Given T 2 of Example 2, we get: tight (legs j >)[:95; Here, the interval "[:95; 1]" comes from the tight logical consequence (fly k penguin)[0; :05], Example 12 Given T 5 of Example 5, we get: tight (fly j Here, the interval "[0; :05]" comes from the tight logical consequence metal wingsj>)[1; 1]g Conditional Entailment We next extend Geffner's conditional entailment (Geffner 1992; Geffner & Pearl 1992) to conditional constraints. In the sequel, let D) be a probabilistic default theory. We first define priority orderings on D as follows. A priority ordering on D is an irreflexive and transitive binary relation on D. We say is admissible with T iff each set of defaults D 0 D that is under P in conflict with some default d 2D contains a default d 0 such that d 0 d. We say T is -consistent iff there exists a priority ordering on D that is admissible with T . Example 13 Consider the probabilistic default theory 2. A priority ordering on D 2 that is admissible with T 2 is given by (fly k bird)[:9; :95] (fly k penguin)[0; :05]. 2 The existence of an admissible default ranking implies the existence of an admissible priority ordering. Lemma 14 If T is -consistent, then T is -consistent. We next define a preference ordering on probabilistic interpretations as follows. Let Pr and Pr 0 be two probabilistic interpretations and let be a priority ordering on D. We say that Pr is -preferable to Pr 0 iff fd 2D j Pr 6j= dg 6= fd 2D j Pr 0 6j= dg and for each d 2D such that Pr 6j= d and there exists some default d 0 2D such that d d 0 , Pr model Pr of a set of probabilistic formulas F is a -minimal model of F iff no model of F is -preferable to Pr . A model Pr of a set of probabilistic formulas F is a conditionally minimal model of F iff Pr is a -minimal model of F for some priority ordering admissible with T . We finally define the notion of conditional entailment. A strict probabilistic formula F is a conditional consequence of KB , denoted KB k ce F , iff each conditionally minimal model of P [ fKBg satisfies F . A strict conditional constraint ( j)[l; u] is a tight conditional consequence of KB , denoted KB k ce tight ( j)[l; u], iff l (resp., u) is the infimum (resp., supremum) of Pr( j) subject to all conditionally minimal models Pr of P [ fKBg with Pr() > 0. Coming back to Examples 1-6, we see that among all introduced notions of entailment, conditional entailment is the one with the nicest nonmonotonic properties (see Table 2). In detail, in the given examples, conditional entailment ignores irrelevant information, shows property inheritance to nonexceptional subclasses, and respects the principle of specificity. Moreover, it does not block property inheritance, and it does not have the drowning problem. Finally, conditional entailment preserves ambiguities and may also handle purely probabilistic evidence. The following examples illustrate how conditionally entailed tight intervals are determined. Example 15 Given T 2 of Example 2, we get: ce tight (legs j >)[:95; Here, the interval "[:95; 1]" comes from the tight logical consequence (fly k penguin)[0; :05], Example Given T 5 of Example 5, we get: ce tight (fly j >)[0; Here, the interval "[0; 1]" is the convex hull of the intervals "[0; :05]" and "[:95; 1]", which come from the tight logical consequences metal wings j>)[1; 1]gj= tight (fly j >)[0; :05] and P 5 [f(fly k bird)[:95; 1], (fly k metal wings)[:95; 1], (penguin ^ metal wings Relationship to Classical Formalisms We now analyze the relationship to classical default reasoning from conditional knowledge bases and to classical probabilistic reasoning with conditional constraints. logical formula is a probabilistic formula that contains only conditional constraints of the kind ( j)[1;1] or strict logical formula is a strict probabilistic formula that contains only strict conditional constraints of the form ( j)[1; 1]. A logical default theory T is a probabilistic default theory that contains only logical formulas. A logical knowledge base KB is a strict logical formula. We use the operator on logical formulas, sets of logical formulas, and logical default theories, which replaces each strict conditional constraint ( j)[1; 1] (resp., defeasible conditional constraint ( k)[1; 1]) by the classical implication Given a logical ce ) to denote the classical notion of z-, (resp., lexicographic, con- ditional) entailment with respect to The following result shows that the introduced notions of z-, lexicographic, and conditional entailment are generalizations of their classical counterparts. Theorem 17 Let D) be a logical default theory and let KB be a logical knowledge base. Then, for every semantics s 2 fz; lex ; ceg: The next result shows that, when the union of generic and concrete probabilistic knowledge is satisfiable, the notions of z-, lexicographic, and conditional entailment coincide with the notion of 1-entailment. Theorem D) be a probabilistic default theory and let KB be a probabilistic knowledge base such that P [D[ fKBg is satisfiable. Then, for every semantics 1. KB k s F iff P [D[ fKBg 2. KB k s tight ( j)[l; u] iff P[D[fKBgj= tight ( j)[l; u]. Summary and Outlook We presented the notions of z-entailment, lexicographic en- tailment, and conditional entailment for conditional con- straints, which combine the classical notions of entailment in system Z, Lehmann's lexicographic entailment, and Geffner's conditional entailment with the classical notion of logical entailment for conditional constraints. We showed that the introduced notions for probabilistic default reasoning with conditional constraints have similar properties like their classical counterparts. Moreover, they properly extend both their classical counterparts and the classical notion of logical entailment for conditional constraints. An interesting topic of future research is to extend other formalisms for classical default reasoning to the probabilistic framework of conditional constraints. Acknowledgments I am very grateful to the referees for their useful comments. This work was supported by a DFG grant and the Austrian Science Fund Project N Z29-INF. --R The Logic of Conditionals Constraint propagation with imprecise conditional probabilities. From statistical knowledge bases to degrees of beliefs. Inconsistency management and prioritized syntax-based entailment Representing default rules in possibilistic logic. Belief functions and default reasoning. Logical Foundations of Probability. Chicago: University of Chicago Press. Theory of Probability. On fuzzy syllogisms. Computational Intelligence Possibilistic logic Conditional objects as non-monotonic consequence relationships Qualitative reasoning with imprecise probabilities. Journal of Intelligent Information Systems Inference with imprecise numerical quantifiers. Complexity results for default reasoning from conditional knowledge bases. A logic for reasoning about probabilities. Plausibility measures and default reasoning. Anytime deduction for probabilistic logic. Conditional entailment: Bridging two approaches to default reasoning. Reasoning: Causal and Conditional Theories. Qualitative probabilities for default reasoning A maximum entropy approach to nonmonotonic reasoning. Random worlds and maximum entropy. Probabilistic description logics. Column generation methods for probabilistic logic. A promenade from monotonicity to non-monotonicity following a theorem prover What does a conditional knowledge base entail? What does a conditional knowledge base entail? Another perspective on default reason- ing Probabilistic logic programming. Local probabilistic deduction from taxonomic and probabilistic knowledge-bases over conjunctive events Probabilistic deduction with conditional constraints over basic events. Probabilistic default reasoning with strict and defeasible conditional constraints. Computation of best bounds of probabilities from uncertain data. Probabilistic semantics for nonmontonic reasoning: A survey. System Z: A natural ordering of defaults with tractable applications to default reasoning. Nomic Probabilities and the Foundations of Induction. Theory of Probability. A semantical approach to nonmonotonic logics. Ordinal conditional functions: A dynamic theory of epistemic states. Towards precision of probabilistic bounds propagation. --TR --CTR Thomas Lukasiewicz, Nonmonotonic probabilistic logics under variable-strength inheritance with overriding: Complexity, algorithms, and implementation, International Journal of Approximate Reasoning, v.44 n.3, p.301-321, March, 2007 Donald Bamber , I. R. Goodman , Hung T. Nguyen, Robust reasoning with rules that have exceptions: From second-order probability to argumentation via upper envelopes of probability and possibility plus directed graphs, Annals of Mathematics and Artificial Intelligence, v.45 n.1-2, p.83-171, October 2005 Veronica Biazzo , Angelo Gilio , Thomas Lukasiewicz , Giuseppe Sanfilippo, Probabilistic logic under coherence: complexity and algorithms, Annals of Mathematics and Artificial Intelligence, v.45 n.1-2, p.35-81, October 2005 Thomas Lukasiewicz, Weak nonmonotonic probabilistic logics, Artificial Intelligence, v.168 n.1, p.119-161, October 2005 Angelo Gilio, Probabilistic Reasoning Under Coherence in System P, Annals of Mathematics and Artificial Intelligence, v.34 n.1-3, p.5-34, March 2002 Gabriele Kern-Isberner , Thomas Lukasiewicz, Combining probabilistic logic programming with the power of maximum entropy, Artificial Intelligence, v.157 n.1-2, p.139-202, August 2004
conditional constraint;system Z;probabilistic default reasoning;conditional entailment;lexicographic entailment
590522
Coloration Neighbourhood Search With Forward Checking.
Two contrasting search paradigms for solving combinatorial problems are i>systematic backtracking and i>local search. The former is often effective on highly structured problems because of its ability to exploit consistency techniques, while the latter tends to scale better on very large problems. Neither approach is ideal for all problems, and a current trend in artificial intelligence is the hybridisation of search techniques. This paper describes a use of forward checking in local search: pruning coloration neighbourhoods for graph colouring. The approach is evaluated on standard benchmarks and compared with several other algorithms. Good results are obtained; in particular, one variant finds improved colourings on geometric graphs, while another is very effective on equipartite graphs. Its application to other combinatorial problems is discussed.
Introduction Graph colouring is an NP-hard combinatorial optimisation problem with real-world applications such as timetabling, scheduling, frequency assign- ment, computer register allocation, printed circuit board testing and pattern matching. A graph E) consists of a set V of vertices and a set E of edges between vertices. Two vertices connected by an edge are said to be adjacent . The aim is to assign a colour to each vertex in such a way that no two adjacent vertices have the same colour. A graph colouring problem for a graph G is the problem of nding a k-colouring with k as small as possible. The chromatic number (G) of a graph is the minimum number of colours required to colour it. Many algorithms have been proposed for the graph colouring problem. Systematic backtracking gives good results on small graphs but scales poorly to large problems. Most colouring algorithms are stochastic in nature, searching in a non-systematic way with a variety of heuristics. The simplest type of stochastic search is local search: hill climbing, often augmented with heuristics for escaping local minima. Local search explores the neighbourhood of a point in a space by making local moves. The neighbourhood consists of the set of points 0 that can be reached by a single local move. The aim is to minimise (or equivalently to maximise) some objective function f() on the space. A local move ! 0 can be classied as backward , forward or sideways, depending on whether f( 0 ) f() is positive, negative or zero. Some algorithms choose a forward move that yields the greatest reduction in value, a strategy sometimes called greedy or steepest descent . A draw-back of local (and other stochastic) search is that it may converge on a local minimum: a point that has lower value than all its neighbours but is not a global minimum. The aim of backward moves is to escape from local minima by providing noise, while sideways moves are often used to traverse function plateaus. We might classify colouring algorithms by the search spaces they explore. The space of total colorations consists of the possible colour assignments to all the vertices of a graph, using a xed number of colours. This approach is used by the TABU algorithm, which tries to minimise the number of con icts. A con ict is an adjacent pair of vertices with the same colour. TABU generates a selection of possible single-vertex re-assignments and selects the best, even if this leads to more con icts. It also maintains a list of recent moves and avoids reversing them, which helps it to escape local minima. The consistent total colorations are the total colorations that contain no con icts. This space is explored by the Greedy (or sequential) algorithm, which tries to colour each vertex with a colour already used for a previous vertex; if this is not possible then a new colour is used. Heuristics control the vertex order and colour selection. The Iterative Greedy algorithm iteratively applies the Greedy algorithm, using vertex orderings that are guaranteed to generate a sequence of colorations using a non-increasing number of colours. Brelaz's DSATUR algorithm [2] explores a similar space. It orders vertices dynamically by maximum saturation (number of distinct colours assigned to adjacent vertices), breaking ties by choosing a vertex of greatest forward degree. The degree of a vertex is the number of its adjacent vertices, and its forward degree is the number of its uncoloured adjacent vertices. DSATUR has also been extended by backtracking. Mehrotra & Trick's version [15] uses full (exhaustive) search and begins by computing a clique which is never re- coloured. Another version described by Culberson, Beacham and Papp [3] uses limited backtracking. The space of consistent partial colorations consists of consistent colorations of subsets of the vertices that do not use more than a specied number of colours. This space is explored by the IMPASSE colouring algorithms. The objective function to be minimised is the sum of the forward degrees of the uncoloured vertices. Two such algorithms are Morgenstern's distributed IMPASSE [17] and Lewandowski & Condon's parallel IMPASSE [14]. Distributed IMPASSE performs limited searches on a distributed architecture, each search starting from previous good coloration, which are maintained in a pool. Parallel IMPASSE is a hybrid of IMPASSE and systematic search; the two execute in parallel and communicate colouring improvements to each other. Finally, the independent sets of a graph are the sets of pairwise non-adjacent vertices. They exploit the fact that all the vertices of an independent set can be assigned the same colour. The space of independent sets is explored by algorithms such as Johnson et al.'s XRLF [10] and Culberson using (exhaustive or restricted) backtracking and iteration. Mehrotra & Trick's LPCOLOR [15] uses column generation and branch-and-bound to explore this space. The motivation for this classication is that we shall describe a local search algorithm that explores a new space. This is a subspace of the consistent partial colorations, reduced by applying the Constraint Programming technique of forward checking . We describe the new algorithm in Section 2 and evaluate its performance in Section 3. The algorithm, results and related work are discussed in Section 4. This paper is an extension of earlier work [19], but the algorithm is described in greater detail and evaluated more systematically. FC-consistent partial coloration neighbourhood search We rst discuss a simple consistency technique from constraint programming: forward checking (FC). FC is commonly used with systematic backtracking [9], and this combination can be applied to graph colouring as follows. Each vertex has an associated domain of possible colours, initialised to the full set of available colours. On colouring a vertex with one of the colours in its domain, that colour is deleted from the domains of the adjacent vertices. No colour assignment is permitted if it causes the domain of some uncoloured vertex to become empty. This domain wipe-out often occurs long before the vertex in question is due to be coloured, greatly reducing the search space. On backtracking, a vertex is uncoloured and its assigned colour restored to the domains of adjacent vertices. An analogue to the Brelaz heuristic may be used to select vertices for colouring: select a vertex with smallest domain, breaking ties by selecting a vertex with greatest forward degree. FC is a simple and inexpensive algorithm, and sometimes out-performs theoretically more powerful techniques. FC with backtracking also has the advantage of completeness. That is, all k-colourings will eventually be found, and if there is no k-colouring then the algorithm will eventually prove this by terminating without nding a solution. However, systematic backtracking algorithms often suer from poor scalability. For example the FC algorithm is eective on small N-queens problems, but cannot solve problems with more than approximately 100 queens; in contrast, a local search algorithm solves up to 10 6 queens in linear time [16]. Backtracking and local search are complementary search techniques for solving colouring and other combinatorial problems, and considerable research has been devoted to combining advantages of the two. We describe how to exploit FC within a local search algorithm for graph colouring. The idea is to explore the subspace of the consistent partial col- orations that are also consistent under forward checking. That is, for each of the currently uncoloured vertices, there is at least one available colour that can be consistently assigned to it; colorations causing vertex domain wipe-out are avoided. We shall call this subspace the FC-consistent partial colorations. Our reasoning is that by reducing the search space we may avoid some local minima. Before describing the particular algorithm used to explore this space, we discuss a complication that arises when applying local search to it. In systematic backtracking it is simple to maintain vertex domains: the order in which colours are restored from domains on backtracking is the reverse of the order in which they were deleted during assignment. It is su-cient to maintain a boolean variable for each colour in each domain, denoting whether or not the colour is currently in the domain. However, local search is non-systematic, and from a given coloration we may wish to uncolour any vertex, not just the most recently coloured one. To do this we need a new way of maintaining domains. A number we shall call a con ict count is maintained for each vertex-colour pair (v,c) recording how many currently coloured vertices the assignment would con ict with; initially all con ict counts are zero. A colour is classed as deleted from a vertex domain if and only if its con ict count is greater than zero. A domain's size is the number of its non-deleted colours. The memory requirement for con ict counts is not excessive: for n vertices and k colours k n con ict counts are needed, which is roughly the amount of memory required to represent the problem. They may be updated incrementally: on colouring/uncolouring a vertex, the con ict count for that colour in each adjacent vertex is incremented/decremented. However, they cause a signicant runtime overhead compared to standard forward check- ing, because they are updated for the domains of uncoloured and coloured vertices. We can now design a local search algorithm on FC-consistent partial col- orations. The algorithm chosen is rather simple, starting exactly as standard FC: it selects a vertex for colouring, nds a colour that can be used without causing domain wipe-out, colours the vertex, updates the domains of adjacent vertices, and repeats by selecting another vertex. The only dierence so far is that domains are maintained by con ict counts. However, on reaching a dead-end , where the selected vertex cannot be coloured, the new algorithm behaves dierently to standard FC. It heuristically selects a vertex to be un- coloured, instead of selecting the most recently coloured vertex. No attempt is made to backtrack systematically, so completeness is lost. Because there is now no obvious criterion for deciding when to stop backtracking and start colouring vertices again, we introduce a parameter B 1. On reaching a dead-end B vertices are uncoloured, and colouring resumes. Note that the vertices selected for colouring and uncolouring may follow dierent heuris- tics, so that the set of coloured vertices may change rapidly during search. B plays the part of a noise parameter (or the temperature in simulated an- nealing), controlling the permitted disruption to the state on reaching a local minimum. It remains to ll in details by describing three heuristics: selecting B coloured vertices for uncolouring (CVERTEX), selecting an uncoloured vertex for colouring (UVERTEX), and selecting colours to try when colouring a vertex (COLOUR). We consider two alternative UVERTEX rules: select the vertex with smallest current domain; break ties by selecting the vertex adjacent to the greatest number of uncoloured ver- tices; break further ties randomly. Nonsingleton: randomly select a vertex with more than one colour in its current domain; if none exists then select a vertex randomly. The Brelaz heuristic is an obvious choice. The idea behind the Nonsingleton heuristic is to emulate the MAXIS algorithm, which constructs independent sets of vertices, whereas DSATUR constructs cliques. By selecting vertices using an inverse of the Brelaz heuristic, and thus focusing on vertices that are as independent as possible from those currently coloured, we might expect to obtain a forward-checking analogue to MAXIS. This was tested and, perhaps surprisingly, performed rather well on random graphs, whereas the Brelaz heuristic performed poorly. However, the weaker Nonsingleton heuristic performs better, possibly because of its greater exibility in selecting a vertex. It is discussed further in Section 4. Given a free choice of vertices for uncolouring, which should be selected? An obvious idea is to use an inverse of Brelaz: uncolour a vertex with large domain and small degree (note that because con ict counts are updated irrespective of whether a vertex is coloured, coloured vertices also have domain sizes). In tests this often caused stagnation, but the weaker Nonsingleton heuristic (applied to coloured vertices) works well. To further reduce stagna- tion, with probability 1=n (where n is the number of vertices in the graph) the CVERTEX rule selects a vertex randomly instead of by domain size. A random ordering on domain values works well, but performance can be improved by remembering the previous colour of each vertex (if it was coloured earlier). The COLOUR rule ips between two modes: initially it prefers dierent colours to those remembered for each vertex; if a dierent colour is successfully used, the rule ips to preferring the remembered colour; when CVERTEX is next invoked it ips back to preferring a dierent colour. The aim of this rule is to minimise disruption to colorations as the set of coloured vertices changes, while avoiding null local moves. function for while U 6= fg let let colouring u to d does not cause domain wipe-outg for i=1 to min(B; jCj) uncolour c and update domains else colour u to COLOUR(D) and update domains return coloration Figure 1: FC partial coloration neighbourhood search for xed k The new algorithm FCNS (FC-consistent partial coloration Neighbourhood is shown in Figure 1. k 1 is the permitted number of colours and B 1 is the noise parameter. C is the current set of coloured vertices, initialised to fg. U is the current set of uncoloured vertices, initialised to the full set of n vertices g. Each vertex has a domain of colours that are FC-consistent with the current partial coloration, initialised to the full set of colours kg. The algorithm proceeds by selecting uncoloured vertices using the UVERTEX rule, and colours them using the COLOUR rule. On reaching a dead-end (D = fg) it uncolours B vertices, each selected by the CVERTEX rule. Termination is not guaranteed but occurs if all vertices are coloured The algorithm can be used to nd a near-optimal colouring by applying it iteratively in an obvious way: start with large k (for example and apply the algorithm; on nding a total coloration using k 0 k colours, restart the algorithm with k 0 1 colours; repeat until reaching a target number of colours or a specied time. Performance is improved by starting each iteration with a coloration similar to the previous one: colour assignments are stored between iterations, and until the rst dead-end occurs each vertex is assigned its previous colour where possible. It is also possible to speculatively reduce k further in the hope of nding better colourings more quickly. However, this aspiration approach does not always speed up search, because inadvertently choosing k less than the chromatic number of the graph runs the risk of spending a long time in fruitless search. Aspiration is not used in current FCNS implementations. 3 Experimental results FCNS is now evaluated using published results for several other algorithms on the well-known DIMACS [11] 1 benchmarks. They are Culberson & Luo's Iterated Greedy (IG) [4], Morgenstern's distributed IMPASSE [17], Wheel Optimization (SWO) [12] and Glover, Parker & Ryan's TABU [8]. The TABU algorithm combines the TABU meta-heuristic with branch-and- bound. SWO operates in two search spaces: a solution space and a prioritisation space. Both searches in uence each other: each solution is analysed and used to change the prioritisation, which guides the search strategy used to nd the next solution, found by restarting the search. We use a standard set of benchmarks taken from the DIMACS web site. Geometric graphs Rx.y and DSJRx.y are generated by randomly placing x vertices in a unit square, then assigning edges between any two vertices with Euclidean distance less than y=10 between them; a graph denoted by Gc is the complement of the graph G. The names R and DSJR re ect dierent sources, but are (we believe) the same type of graph. Random graphs Cn.p and DSJCn.p have n vertices, an edge being assigned between any two vertices with a xed probability p=10. The names C and DSJC again re ect dierent sources. Flat graphs contain colorations that are hidden in such a way as to mislead Brelaz-style heuristics; a graph atn c x contains vertices and a known hidden (though not necessarily optimal) c-colouring. Leighton graphs le450 15x are derived from scheduling, and have 450 edges and known chromatic number 15. Graph colouring is closely related to the timetabling problem and there are two timetabling graphs; the school1 problem is derived from timetabling data from a real high school with around 500 students; the school1 nsh problem is derived from the same data but ignores study halls. Register allocation graphs are used in compilers to assign variables to registers, with the aim of avoiding the placement of two variables in the same register when both may be active; there is one such graph, mulsol.i.1. The latin square graph latin sqr 10 is derived from a standard problem in design theory. Figure reproduces published results for SWO, IG, d-IMP (distributed IMPASSE), p-IMP (parallel IMPASSE) and TABU, and Figure 3 shows results for FCNS with the Brelaz (FCNS-b) and Nonsingleton (fCNS-n) heuris- tics. All times are normalised to our machine (a 300 MHz DEC Alphaserver 1000A 5/300 under Unix) using benchmark timings from [11]; the DIMACS benchmark program dfmax r500.5 takes 46.2 seconds on our machine. The times for parallel IMPASSE were not normalised because of its parallel platform (a 32-processor CM-5). In both tables k is the number of colours used and t is the time taken in seconds. In Figure 3 B is the value used for the parameter. Its value was chosen after a few runs to nd an appropriate setting. This ad hoc approach is unfortunately necessary with many local search algorithms; TABU has a list length parameter, and some algorithms have more than one parameter (for example several local search algorithms for the satisability problem). The initial number of colours k 0 for FCNS was set to the worst k found by the other algorithms in each case (except where our algorithms were even worse, when higher values were used). FCNS was halted on reaching the target k, which was selected after a few experimental runs. Times shown for FCNS are mean times taken to reach k from k 0 averaged runs (more for short times). Experimental details for the other algorithms vary (for details see the cited papers). Brie y, SWO was terminated after 1000 iterations, IG after 1000 iterations without improvement, TABU after an hour or sooner if a lack-of-progress condition was satised, distributed IMPASSE used conditions depending on the problem but always halted on reaching a specied target k, and parallel IMPASSE ran for 3 hours then reported the time taken to nd the best solution. The use of a time limit instead of a target number of colours explains the occasional fractional values of k. First we discuss FCNS-b, which is clearly the best algorithm on the geometric graphs. On R1000.5 and DSJR500.5 it nds the best reported colour- ings, and on most of the others it nds equally good colourings in shorter times. The geometric graphs are randomly generated but closely related to a real problem: frequency allocation [10]. FCNS-b is therefore a promising algorithm for solving such problems, and this is an area for future research. It also performs very well on the school and mulsol graphs, roughly matching SWO IG d-IMP p-IMP TABU problem school1 nsh 14 3.9 14.1 4.8 14 <0.24 14 66.4 26 16.8 mulsol.i.1 at300 26 0 35.8 6.4 37.1 4.1 26 5.4 32.4 6637 41 1849 at300 28 0 35.7 6.4 37 5.2 31 1028 33 1914 41 1849 Figure 2: Previous results for DIMACS benchmarks FCNS-b FCNS-n problem Figure 3: FNCS results for DIMACS benchmarks the performance of distributed IMPASSE. We also tested Mehrotra & Trick's implementation 2 on the geometric graphs because it is known to perform well on such graphs. On those with edge probability 0.1 it found the same colourings in a slightly shorted time than FCNS-b. On those with edge probability 0.5 it quickly found good colourings but then made no further progress for a long time. For R125.5 it found a 36-colouring in 63.4 seconds, for R250.5 a 66-colouring in 2.9 seconds, for DSJR500.5 a 130-colouring in 17.6 seconds, and for R1000.5 a 246-colouring in 75.8 seconds; no further progress was made after several minutes. FCNS-b clearly scales better than DSATUR, nding better colourings on the larger graphs. However, it is very poor on the random, at and latin square graphs, and mediocre on the Leighton graphs. Next we discuss FCNS-n. On the geometric and school graphs it is poor, sometimes the worst algorithm, and (like FCNS-b) mediocre on the Leighton graphs, but on the random, at and latin square graphs it is beaten only by distributed IMPASSE. This is presumably due to the use by distributed IMPASSE of the XRLF algorithm [10] to generate initial colorations: parallel IMPASSE does not use XRLF and is beaten by FCNS-n on random and at graphs. However, other algorithms are also better than FCNS-n on random graphs. For example on G 1000;0:5 graphs the best algorithms nd colourings in the low- or mid-80s. To further investigate FCNS-n we applied it to equipartite graphs, which have been studied by several researchers. A k-colourable equipartite graph is generated by partitioning its vertices into k subsets, which are as equally-sized as possible, the smallest being no more than 1 vertex smaller than the largest. Edges are assigned with probability p, disallowing edges between vertices in the same subset. This guarantees a k-colouring but does not preclude better colourings. Eiben, van der Hauw & van Hemert [5] apply evolutionary algorithms to 3-colourable equipartite graphs with 200 vertices. They report low success rates on graphs with low density, especially around where a phase transition occurs. Minton et al. [16] also report that the Min- Con icts local search algorithm has di-culties with similar problems, but that a backtracking version of DSATUR solves them easily for 3 n 180. Yugami et al. [27] apply local search with constraint propagation to the same problems and obtain improved results. FCNS-n solves these problems easily: nds 3-colourings in approximately 3 seconds. Moving to larger problems, Culberson et al. [4] show that IG can nd hidden k-colourings for G 1000;0:5 equipartite graphs with k 60. FCNS-n is also able to do this and can go a little further. The algorithm was quite insensitive to B until k 55, after which it became more sensitive. The optimal value for was approximately the problems rapidly became harder and the optimal value of B fell. It found a hidden 67-colouring after several hours computation with but failed to nd a hidden 68-colouring. So far as we known 67 is the highest value of k solved for this class of graph. In further experiments FCNS-n also managed to nd the hidden colourings in at1000 f50,60g 0, by setting starting with the target colouring (50 or 60) specied as the initial colouring. However, these results took much trial and error to achieve, so they were not included in our table. It is perhaps surprising that the Nonsingleton heuristic should be successful at all, let alone competitive. In particular, if a vertex has domain size 1 (hence only one possible colour) then Brelaz will select it before a vertex with larger domain (hence more than one possibility), but Nonsingleton will delay colouring such vertices as long as possible. To investigate the eect of adding \forced moves" another variant was tried: select a vertex with domain size 1 if one exists, otherwise select one with maximum domain size. However, this variant was inferior to both Brelaz and Nonsingleton. We speculate that Nonsingleton causes FCNS to behave in a similar way to independent set-based algorithms such as MAXIS, by focusing search on low-degree vertices. A better algorithm might be obtained by explicitly designing it to nd independent sets, and applying forward checking. Another research direction is the design of new vertex orderings, with the aim of improving FCNS's performance on random, Leighton and latin square graphs. The noise parameter B unfortunately requires tuning to each graph. As with any noise parameter, the eect of setting B too low is stagnation: FCNS will never nd a colouring because it becomes trapped in a local minimum. The eect of setting B too high is less serious, simply increasing the time taken to nd a solution, but the increase depends on the problem. The performance of FCNS-n seems to be fairly insensitive to the value of B when nding an 18-colouring for DSJC125.5, while on DSJC1000.5 increasing B from 1 to 2 slows it down greatly | or equivalently, prevents it from nding good colourings in the same time. FCNS-b seems less sensitive, but can still be slowed down signicantly by too much noise. We experimented with variable noise levels to try to reduce sensitivity to noise, but with inconclusive results. A slightly surprising feature of B is that, on several graphs (for example R250.5), best results were obtained for FCNS-b and FCNS-n using dierent values of B. However, perhaps this should not be surprising, because the two algorithms focus on dierent regions of the graphs and therefore might be expected to encounter local minima of dierent depths. The main dierence between FCNS and other stochastic colouring algorithms is that it performs forward checking. It also has an additional advantage over IG and SWO: incrementality . IG and SWO are not incremental because restarting is an expensive move, whereas IMPASSE, TABU and FCNS make small, cheap moves in the search space. This is pointed out as a source of ine-ciency by Joslin & Clements [12], and they propose hybrids of SWO with local search for future work. Graph colouring is a binary constraint satisfaction problem (given xed k), and the use of con ict counts to perform forward checking in local search is easily generalised to other such problems. It can be further generalised to non-binary constraint problems, and this has been done for propositional satisability (SAT). Experimental results are very promising: on some large, structured SAT problems it out-performs current systematic and local search algorithms [20]. In fact our colouring and SAT algorithms are instances of a general-purpose approach to combinatorial optimisation and constraint satisfaction, which we call Constrained Local Search (CLS). The aim of CLS is to enhance local search with constraint programming techniques used in systematic search. It has also given good results on other SAT problems [20], maximum clique problems [22], Golomb rulers [22] and a hard optimisation problem (the generation of low-autocorrelation binary sequences) [21]. The general approach is to take an eective backtracking algorithm and replace systematic by randomised backtracking, usually improving its scalability at the expense of completeness. It might be argued that FCNS (or more generally CLS) is not a local search algorithm, but simply a randomised backtracker. It certainly is a randomised backtracker and has much in common with Dynamic Backtracking (DB), which also allows the removal of early assignments without aecting the assignments made since. This increased exibility of backtracking was a stated aim in the design of DB, and a later hybrid algorithm called Partial Order Dynamic Backtracking [7] achieved even greater exibility. Is FCNS simply an inferior version of DB, sacricing completeness to no good pur- pose? A counter-example to this view is the random 3-SAT problem, on which DB is slower than depth-rst search [1] while CLS scales precisely as local search [21]. Our view is that FCNS stochastically explores a space of FC-consistent partial colorations by local search; the objective function it minimises is the number of uncoloured vertices. However, to some extent the question is academic: even if FCNS is not local search, experimental results show that it captures its essence, successfully solving problems that are beyond the range of systematic backtracking. There are several other hybrids of local search and constraint techniques. The simplest hybrid is a parallel or distributed implementation of more than one algorithm, as in the IMPASSE algorithms used for colouring. Schaerf's timetabling algorithm [24], extended to constraint satisfaction problems, searches the space of all partial assignments (not only the consistent ones) using an objective function that includes a measure of constraint violation. This is a dierent space again than those searched by current colouring algorithms and FCNS. In graph colouring terms this space may be called the partial colorations as opposed to the consistent partial colorations explored by IMPASSE. Jussien & Lhomme's Path-Repair Algorithm [13] is described as a generalisation of Schaerf's approach that includes learning (allowing complete versions to be devised) and a TABU list. Yugami, Ohta & Hara's EFLOP algorithm [27] uses constraint propagation to escape local minima, while allowing some constraint violation. However, forward checking is not maintained throughout the search, as it is in FCNS. Ginsberg & McAllester's Partial-order Dynamic Backtracking [7] is a hybrid of the Dynamic Back-tracking algorithm with a local search algorithm [6], enabling it to follow local gradients in the search space. Pesant & Gendreau [18] apply systematic branch-and-bound search to e-ciently explore local search neighbourhoods. The two-phase algorithm of Zhang & Zhang [28] searches a space of partial variable assignments, alternating backtracking search with stochastic local search on the same data structure. It can be tuned to dierent problems by spending more time in either phase. Yokoo's Weak Commitment Search [26] (WCS) greedily constructs consistent partial assignments. On reaching a dead-end it randomly restarts, and uses learning to maintain complete- ness. Richards & Richards [23] describe a SAT algorithm called learn-SAT based on WCS. Shaw [25] describes a vehicle routing algorithm called Large Neighbourhood Search. It performs local search and uses backtracking with constraint propagation to test the legality of moves. Each of these algorithms either permits constraint violation or uses learn- ing, or both. Constraint violation implies, in the view of this author, that constraints are being under-used. This may be a drawback when solving highly structured problems: the best graph colouring results for structured problems are obtained by algorithms such as IMPASSE and FCNS, which do not violate constraints. The use of learning is a drawback when solving large problems. It can be restricted to use only polynomial memory, but combinatorial problems may be very large. The FCNS approach combines constraint handling and local search, making cheap local moves and avoiding memory-intensive learning techniques. We believe that this combination of features makes it ideal for large, highly constrained problems. --R The hazards of fancy backtracking Exploring the k-colorable landscape with iterated greedy Graph coloring with adaptive evolutionary algorithms Journal of Arti GSAT and dynamic backtracking Coloring by tabu branch and bound Increasing tree search e-ciency for constraint satisfaction problems Optimization by simulated annealing: an experimental evaluation Journal of Arti The path-repair algorithm Experiments with parallel graph coloring heuristics and applications of graph coloring A column generation approach to graph colouring Minimizing con- icts: a heuristic repair method for constraint satisfaction and scheduling problems Distributed coloration neighborhood search A view of local search in constraint pro- gramming Using an incomplete version of dynamic backtracking for graph colouring Stochastic local search in constrained spaces A hybrid search architecture applied to hard random 3-SAT and low-autocorrelation binary sequences Trading completeness for scalability: hybrid search for cliques and rulers Combining local search and look-ahead for scheduling and constraint satisfaction problems Using constraint programming and local search methods to solve vehicle routing problems Improving repair-based constraint satisfaction methods by value propagation Combining local search and backtracking techniques for constraint satisfaction --TR --CTR Steven Prestwich, SAT problems with chains of dependent variables, Discrete Applied Mathematics, v.130 n.2, p.329-350, 15 August Marco Chiarandini , Thomas Sttzle, Stochastic Local Search Algorithms for Graph Set T-colouring and Frequency Assignment, Constraints, v.12 n.3, p.371-403, September 2007
forward checking;graph colouring;coloration neighbourhood
590526
Approximate Qualitative Temporal Reasoning.
We partition the time-line in different ways, for example, into minutes, hours, days, etc. When reasoning about relations between events and processes we often reason about their location within such partitions. For example, i>x happened yesterday and i>y happened today, consequently i>x and i>y are disjoint. Reasoning about these temporal granularities so far has focussed on temporal units (relations between minute, hour slots). I shall argue in this paper that in our representations and reasoning procedures we need into account that events and processes often lie skew to the cells of our partitions. For example, happened yesterday does not mean that i>x started at 12 a.m. and ended 0 p.m. This has the consequence that our descriptions of temporal location of events and processes are often approximate and rough in nature rather than exact and crisp. In this paper I describe representation and reasoning methods that take the approximate character of our descriptions and the resulting limits (granularity) of our knowledge explicitly into account.
Introduction Every temporal object and every spatio-temporal object is located at a unique region of time bounded by the begin and the end of its existence. In every moment of time a spatio-temporal object, o, is exactly located at a single region, x, of space (Casati & Varzi 1995). This region is the exact or precise location of o at the time point t, at t. Spatio-temporal wholes have temporal parts, which are located at parts of the temporal regions occupied by their wholes. Consider, for example, the region of time, y, where the object, o, is located tempo- rally, while being spatially located at the spatial region x. If y is a maximal connected temporal region, i.e., once spatially located at x for a while, left and never came back, then y is bounded by the time instances (points) t 1 and t 2 . Since time is a totally ordered set of time points forming a directed one-dimensional space (McTaggart 1927; Geach 1966), we have t 1 < t 2 . In knowledge representation we are interested in representing spatio-temporal reality as experienced by human beings. In this context it is essential to represent spatio-temporal location (Galton 1997). In this paper I concentrate on representing temporal location. One way of representing temporal location is to represent qualitative relationships between temporal regions occupied by temporal and spatio-temporal objects and their parts (Allen 1983). Copyright c 2000, American Association for Artificial Intelligence (www.aaai.org). All rights reserved. Human knowledge is gained by observations and reasoning about observations. (Bittner 1999) argued 1 that by means of observation and measurement (a precise form of humans cannot know the exact spatial and, hence, exact temporal location. Observations and measurement yield knowledge about approximate spatio-temporal location, i.e., knowledge about relations between spatio-temporal objects and cells of regional partitions of space and time. Regional partitions are sets of regions (cells) that do not overlap but sum up the whole space. Regional partitions are created by measurement and observation processes. Approximate location can be known by observing (qualitative) relations between objects and the cells of the underlying regional partitions. Consider, for example, the measurement of temporal loca- tion. Measurement of temporal location is based on clocks. A clock creates a regional partition of the time-line. The cells forming this partition are time intervals separated by 'clock ticks'. Measurement of temporal location involves counting time intervals and observing relationships between time intervals and (temporal parts of) temporal or spatio-temporal objects. No matter how fine the resolution of the partition there are always partition cells that are disjoint to (the exact region of) the observed object, there may be partition cells that are completely covered by (the exact region of) the observed object, and there always are partition cells that are partly covered by (the exact region of) the observed object. Consequently, observing spatio-temporal location means observing relations between partition cells and regions occupied by spatio-temporal objects, i.e., observing approximate location rather than exact location. Other examples of regional partition in which approximate temporal location is observed is the partition of the time line into past and future separated by the present moment, the hours of the day, forenoon and afternoon, the four seasons. Con- sequently, in the context of knowledge representation reasoning about approximate spatio-temporal location, i.e., approximations of spatial and temporal regions, is more important than reasoning about exact location, i.e., spatial and temporal regions themselves. In the remainder of this paper I omit the distinction between objects and spatial and temporal regions and the 1 Based on (Carnap 1966). (functional) relation of (exact) location between them and concentrate on the approximation of the exact regions (of objects) with respect to an underlying regional partition. Moreover, I concentrate on temporal regions and approximations of temporal regions. This paper builds on (Bittner & Stell 1998) and (Bittner & Stell 2000), in which various ways of providing qualitative approximations of regions with respect to a partition of the plane as well as reasoning about those approximations were described. (Bittner & Stell 2000) showed that approximate qualitative reasoning is based on: (1) Jointly exhaustive and pair-wise disjoint sets of qualitative relations between exact re- gions. These relations need to be defined in terms of the meet operation of the underlying Boolean algebra structure of the domain of regions. As a set these relations must form a lattice with bottom and top element. (2) Approximations of regions with respect to a regional partition of the underlying space. (3) Pairs of join and meet operations on those ap- proximations, which approximate join and meet operations on exact regions. This this is reflected by the structure of this paper: I start with the definition of qualitative relations between temporal regions. I distinguish boundary sensitive and boundary insensitive sets of relations and relations between regions in a directed and non-directed underlying one-dimensional space. Based on the definition of approximations of temporal regions with respect to an underlying regional partition I then generalize the definitions of relations between temporal regions to definitions of relations between approximations of regions. This provides the formal basis for qualitative temporal reasoning about approximate location in time. The conclusions are given in the end. Relations between one dimensional regions Boundary insensitive relations RCC5 relations Given two regions x and y boundary insensitive topological relation (RCC5 relations 2 them can be determined by considering the triple of boolean values (Bittner & Stell 2000): The correspondence between such triples and boundary insensitive relations between regions on an undirected line is given in the following table (Bittner & Stell 2000). I use the notion RCC in order to stress the correspondence between the relations defined in this paper and relations defined by Cohn and his co-workers in terms of the region connection calculus (RCC) (Cohn et al. 1997). Correspondence in this context means that I am talking about regular regions that satisfy the RCC-axioms (Randell, Cui, & Cohn 1992) and that similar relations could be defined or have been defined in terms of RCC, e.g., (Randell, Cui, & Cohn 1992; Cohn, Gooday, & Bennett 1994; Cohn et al. 1997). I am going to use sub- and superscripts (e.g., RCC where the superscript refers to the number of relations in the denoted set and the subscript refers to the dimension of the regions and the embedding space. The set of triples is partially ordered by setting where the Boolean values are ordered by F < T. The Hasse diagram is given in the diagram below. (Bittner & Stell 2000) call this graph the RCC5 lattice to distinguish it from the conceptual neighborhood graph (Goodday & Cohn 1994). I@ @ @ I@ @ @ relations Given two one dimensional regions x and y. I assume that x and y are maximal connected one dimensional regions, i.e., intervals. Boundary insensitive topological relation between intervals x and y on a directed line (RCC 9 1 relations) can be determined by considering the triple of truth values: where where x y :(y x) x y :(y x) and where x y x y with L(x) (R(y)) is the one dimensional region occupying the whole line left (right) 3 of x. The intuition behind FLO and FLI (FRO and FRI) is "false because of parts 'sticking out' to the left (right)'' 4 . The triples formally describe jointly exhaustive and pair-wise disjoint relations under the assumption that x and y are intervals in a one dimensional directed space. The correspondence between the triples and the boundary insensitive relations between intervals is given in the following table. FRO FRO FRO DRR For example. The intuition behind DRL(x; y) is that x and y do not overlap and x is left of y. The intuition behind POL(x; y) is that x and y do overlap without containing each other and the non overlapping parts of x are left of y. The intuition behind PPL(x; y) is that x is contained in y but x does not cover the very right parts of y. Possible geometric interpretations of the relations defined above are given in Figure 1. Assuming the ordering FLO < FLI < T < FRI < FRO a lattice is formed, which has (FLO;FLO;FLO) as minimal element and (FRO;FRO;FRO) as maximal element. x y POR(x,y) Figure 1: Possible geometric interpretations of the RCC 9 relations. Boundary sensitive relations RCC8 relations In order to describe boundary sensitive relations between regions x and y (Bittner & Stell 2000) use 3 I use the spatial metaphor of a line extending from the left to the right rather than the time-line extending from the past to the future in order to focus on the aspects of the time-line as a one-dimensional directed space. Time itself is much more difficult. For example, it is not clear if the future already exists yet (Broad 1923). 4 Notice that in the case of FLI and FRI this dose not exclude that there are also 'parts sticking out' to the opposite side. a triple, where the three entries may take one of three truth values rather than the two Boolean ones. The scheme has the form: where T if the interiors of x and y overlap; if only the boundaries x and y overlap; F if there is no overlap between x and and where x^y x => > > > > > > > > > > > < T if x is contained in the interior of y and the boundaries are either disjoint or identical, i.e.,x if x is contained in y and the boundaries are not disjoint and not identical, F if x is not contained within and similarly for x y. The correspondence between such triples and boundary sensitive topological relations is given in the following table (Bittner & Stell 2000). (Bittner & Stell 2000) define F < M < T and call the corresponding Hasse diagram (diagram below) RCC8 lattice to distinguish it from the conceptual neighborhood graph (Goodday & Cohn 1994). I@ @ @ In this paper I concentrate on regions of one-dimensional space and relations between them. In order to distinguish sets of relations between one dimensional regions from relations between regions of higher dimension I use the notion rather than RCC8. Possible geometric interpretations of their RCC 8 1 relations are given in Figure 2. y x Figure 2: Geometric interpretations of RCC 8 relations between one-dimensional regions of a non-directed line. 1 relations In order to describe boundary sensitive relations between intervals on a directed line (RCC 15 define the relationship between x and y by using a triple, where the three entries may take one of four truth values. The scheme has the form where and where and similarly for x y. The correspondence between such triples, boundary sensitive topological relations between intervals on a directed line, and the 13 relations defined by (Allen 1983) is given in the table below. 5 To be distinguished from RCC15 relations (Cohn et al. 1997) between concave regions of higher dimension. FLO FLO FLO DCL before FRO FRO FRO DCR after MLO FLO FLO ECL meets MRO FRO FRO ECR meets i starts during during during i during i We define FLO < MLO < FLI < MLI < T < MRI < FRI < MRO < FRO and call the corresponding Hasse diagram RCC 15 1 lattice to distinguish it from the conceptual neighborhood graph (Freksa 1992). Possible geometric interpretations of the lower RCC 15 1 relations are given in Figure 3. x y Figure 3: Geometric interpretations of the lower relations between connected intervals. Approximations Approximating regions Boundary insensitive approximation Consider the set of regions, R, of a one-dimensional space. By imposing a partition, G, on R we can approximate elements of R by elements of G 3 (Bittner & Stell 1998). That is, we approximate regions in R by functions from G to the set nog. The function which assigns to each region r 2 R its approximation will be denoted 3 G 3 . The value of ( 3 r)g is fo if r covers all the of the cell g, it is po if r covers some but not all of the interior of g, and it is no if there is no overlap between r and g. (Bittner & Stell 1998) call the elements of G 3 the overlap & containment sensitive approximations of regions r 2 R with respect to the underlying regional partition G. Boundary sensitive approximation Consider one dimensional non-directed space. We can further refine the approximation of regions R with respect to the partition G by taking boundary points shared by neighboring partition regions into account. That is, we approximate regions in R by functions from GG to the set nog. The function which assigns to each region r 2 R its boundary sensitive approximation will be denoted 4 GG 4 . The value of ( 4 r)(g covers all of the cell g i , it is bo if r covers the boundary point, (g shared by the cell g i and g j and some but not all of the interior of g i , it is nbo if r does not cover the boundary point (g some but not all of the interior of g i , and it is no if there is no overlap between r and g i . The Semantic of approximate regions Each approximate region XG stands for a set of precise regions, i.e., all those precise regions having the approximation X . This set which will be denoted [[X a semantic for approximate regions. Where ever the context is clear the superscript is omitted. Approximate operations The domain of regions is equipped with join and meet op- erations, _ and ^. (Bittner & Stell 1998) showed that join meet operations on regions can be approximated by pairs of greatest minimal and least maximal operations on approx- imations. In this paper I discuss the operations on boundary insensitive approximations and boundary sensitive approximations. A detailed discussion can be found in (Bittner & Stell 1998). Boundary insensitive operations Firstly we define operations on the set nog. no no no no po no no po no po fo no no no no po no po po no po fo These operations extend to elements of G 3 (i.e. the set of functions from G to and similarly for ^ . Boundary sensitive operations We define the operations on the set no no no no no nbo no nbo nbo nbo bo no nbo bo bo no nbo bo fo These operations extend to elements of GG 4 (i.e. the set of functions from G G to The definition of the operations ^ is slightly more com- plicated. In this case we need to take the approximation values referring to both boundary points (g account. Let be the set of pairs of approximation values of X and Y with respect to g i . We define the operation X ^Y as follows: defined as no no no no no nbo no bo no no nbo bo fo and (N) is defined as no if (bo; bo) 62 N nbo if (bo; bo) This definition corresponds to the definitions of operations on boundary sensitive approximations of two-dimensional regions in the plane discussed in (Bittner & Stell 1998). Semantic and Syntactic Generalizations of RCC* (Bittner & Stell 2000) showed that there are two approaches to generalizing RCC relations between precise regions to approximate ones: the semantic and the syntactic. Semantic We can define the RCC relationship between approximate regions X and Y to be the set of relationships which occur between any pair of precise regions approximated by X and Y . That is, we can define Syntactic We can take a formal definition of RCC in the precise case, which uses operations on R, and generalize this to work with approximate regions by replacing the operations on R by analogous ones for G or GG . In the remainder of this section I discuss syntactic and semantic generalizations for RCC5 , RCC 8 1 , and 1 . Generalization of RCC5 relations Syntactic generalization If X and Y are approximate regions (i.e. functions from G to 3 ) we can consider the two triples of Boolean values (Bittner & Stell 2000): In the context of approximate regions, the bottom element, ?, is the function from G to 3 which takes the value no for every element of G. Each of the above triples defines an RCC5 relation, so the relation between X and Y can be measured by a pair of RCC5 relations. These relations will be denoted by R(X;Y ) and R(X;Y ). Theorem 1 ((Bittner & Stell 2000)) The pairs (R(X; Y ); R(X;Y )) which can occur are all pairs (a; b) where a b with the exception of (PP; EQ) and Correspondence of semantic and syntactic generalization Let the syntactic generalization of RCC5 be defined by where R and R are as defined above. Theorem 2 ((Bittner & Stell 2000)) For any approximate regions X and Y syntactic and semantic generalization of RCC5 are equivalent in the sense that where RCC5 is the set fEQ; PP; PPi; PO;DRg, and is the ordering in the RCC5 lattice. Generalization of RCC 8relations Syntactic generalization Let X and Y be boundary sensitive approximations of regions x and y. The generalized scheme has the form where and where and similarly for X ^Y Y , X,and X ^Y Y . In this context the bottom element, ?, is either the value no or the function from G G towhich takes the value no for every element of GG. Assume the partial order of the RCC 8 only if the least relation 1 -relation that can hold between x 2 [[X boundary intersection of -(x) and -(y) at a boundary point, (g of the underlying partition G. only if the greatest RCC 8 relation that can hold between x 2 [[X boundary intersection at a boundary point in G. For a detailed discussion of the 2D case see (Bittner & Stell 2000). Each of the above triples defines a RCC 8 the relation between X and Y can be measured by a pair of RCC 8 1 relations. These relations will be denoted by R 8 (X; Y ) and R 8 (X; Y ). Let X and Y be approximations of one dimensional regions in one dimensional space. Then the following holds: Theorem 3 The pairs (R 8 (X; Y can occur are all pairs (a; b) where a b with the exception of (TPP; EQ), (TPPi; EQ),(NTPP; EQ), (DC; EC), (DC; TPP), (DC; TPPi), EC; NTPP), 6 This is an application of theorem 5 in (Bittner & Stell 2000) to the one-dimensional case. Correspondence of syntactic and semantic generalization Let SEM(X;Y ) be a set of RCC 8 relations defined as ]]g. Theorem 4 If there are G such that (X(g Assume (X(g possibly This conflicts with We define the semantically corrected syntactic generalization of RCC 8 as: (R 8 there are that (X(g otherwise. The semantic generalization of 1 relations is defined as SEM(X;Y R 8 )g. Theorem 5 For any boundary sensitive approximations X and Y of regular one dimensional regions, the syntactic and semantic generalization of RCC 8 are equivalent in the sense that SYN(X;Y Generalization of RCC 9relations Syntactic generalization Let X and Y be boundary sensitive 9 approximatons of regions x and y. Then we can consider the two triples of Boolean values: where and where 7 This is an application of theorem 6 in (Bittner & Stell 2000) to the one-dimensional case. 8 This is an application of theorem 7 in (Bittner & Stell 2000) to the one-dimensional case. 9 We need boundary sensitive approximations since we need to approximate intervals, i.e., maximally connected temporal regions. and similarly for and X ^Y Y . We define X Y as and similarly X Y using R(X) and R(Y ), where L and R are defined as follows. Firstly, we define the complement operation no nbo bo fo Secondly, assuming that partition cells g i are numbered in increasing order in direction of the underlying space, we define L(Y ) as no otherwise and R(Y ) is defined as (R(Y no otherwise Each of the above triples defines an RCC 9 the relation between X and Y can be measured by a pair of RCC 9 1 relations. These relations will be denoted by R 9 (X; Y ) and R 9 (X; Y ). Theorem 6 The pairs that can occur are all pairs (a; b) where a b EQ and EQ a b with the exception of (PPL; EQ), (PPR; EQ), The pairs (PPL; EQ), (PPR; EQ), (PPiL; EQ), (PPiR; EQ) cannot occur, since they are refinements of the relations (PP; EQ), (PPi; EQ), which cannot occur in the undirected case. The pair (EQ; DRR) cannot occur due to the non-symmetric definition of FL and FR. The pair (DRL; EQ) represents the most indeterminate case. Since (DRL; EQ) is consistent with (EQ; DRR) and (DRL; EQ) was chosen arbitrarily, (DRL; EQ) is corrected syntactically to (DRL; DRR). The corrected relation will be denoted by R 9 Correspondence of semantic and syntactic generalization Let the syntactic generalization of RCC 9 1 be defined by R 9 and R 9 c are defined as discussed above. Proposition 1 For approximations X and Y syntactic and semantic generalization of RCC 9 1 relations are equivalent in the sense that 1 is the set RCC 9 1 relations and is the ordering in the RCC 9 lattice. Generalization of RCC 15relations Syntactic generalization If X and Y are boundary sensitive approximations of intervals x and y in a directed one-dimensional space then we can consider the two triples of Boolean values: where and similarly for Each of the above triples provides a RCC 15 the relation between X and Y can be measured by a pair of 1 relations. These relations will be denoted by R 15 and R 15 (X; Y ). The pairs of relations that can occur are all pairs (a; b) where a b EQ and EQ a b with the exception of pairs of relations that are refinements of pairs of relations that cannot occur in the undirected case (RCC 9 theorem or that cannot occur in the boundary insensitive case (RCC 8 theorem 3). Correspondence of semantic and syntactic generalization Corresponding to the generalization of the RCC 8 1 and the RCC 9 relations syntactic corrections are needed in order to generalize RCC 15 relations between intervals, x and y, to pairs of RCC 15 relations between approximations X Firstly. Corresponding to the RCC 8 1 case we define R 15 are G such that (X(g R 15 to the RCC 9 1 case the pair (DCL; EQ) represents the most indeterminate case. Since (DCL; EQ) is consistent with (EQ; DCR) and (DCL; EQ) was chosen arbitrar- ily, (DCL; EQ) is corrected syntactically to (DCL; DCR). The corrected relation will be denoted by R 15 Let the syntactic generalization of RCC 15 1 be defined by c and R 15 c are defined as discussed above. Proposition 2 For approximations X and Y syntactic and semantic generalization of RCC 15 1 relations are equivalent in the sense that maxfR 15 where RCC 15 1 is the set RCC 15 1 relations and is the ordering in the RCC 15 lattice. Conclusions In this paper I defined methods of approximate qualitative temporal reasoning. Approximate qualitative temporal reasoning is based on: 1. Jointly exhaustive and pair-wise disjoint sets of qualitative relations between exact regions, which are defined in terms of the meet operation of the underlying Boolean algebra structure of the domain of regions. As a set these relations must form a lattice with bottom and top element. 2. Approximations of regions with respect to a regional partition of the underlying space. Semantically, an approximation corresponds to the set of regions it approximates. 3. Pairs of meet operations on those approximations, which approximate the meet operation on exact regions. Based on those 'ingredients' syntactic and semantic generalizations of jointly exhaustive and pair-wise disjoint relations between exact one-dimensional regions were defined. Generalized relations hold between approximations of regions rather than between (exact) regions themselves. Syntactic generalization is based on replacing the meet operation defining relations between exact regions by its minimal and maximal counterparts on approximations. Se- mantically, syntactic generalizations yield upper and lower bounds (within the underlying lattice structure) on relations that can hold between the corresponding approximated exact regions. In the temporal domain I defined four sets of topological relations between one dimensional regions: RCC5 Boundary insensitive binary topological relations between regions in a non-directed one-dimensional space. RCC 9 Boundary insensitive binary topological relations between maximally connected regions (intervals) in a directed one-dimensional space. Boundary sensitive binary topological relations between regions in a non-directed one-dimensional space. Boundary sensitive binary topological relations between maximally connected regions (intervals) in a directed one-dimensional space. For each of these sets of relations between exact regions I discussed the syntactic and semantic generalization for the corresponding approximations and showed the equivalence of syntactic and semantic generalization. This provides the formal basis for qualitative temporal reasoning about approximate location in time. Acknowledgements This research was financed by the Canadian GEOID net- work. This support is gratefully acknowledged. --R Maintaining knowledge about temporal intervals. A boundary-sensitive approach to qualitative location Approximate qualitative spatial reasoning. On ontology and epistemology of rough location. Scientific Thought. An Introduction to the Philosophy of Science. The structure of spatial localization. Qualitative spatial representation and reasoning with the region connection calculus. A comparison of structures in spatial and temporal logics. Temporal reasoning based on semi- intervals Some problems about time. Conceptual neighborhoods in temporal and spatial reasoning. A spatial logic based on regions and connection. --TR --CTR Thomas Bittner , John G. Stell, Approximate qualitative spatial reasoning, Spatial Cognition and Computation, v.2 n.4, p.435-466, 2001
approximate reasoning;ontology;temporal relations;qualitative reasoning;granularity
590552
Refreshment policies for web content caches.
Web content caches are often placed between end users and origin servers as a mean to reduce server load, network usage, and ultimately, user-perceived latency. Cached objects typically have associated expiration times, after which they are considered stale and must be validated with a remote server (origin or another cache) before they can be sent to a client. A considerable fraction of cache "hits" involve stale copies that turned out to be current. These validations of current objects have small message size, but nonetheless, often induce latency comparable to full-fledged cache misses. Thus, the functionality of caches as a latency-reducing mechanism highly depends not only on content availability but also on its freshness. We propose policies for caches to proactively validate selected objects as they become stale, and thus allow for more client requests to be processed locally. Our policies operate within the existing protocols and exploit natural properties of request patterns such as frequency and recency. We evaluated and compared different policies using trace-based simulations.
INTRODUCTION Caches are often placed between end-users and origin servers as a mean to reduce user-perceived latency, server load, and network usage (see Figure 1). Among these dierent performance objectives of caches, improving end-user Web experience is gradually becoming the most pronounced. Many organizations are deploying caching servers in front of their LANs, mainly as a way to speed up users Web access. Gen- erally, available bandwidth between end-users and their Internet Service Providers (ISPs) is increasing and is complemented by short round trip times. Thus, the latency bottleneck is shifting from being between end-users and cache to being between cache and origin servers. From the view-point of Web sites and Content Distribution Networks (like decreasing costs of server-machines and back- An earlier version of the paper appeared in the Proceedings of Infocom '01 [1]. bone connectivity bandwidth along with increasing use of the Web for commercial purposes imply that server and net-work load are gradually becoming a lesser issue relative to end-user quality of service. At the limit, these trends indicate that communication time between local caches and remote servers increasingly dominates cache service-times and user-perceived latency, and that technologies which provide tradeos between tra-c-increase and latency-decrease would become increasingly worthwhile for both Web sites and ISPs. Servicing of a request by a cache involves remote communication if the requested object is not cached (in which case the request constitutes a content miss). Remote communication is also required if the cache contains a copy of the object, but the copy is stale, that is, its freshness lifetime had expired and it must be validated (by the origin server or a cache with a fresh copy) prior to being served. If the cached copy turns out to be modied, the request constitutes a content miss. Otherwise, the cached copy is valid and we refer to the request as a freshness miss. Validation requests that turn out as freshness misses typically have small-size responses but due to communication overhead with remote servers, often their contribution to user-perceived latency is comparable to that of full-edged content misses. Thus, cache service times can be improved by reducing both content and freshness misses. The content hit rate is measured per object or per byte and sometimes weighted by estimated object fetching cost. It is dictated by the available cache storage and the replacement policy used. Replacement policies for Web caches were extensively studied (e.g. [3, 4, 5, 6, 7, 8, 9, 10]). Policies that seem to perform well are Least Recently Used (lru, which evicts the least recently requested object when the cache is full), Least Frequently Used (lfu, which evicts the least-frequently requested object), and Greedy-Dual-Size (which accounts for varying object sizes and fetching costs). Squid [11], a popular caching server software, implements the lru policy. Under these replacement policies, however, and due to decreasing storage cost, cache hit rate is already at a level where it would not signicantly improve even if unbounded storage is made available. Content availability can be improved by prefetching [12, 13], but prefetching of content involves more involved predictions and induces signicant bandwidth overhead. The freshness hit rate of a cache is not directly addressed by replacement policies or captured by the content hit rate metric. clients origin servers cache Figure 1: Schematic conguration of Cache, clients, and origin servers The expiration time of each object is determined when it is brought into the cache, according to attached HTTP response headers provided by the origin server. Expired content must be validated before being served. Most current caching platforms validate their content passively i.e. only when a client request arrives and the cached copy of the object is stale. They perform validation via a conditional GET request (typically this is an If-Modified-Since (IMS) Get request). This means that validation requests are always performed \online," while the end-user is waiting. Here we promote proactive refreshment where the cache initiates unsolicited validation requests for selected content. Such \of- ine" validations extend freshness time of cached objects and more client requests can be served directly from the cache. Our motivation is that the most signicant cost issue associated with freshness misses is their direct eect on user-perceived latency rather than their eect on server and network load, and thus it is worth performing more than one \oine" validation in order to avoid one performed \on- line." We formalize a cost model for proactive refreshment, where overhead cost of additional validation requests to origin servers is balanced against the increase in freshness hits. We propose and evaluate refreshment policies, which extend freshness periods of selected cached objects. The decision of which objects to renew upon expiration varies between policies and is guided by natural properties of the request history of each object such as time-to-live (TTL) values, popularity, and recency of previous requests. Refreshment policies can also be viewed as prefetching fresh- ness. Their methodology and implementation, however, is closer to replacement policies than object prefetching al- gorithms. Our refreshment policies resemble common replacement policies (such as lru and lfu) in the way objects are prioritized. First, policies prefer renewing recently- or frequently-requested objects. Second, implementation is similar since object value is determined only by the request- history of the object rather than by considering request history of related objects. Another dierence of refreshment and document prefetching is that validations typically have considerably smaller response-sizes than complete doc- uments, but due to communication overhead, the latency gap is not nearly as pronounced. Hence, refreshment potentially provides considerably better tradeos of bandwidth vs. reduced latency compared to object prefetching. Our experimental study indicates that the best among the refreshment policies we have studied can eliminate about half of the freshness misses at a cost of 2 additional validation requests per eliminated freshness miss. Freshness themselves constitute a large fraction (30%-50%) of cache hits in a typical cache today. Therefore we conclude that one can considerably improve cache performance by incorporating a refreshment policy in it. Overview The paper proceeds as follows. We discuss related work in Section 2. We then provide a brief overview of HTTP freshness control in Section 3. In Section 4 we discuss and analyze our log data. In Section 5 we present the dierent refreshment policies. In Section 6 we describe our methodology for trace-based simulations and the simulation results for the performance of the dierent policies. We conclude in Section 7 with a summary and future research directions. 2. RELATED WORK Recent work addressed validation latency incurred on freshness misses, including transferring stale cached data from the cache to the client's browser while the data validity is being veried [14] or while the modied portion (the \delta") is being computed [15]. These schemes, however, may require browser support and are eective only when there is limited bandwidth between end-user and cache (such as with modem users). Some or all freshness misses can be eliminated via stronger cache consistency. Cache consistency architectures include: server-driven mechanisms where clients are notied by the server when objects are modied (e.g. [16, 17]); client-driven mechanisms, where the cache validates with the server objects with a stale cached copy; and hybrid approaches where validations are initiated either at the server or the client. Hybrid approaches include the server piggybacking validations on responses to requests for related objects [18, 19, 20], which are used by the cache to update the freshness status of its content. Another hybrid approach is leases where the server commits to notify a cache of modication, but only for a limited pre-agreed period [21, 22, 23, 24]. Server-driven mechanisms provide strong consistency and can eliminate all freshness misses. Hybrid approaches can provide a good balance of validation overhead and reduced validation traf- c. None of these mechanisms, however, is deployed or even standardized. Implementation requires protocol enhancements and software changes not only at the cache, but at all participating Web servers, and may also require Web servers to maintain per-client state. Hence, it is unlikely they become widely deployed in the near future. Mean- while, except for some proprietary coherence mechanisms deployed for hosted or mirrored content [2, 25] (which require control of both endpoints), the only coherence mechanism currently widely deployed and supported by HTTP/1.1 is client-driven based on TTL (time-to-live). Our refreshment approach utilizes this mechanism. We considered refreshment policies similar to the ones proposed here in [26] for caches of Domain Name System (DNS) servers. Each resource record (RR) in the domain name system has a TTL value initially set by its authoritative server. A cached RR becomes stale when its TTL expires. Client queries can be answered much faster when the information is available in the cache, and hence, refreshment policies which renew cached RRs oine increase the cache hit rate and decrease user-perceived latency. Although not yet supported by some widely deployed Web caching platforms (e.g., Squid [11]), proactive refreshment is oered by some cache vendors [27, 28]. Such products allow refreshment selections to be congured manually by the administrator or integrate policies based on object popularity. Our work formalizes the issues and policies and systematically evaluates them. 3. FRESHNESS CONTROL We provide a simplied overview of the freshness control mechanism specied by HTTP and supported by compliant caches. For further details see [29, 30, 11, 31, 32]. Caches compute for each object a time-to-live (TTL) value during which it is considered fresh and beyond which it becomes stale. When a request arrives for a stale object, the cache must validate it before serving it, by communication either with an entity with a fresh copy (such as another cache) or with the origin server. The cachability and TTL computation is performed using directives and values found in the object's HTTP response headers. When the cache receives a client request for an object then it acts as follows: If the object is cached and fresh, the request constitutes content and freshness hit (chit and fhit respec- tively) and the cached copy is immediately returned to the client. If the object is cached, but stale, the cache issues a conditional HTTP GET request to the origin server (or another appropriately-selected cache). The conditional GET uses the entity tag of the cached copy (with HTTP/1.1 and if an E-tag value was provided with the response) or it issues an If-Modified-Since request with the Last-Modified response header value (indicating last modication time of the object). If the source response is Not-Modified then the request constitutes a content hit (chit) but a freshness miss (fmiss). If the object was modied, the request is a content miss (cmiss-r). If the item is not found in the cache, then it is fetched from the origin (or another cache) and the request constitutes a content miss (cmiss-d). If the request arrives from the client with a no-cache header then the cache forwards it to the origin server. The cache must forward the request even if it has a fresh copy. The cache uses the response to replace or refresh its older copy of the object. We refer to such requests as no-cache requests. The TTL calculation for a cachable object as specied by HTTP/1.1 compares the age of the object with its freshness lifetime. If the age is smaller than the freshness lifetime the object is considered fresh and otherwise it is considered stale. The TTL is the freshness lifetime minus the age (or zero if negative). The age of an object is the dierence between the current time (according to the cache's own clock) and the timestamp specied by the object's Date response header (which is supposed to indicate when the response was generated at the origin). If an age header is present, the age is taken to be the maximum of the above and what is implied by the age header. Freshness lifetime calculation then proceeds as follows. First, if a max-age directive is present, the value is taken to be the freshness lifetime. Otherwise, if Expires header (indicating absolute expiration time) is present, the freshness lifetime is the dierence between the time specied by the Expires header and the time specied by the Date header (zero if this dierence is negative). Thus, the TTL is the dierence between the value of the Expires header and the current time (as specied by the cache's clock). Otherwise, no explicit freshness lifetime is provided by the origin server and a heuristic is used: The freshness lifetime is assigned to be a fraction (HTTP/1.1 mentions 10% as an example) of the time dierence between the timestamp at the Date header and the time specied by the Last-Modified header, subject to a maximum allowed value (usually 24 hours, since HTTP/1.1 requires that the cache must attach a warning if heuristic expiration is used and the object's age exceeds 24 hours). Before concluding this overview, we point on two qualitative issues with the actual use of freshness control and their relation to our refreshment approach. In [33] we analyze the distribution of dierent freshness control mechanisms for objects in the traces we used and it shows that the large majority of cachable objects do not have explicit directives. For these objects, the heuristic calculation is used to determine the freshness lifetime, and thus, tradeos between freshness and coherence can be controlled by tuning parameter values and URL lters [31]. Our refreshment policies are applied on top of this heuristic, take it as a given, and attempt to reduce freshness misses without further compromising coherence. Another important issue suggested by recent studies is that cache control directives and response header timestamp values are often not set carefully or ac- curately. These practices may skew freshness calculations away from the original intent [32, 34, 33]. This issue is also orthogonal to our approach since our policies, like caches, take these settings at face value. 4. We used two 6 days NLANR cache traces [35] collected from the UC and SD caches from January 20th till January 25th, 2000. The NLANR caches run Squid [11] which logs and labels each request with several attributes such as the request time, service time, cache action taken, the response code returned to the client, and the response size. The data analysis below considered all HTTP GET requests such that a 200 or 304 response codes (OK or Not-Modified) were returned to the client. We classied each of these requests as fhit, fmiss, cmiss-r, cmiss-d, or no-cache using the Squid logging labels as follows. Content hits: { fhit: (freshness hit) the cache had a fresh cached copy. Squid labels TCP HIT, TCP MEM HIT, { fmiss: (freshness miss) the cache had a stale cached copy and validated it. Squid label TCP REFRESH HIT. Content misses: { cmiss-r: the cache had a stale cached copy, issued an IMS request, and got a new copy with a Modied response. Squid label TCP REFRESH MISS. { cmiss-d: there was no cached copy of the object. Squid label TCP MISS. no-cache: the request arrived with a no-cache request header. Squid label TCP CLIENT REFRESH MISS. The table of Figure 2 shows the fraction of requests of each type. Requests classied as fmisses, cmisses, or no-cache involve communication with the origin server or another cache. Freshness misses, targeted by refreshment policies, constitute 13% (UC) and 19% (SD) of all such requests. Moreover, it is evident that freshness misses constitute 31% (UC) and 43% (SD) of all content hits (requests classied as fmisses or fhits). These NLANR caches directed most validation requests (fmisses and cmisses-r) to the origin server (100% in the UC cache and 99.3% in the SD cache). It is also apparent that the vast majority (90% for UC and 95% for SD) of validation requests return Not-Modified (are classied as fmisses). The NLANR traces also recorded the service time of each request. That is, the time from when the HTTP request is received to when the last byte of the response is written to the client's socket. Note that this is usually one Round Trip Time (RTT) less than from the client's viewpoint. Figure 3 plots the Cumulative Distribution Function (CDF) of service time of requests, broken down by the cache ac- tion. The gap between freshness misses to freshness hits indicates the potential benet, in terms of latency, of eliminating fmisses. The gap between freshness misses and content misses is in part due to the additional time required to transfer a larger-size response. Another explaining factor is that content misses exhibit less locality of reference in the sense that the elapsed time since the preceding request to the server is more likely to be longer. The decreased locality implies longer DNS resolutions of the server's hostname, since the required DNS records are less likely to be cached, and longer response time for the HTTP request, since the origin server is more likely to be \cold" with respect to the cache 1 . The similar service time distribution for no-cache and freshness misses suggests that most no-cache requests are made to popular cached content. Figure 4 plots the CDF of service times on freshness misses and freshness hits, further broken down by the response code that the cache returned to the client. HTTP response code 200 indicates that content was returned whereas response code 304 (Not-Modified) indicates that the client issued an IMS GET request and that its copy was validated by the cache. Responses with code 304 are typically smaller-size than responses with code 200. We can see that freshness hits with a 304 response to the client had minimal service time whereas freshness hits with 200 responses re ect RTTs Our study in [36] indicates that a rst HTTP request to a server in a time period is more likely to take longer than subsequent ones. between the cache and its clients and additional processing. The gap for freshness misses between 200 and 304 responses is also similar and re ects the additional communication between cache and clients due to the larger transmitted response size. 5. REFRESHMENT POLICIES Refreshment policies associate with every cached object a renewal credit (a nonnegative integer). When a cached copy is about to expire (according to its respective TTL interval), and it has nonzero renewal credit, a renewal request is sent to the respective authoritative server, and the renewal credit is decremented. The association of renewal credits to objects is governed by the particular policy. The policies we consider may increment the renewal credit only upon a client request. Renewal scheduling can be implemented via a priority queue, grouping together objects with close-together expiration, or possibly by an event-triggered process. We discuss two types of renewal requests: 1. Conditional fetch: The cache noties the server of last modication times or entity tag(s) of cached ver- sion(s) of the object, and requests either a validation of its current version or a new valid version. (This is supported by HTTP/1.0 by an If-Modified-Since GET request) 2. Pure validation request: Test whether the cached version of the object is valid, without requesting a valid copy if it is no longer valid. (Under HTTP/1.0 this is performed by issuing an If-Modified-Since HEAD request. HTTP/1.1 provides additional mechanisms, e.g., range requests.) Policies that use pure validation requests stop renewing a copy once it is invalidated, even if its renewal credit is positive. Pure validation renewals generally use less bandwidth than conditional fetches, since if the object was modied, only the object header is transmitted. On the other hand, conditional fetches result in a fresh cached copy even when the object was modied. Thus, policies that only perform pure validation renewals target only freshness misses whereas policies that allow conditional fetches also address some content misses (cmisses-r). For small-size objects that can t on a single packet, however, the overhead of pure validations is similar to that of conditional fetches so conditional fetches would be more eective. Ultimately, the type of renewal request can be determined by the policy on a per-request or per-object basis, according to (previous) content length and modication patterns. The data analysis in Section 4 shows that only a small fraction (5%-10%) of IMS requests result in invalidations and content transmission. This suggests that the additional overhead of performing conditional fetches over pure validations is typically amortized over 10- renewals. Therefore, for moderate size objects, the total bandwidth usage is similar under both choices of renewal actions. For the sake of simplicity, we evaluated only policies with pure validation renewals. Since the likelihood of modication is low, pure validation renewals capture most trace total req. 200+304 req. fhits fmisses cmisses-d cmisses-r no-cache UC 7.5M 6.3M 23% 10% 56% 1% 10% SD 5.6M 4.4M 19% 15% 56% 3% 7% Figure 2: Classication of the requests in the UC and SD traces.0.10.30.50.70.90 500 1000 1500 2000 2500 3000 3500 4000 fraction below x milliseconds UC: CDF of request service times fhits no-cache cmisses-r cmisses-d0.10.30.50.70.90 500 1000 1500 2000 2500 3000 3500 4000 fraction below x milliseconds SD: CDF of request service times fhits no-cache cmisses-r cmisses-d Figure 3: CDF of the service times, broken down by cache action of the potential of refreshments and provide a good indication for the full potential. We believe, however, that the incorporation of conditional fetches in the policies deserves further study. Another design issue that may require explaining is that we chose not to consider predictive policies which renew or prefetch long-expired objects. Predictive renewals and document prefetching are typically eective if activity is traced at a per-user basis, where future requests are predicted according to current requests made by the same user to related objects. Our refreshment approach diers from predictive-renewal in that we consider the aggregate behavior across users on each object. Our policies use minimal book-keeping, simple implementation, and do not require Web server support. Ultimately, it may be eective for refreshment policies to co-exist with predictive renewals and content prefetching, but we believe that the basic dierences between these techniques call for separate initial evaluations. One last important point that we would like to make explicit is that our policies were designed (and evaluated) for caches that forward requests to origin servers. For exam- ple, top-level caches in a hierarchy and caches that are not attached to a hierarchy. Directing renewal requests to an authoritative source assures a maximum freshness time for the response. Caches that are congured to forward requests to a parent cache may also benet from deploying a refreshment policy. The potential gain, however, is limited since renewals would often obtain aged responses (i.e. objects that have already been cached for a time period at the higher level cache). We studied age eects on performance of cascaded caches in [37, 38]. We proceed by describing the dierent policies. An illustrated example is provided in Figure 5. In the next section we evaluate and compare their performance using trace-based simulations. Policies passive: passive validation, objects are validated only as a result of a freshness miss i.e. when a request arrives and there is a stale cached copy. This is the way most caches work today. opt(i): An approximation to the optimal omniscient oine policy. 2 This policy assumes knowledge of the time of the subsequent request for the object and whether the object copy would still be valid then. If the subsequent request is such that the current copy remains valid and it is issued within c i freshness-lifetime- durations after the expiration of the current copy, then the renewal credit is set to c. Otherwise, no renewals are performed. recency(k): The renewal credit is reset to k following any request for the object, including no-cache requests. recency(k), similarly to the cache replacement policy lru, exploits the recency property of request se- quences, which states that future requests are more likely to be issued to recently-requested objects. recency (k): A variant of recency(k) that resets the renewal credit to be k following any request for the object, except for no-cache requests. increment the renewal credit by j for any request that would have been a freshness miss under passive. In other words, we add j to the renewal credit We specify optimal oine algorithm in [26]. fraction below x milliseconds UC: CDF of request service times 304 fhits 200 fhits 200 fmisses0.10.30.50.70.90 500 1000 1500 2000 2500 3000 3500 4000 fraction below x milliseconds SD: CDF of request service times 304 fhits 200 fhits 200 fmisses Figure 4: CDF of the service times for content-hits, broken down by response code to the client recency(2) time request no-cache request refresh miss hit m2.501.671.80overhead passive Figure 5: Behavior of dierent refreshment policies on an example sequence of 9 requests. The time line is in units of freshness-lifetime durations. All policies incur at least two misses: The rst request is a cold-start miss and the 8th request, which has a no-cache request header, is always a miss. The gure also summarizes the number of misses and renewals performed by each policy. passive, for example, incurs 7 misses and performs no renewals, while th-freq(0:5; incurs only 2 misses and performs 13 renewals. The policy opt(1) is the most e-cient in the following sense: it performs the least renewals among all policies that incur at most 4 misses (e.g., recency(1) and freq(1; 0)). This example illustrates how the coverage (fraction of misses eliminated) of the policies recency(i) and freq(i) increases with i. The overhead (number of \extra" requests per eliminated miss), however, typically (but not always) increases as well. For example, recency(1) eliminates 3 misses with respect to passive but performs 8 renewals. Thus, it issues 8+4 (renewals and misses) than passive and has overhead of 5=3 1:67. Similarly, recency(2) eliminates 4 misses while issuing 3 than passive, and thus, has overhead of upon any request which is issued more than freshness- lifetime-duration units of time after a previous request that caused the passive policy to contact the origin server. In addition, upon any request (except for no-cache requests) the renewal credit is set to m if it is less than m. With policy freq(j; purely exploits the frequency property that states that objects that were frequently requested in the past are more likely to be requested soon. replacement policy that exploits the frequency property is lfu.) For the policy freq(j; m) is a hybrid of freq(j; and recency(m). th-freq(th; m) keep renewing objects until the ratio of would-have been passive freshness misses to number of freshness-lifetime-durations since beginning of log drops below a threshold. In other words, upon each request which would have been a freshness miss for passive we increment the renewal credit such that we would keep renewing until the ratio drops below a threshold. In addition, upon any request (except for no-cache requests) the renewal credit is increased to m if it was less than m. This policy exploits the frequency property, and normalizes it by freshness- lifetime-duration. It also naturally provides a more continuous range of tradeos, since th is not necessarily an integer. With the policy is purely frequency-based whereas higher values of m correspond to hybrids with recency(m) policy. 6. EXPERIMENTAL EVALUATION We conducted trace-based simulations in order to evaluate cache performance under the dierent refreshment policies. We outline our methodology and then proceed to present and discuss the simulation results. 6.1 Methodology The traces included the cache action taken on each request for the currently-implemented passive refreshment. In order to simulate other policies, however, we had to obtain response header values or TTL values for requested objects. Unfortunately, this data is not available in the recorded trace. We therefore separately issued GET requests for the URLs in the trace shortly after downloading it. We processed the response headers and extracted cache directives and values of relevant header elds. For cachable objects (objects without a no-cache directive in the response header), we applied the Squid object freshness model [11] (HTTP/1.1 compliant) described in Section 3 to calculate TTLs using the values extracted above. When an explicit freshness- lifetime duration was not provided by an Expires or max-age header, we applied Squid's heuristic. We used 10% of the time dierence between the time specied by the Date header and the time specied by the Last-Modified header subject to a maximum value of one day. We issued more than a single GET for a sample of the objects and repeated the TTL calculation. We found that cache control directives and freshness-lifetime values do not change frequently. This indicated that our estimates re ected reasonably well values that would have been obtained from the origin server at the time requests were logged. We then ran simulations using the original sequence of requests and the extracted TTL values. To put all policies on equal ground and eliminate the eect of boundary conditions we also simulated passive, so the performance gures provided later correspond to the simulated passive rather than the one re ected by the original labels (given in Section 4). We rst discarded all requests that were not labeled as HTTP GET. The simulation was only applied to requests for URLs on which we obtained a 200 (OK) response on our separately-issued requests. Note that 302 responses (HTTP redirect) are not cachable and hence requests of URLs for which we obtained such a response were discarded. The simulation then utilized the logged information in the following way: (i) All logged requests for each considered URL were used along with the request time to determine the status of the object in the simulated cache and the resulting cache actions, (ii) the original cache-action label was used to identify requests which arrived with a no-cache header (our simulation accounted for such requests by resetting the TTL to the freshness-lifetime duration even if a fresh copy was present in the simulated cache), and (iii) the original labels were also used in a heuristic that estimated at which points objects were modied. The modication heuristic considered various recorded labels of the requests. Clearly, when a successful request was classied by the labels in the trace as a cmiss-r (i.e. labeled TCP REFRESH MISS in the trace, see Section contents had changed. For requests labeled cmisses-d we did not have an explicit indication whether content had changed so we used a heuristic based on the logged size of the response to the client. In order to simulate performance of refreshment policies, we also had to estimate at which point in the interval between requests the modication had actually occurred (since refreshment policies stop refreshing once the server invalidates the object). The simulation assumed that when a modication had occurred between two consecutive requests (and therefore incur a content miss on the later it happened at the midpoint of the time interval between the two requests. If more modications had happened, it is likely that the rst one occurred earlier in the interval, and hence our assumption means that more unproductive renewals occur in the simulation than in reality, and thus, the simulated policies would exhibit somewhat worse tradeos. Since the majority of validations return not-modied, however, this assumption could not have had a signicant eect on our results. Since we could not issue GET requests for all the URLs present in the trace in a reasonable time without adversely aect- ing our environment, we selected a subset and then scaled the results up to factor out the sampling. We applied non-uniform sampling with denser samples from more frequently- requested URLs. In particular, we included all URLs that were requested more than 12 times. In total we fetched about 224K distinct URLs. The reason for non-uniform sampling is the Zipf-low relation between requests and URLs where many requests are issued to a small set of popular URLs. The original logs had about 5 million dierent URLs, most of them requested only once. Thus, a same-size sample obtained through uniform sampling over URLs would have yielded lower-condence estimates than the non-uniform sample we used. The sampling bias was factored out by scaling each frequency group individually. The simulations assumed innite cache storage capacity. This is consistent both with current industry trends and with the actual traces we used, since objects requested twice or more in the 6 day period were not likely to be discarded by the replacement policy used in Squid. In the performance metric used in the simulations we counted all requests that constituted content hits. Content hits that occurred more than a freshness-lifetime duration past the previous (simulated) server contact were counted as freshness misses and requests occurring within the duration were counted as freshness hits. Content hits exclude requests to explicit noncachable objects, requests with no-cache request headers, and requests when the content had changed. Since we had no information on such requests, we did not classify the rst request of each object. Appropriate requests for objects with freshness-lifetime-duration of 0 were counted as content hits, but were all considered freshness misses. Note that renewals are not performed on such objects and hence, the number of freshness misses incurred on these objects is not reduced by refreshment policies. 6.2 Simulation results Under the simulated baseline policy passive (where objects are refreshed only as a result of client requests), 48% of content hits constituted freshness misses on the UC trace, and 53% were freshness misses on the SD trace. We recall that the respective numbers according to labels on the recorded trace provided in Section 4 are 31% and 43%. The gap is mainly due to simulating a shorter trace, dierent boundary conditions, and the conservative heuristic used to identify content hits. These factors should aect all policies in a similar manner. So in order to put passive on equal grounds with all other policies we chose to simulate it rather than using the labels of the trace. We evaluated the performance of the dierent refreshment policies by the tradeo of overhead vs. coverage. The coverage (reduction in freshness-misses) is calculated as the fraction of the number of freshness misses according to passive that are eliminated by the respective policy. More precisely, if x denotes the number of freshness misses of passive, and y the number of freshness misses of a policy P then the coverage of P is the fraction (x y)=x. We calculated the overhead of policy P as the dierence between the number of validation requests issued by P and the number of validation requests issued by passive. Recall that the cache issues a validation request for each freshness miss. Hence, the request overhead is the total number of renewals performed minus the number of freshness misses eliminated (converted to freshness hits). We normalize the overhead by dividing it with the total number of freshness misses that were converted to freshness hits. To obtain the coverage/overhead tradeo for each type of policy, we swept the value of its respective parameter. For example, the points on the curve of recency correspond to runs with recency(1); recency(2); the points for freq were obtained on runs with freq(1; . Note that opt(0), recency(0), recency (0), are in fact passive. 3 These tradeos are shown in Figure 6. The performance of recency and recency policies was almost identical, hence we omitted the recency curve. This similarity shows that requests without a no-cache header are about as likely to follow a request with a no-cache header as one without a no-cache header. Under all types of policies, the coverage peaks at about 63%- 67%. The remaining 33%-37% of freshness misses mostly occur on objects with freshness-lifetime-duration of 0, on which refreshment is not eective. The opt policies eliminates all \addressable" freshness misses with overhead of about 1.3 requests per miss, and eliminates the bulk of these misses with an overhead of 0.5. These numbers bound the potential of refreshment policies. They also indicate on the performance loss by the restriction to the refreshment frame- work. 4 The simulation results for opt and recency show the locality eect, where most freshness misses that can be eliminated occur within a small number of freshness-lifetime- durations after a previous request. The left-most point on the curves of recency and opt, which correspond to recency(1) and opt(1), show that about 30% of freshness misses occur within one freshness-lifetime-duration after the expiration of the cached copy. The second points on the left correspond to recency(2) and opt(2) and indicate that about additional 15% of freshness misses occur between one and two freshness-lifetime durations passed the expiration. We note that the observed fact that a very small number of freshness occur more than 10 freshness-lifetime-durations passed the expiration is not only due to locality but also re ects the interaction of the log duration of 6 days and the most common freshness-lifetime-duration of 24 hours [33]. The fact that coverage of recency and the frequency-based policies peaks at about the same place indicates that a very small fraction of freshness misses are incurred on very infrequently requested objects (since the frequency-based policies do not perform any renewals on the rst request and thus can not eliminate misses incurred on the second re- quest. The correspondence in peak coverage of opt and other policy-types is due to a \threshold phenomenon" where most freshness misses occur on objects with a freshness- lifetime-duration of 0 or occur within a small number of freshness-lifetime-durations following a previous request. The frequency-based policies freq(j; 0) and th-freq(th; signicantly outperformed recency(k). This re ects the fact that the vast majority of freshness misses which can be eliminated occur on the more popular URLs. The gap is caused by the very large number of cachable URLs that were 3 We remark that these policies are able to achieve \contin- uous" tradeos by mixing two consecutive integral values of the respective parameter, for example, applying recency(1) on some URLs and recency(2) on others. 4 Recall that under the refreshment framework, objects must be kept fresh continuously till the following \hit." The optimal unrestricted policy, which validates objects just before they are requested, incurs no overhead. request-overhead per eliminated fmiss fraction of fmisses eliminated UC trace: performance of refreshment policies OPT RECENCY request-overhead per eliminated fmiss fraction of fmisses eliminated SD trace: performance of refreshment policies OPT RECENCY Figure Performance of the dierent refreshment policies when simulated on the UC and SD traces requested only once. The recency(k) policy performed up to k unproductive renewals 5 on each such request. The hybrid policies freq and th-freq with m > 0 performed considerably worse than the pure frequency-based policies (that correspond to hence only results for are shown. This behavior is not surprising given that recency yielded much worse tradeos. Overall, the results indicate that frequency-based object prioritization is more eective than recency-based prioritization. The domination of frequency-based policies is also consistent with studies of cache replacement policies for Web contents [8, 7], since dierent URLs tend to have a wide range of characteristic popularities, a property that is captured better by frequency-based policies. It is interesting to contrast these results with a related study of refreshment policies that we performed for DNS records. In contrast with our nd- ings here, for DNS caches the recency and frequency-based policies exhibited similar performance [26]. Our explanation is that at the hostname level, there is signicantly smaller fraction of \objects" that are resolved only once. The performance of freq and th-freq is similar, although th-freq, which normalizes the frequency by freshness-lifetime- duration, performed somewhat better. The similarity is mostly explained by the fact that the large majority of freshness- lifetime-durations are of the same length (one day) and also because for shorter durations, frequency of requests is correlated with the duration. The policy th-freq provides a spectrum of tradeos and better performance, particularly in the low-overhead range. th-freq, however, may require more book-keeping than freq. The particular tradeos obtained by the frequency-based policies shows that signi- cant improvements can be obtained with fairly low overhead. About 10% of freshness misses can be eliminated with the overhead of half of a validation request per eliminated freshness miss; 25% of freshness misses can be eliminated with overhead of a single request per eliminated miss; 50% can be eliminated with overhead of two; and 65% of freshness misses can be eliminated with overhead of three. 5 It could be less than k since we did not perform renewals passed the termination time of the log. 7. CONCLUSION A large fraction (30%-50%) of cache hits constitute freshness misses, that is, the cached copy was not fresh, but turned out to be valid after communication with the origin server. Validations are performed prior to sending responses to users, and signicantly extend cache service time. There- fore, freshness misses impede the cache ability to speed-up Web access. An emerging challenge for Web content caches is to reduce the number of freshness misses by proactively maintaining fresher content. It seems that some cache vendors had already implemented ad-hoc refreshment mechanisms. Our proposed refreshment policies are a relatively low-overhead systematic solution. Refreshment policies extend freshness lifetime by selectively validating cached objects upon their expiration. Since cache freshness is increased, requested objects are more likely to be fresh and thereby are serviced faster. We demonstrated that a good refreshment policy can eliminate about 25% of freshness misses with an overhead of a single validation requests per eliminated miss, that is, two \oine" validation requests replace one \online" request. For future work, we propose ways to further reduce renewal overhead. We rst propose that renewals of objects located at the same host are batched together, and thus, decrease overhead by sharing the same persistent connection. Batching can be natural as co-located objects often share the same cache-control mechanism and subsets of such objects (that are embedded on the same page or related pages) are often requested together. A second proposal is to perform renewals at o-peak hours. The most common freshness- lifetime duration of 24 hours provides su-cient scheduling exibility to do so (e.g., performing renewals due at 10am EST at the signicantly less-busy time of 7am EST). As a next step in the evaluation of the eectiveness of a refreshment policy we hope to incorporate one such policy in a popular caching server software such as Squid. Our results indicate that the integration of refreshment would not impose a signicant computational overhead, and would boost performance in terms of user-perceived latency. Acknowledgment Our experiments would not have been possible without the collection and timely availability of the NLANR cache traces. We thank Duane Wessels for answering questions with regard to a Squid logging bug. 8. --R --TR Scale and performance in a distributed file system Leases: an efficient fault-tolerant mechanism for distributed file cache consistency Web cache coherence Removal policies in network caches for World-Wide Web documents Improving end-to-end performance of the Web using server volumes and proxy filters Exploiting regularities in Web traffic patterns for cache replacement On-line file caching Aging through cascaded caches Volume Leases for Consistency in Large-Scale Systems Image-based Rendering with Controllable Illumination Evaluating Server-Assisted Cache Replacement in the Web Proactive Caching of DNS Records Application-level document caching in the Internet Maintaining Strong Cache Consistency in the World-Wide Web --CTR Edith Cohen , Haim Kaplan, Proactive caching of DNS records: addressing a performance bottleneck, Computer Networks: The International Journal of Computer and Telecommunications Networking, v.41 n.6, p.707-726, 22 April Timo Koskela , Jukka Heikkonen , Kimmo Kaski, Web cache optimization with nonlinear model using object features, Computer Networks: The International Journal of Computer and Telecommunications Networking, v.43 n.6, p.805-817, 20 December
validation requests;HTTP;web caching;user-perceived latency;refreshment policies
590776
Overcoming the Myopia of Inductive Learning Algorithms with RELIEFF.
Current inductive machine learning algorithms typically use greedy search with limited lookahead. This prevents them to detect significant conditional dependencies between the attributes that describe training objects. Instead of myopic impurity functions and lookahead, we propose to use RELIEFF, an extension of RELIEF developed by Kira and Rendell &lsqb;10, 11&rsqb;, for heuristic guidance of inductive learning algorithms. We have reimplemented Assistant, a system for top down induction of decision trees, using RELIEFF as an estimator of attributes at each selection step. The algorithm is tested on several artificial and several real world problems and the results are compared with some other well known machine learning algorithms. Excellent results on artificial data sets and two real world problems show the advantage of the presented approach to inductive learning.
Introduction Inductive learning algorithms typically use a greedy search strategy to overcome the combinatorial explosion during the search for good hy- potheses. The heuristic function that estimates the potential successors of the current state in the search space has a major role in the greedy search. Current inductive learning algorithms use variants of impurity functions like information gain, gain ratio[25], gini-index[1], distance ever, all these measures assume that attributes are conditionally independent given the class and therefore in domains with strong conditional dependencies between attributes the greedy search has poor chances of revealing a good hypothesis. Kira and Rendell [10], [11] developed an algorithm called which seems to be very powerful in estimating the quality of attributes. For example, in the parity problems of various degrees with a significant number of irrelevant (ran- dom) additional attributes RELIEF is able to correctly estimate the relevance of all attributes in a time proportional to the number of attributes and the square of the number of training instances (this can be further reduced by limiting the number of iterations in RELIEF). While the original RELIEF can deal with discrete and continuous at- tributes, it can not deal with incomplete data and is limited to two-class problems only. We developed an extension of RELIEF called RELIEFF that improves the original algorithm by estimating probabilities more reliably and extends it to handle incomplete and multi-class data sets while the complexity remains the same. RELIEFF seems to be a promising heuristic function that may overcome the myopia of current inductive learning algorithms. Kira and Rendell used RELIEF as a preprocessor to eliminate irrelevant attributes from data description before learn- ing. RELIEFF is general, relatively efficient, and reliable enough to guide the search in the learning process. In this paper a reimplementation of Assistant learning algorithm for top down induction of decision trees [4] is described, named Assistant- R. Instead of information gain, Assistant-R uses RELIEFF as a heuristic function for estimating the attributes' quality at each step during the tree generation. Experiments on a series of artificial and real-world data sets are described and the results obtained using RELIEFF as a selection criterion are compared to results of some other ap- proaches. The following approaches are compared: ffl the use of information gain as a selection criterion ffl LFC [27], [28] that tries to overcome the myopia of information gain with a limited lookahead; ffl the naive Bayesian classifier, that assumes conditional independence of attributes; ffl the k-nearest neighbors algorithm. The paper is organized as follows. In the next section, the original RELIEF is briefly described along with its interpretation and its extended version RELIEFF. In Section 3, we present the reimplementation of Assistant called Assistant-R. In Section 4.1 we briefly describe the other algorithms used in our experiments. In Section 4.2 we describe the experimental methodology. Section 5 describes experiments, and compare the results of the different algorithms. We show that Assistant-R performs at least as well as Assistant- I and sometimes much better. In conclusion, the potential breakthroughs are discussed on the basis of the excellent results on artificial data sets. Finally, integration of the compared algorithms is proposed. 2. RELIEFF 2.1. RELIEF The key idea of RELIEF is to estimate attributes according to how well their values distinguish among the instances that are near to each other. For that purpose, given an instance, RELIEF 1. set all weights W[A] := 0.0; 2. for i := 1 to n do 3. begin 4. randomly select an instance R; 5. find nearest hit H and nearest miss M; 6. for A := 1 to #all attributes do 7. W[A] := W[A] - diff(A,R,H)/n 8. 9. end; Figure 1 The basic algorithm of RELIEF searches for its two nearest neighbors: one from the same class (called nearest hit) and the other from a different class (called nearest miss). The original algorithm of RELIEF [10], [11] randomly selects n training instances, where n is the user-defined parameter. The algorithm is given in Figure 1. Function diff(Attribute,Instance1,Instance2) calculates the difference between the values of Attribute for two instances. For discrete attributes the difference is either 1 (the values are different) or 0 (the values are equal), while for continuous attributes the difference is the actual difference normalized to the interval [0; 1]. Normalization with n guarantees all weights W [A] to be in the interval [\Gamma1; 1], however, normalization with n is an unnecessary step if W [A] is to be used for relative comparison among attributes. The weights are estimates of the quality of attributes. The rationale of the formula for updating the weights is that a good attribute should have the same value for instances from the same class (subtracting the difference should differentiate between instances from different classes (adding the difference The function diff is used also for calculating the distance between instances to find the nearest neighbors. The total distance is simply the sum of differences over all attributes. In fact original RELIEF uses the squared difference, which for discrete attributes is equivalent to diff. In all our experiments, there was no significant difference between results using diff or squared differ- ence. If N is the number of all training instances then the complexity of the above algorithm is O(n \Theta N \Theta #all attributes). 2.2. Interpretation of RELIEF's estimates The following derivation shows that RELIEF's estimates are strongly related to impurity functions. It is obvious that RELIEF's estimate W [A] of attribute A is an approximation of the following difference of probabilities: (different value of Aj nearest instance from different class) \GammaP (different value of Aj nearest instance from same class) (1) If we eliminate from (1) the requirement that the selected instance is the nearest, the formula becomes: (different value of Ajdifferent class) \GammaP (different value of Ajsame class) (2) If we rewrite (equal value of classjequal value of we obtain using Bayes rule: For sampling with replacement in strict sense the following equalities hold: Using the above equalities we obtain: const \Theta where is highly correlated with the gini-index gain [1] for classes C and values V of attribute A. The difference is that instead of factor the gini-index gain uses Equation (3) shows strong relation of RE- LIEF's weights with the gini-index gain. The probability that two instances have the same value of attribute A in eq. (3) is a kind of normalization factor for multi-valued at- tributes. Impurity functions tend to overestimate multi-valued attributes and various normalization heuristics are needed to avoid this tendency (e.g. gain ratio [25], distance measure [16], and binarization of attributes [4]). Equation (3) shows that RELIEF exhibits an implicit normalization effect. Another deficiency of gini-index gain is that its values tend to decrease with the increasing number of classes [14]. Denominator which is constant factor in equation (3) for a given attribute again serves as a kind of normalization and therefore RELIEF's estimates do not exhibit such strange behavior as gini-index gain does. The above derivation eliminated the "nearest instance" condition from the probabilities. If we put it back we can interpret RELIEF's estimates as the average over local estimates in smaller parts of the instance space. This enables RELIEF to take into account the context of other attributes, i.e. the conditional dependencies between attributes given the class value which can be detected in the context of locality. From the global point of view, these dependencies are hidden due to the effect of averaging over all training instances, and exactly this makes impurity functions myopic. Impurity functions use correlation between the attribute and the class disregarding the context of other attributes. This is the same as using the global point of view and disregarding the local peculiarities. The example data set given in Table 1 illustrates the difference between myopic estimation functions and RELIEF. We have three attributes and eight training instances. The class value is determined with XOR function on attributes A1 and A2, while the third attribute A3 is randomly generated. RELIEF (equation (1)) correctly estimates that attributes A1 and A2 are the most important while the contribution of attribute A3 is poor. On the other hand, W'[A] (equation (3)), Table Example data set and the estimated quality of attributes function A1 A2 A3 Class information gain [9] 0.000 0.000 0.049 gain-ratio [25] 0.000 0.000 0.051 distance [16] 0.000 0.000 0.026 Ginigain' (equation (4)), original gini-index gain [1], information gain [9], gain ratio [25], and distance measure [16] estimate that the contribution of A3 is the highest while attributes A1 and A2 are estimated as completely irrelevant. Hong [8] developed a procedure similar to RELIEF for estimating the quality of attributes, where he directly emphasizes the use of contextual information. The difference to RELIEF is that his approach uses only information from nearest misses and ignores nearest hits. Besides, Hong uses the normalization to penalize the contribution of nearest misses that are far away from a given instance. 2.3. Extensions of RELIEF The original RELIEF can deal with discrete and continuous attributes. However, it can not deal with incomplete data and is limited to two-class problems only. Equation (1) is of crucial importance for any extensions of RELIEF. It turned out that the extensions of RELIEF are not straightforward unless we realized that RELIEF in fact approximates probabilities. The extensions should be designed in such a way that those probabilities are reliably approximated. We developed an extension of RELIEF, called RELIEFF, that improves the original algorithm by estimating probabilities more reliably and extends it to deal with incomplete and multi-class data sets. A brief description of the extensions follows. Reliable probability approxima- tion: The parameter n in the algorithm RE- LIEF, described in Section 2.1, represents the number of instances for approximating probabilities in eq. (1). The larger n implies more reliable approximation. The obvious choice, adopted in RELIEFF for relatively small number of training instances (up to one thousand), is to run the outer loop of RELIEF over all available training instances. The selection of the nearest neighbors is of crucial importance in RELIEF. The purpose is to find the nearest neighbors with respect to important attributes. Redundant and noisy attributes may strongly affect the selection of the nearest neighbors and therefore the estimation of probabilities with noisy data becomes unreliable. To increase the reliability of the probability approximation RELIEFF searches for k nearest hits/misses instead of only one near hit/miss and averages the contribution of all k nearest hits/misses. It was shown that this extension significantly improves the reliability of estimates of attributes' qualities[13]. To overcome the problem of parameter tuning, in all our experiments k was set to 10 which, empirically, gives satisfactory results. In some problems significantly better results can be obtained with tuning (as is typical for the majority of machine learning algorithms). Incomplete data: To enable RELIEF to deal with incomplete data sets, the function diff(Attribute,Instance1, Instance2) in RELIEFF is extended to missing values of attributes by calculating the probability that two given instances have different values for the given attribute: ffl if one instance (e.g. I1) has unknown value: ffl if both instances have unknown value: The conditional probabilities are approximated with relative frequencies from the training set. nearest neighbors correlation coefficient independent atts parity problems Figure 2 The correlation of the RELIEFF's estimates with the intended quality of attributes on data sets with conditionally independent and strongly dependent attributes. This approach assumes that conditional probabilities of attribute-values given the class are applicable without the context of any other at- tribute. This may in some cases be too naive, however including the context of other atributes is far too inefficient. Multi-class problems: Kira and Rendell that RELIEF can be used to estimate the attributes' qualities in data sets with more than two classes by splitting the problem into a series of 2-class problems. This solution seems unsatisfactory (in Section 4.1 we discuss the performance of this approach and compare it with the extension described below). To use it in prac- tice, RELIEF should be able to deal with multi-class problems without any prior changes in the knowledge representation that could affect the final outcomes. Instead of finding one near miss M from a different class, RELIEFF searches for k near misses for each different class C and averages their contribution for updating the estimate W [A]. The average is weighted with the prior probability of each class: The idea is that the algorithm should estimate the ability of attributes to separate each pair of classes regardless of which two classes are closest to each other. The normalization if prior probabilities of classes is necessary as k near misses from each different class would tend to exaggerate the influence of classes with small number of cases. Note that the time complexity of RELIEFF is O(N 2 \Theta #attributes), where N is the number of training instances. 2.4. RELIEFF's estimates and attribute's qualit To estimate the contribution of parameter k (# nearest hits/misses) on RELIEFF's estimates of attribute's quality Kononenko [13] compared the intended information gain of attributes with the estimates, generated by RELIEFF, by calculating the standard linear correlation coefficient. The correlation coefficient can show how is the intended quality and the estimated quality of attributes related. A typical graph for data sets with conditionally independent attributes and with strongly dependent attributes (parity problems of various de- grees) is shown in Figure 2. For conditionally independent attributes, the quality of the estimate monotonically increases with the number of nearest neighbors. For conditionaly dependent at- tributes, the quality increases up to a maximum but later decreases as the number of nearest neighbors exceeds the number of instances that belong to the same peak in the distribution space for a given class. Note that, if attributes were evaluated with the myopic impurity functions, like the gini-index and the information gain, the quality of the estimates would be high for conditionally independent attributes and poor for strongly dependent attributes. This corresponds to the estimates by RELIEFF with very large number of nearest hits/misses. To test the effect of the normalization factor in eq. (3) we run RELIEFF also on one well known medical data set, "primary tumor", described in 6 THE AUTHORS??? Section 5.3. The major difference between the estimates by impurity functions and the estimates by RELIEFF in the "primary tumor" problem is in the estimates of two most significant attributes. Information gain and gini-index overestimate one attribute with 3 values (by the opinion of physicians specialists). RELIEFF and normalized versions of impurity functions correctly estimate this attribute as less important. 3. Assistant-R Assistant-R is a reimplementation of the Assistant learning system for top down induction of decision trees[4]. The basic algorithm goes back to CLS (Concept Learning System) developed by Hunt et al. [9] and reimplemented by several authors (see [25] for an overview). In the following we describe the main features of Assistant. Binarization of attributes: The algorithm generates binary decision trees. At each decision step the binarized version of each attribute is selected that maximizes the information gain of the attribute. For continuous attributes a decision point is selected that maximizes the at- tribute's information gain. For discrete attributes a heuristic greedy algorithm is used to find the locally best split of attribute's values into two subsets. The purpose of the binarization is to reduce the replication problem and to strengthen the statistical support for generated rules. Decision tree pruning: Prepruning and postpruning techniques are used for pruning off unreliable parts of decision trees. For preprun- ing, three user-defined thresholds are provided: minimal number of training instances, minimal attributes information gain and maximal probability of majority class in the current node. For postpruning, the method developed by Niblett and Bratko [22] is used that uses Laplace's law of succession for estimating the expected classification error of the current node commited by pruning/not pruning its subtree. Incomplete data handling: During learning, training instances with a missing value of the selected attribute are weighted with probabilities of each attribute's value conditioned with a class label. During classification, instances with missing values are weighted with unconditional probabilities of attribute's values. Naive Bayesian classifier: For each internal node in a decision tree eventually a third successor appears labeled with attribute's values for which no training instances are available. For such "null leaves", the naive Bayesian formula is used to calculate the probability distribution in the leaf by using only attributes that appear in the path from the root to the leaf: Y A Note that this calculation is done off-line, i.e. during the learning phase. For classification, the "null" leaves are already labeled with the calculated class probability distribution and are used for classification in the same manner as ordinary leaves. The main difference between Assistant and its reimplementation Assistant-R is that RELI- EFF is used for attribute selection. In addi- tion, wherever appropriate, instead of the relative frequency, Assistant-R uses the m-estimate of probabilities, which was shown to often significantly increase the performance of machine learning algorithms[2], [3]. For prior probabilities Laplace's law of succession is used: of possible outcomes where N is the number of all trials and N (X) the number of trials with the outcome X. These prior probabilities are then used in the m-estimate of conditional probabilities: The parameter m trades off between the contributions of the relative frequency and the prior probability In our experiments, the parameter m was set to (this setting is usually used as default and, em- pirically, gives satisfactory results [2], [3] although with tuning in some problem domains better results may be expected). The m-estimate is used in the naive Bayesian formula (5), for postpruning instead of Laplace's law of succession as proposed by Cestnik and Bratko[3], and for RELIEFF's estimates of probabilities. In eq. (1) we can use probabilities from the root of the tree as an estimate of prior probabilities for a lower internal node t with n(t) corresponding training instances: of Ajnearest miss; Ajnearest miss; root) of Ajnearest hit; Ajnearest hit; root) 4. Experimental environment 4.1. Algorithms for comparison We performed a series of experiments with Assistant-R and compared its performance to the following algorithms: Assistant-I: A variant of Assistant-R that instead of RELIEFF uses information gain for the selection criterion, as does Assistant. How- ever, the other differences to Assistant remain (m-estimate of probabilities). This algorithm enables us to evaluate the contribution of RE- LIEFF. The parameters for Assistant-I and Assistant-R were fixed throughout the experiments (no prepruning, postpruning with 2). LFC: Ragavan et al. [27], [28] use limited lookahead in their LFC (Lookahead Feature Con- struction) algorithm for top down induction of decision trees to detect significant conditional dependencies between attributes for constructive induction. They show interesting results on some data sets. We reimplemented their algorithm [29] and tested its performance. Our results, presented in this paper, show some drawbacks of the experimental comparison described by Ragavan and Rendell and confirm the advantage of the limited lookahead for constructive induction. LFC generates binary decision trees. At each node, the algorithm constructs new binary attributes from the original attributes, using logical operators (conjunction, disjunction, and negation). From the constructed binary at- tributes, the best attribute is selected and the process is recursively repeated on two sub-sets of training instances, corresponding to the two values of the selected attribute. For constructive induction a limited lookahead is used. The space of possible useful constructs is re- stricted, due to the geometrical representation of the conditional entropy which is the estimator of the attributes' quality. To further reduce the search space, the algorithm also limits the breadth and the depth of search. AS LFC uses lookahead it is less myopic than the greedy algorithm of Assistant. The comparison of results may show the performance of the greedy search in combination with RELI- EFF versus the lookahead strategy. To make results comparable to Assistant-R we equipped LFC with pruning and probability estimation facilities as described in Section 3. All tests were performed with a default set of parameters (depth of the lookahead 3, beam size 20), although in some domains better results may be obtained by parameter tuning. However, higher values of the parameters may combinatorially increase the search space of LFC, which makes the algorithm impractical. Naive Bayesian Classifier: A classifier that uses the naive Bayesian formula (5) to calculate the probability of each class given the values of all attributes and assuming the conditional independence of the attributes. A new instance is classified into the class with maximal calculated probability. The m-estimate of probabilities was used and the parameter m was set to 2 in all experiments. The performance of the naive Bayesian classifier can serve as an estimate of the conditional independence of attributes k-NN: The k-nearest neighbor algorithm. For a given new instance the algorithm searches for nearest training instances and classifies the instance into the most frequent class of these k instances. For the k-NN algorithm the same distance measure was used as for RELIEFF (see Section 2.1). The presented results were obtained with Manhattan-distance. The results using Euclidian distance are practically the same. The best results with respect to parameter k are pre- sented, although for fair comparison such parameter tuning should be allowed only on the training and not the testing sets. We selected the naive Bayesian classifier and the k-NN algorithm for comparison because they are both well known, simple, and they both perform well in many real-world problems. The performance of these two algorithms may show the nature of the classification problems. 4.2. Experimental methodology Each experiment on each data set was performed times by randomly selecting 70% of instances for learning and 30% for testing and the results were averaged. Each system used the same subsets of instances for learning and for testing in order to provide the same experimental conditions. To verify the significance of differences we used the one-tailed t-test with confidence level) and the null hypothesis stating that the difference is zero[5]. All the differences in results having the value of statistic t above the threshold are considered significant. The exception from the above methodology were the experiments in the finite element mesh design problem, where the experimental methodology was dictated by previous published results, as described in Section 5.4. Besides the classification accuracy, we measured also the average information score[15]. This measure eliminates the influence of prior probabilities and appropriately treats probabilistic answers of the classifier. The average information score is defined as: #testing instances #testing instances where the information score of the classification of i-th testing instance is defined by: is the class of the i-th testing instance, P (Cl) is the prior probability of class Cl and the probability returned by a classifier. If the returned probability of the correct class is greater than the prior probability the information score is positive, as the obtained information is correct. It can be interpreted as the prior information necessary for correct classification minus the posterior information necessary for correct classi- fication. If the returned probability of the correct class is lower than the prior probability the information score is negative, as the obtained information is wrong. It can be interpreted as the prior information necessary for incorrect classification minus the posterior information necessary for incorrect classification. The main difference between the classification accuracy and the information score can be illustrated with the following example. Let the prior distribution of classes be P let the posterior distribution returned by the classifier be P If the correct class is C 1 then the information score is positive while the classification accuracy treats the given posterior distribution as wrong answer. If the correct class is C 2 then the information score is negative while the classification accuracy treats the given posterior distribution as correct answer. Classification accuracy may in some special cases exhibit high variance while information score is much more stable. In a very special case where we have a data set with irrelevant attributes and exactly 50% of instances from one class and 50% of instances from the other class, the leave- one-out testing for a probabilistic classifier would give the approximate accuracy of 50%, while for the"default" classifier, that classifies every instance into the majority class, the accuracy would be 0%. A slight modification of the distribution of training instances would drastically change the latter accuracy to approximately 50%. A more drastic modification of the distribution, say 80% of cases for one class and 20% for the other, would increase the accuracy of the "default" classifier to 80%, while the accuracy of the probabilistic classifier would be approximately 0:8 \Theta 0:8+0:2 \Theta 0:2 = 68%. However, for both classifiers the information score would in all scenarios remain approximately 0 bits which would indicate, that both classifiers are unable to extract any useful information from attributes. 5. Experimental results In this section we give results on several artificial and real-world data sets. The presentation of the experiments is divided into four parts according to four groups of data sets: artificial data sets with the controlled conditional dependency between attributes, some other benchmark artificial data sets, medical data sets, and other real-world data sets. For each group we give a brief description of data sets followed by the results. The results in tables include averages over several runs and standard errors. 5.1. Artificial data sets We generated several data sets in order to compare the performance of various algorithms: INF1: Domain with three conditionally independent informative binary attributes for each of the three classes and with three random binary attributes. The learner should detect which three attributes are informative which is a relatively easy task. All five algorithms should be able to solve this problem. INF2: Domain obtained from INF1 by replacing each informative attribute with two attributes whose values define the value of the original attribute with XOR relation. For this prob- lem, the learner should detect six important attributes and the fact that attributes are pair-wise strongly conditionally dependent. This is a fairly complex problem and cannot be solved with the myopic heuristics. This data set should show the advantage of LFC and Assistant-R. TREE: Domain whose instances were generated from a decision tree with 6 internal nodes, each containing a different binary attribute. 5 random binary attributes were added to the description of instances. This problem should be easy for greedy decision tree learning algorithms while other approaches may have difficulties due to an inappropriate knowledge representation of the target concept. PAR2: Parity problem with two significant binary attributes and 10 random binary at- tributes. 5% of randomly selected instances were labeled with wrong class. This problem is hard as there is a lot of attributes with equal score when evaluated with a myopic evaluation function, such as information gain. PAR3: Same as PAR2 except that there were three significant attributes for the parity relation which makes the problem harder. PAR4: Same as PAR2 except that there were four significant attributes for the parity relation which makes the problem the hardest among the parity problems used in our experiments. The basic characteristics of the artificial data sets are listed in Table 2. Characteristics include the percentage of the majority class (which can be interpreted as "default accuracy") and the class entropy which gives an impression of the complexity of the classification problem. The results of the learning algorithms LFC, Assistant-I and Assistant-R, as well as the naive Bayesian classifier and the k-NN algorithm, are given in Table 3 (classification accuracy) and Table (information score). The results are as expected and show that: ffl All classifiers perform well in a (relatively sim- ple) domain with conditionally independent attributes ffl Both versions of Assistant perform well in the problem of the reconstruction of a decision tree (TREE), while the other classifiers are significantly worse. Only Assistant-R and LFC are able to successfully solve the problems with strong conditional dependencies between attributes (INF2, PAR2- 4). However, of these two, Assistant-R performs better, especially in the case of the hardest problem (PAR4). Note that LFC can solve PAR4 if the depth of the lookahead is increased, Table Basic description of artificial data sets domain #class #atts. #val/att. # instances maj.class (%) entropy(bit) INF2 3 21 2.0 200 36 1.58 Table 3 Classification accuracy of the learning systems on artificial data sets domain LFC Assistant-I Assistant-R naive Bayes k-NN INF1 86.0\Sigma5.1 90.1\Sigma3.5 88.8\Sigma3.8 91.6\Sigma3.1 89.0\Sigma3.6 INF2 67.1\Sigma6.3 55.4\Sigma9.8 68.7\Sigma7.8 32.1\Sigma4.5 56.8\Sigma6.3 TREE 75.8\Sigma5.4 79.2\Sigma5.7 78.8\Sigma6.2 69.0\Sigma5.9 68.2\Sigma5.3 93.6\Sigma3.3 74.9\Sigma7.9 95.7\Sigma2.8 56.7\Sigma5.7 79.4\Sigma4.3 however, the time complexity of the lookahead increases exponentially with its depth. On the other hand, Assistant-R solves all parity problems equally quickly. ffl The information score of the naive Bayesian classifier in the problems with strong conditional dependencies between attributes is poor which indicates that this classifier failed to find any regularity in these data sets. 5.2. Benchmark artificial data sets Besides the artificial data sets from the previous subsection, we used also the following benchmark artificial data sets used by other authors (note that results of other authors can not be directly compared to our results as experimental conditions (training/testing splits) were not the same): BOOL: Boolean function defined on 6 attributes with 10% of class noise (optimal recognition rate is 90%). The target function is: This data set was used by Smyth et al. [31]and they report 67.2\Sigma1.7% of the classification accuracy for naive Bayes, 82.5\Sigma1.1% for back- propagation, and 85.9\Sigma0.9% for their rule-based classifier. LED: LED-digits problem with 10% of noise in attribute values. The optimal recognition rate is estimated to be 74%. Smyth et al. [31] report 68.1\Sigma1.7% of the classification accuracy for naive Bayes, 64.6\Sigma3.5 for backpropa- gation, and 72.7\Sigma1.3 for their rule-based classi- fier. This data set can be obtained from Irvine database[21]. KRK1: The problem of legality of King-Rook- King chess endgame positions. The attributes describe the relevant relations between pieces, such as "same rank" and "adjacent file". Originally the data included five sets of 1000 examples (1000 for learning and 4000 for testing) and was used to test Inductive Logic Programming algorithms[7]. The reported classification accuracy is 99.7\Sigma0.1 %. We used only one set of 1000 examples (i.e. 700 instances for training). KRK2: Same as KRK1 except that the only available attributes are the coordinates of pieces. The same data set was used by Mladeni-c[19]. The reported results are about 69% accuracy for her ATRIS system and 64% for Assistant. The basic description of data sets is provided in Table 5 and results are given in Tables 6 and 7. It is interesting that in the LED domain, the naive Bayesian classifier and the k-NN algorithm reach the estimated upper bound of the classification accuracy. This suggests that all attributes should be considered for optimal classification in this domain. In this problem the attributes are conditionally independent given the class, therefore the good performance of the naive Bayesian classifier is not surprising. However, in the other three domains the performance of the naive Bayesian classifier is poor, due to the strong Table 4 Average information score of the learning systems on artificial data sets domain LFC Assistant-I Assistant-R naive Bayes k-NN Table 5 Basic description of some benchmark artificial data sets domain #class #atts. #val/att. # instances maj.class (%) entropy(bit) Table 6 Classification accuracy of the learning systems on artificial data sets domain LFC Assistant-I Assistant-R naive Bayes k-NN LED 70.8\Sigma2.3 71.1\Sigma2.4 71.7\Sigma2.2 73.9\Sigma2.1 73.9\Sigma2.1 KRK1 98.7\Sigma1.2 98.6\Sigma1.2 98.6\Sigma1.2 91.6\Sigma1.4 92.2\Sigma1.9 KRK2 86.0\Sigma2.1 66.6\Sigma3.1 70.1\Sigma3.3 64.8\Sigma2.1 70.7\Sigma1.7 Table 7 Average information score of the learning systems on artificial data sets domain LFC Assistant-I Assistant-R naive Bayes k-NN LED 2.13\Sigma0.07 2.11\Sigma0.06 2.12\Sigma0.07 2.33\Sigma0.05 2.22\Sigma0.05 conditional dependencies between attributes. The information score (see Table 7) shows that the naive Bayesian classifier provides (on the average) no information in the BOOL and KRK2 domains. The performance of the different variants of Assistant is almost the same, except for the KRK2 domain, where the performance of Assistant-I is poor (note that the default accuracy in KRK2 is 67%). The performance of Assistant-R and the k-NN algorithm is significantly better (99.95% confidence level). However, the information score shows that both, Assistant-R and k-NN, are not very successful in this problem. As expected, without constructive induction it is not possible to reveal regularities in the chess positions described only with the coordinates of pieces. LFC is able to construct important attributes in this domain, which enables it to achieve significantly better results than the other algorithms. 5.3. Medical data sets We compared the performance of the algorithms on several medical data sets: ffl Data sets obtained from University Medical Center in Ljubljana, Slovenia: the problem of locating of primary tumour in patients with metastases (PRIM), the problem of predicting the recurrence of breast cancer five years after the removal of the tumour (BREA), the problem of determining the type of the cancer in lymphography (LYMP), and diagnosis in rheumatology (RHEU). ffl HEPA: prognostics of survival for patients suffering from hepatitis. The data was provided by Gail Gong from Carnegie-Mellon University. ffl Data sets obtained from the StatLog database[18]: diagnosis of diabetes (DIAB) and diagnosis of heart diseases (HEART). For the DIAB data set, Ragavan & Rendell [27]report 78.8% classification accuracy with their LFC al- gorithm. They also report poor performance of Table 8 Basic description of the medical data sets domain #class #atts. #val/att. # instances maj.class (%) entropy(bit) PRIM 22 17 2.2 339 25 3.89 Table 9 Classification accuracy of the learning systems on medical data sets domain LFC Assistant-I Assistant-R naive Bayes k-NN BREA 76.1\Sigma4.3 76.8\Sigma4.6 78.5\Sigma3.9 78.7\Sigma4.5 79.5\Sigma2.7 LYMP 82.4\Sigma5.2 77.0\Sigma5.5 77.0\Sigma5.9 84.7\Sigma4.2 82.6\Sigma5.7 HEPA 79.0\Sigma5.3 77.2\Sigma5.3 82.3\Sigma5.4 86.1\Sigma3.9 82.6\Sigma4.9 HEART 77.3\Sigma5.2 75.4\Sigma4.0 77.6\Sigma4.5 84.5\Sigma3.0 82.9\Sigma3.7 Table Average information score of the learning systems on medical data sets domain LFC Assistant-I Assistant-R naive Bayes k-NN several other algorithms without constructive induction (up to 58%). However, our results (see below) and results of the StatLog project [18] show that the poor results of the other algorithms in this domain are not due to the lack of constructive induction. In our experiments, on DIAB dataset, all classifiers perform equally well, with the exception of the naive Bayesian classifier which is significantly better. The basic characteristics of the above medical data sets are given in Table 8. The results of experiments on these data sets are provided in Tables 9 and 10. In medical data sets, attributes are typically conditionally independent given the class . There- fore, it is not surprising that the naive Bayesian classifier shows clear advantage on these data sets[12]. It is interesting that the performance of the k-NN algorithm is good in these domains, although worse than the performance of the naive Bayesian classifier. The information score (Table 10) for BREA data set indicates that no learning algorithm was able to solve this problem. This suggests that the attributes are not relevant. Both versions of Assistant have similar per- formance, except in the HEPA domain where, Assistant-R has significantly better performance confidence level). A detailed analysis showed that in this problem RELIEFF discovered a significant conditional interdependency between two attributes given the class. These two attributes score poorly when considered indepen- dently. That is why Assistant-I was not able to discover this regularity in data. On the other hand, other attributes are available that contain similar information as these two attributes together. This is the reason why the naive Bayesian classifier performs better. We tried to provide the naive Bayesian classifier with an additional attribute by joining the two conditionally dependent attributes. How- ever, the performance remained the same. achieved significantly better results than the other two inductive algorithms in the LYMP domain, where constructive induction seems to be useful. However, LFC performed significantly worse in the RHEU domain while in the other domains the three inductive algorithms perform equally well. 5.4. Non-medical real-world data sets We compared the performance of the algorithms also on the following non-medical real world data sets (SOYB, IRIS, and VOTE are obtained from the Irvine database[21], SAT is obtained from the StatLog database [18]): SOYB: The famous soybean data set used by IRIS: The well known Fisher's problem of determining the type of iris flower. MESH3,MESH15: The problem of determining the number of elements for each of the edges of an object in the finite element mesh design problem[6]. There are five objects for which experts have constructed appropriate meshes. In each of five experiments one object is used for testing and the other four for learning and the results are averaged. The results reported by D-zeroski [7] for various ILP systems are 12% classification accuracy for FOIL, 22% for mFOIL and 29% for GOLEM and the result reported by Pompe et al. [23] is 28% for SFOIL. The description of the MESH problem is appropriate for ILP systems. For attribute learners only relations with arity 1 (i.e. attributes) can be used to describe the problem. Note that in this domain the training/testing splits are the same for all algorithms. The testing methodology is a special case of leave-one-out, therefore, the results in the tables for this problem have no standard deviations. Quinlan [26] reports results of some ILP systems that achieved over 90% in that domain testing on positive and negative instances. However, those results are misleading. Each positive instance has ten negative instances in average. Therefore we have 11 copies of the same instance and any classification of this instance is correct at least for 9 out of 11 copies which gives 82% classification accuracy for a classifier that always classifies into wrong class. MESH3 contains the three basic attributes from the original database and ignores the relational description of objects. Therefore, in the domain attribute learners are given less information than ILP learners. contains, besides the 3 original at- tributes, 12 attributes derived from the relational background knowledge. In this prob- lem, attribute learners have advantage as they are already provided with additional attributes. The provided description of objects for ILP learners is actually more informative. In princi- ple, the same attributes and a number of additional attributes could be derived by (extremely cleaver) ILP learners from the relational description of the background knowledge. How- ever, this is a fairly complex task. Therefore attribute learners with MESH15 data set have better chances than ILP learners to reveal a good hypothesis. SAT: The database consists of multi-class spectral values of pixels in 3 \Theta 3 neighborhoods in a satellite image, and the classification of the central pixel in each neighborhood. The results of the StatLog project[18] are 90.6% classification accuracy for the k-NN algorithm, 86.1% for backpropagation, 85.0% for C4.5, 84.8% for CN2 and 69.3% for the naive Bayesian classifier (using relative frequencies and not the m-estimate of probabilities). VOTE: The voting records are from a session of the 1984 United States Congress. Smyth et al. [31] report 88.9% of classification accuracy for the naive Bayesian classifier, 93.0% for backpropagation and 94.9% for their rule-based classifier. The basic characteristics of non-medical real world data sets are presented in Table 11. Tables 12 and 13 give the results. On SOYB and IRIS data sets, all classifiers perform equally well. The results of the naive Bayesian classifier indicate that the attributes are conditionally relatively independent in these data sets, which is in agreement with previously published results. On the SAT data set, k-NN significantly out-performs other algorithms which is in agreement with the results of the StatLog project [18]. How- ever, the naive Bayesian classifier with the m-estimate of probabilities reaches the classification accuracy of inductive learning algorithms. The results of the naive Bayesian classifier used in the 14 THE AUTHORS??? Table Basic description of the non-medical real-world data sets domain #class #atts. #val/att. # instances maj.class (%) entropy(bit) Table Classification accuracy of the learning systems on non-medical real-world data sets domain LFC Assistant-I Assistant-R naive Bayes k-NN IRIS 95.0\Sigma3.8 95.7\Sigma3.7 95.2\Sigma2.6 96.6\Sigma2.6 97.0\Sigma2.1 Table Average information score of the learning systems on non-medical real-world data sets domain LFC Assistant-I Assistant-R naive Bayes k-NN StatLog project are much worse. Cestnik [2] has shown that the m-estimate significantly increases the performance of the naive Bayesian classifier which is also confirmed with our experiments. Both versions of Assistant perform the same on all data sets except on the SAT data set where Assistant-R and LFC achieve significantly better result (99.95% confidence level). This result confirms that RELIEFF estimates the quality of attributes better than the information gain. On the VOTE data set the naive Bayesian classifier is the worst, while both versions of Assistant are comparable to the rule based classifier by Smyth et al. [31]. The most interesting results appear in the domains. Although attribute learners in MESH3 have less information than ILP systems, they all outperform the results by ILP systems as reported by D-zeroski [7] and Pompe et al. [23]. With 12 additional attributes in MESH15, the results of inductive learners are significantly improved. All inductive learning systems significantly outperform the naive Bayesian classifier and the k-NN algorithm. A detailed analysis showed that this excellent result by both versions of Assistant is due to the use of the naive Bayesian formula to calculate the class probability distribution in "null" leaves (see Section 3). Namely, for this problem it often happens that testing instances fall into a "null" leaf because there are no training instances that have the same values of significant attributes as the testing instances. The naive Bayesian classifier efficiently solves this problem. LFC generates no "null" leaves as all constructed attributes are strictly binary with values true and false. Therefore, the classification of objects with a different value of the original attribute than all training instances always proceeds to the branch labeled false. The effect of this strategy is that, for a given testing instance, the corresponding leaf contains training instances with same or similar values for most of the attributes that appear on the path from the root to the leaf. This strategy also works well in MESH problems. 6. Discussion Note that the null leaves of both versions of Assistant had no influence on the performance on arti- ficial data sets as there is no missing values in the data. Also, in MESH15 problem the performance of LFC is good although it does not generate null leaves. Therefore, the use of null leaves is not the crucial difference between Assistant and LFC. Equation (3) shows an interesting relation between the RELIEF's estimates and impurity func- tion. RELIEF can efficiently estimate continuous and discrete attributes. The implicit normalization in eq. (3) enables RELIEF to appropriately deal with multivalued attributes. However, if Assistant-I would use eq. (3) instead of the information gain, it would still be myopic. For ex- ample, in PAR2-4 problems, eq. (3) would estimate all attributes as equally non-important. Therefore, the reason of the success of Assistant-R is in the "nearest instances" heuristic which influences the estimation of probabil- ities. This heuristic enables RELIEF to detect strong conditional dependencies between the attributes which would be overlooked if the estimates of probabilities would be done on randomly selected instances instead of the nearest instances. RELIEFF is an efficient heuristic estimator of attribute quality that is able to deal with data sets with conditionally dependent and independent at- tributes. The extensions in RELIEFF enable it to deal with noisy, incomplete, and multi-class data sets. With increasing the number (k) of nearest hits/misses the correlation of RELIEFF's estimates with other impurity functions also increases unless k is greater than the number of instances in the same peak of the instance space. The study reported in [14] showed that RELIEFF has an acceptable bias with respect to other measures when estimating attributes with different number of values The myopia of current inductive learning systems can be partially overcome by replacing the existing heuristic functions with RELIEFF. Assistant-R, a variant of top down induction of decision trees algorithms that uses RELIEFF for estimating the quality of the attributes, significantly outperforms other classifiers in domains with strong conditional dependencies between at- tributes. The myopia of other inductive learners may cause them to overlook significant relations. While this can be easily demonstrated with artificial data sets, it was also shown in two real world problems: HEPA and SAT. In these data sets RELIEFF detected significant conditional interdependencies between attributes, that resulted in a significantly better result by Assistant-R than the result by Assistant-I. One feature of RELIEF not addressed in this paper is that if the same attribute is replicated in a data set, all replications will get the same estimate. With the increasing number of replications the quality of estimates will descrease as the replicated attribute affects the distances between instances. For constructive induction LFC uses a limited lookahead to detect significant conditional dependencies between the attributes. LFC shows similar advantage over other algorithms as Assistant- R does. In one artificial problem (KRK2) and one real world problem (LYMP) LFC performs significantly better due to constructive induction. However, in some cases the constructive induction may spoil the results as is the case with RHEU data set. LFC performs well in most of the prob- lems, which suggests that the limited lookahead is a good search strategy in most real-world prob- lems. The lookahead, however, should have a reasonable limit as the time complexity exponentialy increases with the lookahead depth. Although RELIEFF may overcome the myopia, it is useless in Assistant-R when the change of representation is required. In such cases the constructive induction should be applied. For example, in the KRK2 problem, Assistant-R achieves good result which can not be further improved without constructive induction. A good idea for constructive induction may be to use RELIEFF instead of or in the combination with the lookahead. The naive Bayesian classifier has obvious advantage in domains with conditionally relatively independent attributes, such as medical diagnostic problems. In such domains, the naive Bayesian classifier is able to reliably estimate the conditional probabilities and is also able to use all at- tributes, i.e all available information. It would be interesting to appropriately combine the power of RELIEFF and the naive Bayesian classifier. Current ILP systems [20] are not able to use the attributes appropriately. This was demonstrated in the MESH3 domain where all attribute learn- ers outperformed existing ILP systems. To enable ILP systems to deal with the attribute-value rep- resentation, a combination with the (semi) naive Bayesian classifier could be useful. On the other hand, current ILP systems use greedy search techniques and the heuristics that guide the search are myopic. Pompe and Kononenko [24] implemented an adapted version of RELIEFF in the FOIL like ILP system called ILP-R and prelemi- nary experiments show similar advantages of this system over other ILP systems as Assistant-R has over Assistant-I. 7. Conclusion RELIEFF is an efficient heuristic estimator of attribute quality that is able to deal with data sets with conditionally dependent and independent at- tributes, with noisy, incomplete, and multi-class data sets. The myopia of current inductive learning systems can be partially overcome by replacing the existing heuristic functions with RELI- EFF. The acceptable increase in computational complexity may in certain domains payoff with eventual discovery of strong conditional dependencies between attributes, which cannot be detected using the myopic impurity measure to guide the greedy search. The experimental results indicate that in the majority of real world problems the myopia has no or only marginal effect. One may wonder whether myopia is really worth much attention at all. However, when faced with a new data set it is unreasonable to try only myopic algorithm unless it is know in advance that in the data set there are no strong conditional dependencies between attributes. Any serious application of machine learning on new data should try to discover as much regularities in the data as pos- sible. Therefore, non-myopic approaches, such as one described in this paper, should be used as indispensable tools for analysing the data. Acknowledgements The use of m-estimate in equation (1) was proposed by Bojan Cestnik. We thank Matja-z Zwitter for the PRIM and BREA data sets, Milan Sokli-c for LYMP, Gail Gong for HEPA, Padhraic Smyth for BOOL and LED, Sa-so D-zeroski for KRK1 and MESH, Bob Hen- ery for the DIAB, HEART, and SAT data sets from the StatLog database at Strathclyde University, and Patrick Murphy and David Aha for the data sets from the Irvine database. We are grateful to our colleagues Sa-so D-zeroski, Matev-z Kova-ci-c, Matja-z Kukar, Uro-s Pompe, and Tanja Urban-ci-c and to anonymous reviewers for their comments on earlier drafts that significantly improved the paper. This work was supported by the Slovenian Ministry of Science and Technology --R Wadsworth International Group Estimating probabilities: A crucial task in machine learning. On estimating probabilities in tree pruning. ASSISTANT 86. General Statistics. The application of inductive logic programming to finite element mesh de- sign Handling noise in inductive logic pro- gramming Use of contextual information for feature ranking and discretization. Experiments in In- duction A practical approach to feature selection. The feature selection prob- lem: traditional methods and new algorithm Inductive and Bayesian learning in medical diagnosis. On biases when estimating multivalued attributes. Information based evaluation criterion for classifier's performance. ID3 Revisited: A distance based criterion for attribute selection. Learning by being told and learning from examples: An experimental comparison of the two methods of knowledge acquisition in the context of developing an expert system for soybean disease diagnosis. Combinatorial optimization in inductive concept learning. UCI Repository of machine learning databases Learning decision rules in noisy domains. Linear space induction in first order logic with RELIEFF Induction of decision trees. The minimum description length principle and categorical theories. Lookahead feature construction for learning hard concepts. Learning complex real-world conceptsthrough feature construction Constructive induction with decision trees. Rule induction using information theory. --TR --CTR Xin Jin , Rongyan Li , Xian Shen , Rongfang Bie, Automatic web pages categorization with ReliefF and Hidden Naive Bayes, Proceedings of the 2007 ACM symposium on Applied computing, March 11-15, 2007, Seoul, Korea Use of Contextual Information for Feature Ranking and Discretization, IEEE Transactions on Knowledge and Data Engineering, v.9 n.5, p.718-730, September 1997 Marko Robnik-ikonja , Igor Kononenko, Theoretical and Empirical Analysis of ReliefF and RReliefF, Machine Learning, v.53 n.1-2, p.23-69, October-November David A. Bell , Hui Wang, A Formalism for Relevance and Its Application in Feature Subset Selection, Machine Learning, v.41 n.2, p.175-195, November 2000 Llus Mrquez , Llus Padr , Horacio Rodrguez, A Machine Learning Approach to POS Tagging, Machine Learning, v.39 n.1, p.59-91, April 2000 Saher Esmeir , Shaul Markovitch, Anytime Learning of Decision Trees, The Journal of Machine Learning Research, 8, p.891-933, 5/1/2007 Huan Liu , Hiroshi Motoda , Lei Yu, A selective sampling approach to active feature selection, Artificial Intelligence, v.159 n.1-2, p.49-74, November 2004 Foster Provost , Venkateswarlu Kolluri, A Survey of Methods for Scaling Up Inductive Algorithms, Data Mining and Knowledge Discovery, v.3 n.2, p.131-169, June 1999
impurity function;learning from examples;RELIEFF;estimating attributes;empirical evaluation
590793
Learning Structure from Data and Its Application to Ozone Prediction.
In this paper we propose an algorithm for structure learning in predictive expert systems based on a probabilistic network representation. The idea is to have the simplest structure (minimum number of links) with acceptable predictive capability. The algorithm starts by building a tree structure based on measuring mutual information between pairs of variables, and then it adds links as necessary to obtain certain predictive performance. We have applied this method for ozone prediction in Mxico City, where the ozone level is used as a global indicator for the air quality in different parts of the city. It is important to predict the ozone level a day, or at least several hours in advance, to reduce the health hazards and industrial losses that occur when the ozone reaches emergency levels. We obtained as a first approximation a tree-structured dependency model for predicting ozone in one part of the city. We observe that even with only three parameters, its estimations are acceptable.A causal network representation and the structure learning techniques produced some very interesting results for the ozone prediction problem. Firstly, we got some insight into the dependence structure of the phenomena. Secondly, we got an indication of which are the important and not so important variables for ozone forecasting. Taking this into account, the measurement and computational costs for ozone prediction could be reduced. And thirdly, we have obtained satisfactory short term ozone predictions based on a small set of the most important parameters.
Introduction Learning is defined as "any process by which a system improves its performance" [1]. Since the first days of research in artificial intelligence, the ability to learn has been considered as one of the essential attributes of an "intelligent system", and a considerable amount of research has been done in this area. Learning has focused in acquiring concepts from examples in what is called inductive learning. The development of expert systems has motivated further research in learning to automate the process of knowledge acquisition. This is considered one of the main problems for the construction of knowledge-based systems. An important aspect in inductive learning is to obtain a model which represents the domain knowledge, and is accessible to the user. In par- ticular, it is useful to obtain the dependency information between the variables involved in the phenomena. That is, those factors that are important for certain variable, and those that are not. This is of particular interest in predictive expert systems, when we want to forecast some variables based on known parameters. It is useful to know the parameters which have more incidence in the unknowns, and the ones that have not much influ- ence. A knowledge representation paradigma that captures this dependency information is a probabilistic network. Probabilistic networks (PN) [3], also known as Bayesian networks, causal networks or probabilistic influence diagrams, are graphical structures used for representing expert knowledge, drawing conclusions from input data and explaining the reasoning process to the user. A PN is a directed acyclic graph (DAG) whose structure corresponds to the dependency relations of the set of variables represented in the network (nodes), and which is parameterized by the conditional probabilities (links) required to specify the underlying distri- bution. The structure of the network makes explicit the dependence and independence relations between the variables, which are important: (i) in representing the knowledge of the domain, and (ii) for efficient probability propagation. Fig. 1. A probabilistic network. If we use a PN representation, learning is divided naturally into two aspects: parameter learning and structure learning [2]. Parameter learning has to do with obtaining the required probability distributions for a certain structure. Structure learning has to do with obtaining the topology of the network, including which variables are relevant for a particular problem, and their depen- dencies. We are interested in this second aspect, that is in obtaining the dependency structure of certain phenomena, to get a better understanding of it and to use it as a predictive tool. In section 2 we give a brief introduction to probabilistic networks. Section 3 reviews previous work on structure learning, and section 4 introduces our methodology for obtaining a dependency structure for predictive systems. In section 5 we describe the problem of Ozone prediction in Mexico City, and we present some experimental results in section 6. Finally, we give some conclusions and possible directions for future work. 2. Probabilistic Networks A probabilistic network is a graphical representation of dependencies and independencies for probabilistic reasoning in expert systems. Each node represents a discrete random variable and each arc a probabilistic dependency. The variable at the end of a link is dependent on the variable(s) at its origin, e.g. C is dependent on A in the PN in figure 1, as indicated by link 1. We can think of the graph in figure 1 as representing the joint probability distribution of the variables A; B; :::; G as: Equation (1) is obtained by applying the chain rule and using the dependency information represented in the network. The topology of a PN gives direct information about the dependency relationships between the variables involved. In particular, it represents which variables are conditionally independent given another variable. By definition, A is conditionally independent of B, given C, if: This is represented graphically by node C "sep- arating" A from B in the network. In general, C will be a subset of nodes from the network that if removed will make the subsets of nodes A and B "disconnected". Independence in a PN network is tested with a criteria called D-separation [2]. A DAG representation G of a probability distribution P is an I-map [2] if all the independencies represented in G are present in P . It is a minimal I-map if it is an I-map with the minimum number of links, that is, if any link is removed, there will be an independency relation in G that is not present in P . Formally, a probabilistic network is minimal I-map for a joint probability distribution P [2]. In other words, it is a graph with the minimum number of links that faithfully represents all the probabilistic independencies for a set of random variables. Given a knowledge base represented as a probabilistic network, it can be used to reason about the consequences of specific input data, by what is called probabilistic reasoning. This consists in instantiating the input variables, and propagating their effect through the network to update the probability of the hypothesis variables. In contrast with previous approaches (e.g., MYCIN and (c) (a) (b) Fig. 2. Network structures: (a) tree, (b) polytree, and (c) multiply-connected. Prospector [4]), the updating of the certainty measures is consistent with probability theory, based on the application of Bayesian calculus and the independencies represented in the network. Probability propagation in a general network is a complex problem, but there are efficient algorithms for certain restricted structures, and alternative approaches for more complex net- works. Pearl [3] developed a method for propagating probabilities in networks which are tree- structured, i.e. each node has only one incoming link or one parent. An example of a probabilistic tree is shown in figure 2 (a). In a probabilistic tree every node has only one parent except one node, denoted root, which has no incoming links. Given certain evidence V , represented by the instantiation of the corresponding variables, the posterior probability of any variable taking its i value (B i ), by Bayes' theorem will be: Given the dependencies represented in the tree, separates it into two independent subtrees, one formed by all the descendants of B and the other by every other node. Thus, we can decompose the evidence variables into two sets, V \GammaB which represents all the data rooted at B, and V +B for the data contained in the rest of the network. So (3) can be written as: Given that the two subtrees are conditionally independent given B: Substituting (5) in (4), and applying some alge- bra, we obtain: Where ff is a normalizing constant. Equation (6) provides a product rule for updating the probability on every node in the network by combining the evidence coming from its descendants with the one coming from its parent. It shows that a prior probability is not required, except for the root node (A) for which P the following terms [2]: Then we can write (6) as: Equation constitutes the basis for the propagation mechanism in a probabilistic tree. For this we only need to store the vectors - and - in each node, and update them with the corresponding parameters from its neighbors; and the fixed conditional probability matrix P for that node. This can be implemented by a message passing scheme in which each node acts as a simple process which communicates with its neighbours (fa- ther and sons). Initially the network is in equi- librium. When information arrives, some nodes called data nodes, are instantiated and the information is propagated through the network by each node sending messages to its parent and sons. Each node uses this information to update its local parameters, and update its posterior probability if required. After the messages reach the root node, they will propagate top-down until they reach the leaf nodes, where the propagation terminates and the network comes to a new equilibrium. So the information propagates through the tree in a single pass, in a time (in parallel) proportional to the diameter of the network. An extension for polytrees, was proposed by Kim and Pearl [5]. In a polytree each node can have multiple parents, but it is still a singly connected graph. A polytree is depicted in figure 2 (b). The main difference with the algorithm for trees, is that for multi-parent nodes, the conditional distribution given all their parents is re- quired. The time for propagation is still linearly proportional to the diameter of the network. For more complex, multiply connected net- works, see fig. 2 (c), there are alternative techniques for probability propagation, such as clustering [6], conditioning [2], and stochastic simulation [2]. These methods are efficient for certain types of structures, mainly sparse networks. But in general, probability propagation in a complex network is an NP-hard problem [7]. Thus, for efficiency reasons, and also for clarity and expressiveness, it is important to obtain the simplest structure, with the minimum number of links, which models appropriately the phenomena of interest. A complete graph will be a trivial Imap of any probability distribution, but it would not be useful in terms of knowledge representation or computational efficiency. 3. Structure Learning Approaches Structure learning consists in finding the topology of the network, that is the dependency relationships between the variables involved. Most expert systems obtain this structure from the expert, representing in the network the expert's knowledge about the causal relations in the domain. But for complex problems there might be no expert that has a complete understanding of the domain to obtain all these dependency (and independency) relations, and if so, her/his knowledge could be de- ceiving. Also, knowledge acquisition could be an expensive and time consuming process. So we are interested in using empirical observations to obtain and improve the structure of a probabilistic network. Some previous research has been done on inducing the structure of a PN from statistical data. Chow and Liu [8] presented an algorithm for representing a probability distribution as dependency tree, and this was later extended by Rebane and Pearl [9] for recovering causal polytrees. Chow and Liu's [8] motivation was reducing the memory requirements for storing a n-dimensional discrete probability distribution. For this they developed a method for approximating a probability distribution by a product of second-order dis- tributions, which is equivalent to a probabilistic tree. Thus, the joint probability distribution will be represented as: Y Where X j(i) is the cause or parent of X i . Each variable has one parent, except one (the root) which has no parent, so the method is restricted to a tree structure. They considered the approximation of the original distribution by a dependency tree as an optimization problem, and used a quantity that measures the difference in information contained in the two distributions. That is: x Where the problem is reduced to finding the tree dependency distribution that approximates the original distribution P such that I(P; P ) is minimal. To find the optimum tree, they use the entropy measure for the mutual information between two variables, defined as: x log(P If we assign to every branch in the tree the weight that corresponds to the mutual information between the variables connected by that link, then the weight of the tree is the sum of the weights for all the branches. It can be shown [8] that maximizing the total branch weight is equivalent to minimizing the closeness measure so the tree with the maximum weight will be the optimum tree dependency approximation of P . This result makes it possible to find the optimum tree structure by a simple algorithm that uses only the n(n \Gamma 1)=2 second-order distributions that correspond to all the possible branches for n variables. These are ordered according to their weight, and the two with maximum weight are selected as the first two branches in the tree. Then the other branches are selected in decreasing order whenever they do not form a cycle with the previously selected ones, until all the variables are covered branches). Thus, to obtain a tree-structured PN from sample data, we just need to estimate the joint frequencies and mutual information for all pairs of variables, and then construct the optimum tree by the previous algorithm. Rebane and Pearl [9] extended Chow's method, developing a similar algorithm for the construction of a polytree from statistical data. A polytree is a singly connected network in which each each node can have multiple parents. So the joint probability distribution can be expressed as: Y Where fX is the set of parents of the variable X i . The algorithm for constructing a polytree starts by using the tree recovering algorithm for constructing the skeleton, that is the network without the directionality in the links. Then it checks for the local dependencies between variables and uses this information to determine the directionality of the branches. The local dependency tests is applied to all connected variable triples and, by checking if all variable pairs are dependent or independent, it can partially determine the directionality of the corresponding links. This test is applied to all nodes starting from the outermost ones (leafs) in- wards, until all possible directionalities are found. In general, it is not possible to find the direction of all the branches, and external semantics are needed for completion [2]. Recent work has focused in two aspects: to combine statistical data with expert knowledge; and to induce multiply-connected networks from data. The first approach is based on combing expert knowledge and data to overcome the limitations of the previous techniques, and obtain a more general and complete dependence structure. Sucar et al. [10] start from a structure derived from subjective rules as an initial topology. Then they develop a methodology based on statistical techniques to improve the structure by testing the independence assumptions, and altering the structure if any of them is not satisfied. Kwoh and Gillies [11] extended this work, by creating hidden nodes to improve the structure of a Bayesian tree when the independence assumptions do not hold. Srinivas et al. [12] combine expert knowledge and dependence information (which can be obtained from statistical tests) in an iterative algorithm for approximating the structure of a Bayesian net- work. The expert knowledge they use includes which variables are hypothesis (root nodes), which are evidence (leaf nodes), and partial knowledge 6 THE AUTHORS??? about causality and independence between some of the variables. In the second approach, Cooper and Herskovits [13] developed a Bayesian Method for the induction of probabilistic networks from data. Given certain assumptions about the probability distri- butions, they developed an algorithm for obtaining the most probable Bayesian network given a database of cases. With this algorithm the probability of certain structure given the data can also be obtained, and it can handle missing data and hidden variables. Recently, Lam and Bacchus [14] developed and alternative technique for inducing multiply-connected networks based on Rissa- nen's minimal description length (MDL) principle. Their algorithm tries to make a trade off between the accuracy and complexity of the structure ob- tained. That is, favoring simpler structures even if they are not as accurate as a more complex one. Chow and Liu's algorithm has two important limitations: (i) it is restricted to tree structures, and (ii) it does not obtain the directions of the links (causality). Rebane and Pearl's extension is still restricted in both aspects, generality and directionality. As Lam and Bacchus [14] point out, the approach in [13] assumes a uniform distribution over all possible network structures, so it could favor a much more complex structure even if it is only slightly more accurate. The approach based on the MDL principle [14] overcomes this difficulty by considering both, accuracy and com- plexity. Still, it is based on certain heuristics so that it can not always obtain the optimum solu- tion. The other approaches assume the existence of expert knowledge which is not always available. The special case of predictive systems have certain characteristics, as we explain in the following section, that make the previous algorithms inap- propriate. In particular, most previous techniques consider all the variables at the same level, while in predictive systems accuracy in terms of the un- is the most important factor. 4. Structure Learning for Predictive System In a predictive system there is one (or a few) variable whose value is unknown, and which is estimated based on other known variables. It is possible to have spatial or temporal predictions. In the first case, the unknown is not observable and is estimated from other measured parameters. In the second case, the unknown is in the future and is predicted form present, and past, measure- ments. For instance, in pollution prediction we might want to estimate the pollution level in some part where there are no measurements, or one day in advance. We are interested in obtaining dependence structures for predictive systems, which have some special characterists: ffl There is usually only one variable which we want to predict, so it can be considered the hypothesis or root node. ffl The other variables are evidence nodes which can have different levels of influence in the hypothesis ffl Not all the evidence nodes have direct influence in the hypothesis, but there influence could be through other evidence nodes. Thus, we propose an algorithm for structure learning in predictive expert systems based on the previous observations. The idea is to have the "simplest" structure (minimum number of links) with acceptable predictive capability. Our approach is to start with a PN with the minimum possible number of links that connects all the variables involved. For N variables, the smallest connect graph is a tree, with arcs. This will constitute the "skeleton" of the network. If the predictive accuracy of a tree is good enough then we consider this structure. Otherwise, we start to add links, according to certain criteria, until we obtain the desired performance. The algorithm is the following: 1. Obtain an initial tree structure by Chow and Liu's algorithm. 2. Make the hypothesis variable the root node. This fixes the directions of the links. 3. Produce an ordering of the variables Xng starting from the root, and following the tree according to the order of mutual information between variables. 4. Test the predictive capability of the network: 4.1 If it is satisfactory, stop(1). play outlook temperature humidity windy Tree links Other links Fig. 4. Initial probabilistic tree for the golf example. 4.2 If not, and the number of links is less than a maximum, add a link to the network and to 4. A link is added with the highest mutual information such that: (i) it does not produce a cycle, (ii) the node at the start of the link is a predecessor of the node at the end, according to the previous ordering. 4.3 If not, and the number of links is equal to the maximum, stop(0). Stop (1) indicates successful termination, and stop (0) that it could not achieve the desired predictive performance. The predictive capability is tested statistically, by performing predictions on different data than the training set, and comparing the predictions with the actual values for the unkown. Maximum is the number of links for a completely connected graph, N (N \Gamma 1)=2. The theoretical justification for step (4.2) in the algorithm is based on a general procedure for obtaining a minimal I-Map (a PN in which every independence relation represented in the network is valid) [15]. It consists on defining an ordering of the variables, and constructing a graph such that the "parents" of each variable are a subset of its predecessors that makes it independent from the rest of its predecessors. If the number of arcs reaches the maximum, we obtain a totally connected graph, which represents the joint probability distribution of the N variables without any independence assumptions. As we mention before, this will be a trivial I-map for any distribution. If the complete graph still does not have the desired predictive accuracy, it means that the training data is inadequate for generating an appropriate structure. Either more cases are required (larger sample), or other parameters not included in the set of variables need to be considered To illustrate the procedure, we use a small, hypothetical example for "predicting when to play Table 1 shows the variables and their values for this examples, table 2 shows the set of examples used for training, and table 3 the dependency links (variable pairs) ordered by mutual in- Table 1. Variables for the golf prediction example. Variable Values play Play, Don't Play outlook sunny, overcast, rain temperature continuous humidity continuous windy true, false Table 2. Set of training data for the golf prediction example outlook temp. hum. windy play sunny 85 85 false false sunny overcast rain 70 96 false true rain 68 80 false true rain overcast sunny 72 95 false false sunny 69 70 false true rain sunny overcast 72 90 true true overcast 81 75 false true rain 71 96 true false Table 3. Set of links (variable pairs) ordered by mutual information for the data in table ??. Link Variables 9 windy - play formation. This is a very small data set just used to illustrate the ideas. The initial tree structure, overimposed in the complete graph is depicted in figure 4. A possible ordering for the variables in these case will be: fplay, outlook, temperature, humidity, windyg. In figure 3, the next 3 steps in the algorithm are shown, assuming that the tree structure is not good enough (for this small example we did not test the predictive accuracy). In each one a new link is added according to the mutual informa- tion, and its direction is determined by the variable ordering. The algorithm will terminate when we obtain the desired accuracy, or generate the complete graph (10 links in this example). An interesting aspect to notice is that, unless we need the complete graph, we can usually eliminate some variables for predicting the unknown. For a tree structure, we can eliminate all the variables but the ones directly connected to the hypothesis (root) node. This is because of the independence relations that are represented in the PN. In a tree, a node is independent of all the other variables given its direct parent and sons. For the root node, these are only its direct sons. In general, if the network is not complete (all links present), a subset of nodes will make a node independent of the remaining nodes. Thus, if all the variables but one are known, we can use these independence information to eliminate parameters and simplify the estimation problem. In the following section we introduce the problem of Ozone prediction in Mexico City, and apply the previous algorithm to obtain the dependence structure of this phenomena. play outlook temperature humidity windy play outlook temperature humidity windy play outlook temperature humidity windy (a) (b) (c) Fig. 3. Structures produced by the second stage in the structure learning algorithm for the golf example: (a) first additional link, (b) second, (c) third. 5. Ozone Prediction in Mexico City Air quality in M'exico City is a major problem. Air pollution is one of the highest in the world, with high average daily emissions of several primary pollutants, such as hydrocarbons, nitrogen oxides, carbon monoxide and others. The pollution is due primarily to transportation and industrial emissions. When the primary pollutants are exposed to sunshine, they undergo chemical reactions and yield a variety of secondary pollu- tants, ozone being the most important. Besides the health problems it may cause, ozone is considered as an indicator of the air quality in urban areas. The air quality is monitored in M'exico City in stations, with five of these being the most com- plete. Nine variables are measured in each of the 5 main stations, including: wind direction and velocity, temperature, relative humidity, sulphur dioxide, carbon monoxide, nitrogen dioxide and ozone. These are measured every minute 24 hours a day, and are averaged every hour. It is important to be able to forecast the pollution level several hours, or even a day in advance for several reasons, including: 1. To be able to take emergency measures if the pollution level is going to be above certain threshold. 2. To help industry to make contingency plans in advance to minimize the cost of the emergency measures. 3. To estimate the pollution in an area where there are no measurements. 4. To take preventive actions in some places, as in schools, to reduce the health hazards produced by high pollution levels. In M'exico City, the ozone level is used as a global indicator for the air quality in the different parts of the city. The concentrations of ozone are given in IMECA (Mexican air quality index). So it is important to predict the ozone level a day, or at least several hours in advance using the other variables measured in different stations. Previous work [16] has been done in using neural network techniques to forecast ozone in M'exico City. The results are encouraging for estimating the ozone level up to 4 hours in advance. The problem with these techniques is that we do not get any insight into the structure of the phenom- ena. It will be useful to know the dependencies between the different variables that are measured, and specially their influence in the ozone concen- tration. This will provide a better understanding of the problem with several potential benefits: ffl Determine which factors are more important for the ozone concentration in M'exico City. ffl Simplify the estimation problem, by taking into account only the relevant information. ffl Find out which are the most critical primary causes of pollution in M'exico City which could help for future plans to reduce it. 6. Experimental Results We started by applying the learning algorithm to obtain an initial structure of the phenomena. For this we considered 47 variables [17]: 9 measurements in 5 stations, plus the hour and month in which they were taken. We used nearly 400 random samples, and applied the first step in our algorithm to obtain the tree structure that best approximates the data distribution. This tree-structured Bayesian network is shown in figure 5. We then considered the ozone in one station (Pedregal) as unknown, to estimate it one hour in advance using the other measurements. So we make ozone-Pedregal the hypothesis variable and consider it as the root in the probabilistic tree, as shown in figure 5. From this initial structure we can get an idea of the relevance or influence of the other variables for estimating ozone-Pedregal. The nodes "closest" to the root are the most important ones, and the "far-away" nodes are not so important. In this case we observe that there are 3 variables (ozone-Merced, ozone-Xalostoc, and wind velocity in Pedregal) that have the greatest influence in ozone-Pedregal. What is more, if the tree structure is a good approximation to the "real" structure, these 3 nodes make ozone-Pedregal independent from the rest of the variables (see figure 7). Thus, as a first test of this structure, we estimated ozone-Pedregal using only these 3 variables. The estimation is done with the probability propagation algorithm for trees presented in section 2. This algorithm works with discrete variables only, so continuos variables are discretized in fixed size intervals. We made two experiments: (1) estimate ozone-Pedregal using 100 random samples taken from the training data, and (2) estimate ozone- Pedregal with other 100 samples taken from other data, not used for training. The results for a sub-set of 20 representative samples in each case are shown in figures 6 and 8. We observe that even with only three param- eters, the estimations are quite good. For training data the average error (absolute difference between real and estimated ozone concentration) is 11.2 IMECA or 12.1%, and for not-training data it is 26.8 IMECA or 22.1%. This results should be judged taking into account that this is the first approximation to a dependency model, and that we are only considering 3 variables for estimating the ozone at Pedregal. The neural network model [16] with 46 inputs, has an average error of with a similar set of test (not-training) data. O3_T O3_L O3_Q VV_T CO_T HORA RH_F TMP-T O3_F RH_L RH_T CO_Q CO_L CO_F Fig. 5. A Bayesian tree that represents the ozone phenomena in 5 stations in M'exico City. The nodes represent the variables according to the following nomenclature. For the measured variables, each name is formed by two parts, "measurement-station", using the following abbreviations: the measurements, O3-ozone, SO2-sulphur dioxide, CO-carbon monoxide, NO2-nitrogen dioxide, NOX-nitrogen oxides, VV-wind velocity, DV-wind direction, TMP-temperature, RH- relative humidity; the monitoring stations, T-Pedregal, F-Tlanepantla, Q-Merced, L-Xalostoc, X-Cerro de la Estrella. The other two variables correspond to the time when the measurements were taken, and they are: HORA-hour, MES-month. The same data was used to train and test C4.5 [18]. The error on the test set was of 17.64%. The tree produced by C4.5 is given in figure 9. It is interesting to note that C4.5 considered the wind direction in Pedregal as its principal attribute. A north-south wind direction (? 120 increases the levels of ozone, whether in a south-north direction the ozone levels are mainly located within the first intervals (below 70 IMECAS). We then tested the accuracy of C4.5 by pruning its tree at different depths, that is, considering only the most relevant attributes. The leaves were labeled with the may- ority class of the training set at that level. At depth 4, the accuracy of C4.5 is 21.86%. Considering only the three most important attributes (i.e., at depth 2), and the mayority class for that branch of the tree, C4.5 has an error of 24.80%. Ozone Pedregal Ozone Merced Ozone Xalostoc Pedregal Fig. 7. Reduced tree for predicting Ozone-Pedregal. The accuracy of C4.5 with the complete tree is higher than with our reduced dependency model, with the tree pruned at depth 4 it is about the same, and it is lower with 3 attributes. It is difficult to compare both algorithms because they use a different representation, so a decision node in a decision tree and a variable node in a Bayesian network are not the same. Still, we can consider that the accuracy is similar with these two different representations and learning algorithms, and will expect a higher accuracy if the dependency model is extended with more variables and relations An advantage of the dependency model is that it is generally easier to understand. The relevance of each attribute for predicting certain variable is explicitly represented. This is more difficult to obtain from a decision tree, where an attribute can be repeated as different nodes at different depths. A second advantage is that it gives a probability measure for each value (range) of the hypothesis, which is, in general, not available with a decision tree. Finally, a Bayesian network can be used to predict any variable with any subset of attributes known, while a decision tree is for one variable and with all (or most) of the attributes known. For practical purposes, the ozone measured in IMECAS is divided in several intervals, each of size 50. The air quality and corresponding emergency measures are based on these intervals. In our experiments with the probabilistic model, aprox. 90% of the predictions fall in the same 50 IMECAS interval as the measured ozone level. Fig. 6. Real vs. estimated levels for ozone-Pedregal using 3 variables and training data. 7. Conclusions and Future Work A causal network representation and the structure learning techniques produced some very interesting results for the ozone prediction problem. Firstly, we got some insight into the dependence structure of the phenomena. For example, the ozone in Pedregal is influenced by the ozone in other stations and the wind velocity. This is due to the fact that the pollution in the south (Pedre- gal) of M'exico City, is, in large part, produced by the industrial plants in the north and a dominant north-south wind direction. Secondly, we got an indication of which are the important and not so important variables for ozone forecasting. Taking this into account could reduce the measurement and computational costs for ozone predic- tion. Thirdly, this dependency information could be used for improving other alternative prediction techniques, such as neural networks. With respect to ozone prediction in M'exico City, we plan to continue this work in several aspects ffl Improve the structure of the Bayesian network by using the second part of our algorithm. ffl Obtain the dependence structure for other variables of interest, in particular the ozone in other stations. ffl Test its predictive capability using other vari- ables, assuming that the most influential ones are unknown or not reliable. ffl Improve the longer term predictions by using additional information, such as weather forecasting variables. In structure learning in Bayesian networks in general, there are several research issues which remain to be addressed. Firstly, there is the problem of obtaining the optimal structure in the general case, considering the model's accuracy and computational complexity. Secondly, there is no general algorithm for obtaining the directions of all the links in the network. And thirdly, most structure learning algorithms only consider the observable variables. But, in many cases, the introduction of other variables (called hidden or virtual nodes) can produce simpler structures with an improved predictive capability. We will be addressing these issues in our future research work. Fig. 8. Real vs. estimated levels for ozone-Pedregal using 3 variables and not-training data. RH_F RH_L HORA 484 28687219227104323177Fig. 9. Decision tree for ozone-Pedregal produced with C4.5. Each node represents a decision over the value of a variable: if the value is less than the value shown under the node, then the left branch is followed, otherwise the right one. The leaves represent different classes for ozone-Pedregal (from C0 to C19), with the same discretization used in the Bayesian network. Notes 1. Data taken from public domain file --R "Why Machines should Learn?" Probabilistic Reasoning in Intelligent Sys- tems "On Evidential Reasoning on a Hierarchy of Hypothesis" "Uncertainty Management in Expert Systems" "A Computational Model for Combined Causal and Diagnostic Reasoning in Inference Systems" "Local Computations with Probabilities on Graphical Structures and their Application to Expert Systems" "The Computational Complexity of Probabilistic Inference Using Bayesian Networks" "Approximating Probability Distributions with Dependence Trees" "Objective Probabilities in Expert Systems" "Using Hidden Nodes in Bayesian Networks" "Automated Construction of Sparse Bayesian Networks from Unstructured Probabilistic Models and Domain Information" "A Bayesian Method for the Induction of Probabilistic Networks from Data" "Learning Bayesian Net- works: An Approach Based on the MDL Principle" "Causal Networks: Semantics and Expressiveness" "In- duction of Dependence Structures from Data and its Application to Ozone Prediction" --TR --CTR Elias Kalapanidas , Nikolaos Avouris, Feature selection for air quality forecasting: a genetic algorithm approach, AI Communications, v.16 n.4, p.235-251, December Ciprian-Daniel Neagu , Nikolaos Avouris , Elias Kalapanidas , Vasile Palade, Neural and Neuro-Fuzzy Integration in a Knowledge-Based System for Air Quality Prediction, Applied Intelligence, v.17 n.2, p.141-169, September-October 2002
structure learning;bayesian networks;predictive systems;decision trees;atmospheric pollution
590817
Evolving Neural Networks to Play Go.
Go is a difficult game for computers to master, and the best go programs are still weaker than the average human player. Since the traditional game playing techniques have proven inadequate, new approaches to computer go need to be studied. This paper presents a new approach to learning to play go. The SANE (Symbiotic, Adaptive Neuro-Evolution) method was used to evolve networks capable of playing go on small boards with no pre-programmed go knowledge. On a 9 9 go board, networks that were able to defeat a simple computer opponent were evolved within a few hundred generations. Most significantly, the networks exhibited several aspects of general go playing, which suggests the approach could scale up well.
Introduction Go is hard. For computers at least, this is true. Though the game has not received the level of attention that computer chess, for example, has received, considerable effort has gone into trying to create strong go playing programs. Yet, despite this effort, the best computer programs are still relatively weak. There are a number of reasons why go is hard for traditional computer game playing techniques: the branching factor is prohibitively large, the game is pattern oriented, and there are multiple interacting goals. In fact, the game is so difficult that new techniques are probably going to be needed before go programs are as strong as those that play checkers, chess, or Othello. This paper explores the usefulness of neuro-evolution as a mechanism for learning to play go. The SANE (Symbiotic, Adaptive Neuro-Evolution [7, 8, 9]) algorithm demonstrates that networks that display a general ability in playing go on small boards can be evolved without To appear in Applied Intelligence. y This research was supported in part by NSF under grant #IRI-9504317. \Omega \Gamma\Omega\Gamma(a) Four liberties. \Theta\Gamma\Omega \Gamma\Omega\Gamma\Gamma\Omega \Gamma\Omega \Gamma\Omega \Theta\Gamma\Omega \Gamma\Omega\Gamma(b) One liberty. \Gammaff\Delta\Delta\Gamma\Omega \Theta\Gamma\Omega \Gamma\Omega\Gamma(c) No liberties. Figure 1: The group in (a) has four liberties, or adjacent free positions, while the group in (b) has one. After that last liberty is lost (c), the group is said to be captured and is removed from the board. any prior knowledge about the game. This result forms a promising foundation for scaling up to full-scale go. The paper begins with an introduction to the rules of go followed by a brief word on computer go and why neural network techniques might be useful for go programs. Next the SANE neuro-evolution algorithm is is reviewed, and details of the architecture and the experiments given. The paper concludes with an analysis of the strategies evolved and suggestions for future research. 2 The Game of Go Although the term go is taken from the Japanese word for the game, go is believed to have originated in China more than 3,000 years ago, making it one of the oldest board games still actively played in modern times. Go is an appealing game because it appears simple yet features strategy and tactics that rival games such as chess. Go is played on a square grid 19 intersections across. Smaller boards are often used for teaching purposes. The two players, black and white, alternate placing stones of their respective colors on the intersections of the grid. Game play continues until both players pass, at which time the score is calculated and a winner is determined. Game play is deceivingly simple. Stones can be placed on empty intersections. Once played, a stone cannot be moved to another location. However, a stone or a group of stones can be captured and removed from the board. A liberty is an empty point adjacent to a group of stones. Any group that has no liberties is said to be dead and the stones in that group will be removed from the board. For example, the black group in figure 1a has 4 liberties. Figure 1b shows the same black group with further white stones placed such that the black group now has only 1 liberty. If white were to play an additional move at 1, the black stones would be reduced to 0 liberties. They would be considered dead and removed from the board, as shown in figure 1c. The liberty rule gives rise to the simple concept of an eye. Any group of stones that completely surrounds some interior space is said to have an eye. In figure 2a, the black group has one eye, at point "a". The black group in figure 2b has 2 eyes, at points "b" and "c". The first group can easily be captured if white plays at point "a". However, the second black group cannot be captured by white as white would need to simultaneously occupy \Omega \Gamma\Omega\Gamma \Omega \Gamma\Omega\Gamma \Omega \Gamma\Omega\Gamma \Omega \Gamma\Omega\Gamma \Omega \Gamma\Omega\Gamma \Delta\Delta\Delta\Delta\Delta\Delta\Delta(a) A group with one eye. \Delta\Gamma\Omega \Gamma\Omega \Gamma\Omega \Gamma\Omega \Gamma\Omega \Gamma\Omega \Gamma\Omega\Gamma \Delta\Gamma\Omega \Gamma\Omega \Gamma\Omega \Gamma\Omega \Gamma\Omega \Gamma\Omega \Gamma\Omega\Gamma \Delta\Gamma\Omega \Gamma\Omega b \Gamma\Omega c \Gamma\Omega \Gamma\Omega\Gamma \Delta\Gamma\Omega \Gamma\Omega \Gamma\Omega \Gamma\Omega \Gamma\Omega \Gamma\Omega \Gamma\Omega\Gamma \Delta\Gamma\Omega \Gamma\Omega \Gamma\Omega \Gamma\Omega \Gamma\Omega \Gamma\Omega \Gamma\Omega\Gamma \Delta\Delta\Delta\Delta\Delta\Delta\Delta\Delta\Delta(b) A group with two eyes. Figure 2: Eye space determines life and death of a group of stones. The group in (a) has one eye and can be killed by placing a stone in the eye (position "a"), while the group in (b) has two eyes, "b" and "c", and cannot be killed. \Omega \Gamma\Omega\Gamma \Omega a -\Omega \Gamma\Omega \Phi\Pi\Gammaffi\Gammaffi\Pi Figure 3: Repetition of full board positions is not allowed. If black has just played the marked stone to capture a white stone at "a", white would not be allowed to play at "a" immediately to capture the black stone, because that would recreate the board full board position before the black's move. both points "b" and "c" to reduce the black group to 0 liberties. This is not possible, and therefore the group cannot be killed. This example demonstrates the most common and simplest form of a living group. Forming living groups is one of the primary goals of the game. Previous full-board positions cannot be repeated. This rule is known as the ko rule. Figure 3 shows an example of how the rule applies. Suppose black has just played the marked stone to capture a white stone at point "a". White would not be allowed to immediately play at point "a" to capture the marked stone because it would recreate the board position before black's move. White must instead play elsewhere, and thereby create a new full-board position. If black does not move to point "a", white would then be allowed to play at "a" on a subsequent move because the full-board position would no longer be repeated. Play continues until there are no more moves of value to be made and both players pass. Each side receives a score where all stones and all locations completely surrounded by groups of the same color are counted as points. A komi, 5.5 points in a typical even game, is added to white's score to offset black's first move advantage. The player with the highest score wins. Because komi is not an integer, a tie cannot occur. Unlike many other board games, go provides an easy way to handicap games so that players of different ranks can play an even game. White will give black a certain number of free moves at the beginning of the game. This advantage is generally well defined, and go ranks are based on it. A player who is ranked 2 stones above another player should be able to give the other player 2 free moves in order to play an even game. 3 Computers and Go Despite the relatively simple game play, computers have had little success mastering the game. Whereas in games like chess, Othello and checkers the traditional game playing techniques such as minimax search and its variations are competitive even at the master level [4, 5, 13], those techniques, when applied to go, have not been able to produce programs that can challenge even weak amateur players. The best computer programs in the world are ranked 6-8 stones below what would be considered master level. Progress continues to be made; however, the gap is so large that traditional techniques are unlikely to reach even the weakest master levels for some time to come. A game of go can be divided into three general phases: the opening, the midgame and the endgame. Computers have had varying success in each of the these stages, revealing insight into what can and cannot be achieved with computational methods. 3.1 The Opening The opening stage of the game is referred to as the fuseki. Typically play starts in the corners. Specific sequences in the corners are referred to as joseki, and they are similar to book openings in chess. However, the fuseki typically refers to the direction of play as it relates to all 4 corners. Good go programs tend to have joseki move databases that range from 5,000 to 50,000 moves, and current programs do not have any difficulty in playing through a database of joseki sequences. Choice of joseki and choosing between variations is more troublesome; however, play by the computers is not advanced enough to consider such issues. Even with such limited techniques, some programs are capable of playing very good openings. 3.2 The Midgame As play moves into the midgame, search-based programs begin to have difficulty. One reason is the sheer size of the game. On a 19 \Theta 19 board, there are typically 200-300 potential moves available from a midgame position, so brute force searching of the game tree is not a viable option. Current go programs apply a wide variety of techniques to control the complexity of the midgame. Typically, a move generation facility is used to generate a number of candidate moves from a position using techniques such as pattern libraries, tactical analysis, and rule-based expert systems. Then, the candidate moves are evaluated, usually through static board evaluation functions. Some programs rely solely on the evaluation function for choice of moves while others attempt limited global search using traditional search techniques. Search plays a more prominent role in local tactical analysis where the number of moves and the size of the search tree are significantly smaller. There are several problems with the current midgame techniques. They tend to be difficult to apply and error-prone. Because most evaluation techniques are static, it is difficult to achieve general play, and advance can only be made by laborous human tweaking of rules, patterns and databases. 3.3 The Endgame The endgame is typically significantly easier for computers. During the endgame, programs are capable of playing very well because the branching factor is significantly reduced and patterns are local in nature. However, in the endgame the score is often already mostly decided and fewer points are usually at stake. Even perfect endgame play might only change the score by a small number of points. So, the most important and most difficult part of playing go is in the midgame where the traditional techniques are the weakest. 4 Neural Networks and Go Go is largely pattern based; as a matter of fact, go players often refer to board positions as shape. Groups of stones can be said to have good shape or bad shape depending on the shape's potential of creating a living group and of efficiently capturing territory. Human players instinctively know where to search for moves based largely on learned knowledge of shape. Although there are many techniques that highly skilled players use and computer programs do not, viewing the board as a search node instead of a collection of shapes and patterns is probably the most significant factor holding computer go programs back. Neural networks are very good at pattern transformation tasks, and thus could well be applicable to go. A network could be trained to compute a mapping between the input space, that is, the current board position, and the output indicating the next move. The main problem with this approach is the credit assignment problem. Suppose a standard backpropagation neural network [12] were being trained to play go. For backpropagation to work, advance knowledge about the best move at any given position would be required. Such knowledge is difficult to come by. In reality, only the final game result is available. The credit assignment problem is the problem of determining which of the many moves played were good and deserve credit for a win, and which were bad and deserve to be blamed for a loss. In go, this problem is severe enough that standard learning techniques such as backpropagation cannot be effectively applied. Adaptive Neuro-Evolution [7, 8, 9]) solves the credit assignment problem by using evolutionary algorithms to search for effective neural networks. Instead of punishing or rewarding individual moves, networks are evaluated, selected, and recombined based on their overall performance in the game. Evolutionary algorithms perform a global, parallel 1 SANE is described in more detail in [8], and the source code can be obtained from http://www.cs.utexas.edu/users/nn/. Label Weight-0.7-0.6 -1.2 Input Layer Output Layer Figure 4: A three-layer feedforward network is created from 3 neurons. The neurons are shown on the left, and the corresponding network is shown on the right. Labels indicate which input or output unit a connection corresponds to, while the weight indicates the strength of the connection. search and are guided by a fitness function that measures the goodness of a particular solution. The search tries to maximize the goodness level throughout the search space to find the best solution. SANE differs from other approaches to neuro-evolution systems where each individual in the population represents a complete neural network. In SANE, two separate populations are maintained and evolved: a population of neurons and a population of neural network blueprints. The neuron level evolution explicitly decomposes the search space and maintains a high level of diversity throughout evolution. The blueprint population maintains and exploits effective combinations of individuals in the neuron population. Conjunctively, the two levels of evolution provide an efficient genetic search that is capable of solving difficult real-world decision problems with minimal domain information [8]. In the neuron population, SANE evolves a large population of hidden neuron definitions for a three-layer feedforward network (figure 4). A neuron is represented by a series of connection definitions that describe the weighted connections of the neuron from the input layer and to the output layer. Each neuron has a fixed number of connections, but may allocate them arbitrarily among the units in the input and output layers. A connection definition consists of a label and weight pair. The label is an integer value that specifies a specific input or output unit, and the weight is a floating point number that specifies the strength of the connection. Figure 4 gives three example hidden neuron definitions and the resulting neural network. The activation of a neuron is computed as the sum of all the connected input units multiplied by their weights and passed through the sigmoid activation function application, the output units are linear so that both positive and negative values and be generated. Neural networks could be formed by randomly choosing neurons from the neuron popu- lation. In fact, this approach performs well in simpler problems [7, 10]. However, random participation does not retain knowledge of the best combinations of neurons and can often l l l l l l l l l l l l l l l Network Blueprint Population l l l l l l l l l Neuron Population Figure 5: The network blueprints consist of a set of neurons in the neuron population. A neural network is formed from a blueprint by following its neuron pointers and decoding the respective neurons. stall the search in more difficult problems [8]. To focus the search on the best neuron com- binations, SANE maintains and evolves a separate population of good neuron combinations called neural network blueprints. The blueprints are made up of a series of pointers to members of the neuron population and define an effective neural network from a previous generation. Figure 5 shows how the network blueprint population and the neuron population are related. SANE integrates the neuron and blueprint populations in a generational evolutionary algorithm that iterates over two phases: an evaluation phase and a reproduction phase. In the evaluation phase, SANE simultaneously evaluates the blueprints and the neurons. A blueprint is evaluated by the performance of the network that it specifies. A neuron is evaluated based on the performance of the networks in which it participates. The basic steps in the SANE evaluation phase are shown in the following pseudo code: for each neuron n in population P n n:fitness n:participation for each blueprint b in population P b neuralnet / decode(b) b:fitness / task(neuralnet) for each neuron n in b n:fitness b:fitness n:participation for each neuron n in population P n n:fitness / n:fitness / n:participation Neural networks are formed from each blueprint and evaluated in the task environment. The evaluation score is given to each blueprint and is added to each neuron's fitness variable. After all blueprints have been evaluated, each neuron's fitness is normalized by dividing the sum of its scores by the total number of networks in which it was a participant. In the reproduction phase, SANE uses common genetic operators such as selection, crossover, and mutation to obtain new blueprints and neurons. Each population is ranked based on fitness and a mate is selected for each of the elite individuals. In this application, the elite parameter is defined as the top 15% in the blueprint population and the top 25% in the neuron population. The mate for each elite individual is selected from the other elite individuals. SANE uses a one-point crossover to mate two individuals, which creates two offspring. The offspring from each of the crossover operations replace the worst performing individuals (according to fitness) in the population. All individuals that are not explicitly replaced by offspring remain in the population, although they may be mutated. A conservative mutation rate of 1% per chromosome position is used on the neuron pop- ulation, because neuron evolution automatically maintains high diversity (good networks require serveral different types of neurons). A more aggressive, two-tiered strategy is used on the blueprint level. First, a small number (approximately 1%) of neuron pointers in each blueprint are swapped with randomly selected neurons in the neuron population. Second, pointers to breeding neurons are replaced by pointers to their offspring with a 50% prob- ability. The second mutation component promotes utilization of offspring neurons, which has two advantages. First, it creates diversity in the blueprint population, and second, it explores new structures created by the neuron population. 6 Applying SANE to Go SANE has previously been shown effective in several sequential decision tasks including robot control [7, 8, 9], constraint satisfaction [10], pursuit and evasion [3], and the game of Othello [6, 8, 10]. This paper will evaluate the usefulness of SANE in learning to play go. SANE is used to evolve networks to play on small boards against a simple computer opponent, and the scale-up properties are evaluated. In order to apply SANE, three aspects of the architecture must be specified: the network parameters, evolution parameters and the evaluation function. Let us look at each one of these in turn in the go task. 6.1 Network Parameters SANE evolves standard three-layer feed-forward networks. The network architecture is fixed; only the associated weights and connections are evolved. The number of units depends on the board size. There are 2 input units and one output unit for each board position. The Board size Neuron Population Blueprint Population Number of neurons per network 5 \Theta 5 2000 200 100 7 \Theta 7 3000 200 300 9 \Theta 9 4000 200 500 Table 1: Network definitions used for evolving networks for various small board sizes. first input unit indicates whether a black stone is present at that location, and the second unit whether a white stone is present. Since only one stone can occupy any given board position at a time, both input units cannot be active simultaneously for any position. The output units are signed floating point values. A positive value indicates a good move. The larger the value, the better the move. Negative (or output indicates that the move is not suggested. 6.2 Evolution Parameters Most aspects of SANE are easily tunable. Some experimentation was done to find good values, however, it was not necessary to find optimum values as SANE operates well as long as the values are withing reasonable ranges. The neurons evolved are 312 bits long and represent a set of 12 weights connecting either from input layer to hidden layer or from the hidden layer to the output layer. Table 1 shows the population and network sizes used in conjunction with the various board sizes. Each generation, 200 networks were formed. This allowed each neuron on average to participate in 10-25 networks per generation. Mutation occurred at a rate of 0.1% The crossover operation was a one-point crossover between neurons or networks in the breeding population. The top 25-30% of the population were allowed to breed. 6.3 Evaluation Function The most difficult aspect of the evaluation function was deciding on a set of evaluation criteria that could be computed completely without human intervention. The first difficulty is in determining the end of the game. When humans play, the end of the game is decided by agreement. When the players feel the game is over, they pass their turn. Stones that are mutually agreed to be dead are removed from the board. If there is a dispute, play can be resumed to settle the issue. After the status of each group is determined, a final score is calculated. Since there is no separate output unit for pass, the network can pass only when none of its positive output units (if any) correspond to a legal move. Because there is no arbitration phase for disputed groups, a series of 3 passes is required to end the game. This simplifies certain endgame situations where ko (i.e. repetition) might occur. Removing dead stones is more difficult. Rather than defining a separate protocol for this task, the evaluation function requires a player to explicitly kill any stones it thinks are dead. Human players find this process tedious, so those groups that are obviously dead are simply removed from the board at the end of the game. For computer opponents, killing groups is not so tedious. If all stones on the board are considered alive, the need for settling disputes after the game is over is eliminated, and the task of scoring is greatly simplified. An upper bound is placed on the number of moves, so that it is not actually necessary to check for repetition of entire board positions. It is enough that only the simple ko (demonstrated in figure 3) is checked. An upper bound also ensures that non-repeating but prohibitively long sequences are not followed. Games between human players do not involve such sequences. However, they may occur in a game by an unskilled program. An example of such a sequence would be the filling in of a player's own eye space, which allows a previously alive group to be killed. If two unskilled opponents play in this manner, excessively long sequences of play might result. Such play is punished by counting excessively long games as a loss for the network. Because this behavior is selected against, the networks become less likely to develop it. Since all stones still on the board are presumed to be alive, determining the score becomes a straightforward task. Simple Chinese scoring, where all stones of each color and all locations completely surrounded by stones of that color are counted as points, is used. The evaluation function must produce a fitness level for the network, and it should be a fine-grained value so that slight improvements in the network's play can be rewarded. In our experiments, the difference in score between SANE and its opponent (for example +10.5 or -7.5) is summed over N games, which allows for good resolution in determining improvement. 7 Results SANE was tested with various board sizes. The opponent used was wally (written by Bill Newman), a simple publicly available go program. Wally is a good choice for an opponent for several reasons. First, wally is one of the few go playing programs available in source code. This turns out to be particularly helpful when trying to adjust parameters, like the degree of randomness, to make the opponent more useful as a training partner. Second, wally's skill level is appropriate for a first training partner. It is strong enough to be a challenge to an unskilled network without being so strong that progress cannot be made. 7.1 Evolution Efficiency was able to evolve a network that could defeat wally on small boards. On a 5 \Theta 5 board, SANE needed only 20 generations, on a 7 \Theta 7 board, 50 generations, and for a 9 \Theta 9 game, 260 generations. These numbers are averages over 100 - 1,000 simulations, requiring the ability to beat the opponent 75% of the time. The network was playing black without a komi, which is an equivalent to a 1-stone handicap for the network. Although these results were relatively easy to get, they take a lot of CPU time (up to 5 days for the 9 \Theta 9 board). Moreover, the training times increase with board size quite rapidly. It can be estimated that for a 13 \Theta 13 board, several thousand generations would be required, and for a full-size 19 \Theta 19 board, perhaps tens of thousands. The CPU time for such simulations could be more than a year with the current CPU speeds, and was not available. However, it is still possible to get an idea of how well such a network plays go by studying the effect of nondeterminism and handicap in the opponent. 7.2 Effect of Nondeterminism An important issue with developing general go playing is the degree of determinism of the opponent. SANE actually manages to learn to defeat more deterministic opponents very rapidly. However, in those cases the network learned little about playing go and only learned what was necessary to win against that particular opponent. To force the network to learn more diverse solutions, 10% non-determinism were applied to wally. This means that 10% of the time, instead of making the normal move, a random legal move would be played instead. The 10% value was chosen experimentally to be a reasonable value. Smaller values did not significantly increase the diversity of the games played nor the solutions learned, and larger values made the opponent behave too randomly and easy to beat. As a test of generality, one network was evolved against the original wally on a 7 \Theta 7 board, while another was evolved against wally playing with 10% randomness. An otherwise deterministic player playing occasional random moves should be weaker in absolute terms. However, when playing a series of games against a learning opponent, the deterministic player turned out to be easier to beat. The first network learned to defeat wally very rapidly. However, it would be defeated easily by the weaker but less-deterministic wally. In fact, it would even lose some games against the randomly moving opponent. The network's behavior in this case was not diverse enough to be useful against other opponents. Instead of learning moves that represent general go-playing ability, the network simply learned tricks and simple sequences that utilize flaws in the static opponent. The network playing against the less deterministic opponent required more generations to train. However, the solution evolved was capable of defeating wally at various levels of determinism, including its normal mode of play, and did not lose to the random opponent. 7.3 Effect of Handicaps Since few stronger go playing programs are freely available, there was no good opportunity to evolve networks against other opponents. However, the go handicapping mechanism does provide a way to modify the difficulty of the game against a given opponent. Networks were evolved on the 7 \Theta 7 board while giving wally differing handicaps. Initially, the networks were evolved to play black and make the first move. After about 50 generations, a network evolves to defeat wally. The networks were then evolved with wally playing the first move. After 130 more generations, a network was able to beat wally 75% of the time. \Omega \Omega \Omega \Gammaj\Pi\Pi-ffi\Gammaffi(a) \Gammaff\Gamma\Omega \Gamma\Omega\Gamma4 a \Gamma\Omega \Gamma\Omega\Gamma4 \Gammaff\Gamma\Omega \Gamma\Omega \Gamma\Omega \Gamma\Omega -ff\Gamma\Omega \Gamma\Omega \Gamma\Omega \Gamma\Omega \Theta\Gamma\Omega \Gamma\Omega \Gamma\Omega \Gamma\Omega \Gammaj\Pi b \Gammaffi\Gammaffi(b) Figure This position is from a game played between wally (white) and the network (black). After white plays the marked stone in (a), black should be dead in the left corner as it should not be able to make the two eyes necessary to live. But the network has learned a trick. It plays the marked stone in (b), threatening to capture the white stone below "a". Instead of playing at "b", which would ensure the death of the black group, it plays the move at "a" to defend the single stone below it. As a result, the network can move to "b" and live. At this level, it could give wally a 2-stone handicap and still win 45% of the games. With these results, we can get a rough idea of the level of play evolved. Handicap stones on a small board represent a larger difference in skill than on a larger board. For example, a single handicap stone on a 13\Theta13 board represents approximately 3 stones on a 19\Theta19 board. On a 9 \Theta 9 board, the difference is about 4-5 stones. Thus, the 2 stone handicap on a 7 \Theta 7 board may represent a difference in skill of about 10 stones, which is quite significant and would allow a good amateur compete with master-level player on the full board. Although this is just a rough estimate, it shows that quite powerful go play can be achieved through neuro-evolution methods. 8 Strategies Evolved Given that the networks started with no prior knowledge on how to play go, an important question is: what kind of strategies did they evolve? One peculiar problem with the evolutionary approach is that the strategy evolved often exploits weaknesses found in the particular opponent rather than representing good general go playing abilities. Figure 6 is an example of such a situation taken from an actual game played by a network against wally on a 7 \Theta 7 board. The network is playing black and wally is white. In this situation, white plays the marked stone in 6a. This is a move that should effectively kill the black group in the lower left corner because the black group would no longer be able to make two eyes. However, black has learned that it can actually win in this situation against wally. It plays the marked move in 6b, which makes a single eye and threatens to capture the single white stone above it at "a". The correct move for white is to play "b" next. Allowing black to play at "b" would give black life. However, this particular computer opponent does not realize this and picks the tiny defensive move at "a" over the large offensive move at "b". The network has learned to take advantage of this weakness and moves to "b" as its next move. That such strategies would evolve is understandable considering that the opponent is the only source of information the network has about the game. The network is never explicitly taught about living or dead groups. It's concept of a living group is any group that the opponent cannot kill. In this case, since the opponent cannot kill this group, the network learns it as a favorable position. This example emphasizes the importance the training partner has on the strategies learned. One possible method for forcing the network to rely less on exploiting this type of weakness in specific opponents is to evolve against a variety of opponents. The network would be less likely to learn these kinds of techniques because it is less likely that the same flaw will be present in multiple opponents. This type of evolution would have a better chance of producing a well-rounded go playing program. However, the lack of multiple freely available go playing programs makes this approach impractical at present time. The neuro-evolution system is clearly learning enough to defeat a simple opponent, but are the networks evolving to play go in any general sense? A closer inspection of game transcripts shows that especially when evolved against a nondeterministic opponent, the networks demonstrated a reasonable amount of diversity and were able to cope with variations in play from the opponent. At the beginning of evolution, the networks' outputs are essentially random. After a few generations, they start to make simple living groups. Typically, they evolve the capability to make one or two such groups along the edges or in the corners, and to extend them from there. As evolution continues, the networks become more flexible and capable of developing a greater variety of living positions. Such a strategy is valid, although not particularly strong. Since this type of strategy is all that is required to defeat the computer opponent, the network really does not need to develop more advanced strategies. Against more powerful opponents, the situation would be different. The experiments with handicaps show that in such cases, more powerful strategies are likely to develop. Some well-known general go strategies were also evolved. For example, consider choosing the first move. In the first few generations, the network plays quite randomly, and therefore its first move tends to be on the edge or the outer lines of the board since they comprise 56 of the 81 positions on a 9 \Theta 9 board. Such a move is not a good idea, however, because it is likely to lead to a losing position. Indeed, in a few generations the networks start to make more opening moves near the center of the board. Since the earlier strategy led to losses, the networks that did not use that strategy are now more prevalent in the population. Later on in evolution all the best networks open at or near the center of the board, which is exactly the strategy good go players use. Remarkably, the evolution system discovered it entirely on its own, based on what moves led to wins and losses. This result suggests that the neuro-evolution method is capable of developing good go playing strategies without preprogrammed knowledge, directed by the sparse reinforcement of the game outcomes only. \Omega \Gamma\Omega\Gamma \Phi\Pi\Pi\Pi\Pi\Pi\Pi(a) \Omega \Upsilon\Omega \Xi\Omega\Gamma3 \Omega \Pi\Omega \Omega \Gamma\Omega\Gamma \Theta\Delta\Delta\Theta\Omega \Delta\Omega \Gamma\Omega \Gamma\Omega \Theta\Delta\Delta\Delta\Gamma\Omega \Gamma\Omega\Gamma jffffl\Omega fi\Omega\Gamma443 ffifffl\Omega ff\Omega \Phi\Omega\Gamma45 \Theta\Omega \Omega \Psi\Omega \Upsilon\Omega \Xi\Omega\Gamma4 \Theta\Delta\Sigma\Omega \Pi\Omega \Omega \Gamma\Omega\Gamma \Theta\Delta\Delta\Theta\Omega \Delta\Omega \Gamma\Omega \Gamma\Omega \Theta\Delta\Delta\Delta\Gamma\Omega \Gamma\Omega\Gamma \Phi\Pi\Pi\Pi\Pi\Pi\Pi(c) Figure 7: Figure (a) shows a position known as a ladder, which retains its shape as it is grown in successive moves. The life or death of the white stone depends on groups far away on the board. Figures (b) and (c) show two such situations: in (b) the white stone in location "b" allows white to escape. In (c) the ladder reaches the edge of the board and the white group is killed. In real games, such variations can span the whole board and are difficult to evaluate with only local methods. 9 Future Work There are three main issues for future work: scaling up to larger boards, enhancing the network architecture, and evolving against stronger opponents. 9.1 Larger Boards Ideally, a network should be able to play on any board size. Currently the networks can only play on the board on which they were evolved. For example, a network that was evolved on a 9 \Theta 9 board is not able to play on a 7 \Theta 7 board. One possibility would be to design a representation that is independent of board size. Another would be to evolve solutions that only consider a portion of the board at a time. This type of evaluation function could then be extended to cover boards of varying sizes. However, considering only local board positions instead of the whole board tends to result in weaker play. To see why, consider the position shown in figure 7, known as a ladder. Based on only the local position, it is impossible to tell whether or not the white group can escape. White can play at point "a", and may live or die depending on stones that are on the other side of the board. If there is a white stone at point "b", for example, white can easily live. The extra stone allows white to break out of the ladder, as can be seen in figure 7b. However, if there are no stones on the area, white cannot live. Eventually the ladder position faces the edge of the board, where it is a losing position for white as can be seen in figure 7c. This way stones that are far away from the current play can transform the position drastically. These types of positions can span the entire board, not merely one corner. Recognizing such distance relationships is essential for playing go on larger boards, yet they cannot be captured with methods that consider only part of the board at a time. It is likely that other types of network architectures need to be employed before play on full boards becomes practical. Possibilities include architectures that use preprogrammed features or are hierarchically organized, 9.2 Advanced Architectures Evolving simple unstructured neural network architectures without any prior knowledge demonstrates the feasibility of the neuro-evolution approach. There are several ways the architecture could be enhanced to make it more effective, including preprogrammed feature detectors and hierarchies of networks. Since the networks are not given any prior knowledge about what features are relevant to playing go, they are forced to discover useful features themselves. Allowing the network to access a pre-defined feature space instead of looking at the raw board might make the task easier [2]. Such features could include common patterns and positions such as an eye or a group or even complicated constructs such as the ladder. These features would then be used as inputs to the neural network [1, 11]. It would still probably be useful to let the network develop its own features as well, but the pre-programmed features might allow it to learn faster and deal with more complex patterns. SANE demonstrates the feasibility of evolving structures on more than one level at the same time. It should be possible to extend this idea and evolve a hierarchy of networks, where the lower levels would provide the inputs for networks at higher levels. In effect, evolution would be searching for an effective combination of networks, much the same way it is searching for an effective combination of weights and neurons now. When the task of playing go is broken into such subtasks, it may be the case that the number of generations required will increase with the number and size of the networks evolved and not with the size of the board. If this is the case, then evolving networks that play on full-size boards would no longer be computationally prohibitive. 9.3 Stronger Opponents Even with more sophisticated architectures, stronger opponents are necessary in order to achieve truly high levels of play. The ability to use handicaps to simulate stronger opponents is a useful technique but not enough alone. The techniques used in handicap games are different than those that would be used against a stronger player in an even game. Handicap stones allow the weaker player to build stronger positions, but it still continues weak play from these positions. On the other hand, in an even game the opponent may play brilliant moves that would never be seen in a handicap game. If evolution is never exposed to such moves, it cannot develop comprehensive go skills. A variety of stronger opponents would allow for a greater generality of play to evolve. However, it is not known how great the difference in play would be nor what the effect on the time required to evolve the networks would be. It is also not yet clear how much diversity is necessary to achieve general play. One problem with using stronger opponents is that they tend to take considerably longer to generate moves than weaker programs. Given the large number of trial games generated every generation, it may not be possible to evolve against a slow opponent in a reasonable amount of time. However, the evaluation function might be modified to compensate for the lack of time. Each generation, a significant number of the networks evolved are significantly weaker than the networks of the previous generation. It should be possible to distinguish the weaker networks from the stronger networks by the use of a faster but weaker opponent. Only those networks that appear promising need be evaluated fully against the slower opponent. Even if stronger computer opponents are used, eventually they would be exhausted. It would be necessary to find a way to evolve networks against actual human players. Given the popularity of internet-based go servers, there is no shortage of human players. However, there would be difficulties, particularly in the generation of fitness values. Fitness is used to distinguish the better networks from the worse networks in any given generation. It requires that the evaluation function be consistent for all networks evaluated. Since it would be unlikely for many different networks from the same generation to play the same human opponent, it would be difficult to assign a fair fitness value. The problem is compounded in that the strength of the human opponent is not always known and cannot be reliably used to weight game results against the strength of the human opponent. Nevertheless, good results have been reported in neuro-evolution with noisy evaluation functions[8], suggesting that the problems could be overcome. This way perhaps go-playing programs could finally be evolved that were able to compete with the best humans. Conclusions Traditional artificial intelligence techniques have been insufficient for building go programs that would be competitive at high levels of play. It appears new techniques based on pattern recognition and learning will be required to reach these levels. The SANE neuro-evolution approach is one such promising direction. Networks were evolved to defeat a publicly available go program on small boards with no pre-programmed knowledge of the game, and they exhibited several aspects of general go strategies. --R The Golem go program. The integration of a priori knowledge into a go playing neural network. Incremental evolution of complex general behavior. A grandmaster chess machine. The development of a world-class othello program Discovering complex Othello strategies through evolutionary neural networks. Efficient reinforcement learning through symbiotic evolution. Symbiotic Evolution of Neural Networks in Sequential Decision Tasks. Evolving obstacle avoidance behavior in a robot arm. Learning sequential decision tasks. Exploratory learning in the game of go. Learning internal representations by error propagation. CHINOOK: The world man-machine checkers champion --TR --CTR Karen T. Sutherland, Book reviews, intelligence, v.11 n.3, p.47-54, Sept. 2000 Khosrow Kaikhah , Ryan Garlick, Variable Hidden Layer Sizing in Elman Recurrent Neuro-Evolution, Applied Intelligence, v.12 n.3, p.193-205, May-June 2000 A. Agogino , K. Stanley , R. Miikkulainen, Online Interactive Neuro-evolution, Neural Processing Letters, v.11 n.1, p.29-38, Feb. 2000
neuro-evolution;game playing;sequential decision making;symbiotic evolution
590834
Multiple Adaptive Agents for Tactical Driving.
Recent research in automated highway systems has ranged from low-level vision-based controllers to high-level route-guidance software. However, there is currently no system for tactical-level reasoning. Such a system should address tasks such as passing cars, making exits on time, and merging into a traffic stream. Many previous approaches have attempted to hand construct large rule-based systems which capture the interactions between multiple input sensors, dynamic and potentially conflicting subgoals, and changing roadway conditions. However, these systems are extremely difficult to design due to the large number of rules, the manual tuning of parameters within the rules, and the complex interactions between the rules. Our approach to this intermediate-level planning is a system which consists of a collection of autonomous agents, each of which specializes in a particular aspect of tactical driving. Each agent examines a subset of the intelligent vehicles sensors and independently recommends driving decisions based on their local assessment of the tactical situation. This distributed framework allows different reasoning agents to be implemented using different algorithms.When using a collection of agents to solve a single task, it is vital to carefully consider the interactions between the agents. Since each reasoning object contains several internal parameters, manually finding values for these parameters while accounting for the agents possible interactions is a tedious and error-prone task. In our system, these parameters, and the systems overall dependence on each agent, is automatically tuned using a novel evolutionary optimization strategy, termed Population-Based Incremental Learning (PBIL).Our system, which employs multiple automatically trained agents, can competently drive a vehicle, both in terms of the user-defined evaluation metric, and as measured by their behavior on several driving situations culled from real-life experience. In this article, we describe a method for multiple agent integration which is applied to the automated highway system domain. However, it also generalizes to many complex robotics tasks where multiple interacting modules must simultaneously be configured without individual module feedback.
Introduction The task of driving can be characterized as consisting of three levels: strategic, tactical and operational [13]. At the highest (strategic) level, a route is planned and goals are determined; at the intermediate (tactical) level, maneuvers are selected to achieve short-term objectives - such as deciding whether to pass a blocking vehicle; and at the lowest (operational) level, these maneuvers are translated into control operations. Mobile robot research has successfully addressed the three levels to different degrees. Strategic-level planners [18, 24] have advanced from research projects to commercial products. The operational level has been investigated for many decades, resulting in systems that range from semi-autonomous vehicle control [7, 11] to autonomous driving in a variety of situations [4, 15]. Substantial progress in autonomous navigation in simulated domains has also been reported in recent years [17, 3, 16]. However, the decisions required at the tactical level are difficult and a general solution remains elusive. Consider the situation depicted in Figure 1. Car A, under computer control, is approaching its desired exit when it comes upon a slow moving blocker (Car B), in its lane. Car A's tactical reasoning system must determine whether to pass Car B and risk missing the exit. Obviously, the correct decision depends on a number of factors such as the distance to the exit, Car A's desired velocity, and the density and speed of surrounding traffic. Such scenarios are of particular relevance to intelligent vehicles operating in mixed-traffic en- vironments. In these environments, computer- and human-controlled cars share the roadway, and tactical decisions must be made without relying on communication-based protocols. This short-term planning problem is challenging because real-time decisions must be made based on incomplete, noisy information about the state of the world. Furthermore, the penalty for bad decisions is severe since errors in judgment may result in high-speed collisions. SAPIENT, described in Section 3, is a tactical reasoning system designed to drive intelligent ve- hicles, such as the Carnegie Mellon Navlab [23], in mixed-traffic environments. In SAPIENT, decisions are made by a collection of independent agents, termed reasoning agents, each of which specializes in a particular aspect of the tactical driving task. This article focuses on how these agents automatically configure themselves to optimize a user- specified evaluation function using a novel evolutionary algorithm termed Population Based Incremental Learning (PBIL). This article is organized as follows. Section 2 presents the simulated highway environment used to train the SAPIENT agents. Section 3 details the SAPIENT architecture, describing the reasoning agents and their voting language. Section 4 introduces PBIL, and explains the encoding scheme used to represent agent parameters. Subsequent sections present our results, both on small-scale tactical scenarios (such as the one shown in Figure 1), and on larger highway configurations. Finally, Section 8 summarizes the research and outlines areas for further research. A GOAL Fig. 1. An example of tactical-level reasoning. Car A is approaching its desired exit behind a slow vehicle B. Should Car A attempt to pass? Agents for Tactical Driving 3 Fig. 2. SHIVA: A design and simulation tool for developing intelligent vehicle algorithms. 2. The SHIVA Simulator Simulation is essential in developing intelligent vehicle systems because testing new algorithms in real traffic is expensive, risky and potentially disastrous. SHIVA 1 (Simulated Highways for Intelligent Vehicle Algorithms) [22, 21] is a kinematic micro-simulation of vehicles moving and interacting on a user-defined stretch of roadway that models the elements of the tactical driving domain most useful to intelligent vehicle designers. The vehicles can be equipped with simulated human drivers as well as sensors and algorithms for automated control. These algorithms direct the vehi- cles' motion through simulated commands to the accelerator, brake, and steering wheel. SHIVA's user interface provides facilities for visualizing and specifying the interactions between vehicles (see Figure 2). The internal structure of the simulator is comprehensively covered in [22], and details of the design tools may be found in [21]. All simulated vehicles are composed of three subsystems: perception, cognition, and control. The perception subsystem consists of a suite of simulated functional sensors (e.g., global positioning systems, range-sensors, lane-trackers), whose outputs are similar to real perception modules implemented on the Navlab vehicles. Simulated vehicles use these sensors to obtain information about the road geometry and surrounding traffic. Vehicles may control the sensors directly, scanning and panning the sensors as needed, encouraging active perception. Some sensors also model occlusion and noise, forcing cognition routines to be realistic in their input assumptions. While a variety of cognition modules have been developed in SHIVA, this article only discusses two types: rule-based reasoning and SAPIENT. The rule-based reasoning system, which was manually designed, is implemented as a monolithic decision tree. It consists of a collection of tactical driving rules such as: "Initiate a left lane change if the vehicle ahead is moving slower than f(v) m/s, and is closer than h(v), and if the lane to your left is marked for legal travel, and if there are 4 Sukthankar, Baluja, Hancock no vehicles in that lane within g(v) meters, and if the desired right-exit is further than where: f(v) is the desired car following velocity, h(v) is the desired car following distance (head- way), g(v) is the required gap size for entering an adjacent lane, and e(x; y; v) is a distance threshold to the exit based on current lane, distance to exit and velocity. While this system performs well on many scenarios, it suffers from four disadvan- tages: 1) as the example above illustrates, realistic rules require the designer to account for many fac- tors; 2) modification of the rules is difficult since a small change in desired behavior can require many non-local modifications; hand-coded rules perform poorly in unanticipated situations; implementing new features requires one to consider an exponential number of interactions with existing rules. Similar problems were reported by Cremer et al. [3] in their monolithic state-machine implementation for scenario control. The SAPIENT distributed architecture, discussed in the next sec- tion, was developed to address some of these problems The control subsystem is compatible with the controller available on the Carnegie Mellon Navlab II robot testbed vehicle. Commands to the controller are issued by the cognition modules at a rate of 10 Hz. 3. SAPIENT SAPIENT (Situation Awareness Planner Implementing Effective Navigation in Traffic) [20] consists of a collection of independent modules (termed reasoning agents), each of which is an expert on a specific aspect of the tactical driving task. Each agent is assigned to monitor a relevant physical entity in the environment and is responsible for assessing the repercussions of that entity on the intelligent vehicle's upcoming actions (see Figure 3). For example, the reasoning agent associated with a vehicle ahead monitors the motion of that vehicle and determines whether to continue car following, initiate a lane change, or begin brak- ing. Similarly, a reasoning agent associated with an upcoming exit is concerned with recommending the lane changes and speed changes needed to successfully maneuver the intelligent vehicle to the off-ramp. 3.1. System Overview The SAPIENT architecture is shown in Figure 4. The perception modules (depicted as ellipses) are connected to the intelligent vehicle's sensors, and perform functions such as lane tracking or vehicle detection. Wherever possible, they correspond to existing systems available on the Carnegie Mellon Navlab (e.g., the lane tracker is based on ALVINN [15]). Each reasoning agent (shown as a dark rectangle) obtains information about the situation from one or two perception modules, and independently calculates the utility of various courses of action. This information is then sent to the voting arbiter, which integrates these recommendations and selects the appropriate response. Finally, the tactical action is translated into steering and velocity commands and executed by the operational-level controller. As seen in Figure 4, reasoning agents can be classified into classes based on their area of spe- cialization. SAPIENT's loosely-coupled architecture allows new classes to be developed without modifying the existing reasoning agents. Our current implementation spans the following tactical- level aspects: ffl Road properties: local geometry, legal lanes, speed limits, etc. ffl Nearby vehicles: sizes, positions, and veloci- ties ffl Exits: distance, exit lane, speed constraints ffl Self-state: current velocity, lateral position, explicit goals Each reasoning agent tracks the associated physical entity's attributes by monitoring the appropriate sensors. For example, a reasoning agent associated with a nearby vehicle normally tracks its longitudinal position and velocity, and its lateral position (mapped into road coordinates). The tracking has two important implications. First, it allows the reasoning agent to obtain a better estimate of the relevant attribute. Second, the reasoning agent can accumulate statistics that can help influence decisions. For instance, based on the ir- Agents for Tactical Driving 5 A Front vehicle tracker region of interest. Exit Finder Velocity Preference Reasoning Agent Exit Reasoning Agent Lane Reasoning Agent Vehicle Reasoning Agent Fig. 3. SAPIENT reasoning agents are associated with relevant physical entities in the environment. In this situation, the intelligent vehicle (A) is following a car and approaching its desired exit. regular lane-keeping performance of a nearby vehicle (an indication of an inexperienced or intoxicated driver), the reasoning agent associated with that vehicle could favor actions that maintain a greater distance from that vehicle. Thus, SAPIENT is not a purely reactive system; the local state associated with each reasoning agent allows SAPIENT to make decisions based on past his- tory. The relevant history is maintained by each agent. Externally, all reasoning agents share a similar structure - each agent accepts inputs from a subset of the intelligent vehicle's perception modules and sends outputs to the voting arbiter as a set of votes over the entire action space (See Section 3.2). Internally, however, SAPIENT's reasoning agents are heterogenous, maintaining local state and using those representations that are most applicable to the assigned subtask. For ex- ample, the reasoning agents responsible for exit management are rule-based while the reasoning agent monitoring other vehicles use generalized potential fields [9, 10]. The different reasoning agent types and their associated algorithms are detailed in [20]. reasoning agents are myopic in their outlook. For example, the Exit Reasoning Ob- ject's votes are not influenced by the presence of the blocking vehicle; conversely, the reasoning agent associated with the blocking vehicle is oblivious to the exit. Finally, the arbiter is completely ignorant of the driving task. Yet the combination of these local reasoning schemes leads to a distributed awareness of the tactical-level situation. Before discussing how a knowledge-free arbiter can combine these local views of the tactical driving task, a closer look at the action space is warranted. 3.2. Actions Tactical maneuvers (such as lane changing) are composed by concatenating several basic actions. Reasoning agents indicate their preference for a basic action by assigning a vote to that action. The magnitude of the vote corresponds to the intensity of the preference and its sign indicates approval or disapproval. Each reasoning agent must assign some vote to every action in the action space. All actions have velocity (longitudinal) and lane-offset (lateral) components; for exam- ple, "brake hard while changing left" or "increase speed and maintain your current lane position". Since different reasoning agents can return different recommendations for the next action, conflicts must be resolved. SAPIENT uses a voting arbiter to perform this integration. At each time- step, the reasoning agents synchronously submit votes or vetoes for each action in the action space (see Table 1). During arbitration, all of the votes for a given action are summed together (after 6 Sukthankar, Baluja, Hancock Operational Controller Lane Tracker Exit Finder Car Detection Modules Perception Cognition Control Voting Arbiter Velocity Agent Lane Agent Exit Agent Front Left Car Agent Front Right Car Agent Back Right Car Agent Back Left Car Agent Hysteresis Agent Front Car Agent Fig. 4. SAPIENT consists of a collection of reasoning agents which recommend actions based upon local considerations. Each reasoning agent monitors a subset of the vehicle's sensors and votes upon a set of possible actions. The hysteresis reasoning agent is responsible for maintaining consistency over time (especially in cases where multiple actions are equally advantageous); this is done by voting in favor of the action selected in the previous time step. Action fusion is performed by a domain-independent, voting arbiter. Table 1. The action space is a 3 \Theta 3 discretization of the lateral/longitudinal space. The labels are translated at the operational level into specific numbers. Thus, "left" and "right" map to lateral positions (e.g., move left/right by 0.1 lane while "accelerate" and "decelerate" map to changes in velocity (e.g., speed up/slow down by 0.1 m/s). accelerate/shift-left accelerate/straight accelerate/shift-right coast/shift-left coast/straight coast/shift-right decelerate/shift-left decelerate/straight decelerate/shift-right being scaled by the reasoning agent's influence weight), and the action with the most accumulated votes (which has not been vetoed by any agent) is executed. The actions used in the implementation described in this article are summarized in Table 1. Finer discretizations and alternate action spaces are discussed in [20]. Although the action space restricts reasoning objects to voting on their adjacent lanes, the reasoning agents can internally plan longer-range courses of action. For example, the exit agent can vote for lane changes towards the exit, even when the exit is several lanes away. Agents for Tactical Driving 7 car following (external) exit weight (external) bits parameters * 3 { { desire to exit (internal) ={ car following (internal) 011 . 010100101110 . 101 Fig. 5. The three-bit encoding scheme used to represent parameters in the search space: internal parameters are linearly scaled while external ones are exponentially scaled. 3.3. Parameters Different reasoning agents use different internal al- gorithms. Each reasoning agent's output depends on a variety of internal parameters (e.g., thresh- olds, gains, etc. Before going to the arbiter, each agent's outputs are scaled by its influence weight (external parameters). When a new reasoning agent is being imple- mented, it is difficult to determine whether a ve- hicle's poor performance should be attributed to a bad choice of parameters in the new agent, a bug in the logic of the new reasoning agent or, more seriously, to a poor representation scheme (inadequate configuration of reasoning agents). To overcome this difficulty, we have implemented a method for automatically configuring the parameter space. A total of twenty parameters, both internal and external, were selected for the tests described here, and each parameter was discretized into eight values (represented as a three-bit string). For internal parameters, whose values are expected to remain within a certain small range, we selected a linear mapping (where the three bit string represented integers from 0 to 7); for the external parameters, we used an exponential representation (with the three-bit string mapping to eight values in the range 0 to 128). The latter representation increases the range of possible weights at the cost of sacrificing resolution at the higher magnitudes. A representation with more bits per parameter would allow finer tuning but increase the training time. The encoding is illustrated in Figure 5. In the next section, we describe the evolutionary algorithm used for the learning task. 4. Population-Based Incremental Learning Population-Based Incremental Learning (PBIL) is a combination of genetic algorithms (GAs) [8] and competitive learning [1, 2]. The PBIL algorithm attempts to explicitly maintain statistics about the search space and uses them to direct its ex- ploration. The object of the algorithm is to create a real valued probability vector which, when sampled, reveals high quality solution vectors with high probability. The full algorithm is presented in Figure 6. Initially, each element of the PBIL probability vector is initialized to 0.5. Sampling from this vector yields random solution vectors because zeros and ones are generated with equal probability in each bit position. As training progresses, the values in the probability vector gradually shift to represent high evaluation solution vectors through the following process. A number of solution vectors are generated based upon the probabilities specified in the probability vector. The probability vector is pushed towards the generated solution vector with the highest evaluation. After the probability vector is updated, a new set of solution vectors is produced by sampling from the updated probability vector, and the cycle is continued. As the search progresses, the entries in the probability vector move away from their initial settings of 0.5 towards either 0.0 or 1.0. The best solution ever generated in the run is returned as the final solution. Note that because the algorithm only returns the best solution generated during the run, convergence of the probability vector is not a prerequisite for the success of the algorithm. 8 Sukthankar, Baluja, Hancock for to LENGTH do while (NOT termination condition) * Generate Samples * for to SAMPLES do best_vector := find_vector_with_best_evaluation( sample_vectors, evaluations ); * Update Probability Vector towards best solution * for to LENGTH do * Mutate Probability Vector * for to LENGTH do if (random (0,1) < MUT_PROBABILITY) then if (random (0,1) > 0.5) then mutate_direction := else mutate_direction := 0; mutate_direction * (MUT_SHIFT); return the best solution found in run; USER DEFINED CONSTANTS (Values Used in this Study): SAMPLES: the number of vectors generated before update of the probability vector (100) LR: the learning rate, how fast to exploit the search performed (0.1). LENGTH: the number of bits in a generated vector (3 * 20) MUT_PROBABILITY: the probability of a mutation occuring in each position (0.02). MUT_SHIFT: the amount a mutation alters the value in the bit position (0.05). Fig. 6. The PBIL algorithm used to train SAPIENT reasoning agent parameters. Here, the explicit preservation of the best solution from the previous generation (elitist selection) is not shown. However, empirically, the probability vector has converged in all of the runs conducted. The probabilistic generation of solution vectors does not guarantee the creation of a good solution vector in every iteration. This problem is exacerbated by the small population sizes used in our experiments. Therefore, in order to avoid moving towards unproductive areas of the search space, the best vector from the previous population is included in the current population (by replacing the worst member of the current population) - in GA literature, this is termed elitist selection [8]. Since space limitations preclude a complete discussion about the relationship between GAs and PBIL, we can only provide a brief intuition. Diversity in the population is crucial for GAs. By maintaining a population of solutions, the GA is able - in theory at least - to maintain samples in many different regions. In genetic algorithms, crossover is used to merge these different solu- tions. However, when the population converges, crossover is deprived of the diversity that it needs to be an effective search operator. When this hap- pens, crossover begins to behave like a mutation operator that is sensitive to the convergence of the value of each bit. If all individuals in the population converge at some bit position, crossover leaves those bits unaltered. At bit positions where individuals have not converged, crossover will effectively mutate values in those positions. Therefore, crossover creates new individuals that differ from the individuals it combines only at the bit positions where the mated individuals disagree. This is analogous to PBIL which creates new trial solutions that differ mainly in bit positions where prior good performers have disagreed. More details can be found in [1]. Our application challenges PBIL in a number of ways. First, since a vehicle's decisions depend on the behavior of other vehicles which are not under Agents for Tactical Driving 9 its control, each simulation can produce a different evaluation for the same bit string. We evaluate each set of vehicle parameters multiple times to compensate for the stochastic nature of the envi- ronment. Second, the PBIL algorithm is never exposed to all possible traffic situations (thus making it impossible to estimate the "true" performance of a PBIL string). Third, since each evaluation takes considerable time to simulate, minimizing the total number of training evaluations is important. 5. Training Specifics All of the tests described below were performed on the track shown in Figure 8, known as the SHIVA cyclotron. While this configuration does not resemble a real highway, it has several benefits as a testbed: 1) It is topologically identical to a highway with equally spaced exits; 2) Taking the nth exit is equivalent to traveling n laps of the course; One can create challenging traffic interactions at the entry and exit merges with only a small number of vehicles. During training, each scenario was initialized with one SAPIENT/PBIL vehicle, and eight rule-based cars (with hand-crafted decision trees). The SAPIENT car was directed to take the second exit while the other cars had goals of zero to five laps. Whenever the total number of vehicles on the track dropped below nine, a new vehicle was injected at the entry ramp to maintain the desired traffic density. Only one SAPIENT vehicle was permitted on the course at a time. At the start of the run, the PBIL algorithm suggested a candidate bit-string which was converted into SAPIENT parameters, and instantiated as a simulated vehicle. Each evaluation of a PBIL parameter string required one run of a simulated ve- hicle. At the end of the vehicle's run, the score that it received was sent to PBIL as the evaluation of that candidate bit-string. It should be noted that the population size in PBIL affected the number of evaluations required in each generation of the PBIL algorithm. The population size does not correspond to the number of SAPIENT vehicles present on the track since each candidate vehicle was independently evaluated (as stated earlier, only one SAPIENT vehicle was permitted on the track at a time). Whenever a SAPIENT vehicle left the scenario (upon taking an exit, or crashing 10 times), its evaluation was computed based on statistics collected during its run. This score was used by the PBIL algorithm to update the probability vector thus creating better SAPIENT agents in the next generation. While the definition of good driving is largely subjective, the following characteristics are strongly correlated with bad driving: 1) collisions; taking the wrong exit; 3) deviating from the desired speed; weaving (poor lane tracking). Many possible evaluation functions could be constructed from these characteristics. For our evaluation function, we combined them in a simple weighted sum, to be maximized: \Gamma(1000 \Theta num-crashes) \Gamma(500 \Theta if-wrong-exit) \Gamma(0:02 \Theta speed-deviation) \Gamma(0:02 \Theta lane-deviation) +(dist-traveled) where: all-veto indicates that the SAPIENT vehicle objects to all actions (with good param- eters, this should never happen); num-crashes is the number of collisions involving the SAPIENT vehicle; if-wrong-exit is a flag - true if and only if the SAPIENT vehicle exited prema- turely, or otherwise missed its designated exit; speed-deviation is the difference between desired and actual velocities, integrated over the entire run; lane-deviation is the deviation from the center of a lane, integrated over the entire run; dist-traveled is the longitudinal distance covered by the vehicle, in meters (an incremental reward for partial completion) While the evaluation function is a reasonable measure of performance, it is important to note that there can be cases when a "good" driver becomes involved in unavoidable accidents; con- versely, favorable circumstances may enable "bad" vehicles to score well on an easy scenario. To minimize the effects of such cases, we tested each candidate string in the population on a set of four sce- narios. In addition to traffic, these test cases in- cluded some pathological situations with broken- vehicles obstructing one or more lanes. 6. Training We performed a series of experiments using a variety of PBIL population sizes, evaluation functions and initial conditions. More details about individual experiments are presented in the next section. This section focuses on evaluation metrics for the training algorithm. Figure 7 shows the results of a training run with the evaluation function described earlier, and a PBIL population size of 100. These 3-D histograms display the distribution of scores for each generation. It is clear that as the parameters evolve in successive generations, the average performance of SAPIENT vehicles increases, and the variance of evaluations within a generation de- creases. In the experiments with population size 100, good performance of some vehicles in the population is achieved early (by the fifth generation) although consistently good evaluations are not observed until generation 15. The number of vehicles scoring poor evaluations drops rapidly until generation 10, after which there are only occasional low scores. The PBIL strings converge to a stable set of SAPIENT parameters, and by the last gen- eration, the majority of the vehicles are able to take the proper exit, and avoid crashes in all sce- narios. The results of experiments with different population sizes were similar. Figure 8 shows a scenario on the cyclotron track. This scenario is pathological, in that it contains many broken-down vehicles, scattered over the roadway. The trace shows a trained SAPIENT vehicle successfully navigating the course by avoiding the obstacles. Above, we described the overall performance of the SAPIENT vehicles in terms of a global evaluation function. Here, we examine how the individual components of the scoring metric improve over time. Three observable quantities play a significant role in the SAPIENT training evaluation function: , the total number of near-collisions; fi, whether the vehicle made its exit; and, i , the distance traveled by the intelligent vehicle in the sce- nario. Thus, for a given population of SAPIENT vehicles, the quantities: vehicles, v, in the popula- tion), reflect the "goodness" of the population. The three graphs in Figure 9 show how K, B, and Z change over successive generations. Each PBIL population contains 40 vehicles, and each vehicle is evaluated on four different scenarios. The graphs show that: ffl The number of near-collisions, K, drops steadily as PBIL tunes the SAPIENT reasoning agent parameters. In the final generation, none of the vehicles in the population are involved in any near-collisions over the entire set of four scenarios. ffl The fraction of vehicles in the population which missed their exit also decreases steadily over time as the SAPIENT vehicles learn. This too is zero in the final generation. ffl The third quantity, Z, reflects the incremental improvement in performance of the vehicles during training. It can be seen that the early vehicles are eliminated from the scenario (ei- ther through timeout, taking the wrong exit, or crashing) before they travel very far. Over time, the vehicles are able to travel greater distances. Note that Z has an upper bound cannot be greater than the distance to the desired exit). To investigate the robustness of this training method, two additional sets of experiments were performed, where the coefficients in the evaluation function were perturbed. In the first set, six experiments were conducted, and in each experi- ment, one coefficient was multiplied by 10. Some of these results are shown in Figures 10 and 11. Somewhat surprisingly, the SAPIENT reasoning agents generated from these perturbed evaluation functions are still successful. We hypothesize two reasons for this: 1) the tactical-level sub-tasks are closely linked: it is quite likely that a vehicle which makes the correct exit has also learned to avoid collisions - otherwise it would have been eliminated in a collision earlier on the track; 2) although PBIL is responsible for setting the internal and external parameters for each reasoning agent, the underlying algorithms are predefined; thus, a small perturbation in reasoning agent parameters does not cause catastrophic failures in the system. Agents for Tactical Driving 11 -5000 Evaluation1030Generation2060Number of Cars -5000 Evaluation Fig. 7. 3-D Histogram showing increase of high-scoring PBIL strings over successive generations. Population size is 100 cars in each generation. Fig. 8. A pathological scenario on the cyclotron track with 15 obstacles. The trace shows a SAPIENT vehicle successfully navigating the course by avoiding the obstacles. The second set of experiments explored the limits of this robustness. This time, the coefficients were perturbed by a factor of 1000. While this tended to create bad cars in general, some of the coefficients, even when multiplied by 1000, still of near-crashes Generations of exits missed Generations B60000100000140000180000 Distance covered Generations Z Fig. 9. This graph shows how the number of near-collisions (K), number of missed-exits (B), and distance traveled (Z) in a population of learning SAPIENT vehicles varies with successive generations. In these tests, the population size was set to 40, and each vehicle was evaluated on four scenarios. The graphs show the accumulated statistics for all of the vehicles in the given generation, over all four scenarios. Note that both K and Z decrease to zero, while Z, the incremental reward, rises. Agents for Tactical Driving 13501502500 of near-crashes Generations of near-crashes Generations of near-crashes Generations of near-crashes Generations Fig. 10. This graph shows that the SAPIENT parameters learned by PBIL converge to competent vehicles despite variations in the evaluation function used. Each of these graphs shows the total number of near-collisions (K), in a population of exits missed Generations of exits missed Generations of exits missed Generations of exits missed Generations Fig. 11. This graph shows that the SAPIENT parameters learned by PBIL converge to competent vehicles despite variations in the evaluation function used. Each of these graphs shows the total number of missed exits (B), in a population of Agents for Tactical Driving 15 A Fig. 12. This scenario tests if the tactical reasoning system can overtake a slower-moving vehicle. Lateral position (lane units) Time (0.1 sec intervals) Mono Velocity (m/sec) Time (0.1 sec intervals) Mono Poly Fig. 13. Lateral displacement (left) and velocity (right) as a function of time, for rule-based (denoted as Mono) and SAPIENT (denoted as Poly) vehicles on the overtaking scenario (See Figure 12). See the text for a discussion of these graphs. generated competent vehicles. For example, increasing the penalty of a collision from -1000 to -1000000 does not affect vehicles since they learn how to avoid all collisions. By contrast, radically increasing the penalty for speed deviations in a similar manner leads to vehicles that are willing to collide with others in a desperate effort to avoid the penalties incurred in dropping below the target velocity. 7. Scenario-Based Evaluation of Tactical Driving Scenarios are widely used in driving research to evaluate the performance of human subjects [12, 14]. Similar techniques are also used to measure situation awareness in other domains [6, 5, 19]. Here, we use micro-scenarios to examine the performance of SAPIENT's reasoning agents in situations where tactical-level decisions are required. A more comprehensive discussion of these scenarios is available in [20]. In each of the following scenarios, we focus on the vehicle marked A in the respective dia- grams. SAPIENT's performance is compared to the behavior of the default rule-based vehicle. In the accompanying graphs, the monolithic, hand- coded, rule-based vehicle is denoted as Mono, while the multi-agent, adaptive, SAPIENT system is marked Poly. It should be emphasized that the SAPIENT vehicles have not been exposed to any tactical scenarios - they were trained (us- ing PBIL) exclusively on obstacle courses in the cyclotron environment. The first scenario (See Figure 12) involves a simple overtaking maneuver, and is a common occurrence on the highway. Initially, both vehicles are moving at normal highway speeds, but the lead vehicle begins braking (as it approaches its exit, for example). There is no other traffic, so Car A should safely overtake. As seen in the lateral displacement and velocity profiles (See Figure 13), both types of cognition module are able to solve A L32.03.04.0Fig. 14. Exit scenarios add complexity to the tactical driving domain because they introduce additional strategic-level goals. The conflicts between two strategic-level goals leads to interesting tactical Lateral position (lane units) Time (0.1 sec intervals) Mono Velocity (m/sec) Time (0.1 sec intervals) Mono Poly Fig. 15. Lateral displacement (left) and velocity (right) as a function of time, for rule-based and SAPIENT vehicles on the exit scenario (See Figure 14). See the text for a discussion of these graphs. this scenario successfully. However, note that the SAPIENT vehicle is more aggressive, maintaining a smaller headway during the maneuver than the hand-tuned, rule-based vehicle. This is because the SAPIENT reasoning agent responsible for car following has tuned its generalized potential fields relative to the other vehicle based on a time-to-impact metric, as opposed to using a constant headway. The other notable feature is the oscillation in the rule-based vehicle's velocity profile. This is caused by a combination of two factors: a discrete velocity controller and brittle car-following rules. Note that the SAPIENT vehicle is not perfectly centered in the passing lane during the overtaking maneuver. This is because the potential field surrounding the obstacle votes for additional space, and since there is sufficient space in the target lane, the SAPIENT vehicle is able to drive off-center. This behavior can also be observed in the other scenarios. The second scenario (See Figure 14) introduces a second (possibly conflicting) strategic goal: taking an exit; also, ambient traffic is introduced. Vehicle A must now change lanes to make its desired exit without colliding with other cars. Figure 15 shows an interesting difference in driving behav- ior. The rule-based car slows down until it can find a gap in the exit lane, and then changes lanes. In contrast, SAPIENT speeds up to overtake the vehicle in the exit lane. This maneuver allows it to maintain its desired speed while making the exit. The final scenario, shown in Figure 16, is identical to the one discussed in the Introduction. Recall that Car A may take its desired exit by either staying behind the slow blocker, or by passing. Unlike the situation shown in Figure 14, chang- Agents for Tactical Driving 17 A GOAL L22.03.04.0Fig. 16. This exit scenario is difficult because the lane changes are optional. To address the strategic-level goal of maintaining speed, the intelligent vehicle must decide whether or not to attempt the overtaking maneuver at the risk of missing the desired exit.1.622.42.83.2 Lateral position (lane units) Time (0.1 sec intervals) Mono Velocity (m/sec) Time (0.1 sec intervals) Mono Poly Fig. 17. Lateral displacement (left) and velocity (right) as a function of time, for rule-based and SAPIENT vehicles on the more difficult exit scenario (See Figure 16). ing lanes is not mandatory; in fact, should Car A decide to pass, it will have to complete two lane changes before exiting. Once again, the two different vehicle types choose differently. The rule-based vehicle opts to stay in its lane, based solely on a rule which depends on the distance to the exit. On the other hand, the SAPIENT vehicle chooses to overtake the blocker. In the final set of experiments, vehicles were injected into an initially empty cyclotron track from the on-ramp at regular intervals, . Each vehicle was given two strategic-level goals: 1) make exactly one circuit of the track before exiting; 2) maintain the speed at which it was injected whenever possible. The aim of the experiment was to see how the two tactical driving systems, rule- based, and SAPIENT, would behave as the roadway became more congested. Three sets of experiments with different traffic configurations were performed: all-rule-based cars, all-SAPIENT cars, and a uniform mix of rule-based and SAPIENT cars. As expected, the number of vehicles on the roadway increased until the rate of vehicles entering the track was equal to the rate of vehicles leaving (either because the vehicles had successfully completed their circuit, or because the vehicles were unable to merge into the traffic stream). At low rates of traffic flow (e.g., ? 6 seconds), all of the three traffic configurations safely negotiated the scenario (with no missed exits). However, once the traffic flow was increased, the behavior of the three traffic configurations diverged. of cars on track Time (secs) Mono Poly Fig. 18. This graph shows how the number of vehicles on the cyclotron varies as a function of time in heavy traffic (traffic injection rate cars). Note that only six rule-based vehicles were able to merge onto the cyclotron loop; by contrast, all of the SAPIENT vehicles were able to merge and complete the scenario. The graph in Figure shows how the number of vehicles varies as a function of time when seconds for the all-rule-based and all-SAPIENT cases. Even in heavy traffic, neither of the pure traffic types have any collisions. Although both types of vehicles perform well initially (when the roadway is clear), once the number of vehicles on the track increases to about 6, the conservative rule-based drivers are unable to merge into the traffic stream (since they require a guaranteed headway of two seconds on both sides of the gap). Thus they are unable to change lanes, and exit the scenario prematurely. To make matters worse, the rule-based vehicles that were already on the roadway become trapped in the inner loops of the cyclotron (due to the high rate of traffic in the entry/exit lane). The all-SAPIENT traffic, on the other hand, is able to drive successfully. This can probably be attributed to two factors: 1) the aggressive driving style, relying on time-to-impact reasoning agents, is willing to merge into smaller gaps; 2) the distributed reasoning system is better at making tradeoffs - the negative votes for merging into a potentially unsafe gap are tolerated since the alternative (missing the exit) is seen to be worse. The brittle decision tree used in the rule-based cars, on the other hand, rejects these gaps outright Interleaving rule-based and SAPIENT cars in the heavy traffic scenario leads to a stable heterogenous behavior with no collisions. While the more aggressive SAPIENT vehicles still miss fewer exits, even the rule-based vehicles perform better than they did in the pure-rule-based case because of the reduced congestion. This may have positive implications for the deployment of automated vehicles in mixed traffic conditions. 8. Conclusion and Future Directions Our experiments have demonstrated: 1) The potential for intelligent behavior in the tactical driving domain using a set of distributed reasoning agents; 2) The ability of evolutionary algorithms to automatically configure a collection of these modules for addressing their combined task. While the evaluation sections compared SAPI- ENT's performance with a rule-based vehicle, the results should not be taken out of context: clearly it is possible to encode SAPIENT's current knowledge in the form of rules to create a more competent rule-based vehicle. The difference is that Agents for Tactical Driving 19 creating a monolithic rule-based vehicle is a much more difficult task due to the interactions between large number of rules, the manual tuning of parameters within the rules, and the complex interactions between the rules. In this study, we used a simple evaluation func- tion. By introducing alternative objective func- tions, we plan to extend this study in at least two directions. First, for automated highways, we would like the cars to exhibit altruistic be- havior. In a collection of PBIL vehicles, optimizing a shared evaluation function (such as highway may encourage cooperation. Second, we are developing reasoning agents to address additional complications which will arise when these vehicles are deployed in the real world, such as complex vehicle dynamics and noisy sensors. Our system, which employs multiple automatically trained agents, can competently drive a ve- hicle, both in terms of the user-defined evaluation metric, and as measured by their behavior on several driving situations culled from real-life expe- rience. In this article, we described a method for multiple agent integration which is applied to the automated highway system domain. However, it also generalizes to many complex robotics tasks where multiple interacting modules must simultaneously be configured without individual module feedback. 9. Acknowledgments The authors would like to acknowledge the valuable discussions with Dean Pomerleau and Chuck Thorpe which helped to shape this work. Thanks also to Gita Sukthankar for the data processing scripts and graphs. This research was partially supported by the Automated Highway System project, under agreement DTFH61-94-X-00001, and was started while Shumeet Baluja was supported by a graduate student fellowship from NASA, administered by the Lyndon B. Johnson Space Center. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the AHS Consortium or NASA. Notes 1. More information and an interactive demo are available at: !http://www.cs.cmu.edu/rahuls/shiva.html? --R Removing the genetics from the standard genetic algorithm. The software architecture for scenario control in the Iowa driving simulator. A curvature-based scheme for improving road vehicle guidance by computer vision Towards a theory of situation awareness. Automatic car controls for electronic highways. Genetic Algorithms in Search A generalized potential field approach to obstacle avoidance control. Integrated path planning and dynamic steering control for autonomous vehi- cles Driver education and task analysis Volume A critical view of driver behavior models: What do we know Coaching the experienced driver II Neural Network Perception for Mobile Robot Guidance. Using genetic algorithms to learn reactive control parameters for autonomous robotic navigation. Selective Perception for Robot Driving. Advanced driver information systems. How in the world did we ever get into that mode? Situation Awareness for Tactical Driving. Also available as CMU Tech Report CMU- RI-TR-97-08 A simulation and design system for tactical driving algorithms. SHIVA: Simulated highways for intelligent vehicle al- gorithms Vision and navigation for the Carnegie Mellon Navlab. Dynamic route guidance and interactive transport management with ALI-Scout --TR --CTR Antonio Pellecchia , Christian Igel , Johann Edelbrunner , Gregor Schoner, Making Driver Modeling Attractive, IEEE Intelligent Systems, v.20 n.2, p.8-12, March 2005
distributed AI;evolutionary algorithms;intelligent vehicles;simulation
590846
Incremental Feature Selection.
Feature selection is a problem of finding relevant features. When the number of features of a dataset is large and its number of patterns is huge, an effective method of feature selection can help in dimensionality reduction. An incremental probabilistic algorithm is designed and implemented as an alternative to the exhaustive and heuristic approaches. Theoretical analysis is given to support the idea of the probabilistic algorithm in finding an optimal or near-optimal subset of features. Experimental results suggest that (1) the probabilistic algorithm is effective in obtaining optimal/suboptimal feature subsets; (2) its incremental version expedites feature selection further when the number of patterns is large and can scale up without sacrificing the quality of selected features.
Introduction Feature selection is about finding useful (relevant) features that describe an application domain. The problem of feature selection can formally be defined as selecting a minimum set of M relevant features from N original features where M N such that the probability distribution of different classes given the values for these M features is as close as possible to the original distribution given the values for N features. Mathematically, if FN is the original feature set and FM is the chosen feature subset, then the conditional probability, P(C j should be as close as possible to P(C j possible classes, C. Here f M and f N are value vectors of respective feature vectors FM and FN [11]. As the dimensionality of a domain expands, the number of features increases. In general, the role of feature selection is three-fold: 1. simplifying data descrip- tion; 2. reducing the task of data collection; and 3. improving the quality of problem solving. For the same problem, a representation by three features is generally simpler than one by six features. The benefits of having a simple representation are abundant such as easier understanding of problems, and better and faster problem solving. In the context of data collection, having more features means that more data should be collected. In many applications, this could be time consuming and costly. The quality improvement of problem solving resulting from feature selection can be illustrated via a classical supervised learning task - pattern classification problem: given a training set of labeled patterns, induce a classification model that can predict the class label for a set of previously unseen patterns (the so-called testing set). Although having more features enhances discriminating power in representation, having excessive features would introduce many difficulties for induction algorithms [11]. First, the time required by an induction algorithm often grows dramatically with the number of features, rendering the algorithm impractical for problems with a large number of features. Second, many learning algorithms can be viewed as performing estimation of the probability of the class label given a set of fea- tures. With many features, this distribution is of high dimension and becomes very complex. Unless exponential amounts of data are available, it is difficult to obtain a good estimation from a training dataset. Third, irrelevant and redundant features may confuse a learning algorithm by obscuring the distribution of the small set of truly relevant features. In addition, irrelevant and redundant features require an exponential increase in data storage requirements [1]. This is because with more features, much more data is required for effective induction. For instance, in a binary domain, the extra m irrelevant/relevant features would times more patterns to describe the whole data. For an induction algorithm, the reduced features can also result in a simpler induction model such as shorter and fewer classification rules. For N features, if d of them are relevant, an exhaustive approach to finding the optimal d features would require examining subsets. The number of possible subsets grows exponentially. Researchers have designed different strategies in search of optimal subsets of d features (Branch and Bound [20] and its variations [26], many heuristic and stochastic methods [5, 7]). If we view these feature selection algorithms from the perspective of using an induction algorithm, as pointed out in [8], the work on feature selection can be divided into filter and wrapper models. In a filter model, a feature selector is independent of an induction algorithm and serves as a filter to sieve the irrelevant and/or redundant features; in a wrapper model, a feature selector wraps around an inductive algorithm relying on which relevant features are determined. One problem with the wrapper model is that it is restricted by the time complexity of a learning algorithm [12]. This time complexity is dependent on the number of features. Often the wrapper methods are prohibitively expensive to run and can be intractable for a very large number of features. Recall that, in many cases, feature selection is performed because of the excessive number of features and because a favorite induction algorithm has difficulties in handling so many features. Different models, however, suit for different applications. If a classifier is chosen and it can run for an application at hand, then it may be wise to choose a wrapper model since both feature selection and classifier induction use the same bias. This work considers the cases in which learning a classifier becomes cumbersome or ineffective due to the large size of a dataset. The largeness can be defined by both the number of features (N ) and the number of patterns (P ). It is the latter that makes some induction algorithms falter. Hence, large datasets in terms of P are the main concern. Naturally, the filter model is adopted. Our aim is to provide a simple and practical method that can select features for large datasets. In the following, we first review related work on feature selection. Related Work The problem of feature selection has long been an active research topic within statistics and pattern recognition [30, 6, 7], but most work in this area has dealt with linear regression [12] and is under assumptions that do not apply to most machine learning algorithms [8]. Researchers [12, 8] pointed out that the most common assumption is monotonicity that increasing the number of features can only improve the performance of a learning algorithm 1 . Recently feature selection has received considerable attention from researchers in machine learning and knowledge discovery who are interested in improving the performance of algorithms and in cleaning data. In handling large databases, feature selection is even more important since many learning algorithms may falter or take too long time to run before data is reduced. Most feature selection methods [9, 12, 8] can be grouped into two categories: exhaustive or heuristic search for an optimal set of M features. For example, Almuallim and Dietterich's FOCUS algorithm [2] starts with an empty feature set and carries out exhaustive search until it finds a minimal set of features that is sufficient to construct a hypothesis consistent with a given set of examples. It works on binary, noise-free data. Its time complexity is O(min(N 1 The monotonicity assumption is not valid for many induction algorithms used in machine learning. An example is dataset 1 (CorrAL) in Section 5 which is reproduced from [8]. They proposed three heuristic algorithms to speed up the searching [2]. This is because selecting a minimal subset is a known intractable problem, and in practice, we often have to trade off the optimality of a solution for less time spent on searching. There are many heuristic feature selection algorithms. The Relief algorithm [9] assigns a "relevance"weight to each feature, which is meant to denote the relevance of a feature to the target concept. Relief samples patterns randomly from the training set and updates the relevance values based on the difference between the selected pattern and the two nearest patterns of the same and opposite classes. According to [9], Relief assumes two-class classification problems and does not help with redundant features. If most of the given features are relevant to the concept (including redundant features), it would select most of them even though only a fraction of them is necessary for concept description. The PRESET algorithm [19] is another heuristic feature selector that uses the theory of Rough Sets to rank the features heuristically, assuming a noise-free binary domain. In order to consider higher order relations among the features, Liu and Wen [16] suggest the use of high order information gains to select features. Since the last two algorithms do not try to explore all the combinations of features, it is certain that they fail on problems whose features are highly interdependent such as the parity problem where combining a small number of features does not help in finding the relevant ones. Another common understanding is that some learning algorithms have built-in feature selection, for example, ID3 [23], FRINGE [21] and C4.5 [24]. The results in [2] suggest that one should not rely on ID3 or FRINGE to filter out irrelevant features. A more detailed survey can be found in [5]. The latest development of feature selection in pattern recognition can be found in [7]. To sum up, the exhaustive search approach is infeasible in practice; the heuristic search approach can reduce the search time significantly, but will fail on hard problems (e.g., the parity problem) or cannot remove redundant features. A probabilistic approach is proposed as an alternative [15] in selecting the op- timal/suboptimal subset(s) of features. In the context of large sized databases, however, it would still take considerably long time to check if a subset is valid or not 2 . We had first-hand experience of this problem when our probabilistic system was dispatched to a local institute for on-site usage. All the evidence showed that reducing data size can significantly speed up the selection of features (see a case study in Section 3.3). Hence, the incremental probabilistic method is designed and implemented. In the following, we describe the probabilistic method first, then the incremental one, followed by an empirical study in which the effectiveness of the algorithms is verified. At the end of the paper, we provide relevant discussion. 2 The checking can be done in O(P ), where P is the number of patterns, by using a hashing method. 3 Probabilistic Feature Selection The proposed probabilistic approach is a Las Vegas Algorithm [4]. Las Vegas algorithms make probabilistic choices to help guide them more quickly to a correct solution. One kind of Las Vegas algorithms uses randomness to guide their search in such a way that a correct solution is guaranteed even if unfortunate choices are made. As we mentioned earlier, heuristic search methods are vulnerable to datasets with high interdependency among their features. Las Vegas algorithms free us from worrying about such situations by evening out the time required on different situations. Another similar type of algorithms is Monte Carlo algorithms in which it is often possible to reduce the error probability arbitrarily at the cost of a slight increase in computing time (refer to page 341 in [4]). In this work, LVF (Las Vegas Filter) 3 is more suitable since probabilities of generating distinct subsets are the same. The time performance of a Las Vegas algorithm may not be better than that of some heuristic algorithms. LVF algorithm Input: MAX-TRIES, fl - allowed inconsistency rate, dataset of N features; Output: sets of M features satisfying the inconsistency criterion best for i=1 to MAX-TRIES best and InconCheck(S; D) ! fl) else if best ) and printCurrentBest(S) end for The LVF algorithm generates a random subset, S, from N features in every round. If the number of features (C) of S is less than the current best, i.e., best , the data D with the features prescribed in S is checked against the inconsistency criterion. If its inconsistency rate (defined later) is below a pre-specified one (fl), C best and S best are replaced by C and S respectively; the new current best (S) is printed. If best and the inconsistency criterion is satisfied, then an equally good current best is found and printed. MAX TRIES in the algorithm is used to control the number of loops. A value of MAX TRIES can be defined according to applications or based on the experience from exper- imentation. Too small or too big a MAX TRIES will affect the performance of LVF. The compromise is made for good and fast solutions. The longer LVF runs, the better its results are. Refer to the analysis in Section 3.2. MAX TRIES is set to 77\ThetaN in our experimental study 4 following the rule-of-thumb that the Its counterpart is LVW - a wrapper feature selector applying a Las Vegas algorithm. 4 We tried first a constant c alone instead of c \Theta N , then linked it to N . 77 was chosen for c more features a dataset has (in other words, the larger N is), the harder the problem of feature selection (parity-5 is more difficult than parity-2, e.g.), and hence more tries are needed. When LVF loops MAX TRIES times, it stops. An alternative to this stopping criterion is to let LVF run forever to take full advantage of its "anytime algorithm" nature (more in Section 6). The function randomSet(seed) returns a set of features randomly. When the seed is changed dynamically, a different set is generated. The function numOfFeatures(S) returns the cardinality of set S. InconCheck(S; D) returns the inconsistency rate of data D with selected features specified in S. printCurrentBest(S) prints out subset S. A more sophisticated version of LVF is like this: since we know the cardinality of a better subset can only be smaller than C best - the cardinality of the current best subset, we just need to randomly generate subsets whose cardinalities are smaller than C best . For a new round of selection, we sample features without replacement. 3.1 Measure of feature goodness The inconsistency criterion (InconCheck(S; D) ! fl) is the key to LVF. The criterion specifies to what extent the dimensionally reduced data is acceptable. The inconsistency rate of the data described by the selected features is checked against a pre-specified rate (fl). If it is smaller than fl, it means the dimensionally reduced data is acceptable. The default value of fl is 0 unless specified. The inconsistency rate of a dataset is calculated as follows: (1) two patterns are considered inconsistent if they match all but their class labels; (2) the inconsistency count is the number of all the matching patterns minus the largest number of patterns of different class labels: for example, there are n matching patterns, among them, c 1 patterns belong to label 1 , c 2 to label 2 , and c 3 to label 3 where 3 is the largest among the three, the inconsistency count is and (3) the inconsistency rate is the sum of all the inconsistency counts divided by the total number of patterns (P ). It can be easily shown that if the inconsistency rate is 0 for both datasets with M and N features, and FN is the original feature set and FM is the chosen feature subset, then the conditional probability P(C exactly equals P(C j for different possible classes, C, where f M and f N represent vectors of values of respective feature vectors FM and FN . The inconsistency criterion is a conservative way of achieving the "class separability" which is commonly used in pattern recognition as the basic selection criterion [6]. A limited version of this was first proposed by [2] as the MIN-FEATURES bias on a binary domain. Instead of aiming to maximize the class separability, our measure tries to maintain the original class separability of the data. The inconsistency criterion is also in line with information-theoretic in all the experiments in this paper. We tried not to use too large c so that for all (small and large) datasets, we could use just one fixed MAX TRIES. The reader may do as we have done in another version of LVF to link MAX TRIES to the percentage of the total search space according to the desired quality of selected features. considerations [28] which suggest that using a feature that is good for discrimination provides compact descriptions of each of the two classes, and that these descriptions are maximally distinct. Geometrically, this constraint can be interpreted [17] to mean that (i) such a feature takes on nearly identical values for all examples of the same class, and (ii) it takes on some different values for all examples of the other class. The inconsistency criterion aims to retain the discriminating power of the data for multiple classes after feature selection. 3.2 Theoretical analysis Our analysis shows that LVF can give a good solution, or an optimal solution if MAX TRIES is sufficiently large. With a good pseudo random number generator [22], selecting an optimal subset of M features can be considered as sampling without replacement. The probability of finding the optimal subset at the (k+1)th experiment is 1 , and the probability of having to conduct (k+1) experiments before finding the optimal subset is \Theta ::: \Theta 1 where N is the number of original features. When N is large, MAX TRIES 2 N . Here we assume there is only one optimum. If there exist l optima as in many applications, at the (k 1)th tossing, the probability of finding one optimum is l . Roughly, when the number of optima is doubled, the number of run times can be halved. Referring to the LVF algorithm, we notice that the inconsistency criterion is checked only when C C best . Thus, when C best is reduced due to the random search, the number of inconsistency checking is also reduced. As shown in Section 5.1, for the real-world datasets, C best can be as few as one fifth of the original number of features (Mushroom). In addition, the time complexity of the checking is O(P ). Hence, LVF is expected to run fast. If the equivalently good subsets are not required, the last two lines inside the for-loop of the LVF algorithm can be removed, LVF can be made even faster. 3.3 Applying LVF to huge datasets (a practical case) Feature selection is particularly useful when datasets are huge since many learning algorithms may encounter difficulties. As mentioned earlier, feature selection can help reduce the dimensionality of the datasets so that learning algorithms can be used to induce rules. Hence, huge datasets are also an ultimate test for a feature selection algorithm. LVF had an opportunity to undergo a real test of huge datasets. In Section 5 (Empirical Study) below, the results of LVF on the benchmark datasets are reported. The datasets involved are related to the service industry. LVF was given to a local institution 5 , which was in need of a method to reduce the number of features before applying some machine learning algorithms to the datasets due to the huge size of the datasets. Because the datasets are confidential, we have no access to them. The users at the institution ran LVF independently 5 Japan-Singapore AI Centre, Singapore. and without modification and provided the following account: One dataset (let us call it HD1) has 65,000 patterns and 59 features; the other (HD2) has 5,909 patterns and 81 features. Both datasets are discrete, feature values range from 2 to 13. LVF found that 10 and 35 features were relevant for describing HD1 and HD2 respectively without sacrificing their discriminating power, after hours of running LVF on a Sun Sparc workstation. Due to the long waiting time, they did another experiment in which only 10,000 patterns of HD1 were used, it took LVF about 5 minutes to complete its run and obtained the same results. The results are summarized in the table below. The stark difference between hours and minutes inspired us to extend the work of LVF. In short, their findings manifest two points: (1) LVF significantly reduced the number of features; and (2) reducing the number of patterns significantly reduced the run time. It is the second finding that leads us to incremental feature selection. Data #Features #Patterns #Selected Time hours hours The largeness of a dataset can be differentiated into two types: (1) horizontal largeness - the number of features, and (2) vertical largeness - the number of patterns. In our implementation of LVF, we have considered overcoming the horizontal largeness by applying a Las Vegas algorithm in order to avoid exhaustive search and attack the vertical largeness by using a hash mechanism in order to speed up. However, the above practical case shows that more can be done in overcoming the vertical largeness. Hence, in the following, when we mention largeness, we mean the vertical one (P ). 4 Incremental Probabilistic Feature Selection Although LVF can generate optimal/suboptimal solutions (see the experimental results below), when datasets are huge, as shown in Section 3.2, checking whether a dataset is consistent still takes time due to its O(P ) complexity. It is only natural to think about an incremental version of LVF that can significantly reduce the number of inconsistency checkings. Studying the LVF algorithm, we notice that if we reduce the data, we can decrease the number of checkings. However, features selected from the reduced data may not be suitable for the whole data. The following algorithm is designed to achieve that features selected from the reduced data will not generate more inconsistencies than those from the whole data. Furthermore, this is done without sacrificing the quality of feature subsets which is measured by the number of features and by their relevance. LVI algorithm percentage of the data used for feature selection, dataset of N features, fl - allowed inconsistency rate; Output: sets of M features satisfying the inconsistency criterion of D randomly chosen ; the rest data */ loop if (checkIncon(subset, D 1 , inconData) != return subset; else remove(inconData, loop In LVI, checkIncon() is similar to InconCheck() in LVF. In addition, it saves the inconsistent patterns of D 1 to inconData. The experiments below are designed to demonstrate the claims made above on LVI. The incremental algorithm (LVI) starts with a portion of data (p%) and an acceptable inconsistency rate (fl) which is usually set to 0 if there is no prior knowledge, or the minimum value of fl can be obtained from applying InconCheck(F; D) in LVF where F is the set of N features. LVI splits the data D into D 0 and D 1 where D 0 is p% of D and D 1 is the remaining. LVI uses a subset of features (subset) for D 0 found by LVF to check subset on D 1 . The actual inconsistency found in D 0 is fi fl. If the inconsistency rate on D 1 does not exceed stops. Otherwise, it appends those patterns (inconData) of D 1 , which cause the additional incon- sistency, to D 0 , and deletes inconData from D 1 . The selection process repeats until a solution is found. If no subset is found, the whole set is returned as a solution. 5 Empirical Study The error probability plays the most important role in the feature selection algorithms. Ultimately, it is always used as a meta-selection criterion [25]. That is, regardless of different feature selection algorithms, the subset with the lowest estimated error will always be selected for classification tasks. An error is caused by a wrongly classified pattern. The number of errors divided by the number of total patterns in the set gives us the error rate. Each dataset is split into two sets (training vs. testing). Error rates are obtained for both sets. It is the error rate of the testing set that estimates the performance of a classification algorithm. In order to check error rates before and after features selection, both artificial and real-world datasets are used in the study of the effectiveness of LVF and LVI. These datasets are either commonly used in comparison or having known relevant features. All but two (CorrAL and Parity5+5) datasets can be obtained from the UCI Repository [18]. Artificial ffl CorrAL The data was designed in [8]. There are six binary features, I is irrelevant, feature C is correlated to the class label 75% of the time. The Boolean target concept is chose feature C as the root. This is an example of datasets in which if a feature like C is removed, a more accurate tree will result. ffl Monk1, Monk2, Monk3 The datasets were taken from [27]. They have six features. The training datasets provided were used for feature selection. Monk1 and Monk3 only need three features to describe the target concepts, but Monk2 requires all the six. The training data of Monk3 contains some noise. These datasets are used to show that relevant features should always be selected. ffl Led17 This data is generated artificially by a program at the UCI data mining repository. It generates 24 features among which the first 7 are used to display a value between 0 - 9 in the seven segment display system. The remaining 17 features are generated randomly. All the values are binary except the class which takes a value between 0 and 9 (representable in seven segments). The number of patterns to be generated is determined by the user. 20000 patterns were generated for our experiments. ffl Parity5+5 The target concept is the parity of five bits. The dataset contains of them are uniformly random (irrelevant). The training set contains 100 patterns randomly selected from all 1024 pat- terns. Another independent 100 patterns are drawn to form the testing set. Most heuristic feature selectors will fail on this sort of problems since an individual feature does not mean anything. Real-World ffl LungCan The Lung Cancer data describes 3 types pathological lung cancers found in UCI repository. This data contains only patterns and 56 features taking the values 0-3. ffl SoybeanL In the UCI machine learning repository, we found training and testing datasets in two separate files containing 307 and 376 patterns respectively. It contains 35 features describing symptoms of 19 different diseases in soybean plant. ffl Vote This dataset includes votes from the U.S. House of Representatives Congress-persons on the 16 key votes identified by the Congressional Quarterly Almanac Volume XL. The dataset consists of 16 features, 300 training patterns and 135 test patterns. Table 1: Notations: C - the number of distinct classes, N - the number of features, S - the size of the dataset, S d - the size of the training data, S t - the size of the testing data. Training and testing datasets are split randomly if not specified. Dataset C N S S d S t LungCan 3 56 Mushroom 2 22 8125 7125 1000 ffl Mushroom The dataset has a total of 8124 patterns, of which 1000 patterns are randomly selected for testing, the rest are used for training. The data has 22 discrete features. Each feature can have 2 to 10 values. ffl Krvskp This is the data for Chess End-Game - King+Rook versus King+Pawn on a7. The Pawn on a7 means its one square away from queening. Its the King+Rook's side (white) to move. The data contains 3196 patterns and 36 features. The class value 1 indicates white can win, which means white can check the black pawn not to advance and vice versa. Each pattern is a board description for this chess end-game. The first 36 features describe board and the last one is the classification. The major measurements of these datasets are summarized in Table 1. Since most datasets in the first group do not have a large number of patterns, we choose Vote and Mushroom plus ParityMix, Led17 and Krvskp to form the second group of datasets to show the effectiveness of LVI in relation to the size of datasets (small, medium, large). ParityMix is composed by having two Parity5+5's side by side so that there are 20 features in total. 5.1 Effectiveness of LVF For the artificial datasets, the evaluation of LVF is simple since the relevant features are known. However, for the real-world datasets, it is not clear what the relevant features are. Therefore, whether the selected features are relevant or not can be only determined indirectly. One way is to see the effect of fea- Table 2: Results of 100 runs of LVF on the datasets with one example of the minimum set of features for each dataset. N - number of original features, M - number of selected features, F - frequency. Vote Mushroom 22 4 (57), 5 (43) A4, A5, A12, A22 6 Allowing 5% inconsistency. If not, four features are selected: the above chosen 3 plus A1. ture selection through a learning algorithm. Among many choices, we chose C4.5 [24] and NBC [29] in our experiments because (1) C4.5 is a decision tree induction algorithm that works well on most datasets as reported by many re- searchers; and (2) it employs a heuristic to find the simplest tree structures. (Naive Bayesian Classifier) employs the Bayes rule by assuming features are independent of each other and is an approximation of Bayesian Classifiers - the optimal classifier. NBC is chosen because it works in a different way from that of C4.5. LVF is run 100 times on each training dataset. The numbers of selected features and frequencies are reported in Table 2 under the condition that the inconsistency criterion be satisfied. Also reported is a sample of these selected features for each dataset which can directly be used by readers in their experiments. For the artificial datasets, the relevant features are always selected, albeit a few of irrelevant ones are also chosen sometimes. For the problem like Parity5+5, LVF correctly identifies the correct features all the time, plus one irrelevant feature sometimes. For the real-world datasets, the number of features is reduced at least by half to less than one fifth of the original. Table 2 shows that those features in the last column are necessary in order to satisfy the inconsistency criterion (the inconsistency rate is 0 except for Monk3). These features are used by C4.5 and NBC to test if its performance improves compared to using all features. Ten-fold cross validation is usually recommended, t-test is used instead of Z-test in calculation of P-values since we need to take into account the small sample effect (10 data in each sample for 10-fold cross validation). The default settings of C4.5 are used in the experiments. For the experiments "after" feature selection, only the features shown in the last column of Table 2 are used. Given in Tables 3 and 4 are the average accuracy rates of C4.5 before and after applying feature selection to the datasets. The same applies to NBC: instead of reporting the tree size, we report the table size. Table 3: 10-fold cross validation results on Tree Size and Error Rates of NBC before and after applying LVF to the datasets. P-val stands for P-value of t-test; and "-" means that the pooled variances of "before" and "after" are zero. Table Size Dataset Before After P-val Before After P-val Monk2 36.0 36.0 - 37.4 37.4 1 Monk3 36.0 22.0 - 3.63 3.63 1 LungCan 722.8 95.4 0.0 56.66 63.33 .6685 Vote 98.0 62.0 - 9.9 9.9 1 Mushroom 236.0 74.0 - 0.33 1.16 .0004 In cases indicated by "-", the comparison between "before" and "after" is obvious. Tables 3 shows that results are consistent with the known fact that there are no bad features from the standpoint of Bayesian decision rules [26]. In all the datasets tested using NBC, only table sizes are all reduced (except Monk2) due to feature selection; error rates are not significantly changed in seven out of nine datasets. For the two datasets (SoybeanL and Mushroom), the latter's error rate increases a little in absolute percentage, but SoybeanL's error rate is much worse after feature selection. This is because the training dataset has fewer patterns than the test dataset (recall that the division is done by the data contributor and because features were selected based on the training data, then both datasets were put together to run 10-fold cross validation). To verify this conjecture, we did another experiment in which features were selected using both data sets (training and testing). Fifteen (instead of fourteen) features were chosen (they are A and results of 10-fold cross validation on NBC and C4.5 are 14.4% (7.0% before) and 9.7% (7.3%) respectively. Thus, error rates are lower with all data used for feature selection. The only improvement of NBC's performance is on Parity5+5, but it is not statistically significant. Results in Table 4 suggests that the performance of C4.5 improves in general. That is, the tree size is getting smaller and the error rate lower. For the artificial datasets, this experiment further shows with the relevant features, C4.5 does better than that with the full set of features. For the real-world datasets, C4.5 is also doing better with the selected features. This indicates that LVF has selected relevant features for these datasets. In particular, C4.5 did poorly on Parity5+5 Table 4: 10-fold cross validation results on Tree Size and Error Rates of C4.5 before and after applying LVF to the datasets. P-val stands for P-value of t-test; and "-" means that the pooled variances of "before" and "after" are zero. Tree Size Dataset Before After P-val Before After P-val Monk1 41.9 41.0 .0782 1.3 0.0 .1937 Monk2 14.3 14.3 1 35.4 35.4 1 Monk3 19.0 19.0 - 1.1 1.1 1 LungCan 18.3 16.6 .03 56.7 57.5 .2627 7.3 15.2 .0001 Vote 14.5 6.1 .0001 5.3 5.5 .8357 before feature selection. Nevertheless, C4.5 did as well on Mushroom with 22 features as with 4 features. This demonstrates that C4.5 does select relevant features for some datasets, though not for all. The only serious deterioration of C4.5's performance is seen in the results for SoybeanL. The reason is given above in explaining NBC's poor performance on the dataset. The gain from feature selection differs for NBC and C4.5. The difference is due to the way in which features are used to induce a classifier. C4.5 is a selective induction algorithm that selects the best feature at each test for tree branching. NBC uses all features' conditional probabilities in determining a pattern's class. Since NBC assumes that features are conditionally independent given the class, the conditional probabilities of an irrelevant feature given the class will be approximately the same, so it is not a good discriminant. 5.2 Effectiveness of LVI For this set of experiments, we want to verify four claims: (1) LVI may not be suitable for small datasets; (2) LVI can run faster than LVF on large datasets; (3) LVI does not sacrifice the quality of the selected features; and (4) if no solution can be found by LVF in the earlier runs, neither it can in later runs (earlier runs start with less data). Five datasets in the second group are chosen for experiments. They are (1) Vote, (2) Mushroom, (3) ParityMix, (4) Krvskp, and (5) Led17. The experiments are conducted as follows. For each dataset, starting with 10% of the data (D 0 ) for feature selection, we run LVI 10 times, recording the number of features, features, and selection time in each run. Subsequently, we do the same experiments with 20%, 30%, ., 90%, and 100% of the data. The average time and number of features are computed for each experiment. Using 100% of data as the reference, we calculate the P-values for each sized D 0 . A low P-value (e.g., ! 5%) suggests that the NULL hypothesis that the two averages are the same be rejected. Refer to Figures 1 and 2 for varied P-values shaded differently. We summarize the findings from the experiments as follows. ffl The effectiveness of LVI becomes more obvious when the data size is larger. LVI performs well on all the three datasets. If the data size is small (around a few hundred) as in Vote even the time saving for the best D 0 is not much. However, the saving is significant in the case of ParityMix and a clear trend can be observed. This is due to the overheads required by incremental feature selection. Since our inconsistency checking is fast (O(P )), if P is not sufficiently large, the time saving will not be apparent, it may even be negative if P is too small. This is why LVI is more suitable for large sized datasets. ffl Another issue is the number of patterns with which LVI should start for a dataset. Having either too few or too many patterns affects the LVI's per- formance. If too few patterns are used (D 0 is too small), LVF could select few features that cannot pass the inconsistency check on the remaining data (D 1 ), that is, inconData can be large. The worst case is that after the first loop, D 0 becomes D (the whole dataset). One case of too small a D 0 can be observed in Figure 1 for Vote when 10% of the data was used; it took longer time than that using 20% - 50% of D. If too many patterns are used (D 0 is large), the overheads (i.e., the time spent on those steps inside the loop of the LVI algorithm after the LVF call) plus the time on LVF may exceed the time of simply running LVF on D. In the cases of Vote and Mushroom, the difference in times is not statistically significant when 70% or more of D is used. ffl The incremental algorithm does not have to sacrifice the quality of feature selection. The time saving is mainly due to (1) small D 0 which is usually a portion (say 10%) of D, and (2) by remembering inconsistent patterns, LVI can avoid checking wrong guesses twice. The quality is measured here in two dimensions. One is the number of features, and the other is the relevance of the features. As shown in Figure 2, if there is some statistically significant difference in the number of features selected between various sized D 0 's against D, the numbers of features are lower than using 100% of the data; otherwise, there is no statistically significant difference according to t-test. For ParityMix, the relevant 5 features are always selected plus 1 or 2 irrelevant/redundant ones. For the Vote and Mushroom, the relevance test is done through a learning algorithm (C4.5 and NBC here). If their performances do not deteriorate or even improve, we conclude that these features are relevant. Experimental results shown in Tables 3 and 4 have verified the sets of features for the two datasets via 10-fold cross validations. When the quality of the selected features can be warranted, the reduction can simplify data analysis, rule induction, as well as data collection in future. ffl LVI can scale up. Time complexity of a feature selection algorithm can be described along two dimensions: number of features (N ) and number of patterns (P ). By approximating MAX TRIES of LVF with 77\ThetaN (reduced from 2 N ), the time complexity of LVF is mainly determined by P since N is relatively very small. The incremental version, LVI, makes it possible to start with a fixed small number of patterns (e.g., a few thousand for D 0 ), no matter how large the original dataset is. The experimental results show that the time saving by so doing is statistically significant when P is large. ffl For data Krvskp, no feature can be removed from 10% data to 100% data used. It indicates the other side of incremental feature selection by LVF: if LVF cannot reduce features based on a smaller portion of data, then more or all data cannot help reduce features either, other things being equal. It will help if we extend the run time, for example, linking MAX TRIES to the percentage of the total search space. 6 Discussion and Conclusion The time performance of LVF is not reported for the first set of data because (1) LVF completes its run fast (in a few seconds of elapsed time); (2) there is not much to compare with among the small datasets; and (3) the time measurements of LVI for the large datasets indicate the time performance of both LVF and LVI. Both algorithms are simple to implement and fast to obtain results. By predefining fl according to prior knowledge, LVI and LVF can handle noisy data, as shown in the case of Monk3. Both can deal with multiple class values. Another feature of LVF is related to so-called anytime algorithms [3] that are algorithms whose quality of results improves gradually as computational time increases. LVF prints out a possible solution whenever it is found; afterwards LVF reports either a better subset or equally good ones. This is a really nice feature because while it works hard to find the optimal solution, it provides near optimal solutions. There is no need for a user to wait for results until the end of search for optimal/suboptimal solutions as other types of search do. The longer LVF runs, the better the solutions it produces. One salient feature of LVI is its scaling capability without losing the quality of selected features. The suggested modification - sampling subsets of features without replacement of selected features and constraining subset generation by the newly found minimum number of features should allow LVF to work faster. In order to verify that a filter feature selector can easily be turned to a wrapper one, LVW is built to prove the case [14]. If a favorite induction algorithm is available, LVF can be easily transformed into LVW. The experimental results show that LVW is much slower than LVF. This finding is consistent with the results reported in [10]. There may be a problem with using inconsistency as a feature selection criterion when one feature alone (such as social security number) can guarantee that there is no inconsistency in the data. Obviously, this feature is irrelevant for rule induction. The problem can be solved by leaving this feature out of the feature selection process. If there is no prior knowledge, it will just take one run of LVF to locate this kind of features 7 . Another run of LVF with the other features will identify the right set of features. LVF only works on discrete features since it relies on the inconsistency cal- culation. One way is to apply a discretization algorithm (e.g., Chi2 [13]) to discretize the continuous features first before one runs LVF. Other possibilities are (1) to simply treat a continuous feature as a discrete one in some cases; and (2) to apply LVF only to the discrete features when the number of features is large. More work is needed. The search for new criteria in addition to the inconsistency continues. LVI and LVF can find other uses as well. As mentioned earlier that Las Vegas algorithms may not be as fast as some domain specific heuristic methods, LVI can still play a role as a reference in design of a domain specific heuristic method. This is because it is not an easy task to verify a heuristic method, especially when datasets involved are huge. LVI can be most helpful in this case to validate feature subsets found by the heuristic method. Another feature is that LVF may produce a number of equally good solutions for one dataset based on the inconsistency criterion. One solution can be chosen according to its predictive accuracy of a learning algorithm. That is, we choose a solution that generates the best accuracy. This suggests a straightforward extension of this work, i.e., a combined filter and wrapper model of this incremental probabilistic algorithm. The significant advantage of LVI is that we move one step further in handling the large sized datasets. It is a necessary addition to the present repertoire 8 . So far, all algorithms are of automated feature selection. We have not mentioned another important practical issue - using domain knowledge in feature selection. Domain knowledge or expert's understanding of the data can help tremendously in feature selection. For instance, domain knowledge can be used to verify the finding of automated feature selection; domain knowledge can be used to remove some obviously irrelevant or redundant features; and domain knowledge can also help in designing heuristics. When expertise is available, one should always start feature selection from what is known first, and apply automated selection algorithms next. Acknowledgments The authors would like to thank H.Y. Lee for the suggestions on an earlier version of this paper and H.L. Ong and A. Pang for providing the results on their applying LVF to huge datasets at Japan - Singapore AI Center. Thanks also go to Manoranjan Dash and Farhad Hussain for conducting some experiments using 7 Recall that one run of LVF has MAX TRIES loops. 8 Both LVI and LVF are available for research purposes upon request. LVF and LVI, and Jian Shu for implementing NBC used in the experiments. The suggestions by anonymous referees have also significantly helped improve the paper. --R Tolerating noisy Learning boolean concepts in the presence of many irrelevant features. Deliberation scheduling for problem solving in time-constrained environments Fundamentals of Algorithms. Feature selection methods for classifications. Pattern Recognition: A Statistical Approach. Feature selection: Evaluation Irrelevant feature and the subset selection problem. The feature selection problem: Traditional methods and a new algorithm. Wrappers for performance enhancement and oblivious decision graphs. Toward optimal feature selection. Selection of relevant features in machine learning. Chi2: Feature selection and discretization of numeric attributes. Feature selection and classification - a probabilistic wrapper approach A probabilistic approach to feature selection - a filter solution Concept learning through feature selection. Principled constructive induction. UCI repository of machine learning databases. Feature selection using rough sets theory. A branch and bound algorithm for feature subset selection. Boolean feature discovery in empirical learning. Numerical Recipes in C. Induction of decision trees. Inductive Pattern Classification Methods - Features - Sen- sors On automatic feature selection. The Monk's problems: A performance comarison of different learning algorithms. Pattern Recognition: Human and Mechanical. Computer Systems That Learn. A critical evaluation of intrinsic dimensionality algorithms. --TR --CTR Stergios Papadimitriou , Seferina Mavroudi , Liviu Vladutu , G. Pavlides , Anastasios Bezerianos, The Supervised Network Self-Organizing Map for Classification of Large Data Sets, Applied Intelligence, v.16 n.3, p.185-203, May-June 2002 Myung-Kuk Park , Ki K. Lee , Key-Mok Shon , Wan C. Yoon, Automating the Diagnosis and Rectification of Deflection Yoke Production Using Hybrid Knowledge Acquisition and Case-Based Reasoning, Applied Intelligence, v.15 n.1, p.25-40, July-August 2001 Wei-Chou Chen , Ming-Chun Yang , Shian-Shyong Tseng, The bitmap-based feature selection method, Proceedings of the ACM symposium on Applied computing, March 09-12, 2003, Melbourne, Florida Huilin Ye , Bruce W. N. Lo, Feature Competitive Algorithm for Dimension Reduction of the Self-Organizing Map Input Space, Applied Intelligence, v.13 n.3, p.215-230, November-December 2000 Wei-Chou Chen , Ming-Chun Yang , Shian-Shyong Tseng, A novel feature selection method for large-scale data sets, Intelligent Data Analysis, v.9 n.3, p.237-251, May 2005 Samuel H. Huang, Dimensionality Reduction in Automatic Knowledge Acquisition: A Simple Greedy Search Approach, IEEE Transactions on Knowledge and Data Engineering, v.15 n.6, p.1364-1373, November Ki K. Lee , Wan C. Yoon, Adaptive classification with ellipsoidal regions for multidimensional pattern classification problems, Pattern Recognition Letters, v.26 n.9, p.1232-1243, 1 July 2005 Xindong Wu , Shichao Zhang, Synthesizing High-Frequency Rules from Different Data Sources, IEEE Transactions on Knowledge and Data Engineering, v.15 n.2, p.353-367, February Bill B. Wang , R. I. Bob Mckay , Hussein A. Abbass , Michael Barlow, A comparative study for domain ontology guided feature extraction, Proceedings of the twenty-sixth Australasian conference on Computer science: research and practice in information technology, p.69-78, February 01, 2003, Adelaide, Australia
dimensionality reduction;pattern recognition;feature selection;machine learning
590876
A Neural Network Diagnosis Approach for Analog Circuits.
This paper presents a neural network system for the diagnosis of analog circuits and shows how the performance of such a system can be affected by the choice of different techniques used by its submodules. In particular we discuss the influence of feature extraction techniques such as Fourier Transforms, Wavelets and Principal Component Analysis. The system uses several different power supplies and as many neural networks in parallel. Two different algorithms that can be used to combine the candidate sets produced by each network are also presented. The system is capable of diagnosing multiple faults even if trained on single ones.
Introduction During the past years, the authors have been involved in several projects on analog circuit diagnosis and quality control of electrical components. The aim of this paper is to present the diagnostic system developed by them and to show how the performance of such a system can be affected by the choice of different techniques used by its submodules. The system is based on neural networks , and is used for off-line diagnosis of analog circuits affected by catastrophic multiple faults . It may handle linear and nonlinear circuits in transient or steady state behavior. 1.1. Analog circuit faults Fault diagnosis of analog circuits is a complex problem. Classical solutions require either a huge amount of calculation if parameter identification methods are used, or a great number of simulations of faulty conditions if fault dictionary methods are used [23, 24]. The faults in analog circuits may be catastrophic faults, that cause a large and sudden variation of the circuit parameter values, and deviation faults, associated to slight variations of the circuit parameter values from their nominal values [2]. Since statistics have shown that 80-90% of analog circuit faults are catastrophic [19], we chose to study faults of this kind, such as short circuits and open circuits between two terminals of a component. In some applications (regulation systems, nuclear plants, etc.) a prompt fault detection is necessary to avoid damaging the controlled process any further. A diagnostic system capable of detecting a fault during its occurrence performs what is called on-line diagnosis. In many other applications (quality control of circuits, post-mortem diagnosis of electronic boards, etc.) the diagnostic procedure may be applied in an off-line fashion, in the sense that the diagnosed device need not be operative. In these cases there is no strict time constraint and even computationally intensive diagnostic sys- tems, such as those based on parameter identification or fault dictionary methods [23, 24], qualitative reasoning [11, 13] model-based and rule-based expert systems [7, 28] etc., may be used. In the case of electric circuits, off-line diagnosis offers an additional advantage: suitable voltage supply configurations may be chosen in order to maximize the observability of the faults [10]. 1.2. Diagnosis as pattern recognition Classically, a pattern recognition system is composed of three modules [12]. A transducer acquires data on a physical device and passes them to a feature extractor whose purpose is to reduce the data by computing certain features (or prop- erties). These features will be used by a classifier to make a final decision on the state of the device. A circuit diagnostic system is a particular pattern recognition system, in which the physical device is an analog circuit, and the state that must be recognized is the set of faulty components. In particular, in the diagnostic system we have de- veloped, the classifier is a neural network. This type of diagnostic system offers some advantage over other classical diagnostic methods. Rule-based systems. These diagnostic systems use compiled sets of rules to associate a symptom to its cause. On the contrary, a neural network automatically derives the symptom- cause correspondence during the training, and does not require an explicit formalization. It is well known that this formalization is the bottleneck of rule-based system technology. Note that there is a small price to pay for this. A rule-based system has a symbolic-heuristic approach to diagnosis and is generally able to justify its deduction from the rules used to compute the diagnosis. A neural network, on the contrary, has a numerical-algorithmic approach and the knowledge is implicitly memorized in the weights of its synapses. Thus, to justify its deduction a neural network requires additional rule extraction techniques [9]. Model-based systems. These diagnostic systems usually require the complete knowledge of the circuit scheme and a model of its be- havior. Using neural networks it is possible to avoid the problems connected with the calculation of circuit parameters and in general to the modeling. Fault dictionary method. This method can be used to identify only those faults whose signature has been previously computed and added to the dictionary. Neural networks on the contrary - as reported in several works - may be able to recognize fault configurations not explicitly included in the training set. In [14, 21, 31] neural networks trained to recognize single faults are successfully used to diagnose multiple faults. In [36] neural networks accurately classify previously unseen fault signatures belonging to a deviation fault class known by a few samples. There have been several works where neural networks have been compared with other pattern classifiers in diagnosis applications. In the domain of single fault diagnosis of circuits, a comparison with Gaussian maximum likelihood and K-nearest neighbors is presented in [26] where neural net- works, once trained, are shown to significantly reduce the time of the diagnosis, although they do not offer improvements in the diagnostic accu- racy. The same result was independently reported in [36], comparing neural networks and K-nearest neighbors classifiers in the diagnosis of deviation faults. In the diagnostic system we present, the transducer is an acquisition board that measures the voltage values in a given set of test points . Other choices are possible as we will discuss in Section 2. As an example, Spence et al. [34, 35] have used nonintrusive circuit measurements (such as infrared images or magnetic field images); however, nonintrusive measurements have been proved to be very ineffective, in the sense that they can only be used to recognize a limited number of faults. Kirkland and Dean [22] obtained good results using current measurements; however, current mea- Neural Diagnosis for Analog Circuits 3 surements are often impractical, since they would require the opening of the circuit, and this is clearly not possible on printed circuits. We have investigated several feature extraction techniques and have studied their influence on the performance of the diagnostic system. In this paper we compare Fourier Transforms [14], Wavelets [8], Principal Components Analysis [15], and Sampling. In [16] Mean and Root-Mean- Square Values of the test point voltages were used as features, but due to the large amount of lost information they could only recognize a limited number of faults. We also observed that the performance of the diagnostic system heavily depends on the choice of power supplies. In particular, it is often the case that a given supply can only lead to the detection of a particular subset of all possible faults. A suitable set of different supplies may be used to build a diagnostic system that combines different diagnoses (one for each supply) dramatically improving the performance of the diagnostic system. In the paper we also present two algorithms that can be used to combine these different diagnoses. The paper is structured as follows. In Section 2 we recall relevant work on the use of neural networks for circuit diagnosis. In Section 3 we describe the architecture of the proposed diagnostic system and discuss the important issue of simulation versus acquisition. In Section 4 we discuss the choice of power supplies and how this affect the diagnosis. In Section 5 we describe different techniques that can be used by the features extraction module to compactly represent the behavior of the circuit. In Section 6 we present the structure of neural network classifier and show how it is trained. In Section 7 we present two algorithms that can be used to combine the diagnosis computed by different networks. In Section 8 we present statistics of the system performance when diagnosing two different circuits: a board part of a DC motor drive, and an oscillator. 2. Relevant work Other approaches to the use of neural networks for circuit diagnosis have recently been published. Keagle et al. [21] discuss how networks trained to recognize single faults may be used to detect multiple faults. Tests are performed on a digital circuit consisting of nine logical gates affected by stuck-at 1 or stuck-at 0. The paper also presents results on the performance of the diagnostic system as a function of the network architecture. Meador et al. [26] compare feedforward neural network performance with other classifiers: gaussian maximum likelihood and K-nearest neigh- bors. In each experiment a single parameter deviation fault on an operational amplifier circuit is considered. The classifiers must separate the input patterns corresponding to the correct behavior and to the faulty one. Parten et al. [29] propose using neural networks as part of a model-based expert system for diagnosing lumped parameter devices. The purpose of the net would be that of solving the equations ruling the behavior of the diagnosed device, modeled as a set of interconnected components. Thompson et al. [38] consider the problem of diagnosing an IC board with approximately components, both analog and digital. They use a backpropagation neural network with a modular structure, i.e., each part of the net recognizes a particular fault. Totton and Limb [39] use neural networks to diagnose a circuit board part of a digital telephone exchange. They observed from historical data that failures on four types of components account for more than 85% of all faults. This led them to construct a network whose four outputs signal the presence of a faulty component of a given type, i.e., the network does not pinpoint the faulty component but simply detects what type of component is faulty. Spence et al. [34] use a different approach to the single fault diagnosis of printed circuit boards (PCB). The difference between the malfunctioning infrared image and the image of a correctly functioning PCB is interpreted by an artificial neural network to diagnose some types of faulty components. In a subsequent work Spence [35] presents a different test method based on the interpretation of the magnetic field close to the PCB. Although these methods can only recognize a limited number of faults, they have the advantage of requiring nonintrusive measurements. Rutkowski [31] was the first to suggest the use of neural networks for the diagnosis of multiple faults on analog DC circuits. In this introductory work, 4 Fanni, Giua, Marchesi and Montisci the main focus is on testing the capability of the network to generalize from single to double fault diagnosis. In the application example presented in the paper, only a limited number of faults are considered. Bernieri et al. [3] use a neural net-work for on-line analysis of dynamic discrete-time systems whose input/output behavior is ruled by equations of the form: y f(y . The network at the k\Gammath instant receives as inputs the value of y and is trained to estimate the value of given parameters that rule the behavior of the system. Parameter deviations over a given threshold are symptoms of faults. Kirkland and Dean [22] have reported using input current measurements as circuit images. Gu et al. [17] combine neural networks and expert systems into a single diagnostic system. To each component is associated a neural network trained to recognize the component's fault. The expert system acts as a coordinator between the different neural networks, supplying suitable inputs to the networks and deriving a diagnosis from the analysis of the networks' output. Spina and Upadhyaya [36] have considered the problem of diagnosing deviation faults in linear circuits. A white noise source is used to automatically generate test patterns. Fault signatures are generated associating to a single component a value equal to the nominal value plus 50%. The network can correctly classify previously unseen patterns corresponding to deviation faults of different magnitude. All these works highlight the prominence in a neural diagnostic system of the aspects related to feature extraction and circuit supplies, thus leading us to a systematic exploration of these issues. The present paper summarizes the results that its authors have obtained throughout a long period of time and that have only partially been presented in the papers referenced in the rest of this section. In [16] is discussed how networks trained to recognize single faults on analog circuits in dynamic behavior may be used to detect multiple faults. The neural network identifies the faulty components from the mean values of the voltage measurements in a given set of test points. In general it was observed that the network is able to diagnose multiple faults on two and three components, although less sharply than in the single fault case, due to the presence of false alarms. The set of multiple faults was chosen among those single faults well recognized by the network. In [14] Fourier transforms are used as features of the circuit image, and multiple neural networks were used in parallel by the diagnostic system. This improved the performance of the diagnostic system with respect to the previous one. In [8] Wavelet transforms are used as features. Wavelets proved to be a good data compression technique when the circuit is studied during a transient. In fact, one can increase the number of wavelets only in particular time intervals depending on the degree of approximation required. In [15] Principal Component Analysis is used in the feature extraction phase. The main advantage of such a technique lies in the fact that it gives a simple automatic procedure to compress the data. 3. Architecture of the diagnostic system The architecture of the proposed diagnostic system is shown in Figure 1. Testing procedure (horizontal path) Given a circuit to diagnose, we apply a suitable power supply and acquire the voltage signals at a given set of test points, constructing the circuit image. We extract significant features, as discussed in Section 5, from the image and use them as inputs to a neural network that has been previously trained to recognize single faults on that circuit. The neural network will generate the candidate set , i.e., the set of components recognized as faulty. In Section 4, we will show that to increase the number of detectable faults it is necessary to use different supplies. Consequently, we will have several neural networks, one for each supply consid- ered. Repeating the procedure described above for all supplies, we obtain several candidate sets. These sets will be combined to derive a single diagnosis using suitable algorithms, as described in Section 7. Training procedure (vertical path) The diagnostic system is built training the neural networks that will be used in the testing procedure Neural Diagnosis for Analog Circuits 5 Each neural network is trained using a set of patterns corresponding to all possible single- faults, as detailed in Section 5 and 6. The training patterns are constructed from the faulty circuit images using the same feature extraction technique that will be used in the testing. It may be possible to obtain each faulty circuit image using an acquisition board. One has to pro- duce, one by one, all single faults on the circuit and then has to acquire the corresponding faulty circuit image. This procedure is not practical in many cases. Thus we resorted to PSpice simulation of the circuit behaviour in faulty conditions. On the contrary, when constructing the circuit image in fault-free condition, both real acquisitions and PSpice simulations are possible. As we will later discuss, several real acquisitions will be used to estimate the magnitude of the measurement noise. Our results showed that if the circuit PSpice model is accurate enough, there is no difference between a network trained with "simulated" patterns and a network trained with "acquired" pat- terns. In fact, the distance between a simulated and an acquired pattern has the same order of magnitude of the distance (due to measurement noise and component parameter tolerance) between two patterns acquired during the same fault condition. 4. Power supplies One of the main problems in the diagnosis of circuits is the presence of undistinguishable and undetectable faults. Consider two (or more) components, say k and k 0 in parallel as in Figure 2.(a). Clearly the behavior of the circuit is the same whenever component k or component k 0 is short circuited. The same problem appears when we consider open circuit faults of series components as in Figure 2.(b). Faults of this kind are called undistinguishable, in the sense that they produce the same voltage configuration at the available test points. A similar problem may arise when a fault is un- detectable. In this case, the measured behavior of the fault-free circuit is the same as the measured behavior of the faulty circuit. The presence of undistinguishable and undetectable faults may have different causes. ffl Topology of the circuit , as in the examples discussed above. ffl Limited number of test points , that may not allow detection of an abnormal behavior of the circuit. ffl Components whose measured behavior is the same when faulty or correctly functioning. We recall some of the possible causes. Operating point of the component. Consider the diode in Figure 2.(c). It is reverse biased and thus for all practical purposes its behavior is the same when the diode is functioning well or when it is affected by an open circuit fault. Frequency content of the supplies. Some frequency components may not be suitable for highlighting a given fault. In DC steady state, for instance, capacitors behave as open circuits and inductors behave as short circuits, as shown in Figure 2.(d) and Figure 2.(e), respectively. Protection subcircuits. The behavior of the protection components is not supposed to affect the overall behavior unless other faults are present. There is little we can do to resolve the ambiguity due to the topology of the circuit or due to the choice of test points. However, we may try to resolve the ambiguity due to the behavior of the circuit by an appropriate choice of power supplies. As an example, a different choice of supply, such a high frequency square voltage, force the diode in Figure 2.(c) to alternatively switch from reverse to forward bias, and the capacitor and inductance in Figure 2.(d),(e) to work in AC. This problem has also been discussed by Dague et al. [10]. These authors add an external stimulation in suitable points so as to disturb the circuit operating conditions. We will train different networks to process the data collected for each different supply configura- tion. Thus, our diagnostic system is composed of several neural networks, each one specialized in detecting a given set of faults. When the system is used to diagnose a circuit, each network will produce a set of candidates, i.e., of possibly faulty 6 Fanni, Giua, Marchesi and Montisci combination of candidate sets power supplies circuit under test feature extractors neural nets PSpice model of circuit under test acquisition board diagnosis feature extractor single faults simulations training Fig. 1. The proposed diagnostic system architecture. components. The overall diagnosis can be computed by means of different algorithms, given in Section 7. 5. Feature extraction techniques We assume that the information on the circuit be- havior, i.e., the circuit image, is given by the voltage measurements in a set of available test points. These points are usually given by the circuit board manufacturer and cannot be arbitrarily chosen. (a) (b) (c) (d) (e) Fig. 2. Examples of undistinguishable and undetectable faults. Since the voltage signal at each test point is a function of time, we need to extract significant features to compactly represent the circuit behavior. Extensive experimental studies showed the influence of the particular feature chosen. The feature used in [16] was the mean value (MV). The diagnostic system performances improved when root- mean-square values (RMSV) or a combination of MV and RMSV were used. When MV or RMSV are used, all the information on the dynamic behavior is lost. Thus other feature extraction techniques are required. We discuss here four different techniques: Fourier Transforms , Wavelets , Principal Components Analysis , and Sampling. During the training , the goal of the feature extraction procedure is to construct an (s\Thetar) matrix X . Each row of this matrix represents the circuit behavior during one of the s acquisitions and each column represents the value of a particular fea- ture. Each row of X is use as a training pattern input for the neural network, hence there will be r nodes in the network input layer, and s training patterns, as discussed in Section 6. During the testing of a circuit, the same feature extraction procedure is used to derive the inputs that will be given to the neural network. In this section, we mainly discuss the feature extraction module as used during the training. Consider a circuit with n components and a given set of m test points. Neural Diagnosis for Analog Circuits 7 The voltage of all test points is measured on a real circuit by an acquisition board during p acquisitions in the absence of faults. These measurements will be used to estimate the magnitude of measurement noise. On the contrary, the faulty circuit images, i.e., the voltage of all test points in presence of a fault, are constructed via PSpice simulation. We consider two single faults for each bipolar component: open circuit and short circuit. We also considered faults on components with more than two terminals. As an example, in the circuit shown in Figure 4, there are trimmers and operational amplifiers. We considered two possible faults on a trimmer (cursor stuck up and cursor stuck down) and just one single fault on an operational amplifier (it was made inoperative by feeding with exceedingly high voltage). In general, let s 0 be the number of the single faults taken into account; then one needs to collect images. 5.1. Fourier transforms A simple technique for compacting the information given by the circuit image without losing the dynamics of the system is given by the Fourier analysis that converts the signals into frequency components [37]. We compute the Fast Fourier Transform of the sampled voltage signal measured at each test point. If we have t voltage samples, we obtain - for each test point - components and we take the amplitude of each component. We are now ready to construct the input pattern matrix . The matrix has initially s rows, one for each acquisition, and m \Delta q columns, one for each feature, i.e., frequency component computed at each test point. Thus the input pattern matrix takes the form X qg. The first p rows of X 0 are associated to the fault-free acquisitions. Matrix X 0 is still unusable because of its high di- mensionality. Domain dependent knowledge may be used to further reduce its number of columns [37]. The data reduction algorithm we propose, requires two phases. 1. Remove features that give no information. We compute for each column j the difference \Delta j between its maximum and minimum element. We also compute the difference ffi j between the maximum and minimum element in the first p rows of the column: this is an index of the numerical uncertainty associated to the value of feature j during the p different fault-free ac- quisitions. Fix a threshold ' ? 1. If \Delta j 'ffi j then the variation of the feature j has the same order of magnitude of the numerical uncertainty and column j will be removed. We used a value of 2. Scale the inputs. To improve separability between patterns we scale the columns of the input pattern matrix in the interval [\Gamma1; 1]. 3. Select a subset of significant features. The idea is to keep only those columns that are necessary to distinguish between different pat- terns. Fix a threshold oe ! 1. If j x 0 oe then the variation of feature j is not large enough to distinguish pattern i from pattern We used a value of We proceed as follows. begin let the initial set of significant features be S := for i := 2; s (* compare each row i of X 0 with all previous ones *) begin ffl let S i;i 0 oeg be the set of those features, i.e., columns, that may be used to distinguish between patterns i and is not empty then add to S the most significant feature, i.e., feature j such that j x 0 We thus obtain a new matrix X of order (s \Theta r) with r m \Delta q. The data reduction algorithm we use with FFT falls into the category of unsupervised feature extraction methods [4], i.e., methods that do not use information on the target data. Note, however, that the data reduction is performed opportunis- tically, by projecting the features onto a subspace 8 Fanni, Giua, Marchesi and Montisci that still contains all information required to separate the input patterns. 5.2. Wavelets The origins of Wavelets date back to 1909, when Haar proposed them as a viable solution to function decomposition problems. In fact Fourier se- ries, as stated in its original formulation, show a non-uniform convergence even for particular continuous functions. Wavelets approach is more suitable than Fourier one, especially when signals are non-stationary. Both "time-frequency" and "time-scale" wavelets are suited to signal analysis ranging from "quasi-stationary" to fractal structure type. Mathematicians speak of "atomic de- composition" of signals, where wavelets are the elementary constituents. The various wavelets are obtained from a single wavelet by scaling and shifting operations. There are several definition of wavelets. One possible is the following [27]: a wavelet is a function y(x) in L 2 (IR) such that 2 j=2 y(2 is an orthonormal basis for L 2 (IR). The most frequently used wavelets are the Grossmann-Morlet wavelets, that are also similar to Daubechies wavelets and to Gabor-Malvar wavelets. The last algorithm is of time-frequency type, while the former is a time-scale algorithm. In the wavelet theory [30, 25] any signal of finite energy can be represented as a linear combination of wavelets whose coefficients represent the features we want to extract, and indicate how close the signal is to a particular basis function. Discrete wavelet transform (DWT) is a relatively recent method whose biggest potential has been found to be signal compression. The two major advantages of the wavelet transformation are that it can zoom in time discontinuity and that it is possible to construct an orthonormal basis, localized in time and frequency. An important issue of wavelet analysis is the choice of the proper type of wavelet and of the methodology to use, i.e., time-scale, time-frequency or a combination of the two. In our diagnostic system, Haar wavelets are chosen to realize data compression of circuit-image information. Decomposition proposed by Haar results as follows: R 1f(t) h i (t) dt , and s n (t) is the n\Gammath order summation which uniformly converges to the signal f(t), and Haar wavelets are defined as: Here, the scaling factor is a power of 2, and k defines the time shift with respect to the basic wavelet H, that is the unit square window func- tion. The various wavelets (n ? are obtained starting from the basic wavelet ing, scaling and shifting operations. It is important to note that the time range has to be limited in the [0; 1] interval. This is not limiting because real signals always have a finite time length and this will become the new time unit. It is also possible to realize a suitable time windowing of the signal. Thus, it is possible to project the time signal onto a set of mutually orthonormal wavelets. The number of the wavelets may be arbitrary, depending on the required approximation in reconstruction or, as in the present case, on the amount of information to extract from the signal. Because a circuit image results from a set of digital acquisition, signals are not continuous in time, but discrete due to sampling. Hence, a discrete transform has to be used and particular care is required to compute the inner products. The construction of the input matrix X using wavelets follows the same procedure presented in Section 5.1 for Fourier transforms and will not be repeated here. 5.3. Principal components analysis Principal Component Analysis (PCA) is another unsupervised feature extraction method. Compression by means of PCA is accomplished by projecting each data vector along the directions of the individual orthonormal eigenvectors of the covariance matrix of data. As the first few eigenvalues of the covariance matrix contain most of the signal energy, the dimensionality of the data can be Neural Diagnosis for Analog Circuits 9 greatly reduced without losing much information on the input data. It may happen that the information associated to the discarded PC subspace is important for the subsequent classification phase [4] and in this case PCA is not suitable. However, PCA is a potentially useful method because it works in many ap- plications. In [1] PCA is used for terrain classifi- cation, and it is shown that it can lead to a significant improvement in the classifier performance. In [6] there is a comparison between Gabor filters and PCA as feature extraction methodologies applied to SAR images segmentation with neural networks. Let s be the number of the circuit behaviors taken into account, and t be the number of samples for each test point voltage. Each circuit image is represented by m \Delta t values. We have a (s \Theta m \Delta t) data matrix X 0 which could be used as input for the neural network. As previously mentioned, preprocessing is necessary to extract from these data the salient fea- tures. We would like to reduce the number of columns of this matrix from m \Delta t to r m \Delta t, with acceptable loss of information. Using PCA [20] this compression is accomplished projecting the s circuit images along the directions of the principal eigenvectors of the covariance matrix of Given the data matrix (X th column represents a circuit im- age, the covariance matrix of these data is the s The eigenvectors of this matrix form an orthonormal basis, and any vector ~x i can be represented with respect to this basis by means of a coefficient vector with m \Delta t elements. To reduce the data dimension, it is possible to consider only those eigenvectors associated to the dominant eigenvalues of C. Fix a threshold c 2 be the ordered set of eigenvalues of C, i.e., j j+1 . We say that there are r dominant eigenvalues if are the eigenvectors associated to the dominant eigenvalues, we may use as compressed representation of a vector ~x i the coefficient vector: We used a value of Thus, the data matrix X 0 is reduced to a (s \Theta r) matrix X . The same compression technique will be used on subsequent circuit images acquired during the test phase. 5.4. Sampling Given the circuit image (i.e., the sampled voltage signals at all test points) one may compact the data retaining just a limited number q of the t samples. Experimental results [16] showed that this is not a viable technique if the circuit is in AC steady-state or if there are many test points. In fact, this leads to a neural network with too many nodes in the input layer, i.e., too many features. This may reduce the performance of the classification system and leads to a higher computational cost of the training. However, this technique was effective when studying short transients on circuits with a limited number of test points. The choice of the samples to retain must be opportunistic, and depends on the signal variation pattern. 6. Neural model As proposed in most of the literature discussed in Section 2, we use a three level neural network with sigmoid activation functions and backpropagation learning with generalized delta rule. 6.1. Fault coding The network has r input nodes, i.e., as many as there are columns in the input pattern matrix X derived with any of the different feature extraction procedures previously described. The output nodes of the network are as many as the number of circuit components n. We construct the s input-output patterns that will be used to train the neural network for the diagnosis of the circuit as follows. Each pattern is given by a pair (~x i ; ~y i ). The vector ~x i is the th row of matrix X while the associated vector ~y i is defined as follows: component k is not faulty during the th acquisition; component k is faulty during the th acquisition. This general scheme must be altered to take into account undistinguishable faults. Topologically undistinguishable faults are easy to deal with. From an inspection of the circuit a list of all sets of parallel components is made. Then, a single short circuit fault acquisition for each set C i of parallel components is considered. There will be a single training pattern (~x such a fault. The vector ~y i is such that y i for all k 2 C i , while all other components have a 0 value. A dual procedure takes care of sets of series components. Two faults i and i 0 are behaviorally undistinguishable if where oe is the threshold introduced in section 5.1. A fault i is behaviorally undetectable if the condition k1 oe is satisfied for all input vectors ~x 0 obtained in the faulty free condition. We combine the patterns of behaviorally undistinguishable faults (as we did for topologically undistinguishable faults) and remove from the training set the patterns associated to undetectable faults. The fault coding here described is different from the one presented in [16], that defined the vector ~y i as follows: component k is short circuited during the th acquisition; 0:5 if component k is not faulty during the th acquisition; component k is open circuited during the th acquisition. The new coding gives sharper identification of the faulty component and is more robust when diagnosing multiple faults because the values of interest (0 and 1) are obtained by "pushing" the sigmoid function toward saturation. Note also that there is a difference with the coding in [31] where each output node is associated to a catastrophic fault and not to a component. Once the net has been trained, it may be used to perform the diagnosis of the circuit. The net must be given the features extracted from the measured test point voltages as input vector ~x. The net will produce an output vector ~y; a value of y(k) close to 1 will pinpoint a fault of component k; a value close to 0 will denote that the component is correctly functioning. Although the net has been trained with the results of single fault acquisitions, it is potentially able to diagnose multiple faults. In this case, two or more elements of ~y will be close to 1. 6.2. Network structure and training The basic architecture we used consists of a three layers backpropagation network. Since the input patterns have been preprocessed to eliminate undistinguishable faults, and thus they are sepa- rable, we are sure that eventually there will be a network capable of correctly learning all patterns. We use early-stopping [4] to avoid overfitting. This consists in measuring, during the training, the error with respect to an independent set of patterns, called validation set , and in stopping the training when this error reaches a minimum. Caruana [5] has shown that if early-stopping is used the number of nodes in the hidden layer may vary without appreciably affecting the performance of a neural network, provided it is sufficiently large. The results of our simulations, not reported in this paper, seem to confirm this general rule. The validation set used for the stopping is independent from the training set. We construct it by performing a new set of PSpice simulations (one for each fault) randomly changing the parameter values of the components within their tolerance range and by adding to the voltage signals of the test points a noise whose magnitude is equivalent to the measurement noise observed during the p fault-free acquisitions. 7. Combining different diagnosis In the diagnosis of circuits, we have underlined the importance of using more than one power supply. In fact, it is often the case that a given supply can only lead to the detection of a particular subset Neural Diagnosis for Analog Circuits 11 of all possible faults. The use of different supplies leads to the use of several neural networks N i , each of which produces its own candidate set A i . The final diagnosis must be computed combining these sets of candidates. The combination of neural networks is a problem that has been discussed in the literature and is reviewed in [32]. In particular, since we use neural networks that are all trained on the same task, our approach falls into the ensemble (or commit- It is clear that the "union" of two sets of candidates magnifies the influence of false alarms, while the "intersection" can be used to filter false alarms at the risk of removing some faulty components from the diagnosis. Keeping this in mind, we propose two different ensemble algorithms. Let us first give the following definitions. For each candidate k let v(k) be the number of votes it receives, i.e., the number of nets that consider k malfunctioning, and let v(k). We consider all non-empty intersections of v candidate sets; assume there are ff of such intersections and denote them R u , with We also define u the index of the intersection R u with the smallest cardinality (should there be more than one such intersection we randomly pick one). Algorithm 1 The first algorithm considers as faulty all those candidates that have received the highest number of votes. The corresponding diagnosis is: ff R u Algorithm 2 The second algorithm considers as faulty all those candidates that have received the highest number of votes and that belong to the intersection with the smallest cardinality. By considering only the smallest intersection we hope to filter out some false alarms. The corresponding diagnosis is: An example is shown in Figure 3. Here D Note that these algorithms do not give different weight to the candidate sets of each network, but simply perform boolean operations on these sets. We are currently investigating the possibility of associating to each candidate set a different weight, depending on how the network has learned to recognize the single fault on each component that belongs to the candidate set. 8. Experimental results We discuss the results obtained by the different diagnostic strategies presented in this paper. Two circuits are studied: a DC motor drive board, and an astable multivibrator. Training As discussed above, we use early stop- ping, hence we need both a training and a validation set of patterns. The training patterns corresponding to each single fault condition are constructed using a PSpice model of the circuit. This choice gives patterns corresponding to a circuit where the component parameters have nominal values and the voltage signals in each test point are noise free. The validation set is constructed by performing a new set of PSpice simulations where parameter tolerance and measurement noise are introduced. Testing During the test phase, we consider a real circuit and the different faults are implemented by manually shortcircuiting or opening each component terminals. The circuit measurements are col- A 3 Fig. 3. Example of diagnosis combination. lected through a National Instrument Corporation acquisition board. Thus the test patterns are determined independently of the training patterns. Furthermore, the test patterns are affected by measurement noise and by the error due to the parameter tolerance of the circuit components. When diagnosing a circuit, we observe the net-work output corresponding to the input pattern derived from the measurements. Let us recall that the network output layer has as many nodes as there are components. During the training phase we have coded a fault on component k assigning a value 1 to the corresponding output node, while a value 0 was assigned to the output node of a fault-free component. In general, during the test phase the value of each output node may take any value between 0 and 1. A value close to 0 (1) of an output node will be interpreted as the absence (presence) of a fault on the corresponding component. Threshold values need to be set to discriminate between these two cases. Let vmax be the maximum value of all output nodes. If vmax ! 0:2 we consider the circuit as fault-free and the candidate set will be empty. If vmax 0:2 we consider the circuit as faulty, and the candidate set will contains all components whose corresponding output node has a value greater than 0:5v max . 8.1. DC motor drive board We present the results obtained diagnosing the circuit in Figure 4, part of a DC motor drive. The same circuit has also been diagnosed in [8, 13, 14, 15, 16]. In the figure, the test points are marked by numbers within circles, while the are labeled by numbers in square brackets. Training There are 70 single faults to consider on this circuit. In fact, the circuit is composed of 36 components but only one fault is considered for each of the two operational amplifiers. Thus, the overall training set should consist of 76 training patterns - the additional six being obtained by acquisitions of the circuit behavior in absence of fault. The following sets contain topologically undistinguishable faults: 28s, 29sg, 17og, 30/31o =f30o, 31og. Here 10s represents a short circuit fault on component 10, 16o represents an open circuit fault on component 16, etc. Thus, the training set is reduced to 62+6 patterns by combining the conflicting patterns as discussed in Section 6.1. We have used three different voltage supplies and thus three different networks. 1. The first network N 1 is trained with patterns acquired when the circuit has close to nominal voltage supplies: V 1 \Gamma12 (V). Fourier The number of significant frequency components is This gives rise to columns in the input matrix X 0 , that are reduced to r = 14 in the matrix X . The set of behaviorally undetectable faults for this net is: f3o, 6o, 11o,12s, 15o, 23o, 24o, 25s, 27o, 29o, 31s, 32s, 33o, 34og. The sets of behaviorally undistinguishable faults are: f4o, 6sg, f21o, 27/28/29s, 30sg, f34/35s, 36og, f 35o, 36sg. Wavelets The number of significant wavelets is 8. This gives rise to m \Delta columns in the input matrix X 0 , that are reduced to r = 15 in the matrix X . The set of behaviorally undetectable faults for this net is: f3o, 6o, 11o, 12s, 15o, 23o, 24o, 25s, 27o, 29o, 31s, 32s, 33o, 34og. The sets of behaviorally undistinguishable faults are: f21o, 27/28/29s, 30sg, f34/35s, 36og, f35o, 36sg. PCA Assuming a threshold 0:999, the number of dominant eigenvalues (i.e., the number of columns of the X matrix) is 25. The set of behaviorally undetectable faults for this net is: f3o, 6o, 11o, 12s, 15o, 23o, 24o, 25s, 27o, 29o, 31s, 32s, 33o, 34og. Neural Diagnosis for Analog Circuits 13 Fig. 4. DC motor drive board. The sets of behaviorally undistinguishable faults are: f4o, 6sg, f21o, 27/28/29s, 30sg, f34/35s, 36og, f35o, 36sg. 2. The second network N 2 is trained with patterns acquired when the circuit has far from nominal periodic voltage supplies: are zero-mean square waves with 160 Hz frequency, 4 V peak-to- Fourier The number of significant frequency components is 8. This gives rise to columns in the input matrix X 0 , that are reduced to r = 15 in the matrix X . The set of behaviorally undetectable faults for this net is: f10o, 11o, 12s, 21o, 22o, 23o, 24o, 25s, 27/28/29s, 28o, 29o, 30s, 31s, 32sg. The sets of behaviorally undistinguishable faults are: f1o, 3sg, f4o, 6sg, f34/35s, 36og. Wavelets The number of significant wavelets is 9. This gives rise to m \Delta columns in the input matrix X 0 , that are reduced to in the matrix X . The set of behaviorally undetectable faults for this net is: f8 up , 10o, 11o, 12s, 14o, 17s, 21o, 22o, 23o, 24o, 25s, 27/28/29s, 28o, 29o, 30s, 31s, 32sg. The sets of behaviorally undistinguishable faults are: f4o, 6sg, f14/15s, 16sg, f18o, 19og, f34/35s, 36og. PCA Assuming a threshold 0:999, the number of dominant eigenvalues is The set of behaviorally undetectable faults for this net is: f10o, 11o, 12s, 21o, 22o, 23o, 24o, 25s, 27/28/29s, 28o, 29o, 30s, 31s, 32s, 34/35s, 36og. The set of behaviorally undistinguishable faults is: f1o, 3sg. 3. The third network N 3 is trained with patterns acquired when the circuit has step voltage supplies: (V). Fourier The number of significant frequency components is 8. This gives rise to columns in the input matrix X 0 , 14 Fanni, Giua, Marchesi and Montisci that are reduced to r = 15 in the matrix X . The set of behaviorally undetectable faults for this net is: f10o, 12s, 14o, 21o, 23o, 25s, 27o, 28o, 29o, 31s, 32s, 34og. The sets of behaviorally undistinguishable faults are: f1o, 3sg, f4o, 6sg, f22o, 27/28/29s, 30sg. Wavelets The number of significant wavelets is This gives rise to m \Delta columns in the input matrix X 0 , that are reduced to r = 17 in the matrix X . The set of behaviorally undetectable faults for this net is:f10o, 12s, 21o, 23o, 25s, 27o, 28o, 29o, 31s, 32s, 34og. The sets of behaviorally undistinguishable faults are: f1o, 3sg, f4o, 6sg, f22o, 27/28/29s, 30sg. PCA Assuming a threshold 0:999, the number of dominant eigenvalues is The set of behaviorally undetectable faults for this net is: f10o, 12s, 21o, 23o, 25s, 27o, 28o, 29o, 31s, 32sg. The sets of behaviorally undistinguishable faults are: f1o, 3sg, f4o, 6sg, f22o, 27/28/29s, 30sg. Testing We are now ready to study the performance of the neural diagnostic systems previously constructed. In the initial phase, we test the systems on a fault-free circuit. We observed that when diagnosing a real circuit in absence of faults, all networks correctly identify this behavior, in the sense that all output nodes have a value less than the assigned threshold of 0:2 and thus the candidate set is always empty. In a second phase, we consider faulty circuits. Table 1 compares the performance (in percent) of the different systems. The first columns of the table shows the diagnosis of N 1 , N 2 , and N 3 and the diagnosis obtained combining the candidate sets of the three nets with Algorithm 1 and Algorithm 2, using Fourier, Wavelets, and PCA, re- spectively. The last two columns show the results obtained combining the candidate sets of the nine nets (three for each feature extraction technique) with Algorithm 1 and Algorithm 2. We consider a fault correctly diagnosed if the candidate set of the net contains a subset of the components associated to this fault, taking into account topologically undistinguishable fault classes. Let us consider some examples in the circuit of Figure 4. The fault 16o belongs to the topologically indistinguishable fault class 16/17o; we say that it is correctly identified if the candidate set is either f16g or f17g or f16, 17g. The fault 16s is correctly identified if the candidate set is f16g. Single faults The first row block of Table 1 shows the diagnosis of the 62 possible single faults. There are three different classes of diagnosis. Class A 1 Faults correctly diagnosed. Undistinguishable faults: these are the faults that we have classified as behaviorally undistinguishable during the training. As an example, in net N 1 with Fourier, we have identified 9 undistinguishable faults, i.e., 14% of the total 62 faults. Class C 1 Undetected faults: these are the faults that we have classified as behaviorally undetectable during the training. Double faults The second row block of Table 1 shows the performance of the different systems when diagnosing double faults. Each double fault consists in the simultaneous presence of two faulty components. Note that not all possible pairs of single faults constitute a double fault: e.g., a bipolar component cannot be simultaneously open- and short- circuited. We have considered a sample of 168 different double faults randomly chosen from the total population. This sample was large enough to satisfy the 2 test for the six different classes of diagnosis. These are the classes considered. Class A 2 Both faults correctly diagnosed. Only one fault correctly diagnosed. Class C 2 At least one fault correctly diagnosed with one or two false alarms. Class D 2 At least one fault correctly diagnosed with more than two false alarms. Only false alarms. Neural Diagnosis for Analog Circuits 15 Triple faults The third row block of Table 1 shows the performance of the different systems when diagnosing triple faults. We have considered a sample of 181 different faults randomly chosen out of the total population. These are the classes considered. Class A 3 All three faults correctly diagnosed. Only one or two faults correctly diagnosed Class C 3 At least one fault correctly diagnosed with one or two false alarms. Class D 3 At least one fault correctly diagnosed with more than two false alarms. Class F 3 Only false alarms. Discussion In the case of multiple faults, we consider correct all diagnoses in class A and in class B. In fact, starting from class B we may use an incremental repair procedure, substituting the faulty components one by one. Diagnosis in class C may also be useful. From the table it can be seen that the use of several networks improves the system performance provided that a good procedure is used to combine the results of the networks. In particular, Algorithm 1 and Algorithm 2 give the same results when diagnosing: (a) single faults; (b) multiple faults using a system composed of many nets in parallel. When diagnosis multiple faults, if the system is composed by a small number of neural nets Algorithm 2 performs better because it exalts the filtering effect of the intersection operator, reducing the number of diagnoses in class A but increasing the total number of diagnoses in class A+B. All three feature extraction techniques give comparable results. PCA performs better than the other two when diagnosing single and double faults, but seems to be less robust when diagnosing three simultaneous faults. Unlike Fourier and Wavelets, PCA requires less data preprocessing in the feature extraction phase, as discussed in Section 5. 8.2. Astable multivibrator Dague et al. in [10] remarked that oscillators are difficult to diagnose because most faults cause the same type of symptoms. This is exactly the case in which a proper choice of the power supplies can improve the diagnosability of the circuit. They proposed using an external "stimulation" and showed the results obtained using their diagnostic expert system on the astable multivibrator shown in Figure 5. In this section we present results obtained using our diagnostic system on the same circuit. Training We chose test-points in the nodes labeled 1 and 2 in the figure. The number of components is single faults have been considered. In fact, we consider 6 faults for each transistor: short circuit between base and emitter (QBEs), base and collector (QBCs),collector and emitter (QCEs); open circuit on the base (QBo), collector (QCo), and emitter (QEo). The circuit does not contain topologically undistinguishable faults. To be able to compare the results of our diagnostic system with the system developed by Dague, we used the same voltage supply proposed in [10]. It consists of the superposition of the nominal supply PS (a continuous voltage signal of +5V) and of an external stimulation EP (a voltage pulse of 10V amplitude, applied in and lasting 1s). Since we consider a unique sup- ply, we use a single neural network for each feature extraction. We used a PSpice model of the oscillator to collect the training and validation patterns for all faulty conditions, as previously described. The circuit has been studied in transient behaviour and the voltage signals in each test point have been collected in the interval 0:9 \Xi 40s with a sampling interval of t 0:1s. This gives rise to a circuit image before feature extraction composed of for each test point. We have used all different feature extraction techniques described in Section 5. Fourier The number of significant frequency components is This gives rise to m \Delta columns in the input matrix Table 1. Diagnosis of the circuit in Figure 4 (in percent). Fourier Wavelets PCA F+W+P Single faults (62 fault cases) Double faults (168 fault cases) Triple faults (181 fault cases) A3 28 19 21 62 74 23 21 23 are reduced to in the matrix X . There are no behaviorally undetectable faults. The sets of behaviorally undistinguishable faults are: fRC2s, Q2CEsg, fRB2s, Q2BEsg. Wavelets The number of significant wavelets is This gives rise to m \Delta columns in the input matrix X 0 , that are reduced to in the matrix X . There are no behaviorally undetectable faults. The sets of behaviorally undistinguishable PS RI Fig. 5. Astable multivibrator. PS is the nominal power generates the stimulation voltage pulse. faults are: fC1o, RC2s, RB2s, Q2CEs, Q2BEsg. PCA Assuming a threshold 0:999, the number of dominant eigenvalues (i.e., the number of columns of the X matrix) is There are no behaviorally undetectable faults. The sets of behaviorally undistinguishable faults are: fRC2s, Q2CEsg, fRB2s, Q2BEsg. Sampling The Sampling feature extraction retains samples (out of a total of for each test point spaced with an hyperbolic law so as to have more samples during the initial phase of the transient and only a few as the circuit reaches the steady state. For Table 2. Diagnosis of the circuit in Figure 5 (in percent). Fourier Wavelets PCA Sampling Single faults (24 fault cases) Double faults (140 fault cases) 28 8 23 41 Neural Diagnosis for Analog Circuits 17 Table 3. A comparison between the diagnosis of the circuit in Figure 5 done with Dague's Expert System and with Neural Network with Sampling. Defect Expert System Neural Network (Dague et al.) with Sampling RC1s RC1, Q1, RC1 fC2g \Theta fPS, CX, C1, EP, Q2, RB1, RB2, RC2, RIg RB1s (*) Q1 RB1 double candidates C1 Q1CEs Q1, C2, Q1 fRC1g \Theta fCX, C1, EP, Q2, RB2, RC2, RIg Q1BEs Q1, C1, RB1 Q1 fC2g \Theta fPS, EP, Q2, RB2, RC2, RIg RC1o RC1, Q1, C2 RC1 RB1o RB1, Q1, C1 RB1 C1o C1, Q2, RC2 fC1, RC2g Q1Eo Q1 Q1 Q1Bo Q1 Q1 Q1Co Q1, C1, RB1, C2, RC1 Q1 (*) Note that a short-circuit on RB1 induces destruction of Q1. Then the ae-th retained sample is the ae -th sample, as shown in Figure 6. There are no behaviorally undetectable faults. The sets of behaviorally undistinguishable faults are: fC1o, RC2og, fRC2s, Q2CEsg, fRB2s, Q2BEsg. Testing In a first phase, we test the systems on a fault-free circuit. We observed that when diagnos- tr r Fig. 6. Sampling of the total measurements. ing a real circuit in absence of faults, all networks correctly identify this behavior. In a second phase, we consider faulty circuits. Table 2 compares the performances (in percent) of the different neural diagnostic systems. The classes of diagnosis are the same defined in the previous example. The first row block of Table 2 shows the diagnosis of the 24 single faults considered. The second row block of Table 2 shows the diagnosis of the 140 double faults considered. Note that in this case the total number of possible double faults is 240. We can see that in this particular case Sampling appears to be the most effective feature extraction technique (and it is also the easiest to implement). As remarked before, however, it is a viable solution only because we have a small number of test points (two in this example) and the circuit has a short transient. Comparison with Dague's Expert System In Table 3 we compare the results obtained using our diagnostic system with sampling feature extraction (Neural Network with Sampling) with the results obtained by Dague's expert system. Note that although only 12 faults are considered by Dague, our diagnostic system has been trained to recognize all 24 single faults. We observe that the neural network has been able to correctly classify 11 faults in class A and only 1 fault (C1o) in class B. The expert system, on the contrary, can very rarely correctly identify the faulty component (only 3 diagnosis in class A including the diagnosis of RB1s). The results show that the neural network performs much better than the expert system. This is a consequence of its ability to exploit the information on each single-fault behavior of that particular circuit and to generalize. This information is not taken into account by the expert system, that reasons on more abstract principles. 9. Conclusions We have shown how a neural network, trained to recognize catastrophic single faults, may be used to diagnose multiple faults on analog circuits. In general we observe that the network is almost always able to learn and recall the single fault patterns presented during the training. Multiple faults on two and three components may also be diagnosed, although less sharply than in the single fault case, due to the presence of false alarms. In most cases, however, the network is able to detect at least one of the malfunctioning components. Thus one may use an incremental repair proce- dure, substituting the faulty components one by one. We consider several different power supplies in order to detect those faults that do not modify the circuit behavior under nominal supplies. We use several neural networks "in parallel", one for each different supply configuration. Each network is specialized in detecting a given set of faults. Thus, it is not necessary to force a network to recognize a fault that is more easily detected by another one. The use of different networks, leads to the problem of composing different sets of candidates into a single diagnosis. We showed that a suitable choice of the composition algorithm may dramatically improve the system performance, especially when diagnosing multiple faults. We compared the results obtained by our system when using different feature extraction tech- niques. In fact, the performance of the diagnostic system is noticeably affected by the choice of features that we consider as representative of the device behavior. Although we have only presented two simple examples of diagnosis, extensive experiments convinced us that this approach is fairly general and that it gives better results than other diagnostic systems, such as expert systems, whenever it can be applied. --R Ghaloum S. Salama A. A Neural Network Approach for Identification and Fault Diagnosis of Dynamic Systems. Neural Networks for Pattern Recogni- tion Extra capacity rarely hurts generalization if yo u use early stopping. SAR Image Segmentation Using Textural Information and Neural Classifiers. A Spectrum of Logical Definitions of Model-Based Diagnosis Wavelet Analysis for Diagnostic Problems. Extracting Comprehensible Models from Trained Neural Networks. Luciani P. Diagnosing Multiple Faults. Hart P. Qualitative Dynamic Diagnosis of Circuits. A Multiple Neural Network Diagnostic System for Analog Circuits Based on Fourier Transforms. Diagnosis of Electrical Circuits Using Neural Networks and Principal Components Analysis. Neural Networks for Multiple Fault Diagnosis in Analog Circuits. Yang Y. Introduction to the Theory of Neural Computation. A DC Approach for Analogue Fault Dictionary Determination. Fundamentals of Digital Image Processing. Murphy J. Dean J. Selected Papers on Analog Fault Diagno- sis Testing and Diagnosis of Analog Circuits and Systems. A theory for multiresolution signal de- composition: the wavelet representation Stategy for Diagnosis. Saeks R. IEEE SP Magazine A Neural Network Approach to Fault Location in Non Linear DC Circuits. On combining Artificial Neural Nets. Artificial Neural Systems. Burris D. Printed Circuit Board Diagnosis Using Artificial Neural Networks and Circuit Magnetic Fields. Linear Circuit Fault Diagnosis Using Neuromorphic Analyzers. A Guide to Neural Computing Appli- cations Sutton J. Limb P. --TR --CTR Francesca Cau , Alessandra Fanni , Augusto Montisci , Pietro Testoni , Mariangela Usai, A signal-processing tool for non-destructive testing of inaccessible pipes, Engineering Applications of Artificial Intelligence, v.19 n.7, p.753-760, October, 2006 Barbara Cannas , Francesca Cau , Alessandra Fanni , Augusto Montisci , Pietro Testoni , Mariangela Usai, Neural NDT by means of reflected longitudinal and torsional waves modes in long and inaccessible pipes, Proceedings of the 5th WSEAS/IASME International Conference on Systems Theory and Scientific Computation, p.94-102, September 15-17, 2005, Malta
multiple fault diagnosis;analog circuits;neural networks
590909
Knowledge Extraction from Transducer Neural Networks.
Previously neural networks have shown interesting performance results for tasks such as classification, but they still suffer from an insufficient focus on the structure of the knowledge represented therein. In this paper, we analyze various knowledge extraction techniques in detail and we develop new transducer extraction techniques for the interpretation of recurrent neural network learning. First, we provide an overview of different possibilities to express structured knowledge using neural networks. Then, we analyze a type of recurrent network rigorously, applying a broad range of different techniques. We argue that analysis techniques, such as weight analysis using Hinton diagrams, hierarchical cluster analysis, and principal component analysis may be useful for providing certain views on the underlying knowledge. However, we demonstrate that these techniques are too static and too low-level for interpreting recurrent network classifications. The contribution of this paper is a particularly broad analysis of knowledge extraction techniques. Furthermore, we propose dynamic learning analysis and transducer extraction as two new dynamic interpretation techniques. Dynamic learning analysis provides a better understanding of how the network learns, while transducer extraction provides a better understanding of what the network represents.
Introduction There has been a lot of interest lately in knowledge structures and their representation in articial neural networks [Holldobler, 1990, Kurfe, 1991, Sperduti et al., 1995, Wermter, 1995, Hallam, 1995, Medsker, 1995, Sun, 1995, Wermter et al., 1996, Elman et al., 1996, Craven, 1996, Wermter, 1999]. Articial neural networks (or connectionist networks) have already demonstrated interesting learning results for various classication tasks. However, it continues to be very di-cult to understand the underlying representations within the connectionist networks which lead to this perfor- mance. A better understanding of the connectionist representations learned is not only important for improving the credibility of a computational technique, but also for improving the net-work performance and the integration possibilities with symbolic representations. Several attempts have been made to interpret connectionist networks, focusing on feedforward networks in particular [Andrews and Diederich, 1996, Abe et al., 1993, Shavlik, 1994]. For in- stance, visualizations of internal activations or weight strengths can be used to get an impression of the internal knowledge [Hinton, 1986, Gorman and Sejnowski, 1988]. Some eort has also been made to reduce the network size in order to simplify the knowledge expressed therein by elimi- 28 Wermter nating very small weights. Furthermore, groups of similar weights can be replaced with their average strength [Shavlik, 1994]. In addition, techniques such as hierarchical cluster analysis have been used to interpret connectionist networks. Never- theless, often the interpretation of the dynamics of the learning process and the underlying knowledge has been neglected, especially in the case of dynamic recurrent neural networks. The interpretation of recurrent networks is more di-cult than that of non-recurrent feedforward networks, since the previous context in recurrent networks has an important dynamic in uence within these networks. The internal states in recurrent networks do not only depend on the input but also on the internal state of the local memory based on previous inputs [Elman, 1995, Giles and Omlin, 1993, Omlin and Giles, 1996]. For this reason, to date the focus has been primarily on smaller recurrent networks and articially generated data. For instance, an interesting current approach interprets the training of a SRN network that has two input, two output and two internal elements in learning the sequence a n b n [Wiles and Elman, 1996]. It has been discovered that the net-work behaved like a spiral which moved to and from a x point. Whereas this seems to be a plausible interpretation of the behavior of recurrent networks trained for the learning of the sequences a n b n , dierent interpretations are required when we move to dierent tasks and data sets closer to real-world scenarios. In the past, we have developed a large \real- world" system for spoken language analysis which makes extensive use of SRN networks [Wermter and Weber, 1997, Wermter and Meurer, 1997]. The spoken input is recognized by a speech recognizer and analyzed at the syntactic, semantic and dialog levels based on an incremental analy- sis, parallel syntactic and semantic interpretation, and robust processing of errors. To date, however it is not yet possible to focus on the interpretation of the learning process and the interpretation of the connectionist knowledge. In this paper, we are primarily concerned with a detailed interpretation of the learning behavior as well as a symbolic interpretation of the learned knowledge after training. In order to carry out such a detailed analysis we will concentrate on a syntactic transformation task as a representative task for our large-scale speech/language system. The task for the recurrent network is to process sentences and associate their syntactic classes at the phrasal level, e.g. noun phrase, prepositional phrase etc. Using this task, we analyze a recurrent neural network using many dierent techniques. We have structured the paper as follows. First, we introduce our representative syntactic transformation task. Then, we dene and illustrate a) dynamic learning analysis, b) weight analysis, c) hierarchical activation analysis, d) component activation analysis, and e) transducer extraction. We rigorously compare these techniques on the same net-work and the same data set and argue that these dierent techniques provide mutually complementary interpretations. The contribution of this paper is a particularly broad and concrete analysis of the knowledge extraction process which has not been done before. Furthermore, we propose dynamic learning analysis and transducer extraction as two new interpretation techniques. Dynamic learning analysis provides a better understanding of how the network learns while transducer extraction provides a better understanding of what the network represents. 2. Extracting structured knowledge using syntactic analysis task In order to examine a number of dierent techniques for extracting structured knowledge from connectionist networks in a rigorous manner, we will focus on a particular task. In our spoken language environment [Wermter and Lochel, 1996, Wermter and Weber, 1997, Wermter and Meurer, 1997], we have trained many variations of SRN networks [Elman, 1991] with many sentences using various corpora of several thousand words each. Based on a corpus of sentences from the domain of scheduling appointments (2355 words), table 1 summarizes the accuracy of label assignment on the unknown test set. The related experiments and results have been reported elsewhere in detail [Wermter and Lochel, 1996, Wermter and Weber, 1997, Wermter and Meurer, 1997, Wermter, 1998]. Here we just want to illustrate the real-world net-work performance in table 1. The focus, how- Knowledge Extraction from Transducer Neural Networks 29 ever, is on an analysis of the process of extracting explicit knowledge from implicitly learned knowl- edge. In this paper, we concentrate on syntactic phrasal assignment (marked by *) in table 1. Table 1. Performance of some networks on the test set of the appointment scheduling corpus Task Accuracy on test set Basic syntactic disambiguation 89% Basic semantic disambiguation 86% Syntactic phrasal assignment* 84% Semantic phrasal assignment 83% Dialog act assignment 79% Word repair detection 94% Phrase repair detection 98% To demonstrate this process of knowledge ex- traction, we will here use 15 of these sentences (containing 76 words) from the domain of appointment scheduling. For illustration purposes, we concentrate on the learning of a syntactic phrasal assignment task where a sequence of basic categories of words is associated with a sequence of abstract syntactic categories. The actually occurring syntactic basic categories are noun (n), verb (v), adverb (a), adjective (j), preposition (r), determiner (d) and pronoun (u). The abstract phrasal categories are noun group (ng), verb group (vg), and prepositional group (pg). The task of the recurrent network is to learn to assign phrasal categories on the basis of basic syntactic categories in order to support a robust at understanding of spontaneously spoken language. Below, we show some example utterances from the corpus, together with the syntactic categories at the basic and the phrasal level. 1. I (u ! ng) thought (v ! vg) in (r ! pg) the 2. That (u ! ng) is (v ! vg) the (d ! ng) Thursday (n ! ng) after (r ! pg) Easter (n Based on these seven basic syntactic and three phrasal syntactic categories, we use an SRN net-work with seven input units, three internal units and three output units (the networks in the actual system contain more categories and have been trained with several thousand words, but for illustration purposes we restrict ourselves to this smaller network). The learning rate was 0.05 and momentum 0.9. The weight updates were performed incrementally after each training pattern. Each training pattern consisted of the basic syntactic category at the input layer and the abstract phrasal category at the output layer. Figure 1 shows a simplied example of such a recurrent network for the task of syntactic phrase assignment. Output Input Context- layer AAAA AAAA AAAA Noun group Prepositional group Verb group Pronoun Noun Adjective Verb Adverb Preposition Determiner Fig. 1. Recurrent network for knowledge extraction for syntactic phrase assignment The activation of an output element O j (t) at time t in SRN networks is computed on the basis of the weighted activation H i (t) of all incoming connections limited by the logistic function f . O The activation of an element on the internal layer H l (t) is computed in a similar manner. Here the activation of the input layer I k (t) at time t is used as the activation of the internal layer at the previous time step t 1. Wermter 3. Dynamic learning analysis: knowledge structuring during lazy learning In the past, most work on knowledge structures and connectionist networks has focused on static connectionist network representations. However, important insights can be gained by examining how certain knowledge structures emerge and develop time before a certain task is learned completely. Frequently, the interpretation of the learning behavior is just demonstrated by means of the learning curve of the overall error reduction over time. However, the learning curve is just the rst step in a more detailed analysis and can only provide preliminary hints about the performance of a network over the training time. Figure 2 shows the learning curve with the overall sum squared error over time. patterns 50000 100000 150000 2000000.20.6Error Fig. 2. Learning curve for syntactic phrasal assignment The learning curve shows that the speed of learning diers substantially over time. Further- more, we can see dierent stages during the learning process. In the beginning, learning proceeds fast, but later learning is slower and it takes longer to make signicant improvements. For instance, between 70000 and 140000 it seems that learning is about to nish before there is a nal signicant improvement. We will now examine how the network reaches its performance. We start the analysis directly after the random initialization of the weights. This is the state before learning starts. We want to give an overview of the overall performance for all input patterns at dierent time steps. To this ef- fect, we show the error for each of the 76 patterns of the demonstration set at dierent time steps. Figure 3 shows the individual error for each of the 76 patterns before training.Individual patterns Fig. 3. Performance for individual patterns before learn- ing Based on the random initialization, all patterns show a relatively high error. At this point, it is to be expected that the values of each output element dier from the desired value 0 or 1 by 0:5. Therefore the expected error for an individual pattern for three output elements is expected error value is conrmed in this gure. As shown in gure 2, the error decreases quickly at the start of the training. The state after 100 patterns of the training set is shown in gure 4. First, we can observe that after 100 training pat- terns, the error for some of the 76 patterns shown could be reduced signicantly. Other patterns still show a high error. Obviously, the network has started to learn patterns selectively. Knowledge Extraction from Transducer Neural Networks 31 Individual patterns patterns other patterns Fig. 4. Performance for individual patterns after 100 training patterns A more detailed analysis revealed that the patterns with a lower error are exactly those patterns which belong to the noun group NG. After only 100 patterns, the network has recognized that the global error can be minimized signicantly by focusing on the NG patterns, since these patterns occur more frequently than, for instance, prepositional groups or verb groups. Therefore, at rst the network has learned a constant mapping of all patterns to the noun group, since this reduces the overall error most at this stage. This explains why certain patterns in gure 4 still exhibit a high error and others a low error. The patterns with a low error are exactly the patterns which have been classied correctly as noun groups. Figure 5 shows the detailed performance after patterns. After the network has learned a constant mapping to NG, we can observe that the performance for the NG patterns has improved even further. However, we also observe that V G patterns have been learned. A more detailed analysis of the output preferences reveals that at this stage, in addition to all NG, all V G patterns have also been learned correctly. This is also demonstrated in gure 5. All the remaining error patterns at this stage are those patterns which should belong to a prepositional group PG but which are still categorized as noun groups NG. All NG patterns and all V G patterns are classied correctly. After the network has learned the most frequent NG patterns, the second most frequent V G patterns are learned. Thus, one could state that the network pursues a conservative lazy learning strategy and learns frequently occurring and simple regularities rst. Individual patterns patterns patterns patterns Fig. 5. Performance for individual patterns after 600 training patterns Afterwards, the network attempts to improve all patterns, especially the remaining patterns for prepositional groups PG. The occurring nouns, pronouns, determiners, and adjectives can either be part of PG patterns or NG patterns. In order to resolve this potential for ambiguity, previous context must be used to learn the correct class assignment. Again, we have an example of the conservative lazy learning strategy of the network, Individual patterns 3 exceptions of PG patterns patterns NG patterns, VG patterns Fig. 6. Performance for individual patterns after 3000 training patterns Wermter since at rst the network has learned patterns which do not need previous context knowledge for the category assignment. Only after the simple non-context-dependent category assignments have been learned, are those patterns learned which require the context of previous pattern assignments. The state of the network after 3000 patterns is shown in gure 6. All patterns are classied correctly with the exception of three. Comparing g- ures 5 and 6, the remaining error for the individual patterns could be reduced signicantly. For the learning of the PG patterns, it was necessary for the network to integrate the local preceding con- text. After 150000 patterns all regularities have been learned as shown in gure 7. In comparison with gure 6 we point out the smaller scaling of the vertical axis. At this stage all patterns have been learned, even though there are dierences between the error rates of individual patterns. In order to reach this 100% correctness on the training set, it may be necessary to give up a reasonably good state at a certain stage in order to reach an even better stage later. This is also re ected in the global learning curve in gure 2. Individual patterns all NG patterns,VG patterns, PG patterns correct Fig. 7. Performance for individual patterns after 150000 training patterns In general, the network pursues a conservative lazy learning strategy. First, simple and frequently occurring generalizations of one category are learned. Only when the network cannot minimize its error signicantly any more, are other frequently occurring categories integrated. Fur- thermore, only when all those patterns have been learned that do not require previous local context, are those patterns learned that require context for the correct category assignment of otherwise ambiguous input. Finally, any remaining exceptions are learned. During this conservative learning process it may be possible that the overall error increases brie y in order to reach a better overall state later. 4. Weight analysis for knowledge extrac- tion Visualizations of internal weight strengths can be used to get an impression of the internal knowl- edge. In our experiments, the training set was learned correctly after 150000 patterns and this is where we start our analysis. We start with such a weight analysis since weights provide the lowest level of interpretation of a connectionist represen- tation. Figure 8 shows the weights of the network for three dierent time steps. It is illustrated how the weights change over time during learning. In this gure the identiers of the source connectionist elements are shown horizontally and the identiers of the goal connectionist elements are shown vertically. We start with the horizontal axis. From left to right, we can see the weights from the threshold element (S), from the input connectionist elements for the syntactic basic categories (n, j, v, a, r, u, d), from the three internal elements and from the three context elements c3). In the vertical axis from top to bottom, we see the weights to the three internal elements and to the output elements representing the abstract syntactic categories (VG, NG, PG). Knowledge Extraction from Transducer Neural Networks 33 after 100 patterns after 600 patterns after 150000 patterns Fig. 8. Weight analysis at the beginning of training (100 patterns), during training (600 patterns) and after training (150000 patterns) White boxes represent positive weights, black boxes negative weights. The size of the boxes corresponds to the size of the weights. The copy connections from the internal layer to the context layer are not changed. Therefore, they are not shown since they are always equal to 1. We start with the analysis of the rst third of gure 8. After random initialization, this rst third shows all weights of the network after 100 patterns. At this point, all NG patterns can be classied correctly, but no other patterns have been learned yet. The network has learned a constant output in order to reduce the overall error as much as possible. We can see in gure 8 why the network produced this constant NG class. 34 Wermter We can see that the weights from the input elements of the syntactic basic categories (n, j, v, a, u, d) to the internal elements are relatively small and similar. The same holds for the weights from the context elements (c1, c2, c3) to the internal elements. This is due to the random initialization at the beginning of the train- ing. The weights from the internal elements to the output elements of the abstract syntactic categories are negative for V G and PG; those from the internal elements to NG are close to 0. This is the reason why the network produces constantly the NG category at this stage. Now we focus on the state of the network after presenting 600 patterns, also shown in gure 8. At this point, all NG and all V G patterns are assigned correctly. This is also re ected in the weights. We observe positive weights from n, and d to the internal elements and positive weights from the internal elements to NG. However, we see negative weights from v to the internal elements and from the internal elements to V G. The PG patterns are not categorized correctly at this point. One reason for this is that the PG patterns depend signicantly on the previous con- text. However, at this point, the network has just learned the obvious preferences and is only just starting to change the weights of the context layer. The network state after 150000 patterns is shown at the bottom of the gure. In the internal layer, a distributed representation has developed. Therefore, a direct interpretation is not easily pos- sible. However, it is observed that the rst internal element is primarily important for PG detec- tion, the second internal element plays an important role in V G assignment and the third internal element is important for NG. Nevertheless, this is a distributed rather than a local representation and there is additional in uence from other ele- ments. Furthermore, the weights of the context layer (from c to h) have changed. This is necessary in order to learn the PG group assignment. Generally speaking, we can explain certain phenomena using this type of weight analysis at the lowest interpretation level of a network. However, it is di-cult to extract explicit knowledge and a deeper understanding of the behavior of the net-work directly from the weights. Reasons for this di-culty include (1) the static representation of the weights which does not show the dynamics of a recurrent network, (2) the distribution of weights and activation, and (3) the number of weights, especially in the case of larger networks. There- fore, some eort could be made to reduce the size of the network by eliminating very small weights. Furthermore, groups of similar weights could be replaced with their average strength. Nonethe- less, weight analysis is still too detailed for larger networks. 5. Component activation analysis for knowledge extraction Weight analysis focuses on the weights and provides a very low-level analysis. One way to address this problem is to move towards activation analysis where the activations of internal elements are analyzed. Since internal elements receive activation from a number of weighted connections, the activation of an internal element integrates several weighted connections and provides a higher abstraction level of analysis. In order to demonstrate how such an analysis is performed, we will use the same SRN network we have introduced in the previous section and store all vector representations of the internal layer for each pattern. These vector representations constitute the input to a cluster algorithm which provides a hierarchical representation in the form of a dendrogram. Vectors with similar vector representations will end up in the same cluster. Figure 9 shows the initial part of patterns as they were clustered according to their internal activations. It can be clearly observed that the internal representations re ect the classication according to the three classes PG. That is, based on the weights, the internal layer has learned representations which particularly support this classication. A single word can appear in dierent contexts and can lead to different internal representations. For instance, the word \the" is shown with two dierent represen- tations. One representation is its use as part of the NG class, and the other as part of the PG class. Therefore, we nd both representations at dierent positions within the dendrogram. Knowledge Extraction from Transducer Neural Networks 35 ALL$U$NG I$U$NG I$U$NG US$U$NG WE$U$NG WEDNESDAY$N$NG MORNING$N$NG IN$R$PG WEEK$N$PG IS$V$VG COME$V$VG MUST$V$VG IS$V$VG Fig. 9. Hierarchical cluster analysis of internal classication representations (for visibility purposes, only a portion is shown 6. Principal component analysis for knowledge extraction Another kind of analysis which can be used for interpreting the internal representations of clas- sications is principal component analysis. Figure shows the result of this analysis for our current task. All vectors from the internal layer and the corresponding identiers provide the input for the principal component analysis. Vectors which dier substantially from each other are depicted in the gure with a large distance. It can also be observed that the internal representations re ect the preference mappings learned for the three category classes. NG, V G, and PG patterns are distributed across dierent areas. Thus, the classication of the internal representations can be clearly seen. This shows that the network has actually learned the classication task well. After learning has been completed, the internal representation characterizes the preference map- ping. According to cluster analysis or principal component analysis, similar internal vector representations are responsible for the representation of similar preference assignments to equal cate- gories. However, the interpretation of the weights by means of Hinton diagrams and of the activations via cluster analysis and principal component analysis only provides a limited form of structuring to the extracted knowledge. 36 Wermter I$U$NG IN$R$PG THURSDAY$N$NG* MORNING$N$NG WE$U$NG UNS$U$NG ON$A$PG THE$D$NG* IS$V$VG WEDNESDAY$N$NG I$U$NG* COME$V$VG MUST$V$VG I$U$NG* THAT$U$NG* IS$V$VG ALL$U$NG* TUESDAY COULD$V$VG US$U$NG LET$V$VG* US$U$NG DAS$U$NG* IS$V$VG MAKE$V$VG* TILL$R$PG* WE$U$NG IN$R$PG MARCH$N$PG OTHER$J$NG THAT$U$NG* IS$V$VG Fig. 10. Principal component analysis of internal classication representations 7. Transducer extraction Words and sequences of words can be represented as syntactic, semantic, and pragmatic category preferences. Then they can be input, for instance, to SRN networks. Each input representing a sequence of category preferences is associated with a sequence of corresponding output preferences. This simple description of sequence analysis is similar to the function of synchronous sequential machines [Booth, 1967, Kohavi, 1970, Shields, 1987], although preferences and learning are not yet considered in such machines. Therefore, we shall focus on extensions of synchronous sequential machines for representing sequential knowl- edge, especially synchronous Moore machines. We start with the basic denition of a synchronous sequential machine which is also called a transducer: Denition of a Synchronous Sequential Ma- chine, Transducer A synchronous sequential machine M is a tuple 1. I , O nite, nonempty sets of input and output 2. S nonempty set of states 3. The function f s : I S ! S is state transition function 4. The function f o is an output function. If the output depends on the state and the input, the machine is a so-called Mealy machine with the output function f O. If the output only depends on the state the machine, the later is a so-called Moore machine with the output function f These synchronous sequential machines are sometimes called transducers. A sequential machine assigns an output and a new state to an input and an old state. This can be done for a whole sequence of inputs and states in discrete time. The set S is not necessarily - nite [Booth, 1967], although this is assumed in the case of nite machines. Whereas automata Knowledge Extraction from Transducer Neural Networks 37 or acceptors of languages decide whether a certain input belongs to the corresponding grammar, these sequential machines are transducers which change their internal states dynamically, depending on the inputs and the previous states, while also providing an output for each input. Mealy and Moore machines are slightly dier- ent from each other. Moore machines determine the state rst and afterwards this state is used to provide the output. In contrast, the output in a Mealy machine depends also directly on the current input. However, it can be shown that for each Moore machine there is an equivalent Mealy machine and vice versa [Booth, 1967, Hopcroft and Ullman, 1979]. In our case, we concentrate on Moore machines since the output in certain neural networks is based on the internal state. This holds, for in- stance, for feedforward networks or SRN networks. Whereas sometimes [Sun, 1995] a sequential machine has been used to model a single element of a neural network, we want to use a sequential machine as a description for a whole network. This is also motivated by the fact that real neuron systems can be seen as physical entities which perform state transitions [Churchland and Sejnowski, 1992]. Now we can specify language knowledge by describing Moore machines and their state transition function f s and output function f o . We can also integrate f s and f o to a function f : IS ! OS. Then f corresponds for instance to the transformation within a SRN network. The specication of a Moore machine could be performed by using state tables. A potential entry for the task of assigning syntactic phrasal categories to syntactic basic categories could be: If verb and current state = prep. group then new state = verbal group and output = verbal group It may not be possible to assign a direct interpretation to a state. For this reason, simple identiers may be used: If verb and current state = 4 then new state = 5 and output = verbal group It is possible to dene state transition tables which assign each combination of input and current state an output and a new state. In this way, a symbolic synchronous sequential machine is specied. If clear regularities are known beforehand and the number is limited, such tables can be composed manually. However, the number of input and state combinations quickly gets so large that automatic procedures become necessary. The above-mentioned state transition tables are discrete symbolic. Therefore, they do not support gradual representations. For instance, the input or the state could be ambiguous and dierent gradual preferences could exist for dierent inter- pretations. For instance \meeting" could have a stronger preference for its syntactic interpretation as a noun and a smaller preference for a verb form. Consequently, we want to use preferences for the input, output, and states of such machines. Preferences of this type should be able to take values from [0; 1] m so that multiple preferences can be represented and integrated. If we extend a single category (as in: if verb) to an n-dimensional preference for the input and an m-dimensional preference for the output then we obtain a new synchronous machine which we will call a preference Moore machine. Now we want to describe such a synchronous sequential preference Moore machine which transforms sequential input preferences to sequential output preferences. We will see that simple recurrent networks or feedforward networks can be interpreted as neural preference Moore machines. Furthermore, we will show how symbolic and neural knowledge can be integrated quite naturally using preference Moore machines. 38 Wermter Denition of a Preference Moore Machine A preference Moore machine PM is a synchronous sequential machine which is characterized by a 4-tuple and S being non-empty sets of inputs, outputs and states. O S is the sequential preference mapping and contains the state transition function f s and the output function f o . Here I , O and S are n-, m- and l-dimensional preferences with values from [0; 1] n , [0; 1] m and [0; 1] l , respectively A generalized version of a preference Moore machine is shown in gure 11 on the left. The preference Moore machine realizes a sequential preference mapping, which uses the current state preference S and the input preference I to assign an output preference O and a new state preference. Preference mapping States AAAA AAAA Output Input Fig. 11. Neural preference Moore machine and its relationship to a SRN network Now we describe a new technique of extracting the knowledge within a recurrent network in the form of a transducer. A symbolic transducer can be extracted from our recurrent network which assigns to each input vector of basic syntactic categories a new output vector of phrasal categories depending on the previous context. In our network, the internal state and the context were represented by a three-dimensional vector. For simplicity, each strict symbolic interpretation of a three-dimensional vector can take 2 3 , that is 8 states. In order to acquire a symbolic interpretation of the network, we presented all patterns from the training set and stored the internal state vectors at the hidden layer of the network. For each output vector and for each state vector the next corner preference was determined using the Euclidean distance metric. Thus the Euclidean distance metric assigned one of three symbolic abstract syntactic phrase categories to each output vector and one of eight state number identiers to each state vector. Knowledge Extraction from Transducer Neural Networks 39000 100010 110011 n:ng r:pg v:vg d:ng d:ng n:ng r:pg u:ng d:ng n:ng n:ng d:pg j:pg n:pg v:vg d:ng a:ng a:pg j:ng n:ng v:vg v:vg r:pg v:vg u:ng r:pg n:ng n:ng d:ng u:ng n:ng r:pg a:pg v:vg v:vg u:ng Fig. 12. Transducer extraction from a recurrent network for the example sentence \That (u:ng) is (v:vg) the (d:ng) Thursday (n:ng) after (r:pg) Easter (n:ng)". Figure 12 shows the knowledge learned by the network as an extracted symbolic transducer. The corner nodes represent the eight strict states, the center node represents the start state of the trans- ducer. At the edges we nd the symbols for the single transductions. Input and output categories are separated by a colon, e.g. d : ng means that - starting from the source state of this edge - a determiner preference d is assigned to a noun group preference ng and the transduction is made to the end state of this edge. In the extracted transducer we can see some clear regularities at certain states. For instance, the transductions to state 100 are primarily responsible for the assignments to the prepositional group pg. Other examples are the transductions to state 010 and to state 000, which are primarily responsible for the verbal group (vg) assignment. Furthermore, gure 12 shows the example transductions for the sentence \That is the Thursday after Easter". Beginning with the start state at the center, we see the transduction ng for the word \That" which assigns the noun group ng to the pronoun u. Then, assigns a verb group vg to the verb \is". Then the transductions ng ng assign the noun group ng to \the Thursday". Finally the transductions assign the prepositional group pg to the sequence \after Easter". Dierent abstract syntactic categories (ng, pg) can be assigned to the same category (n) depending on the learned previous context. n:ng r:pg v:vg d:ng d:ng n:ng r:pg u:ng d:ng n:ng n:ng d:pg j:pg n:pg v:vg d:ng a:ng a:pg j:ng n:ng v:vg v:vg r:pg v:vg u:ng r:pg n:ng n:ng d:ng u:ng n:ng r:pg a:pg v:vg v:vg u:ng Fig. 13. Transducer extraction from a recurrent network for the example sentence \I (u:ng) thought (v:vg) in (r:pg) the (d:pg) next (j:pg) week (n:pg)". More detailed (less detailed) transducers can be obtained if the state and output vectors are mapped to more (fewer) nodes. Thus, the general abstraction level of such a symbolic transducer can be quite variable. The symbolic transducer represents an abstraction of the detailed network knowledge but this abstraction also hides some of the numerical complexity and allows a direct symbolic interpretation which provides a summary of the network behavior. To give an example, gure 13 shows the transductions for the example sentence \I thought in the next week". Beginning with the start state at the center, we see the transduction ng for the word \I", which assigns the noun group ng to the pronoun u. Then, assigns a verb group vg to the verb \thought". Finally the transductions r : pg d : assign the prepositional group \pg" to the word sequence \in the next week". One advantage of this transducer extraction is the higher abstraction level used for the representations of the recurrent network which leads to a better understanding of its function. The original network contains more detailed knowledge in the numerical weights and activations, but it is not possible to see the declarative sequential symbolic knowledge which this network represents. The extraction of a symbolic transducer allows a better understanding of the learned sequential knowledge which is represented in a more explicit manner. Knowledge Extraction from Transducer Neural Networks 41 8. Discussion and Analysis 8.1. Comparison of knowledge extraction technique There has been some previous work on using individual techniques in isolation for interpreting neural networks and extracting structural knowledge from them. In this paper, we have analyzed ve such dierent techniques using the same trained network in order to interpret the network knowl- edge. Such extensive comparisons of detailed net-work knowledge are needed in order to gain a better understanding of the knowledge extraction represented in neural networks. We have also introduced two new techniques here: dynamic learning analysis and transducer extraction. Dynamic learning analysis examines the formation and development of categories over time during learning. Thus, it provides a much deeper understanding of how the neural network arrives at its learned representation. Transducer extraction was developed to represent the sequential processing in a recurrent network at a higher level of abstraction. In general, we found that dierent interpretation techniques provide dierent views of the knowledge contained in a neural network. Thus, there is not a single best technique for all dier- ent aspects of knowledge extraction. The use of a particular technique depends rather on the requirements of the interpretation. In table 2, we illustrate and summarize the general properties of the ve dierent techniques. Dynamic Learning Analysis (DLA) is based on the output representations and provides a high level of understanding based on these known output representations. This technique is easy to interpret and can be used with other network types. On the other hand, it does not particularly support recurrent networks, symbolic integration, and exible knowledge structuring. Furthermore, structural relationships cannot be extracted. Transducer extraction (TE) is a new technique which uses output representations as well as internal activations. The main advantages of this technique are the high level of understanding in the form of an extracted symbolic transducer, the specic support for the sequentiality of recurrent networks and the possibility for extracting structural relationships. Such an extracted transducer can be integrated with other symbolic knowledge, e.g. other coded symbolic transducers. Further- more, dierent transducers can be generated with exibility, based on the number of states used in the internal activation layer. This leads to a relatively straightforward interpretation of the net-work involved compared to the other techniques, but it also requires the additional eort of extracting this symbolic transducer from the internal activations and the output representations. If we compare DLA and and CAA, we can see that DLA and are techniques that specically provide high level interpretations for dynamic learning and processing. We argue that WA, HAA, and CAA are techniques with a tendency towards a general, detailed, but low-level interpretation. DLA and TE, however, are techniques for specialized, high-level, dynamic interpretation. Focusing on output interpretations and the dynamics of recurrent networks provides a new level of understanding. Whereas a lot of previous work has focused on low-levels of in- terpretation, we believe that in the future, higher levels of interpretation and knowledge extraction will be required. 8.2. Related work on transducer extraction and related work Finite state automata and transducers have been widely used in various forms within traditional e.g. [Hopcroft and Ullman, 1979]. Basically, automata and transducers are always in a certain context state and they analyze a certain word (symbol). Then they move to a new state and potentially generate a new word (sym- bol). By using changing states, it is possible to encode the sequential context. Although nite automata or regular languages are not su-cient to describe all possible constructions of natural language completely (see e.g. [Winograd, 1983]), automata still constitute a central minimal requirement for the representation of natural language. Thus, they occupy the lowest level in the Chomsky hierarchy of languages [Hopcroft and Ullman, 1979]. Furthermore, it is possible to design e-cient realizations of nite automata for dierent domains [Kaplan, 1995], e.g. 42 Wermter Table 2. Comparison of dierent knowledge extraction techniques: Dynamic Learning Analysis (DLA), Weight Analysis (WA), Hierarchical Activation Analysis (HAA), Component Activation Analysis (CAA), Transducer Extraction (TE). Further abbreviations: Activations/Weights/Outputs and Low/Medium/High. Network representations used O W A A AO General level of understanding H L M M H Specic support for recurrent networks L L L L H Degree of structural relationships L L M M H Integration with symbolic knowledge L L M M H Flexibility in level of knowledge structuring L L M M H Computational eort L L M M M Easiness of interpretation H L M M H Generality and portability to other networks H H H H M for morphology, lexicon access, information extraction from sentences, syntactic tagging, etc. Recurrent networks have the potential to learn a sequential preference mapping f automatically, based on input and output examples (see gure 11), whereas traditional Moore machines or Fuzzy-Sequential-Functions [Santos, 1973] involve manual encoding. It has been recently illustrated how SRN networks can emulate each symbolic Moore machine and each nite automaton [Kremer, 1995, Kremer, 1996]. It has also been shown however [Goudreau and Giles, 1995, Goudreau et al., 1994] that a recurrent net-work with only a single input layer, one context layer, and one output layer, the so-called Single- layer-rst-order-network, is not su-cient for the realization of arbitrary nite automata. In natural language processing, representations have to be at least as powerful as nite au- tomata. Consequently, Single-layer-rst-order- networks are not appropriate, which is why we have used SRN networks here. These recurrent networks contain nite transducers as a special case, but also support much more powerful properties based on their gradual m-dimensional preference representations. For instance, it could be shown that SRN networks can emulate certain restricted properties of a pushdown automaton, in particular the recursive representation of structures with a limited depth [Elman, 1991, Wiles and Elman, 1996]. Apart from traditional symbolic regular rep- resentations, gradual and learned representations can also be represented. Furthermore, the number of input, state, and output preferences is not necessarily nite. Therefore, neural preference Moore machines are more powerful than nite transduc- ers. Our recurrent neural networks can be seen as learning augmenting a simple nite symbolic transducer with respect to learning within a gradual preference space. From this perspective, symbolic knowledge is a special abstract region in a neural preference space. An important line of research on automata and recurrent networks has been reported in [Giles et al., 1992, Goudreau and Giles, 1995, Tino et al., 1995]. Giles and colleagues studied both nite state automata and neural networks, but there are substantial dierences with our research. They started often with a known nite state automaton, which was used to generate sequences for it. Then these sequences were used for training a second-order neural network. Using a partition algorithm, a nite state automaton was extracted from the network activations, minimized and compared to the original known nite state automaton. In this way, Giles and colleagues could study the computational properties of the extraction particularly well, but the nite state automata also frequently relied on relatively simple 1/0 sequences. Knowledge Extraction from Transducer Neural Networks 43 Our motivation and methodology is dierent from theirs in several respects. We assume that the initial nite state automaton or transducer is not known. Especially for real-world problems, the interesting case is the one where such an automaton is not known in advance. Whereas it is interesting for comparison and sequence gen- eration, generating sequences with a nite state automaton already introduces certain regularities into the training set. Thus, sequence generation has an important in uence on the learning behav- ior, something which we want to rule out. In fact, we are more interested in situations where we do not know the machine which has to be ex- tracted. Especially with noisy real-world learning data, the underlying regularities may be quite disparate from regularly generated sequences. Furthermore, the task of our networks is quite dierent. The second-order networks employed by Giles and colleagues are trained for recognition. The output layer represents state representations which can be fed back to the input layer at the next step. Our recurrent networks perform an assignment task, where a sequence of inputs is associated with a sequence of outputs. We are not determining whether a certain sequence belongs to a certain automaton, but what the simple at structure of this sequence is. That is, we are interested in transducer extraction rather than recognizer ex- traction. In general, there are no designated nal states in our networks, since the network - and the extracted symbolic transducer - produce output as long as input is provided. This transducer behavior is therefore quite dierent from the recognition performance reported in [Giles and Omlin, 1993], which is based on acceptors for articial languages. 9. Conclusion The main contribution of this paper is a particularly broad analysis of knowledge extraction for recurrent networks. In addition, we propose dynamic learning analysis and transducer extraction as two new dynamic interpretation tech- niques. Dynamic learning analysis provides a better understanding of how the network learns, while transducer extraction provides a better understanding of what the network represents. After learning, a conservative \lazy learning" strategy leads to connectionist representations which can be described as symbolic transducers. These transducers allow for a much better interpretation of the sequential network knowledge compared to the standard analysis using hierarchical clustering or Hinton diagrams. Weight analysis, cluster analysis, and principal component analysis are detailed but static. In contrast, our new method for extracting symbolic transducers can describe the learned classication performance much bet- ter, since transducer extraction considers the sequential character of the learned representations in a recurrent network and allows a better symbolic inspection. Possibilities for direct integration with symbolic classiers can be explored in future work. We conclude that dynamic learning analysis and transducer extraction have a lot of potential for improved knowledge structuring based on recurrent networks. --R Extracting algorithms from pattern classi Rules and Networks. Sequential Machines and Automata Theory. The Computational Brain. Extracting Comprehensible Models from Trained Neural Networks. Distributed repre- sentations Language as a dynamical system. Learning and extracted On recurrent neural networks and representing Hybrid Prob- lems Learning distributed representations of concepts. Introduction to Automata Theory Finite state technology. Switching and Finite Automata Theory. On the computational power of Elman-style recurrent networks A theory of grammatical induction in the connectionist paradigm. Hybrid Intelligent Systems. Extraction of rules from discrete-time recurrent neural networks Fuzzy sequential func- tions A framework for combining symbolic and neural learning. An Introduction to Automata Theory. Learning distributed representations for the classi Finite state machines and recurrent neural networks. Hybrid Connectionist Natural Language Processing. The hybrid approach to arti Preference moore machines for neural fuzzy integration. Building lexical representations dynamically using arti SCREEN: Learning a at syntactic and semantic spoken language analysis using arti Learning to count without a counter: A case study of dynamics and activation landscapes in recurrent net- works Language as a Cognitive Process. --TR
symbolic interpretation;SRN networks;analysis of connectionist learning;knowledge extraction;neural network learning
590930
Two-Loop Real-Coded Genetic Algorithms with Adaptive Control of Mutation Step Sizes.
Genetic algorithms are adaptive methods based on natural evolution that may be used for search and optimization problems. They process a population of search space solutions with three operations: selection, crossover, and mutation. Under their initial formulation, the search space solutions are coded using the binary alphabet, however other coding types have been taken into account for the representation issue, such as real coding. The real-coding approach seems particularly natural when tackling optimization problems of parameters with variables in continuous domains.A problem in the use of genetic algorithms is premature convergence, a premature stagnation of the search caused by the lack of population diversity. The mutation operator is the one responsible for the generation of diversity and therefore may be considered to be an important element in solving this problem. For the case of working under real coding, a solution involves the control, throughout the run, of the strength in which real genes are mutated, i.e., the step size.This paper presents TRAMSS, a Two-loop Real-coded genetic algorithm with Adaptive control of Mutation Step Sizes. It adjusts the step size of a mutation operator applied during the inner loop, for producing efficient local tuning. It also controls the step size of a mutation operator used by a restart operator performed in the outer loop, for reinitializing the population in order to ensure that different promising search zones are focused by the inner loop throughout the run. Experimental results show that the proposal consistently outperforms other mechanisms presented for controlling mutation step sizes, offering two main advantages simultaneously, better reliability and accuracy.
INTRODUCTION . Genetic algorithms (GAs) are general purpose search algorithms which use principles inspired by natural genetic populations to evolve solutions for problems ([Goldberg (1989a), Holland (1992)]). The basic idea is to maintain a population of chromosomes, which represent candidate solutions for the specific problem, that evolves over time through a process of competition and controlled variation. The following bibliography may be examined for a more detailed discussion about GAs: [B-ack (1996), B-ack et al. (1997), Goldberg (1989a), Holland (1992), Michalewicz (1992)]. Under their initial formulation, the search space solutions are coded using the binary alpha- bet. However, other coding types have been considered for the representation issue, such as real coding, which would seem particularly natural when tackling optimization problems of parameters with variables on continuous domains. Then a chromosome is a vector of floating point numbers, the size of which is kept the same as the length of the vector, which is the solution to the prob- lem. GAs with this type of coding are called real-coded GAs (RCGAs) (see [Herrera et al. (1998), Surry et al. (1996)]). There are other types of Evolutionary Algorithms (EAs), i.e., implementing the idea of evolution ([B-ack (1996)]), which are based on real coding as well. These are Evolution Strategies ([Schwefel (1995)]) and Evolutionary Programming ([Fogel (1995)]). This paper deals with RCGAs. Population diversity is crucial to a GA's ability to continue the fruitful exploration of the search space ([Li et al. (1992)]). If the lack of population diversity takes place too early, a premature stagnation of the search is caused. Under these circumstances, the search is likely to be trapped in a local optimum before the global optimum is found. This problem, called premature convergence, has long been recognized as a serious failure mode for GAs ([Eshelman et al. (1991)]). The mutation operator may be considered to be an important element for solving the premature convergence problem, since it serves to create random diversity in the population ([Spears (1993)]). Different techniques have been suggested for the control, during the GA's run, of parameters associated with this operator, depending on either the current state of the search or other GA related parameters ([Angeline (1995), Herrera et al. (1996b), Hinterding et al. (1997)]). They try to offer suitable diversity levels for avoiding premature convergence and improving the results. In the case of working with real coding, a topic of major importance involves the control of the proportion or strength in which real-coded genes are mutated, i.e., the step size ([B-ack et al. (1996a)]). The objective of this paper is to formulate a mechanism for the control of mutation step sizes for RCGAs, which should handle and maintain population diversity that in some way helps produce good chromosomes, i.e., useful diversity ([Mahfoud (1995)]). We present TRAMSS, a Two-loop RCGA model with Adaptive control of Mutation Step Sizes that attempts to do this. It is made up by two loops, an inner loop and an outer one: Inner loop. It is designed for processing useful diversity in order to lead the population toward the most promising search areas, producing an effective refinement on them. So, its principal mission is to obtain the best possible accuracy levels. The inner loop performs the selection process and fires the crossover and mutation operators. Furthermore, for achieving its objective, it controls the step size of the mutation operator. ffl Outer loop. It introduces new population diversity, after the inner loop reaches a stationary point where there are no improvements, that helps the next one to reach better solutions. Therefore, it attempts to induce reliability in the search process. The outer loop iteratively performs the inner one, and later, it applies a restart operator that reinitializes the population by mutating all genes, using a step size that is adapted as well, throughout the runs for this loop. For doing this, the paper is set up as follows: in Section 2, we analyze two mutation issues, the ways in which the control of mutation step sizes may be made and the idea of the restart operator; in Section 3, we present TRAMSS, in Section 4, we describe the experiments carried out for determining the efficacy of the proposal; and finally, some conclusions are dealt with in Section 5. In this Section, we explain two issues that will be included as important components in the conceptual foundation of TRAMSS, mutation step size control (Subsection 2.1) and the restart operator (Subsection 2.2). 2.1 Mutation Step Size Control In general, the mechanisms presented for controlling parameters associated with EAs may be assigned to the following three categories ([Hinterding et al. (1997)]): ffl Deterministic Control. It takes place if the values of the parameters to be controlled are altered by some deterministic rule, without using any feedback from the GA. Usually, a time-varying schedule is used. Adaptive Control. It takes place if there is some form of feedback from the GA that is used to determine the direction and/or magnitude of the change to the parameters to be controlled. The rules for updating parameters that are used by this type of control and, by the previous one, are termed absolute adaptive heuristics ([Angeline (1995)]) and, ideally, capture some lawful operation of the dynamics of the EA over a broad range of problems. ffl Self-adaptive Control. The parameters to be controlled are encoded onto the chromosomes of the individual and undergo mutation and recombination. Next, we describe mechanisms for the control of mutation step sizes that belong to each one of these categories. 2.1.1 Deterministic Step Size Control In [Michalewicz (1992)], a mutation operator for RCGAs, called non-uniform mutation, was pre- sented, which is based on the absolute adaptive heuristic "to protect the exploration in the initial stages and the exploitation later". It implements this idea by decreasing the step size as the GA's execution advances. Let us suppose that this operator is applied on a real-coded gene, x 2 [a; b] (a; b 2 !), at generation t, and that T is the maximum number of generations, then it generates a gene, x 0 , as follows: ae with - being a random number that may have a value of zero or one, and where r is a random number from the interval [0; 1] and b is a parameter chosen by the user. This function gives a value in the range [0; y] such that the probability of returning a number close to zero increases as the algorithm advances. The size of the gene generation interval shall be smaller with the passing of generations. This property causes this operator to make an uniform search in the initial space when t is small, and very locally at a later stage, favoring local tuning. The non-uniform mutation operator has been widely used, reporting good results ([Herrera et al. (1996a), P'eriaux et al. (1995), Sefrioui et al. (1996)]). It is considered to be one of the most suitable mutation operators for RCGAs ([Herrera et al. (1998)]). 2.1.2 Adaptive Step Size Control The (1+1)-Evolution Strategy ((1+1)-ES) ([Schwefel (1995)]) is an EA that uses adaptive step size control. It attempts to adapt its mutation step size to the problem according to the absolute adaptive heuristic: "expand the step size when making progress, shrink it when stuck". This heuristic will be denoted as E/S heuristic. (1+1)-ES works using a continuous representation and a mutation operator based on normally distributed modifications with expectation zero and given variance, oe, as the step size. It operates on a vector of variables by applying mutation with identical oe to each variable, so generating a descendant. The better of ancestor and descendant is considered as the new starting point. (1+1)- ES applies the E/S heuristic for adapting oe by means of the 1/5 success rule. This rule uses the results obtained by mutation in the last few generations: if more than one fifth of the mutation have been successful, the step size is increased, otherwise it is decreased. In [De La Maza et al. (1994)], a dynamic hill climbing algorithm is presented, which uses the E/S heuristic as well. We would like to point out that the model proposed in this paper, TRAMSS, uses important ideas that are present in this algorithm. 2.1.3 Self-Adaptive Step Size Control In [Schwefel (1995)], an EA model, called (-Evolution Strategy ((-ES), is developed that uses a mechanism for the self-adaptive step size control. In (-ES, - parents create - offspring by means of recombination and mutation, and the best offspring individuals are deterministically selected to replace the parents. Therefore, - should be greater than -. The main quality of the algorithm is its ability to incorporate the standard deviations (step sizes) and the correlation coefficients of normally distributed mutations into the search process, such that adaptation not only takes place in the object variables, but also in these parameters according to the current local topology of the search space. This property is called self-adaptation ([B-ack (1996), Schwefel (1995)]). Self-adaptation exploits the indirect link between favorable parameter values and fitness function values, being capable of adapting the parameters implicitly, according to the topology of the objective function ([B-ack et al. (1996b)]). Therefore, each population individual consists of three vectors, representing the object variable, the standard deviation and the rotation angle values, respectively. The vector ~x has dimensions, equal to the number of problem variables. The n oe dimensions of a vector ~oe can be up to n (in this case, each object variable x different step size oe i associated to it), and n ff can be up to (2\Deltan\Gamman oe )\Delta(n oe \Gamma1). n ff may be set to zero, indicating that the rotation angles are not considered, as is assumed in this paper. For more information about (-ES refer to [B-ack (1996), Schwefel (1995)]. Other mechanisms for the self-adaptive step size control are to be found in [Fogel (1995), Hinterding (1995), Ostermeier et al. (1994)]. 2.2 Restart Operator Premature convergence causes a drop in the GA's efficiency; the genetic operators do not produce the feasible diversity to tackle new search space zones and thus the algorithm reiterates over the known zones producing a slowing-down in the search process. Under these circumstances, resources may be wasted by the GA searching an area not containing a solution of sufficient quality, where any possible improvement in the solution quality is not justified by the resources used. Therefore, resources would be better utilized in restarting the search in a new area, with a new population et al. (1995)]). This is carried out by means of a restart operator. Next, we review some different approaches to this operator. ffl In [Goldberg (1989b)], it was suggested restarting GAs that have substantially converged, by reinitializing the population using both randomly generated individuals and the best individual from the converged population. ffl In [Eshelman (1991)], upon convergence, the population is reinitialized by using the best individual found so far as a template for creating a new population. Each individual is created by flipping a fixed proportion (35%) of the bits of the template chosen at random without replacement. If several successive reinitializations fail to yield an improvement, the population is completely (100%) randomly reinitialized. ffl In [Maresky et al. (1995)], a selectively destructive restart is proposed that does not completely destroy the converged population; a percentage of the converged genes will survive untouched to begin the next convergence stage. A probability of gene reinitialization, p r , is used: the higher the rate, the more genes are initialized. Experiments carried out with some p r values showed that different problems have different optimal reinitialization probabilities. This model seems to provide an improved method for renewing genetic diversity in GA search. Intuitively, the complete reinitialization of the population forgets the previous solutions, therefore it cannot make use of previously discovered building blocks. ffl In [Grefenstette (1992)], a similar mechanism, called partial hypermutation model, was in- troduced, which replaces, at each generation, a percentage of the population by randomly generated individuals. The percentage is called replacement rate. The intended effect is similar to the one of the previous approach: to maintain a continuous level of exploration of the search space, while trying to minimize disruption for the ongoing search. Other important GA models based on the restart operator are ARGOT ([Shaefer (1987)]), Dynamic parameter encoding ([Schraudolph et al. (1992)]) and Delta coding ([Whitley et al. (1991)]). OF STEP SIZES. In this Section, we present TRAMSS. It uses: ffl An instance of the absolute adaptive E/S heuristic, presented in Subsection 2.1.2, for the adaptive step size control of the mutation operator applied in the inner loop, and ffl An instance of its opposite version, denoted here as S/E heuristic, for the adaptive step size control of the mutation operator used by the restart operator that is executed by the outer loop. Next, in Subsection 3.1, we examine the application of the E/S and S/E heuristics for step size control in RCGAs, and, in Subsections 3.2 and 3.3, we present the TRAMSS inner and outer loops, respectively. 3.1 The E/S and S/E Heuristics Let's suppose that an RCGA is applying a mutation operator with ffi being its step size. If a stationary state is detected (the fitness of the best individual or the average fitness have not been improved during the previous generations), there are two possible causes concerning ffi: 1. It is too high, which does not allow the convergence to be produced for obtaining better individuals, or 2. It is too low, which induces a premature convergence, with the search process being trapped in a local optimum. On the one hand, if we decided to include an adaptive control of ffi based on the instance of the E/S heuristic "increase ffi when making progress, decrease it when stuck", a stationary state caused by (1) would be suitably tackled, since ffi would become lower, so introducing more convergence. However, this heuristic would not be adequate if the stationary state is caused by (2), because it would complicate the problem even more. Precisely, this last circumstance will occur as the number of iterations increases. Since the RCGA will find more difficulties for making progress, the natural trend of the instance of the E/S heuristic will be to lead ffi to lower values, so producing more convergence. The possibility of this problem has been claimed by some authors. For example, in [B-ack et al. (1995)], the following was stated about the 1/5 success rule: ". the 1/5 success rule may cause premature stagnation of the search due to the deterministic decrease of the step size whenever the topological situation does not lead to a sufficiently large success rate". For complex problems, this effect will probably become a premature convergence. This explains the following claim, again about the 1/5 success rule ([Angeline (1995)]): ". this heuristic is especially useful in smooth multimodal environments of the type well studied by the ES community but would be less applicable in discontinuous or extremely rough environments". On the other hand, if we are inclined to use the instance of the S/E heuristic "decrease ffi when progress is made, increase it when there are no improvements", a stationary state produced by (2) will be adequately attacked, since ffi would be greater and so, more diversity is introduced with the possibility of escaping from the local optimum. However, an important problem may occur: as no improvements are made by the RCGA, higher ffi values are tried, so introducing too much diversity and not considering the possibility that convergence may be suffice for improving results. So, all these facts show that serious problems may arise when the E/S and S/E heuristics are applied separately. However, we think that a mechanism applying both of these heuristics would handle the population diversity suitably to avoid the premature convergence problem and improve the behavior of the search process. The adaptive RCGA model proposed, TRAMSS, includes this idea: it uses the E/S heuristic for adapting the step size of a mutation operator applied in the inner loop and the S/E heuristic for adapting the step size of a mutation operator used by a restart operator performed in the outer loop. 3.2 TRAMSS Inner Loop The inner loop performs the usual process (selection, crossover and mutation) over a number of G, called time-interval between observations. Then, depending on the progress of the population mean fitness found throughout these generations, it adjusts the step size of the mutation operator, and calculates a new value for G. Next, we fully describe these steps where a minimization problem is assumed. Selection, Crossover and Mutation (Step 2.2). Over the time-interval between observations, G, the following selection mechanism and crossover and mutation operators are applied. ffl The selection probability calculation follows linear ranking ([Baker (1985)]), with and the sampling algorithm is the stochastic universal sampling ([Baker (1987)]). The elitist strategy ([De Jong (1975)]) is considered as well. It involves making sure that the best performing chromosome always survives intact from one generation to the next. This is necessary since it is possible that the best chromosome disappears, due to crossover or mutation. ffl We have tried different crossover operators, which are presented in Subsection 4.2. ffl The mutation operator used is denoted as Mutation(ffi), where ffi is the step size (0 - ffi - 1). This operator is defined as follows: If x 2 [a; b] is a gene to be mutated, then the gene resulting from the application of this operator, x 0 , will be a random (uniform) number chosen from Clearly, the higher ffi is, the greater changes on x are produced. Adaptive Control of ffi (Step 2.3). After G generations, the ffi parameter used by the mutation operator is adapted following a particular instance of the E/S heuristic: "increase ffi when observing progress on - f (population mean fitness), decrease it when stuck". ffi is kept in the interval [ffi min ; \Delta], where \Delta is a parameter calculated by the outer loop, as described in Subsection 3.3, and ffi min is the minimum threshold defined by the user (in experiments we assume a value of 1.0e-100). The inner loop ends when ffi reaches the ffi min value. By finishing with a fine grained search with small step sizes, we are sure that a local optimum, or the global one, will be located precisely ([De La Maza et al. (1994)]). It will stop as well, when a maximum number of generations is reached. The update rates for ffi depend on the number of previous successive observations that were successful or not successful. Two variables, yes and no, are used for recording these occurrences, respectively. If progress is made during many successive previous observations fOld - fOld being the population mean fitness of the previous iteration), then the increasing rate for ffi is very high (in particular, ffi is multiplied by 2 yes ), whereas if these observations were not successful, then the decreasing rate is high (ffi is divided by 2 no ). In this way, when the search process is located in a local optimum and improvements are still not surely expected by reducing ffi, the inner loop duration will not be too long. Time-interval Calculation (Step 2.4). The time-interval between observations, G, is calculated depending on the current values of ffi with regard to \Delta. If ffi is similar to \Delta, then the time-interval is high (G 0 100 in the experiments), and if it is lower, the time-interval will become like Gmin in the experiments). This allows ffi values similar to \Delta to be used for a long time (\Delta is considered a good starting point for ffi, because it is adapted in the outer loop on the basis thereof, as we will explain in Subsection 3.3). Furthermore, we need to point out that the initial \Delta, \Delta 0 , was assigned to 1 in the experiments, in order to favor exploration during the initial stages of the first inner loop's run. Figure 1 shows the pseudocode algorithm for the whole TRAMSS inner loop. In short, the objective of this loop is to find and refine local optima (or the global one), in an efficient way. TRAMSS INNER LOOP 1. 2. while (ffi - ffi min ) and (not termination-condition) do 2.1. - fOld := - 2.2. perform Selection, Crossover and Mutation(ffi) over G generations; 2.3. if ( - fOld - f) then else 2.4. G := G0 \Delta ffi=\Delta; if Fig. 1. TRAMSS Inner Loop Structure 3.3 TRAMSS Outer Loop The outer loop randomly initializes the population that will be handled throughout the TRAMSS run. It fires the inner loop, and when this one returns, it applies a restart operator based on a step size that is adaptatively controlled throughout its execution. Now, we explain, in depth, the main issues related to this loop. Restart Operator Application (Step 3.4). The outer loop applies a restart operator, called applies Mutation(\Delta) to all the genes in the chromosomes stored in the population. The objective of this operator is similar to the one of the partial restart operators for binary-coded GAs, described in Section 2.2, i.e., to maintain a continuous level of exploration of the search space, while trying to use the promising zones located as a kind of sketch. It attempts to ensure that new and promising genetic material is available in the population for being handled and treated by the next inner loop. Adaptive Control of \Delta (Step 3.3). The outer loop adapts the \Delta parameter, using information obtained after each inner loop run, by means of an instance of the S/E heuristic: "decrease \Delta when observing progress on fBest (fitness of the best element found so far), otherwise increase it. This is implemented by dividing the previous \Delta value by 2 or multiplying it by 2, respectively. The new \Delta value will be the first value for the ffi parameter used in the next inner loop. The pseudocode algorithm for the outer loop is depicted in Figure 2. To sum up, the outer loop attempts to introduce adequate diversity levels for allowing the subsequent inner loop processing to be capable of finding, better local optima, or the global one, every time. For this reason, it uses the f best for the adaptive step size control. When no better local optima are found after the last inner loop runs, the outer loop produces more diversity in order to increase the probability of having access to a better one, which will be refined by the next inner loop. On the other hand, if better solutions are being found by previous inner loop's runs, \Delta becomes low, so avoiding, for the moment, great destructive effects of the restart operator. TRAMSS OUTNER LOOP 1. 2. run Initialize; 3. while (not Termination-condition) do 3.1. fOldBest := fBest ; 3.2. run Inner Loop; 3.3. if (fOldBest - fBest ) then else 3.4. run Restart(\Delta); Fig. 2. TRAMSS Outer Loop Structure Minimization experiments on the test suite, described in Subsection 4.1, were carried out in order to study the behavior of the TRAMSS model. In Subsection 4.2, we describe the algorithms built in order to do this and, in Subsection 4.3, we show the results and discuss some conclusions about them. fSph fRos fSph fSch fGri fSch d fSch fRas ef10 Fig. 3. Test functions 4.1 Test Suite For the experiments, we have considered six test functions used in the GA literature: Sphere model (f Sph ) ([De Jong (1975), Schwefel (1981)]), Generalized Rosenbrock's function (f Ros ) ([De Jong (1975)]), Schwefel's Problem 1.2 (f Sch ) ([Schwefel (1981)]), Griewangk's function ([Griewangk (1981)]), Generalized Rastringin's function (f Ras ) ([B-ack (1992), T-orn et al. (1989)]), and Expansion ([Whitley et al. (1995)]). Figure 3 shows their formulation. The dimension of the search space is 25. fSph is a continuous, strictly convex and unimodal function. fRos is a continuous and unimodal function, with the optimum located in a steep parabolic valley with a flat bottom. This feature will probably cause slow progress in many algorithms since they must permanently change their search direction to reach the optimum. This function has been considered by some authors to be a real challenge for any continuous function optimization program ([Schlierkamp-Voosen et al. (1994)]). A great part of its difficulty lies in the fact that there are nonlinear interactions between the variables, i.e., it is nonseparable ([Whitley et al. (1996)]). fSch is a continuous and unimodal function. Its difficulty lies in the fact that searching along the coordinate axes only gives a poor rate of convergence, since the gradient of fSch is not oriented along the axes. It presents similar difficulties to f Ros , but its valley is much narrower. fRas is a scalable, continuous, separable and multimodal function which is produced from fSph by modulating it with a fGri is a continuous and multimodal function. This function is difficult to optimize because it is nonseparable ([M-uhlenbein et al. (1991)]) and the search algorithm has to climb a hill to reach the next valley. Nevertheless, one undesirable property exhibited is that it becomes easier as the dimensionality is increased ([Whitley et al. (1996)]). is a function that has nonlinear interactions between two variables. Its expanded version, built in such a way that it induces nonlinear interaction across multiple variables. It is nonseparable as well. 4.2 Algorithms We have built five different TRAMSS versions that apply the following crossover operators: Arithmetical ([Michalewicz (1992), Wright (1991)]), Max-min-arithmetical ([Herrera et al. (1997)]), Discrete ([M-uhlenbein et al. (1993)]), Fuzzy recombination ([Voigt et al. (1995)]), and BLX-ff ([Eshel- man et al. (1993)]). The TRAMSS versions are called TRAMSS-AR, TRAMSS-MMA, TRAMSS-DI, TRAMSS-FR and TRAMSS-BLX, respectively. We have implemented five classical RCGAs based on these crossover operators that apply the non-uniform mutation operator (Subsection 2.1.1.) and, the same selection strategy as the one used by the TRAMSS inner loop. They are called RCGA-AR, RCGA-MMA, RCGA-DI, RCGA-FR and RCGA-BLX, respectively. The crossover probability in these RCGAs and in the TRAMSS versions is 0:6, the mutation probability 0:005, and the population size 60 chromosomes. Now, we present the definition of the crossover operators. Let us assume that are two real-code chromosomes that have been selected to apply the crossover operator to them. Below, the effects of the five crossover operators are shown. Arithmetical crossover. An offspring, We have considered Max-min-arithmetical crossover. It generates the following four offspring: In particular, we have considered 0:25. The resulting descendents are the two best of the four aforesaid offspring. Discrete crossover. z i is a randomly (uniformly) chosen value from the set fx g. Fuzzy recombination. The probability that the i-th gene in the offspring has the value z i is given by the distribution p(z i g, where OE x i and OE y i are triangular probability distributions with the following features (d - 0:5): Triangular Prob. Dist. Minimum Value Modal Value Maximum Value d have been set to 0.5 in the experiments. BLX-ff crossover. z i is a randomly (uniformly) chosen number from the interval [Min \Gamma I \Delta We have assumed All these crossover operators may be ordered with regard to the way randomness is used for generating the genes of the offspring: 1) Arithmetical and Max-min-arithmetical crossovers do not use it, 2) Discrete crossover considers discrete probability distributions, where there are only two possibilities introduces uniform continuous probability distributions, and Fuzzy recombination applies triangular continuous probability distributions, and therefore, it may be considered as a hybrid between discrete crossover and BLX-ff. Probability distributions used by BLX-ff and Fuzzy recombination are calculated according to the distance between the genes in the parents (x i and y i ), and the ff and d values, respectively. So, they fit their action range depending on the diversity of the population using specific information held by the parents ([Eshelman et al. (1993)]). We have also implemented two (-ESs, with 100. The first one is denoted as (15; 100)-ES1 and uses n the other, denoted as (15; 100)-ESn, uses n i.e., the number of variables. n for both of them. They apply discrete recombination (random exchanges between parents) on object variables and local intermediate recombination (arithmetic averaging) on standard deviations (see [B-ack (1996)] for a more detailed description of these operators). The number of potential parents involved in the recombination of the object variable is 15, and that for the standard deviations is 15, as well. The standard deviations are initialized to value of 3.0. This parameterization is very usual for (-ES ([B-ack et al. (1995)]). All TRAMSS versions and RCGAs were executed 15 times, each one with 10,000 generations, except TRAMSS-MMA and RCGA-MMA, that performed 5,000 generations, since each Max-min- arithmetical crossover application needs four evaluations. (15; 100)-ES1 and (15; 100)-ESn were executed times, each one with 4,000 generations. In this way, the number of fitness function evaluations required by all algorithms are similar. 4.3 Results For each function, we introduce the average of the best fitness function found at the end of each run (A), the best of these values (B), and the percentage of success with respect to the thresholds shown in Table 1 (S). Table 2 shows the results obtained. Test Thresholds fSph 1.0e-150 fRos 1.0 fSch 1.0e-3 fRas Table 1. Thresholds for the test functions First, we consider the behavior of the TRAMSS algorithms compared with their corresponding RCGAs. Then, we comment on the results obtained by (-ES1 and (-ESn. 4.3.1 RCGAs vs. TRAMSS In general, TRAMSS-AR, -MMA, -DI, -FR and -BLX outperform their corresponding RCGAs, RCGA-AR, -MMA, -DI, -FR and -BLX, respectively, for all functions, with regard to all performance measures. To explain this ability, first we focus on the study of the results on the unimodal functions, fSph , f Ros and fSch , and then, on the multimodal ones, f Ras , fGri and f ef10 . Analysis for unimodal functions. ffl We have observed that during runs of the TRAMSS instances on all unimodal functions, the outer loop performed the inner loop only once. This means that the inner loop has been controlling ffi to properly suit the local nature of the landscape in these functions. In this way, improvements on - f (population mean fitness) predominated, so obtaining very accurate results. Thus, we may say that the implementation of the E/S heuristic used for controlling ffi is highly suitable for dealing with unimodal functions. In fact, we should point out that this heuristic has already been used for designing efficient local search procedures ([De La Maza et al. (1994)]). fSph fRos fSch Algorithms A B S A B S A B S TRAMSS-FR 4.5e-153 2.1e-163 100.0 1.6e+01 2.7e-03 20.0 2.7e-04 2.7e-05 100. 0 TRAMSS-BLX 2.2e-176 2.7e-188 100.0 1.3e+01 4.9e-01 13.3 7.4e-08 2.2e-09 100 .0 (15, 100)-ESn 0.0e+00 0.0e+00 100.0 3.4e+00 7.9e-03 13.3 7.8e+03 3.4e+03 0.0 fRas fGri ef10 Algorithms A B S A B S A B S TRAMSS-MMA 2.3e-14 0.0e+00 33.3 2.7e-02 0.0e+00 13.3 6.9e-01 0.0e+00 66.7 Table 2. Results of experiments ffl For fSph , the A and B measures of the TRAMSS algorithms are very accurate to a higher degree than the ones for the respective RCGAs based on non-uniform mutation (see the very good results of TRAMSS-FR and TRAMSS-BLX for this function). ffl For f Ros and fSch , more complex functions, these measures are better as well. The joint effects of the E/S heuristic and the application of high ffi values during many initial generations allow the inner loop to be capable of tackling the difficulties associated with the localization of the optimum in these functions. For example, we may highlight the good B results obtained by TRAMSS-FR for f Ros , 2.7e-03, and TRAMSS-BLX for fSch , 2.2e-9. These results indicate that the inner loop has been generating useful diversity throughout the runs for the unimodal functions. On the other hand, the non-uniform mutation does not take into account whether the diversity being generated is useful or not. It only decreases the step size depending on the time without observing if improvements are made, or not. For the case of f Ros , this fact does not allow a good search direction to be established for reaching the optimum, and so, no good final results are obtained (see the results of RCGA-AR, RCGA-FR and RCGA-BLX for this Analysis for multimodal functions. ffl The participation of the restart operator in the outer loop allowed reliability to be improved on the multimodal functions, with regard to the RCGAs based on non-uniform mutation. For f Ras and fGri , all TRAMSS implementations reached the global optimum, at least once (see B measure). ffl Other examples of improvements on reliability, for the case of fGri , involve increasing in the S measure (expressed with regard to the fitness of the global optimum) of TRAMSS-AR (60%) in front of RCGA-AR (0%), and TRAMSS-BLX (80%) as contrasted with RCGA-BLX (26.7%). ffl For the case of , the union of the effective local tuning of the inner loop and the introduction of a spread search by means of the restart operator allows very good results to be obtained. For example, the A measure of TRAMSS-FR for this function is the best of all the algorithms executed, 1.4e-14, and the B measure of TRAMSS-BLX is very exact, 1.5e-44. On the other hand, the decreasing of the step size performed by the non-uniform mutation does not allow the search direction to escape from a possible stagnation in a local optimum, when working on multimodal functions. In particular, for the case of using arithmetical and discrete operators, we may observe how this disadvantage induced very poor performance measure values for this type of functions. TRAMSS and Crossover Operators. ffl An important issue that should be highlighted is that the improvements of the TRAMSS implementations on their corresponding RCGAs based on non-uniform mutation operator are more notable for most test functions when using Fuzzy recombination and BLX-ff. The adaptive ffi control performed in the inner loop depends on the changes produced on f , which are determined by the joint effects of the selection process, and the mutation and crossover operators. Let us consider only the interactions between the last ones. The mutation operator generates diversity and the crossover operator would have to use it for creating better individuals. If the crossover operator achieves this task, then the mutation operator would be generating "useful diversity", and so evolution is introduced. Therefore, only if the mutation and the crossover operators are being suitably coupled, the success of the inner loop may be accomplished. Results have shown that in the case of using Fuzzy recombination and BLX-ff, this circumstance is held. In particular, we highlight the very good A, B and S measures of TRAMSS-FR and TRAMSS-BLX for all functions. These facts lead us to think that the associated property of these crossover operators (to fit their action range depending on the diversity of the population) is the main one responsible for making this union so profitable. It would allow these operators to exploit the diversity generated by the mutation operator. ffl We should underline the good results of the TRAMSS version based on the Max-min-arithme- tical crossover as well, TRAMSS-MMA, for the unimodal fSph and the multimodal . It reached the global optimum of the first one for 100% of the runs (see A measure) and was the only algorithm that found the global optimum of the second one (see B measure). The process of selecting the two best offspring from a set of four ones with different properties along with the effects of the TRAMSS model allow the best search regions to be located and refined. 4.3.2 (-ESs vs. TRAMSS performance measures for (15, 100)-ES1 show good behavior. The adaptation mechanism concerning this algorithm is similar to the one in the TRAMSS model to the fact that all genes belonging to the same chromosome are mutated using the same step size. However, all TRAMSS implementations have outperformed this algorithm. On the other hand, (15, 100)-ESn shows a similar behavior for fSch and f Ras than (15, 100)-ES1. Furthermore, it is outperformed by TRAMSS-FR and TRAMSS-BLX in the complex f ef10 . However, its performance for the unimodal functions fSph and f Ros and for the multimodal one fGri was very good. For these functions, (15, 100)-ESn certainly benefited from the greater degree of freedom by working with n different self-adaptive step sizes per individual in contrast to a single one in the case of the TRAMSS algorithms. 5 CONCLUDING REMARKS. In this paper, we have presented TRAMSS, a two-loop RCGA model that adjusts the step size of a mutation operator applied during the inner loop, for producing an efficient local tuning, and controls the step size of a mutation operator used by the outer loop, for reinitializing the population in order to ensure that different promising search zones are focused by the inner loop throughout the run. An instance of the E/S heuristic was used for implementing the adaptive mechanism in the inner loop whereas an instance of its opposite, the S/E heuristic, was considered for the outer loop. Five TRAMSS algorithms were built using five crossover operators for RCGAs, Arithmetical, Max-min-arithmetical, Discrete, Fuzzy recombination and BLX-ff, which represent different ways in which randomness may be used for generating real-coded genes. The principal conclusions derived from the results of experiments carried out are the following: ffl The TRAMSS model allows the control of useful population diversity to be accomplished for improving accuracy in the case of unimodal functions, and, both reliability and accuracy for the multimodal ones, with regard to RCGAs based on the non-uniform mutation operator. ffl The adaptive control of step size performed by TRAMSS couples suitably with Fuzzy recombination and BLX-ff. Their interactions allow TRAMSS to manage useful diversity, so inducing an effective behavior on all test functions. We suggested that this occurs thanks to the fact that these crossovers adjust the intervals for the generation of genes depending on the current population diversity. ffl With the TRAMSS model, the performance of the strategy of selecting the two best offspring from a set of four with different properties (Max-min-arithmetical crossover) is enhanced. Finally, we should point out that TRAMSS extensions may be followed in two ways: 1) control the parameter associated with the Fuzzy recombination and BLX-ff, respectively, in order to exploit, even more, the profitable combination between TRAMSS and these crossover operators, and 2) study the possible application of dynamic crossover operators, such as the Dynamic FCB-crossovers ([Herrera et al. (1996a)]) and Dynamic Heuristic FCB-crossovers ([Herrera et al. (1996c)]), which are based on the same absolute adaptive heuristics as the non-uniform mutation operator. --R Adaptive and self-adaptive evolutionary computations Evolution strategies I: variants and their computational implementation. Genetic Algorithms in Engineering and Computer Science (New York: John Wiley Evolutionary Algorithms in Theory and Practice (Oxford). Artificial Evolution (Berlin: Springer) Handbook of Evolutionary Computation (Oxford University Press). Adaptive selection methods for genetic algorithms. Reducing bias and inefficiency in the selection algorithm. An Analysis of the Behavior of a Class of Genetic Adaptive Systems Dynamic hillclimbing. The CHC adaptive search algorithm: how to have safe search when engaging in nontraditional genetic recombination. Preventing premature convergence in genetic algorithms by preventing incest. Evolutionary Computation: Toward a New Philosophy of Machine Intelligence (Piscataway: IEEE Press). Genetic algorithms for changing environments. Generalized descent of global optimization Genetic Algorithms and Soft Computing (Heidelberg: Physica-Verlag) In 4th International Conference on Parallel Problem Solving from Nature (Berlin: Springer) Fuzzy connectives based crossover operators to model genetic algorithms population diversity. Tackling real-coded genetic algorithms: operators and tools for the behavioural analysis Representation and self-adaption in genetic algorithms Adaptation in evolutionary computation: a survey. Adaptation in Natural and Artificial Systems (London: The MIT Press). Optimization of calibration data with the dynamic genetic algorithm. Niching methods for genetic algorithms. Selectively destructive re-start Genetic Algorithms The parallel genetic algorithm as function optimizer. Predictive models for the breeder genetic algorithm I. A derandomized approach to self-adaptation of evolution strate- gies The ARGOT strategy: adaptive representation genetic optimizer Technique. Numerical Optimization of Computer Models (Chichester: Wiley). Evolution and Optimum Seeking. Dynamic parameter encoding for genetic algorithms. Strategy adaptation by competing subpopulations. Fast convergence thanks to diversity. Crossover or mutation? Real representations. coding: an iterative search strategy for genetic algorithms. Test driving three Genetic algorithms for real parameter optimization. --TR --CTR Manuel Lozano , Francisco Herrera , Natalio Krasnogor , Daniel Molina, Real-coded memetic algorithms with crossover hill-climbing, Evolutionary Computation, v.12 n.3, p.273-302, September 2004
mutation operator;premature convergence;real-coded genetic algorithms
590937
Probabilistic Pattern Matching and the Evolution of Stochastic Regular Expressions.
The use of genetic programming for probabilistic pattern matching is investigated. A stochastic regular expression language is used. The language features a statistically sound semantics, as well as a syntax that promotes efficient manipulation by genetic programming operators. An algorithm for efficient string recognition based on approaches in conventional regular language recognition is used. When attempting to recognize a particular test string, the recognition algorithm computes the probabilities of generating that string and all its prefixes with the given stochastic regular expression. To promote efficiency, intermediate computed probabilities that exceed a given cut-off value will pre-empt particular interpretation paths, and hence prune unconstructive interpretation. A few experiments in recognizing stochastic regular languages are discussed. Application of the technology in bioinformatics is in progress.
INTRODUCTION Language inference is a classical problem in machine learning, and continues to be an important and active research topic. The basic problem is, given a set of example behaviours or strings, automatically infer a corresponding language (grammar, automata, expression,.) which generates or recognizes those examples. Genetic algorithms (GA) and genetic programming (GP) have been applied towards languages inference, with varying degrees of success. Although successful inference is possible, the generic inference problem is not entirely well-suited for solution by evolutionary search. There are a number of reasons for this. For example, some genome encodings do not preserve useful language characteristics during crossover. Even small local changes to such genomes can be catas- trophic, which does not lend itself well to genetic reproduction and evolutionary search. An even more acute weakness is that "all or nothing" problems such as the language inference problem are not entirely natural for GP. An acceptable language inference minimally requires that the solution language correctly recognize all positive test cases, and reject all negative ones. This essential criteria may also be supplemented by efficiency con- cerns, such as a relatively small number of states or grammar rules. The resulting search space is a difficult one to navigate with evolutionary techniques, due to these stringent requirements for language correctness and completeness. On the other hand, it is generally recognized in the GP community that problems which require an "acceptably close" solution are typically the best candidates for successful solution with GP. Pragmatically speaking, giving the fitness function a larger degree of freedom for evaluating a successful solution will substantially increase the chances of the discovery of acceptable solutions. This research addresses the inference of stochastic regular languages using genetic programming. Stochastic languages are formal languages with probability distribution associated with the language set. The stochastic language inference problem is similar to the classical inference problem, with the additional requirement that the distribution of strings recognized by the stochastic language conform to some desired target distribution. At first, this may seem intuitively more complex than non-stochastic language inference, since it is unclear what impact the determination of probability distributions has on the It turns out, however, that the inclusion of string distributions can simplify the inference problem. Hypothesized languages are now allowed to generate erroneous strings so long as they fall within an acceptably small probability of occurence. In other words, the use of language distributions introduces a more generous degree of freedom for generated solutions. This is ideal in a GP setting, as it simplifies the search space substantially for evolutionary search. The target language used here is a probabilistic regular expression language, henceforth called Stochastic Regular Expressions (or SRE). Although theoretically weaker than stochastic context-free languages studied elsewhere, it was nevertheless chosen due to both its amenability to concise GP representation, and its ability to naturally solve the substantial number of problems in the "regular language domain". The stochastic regular expression language is closely related to stochastic regular grammars and stochastic finite automata, the latter commonly referred to as Hidden Markov Models in the literature. Some SRE language implementation issues had to be addressed before GP could be successfully applied to stochastic regular expression problems. Firstly, an efficient implementation of SRE interpretation was necessary. Interpretation of an SRE expression requires that the probability of recognizing a given string is generated. Since intermediate probabilities would be computed during the interpretation of a string, these values can be used to terminate or prune unproductive interpretation paths whose probabilities are smaller than some supplied cut-off probability. Given the extensive testing that is necessary during fitness evaluation, such pruning greatly increases the speed of GP runs. The SRE language is implemented in a grammatical GP system, which permitted the use of syntactic language constraints to further enhance evolution efficiency. Two example experiments proved that probabilistic language inference is indeed possible with SRE and GP. The more complex of these experiments indicated that the complex search space often resulted in premature convergence. A minor language enhancement to this experiment resulted in failed inferences by the GP system. From this experience, it can be deduced that the fitness evaluation strategy used here is not a general purpose solution to all stochastic language problems, but rather, is suitable to a class of stochastic regular languages whose members are structurally related to one another. An outline of the paper is as follows. Related work is reviewed in section 2. Section 3 defines the syntax and semantics of the stochastic regular expression language, and discusses the algorithm for processing SRE expressions. Section 4 outlines the genetic programming system used. Two example experiments are discussed in section 5. A discussion and future directions conclude the paper in section 6. Formal language induction has a long history as a fundamental problem in machine learning and Booth 1975a, Fu and Booth 1975b, Angluin 1992, Sakakibara 1997). The specialized topic of stochastic languages has also been studied for some time (Fu and Huang 1972). A stochastic grammar differs from a conventional grammar in that each grammar rule is marked with a probability associated with its use, and the set of probabilities for a grammar encode a probability distribution for the resulting derived language. (Fu 1982) has an extensive treatment of stochastic grammars, their derivation, and their application in pattern recognition. Stochastic grammars are also more complex than their non-stochastic kin, as the distributions inherent with the language introduce a new dimension of membership criteria. For example, all context free languages are also stochastic context free languages (all probabilities are 1); however, there may be many stochastic context free languages having essentially the same membership set, but vastly different distributions over that set. Language equivalence issues are therefore more discriminating than in a non-stochastic setting. Stochastic context free languages enjoy both expressitivity and tractable properties, for example, the existence of useful inference algorithms (Lari and Young 1990). They have also found practical use in language processing (Charniak 1993). Stochastic regular languages, albeit descriptively weaker than stochastic context-free languages, have also found their practical niche in applications. Regular languages are definable by finite automata, regular grammars, and regular expressions (Hopcroft and Ullman 1979). Similarly, stochastic regular languages are defined by stochastic versions of these three representations. Examples of work in stochastic grammar inference is in (Maryanski and Booth 1977, van der Mude and Walker 1978, Carrasco and Forcada 1996, Carrasco and Oncina 1998). Stochastic finite automata are defined in terms of Hidden Markov Models (HMM) (Rabiner and Juang 1986). An HMM is a finite automaton with probabilities marking the transition links between nodes. Each node is connected to all other nodes, and so the network itself is maximally connected. When particular transitions are not required, the probabilities associated with those nodes are set to zero. HMMs have found extensive use in language and speech processing (Rabiner 1989, Charniak 1993). Strangely enough, stochastic regular expressions have not been extensively studied; one example paper is (Garg et al. 1996). Language inference has been successfully done using genetic algorithms (GA) and genetic programming (GP). The distinguishing difference between GA and GP approaches is one of denotation: a pure GA uses a binary encoding for the genome, while a GP uses a variable-sized parse tree. Some of the following use encodings with characteristics of both approaches. With respect to regular languages, an early work in evolving finite automata is in (Zhou and Grefenstette 1986). They used a GA with a binary encoding of the automata as a set of state transitions, capped at a size of 8 states. A weakness of this encoding is that the represented automata are susceptible to destructive effects during crossover and mutation. Their unspecified fitness function scores language performance (ability to accept positive strings and reject negative examples) and automata size. (Dunay et al. 1994)'s approach is similar to (Zhou and Grefenstette 1986), except that finite automata are denoted in GP-style nested S-expression notation. (Dupont 1994) uses an automata-theoretic partition representation for regular languages. This has the advantage of preserving language properties of chromosomes during GA reproduction, unlike the more fragile FA represention in (Zhou and Grefenstette 1986). His fitness function scores both language performance and automata size. He successfully evolved a large set of regular languages, including the benchmark Tomita languages (Tomita 1982). (Brave 1997) uses an abstract "cellular encoding" representation for deterministic FA's, which builds the network structure of a FA during interpretation. The intention of this denotation is to preserve structural properties of a language during evolutionary re- production. His automata are embellished with boolean operators which permit automata composition. The fitness function tallied the number of correctly classified sentences. All but one of the Tomita languages were successfully inferred using this technique. (Longshaw 1997) uses a straight-forward state-transition representation for au- tomata. However, his GA uses a population seeded with correct but overly general au- tomata. Specialized reproduction operators manipulate automata by duplicating or refining states. The overall intention is to refine the general automata into a more specific one for the language in question. His fitness function scores example classification performance and automata size. (Svingen 1998) uses a GP on regular expressions. Regular expressions are directly encoded as program trees, and fitness is based on correct example classification. He successfully evolved the Tomita languages. Context-free languages have also been studied. (Wyard 1991) uses a GA to evolve context-free grammars. Chromosomes takes the form of lists of production rules, which guarantees correctness at all times. The fitness function scores example classification per- formance. Two simple CFG's were successfully evolved. (Lankhorst 1994) uses a vector encoding to represent productions. His fitness function is more involved than most others, as it scores example classificaton perfor- mance, the length of substrings of examples correctly classified, the degree of determinism of grammars, and the ability of the grammar to generate correct strings not included in the example set. These additional evaluation considerations give the GA more information with which to drive evolution. He applied the GA to a number of CFG and regular languages. (Lucas 1994) uses a binary-encoded normal form for CFG productions, which preserves language properties during reproduction, and promotes convergence. His fitness strategy scores example classification and grammar size. (Sen and Janakiraman 1992) applies a GA towards inferring deterministic push-down automata, which is an alternative to the grammar representation for CFG's. Fitness scores example recognition performance, and whether the PDA attempts to erroneously 'pop' an empty stack. (Lankhorst 1995) extends this idea towards nondeterministic push-down automata. His fitness additionally considers prefix sizes and the stack size after a string has been consumed. (Dunay and Petry 1995) use a Turing machine representation in their GA exper- iments. Although this powerful notation can denote the entire set of languages in the Chomsky, it does not necessary mean that search will be easy to accomplish, given the inherent enormity of the search space in question. To solve some relatively simple examples of regular, context-free and context-sensitive languages, they used a compositional approach, in which the GA had access to TM building blocks evolved in earlier runs. Finally, the evolution of stochastic languages has been studied. (Schwehm and Ost 1995) uses a GA for evolving stochastic regular languages. Two different encodings are studied - production rules with probabilities, and quotient automata. The fitness function uses grammar complexity (number of productions), a modified - 2 test for distribution conformance, and a measure of the grammar's ability to accept prefixes of the target grammar. A few experiments were performed, and their GA performance compares well with standard regular-language inference algorithms. (Kammeyer and Belew 1997) uses a GA to evolve stochastic context-free gram- mars. They use a liberal representation for grammars in which correct grammars are parsed from the genome when evaluated; this permits intron or junk material to be included in chromosomes. The fitness function evaluates the size of test example prefixes consumed by a grammar, and uses cross-entropy to evaluate distribution conformance. They also use a local search technique for finding production probabilities during evolution. A couple of CFG's were successfully evolved. 3.1 Language Definition The target language for the GP system is stochastic regular expressions, or SRE. The language is very similar to one in (Garg et al. 1996), which is used for modeling the qualitative behaviour of stochastic discrete event systems. Amongst other properties, they prove that probabilistic regular language operations such as choice, concatenation, and Kleene-closure forms a closed language, and hence an algebra. Although a few basic properties will be illustrated here, the reader is referred to (Garg et al. 1996) for further details. It is assumed the reader is familiar with basic concepts from formal language theory (Hopcroft and Ullman 1979). Two language variations, SRE and Guarded SRE (or gSRE), are used. We first define SRE. Let ff range over alphabet positive integers (0 - n - 1000), and f range over decimal values with a precision of 2 decimal places (0 - f ! 1:00). The syntax of SRE is recursively defined as: Without loss of generality, the empty string ffl is not included in the alphabet. The operators have the following meaning: 1. Atomic action ff : The action ff is generated. 2. Choice This denotes a probabilistic choice of terms. Each choice expression can be chosen with a probability: For example, given the expression E (5), the term E 1 can be chosen with a probability of 3=8 and E 2 with a probability of 5=8. 3. Concatenation "E followed by that of E 2 . 4. Kleene Closure can be repeatedly executed 0 or more times, and each iteration occurs with a probability of f . The probability of E terminating execution 5. +Closure once, after which it repeatedly executes 0 or more times using the same probability scheme as Kleene closure. +Closure is an abbreviation for the following: The Guarded SRE language is identical to SRE, except that a guarded choice operator is used instead of the general choice in 2 above: 6. Guarded Choice Here, each term in the choice expression is either prefixed with a unique atomic action that is found nowhere else in the expression, or consists of a unique action by itself. This makes guarded choice deterministic, unlike SRE's nondeterministic choice. Note that, even with guarded choice, gSRE is still a nondeterministic language, since the closure operators are nondeterministic. The rest of the discussion in this section pertains equally to both SRE and gSRE. A derivation of a conventional regular expression E is the set of sentences, or strings over the alphabet, derivable from it. This defines the language L(E) of E. This notion of language derivation is similarly applicable to SRE, except that each string has a probability value associated with it, and hence the language itself is associated with a probability distribution of its members. Alternatively, an intuitive way to consider SRE expressions is that every expression defines a specific probability function over strings in \Sigma : Using a denotational semantics style of representation (Stoy 1977), the probability function for SRE expression E is denoted by [[E]], and its application to a particular string s is denoted [[E]]s, which denotes the probability associated with string s in the language L(E). A probability function model of SRE is now given. Let ffl Atomic actions: (1) ffl Choice (including guarded choice): (2) Since every term might recognize s, the overall probability for a choice expression is the sum of all the term probabilities with respect to s. ffl Concatenation: In the first summation, s is decomposed into two substrings, each of which may be consumed by a concatenated expression. Even though one term may recognize its substring argument, if the other term does not recognize its respective substring, then that term returns a probability of 0, and the overall probability for that instance of decomposition is 0. The rest of the formula represents the cases when one entire expression consumes s, while the other consumes ffl. If these other cases do not succeed, then they return 0. ffl Kleene closure: The first formula accounts for empty strings, as the only way an iterated expression should recognize an empty string is by not iterating. The other formula recursively defines the general case. Here, one iteration of E will consume some portion of s, and the rest of s is consumed by further iterations. The final term in this formula represents when the first iteration consumes the entire string. It is assumed that an iteration of a loop always consumes some non-empty string. Otherwise, the semantic model would have to account for Kleene closure iterating indefinitely on an argument, which is not useful behaviour. ffl +Closure: This is similar to the non-empty argument formula for Kleene closure, except that the expression E will consume part of the string before iterations commence. This can be seen by the lack of f value in the formula. The nondeterministic nature of regular expressions is modeled in the above by multiple argument decomposition in the concatenation and closure operators. Nondeterminism can also arise in the (nonguarded) choice operator. Membership in SRE is reflected by SRE expressions returning non-zero probabilities for particular strings: Definition 3.1 All probability functions pf must adhere to the following two characteristic (Subrahmaniam 1979): (i) for all x i in the sample space of the experiment: (ii) For every event (7)Consequently, if SRE expressions are to define well-formed probability functions, all expressions must similarly respect these requirements. Theorem 3.1 The SRE operators are well-formed probability functions. Proof: The proof uses structural induction on SRE expressions. We show conditions (i) and (ii) of Defn 3.1 hold for all operators. Let s 2 \Sigma . (a) Atomic actions: trivially. (b) Choice: i. From equations 2 and By the induction hypothesis, X Thus we have, X ii. From equation 2, the greatest value for the sum occurs when k. In this case, the sum reduces to Equivalently, when all the summation is zero. And when any 0 ! the resulting summation is a fraction between 0 and 1. Hence it is a probability. (c) Concatenation: i. From equation Using equation 3, because s i ranges over all \Sigma , this becomes: This translates to: By the induction hypothesis, this simplifies to: ii. Given a concatenation, By the induction hypothesis, each of E 1 and Hence their product must likewise be a probability. (d) Kleene closure: i. Starting with equation Using equations 4, it translates as follows: By the inductive hypothesis: Doing some algebraic manipulation: f Note that the division by f \Gamma 1 is permitted because f ! 1 by definition. ii. By induction on the length of a string s, it can be shown that The base case is when ffl, in which case the probability is f from the first equation in 4, and 1. For an arbitrary s 6= ffl, the probability from the second equation in 4 is: By incorporating the second term into the first term's summation, this is rewritten: By the inductive hypothesis over s, a probability. Furthermore, by the structural induction of expressions, probability. Hence their product with f is a probability. +Closure: Similar to (c) and (d) above.3.2 Implementation of an SRE Processor Given a regular expression, determining whether particular strings are members of its language is a tractable problem (Sipser 1996, Hopcroft and Ullman 1979). There are different ways in which this may be performed. One technique is to convert the regular expression into an equivalent nondeterministic finite automaton, which can be done in polynomial time. Once this is done, a graph-searching algorithm reads a string character by charac- ter, marking states of the FA that are still elligible as paths towards an acceptance state. An advantage of the FA approach is that the nondeterministic FA can be polynomially- time translated into a deterministic FA, which will then have more efficient recognition characteristics during language recognition. Alternatively, regular expressions can be symbolically interpreted directly. The behaviour of each regular expression operator has a corresponding operational semantics, which can be used to define a regular expression interpreter. This may be done from the perspective of either language generation or language acceptance. One technical requirement of the expression interpretation approach is that the interpreter must be able to handle the nondeterministic nature of expression derivations, since regular expressions are naturally nondeterministic in nature. The expression interpretation is similar to the FA approach, in that there is a mapping between the states of a translated FA and the derivation paths used by the interpreter when processing an expression. Stochastic regular language recognition uses the same basic recognition schemes as conventional regular languages, with the additional requirement that probabilities be computed for strings. For example, if a FA is derived for a stochastic language, then the links are marked with probabilities. The overall probability of accepting a given string is then computed by computing the product of all the transition probabilities used from the start state to the final accepting state. This probabilistic FA is known as a Hidden Markov Model or HMM (Rabiner and Juang 1986). Therefore, given a stochastic regular language as defined by SRE, the formulae of section 3.1 are incorporated into a translated FA or a The SRE recognition system uses the expression interpretation approach described above. The operational semantics use two relations. One relation, \Gamma! over E \Theta (\Sigma; p) \Theta E, where p is a probability, represents single action transitions of expressions. This relation is denoted, The other relation, =) over E \Theta (\Sigma ; p) \Theta E, is the transitive closure of \Gamma! , and models the generation of strings: Figure contains transitional rules for the relations, which define the structural operational semantics of the SRE operators (Hennessy 1990). These inference rules define an abstract interpreter for SRE expressions, and are the basis of an SRE recognizer. In fact, with languages such as Prolog, these rules can be compiled into Prolog statements, and then directly interpreted using Prolog's inference engine (Clocksin and Mellish 1994). Furthermore, multiple solutions are obtained for nondeterministic SRE expressions using Prolog's builtin backtracking mechanism. The actual implementation of the SRE processor uses the above fundamental ideas. The operational semantics implemented are a superset of the rules in Figure 1. The Action ff F (ff;p) +Closure Figure 1: Transitional semantics of SRE implementation uses a logical grammar definition of SRE, which is part of the DCTG-GP system (Ross 1999) (see Section 4). Prolog's backtracking is advantageously used to investigate different paths of an expression's derivation. In addition, string recognition is performed by pattern matching on an argument string and the generated string as shown in the transitional semantics: when a match occurs, the current derivation path is correct, while mismatches cause the current derivation to backtrack and test another possible nondeterministic path. For example, one instance of backtracking may try different terms in a Choice expression, while another may unwrap an iterative expression a varying number of times. Such backtracking is assured of terminating because of the finite size of input strings to be checked, as well as the assertion within the SRE semantics that empty strings ffl can never be generated within the generative component of iterative operators (they can only be generated when the iteration terminates). One advantage of a stochastic language is that the computed probabilities of strings can be used as an efficiency mechanism during expression recognition. The implementation of the SRE recognizer is such that the probability of intermediate strings are always known throughout the interpreter. When the current probability becomes smaller than a user-supplied threshold, the current derivation path can be forced to terminate. This prunes derivations of an expression which yield probabilities too small to be of conse- quence. Of course, setting this threshold too large results in inaccurate probability values for recognized strings, and may even erroneously reject legal strings. However, for many experiments, especially with large strings to be recognized, this speeds up processing significantly 4.1 Grammatical SRE and gSRE The GP system used for the SRE experiments is the DCTG-GP system (Ross 1999). DCTG-GP performs grammar-based genetic programming, in which the target language of the evolved program population is defined in terms of a context-free grammar (Lucas 1994, Whigham 1995, Wong and Leung 1995, Geyer-Shulz 1997). A major advantage of grammatical GP systems is that the search space is syntactically constrained so that evolution is given a helpful push towards program structures that are more sensible for the problem at hand. The grammar used by DCTG-GP is a definite clause translation grammar, or DCTG (Abramson and Dahl 1989). A DCTG is a logical version of a context-free attribute grammar. Each DCTG production has a syntactic component, which defines a context-free syntax rule. In addition, each production can have included with it one or more semantic components. A semantic component defines some characteristic of syntactic component to which it is attached. For example, one important SRE characteristic that is defined in the DCTG grammar is the string recognition algorithm of Section 3.2. Since the operational semantics of the SRE operators are very modular in nature, their recognition behaviours can be encoded with the grammar rules that define the syntax of the operators themselves. The overall result of this is a compact definition of the SRE language, in which the syntax and semantics are conveniently unified together. One syntactic constraint applied to both SRE and gSRE in the experiments in section 5 is the following. Although not specified in the grammar of SRE (or gSRE), the grammatical definition of SRE disallows iterative operators to be directly nested within one another. In other words, expressions such as are not allowed. The reason for this restriction is a pragmatic one. When GP was performed without this restriction, many programs had multiply nested iterative expressions. Such expressions are relatively expensive to interpret, due to the variety of nondeterministic paths possible for interpreting them. In addition, nested iteration typically results in strings with very low probabilities, since there is a probabilistic factor f associated with executing every nested iterative expression. Moreover, the expense of nested iteration is not justified by results, since any of these expressions can be replaced with a semantically equivalent expression that uses only one iterative operator. This restriction does not imply that an expression like is illegal, since the concatenation operator means that the iteration operators are not directly nested. gSRE is also encoded as a syntactic constraint of SRE. Rather than permitting any SRE expression as a term within a choice operator, only uniquely guarded terms are permitted. 4.2 Other GP System Details DCTG-GP uses standard GP strategies, such as tournament or roulette-wheel selection, and steady-state or generational evolution. Relevant experimental parameters will be illustrated in section 5. The system is implemented in Sicstus Prolog 3 on both Windows and Silicon Graphics platforms. 5.1 General Strategy The inference of a stochastic language can be considered to involve two different objec- tives. Given a training set of positive (and possibly negative) examples, one task is to infer a language which correctly classifies the training examples. This is equivalent to non-stochastic language inference. An additional task required for stochastic language in- ference, however, is to ascertain the stochastic distribution of the training examples. One might naively presume that a statistical analysis of the training set could be performed, and the results applied to the inferred language. Unfortunately, the situation is typically more complicated than this, because the representation of the stochastic language as used in the hypothesis will not likely permit a straight-forward application of the final string distributions to its internal encoding. For example, if an HMM representation is used in hypotheses, finding appropriate probability values for intermediate links in the network that will correspond to the example set distribution is a challenging task. The significance of the problem of determining distributions for HMM's and context-free languages has spawned specialized training algorithms (Lari and Young 1990, Charniak 1993). The inference strategy undertaken with the GP experiments is to let evolution determine stochastic distributions in concert with example classification. Since SRE incorporates probability values directly in expressions, treating numeric probability fields is straight-forward in GP. It was found that this approach was sufficient for many experiments undertaken. In fact, it was discovered that evolution using local search for fine-tuning probability parameters lent no advantage over simple evolution of the parameters The training sets used in the GP experiments consist of sets of positive examples for the target language to be inferred. Each member of the set is a string, along with its frequency with respect to the total number of strings in the set (typically 1000). Since the format of the target languages is already known via a stochastic regular expression or gram- mar, generating these sets is straight-forward. Unlike conventional language inference, the probability distributions in training example sets permits stochastic languages to forgo the need for negative examples. This is because the inference of a distribution that matches that of the training set will automatically account for 'negative examples', which have 0 probability in the distribution. Stochastic language inference incorporates an implicit degree of error in any inferred solution. This has ramifications on the GP fitness evaluation described below. It also can be used to boost efficiency of computations performed during inference. As detailed in section 3.2, string recognition can be pre-empted when intermediate probabilities become smaller than some threshold limit set for the experiment. Similarly, the test set can be pruned of strings whose frequency is below some limit set by the user. This limit parameter should be set with the recognition threshold in mind. For example, if the threshold is set to 0:001, then the test set limit could be likewise set to 1 for a test set of size 1000. Of course, there may be many nondeterministic derivations of an expression when recognizing a string, and all the probababilities of these derivations will be summed to an overall probability for that string. The less discriminating the recognition threshold and test set limit, the more precise (albeit slower) the results. Since GP experiments use a steady-state algorithm, there are not any discrete generations. For convenience, however, a new generation is said to have occurred every K reproductions, where K is the population size. Between generations, the test set is regenerated. This prevents overfitting to one set of test data, and reflects the nature of the stochastic languages, as each test set reflects a sampling from the actual distribution. One disadvantage, however, is that a discrete test set is an approximation of the real distribution of the language, and hence this introduces an unavoidable measure of noise. This noise is compensated by the fact that multiple test sets are used during successive generations, and their cumulative effect should reflect a more accurate model of the target distribution. However, the population is not reevaluated for each newly generated test set, and so the fitnesses of much of the population may be legacy values from earlier generations. This is acceptable, because the test sets used for those generations are presumed to be as statistically valid as those from any other generation. The fitness evalution strategy used in the experiments is a modified - 2 test (Press et al. 1992). The known distribution is taken to be the set T of test examples, and the experimental set will be the results of the SRE recognition algorithm on each member Each test set example string is given to the SRE processor, and an overall probility that string is computed. Non-membership is reflected in a probability of 0. The fitness formula is: where d i is the frequency of example t i in test set T , is the maximum prefix of t i recognized. The first term is the - 2 formula, and it is used when the example string t i is completely recognized. The second formula is used when only a prefix of t i is recognized. Its value is inversely proportional to the size of the prefix recognized. Should none of t i be recognized, then this value becomes 2 \Delta d i (a normal formula would use just d i ). This prefix scoring gives credit to expressions that recognize portions of the examples, which helps drive evolution towards expressions that recognize complete examples. 5.2 Experiment 1: Stochastic Iteration The first experiment uses a simple stochastic regular language which can be naturally encoded in SRE. The main intention of this experiment is to test the evolvability of stochastic Kleene closure as modeled in SRE. The target language is a stochastic rendition of a regular language suggested by Tomita 1 from his popular benchmarks for machine learning (Tomita 1982). The target language written in SRE is: This is a non-trivial language, especially in the stochastic domain, as the overall distribution of each a and b term in all the strings should conform to the given probability of 0:5. These terms may also generate empty strings, should iterations terminate immediately. The parameters for the experiment are in Figure 1. Most are self-explanatory, and the fitness function strategy was discussed earlier in Section 5.1. The initial population is oversampled, and the running population is pruned from it using tournament selection. Replacement is done using a reverse tournament selection (a sample of K members are randomly selected, and the member with the lowest fitness is selected to be replaced). His language is a b a b . Table 1: Parameters Parameter Value Target language gSRE Fitness function modified - 2 Generation type steady-state Initial population size 750 Running population size 500 Unique population members yes Maximum generations 50 Probability of crossover 0.90 Probability of mutation 0.10 Probability internal crossover 0.90 Probability terminal mutation 0.75 Probability numeric mutation 0.50 mutation range \Sigma0.1 Max reproduction attempts 3 Initial population shape ramped half&half depth initial popn. 6, 12 Max depth offspring 24 Tournament size 5 Test set size 1000 test string size approx. 20 Min test example frequency 3 probability limit 0.0001 Mutation is performed on either terminal or nonterminals. If a nonterminal is to be mutated, there is a 0:5 probability that it should be a numeric field. When a numeric field is selected for mutation, its current value is perturbed \Sigma10% of the entire range for that numeric type (a range of \Sigma100 for integers, and \Sigma0:1 for probabilities). A test set is generated before every generation. Initially, 1000 strings are generated for L 1 , and their frequencies are tallied. The maximum string size is approximately 20 (some may exceed this length). Should there be less than 3 instances of a given string, it is pruned from the test set. This means that there are typically between 55 to strings in the test set, each of which has its particular frequency for that particular sample of the language. The number of unique strings in the test set is important for - 2 analyses, as it is equivalent to bin size in the - 2 formula. Table 2: Summary L 1 Total runs 15 # unique examples Avg. test set - 2 142.22 (50 cases) Fitness min 89.4 (- 2 =88.8) Generation Fitness Average Figure 2: Fitness curves (avg 15 runs) Summary statistics for the best solutions from 15 runs are given in Figure 2. These values are obtained using a common test set, since each run will have used a different test set during its prior evolution. An average - 2 test of the test sets themselves is included, in order to better evaluate the expression results. 50 pairs of random test sets were generated. One of the pair was fixed as the independent variable, while the other was the dependent variable. The sets were filtered for frequencies below the minimal test example frequency in Figure 1), and the - 2 was computed. The resulting 50 - 2 values were averaged. A performance chart of the best and average population fitness averaged for 15 runs is in Figure 2. It can be seen that convergence to a local optimum has largely occured by generation 10. The best solution found a This is a nearly perfect solution, and the iterative probabilities within the range of what might be expected given the stochastic error inherent with the random test sets. The second best solution (- 2 =89.63) is: a The last term is interesting, in that the erroneous choice of a is not too acute a problem, given the low probability of choosing it (0.11). One of the poorer solutions (- 2 =132.85) is: The inaccuracy occurs with the first term, which erroneously permits b to occur too fre- quently, even though the low probability of 0:25 for the enclosing iteration helps reduce its likelihood. The worst solution obtained (- 2 =203.75) is: (a Note the repetition of particular numeric fields, such as 320 and 0.04, which is a sign of population convergence. Simplifying this expression by removing iterative probabilities less than 0.10 and expanding +Closure terms, it becomes: a which is obviously a suboptimal solution. This example shows the nature of introns within expressions: virtually any expression can be intron code, so long as the associated choice or iterative probability is low enough. a A (0.4) Figure 3: Target language 5.3 Experiment 2: Stochastic Regular Grammar The second experiment evolves a more complex stochastic regular language. The target language L 2 is taken from (Carrasco and Forcada 1996), and is defined by the stochastic regular grammar in Figure 3. Each production has a probability on the right, which denotes the probability that rule is selected with respect to the other productions for that nonterminal. Table 3: Summary L 2 Total runs 50 # unique examples 35 Avg. test set - 2 99.75 (50 cases) Fitness min 66.39 (- 2 =65.06 ) The experimental parameters for these runs are identical to those in Figure 1. The summary for 50 runs are in Figure 3. A performance plot for the best fitness and average population fitness averaged for the 50 runs is in Figure 4. The best solution (- Simplifying by expanding +Closures and removing terms with probabilities less than 0.03, Generation Fitness Average Figure 4: Fitness curves (avg 50 runs) this becomes: It is difficult to see how this expression maps to the target grammar of Figure 3, and an intuitive mapping may not even exist. However, its - 2 is impressive compared to the test set average. 5.4 Limitations Many language inference algorithms are easily thwarted by target languages having characteristics antagonistic to the peculiarities of the algorithm in question. Often, these languages are only subtlely different from ones that the algorithms have no problems inferring The GP paradigm suffers a similar limitation. A variation of the language in section 5.3 was tried, which is just language L 2 with an additional string bbaaabab with probability 10%. 50 runs were performed using the same parameters as figure 1. None of the runs found an acceptably close solution: the best solution had a fitness of 259 and - 40). One reason that GP had problems evolving L 0 2 can be attributed to the linguistic characteristics of SRE. Even though the above definition of L 0is a concise statement of the language, the evolutionary process tries to unify the term bbaaabab and L 2 together in a regular expression. This is difficult to do, because this string is an anomaly with respect to the other strings in L 2 . Considering the stochastic regular grammar used to generate is clear that strings are derived progressively and incrementally from one another, and so strings of L 2 equal in length to bbaaabab are natural extensions of smaller strings of the language. The anomalous string, however, is not derivable from L 2 , and hence a natural model of the union of these languages in SRE cannot be inferred. This is especially true given that bbaaabab has a 10% probability, which makes it a populous member. If it had a smaller probability, it might be ignored as noise. The above must be considered in light of the linguistic nature of all formal lan- guages: some representations more naturally model particular languages than others. Even though regular expressions, finite automata and regular grammars have the same expressive power, some languages are more naturally and concisely denoted by regular expressions than by finite automata, and vice versa. It could be the case that another representation language, for example HMM's, may more naturally denote L 0 than SRE. 6 CONCLUSION This paper presented a new means for evolving stochastic regular languages. Using a probabilistic version of regular expressions as a language for evolution, genetic programming is capable of evolving accurate expressions for stochastic regular languages. However, some stochastic regular languages are more amenable to successful evolution than others. It can be speculated that languages in which members have structural similarities with one another are the most suitable for this paradigm. For more complex languages, more sophisticated evolutionary techniques may be required. It was found during experimentation that SRE had no evolutionary advantage over gSRE with respect to the quality of solutions discovered. On the other hand, SRE expressions were less efficient to process, and runs took much longer than the gSRE ones. The use of SRE in a genetic programming context presents advantages over other evolutionary experiments with stochastic languages. One advantage is that SRE is akin to a programming language, with operators that have syntactic and semantic definitions akin to conventional languages. Since GP is typically applied towards such languages as Lisp, the encoding and processing of SRE within a GP environment is straight-forward. More importantly, however, is that SRE has linguistic advantages over finite automata and regular grammars: some stochastic languages are more naturally encoded in SRE than these other representations. The L 1 experiment is a clear example of this point. The linguistic clarity of L 2 is less apparent, although the solution is not overly complex compared to the target grammar. Like (Svingen 1998), this work uses a regular expression language directly for GP. His work required fairly large populations and parallel populations in order to evolve the Tomita languages. The fitness strategy used here is similar to that used in (Lankhorst 1994, Schwehm and Ost 1995), in that both language recognition performance and prefix consumption are taken into consideration. There are many directions for future work. The GP strategies used here were fairly conventional, and more sophisticated approaches may be more applicable to stochastic languages. In the experiments, the wide degree of qualitative variations between runs indicates that evolution quickly gets stuck at suboptimal solutions. Parallel subpopulations may help in this regard. Although it was found that local search using hill-climbing over numeric fields was not advantageous to evolution, it is worth investigating the utility of more sophisticated local search techniques akin to those used in stochastic context-free languages (eg. the inside-outside algorithm). Currently, the applicability of SRE in bioinformatics problems is being investi- gated. A fundamental problem in DNA and protein sequencing is to determine a common pattern shared amongst a family of sequences (Brazma et al. 1995), which can be used for both search and analytical purposes. A number of techniques, such as HMM's and regular pattern languages, are used for this purpose. SRE is a natural vehicle for this problem area, since its regular expression basis conforms to the pattern languages commonly used (eg. that used in the PROSITE database (Hofmann et al. 1999)), while its stochastic features conveniently model the probabilistic characteristics of DNA sequences themselves. Acknowledgement Thanks to Tom Jenkyns for helpful discussions about probability. This research is supported by NSERC Operating Grant 138467-1998. --R Logic grammars. Computational Learning Theory: Survey and Selected Bibliography. Evolving deterministic finite automata using cellular encoding. Approaches to the automatic discovery of patterns in biosequences. Learning Deterministic Regular Grammars from Stochastic Samples in Polynomial Time. Inferring Stochastic Regular Grammars with Recurrent Neural Networks. Statistical Language Learning. Programming in Prolog (4th Solving Complex Problems with Genetic Algorithms. Regular Grammatical Inference from Positive and Negative Samples by Genetic Search: the GIG method. Syntactic Pattern Recognition and Applications. Stochastic Grammars and Languages. Probabilistic Language Framework for Stochastic Discrete Event Systems. The Next 700 Programming Languages for Stanford University The Semantics of Programming Languages - An Elementary Introduction Using Structural Operational Semantics The database Introduction to Automata Theory Stochastic Context-free Grammar Induction with a Genetic Algorithm Using Local Search Grammatical Inference with a Genetic Algorithm. A Genetic Algorithm for the Induction of Pushdown Automata. The estimation of stochastic context-free grammars using the Inside-Outside algorithm Evolutionary learning of large grammars. Structuring chromosomes for context-free grammar evolution Inference of Finite-State Probabilistic Grammars IEEE Transactions on Computers C26 Numerical Recipes in C. An Introduction to Hidden Markov Models. Recent Advances of Grammatical Inference. Inference of Stochastic Regular Grammars by Massively Parallel Genetic Algorithms. Introduction to the Theory of Computation. Denotational Semantics. A Primer in Probability. Learning Regular Languages Using Dynamic construction of finite automata from examples using hill- climbing On the Inference of Stochastic Regular Grammars. Information and Control Learning Programs in Different Paradigms using Context Free Grammar Induction Using Genetic Algorithms. Induction of Finite Automata by Genetic Algo- rithms IEEE Press. --TR --CTR Ashok Argent-Katwala , Jeremy T. Bradley , Nicholas J. Dingle, Expressing performance requirements using regular expressions to specify stochastic probes over process algebra models, ACM SIGSOFT Software Engineering Notes, v.29 n.1, January 2004 Rolv Seehuus , Amund Tveit , Ole Edsberg, Discovering biological motifs with genetic programming, Proceedings of the 2005 conference on Genetic and evolutionary computation, June 25-29, 2005, Washington DC, USA Brian J. Ross, The evolution of stochastic regular motifs for protein sequences, New Generation Computing, v.20 n.2, p.187-213, April 2002
genetic programming;stochastic regular expressions
590954
Decision-Theoretic Planning for Autonomous Robotic Surveillance.
In this paper, we introduce a decision-theoretic strategy for surveillance as a first step towards automating the planning of the movement of an autonomous surveillance robot. In our opinion, this particular application is interesting in its own right, but it also provides a test-case for formalisms aimed at dealing both with (low-level) sensor, localisation, and navigation uncertainty and with uncertainty at a more abstract planning level. After a brief discussion of our view on surveillance, we describe a very simple formal model of an environment in which the surveillance task has to be performed. We use this model to illustrate our decision-theoretic strategy and to compare this strategy with other proposed strategies. We treat several simple examples and obtain some general results.
Introduction In several projects, robots are employed to perform surveillance tasks. See, for example, [5, 6, 7]. However, in most of these projects, the robots are typically used as a kind of flexible sensor platform controlled by some external human operator who makes the high-level decisions on where to go, and how to use the on-board sensors. We are interested in automating this high-level decision making process to allow an autonomous mobile robot plan itself how to perform the surveillance task. Little work has been done in this area, with the the notable exception of [2, 3]. We believe this autonomous robotic surveillance application is interesting in its own right, but it also provides a test-case for formalisms aimed at dealing both with (low-level) sensor, local- isation, and navigation uncertainty and with uncertainty at a more abstract planning level. In section 2, we give our view on what surveillance is. We then give in section 3 a simple formal model of an environment in which the surveillance task has to be performed. In section 4 we describe several surveillance strategies that have been studied in [3], and introduce our proposed decision-theoretic strategy. Finally, section 5 contains some examples illustrating the various strategies and some preliminary results. Surveillance In dictionaries, surveillance is defined as a close watch kept over something or someone. We feel one should add that the purpose of this close watch is to detect the occurrence of some relevant events. Which events are considered relevant depends on the overall reason for the surveillance. Of course, this close watch is not limited to visual means. In principle, any kind of sensor can be used. Also, the person or thing to be closely watched should be understood to include areas, places, (parts of) aerospace, et cetera. Familiar examples of surveillance tasks include the observation of suspected members of a criminal organization by the police, the detection of airplanes in a no-fly zone by the military, and the look-out for suspicious behaviour of people in a shop by a security-guard. There are at least two disctinct difficulties involved in detecting relevant events. 1. The relevant events have to be in sensor range. 2. Once detected, the events have to be recognised as relevant. For some applications, such as the observation of a place which is in plain view of a (static) camera (see e.g. [1]), the first difficulty does not arise. The difficulty that remains here is to recognise or classify particular events as belonging to the class of relevant events. Such a classification may vary from rather straightforward (e.g. speeding cars) to quite complex (e.g. suspect behaviour of people in a shop). Although the issue of automatic classification of events is extremely important, and essential for artificial intelligence, we concentrate our research on the first difficulty. As soon as the relevant events are not guaranteed to occur within sensor range, it becomes essential to get the available sensors in such a position that the relevant events are (most likely) detected. Mobile robots seem to be excellently suited for automating the process of moving sensors, since robot navigation and self-localization, although still not trivial, are nowadays considered fairly standard tasks. Of course, the difficulty of navigation and self-localization, and therefore of surveillance, depends on the type of environment in which the tasks have to be performed. But some successes have been obtained in a variety of environments, ranging from structured in-door environments, like office buildings, to out-door environments, like aerospace. The use of robots in surveillance tasks is not new, but in most of the existing projects in this area, the robots are typically used as a kind of flexible sensor platform controlled by some external human operator who makes the high-level decisions on where to go, and how to use the on-board sensors. We are interested in automating this planning task to make the surveillance robots more autonomous and less dependent on human control. Our main concern is to develop strategies or algorithms that make use of the available recources (sensors, manoeuvring capabilities, et cetera) such that the probability of relevant events remaining undetected is minimised. (This is only a preliminary description of the goal of surveillance. See below.) The fact that detecting relevant events does not guarantee them to be recognised as relevant, can (and should) be modelled by introducing some uncertainty in the classification of events. When evaluating surveillance strategies one should compare them with alternative strategies, but also with alternatives to robotic surveillance, such as the use of security cameras, smoke detectors, et cetera. Of course, one can also consider combining robotic surveillance with simple static sensors and, for example, use a surveillance robot to check out alarms raised by smoke detectors to decrease the amount of false alarms reported to police and fire departments. An important issue when evaluating surveillance strategies is to clarify the goal of surveillance. We cannot simply say that the goal is to detect all relevant events, since it is not possible to develop strategies that guarantee to achieve this goal in case the relevant events are not necessarily in sensor range. We used as a preliminary description that the probability of relevant events remaining undetected should be minimised. This is similar to the description used in [6]. This description should at least be refined to take into account the fact that some relevant events can have more serious consequences than others. near a storage room of highly inflammable toxic material is more serious than a flooding of a rarely used basement with no electric outlets.) This can be done by saying that not detecting a relevant event involves some cost, which may vary depending on the event, and that one should minimise the expected cost of not detecting relevant events. Using such a decision-theoretic criterion has a number of advantages. For example, at least from a theoretical point of view, it seems obvious how to further refine this criterion to incorporate other types of cost, such as the danger of damage to the robot, or the cost of false alarms (in case the recognition of relevant events is not completely reliable). However, several serious problems remain. The first problem is that the expected cost of the whole surveillance plan or decision policy should be minimised. In general, such a policy cannot be reduced to a sequence of simple decision problems, but involves looking (far) ahead at (many) possible outcomes of (many) future actions. Therefore, in realistic applications one will probably need heuristics and approximate methods to arrive at feasible solutions. Another problem is that the application of decision theory requires a great amount of data. The necessary probabilities of events and costs of possible outcomes of actions are not always readily available. This is a good reason for preferring policies which are robust, in the sense that they are not influenced much by small changes in the data. Also, it can be argued that in realistic applications one should settle for satisfactory policies rather than insist on optimal policies. It should be mentioned that we explicitly adopt an asymmetrical view on surveillance, in the sense that the goal can be described as detecting (or minimising the cost of not detecting) relevant events. In [3], the goal of surveillance seems (at least sometimes) to be understood as maintaining a maximally correct model of the state of the environment with respect to both the presence and the absence of relevant events. We call this the symmetrical view on surveillance. We believe that for a surveillance robot with, for example, detecting fires or intruders as primary goal, the asymmetrical view on surveillance is the most appropriate, and the decision-theoretic formalisation of the goal is the most natural. In [4], the issues mentioned in this section are discussed in more detail. 3 A Simple Environment Model In this section we introduce a very simple, formal model of the environment for the surveillance task. Definition 3.1 An environment E is a tuple hX; A is a set ng of mutually disjoint spatial areas, or locations, A 0 2 X is the start location of the robot, A ' X \Theta X represents the relation of immediate accessibility (for the robot) between locations, C is a function assigning to each location X i 2 X the cost c i associated with not detecting a relevant event occurring at X i , P 0 is a function assigning to each location X that at time 0 an event occurs at is a set of transition probabilities. That is, for every contains denoting the (prior) probability that at time t a relevant event starts at denoting the (prior) probability that the relevant event at stops at time t. For the rest of this paper, we assume that for every A. As a consequence, we can simplify the description of an environment Further, we assume that P t (X do not depend on the time t, and we drop the subscript t. We also assume that the environment is connected, in the sense that for every there is a path We write r t to denote the sensor range of the robot at time t, and we assume that then we say that is visited at time t. The decision strategies should decide which immediately accessible location to visit next. For the moment, we do not take recognition uncertainty into account and assume a relevant event to be detected whenever it is in the sensor range of the robot. The above definition provides a very abstract model of the decision problem. For realistic applications, we have to take into account that a robot can have several sensors, each with its own sensor range, that the sensor range is not necessarily an element of X , that the actions of the robot may include changing its location, its orientation, and possibly manipulating aspects of the environment, such as opening a door, and that the exact state and dynamics of the environment, the exact position of the robot in the environment, and the recognition of relevant events are uncertain, et cetera. See [4] for a more detailed discussion of the simplifying assumptions of the given model. In spite of the many simplifying assumptions, the model is sufficiently general to capture the abstract environment used in [3] to experimentally compare different surveillance strategies. In that experimental set-up, all locations are assumed to be immediately accessible from each other, i.e., . The dynamics at each location is modelled by a Markov process, and P (X are viewed as transition probabilities between possible states of a location. In our opinion, for many typical surveillance applications it is not appropriate to model the dynamic behaviour of a location X i as a Markov process. For example, although the start of a fire can be assumed to have an extraneous cause (nature or some agent other than the surveillance agent), it is typically the case that the detection of a fire influences the probability of the fire being extinguished. The detection of a relevant event should trigger some appropriate response of the surveillance agent, such as raising an alarm, or taking more direct countermeasures against the event. Not all applications of surveillance are meant to trigger intervening responses to the observed relevant event. For example, observations made in the context of a scientific study are primary aimed at information gathering, not at intervening. However, when interventions do play a role their effects should be incorporated in the model of the surveillance problem. Since the particular actions triggered by a detection are them- selves, strictly speaking, not part of the surveillance behaviour of the agent, we will leave them out of our considerations. In our examples, we assume that relevant events do not stop spontaneously, but that they stop immediately when detected by the surveillance agent. Formally, P (X 0, and P t+1 (X It is possible to introduce some time delay for the countermeasures to take effect, but this raises the problem of deciding how important it is to monitor areas where relevant events are known or have been observed to occur. It is also possible to allow P (X and to model the effect of the actions triggered by observing a relevant event as an increase in P (X Our assumption can be viewed as an extreme instance of this possibility. Given our assuptions, it is possible to express P t (X in terms of P (X and the amount of time that has passed since the last visit to X i . Proposition 3.1 be an environment where P implies that P t+1 (X , where t 0 is the largest time point - t such that r t 4 Surveillance Strategies In [3] a surveillance strategy is proposed based on the newly introduced notion of confi- dence, which can be viewed as a second-order uncertainty measure. Whenever sensory information about the state of a location becomes available, the probability of an event occurring at that location at that time is updated, and one is assumed to be very confident about this assessment of the state of the location. This confidence then drops gradually over time during a period in which no fresh sensory information concerning this particular location is obtained. The rate by which the confidence decreases depends on the transition probabilities: the more likely the changes, the higher the decrease rate. Specifically, the factor - p is used as confidence decrease rate, where p is the transition probability leaving from the observed state and - is some unspecified parameter. The actually used computation of confidence is slightly more complicated, due to the fact that some time after the observation it is no longer clear which transition probability should be used in the computation of the decrease rate. In our model, the situation is simpler, since we assumed that when visiting a location X i at time t, the robot either observes that no relevant event occurs at X i or the robot immediately stops the relevant event. In both cases, the robot can be confident that no relevant event is going on at X i after t. This confidence can decrease over time due to the possibility that a relevant event starts after t. This rate of this decrease of course depends on P (X 1). The transition probability P does not play a role. Since the factor - P (X i !1) is meant to be a decrease rate, one can infer that Lemma 4.1 Let my. It follows that if one assumes that by visiting X i the confidence that no relevant event is going on at X i becomes 1, then the location with the lowest confidence at time t is the location X i such that P (X is the time of the last visit to X i . The policy proposed in [3] can be described as follows. maximum confidence Choose the action that changes the sensor range to the neighbouring location which has the lowest degree of confidence attached to it. This policy is experimentally compared to the following policies. random exploration Randomly choose a location as the next sensor range. methodical exploration Choose all the locations, one after the other, and always in the same order, as the sensor range at the next moment. maximum likelihood Choose the action that changes the sensor range to the neighbouring location with maximal uncertainty, where the uncertainty at location X i is measured by min(P (X Notice that both random and methodical exploration, as described above, allow choosing non-neighbouring locations. Actually, in the experiments of [3] it is assumed that all locations are directly accessible from each other X). This is only realistic in case changing attention to a far removed location involves no or only a negligible amount of cost or time. It is of course not difficult to restrict random exploration to choosing randomly between neigbouring locations only, but it is not clear how to put a similar restriction on methodical exploration. One possible policy that can be considerd to be a local variant of methodical exploration is the following. interval Minimise the maximum time interval between visits of locations by choosing the action that changes the sensor range to the neighbouring location which has not been visited for the longest time. We propose to use this minimax interval policy as a kind of reference strategy. Since this strategy does not use information about the uncertainties, it can be used to clarify how much other strategies which do use uncertainty information gain in efficiency. It should be mentioned that in the case of the maximum likelihood policy many uncertainty measures, including, for example, entropy, give rise to the same preferences as min(P (X As we will see in section 5, the maximum likelihood policy seems more appropriate for (symmetrical) surveillance understood as maintaining a maximally correct model of the state of the environment with respect to both the presence and the absence of relevant events, than for (asymmetrical) surveillance aimed at detecting relevant events. In [3], no explicit choice is made between such a symmetrical view on surveillance and the asymmerical view we take. Several criteria are used to evaluate the performance of the strategies in the experiments, including the (symmetrical) criterion of the percentage of erroneous estimations of the state of each location and the (asymmetrical) criterion of the percentage of non detected relevant events. We propose a surveillance strategy based on decision-theoretic considerations. By decision-theoretic surveillance we understand the kind of behaviour guided by the following decision policy. minimum expected cost Choose the action that minimises the expected cost. This decision policy can be interpreted both globally and locally. Under the global interpretation, the action that has to be chosen corresponds to the behaviour of the surveillance agent from the start to the end of the surveillance task. There is not an inherent end to a surveillance task, but in practice each particular task has a limited duration (say, until the next morning when the employees return to the office building, or until the batteries of the robot have to be recharged). The (global) expected cost EC T until time T can be computed by the following formula. Notice that a choice to visit X i at t not only removes the term P t (X the above sum, but it also has some indirect benefits, due to the fact that it reduces t. The behaviour of the surveillance agent from the start to the end of the surveillance task can also be viewed as consisting of a sequence of simpler actions. One can apply the above decision policy locally to choose at each time between the possible simple actions by comparing the consequences of these simple actions, or perhaps by comparing the (expected) consequences of small sequences of simple actions. Let us say that an n-step policy compares the (expected) consequences of sequences of n (simple) actions. Of course, the policy is more easily implemented for small n, whereas, in general, it better approximates the global policy for large n. Since the goal of surveillance should be interpreted as global minimisation of expected cost, it will be interesting to study under what conditions the local policy is already (sufficiently) equivalent to the global policy for small n. None of the policies considered in [3] takes into account a notion of cost. If the cost C(X i ) of not detecting a relevant event is the same for all X i , then minimising the expected cost reduces to maximising the probability of detecting a relevant event. Notice that this is still different from the maximum likelihood policy. 5 Examples and First Results We have defined three decision policies for surveillance which make use of some kind of uncertainty information, namely maximum confidence, maximum likelihood, and minimum expected cost. The following proposition shows that these policies essentially agree if there is no (relevant) uncertainty information to be used. Proposition 5.1 Let Assume that for all Then the maximum confidence policy and the one step minimum expected cost policy both reduce to the minimax interval policy. Also, for sufficiently small transition probabilities P 1), the maximum likelihood policy will agree with the minimax interval policy. It follows that we can expect a difference between the mentioned policies only in the case of varying probabilities (or costs). Our first example illustrates the difference between the various surveillance strategies introduced so far. Example 5.1 Consider an environment is a set consisting of two rooms, 0.8. The strategy based on maximum likelihood will always look at room X 1 (where the uncertainty is maximal), and will never take a look at room X 2 . The strategy based on methodical exploration goes back and forth between both rooms, just as the maximum confidence policy does. The strategy based on one step minimising expected cost is slightly more complicated. again 0.8, since room X 2 was visited at the immediately preceding time step. However, P only increased to 0.75, which is not enough to get chosen. Only at time 3, P increased above the 0.8 probability that a relevant event occurs in room 2. We thus obtain a sequence where room X 1 is only chosen every third time step. See table 1, where for the first six time steps the expected costs (of not visiting a room) are displayed. The one step minimum expected cost policy chooses the room with the (maximal) expected cost printed in boldface. Table 1: The expected costs of example 5.1. In this example, the maximum likelihood policy does not result in an exhaustive exploration of the environment. Both maximum confidence and minimum expected cost behave better in this respect. The problem with the maximum likelihood criterion is that P (X not guaranteed to result in a preference to visit X i rather than X In fact, if P (X then the criterion prefers X j . Notice that not only in the artificial example above, but also in practical applications, with sufficiently many locations and sufficiently high transition probabilities, it can happen that P (X Since the maximum likelihood criterion does not prefer locations where the chance of detecting a relevant event is high, but is more interested in locations where the occurrence of a relevant event is highly unknown, we conclude that the criterion is more appropriate for symmetrical surveillance than for asymmetrical surveillance. In example 5.1, the one step minimum expected cost policy results in a behaviour which seems intuitively appealing, since it clearly reflects the fact that P (X substantially lower than P confidence, just as methodical exploration, treats both rooms the same. However, we will see below that this intuitive appeal may be somewhat misleading. The maximum confidence policy does also take the probabilities P account, since the rate of confidence decrease is a function of these probabilities. However, the decrease rate proposed in [3] does not result in a different treatment of both rooms in the example. Thus one can view the example as an indication that in the minimum expected cost policy the probabilistic information is used more directly and taken more seriously than in the maximum confidence policy. The maximum confidence policy does not consider at all the possibility that for some areas it may be relatively more important to detect events. This is easily implemented in the minimum expected cost policy by letting the cost C(X i ) of not detecting a relevant event depend on the area X i . Such varying costs may cause a problem, since they may prevent the one step mimimum expected cost policy to obtain an exhaustive exploration of the environment. It can be shown that if the cost of not detecting a relevant event is constant over the different areas X i , then, in the long run, the one step mimimum expected cost policy will result in an exhaustive exploration of the environment. More precisely, one can show the following. Proposition 5.2 Let Assume that for all in the long run, every X i 2 X is visited when applying the one step mimimum expected cost policy, and there is a finite upper bound N i on the length of the time interval between visits of X i . If relevant events will not stop spontaneously before they are detected, exhaustive exploration implies that all relevant events will eventually be detected. But even among policies that are 100% successful with respect to (eventually) detecting relevant events there may be a difference in performance if, for example, early detection is considered to be important. Table 2: The expected costs of example 5.2. The following simple modification of example 5.1 shows that, in general, proposition 5.2 is no longer valid (and the one step minimum expected cost policy is no longer guaranteed to result in an exhausite exploration of the environment) if the costs are allowed to vary. Example 5.2 Consider the situation of example 5.1, but now assume C(X 1 3. Then the expected cost of not visiting room 1 has an upper bound of 1, whereas the expected of not visiting room 2 is 2.4 even when it has been visited the previous moment. Therefore, by the one step minimum expected cost policy, room 2 will always be chosen. See table 2, where for the first four time steps the expected costs (of not visiting a room) are displayed. The one step minimum expected cost policy chooses the room with the (maximal) expected cost printed in boldface. This effect of ignoring room 1 can be avoided by allowing the cost of not detecting an event to grow as a function of the time passed since the event has started. However, if the mentioned (expected) costs are correct, then this ignoring of room 1 may be defensible. Intuitively, one should not require a surveillance agent to explore irrelevant or unimportant areas of the environment. This is substantiated by the following. Proposition 5.3 In the environment of example 5.2 the one step minimum expected cost policy minimises the global expected cost. Perhaps a more important problem than possibly preventing an exhaustive exploration of the environment is that the varying cost can form an obstacle to obtaining optimal behaviour using the local (one step) mimimum expected cost policy. Example 5.3 Consider an environment is a set consisting of three rooms, As in example 5.2, the one step minimum expected cost policy will always choose room 2. But now the (possibly justifiable) ignoring of room 1 will make it impossible to visit room 0, and in the long run the expected cost of not visiting room 0 will be very high. Obviously, such problems can be solved theoretically by looking more than one step ahead. Since looking ahead many steps is computationally expensive, it would be useful to develop some methods for assessing the number of steps required to obtain satisfactory behaviour. Even for constant costs, the one step minimum expected cost policy is not guaranteed to globally minimise the expected cost. Proposition 5.4 In the environment of example 5.1 the one step minimum expected cost policy does not minimise the global expected cost. Actually, in example 5.1, the global expected cost of the one step minimum expected cost policy is higher than that of the back and forth behaviour resulting from the methodical exploration and the maximum likelihood policy. The two step minimum expected cost policy already results in the same back and forth behaviour. The cause of the problem with the one step minimum expected cost policy in our model is that visiting a location at time t decreases the probability of a relevant event occurring at that location after t. It is worth investigating whether one can take into account such indirect benefits of visiting a location by adjusting the costs. Proposition 5.5 Let t 0 be the time of the last visit to X i before t, and T be the time of the next visit after t. Then the indirect benefits of a visit to X i at t are equal to the following. then the above expression provides a lower bound of the indirect benefits of a visit to X i at t instead of later. Incorporating this amount of the indirect benefit into the one step minimum expected cost policy is similar to employing a two step minimum expected cost policy, and it result in the back and forth behaviour in the environment of example 5.1. Proposition 5.6 Let t 0 be the time of the last visit to X i before t. Then the indirect benefits of a visit to X i at t have the following upper bound. lim This upper bound can be used to argue that it is not optimal to visit room 1 in the environment of example 5.2. 6 Conclusions and Further Work We do not claim that our discussion of decision-theoretic surveillance establishes the minimum expected cost policy as the best policy for surveillance. To substantiate such a claim, one needs to evaluate the performance of different possible strategies when applied to much more realistic surveillance problems than those discussed thus far. However, as we mentioned before, we believe that the decision-theoretic view on surveillance is sufficiently general to, at least theoretically, incorporate many of the necessary refinements to the model presented in this paper. To properly evaluate surveillance strategies, one should clarify which kind of surveillance task one is considering, since the term 'surveillance' is used to refer to quite different problems. Actually, it may be impossible to find an unqualified best policy for all surveillance problems. For example, the maximum likelihood policy seems to be intimately connected with a symmetrical view on surveillance, whereas the minimum expected cost policy is devised with the asymmetrical interpretation in mind. In gen- eral, we expect the minimum expected cost policy to behave well in situations where the probabilities and costs matter and where early detection is important. We introduced a simple, formal model of an environment in which surveillance tasks can be performed. The main purpose of this environment model is to clarify the distinctions between different surveillance strategies. In the future, we plan to refine the environment model presented in this paper to incorporate the uncertainties a real surveillance robot has to face: sensor uncertainty, navigation uncertainty, et cetera. A simulation program is being developed to experimentally evaluate different surveillance strategies in varying environments. Also, some further theoretical work is needed to clarify the situations under which local strategies can be used to obtain globally optimal (or satisfactory) performance. Acknowledgments The investigations were carried out as part of the PIONIER-project Reasoning with Uncertainty, subsidized by the Netherlands Organization of Scientific Research (NWO), under grant pgs-22-262. --R Visual surveillance in a dynamic and uncertain world. A new approach in temporal representation of belief for autonomous observation and surveillance systems. Repr'esentation Dynamique de l'Incertain et Strat'egie de Perception pour un Syst'eme Autonome en Environnement ' Evolutif. Dissertation, L' ' Ecole Nationale Sup'erieure de l'A'eronautique et de l'Espace Issues in surveillance. On the lookout: The air mobile ground security and surveillance system (AMGSSS) has arrived. SPAWAR Mobile Detection Assessment and Response System. AUV survey design qpplied to oceanic deep convection. --TR
autonomous robots;surveillance;decision-theoretic planning
590960
A Reusable Multi-Agent Architecture for Active Intelligent Websites.
In this paper a reusable multi-agent architecture for intelligent Websites is presented and illustrated for an electronic department store. The architecture has been designed and implemented using the compositional design method for multi-agent systems DESIRE. The agents within this architecture are based on a generic information broker agent model. It is shown how the architecture can be exploited to design an intelligent Website for insurance, developed in co-operation with the software company Ordina Utopics and an insurance company.
Introduction Most current business Websites are mainly based on navigation across hyperlinks. A closer analysis of such conventional Websites reveals some of their shortcomings. For example, customer relationships experts may be disappointed about the unpersonal treatment of customers at the Website; customers are wandering around anonymously in an unpersonal virtual environment and do not feel supported by anyone. It is as if customers are visiting the physical environment of a shop (that has been virtualised), without any serving personnel. Marketing experts may also not be satisfied by the Website; they may be disappointed in the lack of facilities to support one-to-one marketing. In a conventional Website only a limited number of possibilities are provided to announce new products and special offers in such a manner that all (and only) relevant customers learn about them. Moreover, often Websites do not acquire information on the amounts of articles sold (sales statistics). It is possible to build in monitoring facilities with respect to the amount of products sold over time, but also the number of times a request is put forward on a product (demand statistics). If for some articles a decreasing trend is observed, then the Website could even advice employees to take these trends into account in the marketing strategy. If on these aspects a more active role would be taken by the Website, the marketing qualities could be improved. The analysis from the two perspectives (marketing and customer relationships) suggests that Websites should become more active and personalised, just as in the traditional case where contacts were based on humans. Intelligent agents provide the possibility to reflect at least a number of aspects of the traditional situation in a simulated form, and, in addition, enables to use new opportunities for, e.g., one-to-one marketing, integrated in the Website. The generic agent-based architecture presented in this paper offers these possibilities. This generic architecture for active intelligent Websites was first introduced for the application domain of a department store, which has been analysed in co-operation with the software company CMG (cf. [22]). It reuses the generic architecture of information broker agents developed earlier (cf. [21]), which in turn was designed as a specialisation of the generic agent model GAM introduced in [8]. As a second step the reusability of the generic multi-agent architecture for active intelligent Websites has been tested by applying it in a project on an intelligent Website for insurance in co-operation with the software company Ordina Utopics and an insurance company (cf. [20]). The testbed chosen for this application involves information and documents that need to be exchanged between insurance agents and the insurance company main office. The goal of the intelligent Website is to provide insurance agents with an accurate account of all relevant available documents and information. The supporting software agents are able to provide a match (either strict or soft) between demand and available information. They support pro-active information provision, based on profiles of the insurance agents that are dynamically constructed. A prototype system for this application is described in more detail in the second part of the paper. In this paper in Section 2 the global design of a multi-agent architecture for an intelligent Website is presented; the different types of agents participating in the Website are distinguished. In Section 3 their characteristics and required properties are discussed. In Section 4 the compositional generic information broker agent architecture is described and applied to obtain the internal structure of the agents involved in the multi-agent architecture. In Section 5 the insurance application domain is introduced. In Section 6 the application of the architecture to insurance is discussed in more detail and illustrated by some example behaviour patterns. Section 7 concludes the paper by a discussion. Multi-Agent Architecture for Intelligent Websites In this section a global multi-agent architecture, that can be used as a basis for an intelligent Website, is introduced. Although the architecture is generic, for reasons of presentation some of its aspects will be illustrated in the context of the insurance application. The domain has been identified as a multi-agent domain. Therefore, it makes sense to start with the agents as the highest process abstraction level within the system. Four classes of agents are distinguished at the level of the multi-agent system (see Fig. 1): . customers (human agents), . Personal Assistant agents (software agents, denoted by PA), . Website Agents (software agents, denoted by WA), . employees (human agents). In Fig. 1, the shaded area at the right hand side shows the agents related to the Website; the shaded area at the left hand side shows the two agents at one of the customer sites. In this figure, for shortness only two Website Agents, one employee, one Personal Assistant agent and one customer (user of the Personal Assistant) are depicted. Moreover, for the sake of simplicity, the Website itself is left out of the picture. The Website has the role of the external world for the agents; note that is not considered an agent itself. All agents can have interaction with this external world to perform observations. The Website agents and employees can also perform actions in this world, e.g., to change the information on one of the Webpages. user Website Agent2 Website Agent1 Personal Assistant employee Fig. 1. The overall multi-agent architecture Note that the Personal Assistant is involved as a mediating agent in all communication between its own user and all Website Agents. From the user it can receive information about his or her interests and profile, and it can provide him or her with information assumed interesting. Moreover, it can receive information from any of the Website Agents, and it can ask them for specific information. The Website Agents communicate not only with all Personal Assistants, but also with each other and with employees. The customer only communicates with his or her own Personal Assistant. This agent serves as an interface agent for the customer. If a customer visits the Website for the first time this Personal Assistant agent is instantiated and offered to the customer (during all visits). The application domain to illustrate the architecture addresses the design of an active, intelligent Website for a chain of department stores. The system should support customers that order articles via the Internet. Each of the department stores sells articles according to departments such as car accessories, audio and video, computer hardware and software, food, clothing, books and magazines, music, household goods, and so on. Each of these departments has autonomy to a large extent; the departments consider themselves small shops (as part of a larger market). This suggests a multi-agent perspective based on the separate departments and the customers. For each department in the department store a Website Agent can be designed, and for each customer a Personal Assistant agent serves as an interface agent. 3 Requirements for the Software Agents The departments should relate to customers like small shops with personal relationships to customers. The idea is that customers know at least somebody (a Website Agent) related to a department, as a representative of the department and, moreover, this agent knows specific information on the customer. Website Agent - Interaction with the world observation passive observation active its own part of the Website product information - presence of customers/Personal Assistants visiting the Website - economic information - products and prices of competitors - focusing on what a specific customer or Personal Assistant does - search for new products on the market performing actions - making modifications in the Website (e.g., change prices) showing Web-pages to a customer and Personal Assistant creating (personal or general) special offers - modification of assortment Table 1. World interaction characteristics for a Website Agent 3.1 Characteristics and Requirements for the Website Agents Viewed from outside the basic agent behaviours autonomy, responsiveness, pro-activeness and social behaviour such as discussed, for example in [38] provide a means to characterise the agents (see Table 3). Moreover, the following external agent concepts to define interaction characteristics are used: . interaction with the world (observation, action performance) . communication with other agents In Tables 1 and 2 the interaction characteristics for the Website Agents have been specified and illustrated for the case of the department store. Website Agent - Communication incoming from Personal Assistant: - request for information - request to buy an article paying information customer profile information customer privacy constraints from employee: - requests for information on figures of sold articles new product information - proposals for special offers and price changes - confirmation of proposed marketing actions - confirmation of proposed assortment modifications - proposals for marketing actions - proposals for assortment modifications from other Website Agent: - info on assortment scopes customer info outgoing to Personal Assistant: asking whether Website Agent can help providing information on products providing information on special offers special (personal or general) offers to employee: - figures of articles sold (sales statistics) - analyses of sales statistics - numbers of requests for articles (demand statistics) - proposals for special offers - proposals for assortment modifications to other Website Agent: - info on assortment scopes customer info Table 2. Communication characteristics for a Website Agent The following requirements have been imposed on the Website Agents: . personal approach; informed behaviour with respect to customer In the Website each department shall be represented by an agent with a name and face. Furthermore, some of these agents (those who have been in contact with the customer) know the customer and his or her characteristics, and remember what this customer bought previous times. . being helpful Customers entering some area of the Website shall be contacted by the agent of the department related to this area, and asked whether he or she wants some help. If the customer explicitly indicates that he or she only wants to look around without getting help, the customer shall be left alone. Otherwise, the agent takes responsibility to serve this customer until the customer has no wishes anymore that relate to the agent's department. The conventional Website can be used by the Website Agents to point at some of the articles that are relevant (according to their dialogue) to the customer. . refer customers to appropriate colleague Website Agents A customer which is served at a department and was finished at that department can only be left alone if he or she has explicitly indicated to have no further wishes within the context of the entire department store. Otherwise the agent shall find out in which other department the customer may have an interest and the customer shall be referred to the agent representing this other department. . be able to provide product and special offer information For example, if a client communicates a need, then a product is offered fulfilling this need (strictly or approximately), and, if available a special offer. . dedicated announcement As soon as available new products and special offers shall be announced to all relevant (on the basis of their profiles) customers, (they shall be contacted by the store in case they do not frequently contact the store). Website Agent - Basic types of behaviour Autonomy - functions autonomously, especially when no employees are available (e.g., at night) Responsiveness - responds to requests from Personal Assistants - responds to input from employees - triggers on decreasing trends in selling and demands Pro-activeness - takes initiative to contact Personal Assistants takes initiative to propose special offers to customers - creates and initiates proposals for marketing actions and assortment modifications Social behaviour - co-operation with employees, Personal Assistants, and other Website Agents Table 3. Basic types of behaviour of a Website Agent . analyses for marketing The Website Agents shall monitor the amounts of articles sold (sales statistics), communicate them to employees (e.g., every week) and warn if substantially decreasing trends are observed. For example, if the figures of an article sold decrease during a period of 3 weeks, then marketing actions or assortment modifications shall be proposed. . actions for marketing Each Website Agent shall maintain the history of the transactions of each of the customers within its department, and shall perform one to one marketing to customers, if requested. The employees shall be able to communicate to the relevant Website Agents that they have to perform a marketing campaign. The agent shall propose marketing actions to employees. . privacy No profile is maintained without explicit agreement with the customer. The customer has access to the maintained profile. Personal Assistant - Interaction characteristics A. Interaction with the world observation passive observation active observe changes and special offers at the Website - observe the Website for articles within the customer needs performing actions B. Communication with other agents incoming from Website Agent: product info special (personal and general) offers from customer: customer needs and preferences - agreement to buy privacy constraints outgoing to Website Agent: customer needs payment information profile information to customer: product information special offers Table 4. Interaction characteristics for the Personal Assistant 3.2 Characteristics and Requirements for the Personal Assistants For the Personal Assistants the interaction characteristics are given in Table 4, and their basic types of behaviour in Table 5. The following requirements can be imposed on the Personal Assistants: . support communication on behalf of the customer Each customer shall be supported by his or her own Personal Assistant agent, who serves as an interface for the communication with the Website Agents. . only provide information within scope of interest of customer A customer shall not be bothered by information that is not within his or her scope of interest. A special offer that has been communicated by a Website Agent leads to a proposal to the customer, if it fits in the profile, and at the moment when the customer wants such information . sensitive profiling Customers are relevant for a special offer if they have bought a related article in the past, or if the offer fits in their profile as known to the Personal Assistant. . providing customer information for Website Agents every week the relevant parts of the profile of the customer is communicated to the Website Agent, if the customer agrees. . privacy The Personal Assistant shall protect and respect the desired privacy of the customer. Only parts of the profile information agreed upon are communicated. Personal Assistant - Basic types of behaviour Autonomy autonomous in dealing with Website Agents on behalf of customer Responsiveness responsive on needs communicated by customer Pro-activeness initiative to find and present special offers to customer Social behaviour with customer and Website Agents Table 5. Basic types of behaviour for the Personal Assistant 4 The Internal Design of the Information Broker Agents The agents in the multi-agent architecture for intelligent Websites presented in the previous sections have been designed on the basis of a generic model for a broker agent. The process of brokering as it often occurs as a mediating process in electronic commerce involves a number of activities. For example, responding to customer requests for products with certain properties, maintaining information on customers, building customer profiles on the basis of such customer information, maintaining information on products, maintaining provider profiles, matching customer requests and product information (in a strict or soft manner), searching for information on the WWW, and responding to new offers of products by informing customers for whom these offers fit their profile. In this section a generic broker agent architecture is presented that supports such activities. This generic information broker model has been used as a basis for both the Website Agents and the Personal Assistant agents. As these architectures have been designed using the compositional design method for multi-agent systems DESIRE, first a brief overview of DESIRE is presented (Section 4.1), next the generic broker agent model is briefly discussed (Section 4.2), and finally the two types of information broker agents that are used in the generic multi-agent architecture for intelligent Websites are discussed: Website Agent (Section 4.3) and Personal Assistant (Section 4.4). 4.1 Compositional Design of Multi-Agent Systems The emphasis in DESIRE is on the conceptual and detailed design. The design of a multi-agent system in DESIRE is supported by graphical design tools within the DESIRE software environment. The software environment includes implementation generators with which (formal) design specifications can be translated into executable code of a prototype system. In DESIRE, a design consists of knowledge of the following three types: process composition, knowledge composition, and the relation between process composition and knowledge composition. These three types of knowledge are discussed in more detail below. 4.1.1 Process Composition Process composition identifies the relevant processes at different levels of (process) abstraction, and describes how a process can be defined in terms of (is composed of) lower level processes. of Processes at Different Levels of Abstraction Processes can be described at different levels of abstraction; for example, the process of the multi-agent system as a whole, processes defined by individual agents and the external world, and processes defined by task-related components of individual agents. The identified processes are modelled as components. For each process the input and output information types are modelled. The identified levels of process abstraction are modelled as abstraction/specialisation relations between components: components may be composed of other components or they may be primitive. Primitive components may be either reasoning components (i.e., based on a knowledge base), or, components capable of performing tasks such as calculation, information retrieval, optimisation. These levels of process abstraction provide process hiding at each level. Composition of Processes The way in which processes at one level of abstraction are composed of processes at the adjacent lower abstraction level is called composition. This composition of processes is described by a specification of the possibilities for information exchange between processes (static view on the composition), and a specification of task control knowledge used to control processes and information exchange (dynamic view on the composition). 4.1.2. Knowledge Composition Knowledge composition identifies the knowledge structures at different levels of (knowledge) abstraction, and describes how a knowledge structure can be defined in terms of lower level knowledge structures. The knowledge abstraction levels may correspond to the process abstraction levels, but this is often not the case. of knowledge structures at different abstraction levels The two main structures used as building blocks to model knowledge are: information types and knowledge bases. Knowledge structures can be identified and described at different levels of abstraction. At higher levels details can be hidden. An information type defines an ontology (lexicon, vocabulary) to describe objects or terms, their sorts, and the relations or functions that can be defined on these objects. Information types can logically be represented in order-sorted predicate logic. A knowledge base defines a part of the knowledge that is used in one or more of the processes. Knowledge is represented by formulae in order-sorted predicate logic, which can be normalised by a standard transformation into rules. Composition of Knowledge Structures Information types can be composed of more specific information types, following the principle of compositionality discussed above. Similarly, knowledge bases can be composed of more specific knowledge bases. The compositional structure is based on the different levels of knowledge abstraction distinguished, and results in information and knowledge hiding. 4.1.3 Relation between Process Composition and Knowledge Composition Each process in a process composition uses knowledge structures. Which knowledge structures are used for which processes is defined by the relation between process composition and knowledge composition. 4.2 A Generic Information Broker Agent Architecture The generic information broker agent architecture was designed as a refinement of the generic agent model GAM (cf. [8]), supporting the weak agency notion (cf. [38]). First we will briefly describe the generic model GAM and next we discuss how this model was refined to the generic information broker model. 4.2.1 The generic agent model GAM At the highest process abstraction level within the compositional generic agent model GAM introduced in [8], a number of processes are distinguished that support interaction with the other agents. First, a process that manages communication with other agents, modelled by the component agent interaction management in Fig. 2. This component analyses incoming information and determines which other processes within the agent need the communicated information. Moreover, outgoing communication is prepared. Communication is modelled in a first-order logic approach, comparable, for example, to KIF. Communication from agent A to B takes place in the following manner: . the agent A generates at its output interface a statement of the form: . the information is transferred to B; thereby it is translated into If needed, it is not difficult to replace this format by more extensive formats used in KQML or FIPA-ACL. Next, the agent needs to maintain information on the other agents with which it co- operates: maintenance of agent information. The component maintenance of world information is included to store the world information (e.g., information on attributes of products). The process own process control defines different characteristics of the agent and determines foci for behaviour. The component world interaction management is included to model interaction with the world (with the World Wide Web world, in the example application): initiating observations and receiving observation results. The agent processes discussed above are generic agent processes. Many agents perform these processes. In addition, often agent-specific processes are needed: to perform tasks specific to one agent, for example directly related to a specific domain of application. This is the purpose of the component Agent Specific Task. Fig. 2 depicts how the generic agent is composed of its components. communicated observation results to wim observed agent communicated agent Agent task control Own Process Control Maintenance of Agent Information Agent Task Maintenance of World Information Agent Interaction Management World Interaction Management own process info to wim own process info to aim own process info to own process info to mwi info to be communicated communicated info to ast communicated world info observations and actions observed info to ast observed world info action and observation info from ast communication info from ast agent info to opc world info to opc agent info to wim agent info to aim world info to aim world info to wim Fig. 2. Composition within the generic information broker agent model 4.2.2 Refinement of GAM to the generic information broker agent model The refinement of a generic model may involve both specialisation of the process composition and instantiation of the knowledge composition. The specific refinement discussed here only involves instantiation of the knowledge composition. Part of the exchange of information within the generic broker agent model can be described as follows. The broker agent needs input about scopes of interests put forward by agents and information about attributes of available products that are communicated by information providing agents. It produces output for other agents about proposed products and the attributes of these products. Moreover, it produces output for information providers about interests. In the information types that express communication information, the subject information of the communication and the agent from or to whom the communication is directed are expressed. This means that communication information consists of statements about the subject statements that are communicated. Within the broker agent, the component own process control uses as input belief info, i.e., information on the world and other agents, and generates focus information: to focus on a scope of interest to be given a preferential treatment, i.e., pro-active behaviour will be shown with respect to this focus. The component agent interaction management has the same input information as the agent (incoming communication), extended with belief info and focus info. The output generated includes part of the output for the agent as a whole (outgoing communication), extended with maintenance info (information on the world and other agents that is to be stored within the agent), which is used to prepare the storage of communicated world and agent information. Information on attributes of products is stored in the component maintenance of world information. In the same manner, the beliefs of the agent with respect to other agents' profiles (provider attribute info and interests) are stored in maintenance of agent information. The component agent specific task uses information on product attributes and agent interests as input to generate proposals as output. For reasons of space limitation the generic and domain-specific information types within the agent model are not presented; for more details; see [21]. The information broker agent may have to determine proposals for other agents. In this process, information on available products (communicated by information providing agents and kept in the component maintenance of world information), and about the scopes of interests of agents (kept in the component maintenance of agent information), is combined to determine which agents might be interested in which products. 4.3 The Website Agent: Internal Design The broker agent architecture provides an appropriate means to establish the internal design of the two types of agents involved. For the Website Agent, the internal storage and updating of information on the world and on other agents (the beliefs of the agent) is performed by the two components maintenance of world information and maintenance of agent information. In Table 6 it is specified which types of information are used in these components. Profile information on customers is obtained from Personal Assistants, and maintained with the customer's permission. Also identified behaviour instances of the Personal Assistants can give input to the profile. Profile information can be abstracted from specific demands; how this is performed may depend on the application that is made. Website Agent - Maintenance of Information world information - info on products within the Website Agent's assortment - info on special offers agent information - info on customer profiles - info on customer privacy constraints - info on customer preferences in communication - info on which products belong to which other Website Agent's assortments - info on providers of products Table 6. Maintenance information for the Website Agent The component agent interaction management identifies the information in incoming communication and generates outgoing communication on the basis of internal information. For example, if a Personal Assistant agent communicates its interests, then this information is identified as new agent interest information that is believed and has to be stored, so that it can be recalled later. In the component agent specific task specific knowledge is used such as, for example: . if the selling numbers for an article decrease for 3 weeks, then make a special offer with lower price, taking into account the right season . if a customer asks for a particular cheap product, and there is a special offer, then this is proposed . if an article is not sold enough over a longer period, then take it out of the assortment Within this component non-strict (or soft) matching techniques can be employed to relate demands and offers. 4.4 The Personal Assistant: Internal Design In this section some of the components of the Personal Assistant are briefly discussed. For the Personal Assistant, as for the Website Agent, the internal storage and updating of information on the world and on other agents is performed by the two components maintenance of world information and maintenance of agent information. In Table 7 it is specified which types of information are used in these components. Personal Assistant - Maintenance of Information world information - product information special offers agent information - customer needs and profile customer privacy constraints offers personal to the customer - Website Agents assortment scopes Table 7. Maintenance information for a Personal Assistant As in the previous section, the component agent interaction management identifies the information in incoming communication and generates outgoing communication on the basis of internal information. For example, if a Website Agent communicates a special offer, then this information is identified as new agent information that is believed and has to be stored, so that it can be recalled later. Moreover, in the same communication process, information about the product to which the special offer refers can be included; this is identified and stored as world information. 4.5 Profile modelling approaches that can be used within the agents Within the generic architecture for Website Agents and Personal Assistants no commitment has been made to specific approaches to user profiling. In this section a number of these approaches are briefly discussed (for a more detailed treatment, see [11]). The profile of a user can be used to determine how interesting an information item is to that user. It can be used to select and prioritise information items in a personalised manner. The structure and properties of profiling approaches may vary with the application area in which they are used. For example, in multi-attribute decision systems (see [3],[23],[37]) the user profile or preference for an item is defined in terms of values of various attributes of the item and the preferences of the user towards those attributes (i.e., the importance of those attributes). On the other hand, in the area of recommendation systems the profile may as well be defined in terms of statistical correlation between users and their rated items. The preferences of a user towards a set of items can be defined in terms of the content of the items (content information) or the preference of the items by a society of users (collaborative or social information). In the content-based approach a user is defined to have preference for an item if the item is similar in attribute values to other items that are preferred by the user. Also ratings for the (relative) relevance of attributes for a user are often included. In the collaborative-based approach a user is defined to have preference for an item if the user is similar (in preferences of other items) to other users who have preference for the item. Both the content information as well as the collaborative information can be used to construct user profiles. The construction of a profile can be a time consuming matter. For example, in the content-based approach the user may have to express his or her preferences towards various (combinations of) attributes and attribute values in extensive forms. Some systems (e.g., see [12]) derive the preferences of a user by suggesting an item to the user and ask her to correct this suggestion. The user corrects the system's suggestion by indicating why the suggested item does not match his or her needs. Based on these corrections, profiles of users are constructed or updated. In other, collaborative-based applications such as recommendation systems, a user may be asked to rate several, sometimes hundreds, of (other) items before an item can be recommended. A number of systems employ methods to induce the user profile by observing the behaviour of that user over time (e.g., see [16],[26],[29],[31]). These methods are usually not intended to fully model user profiles, but to model the more frequent and predictable user preferences. Applications that require huge efforts from their users may become ineffective (e.g., see [27],[28]). To model user profiles in an application, a balance is to be found between the amount of interaction with the user and the effectiveness of the constructed user profile. Modelling user profiles on the basis of content or collaborative information can be considered as a learning problem where the aim is to learn the so-called preference function for a certain user. The preference function for a user maps items from a certain domain to some values that express the importance of those items for that user. Various types of preference functions may exist. The type of a preference function characterises the structure of profile (e.g., see [23],[37]). Another profile learning approach, based on Inductive Logic Programming, can be found in [5], [6], [11]. Several collaborative-based recommendation systems have been introduced in which the preferences of users are modelled automatically. Examples of online recommendation systems that employ a collaborative approach are MovieFinder [39] and FireFly [13]. The preferences of a user are modelled automatically by observing the behaviour of that user and applying statistical methods to the observed behaviour (e.g., see [4],[16],[17],[33]). In contrast to the collaborative-based approach, the content-based approach can be applied only when items are described in terms of properties and attribute values. The content-based profiling approaches have been used in online recommendation systems such as BargainFinder [1] and Jango [18]. Unlike collaborative-based preference models, the content-based preference models are also used in applications such as integrative negotiation where the utility function is defined in terms of user preferences towards various attribute values (e.g., see [3],[15],[23],[29],[37]). The collaborative-based and content-based approaches do not exclude each other; in fact they can be combined into an integrated approach to model user profiles (see [2]). The effectivity of collaborative-based and content-based approaches to profiling may depend on the application. For example, collaborative-based profiling approaches may be more effective in applications where it is unrealistic to collect a large amount of information about the preferences of an individual user, or where the number of users is too large. Using collaborative-based profiling models is also effective for applications where the content of the items neither is available nor can be analysed automatically by a machine (e.g. items like a picture, video, sound). However, the collaborative-based profiling approaches are less effective for applications like integrative negotiation (e.g., see [3],[14]) in retail Electronic Commerce where negotiation is considered to be a decision making process over items that are described as multiple interdependent attributes. 5 Reuse of the Generic Architecture in the Insurance Domain The reusability of the designed generic multi-agent architecture was tested in a new domain: insurance. In this section this domain is briefly introduced. One of the largest insurance companies in the Netherlands is organised on the basis of (human) mediating insurance agents. To support these agents a Website was created with information about products offered by the company, forms to support administrative actions, and other related information. The Website is structured around four main sections: Store, Desk, Newsstand and Office. The store provides information about the insurance products offered. The various insurance policies can be found here, as well as request forms for more information, brochures, and personalised proposals . From the store a couple of useful programs can be downloaded as well: spreadsheets, an anti-virus toolkit, and an insurance dictionary. When the insurance agent is faced with a problem, he or she can turn to the desk. Apart from a Frequently Asked Questions page also a form is available for specific questions. The desk further contains the editorials that address certain problems in depth. Finally, an address book is available, in which the various departments and teams operating within the company can be found. At the news-stand the visitors of the site can find the most recent information. Newsletters can be found, and a calendar can be checked for upcoming events. Furthermore, various links to other interesting sites and assorted articles are offered here. Whenever new interesting sites or articles are added, the visitor can be notified of this by email. At the office, the sale of insurance products is supported. Here resources to improve the insurance agent's job can be found: telemarketing scripts, newsletter articles, advertisements that only need further filling out and sales letters. Furthermore, the agent can find its personal production figures for the company's products. The Website consists of a collection of variable information sources: images, programs, documents, addresses, phonebooks and personal data. New information is added daily. Keeping up to date with the most recent relevant information, is time-consuming. The multi-agent system has been developed to support the human agent in this task. The aim of the multi-agent system integrated in the Website is to improve the use of the resources offered by the Website. From the visitors point of view, more interesting information can be obtained. The agent, with its knowledge of the user improves the customer experience. Application forms can be offered, already (partially) filled out by the software agent. The employees maintaining the Website can use information collected by the multi-agent system to improve marketing. The appropriate visitor can be contacted about new (possibly personalised) products or offers that are relevant to him or her. 6 Instantiation of the Generic Architecture The generic multi-agent architecture has been instantiated for the new domain of insurance described in Section 5. Application-specific information types and knowledge bases were specified and included in the model. The system is explained for two cases: behaviour initiated by an information request of a user (user initiated), and behaviour initiated by update or addition of information to the Website (Website initiated). In both cases, after initiation a reactivity chain is triggered. In the first case the main reactivity chain follows the path user-PA-WA-PA-user The first half of this path deals with the queries, and the second half (back) with answers on these queries. In the second case the main reactivity chain follows the path The first half deals with voluntarily offered information (one-to-one marketing), and the second half (back) with feedback on usefulness of the offered information (in order to update profile information). In the explanation of these behavioural traces, it is shown which knowledge bases were used to instantiate the generic architecture. 6.1 Information used in the system This system is only a prototype; as such it does not work with the actual information on the Website. Instead a sample of the information objects on the Website was selected and a description of each of these was made. In cooperation with employees from the insurance company the following attributes were selected to describe the information: . Title: The title of the information object. . Author: The department or person that created the information object. . Subject: Subject of the information object. . First Relation: The first related subject. . Second Relation: The second related subject. . Date: Date of creation/availability. . Language: The language used in the information. . Persistency: An indication of how soon the information will be outdated. . Kind: The form of the information object (mailform, text, audio, etc. . Type: The type of information in the information object (e.g., FAQ, newsletter, personal information). . URL: The hyperlink to the actual information object Fig. 3. User interface for asking questions and stating user interest 6.2 Behaviour initiated by a user When a user asks a question, the Personal Assistant agent performs a number of actions. The question is analysed to find similarities to previous questions and if these exist, new interests are created within the user profile. Furthermore, the agent attempts to respond to the information request using information available within the Personal Assistant itself and by contacting the appropriate Website Agents. First it is described how an answer to a question is found. Next, the process of updating the user profile is discussed. Handling a question. The behaviour of the system is first described from the user's point of view. Subsequently, the processes that are invisible for the user a described in more details. The user interaction A trace is described in which a user needs information about car insurance. As a first step the user communicates this question to the Personal Assistant using the interface (Fig. 3): the user selects the subject 'car insurance' in the scrollable list under the heading `Subjects'. The Personal Assistant will start to acquire useful information on behalf of its user. Fig. 4. Display for the answers to questions and offers made (by a Personal Assistant) The Personal Assistant inspects all information it has in store and it contacts appropriate Website Agents for more information. All the gathered relevant information is communicated to the user, using the display in Fig. 4; each title is a link to a description of the information. The user can indicate whether or not he or she evaluates the information as interesting. if query(Q:QUERY_ID, scope(subject, S:SUBJECT)) and object_scope(O:OBJECT_ID, scope(related_subject, S:SUBJECT)) then possible_answer_to_query(O:OBJECT_ID, Q:QUERY_ID); if query(Q:QUERY_ID, scope(A:ATTRIBUTE, V1:VALUE)) and object_scope(O:OBJECT_ID,scope(A:ATTRIBUTE, V2:VALUE)) and not subject and not then rejected_answer_for_query(O:OBJECT_ID, Q:QUERY_ID); if possible_answer_to_query(O:OBJECT_ID, Q:QUERY_ID) and not rejected_answer_for_query(O:OBJECT_ID, Q:QUERY_ID) then selected_answer_to_query(O:OBJECT_ID, Q:QUERY_ID); Table Knowledge involved in user-initiated behaviour The processes within the multi-agent system When Personal Assistant agent receives a question from the user, it identifies the communication as a question in the component agent interaction management. The question is further processed in the task specific component determine proposals of the Personal Assistant. That component matches the request to the information objects available in the memory of the agent (component maintenance of world information). Two types of matching are covered: strict matching and soft matching. For strict matching, attributes need to have exactly the same value, or an overlapping value range. For soft matching, it can be specified when values of attributes are considered close (but not necessarily equal) to each other. This closeness relation may be based on various techniques. In the current prototype the closeness relation for the subject attribute is taken as a point of departure, abstracting from the manner in which it is determined. One of the matching rules is rule r1 in Table 8. The subject of the query is matched with the related subject of the object under consideration. If the rule succeeds, the object is selected as a possible answer. A criterion for this possible answer to become a definite answer is that the object does not differ on other attributes (see rule r2). Rule r3 is used to derive the final answer to the question. Simultaneously, in the same component determine proposals, the relevant Website Agents the are selected. This is done in three steps. First, the Personal Assistant agent looks for a Website Agent that is known to provide information about the subject occurring in the query; see rule r4. Rule 4 makes use of the agent model for the Website Agent that is stored by thePersonal Assistantwithin component maintenance of agent information. Information about the subjects that a Website Agent can provide is expressed by the statement webagent_subject(W:WA, S:SUBJECT) . Rule 4 will not succeed, however, when the question does not contain a subject term or when the Personal Agent does not know a relevant Website Agent. In this case the Personal Assistant agent uses a second method to determine an appropriate Website Agent, by considering another part of the agent models it maintains of Website Agents; see rule r5. Finally as a fail-safe, each Personal Assistant has a default Website Agent it can contact. The name of this default Website Agent is stored in the component own process control and is also available in the component determine proposals. The final selection of the Website Agent is performed by the knowledge specified in rules r6 to r9. if query(Q:QUERY_ID, scope(subject, S:SUBJECT)) and webagent_subject(W:WA, S:SUBJECT) then main_wa_for_answer(W:WA, Q:QUERY_ID) and found_wa_for_query(Q:QUERY_ID); if query(Q:QUERY_ID, S:SCOPE) and can_provide_scope(W:WA, S:SCOPE) then secondary_wa_for_answer(W:WA, Q:QUERY_ID) and found_wa_for_query(Q:QUERY_ID); if main_wa_for_answer(W:WA, Q:QUERY_ID) then selected_wa_for_answer(W:WA, Q:QUERY_ID); if secondary_wa_for_answer(W:WA, Q:QUERY_ID) then selected_wa_for_answer(W:WA, Q:QUERY_ID); if not found_wa_for_query(Q:QUERY_ID) and default_wa(W:WA) then selected_W:WA_for_answer(W:WA, Q:QUERY_ID); Table 9 Knowledge involved in selection of Website Agents Next the selected_wa_for_answer and selected_object_for_query information is transferred to the component agent interaction management where communication to the selected Website Agent(s) is actually initiated (see Table 9. Website Agents handle questions in the same way as the Personal Assistant. The component determine proposals of a Website Agent attempts to find a match with the known information objects. The matches are communicated back to the Personal Assistant. The component agent interaction management of the Personal Assistant passes the received answers on to its user. if communicated_by(query_answer(Q:QUERY_ID, object_scope(O:OBJECT_ID, S:SCOPE)), pos, W:WA) then to_be_communicated_to(query_answer(Q:QUERY_ID, object_scope(O:OBJECT_ID, S:SCOPE)), pos, user); Table 9 Knowledge involved in communication to user The information contained in received answers is also stored by the Personal Assistant: in the future it can supply this information by itself. Update of user profile. The focus of the current prototype lies on the agent interaction and document selection. Profile management had a lower priority. Therefore the mechanisms for profile management used are simple. As stated earlier, the Personal Assistant compares questions to each other. When similarities are found in three questions, these similarities are added to the user profile. This is performed by the (composed) component interest creator. A new question is first compared to all previous questions. A simple method has been chosen to create these candidates: whenever three different questions match on one or more attribute values, these attribute-value pairs are selected as a candidate interest specification; see rule r11 in Table 10. The three query id's are combined to create a temporary candidate if asked(query(Q1:QUERY_ID, scope(A:ATTRIBUTE, V:VALUE))) and asked(query(Q2:QUERY_ID, scope(A:ATTRIBUTE, V:VALUE))) and asked(query(Q3:QUERY_ID, scope(A:ATTRIBUTE, V:VALUE))) and not and not and not then candidate_for_interest(candidate_id(Q1:QUERY_ID, Q2:QUERY_ID, Q3:QUERY_ID), scope(A:ATTRIBUTE, V:VALUE)); if candidate_for_interest(C:CANDIDATE_ID, scope(A:ATTRIBUTE, V1:VALUE)) and belief(interest(I:INTEREST_ID, scope(A:ATTRIBUTE, V2:VALUE)) and not then different(C:CANDIDATE_ID, I:INTEREST_ID); if new_interest_id(I:INTEREST_ID) and approved_candidate(C:CANDIDATE_ID, S:SCOPE) then to_be_created(interest(I:INTEREST_ID, S:SCOPE)); Table Knowledge involved in profile update: user-initiated case. 6.3 Behaviour initiated by the Website The second type of behaviour discussed here is initiated by the Website. First the behaviour to directly serve the user is discussed, and subsequently the behaviour to update the user profile is described in more detail. Offering the user new information. First the behaviour shown to the user is described. Next a more detailed description is given of the processes within the multi-agent system itself. The user interaction The Personal Assistant takes the initiative to notify its user when relevant information has been found, using the display depicted in Fig. 4. Again, the user can click on a title to get more information about the proposal (Fig. 5). Furthermore, the user can choose to accept the proposed information or to reject it. The processes within the multi-agent system When new information becomes available at a Website, the Website Agent identifies possible interested parties. The Website Agent has built a profile of the Personal Assistants it has been in contact with. In the component determine proposals the Website Agent uses this information to match the new information to the Personal Assistants interests; see rule r13 in Table 11. if new_object_scope(O:OBJECT_ID, S:SCOPE) and interest(P:PA, I:INTEREST_ID, S:SCOPE) then partly_matched_new_object(O:OBJECT_ID, P:PA, I:INTEREST_ID); if offered_object_scope(O:OBJECT_ID, S:SCOPE) and interest(I:INTEREST_ID, S:SCOPE) then partly_matched_offer(O:OBJECT_ID, I:INTEREST_ID); if partly_matched_offer(O:OBJECT_ID, I:INTEREST_ID) and not rejected_offer(O:OBJECT_ID, I:INTEREST_ID) then accepted_offer(O:OBJECT_ID, I:INTEREST_ID); Table Knowledge involved in Website-initiated behaviour For each scope in the new object a comparison to the existing interests in the profile is made. When they match, the object is partly selected. However, on another scope, the interest and the new object may differ. Only if all of the scopes of the object match is the object selected. The offer is made by the component agent interaction management. The Personal Assistant receives this offer and compares it to the interests in its user profile. This is performed in the Personal Assistant's determine proposals, as it is done in the Website Agent; see rule r14. Again, when no conflicting scopes can be found between the interest and the offered object, it is selected using rule r15. The selected offer is communicated to the user, who can reply to the offer. Update of user profile. After the user has communicated to the Personal Assistant whether he or she rates the offer interesting or not, a profile update process is initiated, if necessary, by removing those interests repeatedly receiving negative feedback. This feedback is used in the component interest remover to select interests for removal. Similar to the creation of new interests, a simple mechanism is used to select interests for removal. A circular list is kept of the last three responses to offers based on an interest. This list has three objects; when all three objects show a negative response, the interest is marked for removal; see rule r16 in Table 12. if last3_suggestions_response(last_id1, rejected, I:INTEREST_ID) and last3_suggestions_response(last_id2, rejected, I:INTEREST_ID) and last3_suggestions_response(last_id3, rejected, I:INTEREST_ID) then to_be_confirmed(remove(I:INTEREST_ID)) if removal_response(I:INTEREST_ID, confirmed) and believe(interest(I:INTEREST_ID, S:SCOPE)) then to_be_removed(interest(I:INTEREST_ID, S:SCOPE)); Table Knowledge involved in profile update: removal in Website-initiated case An interest marked for removal is not automatically removed. Before actual removal, the user has to give his or her approval. When the user disapproves of the removal, the three last responses to that interest are reset; thus again three rejections in a row must be received before the agent considers the interest for removal. When the user approves, the removal is performed; see rule r17. As for the interest creator, this component reasons about changes in interests and is therefore at a meta-level compared to the component maintenance of agent information. The interest is actually removed by an information link, similar to how interests are created. In this paper a generic, reusable multi-agent architecture for active intelligent Websites is presented. This generic architecture for active intelligent Websites was first designed for one application domain: a department store (cf. [22]). This application reuses the generic architecture of information broker agents developed earlier (cf. [21]), which in turn was designed as a specialisation of the generic agent model GAM introduced in [8]. The model has been designed in such a way that the generic, reusable structures are separated from the application-specific aspects in a transparent manner. The reusability of the generic multi-agent architecture for active intelligent Websites has been tested in a second application: a project on an intelligent Website for insurance in co-operation with the software company Ordina Utopics and an insurance company (cf. [20]). The outcome of this test was clearly positive. With not much effort (an investment of only a few person months) a prototype multi-agent system for an intelligent Website in insurance has been designed and implemented, based on the generic architecture. The actual work concentrated mainly on the specification of the domain concepts and application-specific knowlege bases. A Website, supported by the architecture introduced has a more personal look and feel than the usual Websites. Within the architecture, also negotiation facilities (e.g., as in [38]) can be incorporated. In the agent literature, a number of architectures for (information) broker agents can be found; e.g., [9], [10], [25], [30], [32], [36],. The design of most of these architectures is not formally specified in detail; usually they are only available in the form of an implementation, and at the conceptual level some informal pictures and natural language explanations. In general, the aim for the development of these architectures in the first place is to have a working piece of software for a specific type of application. The design of the generic architecture for intelligent Websites introduced in this paper has a different aim. The generic model was meant as a unified design model, formally specified in an implementation- and domain-independent manner at a high level of abstraction. The (multi) agent architecture described here was designed and implemented in a principled manner, using the compositional design method for multi-agent systems DESIRE [7]. Due to its compositional structure it supports reuse and maintenance; a flexible, easily adaptable architecture results. A success criterion for this aim is the possibility to specialise and instantiate the model to obtain conceptual, formal specifications of design models for different applications. The positive experience in the insurance domain, discussed above, shows that the aim was achieved. Applications of broker agents (addressed in, e.g., [9], [10], [24], [25], [30], [32], [34], [36]), often are not implemented in such a principled manner: without an explicit design at a conceptual level. Compared to, for example, systems designed using CORBA, or other object-based methods, a main difference is that in our approach functionality can be specified at the level of design in an explicit declarative manner (in the form of ontologies and knowledge bases). Especially for applications in knowledge-intensive domains this provides appropriate means to specify a design. The RETSINA approach (cf., [34], [35]) is more comparable to the design method DESIRE as such, and not to the generic architecture for the specific application type of intelligent Websites proposed here. A difference is that DESIRE is based on a formal specification language for design models. The same difference applies to the work on SIMS, described in [24]. However, in [24], also the problem of information integration is addressed (i.e., integration of information expressed in different ontologies), which has not (yet) been addressed in the architecture proposed here. A next step is to refine our model with possibilities for information integration, for example adopted from SIMS. The question whether the approach scales up has not been explicitly investigated in the research reported, by performing experiments. Since an essentially distributed approach has been chosen, the Personal Assistant agents can all be implemented on an own server. Also it is possible to implement different Website Agents on different servers, thus avoiding too much interaction overload of one server. For the particular application in insurance the generic broker agent model has been instantiated with domain ontologies and domain knowledge. In the prototype some of these instantiations have been done in an ad hoc manner, without the intention to propose these instantiations as a generic approach for more domains. Current research addresses more principled manners to use dynamic taxonomies in profile creation and techniques from inductive logic programming to induce profiles from examples. In [11] (see also Section 4.5 above) an overview is given of a number of these profiling approaches and it is shown how they can be incorparated). As an example, in further research a component-based generic agent architecture for multi-attribute (integrative) brokering and negotiation has been developed in co-operation with, among others, Dutch Telecom KPN. The agent architecture was designed as a refinement of the compositional generic agent model GAM. Within the component Maintenance of Agent Information (MAI) within this agent architecture, a profile of the human user of the agent is maintained, which includes . evaluation functions per attribute assigning to each attribute value an evaluation value between 0 and 1, . importance factors (between 0 and 1) for the different attributes. Within such a more sophisticated content-based profile model (which, for example, is also used in [3], it can be expressed, for example, that a car with colour blue is evaluated as 0.9, whereas a yellow colour is evaluated as 0.1, and a CD player of high quality is rated 0.8 whereas a CD player with low quality as 0.2. Moreover, the attribute 'colour' can be assigned, e.g., importance 0.6, whereas the attribute 'CD player' can be assigned importance 0.8. Acknowledgements Pascal van Eck (Vrije Universiteit) supported other experiments with some variants of the broker agent model. Working on a design for the department store application with various employees of the software company CMG provided constructive feedback on the architecture introduced. The authors are also grateful for discussions with employees of Ordina Utopics (in particular, Richard Schut) and an (anonymous) insurance company. --R Recommendation as Classification: Using Social and Content-Based Information in Recommendation Enabling Integrative Negotiations by Adaptive Software Agents. Springer Verlag. Learning Collaborative Information Filters. Lookahead and discretization in ILP. Scaling up inductive logic programming by learning from interpretations. Formal specification of Multi-Agent Systems: a real-world case Compositional Design and Reuse of a Generic Agent Model. An Agent Marketplace for Buying and Selling goods. Modeling User Preferences and Mediating Agents in Electronic Commerce Freuder and Richard J. Latent Class Models for Collaborative Filtering. Design of Collaborative Information Agents. A Multi-Agent Architecture for an Intelligent Website in Insurance Springer Verlag Compositional design and maintenance of broker agents. Information Broker Agents in Intelligent Websites. Decisions with Multiple Objectives: Preferences and Value Trade-offs Agents for Information Gathering. On Using KQML for Matchmaking. Learning to filter news. Machine Learning for Adaptive User Interfaces. User Modeling in adaptive interfaces. An Agent that Assists Web Browsing. Information Brokering in an Agent Architecture. The identification of interesting web sites. Issues in Automated Negotiation and Electronic Commerce: Extending the Contract Network. "Word of Mouth" Designing behaviors for information agents. IEEE Expert 11 Toward a Virtual Marketplace: Architectures and Strategies. Multicriteria Decision-aid --TR
information agent;intelligent website
591005
Multi-Dimensional Modal Logic as a Framework for Spatio-Temporal Reasoning.
In this paper we advocate the use of multi-dimensional modal logics as a framework for knowledge representation and, in particular, for representing spatio-temporal information. We construct a two-dimensional logic capable of describing topological relationships that change over time. This logic, called (Propositional Spatio-Temporal Logic) is the Cartesian product of the well-known temporal logic PTL and the modal logic S4u, which is the Lewis system S4 augmented with the universal modality. Although it is an open problem whether the full PSTL is decidable, we show that it contains decidable fragments into which various temporal extensions (both point-based and interval based) of the spatial logic RCC-8 can be embedded. We consider known decidability and complexity results that are relevant to computation with multi-dimensional formalisms and discuss possible directions for further research.
Introduction It is widely accepted that many kinds of AI applications require high-level reasoning involving spatial and temporal concepts (see e.g. (Hayes, 1979; Hobbs and Moore, 1985; Davis, 1990)). Several approaches have been applied to representing these concepts: some researchers have developed specialised computation-oriented representations such as solid geometry (Requicha, 1980), constraints techniques (Allen, 1981) and spatio-temporal database architectures (Guting et al., 2000); others have employed techniques of formal logics. Typically, logical representations have made use of the expressive power of rst-order languages (e.g. (McCarthy and Hayes, 1969; Clarke, 1981; Allen, 1984; Hayes, 1985; Kowalski and Sergot, 1986; Randell et al., 1992b)); but, since Supported by the EPSRC under grants GR/K65041 and GR/M56807 c 2000 Kluwer Academic Publishers. Printed in the Netherlands. rst-order logic is undecidable, such languages do not provide an eec- tive reasoning algorithm unless supplemented by some special purpose inference mechanism. An alternative to rst-order logic is the framework of propositional modal logic, which extends classical propositional calculus with one or more 'modal' operators, interpreted on some relational or algebraic structures. Appropriate structures may be chosen to correspond to an aspect of reality, such as time or space, which one wishes to describe within a logical representation. For example, one may think of time as a sequence of points represented by the natural numbers hN; <i on which temporal operators | 'some time in the future', `since' etc. | can be dened. For instance is true at a time point t 2 N (in symbols t only if t 0 t. The resulting propositional temporal logic, PTL, is decidable in PSPACE (Sistla and Clarke, 1985) and by Kamp's theorem (Kamp, 1968) it is as expressive (on hN; <i) as the rst-order language with unary predicates P i (x), < and =. Temporal logics of this sort are widely used in computer science (Manna and Pnueli, 1992; Stirling, 1992). Modal representations of space are relatively new to computer sci- ence. However, the possibility of giving topological interpretations of intuitionistic propositional logic has been known since the thirties (Stone, 1937{8; Tarski, 1938). Tang (1938) and later McKinsey and Tarski (1948) interpreted the necessity operator of the Lewis modal system S4 as the interior operator of topological spaces and proved that S4 is sound and complete with respect to this interpretation (for details see e.g. Chagrov and Zakharyaschev (1997)). Since S4 is decidable in PSPACE, this suggests that it might provide a computationally viable language for describing topological information. That S4 really can be used for qualitative spatial representation and reasoning was observed by Bennett (1996, 1998), who embedded the spatial language RCC-8 into extended with the universal modality 8 (for all points in the space). This encoding, which will be explained in Section 2, provides a decision procedure for a spatial language that can express a large class of signicant topological relations including all those in RCC-8. As we have seen, propositional modal and temporal logics may be decidable and hence amenable to automated reasoning (Demri, 1994; Giunchiglia and Sebastiani, 1996; Horrocks, 1998; Dixon et al., 1998; Voronkov, 1999; De Nivelle et al., 2000). This of course means that they are less expressive than rst-order logic. Moreover it is clear that modalities of single type (say, only temporal or only spatial) are not comprehensive enough for many real applications, which require rea- Time Space Space Figure 1. Motion modelled in terms of a Cartesian product of space and time soning with a range of dierent concepts. What we actually need is systems combining several types of modal operator. Such multi-modal logics have been the subject of much recent research | e.g. (Kracht and Wolter, 1991; Spaan, 1993; Marx and Venema, 1997; Gabbay, 1998; Wolter, 2000; Gabbay et al., 2000). Central problems in this eld concern how modalities can be combined, whether interplay between operators preserves decidability and other meta-logical properties, and what is the computational complexity of the resulting systems. Whereas modalities can be successfully combined as logically independent operators (Kracht and Wolter, 1991; Fine and Schurz, 1996), a key feature of most of the more interesting systems is the interaction between dierent modalities. This applies in particular to modal operators relevant to describing spatial and temporal concepts. Consider, for instance, a moving region (the black disk) depicted in Fig. 1. If we can model space as an S4 u -frame (i.e., a quasi-order) F and if the ow of time is represented by a linear order G, then the whole spatio-temporal 'universe' can be viewed as the Cartesian product F G, in which the act 'horizontally' to talk about spatial regions, while the temporal operators act 'vertically' taking care of their movements in time. Cartesian products of Kripke frames are typical examples of multi-dimensional structures that serve as models of multi-dimensional modal logics . The idea of a multi-dimensional modal logic was introduced by presented an axiomatisation for a 2-dimensional extension of S5. Products of modal logics were investigated in (Sheht- man, 1978; Gabbay and Shehtman, 1998; Marx, 1999). In computer science and AI, multi-dimensional modal logics are used e.g. for constructing temporal epistemic logics for multi-agent and distributed systems et al., 1995), temporal logics of parallel processes (Reynolds, 1997), temporal, epistemic, and dynamic description logics (Baader and Ohlbach, 1995; Wolter and Zakharyaschev, 1998, 1999, 2000c). Further details and development of multi-dimensional modal logic can be found in (Gabbay et al., 2000). Although the idea of constructing multi-dimensional modal logics may appear natural and simple, the resulting 'hybrids' often turn out to be very complex or undecidable, even if the one-dimensional components are 'almost tractable'. As with other frameworks for knowledge representation there is a delicate balance between expressive power and computational complexity. Let us consider, for instance, the following nave example of a 'compass' spatial logic (Venema, 1990). In our everyday practice, we often connect spatial structures with this or that system of coordinates. For instance, in geographical maps we have four compass directions: North, South, West, and East. Using these we can say e.g. that Moscow is to the North-East of London. Spatial relations of this sort can be expressed in a modal language with four operators N ('somewhere to the ('somewhere to the South') interpreted in the Cartesian product of two linear orders, say, hR; <i hR; <i in the standard Kripke-style manner: etc. That 'Moscow is to the North-East of London' can be represented in this language by the formula London It is well known from modal logic that the satisability problem for a language with a single operator is NP-complete for the frames hR; <i or hN; <i, (Ono and Nakamura, 1980). However, the satisability problem for a bi-modal language (with N and E ) in hR; <ihR;<i or hN; <ihN;<i turns out to be undecidable, even 1 and Reynolds, 1999; Reynolds and Zakharyaschev, 2001). The interval temporal logic of Halpern and Shoham (1986) can be interpreted in the f(x; y) 2 N N : x yg part of hN; <i hN; <i (see (Venema, 1990)), and this logic is also not recursively enumerable. There are many other examples of undecidable multi-dimensional modal logics, e.g. S5S5S5 (Maddux, 1980) or even any logic between KKK and S5S5S5 (Hirsch et al., 2000). Thus, we see that multi-dimensional modal logics are not easy to deal with, and we need to be careful in constructing eective and expressive spatio-temporal formalisms. For example, the straightforward attack on the problem by means of using the Cartesian products of frames for S4 and the ow of time hN; <i (or any other innite linear has not brought any result yet: whether the logic of such 2-dimensional frames is decidable remains one of the challenging open problems in the eld. In the rest of this paper, we discuss possible uses of multi-dimensional modal logics for spatio-temporal representation and reasoning, giving initial results already obtained in this direction. Thereby we would like both to attract the attention of the knowledge representation community to this novel and promising approach and also to give incentive to logicians to investigate modal languages that may turn out to have practical utility. 2. Region Connection Calculus and a Modal Interpretation The Region Connection Calculus (RCC) is a rst-order theory proposed by Randell et al. (1992a) for qualitative spatial representation and reasoning. 1 The basic language of RCC contains only one primitive predicate C(X; Y ), read as 'region X is connected with region Y '. 2 Many other spatial relations can be dened in terms of the C primi- tive. Of particular signicance are the eight relations depicted in Fig. 2. These form a pairwise disjoint and exhaustive set of relations known as RCC-8. Essentially the same set (but applied to the more restricted class of connected regions) has been independently identied as useful in the context of Geographical Information Systems (Egenhofer and Franzosa, 1991). In English, the relations can be described as: Dis- Connection, External Connection, Partial Overlap, Tangential Proper Part, Non-Tangential Proper Part and Equality. Formal notations for these relations are given under the diagrams. The part relations are asymmetric, so each has an inverse, su-xed by 'i'. The RCC formalism was originally presented as a nave theory in the spirit of Hayes (1979), so no specic model was assumed. However, it has been found that the theory can be interpreted in classical point-set topology (Gotts 1996a; Bennett 1997,1998). Thus, we can take as models of RCC, topological spaces, U is a non-empty set, the universe of the space, and Ian interior operator on U . 3 comprehensive account of this theory and its applications can be found in (Cohn et al., 1997). Randell et al. (1992a) also considered an extended language incorporating an additional primitive function conv(r), which denotes the convex-hull of region r. We will comment brie y on convexity later (Section 6). 3 This means that Imust satisfy the axioms I(X) X, and I(X \ Y a a a a a b a a a b Figure 2. Basic Relations in the RCC Theory Individual variables of RCC range over non-empty regular closed sets of T, i.e., an assignment in T is a map a associating with every variable X a set a(X) U such that is the closure operator on U dual to I. The connection relation C(X; Y ) is then interpreted as meaning that the point sets denoted by X and Y share at least one point: The full rst-order theory of RCC is too expressive to be computationally useful and is in fact undecidable (this follows from (Grzegor- czyk, 1951); results applying more specically to RCC can be found in (Gotts, 1996b) and (Dornheim, 1998)). Fortunately, there are various decidable (and even tractable) fragments of RCC. As was mentioned above, for certain applications one can limit the relations employed to the set RCC-8. Moreover, according to the experiments reported by Knau et al. (1997) the eight predicates turn out to be 'cognitively adequate' in the sense that people indeed distinguish between those relations. Formally, the language of RCC-8 consists of a set of individual variables X called region variables, the eight binary predicates DC, EC, PO, EQ, TPP, TPPi, NTPP, NTPPi, and the Boolean connectives (:, out of which we can construct spatial formulas. RCC-8 is interpreted in topological spaces the region variables range over non-empty regular closed sets in T, and the eight predicates are dened in the following way: For example, according to this denition, EC(X; Y ) means that X and Y share at least one point but do not share any interior point (i.e. they only share boundary points). The main reasoning task for RCC-8 can be formulated as follows: given a nite set of spatial formulas, decide whether is satis- able in a topological space, i.e., whether there exists a topological space T and an assignment a in it such that T j= a . That this satisability problem is decidable was observed by Bennett (1994), who exploited the results of Tarski (1938) to encode RCC-8 into intuitionistic propositional logic. Subsequently, Bennett (1996) gave another embedding of RCC-8 into the more expressive logic containing Lewis's S4 modality and an additional universal modality, whose meaning will be explained below. In the current paper we shall refer to this bi-modal logic as S4 u . The embedding is based on the result of McKinsey and Tarski (1948) according to which S4 is characterised by the class of topological spaces in the following sense: Given a topological space interpret the propositional variables in ' as subsets of U , the Boolean connectives as the corresponding set-theoretic operations, the necessity operator I as I, and the possibility operator C as C (we denote the usual box and diamond of S4 by I and C to emphasize their topological meaning). Now, if the value of ' is the whole space U , no matter what sets are assigned to its variables and what topological space is taken, then ' is provable in S4, and vice versa. According to (Goranko and Passy, 1992), this completeness theorem still holds if we extend S4 with the universal modalities 8 and 9 interpreted in T as 'for all points in T' and `there is a point in T', respectively. That is the value of 8' (under a certain interpretation) is U if the value of ' (under this interpretation) is U , otherwise it is ;; :8:'. It is easy to see that the language of the resulting logic S4 u is expressive enough to encode the topological meaning of the RCC-8 formulas as dened above. Given such a formula ', we replace in it occurrences of RCC-8 predicates with the corresponding modal formulas, e.g. etc. are propositional variables) and add to the result the conjunct for every region variable X i in ' to ensure that the propositional variables are interpreted by non-empty regular closed sets of topological spaces. The resulting modal formula is denoted by ' y . Now, if we recall that some (in particular, all nite) topological spaces are determined by Kripke frames hW; Ri for S4 | every such frame gives rise to the topological space hW; Ii, where for any X W | and that S4 u has the nite model property (Goranko and Passy, 1992), then we immediately obtain Theorem 1. For every RCC-8 formula ', the following conditions are equivalent: (i) ' is satisable in a topological space, (ii) ' y is satisable in a topological space, (iii) ' y is satisable in some nite Kripke frame for S4. Thus we reduce the satisability problem for RCC-8 formulas to the satisability problem for propositional bimodal formulas in Kripke frames for S4 u , which is decidable (Goranko and Passy, 1992). Moreover, one can show that every satisable formula of the form ' y can be satised in a partially ordered Kripke frame each point in which has at most two (incomparable) successors, and the number of worlds in this frame is linear in the number of symbols in 'y (and consequently also linear in the length of the RCC-8 formula '). This result was obtained by Renz (1998); a somewhat more general theorem is proved in (Wolter and Zakharyaschev, 2000a). It follows in particular that the satisability problem for RCC-8 formulas is NP-complete; see (Renz and Nebel, 1997; Renz and Nebel, 1999) where maximal tractable fragments of RCC-8 are also described. It also follows that all satisable RCC-8 formulas can be satised in R n for any n 1. (Note however that S4 u is not complete with respect to R n .) 3. Point-based temporal RCC-8 One approach to constructing spatio-temporal logics is to combine RCC-8 with point-based temporal logics, for instance, the well-known propositional temporal logic PTL interpreted on the ow of time hN; <i and having the temporal operators S (Since) and U (Until). Other standard operators can be dened: ('at the next moment "), (p _ :p) U ' (`at some time in the future '') in the future ''), and similarly for their past counterparts f , and . The spatial regions occupied by the objects under consideration may change with time passing by, but the topological space in which they are moving always remains the same. This nave picture is formalised by the following concept of topological temporal model. Denition 1. A topological temporal model (or tt-model, for short) based on a topological space is a triple of the form a, an assignment in T, associates with every region variable X and every moment of time n 2 N a non-empty regular closed subset of U . For each n, we take a n to be the function dened by a There are several dierent ways of introducing a temporal dimension into the syntax of RCC-8. The most obvious is to allow applications of the operators S and U (along with the Booleans) to spatial formulas. We call the resulting spatio-temporal language ST 0 . Denition 2. For a tt-model ai, an ST 0 -formula ', and dene the truth-relation (M; n) meaning `' holds in M at moment n,' by induction on the construction of ': if ' contains no temporal operators, then (M; n) an '; (M; n) there is k > n such that every l such that n < l < k; (M; n) there is k < n such that every l such that k < l < n. allows us to say many things about changes of spatial relationships over time. For example, means that 'Kosovo will not always be part of Yugoslavia.' It is also expressive enough to constrain movements to be continuous, in so far as one can describe possible continuous transitions among the RCC-8 relations that can hold between two regions (such transitions were identied by Randell et al. (1992a) and the importance has been emphasised more recently by Muller (1998)): etc. However, the expressive power of ST 0 is limited in that one can only employ the RCC-8 relations to compare region variables at the same point in time; so it does not allow one to describe spatial relations between the extensions of a region variable at dierent times. For ex- ample, there is no way to say something like 'The extension of the EU is a part of what its extension will be next year (i.e. at the next time To overcome this limitation we introduce the logic ST 1 , which extends ST 0 by allowing applications of the next-time operator f not only to formulas but also to region variables. Thus, arguments of the RCC- 8 predicates are now region terms, which consist of a region variable that may be prexed by an arbitrarily long sequence of f operators. For instance f f EU could denote the region occupied by the EU in two years time and one can write formulas such as P(EU ; f f EU ). The semantics of ST 1 is just that of ST 0 extended by the following clause: Using ST 1 we can now talk about the changing extensions of individual region variables. For instance that 'the EU will never shrink'. The new construct may also be used to rene the continuity assumption by requiring that i.e., 'regions X and f X either coincide or overlap.' We can also express the condition that the extension of a region variable X is xed for the future: or that it has at most two distinct states, one on even days, another on odd ones: X). Note, that the only in models based on innite topological spaces, whereas for formulas of ST 0 (and of course nite topological spaces always su-ce. It may appear that ST 1 can compare regions that are separated in time only by limited sequences of time points. However, using an auxiliary variable, whose extension is constrained to be constant over time, we can write, for instance, which is satisable i 'someday in the future the present territory of Russia will be part of the EU.' This contrasts with the formula meaning that there will be a day when Russia (its territory on that day | perhaps without Chechnya but with Byellorus- sia) becomes part of the EU. Imagine now that we want to say that every location in Europe will pass through the Euro-zone, but only the land currently occupied by Germany will use the Euro forever. Unfortunately, we do not know which countries will form within Europe in the future, so we can't simply write down all formulas of the form What we need is the possibility of constructing regions containing all the points that will belong to region X at some time in the future and containing just those points included in all future states of X . Then we can write: P(Europe; Similarly, the formula P(Russia; that all points of the present territory of Russia will belong to the EU at some time in the future (but perhaps at dierent moments of time). To permit fully general application of temporal operators to region variables we dene the language ST 2 in which region variables may be prexed by arbitrary strings of the operators f , . The semantics of the new operators are given by: k>n a(t; k)), k>n a(t; k)). Finally, we can make our languages ST i , for even more expressive by allowing applications of the Boolean operations to region terms. Their semantical meaning is dened as follows: i be the resulting family of languages. In these languages we can write formulas such as EQ(UK,Great Britain _ Northern Ireland), meaning that the extension of the UK is the sum of the extensions of Great Britain and Northern Ireland. 4. Modal Encoding of ST +The decidability and complexity results for the ST i languages were proved in (Wolter and Zakharyaschev, 2000b) by embedding them into the two-dimensional propositional modal logic S4 u PTL, the Cartesian product of the logics S4 u and PTL. We call this logic Propositional Spatio-Temporal Logic or PSTL. Its connectives are: the Booleans, the necessity and possibility operators I and C of S4, the universal necessity and possibility operators 8 and 9, and the temporal operators S and U . It is not known whether the full PSTL (or even S4PTL) is decidable. However, its fragments corresponding to the ST i languages are quite manageable. The modal translation from ST into PSTL is dened by extending the y transformation specied in Section 2 to handle region terms containing Boolean and temporal operators. The region terms are already ostensibly formulas of PSTL; however, we must bear in mind that these terms are intended to denote only regular closed subsets of the space. The easiest way to take this into account is to prex CI to every subterm occurring in an ST +-formula. What is the intended semantics of PSTL? As we have seen, when encoding pure RCC-8 into S4 u we can specify the semantics with either topological models or Kripke models and these are equivalent in terms of validity. However, the addition of the temporal component makes the situation more complicated in that the topological and Kripke semantics do not perfectly agree on the class of valid formulas. Let us dene these two types of model structures: Denition 3. A Kripke PSTL-model is a triple is a quasi-order (a frame for S4) and V, a valuation, is a map associating with every propositional variable p and every n 2 N a subset V(p; n) W . For each spatial point u 2 W and each time point N, the truth-relation (u; n) (u; n) (u; n) (u; n) such that uRv, (u; n) there is k > n such that plus the standard clauses for S and the Booleans. A PSTL-formula ' is satised in K if (u; n) Denition 4. A topological PSTL-model is a structure in which is a topological space and U is a map associating with every propositional variable p and every n 2 N a set U(p; n) U . U is then extended to arbitrary PSTL-formulas in the following way: there is a k > n such that x 2 U(; m) for all m; n < m < k. Consequently the dened temporal operators are interpreted by A PSTL-formula ' is satised in N if U('; n) 6= ; for some n 2 N. The sets of PSTL-formulas satisable in Kripke models and topological models turn out to be dierent. Of course, every Kripke model is equivalent to some topological model. But the converse does not hold. A good example is provided by the formula This is valid in every Kripke PSTL-model because from any space-time point in such a structure any given other point is reachable by forward transition along the time line followed by transition along the accessibility relation of F just in case it is reachable by rst moving along the accessibility relation and then forward along the time dimension (since the accessibility relation remains constant for all time points). However, in certain innite topological spaces (e.g. R) one can construct innite sequences X n of closed sets such that S closed. In a topological model based on such a space need not have the same denotation as C + p. This subtlety concerning innite unions is not accounted for by the Kripke approach. The divergence between topological and Kripke PSTL-models is problematic because whereas the topological models correspond to the desired spatial interpretation, most currently known methods of determining decidability and complexity are based on Kripke models. Fortunately, Wolter and Zakharyaschev (2000b) have been able to show that the two semantics agree on the satisability of ST +-formulas as long as we adopt a reasonably natural Finite State Assumption (FSA). The FSA requires that over the innite sequence of time points each region that one can refer to can have only nitely many distinct extensions (but it may change its extension innitely often). Although this restriction rules out many mathematically interesting possibilities, it is perfectly satisfactory for a wide range of practical applications (for example planning tasks where we want to get from an initial to a nal situation | in a nite number of steps). To formalise the FSA we dene the following restricted class of tt-models: Denition 5. Say that a tt-model or is an FSA-tt-model, if for every region term t there are nitely many regular closed sets A U such that fa(t; n) g. It can be shown that an ST 2 -formula is satisable in an FSA-tt- model i it is satisable in an FSA-model based on a nite topological space. Because of the very restricted combinations of operators that result from translating RCC-8 predicates and their region term arguments, the modal translations of ST -formulas form a rather special fragment of the modal language PSTL. As was mentioned above, Renz (1998) showed that an RCC-8 formula ' is satisable i ' y is satisable in a Kripke model based on an S4-frame of depth 1 and width 2 (which means that it contains no chains of more than 2 distinct points, and no point has more than 2 distinct successors); and it turns out that this result can be generalised to formulas of ST +and ST +: Theorem 2. a) An ST -formula ' is satisable in a FSA-tt-model i ' y is sat- isable in a Kripke PSTL-model (which also satises FSA) whose underlying S4-frame is of depth 1 and width 2. 4 -formula ' is satisable in a tt-model i ' y is satisable in a Kripke PSTL-model of depth 1 and width 2. This result makes it possible to use the method of quasi-models (Wolter and Zakharyaschev, 1999) to prove that the satisability problem for all the languages ST in tt-models is decidable. Given the rather weak interaction between time and space in ST 0 it is not hard to show that the satisability problem for ST 0 formulas in tt-models is PSPACE-complete (the same as that of PTL). The satisability problem for ST 1 formulas in tt-models is decidable in EXPSPACE (and becomes NP-complete for the sublanguage of ST 1 with only the temporal operator f ). If we restrict the admissible models to FSA-models then the satisability problem for ST 2 formulas is decidable in EXPSPACE. Moreover, all the complexity results just given remain valid after replacing ST i with ST 4 Actually, since occurrences of + within the region terms of ST are always regularised by adding the prex CI, one might conjecture that the FSA is not necessary. However, at present we cannot envisage how to prove this. The topological temporal models we were considering above are based on the discrete ow of time N. By replacing N with Q, R, or any other strict linear order we can extend our semantics to cover dierent ows of time. But then the question arises as to whether the decidability and complexity results proved for N can be extended to, say, the logic determined by the class of arbitrary strict linear orders, by the reals R, or by the rationals Q. It turns out that for the language of ST 0 we can easily extend the decidability proof for N by using the fact that the propositional temporal logics based on those ows of time are decidable (Gabbay et al., 1994). As concerns ST 1 , observe that the operator f is meaningless for dense linear orders|thus for Q and R this language reduces to ST 0 . Decidability of ST 1 interpreted in arbitrary linear orders is an open problem. And nothing is known about the decidability of ST 2 interpreted in ows of time dierent from N. We conjecture, however, that the methods developed in (Hodkinson et al., 2000) can be used to prove the decidability of ST 2 based on any of the ows of time mentioned above (under FSA). 5. RCC-8 and Interval Temporal Logic We have seen how the RCC-8 theory can be temporalised by combining it with a point-based temporal logic. However, since the region-based approach to spatial reasoning was inspired by and closely mirrors the interval-based approach to temporal reasoning (Allen, 1981; Allen, 1984) (they both take extended entities, rather than points as primi- tives) it would seem far more natural to temporalise RCC-8 by combining it with an interval-based logic. In this section we show that this can indeed be done; and moreover, that the resulting system can in fact be embedded by suitable syntactic denitions within the point-based logic PSTL dened above. We remind the reader that Allen's logic has thirteen basic relations between time intervals | Before(i; j), Meets(i; j), Overlaps(i; j), The set of formulas constructed using these predicates and the Booleans can be regarded as a temporal 'twin' of RCC-8. Following Allen (1984), we write HOLDS(R; i) to say that the relation R holds during some time interval i. Thus HOLDS(PO(X; Y ); i) means that during interval i regions X and Y partially overlap. Let us call an ARCC-8 formula any Boolean combination of basic temporal predicates and formulas of the form HOLDS('; i) where ' is an RCC-8 formula. Here is a simple example of a valid entailment in this unsophisticated language: HOLDS(TPP(Hong Kong HOLDS(DC(Hong Kong ; UK); j); There are dierent views on how intervals should be modelled in dierent time ows. A common interpretation is that the intervals are treated as ordered pairs of distinct points of the domain Q or R. Within our framework we can in fact adopt any of these models; but, for simplicity of presentation it is convenient to employ a semantics where intervals are arbitrary convex non-empty subsets of the time points of an arbitrary time ow. Thus we give the following semantics for ARCC-8: Denition 6. An it-model (interval topological model) is the triple is a strict linear order (modelling the intended ow of time); is a topological space; and assignment a associates with every interval variable i a non-empty convex subset of W and with every region variable X and every moment of time u it associates a regular closed set a(X; u) in T. The truth-relation is dened inductively as follows (clauses for the Booleans are standard): and not M j= a Before(i; j); The other interval relations are dened similarly; M j= a HOLDS('; i) i for every point u 2 a(i) we have T We now show how by using some ideas of Blackburn (1992) the language ARCC-8 can be directly embedded into PSTL. For convenience we dene undirected universal and existential modalities over time: is true at every moment) and We extend the translation function y dened above to encode the interval relations and HOLDS predicate. For example, we replace in ' and the other Allen relations can be encoded in similar fashion. Here the variables t i are just ordinary propositional variables; however, we need to constrain their interpretation so that they can be employed to stand for intervals. Therefore, we add to the resulting formula the conjuncts for every interval variable i occurring in ', thus obtaining ' y . The rst conjunct ensures that t i is non-empty; 5 the second that it is a convex interval of the time series; and the nal conjunct means that the value of t i is constant relative to the spatial dimension. It can be shown that an ARCC-8 formula ' is satisable in an it-model just in case its modal translation ' y is satisable in a Kripke PSTL-model. Moreover, the satisability problem for ARCC-8 formulas in it-models is NP-complete. Thus ARCC-8 has the same computational complexity as both RCC-8 and the constraint language of Allen's temporal interval predicates (Vilain et al., 1986). 6. How Far Can Multi-Dimensional Modal Logic Take Us? We have seen how the 2-dimensional language PSTL provides an expressive formalism for representing spatio-temporal information, which encompasses both topological constraints and linear temporal logic. We now consider what other concepts one might want to represent and note some di-culties that arise due to known complexity results. A concept that is very useful for describing real situations is that of spatial convexity. The addition to the rst-order RCC formalism of a function giving the 'convex-hull' of a region proposed in (Randell et al., 1992a) (such a function and a convexity predicate true of convex regions are interdenable). First-order languages including topological relations and a convexity predicate have been found to be highly expressive (Pratt, 1999). Although satisability of a combination of 5 If we replace this conjunct by out 'intervals' consisting of a single time point. RCC-8 relations and convexity predicates is known to be decidable is has also been shown to be as hard as solving systems of non-linear constraints over R (Davis et al., 1999). It is unclear whether modal logics can contribute to reasoning about convexity. The obvious models of multi-dimensional space within which convexity constraints could be specied involve cross products of linearly ordered innite frames corresponding to coordinate axes of the space. However, it is known that all modal logics of such products are undecidable (Reynolds and Zakharyaschev, 2001). Balbiani et al. (1997) have given a modal logic of incidence geometry within which one can express collinearity and betweenness (and hence convexity) but it is not known whether this logic is decidable. Another concept crucial to real situation descriptions is that of connectedness of regions. The objects we normally think and talk about are connected in the sense that any two points within an object can be joined by a path that lies entirely within the object. One of the most important open problems in spatial reasoning is whether there is a decision procedure for testing satisability of sets of topological relations holding among connected planar regions. A lower exponential complexity bound follows from the results of (Kratochvl, 1991; Kratochvl and Matousek, 1991) concerning an analogous problem for planar graphs; and this result applies even in the case where we deal with only the basic non-disjunctive relations of RCC-8. We speculate that a modal analysis might shed some light on this decision problem. One might also want to introduce more expressive power in the temporal dimension. In order to describe continuous changes more adequately one may want to employ a logic based on a dense model of time (Barringer et al., 1986). Within such a time ow, we must consider whether spatial relationships hold over open or closed intervals. This problem has be examined by Galton (1997,2000), who argues that whether a spatial relation is true at the bounding point of an interval over which it holds depends on the nature of the relation in question. This analysis may allow for more detailed description of changing spatial relationships than is possible within Allen's (1984) treatment. Also, rather than treating time as a linear ordering one might like to model alternative possible histories in terms of some branching structure, such as is described by the logic CTL (Emerson and Halpern, 1986). Semantically it is straightforward to combine the spatial interpretation of S4 u with any reasonable temporal logic; but so far nothing is known about the complexity of combinations with non-linear time- ows. In considering decidability and complexity of multi-modal systems, logicians have almost always looked at the problem for arbitrary formulas of the resulting modal language. However, as we saw in the case of PSTL, sub-languages with severely restricted syntax (e.g. ST (+) may be able to express a signicant vocabulary su-cient for many applications. Thus we suggest that looking for such sub-languages is a research area of great potential. In order to really demonstrate the utility of multi-dimensional logics one would need to develop and implement reasoning algorithms capable of carrying out useful reasoning tasks. Automated reasoning with multi-modal logics is still in its infancy; but recently there have been some successes in developing proof methods (Dixon et al., 1998; Hustadt et al., 2000). A tableau calculus for the local cubic logic LC 2 , which is closely related to S5S5, has been implemented by Marx et al. (1999). 7. Conclusion We have outlined the general structure of a knowledge representation formalism, based on multi-dimensional modal logics. This framework seems to be well suited for representing a large vocabulary of useful high-level spatio-temporal relations. Our decidability and complexity results for the languages ST (+) i and ARCC-8 show that this approach enables one to construct very expressive yet eective spatio-temporal languages. We hope that the reader will take from this paper not only the particular details of the language PSTL but a wider appreciation of the possibilities of applying multi-dimensional modal logics to the development of Articial Intelligence. We would also like to suggest that spatio-temporal reasoning is an area within which cooperation between pure logicians and researchers tackling specic reasoning problems arising in applications can lead to both interesting theorems and powerful practical algorithms. --R Modal Logic. 'A calculus of individuals based on 'connection Representations of Commonsense Knowledge. Principles of Knowledge Representation and Reasoning: International Journal of Geographical Information Systems 5(2) Reasoning About Knowledge. Temporal Logic: mathematical foundations and computational aspects Fibring Logics. Formal Theories of the Commonsense World. Journal of Combinatorial Theory The temporal logic of reactive and concurrent systems. Principles of Knowledge Representation and Reasoning: Journal of Logic and Computation Casopis pro p Revised version in (Weld and De Kleer Readings in Qualitative Reasoning About Physical Systems. The decision problem for combined modal logics. Fundamenta Informaticae Frontiers of Combining Systems 2. --TR --CTR Alfredo Burrieza , Manuel Ojeda-Aciego, A Multimodal Logic Approach to Order of Magnitude Qualitative Reasoning with Comparability and Negligibility Relations, Fundamenta Informaticae, v.68 n.1-2, p.21-46, January 2005 Andrzej Skowron , Piotr Synak, Complex Patterns, Fundamenta Informaticae, v.60 n.1-4, p.351-366, January 2004 Norihiro Kamide, Linear and affine logics with temporal, spatial and epistemic operators, Theoretical Computer Science, v.353 n.1, p.165-207, 14 March 2006 Andreas Schfer, Axiomatisation and decidability of multi-dimensional Duration Calculus, Information and Computation, v.205 n.1, p.25-64, January, 2007 Stphane Demri , Deepak D'Souza, An automata-theoretic approach to constraint LTL, Information and Computation, v.205 n.3, p.380-415, March, 2007 Frank Wolter , Michael Zakharyaschev, Qualitative spatiotemporal representation and reasoning: a computational perspective, Exploring artificial intelligence in the new millennium, Morgan Kaufmann Publishers Inc., San Francisco, CA,
modal logic;multi-dimensional logic;spatio-temporal reasoning
591458
Slipping and Tripping Reflexes for Bipedal Robots.
Many robot applications require legged robots to traverse rough or unmodeled terrain. This paper explores strategies that would enable legged robots to respond to two common types of surface contact error: slipping and tripping. Because of the rapid response required and the difficulty of sensing uneven terrain, we propose a set of reflexes that would permit the robot to react without modeling or analyzing the error condition in detail. These reflexive responses allow robust recovery from a variety of contact errors. We present simulation trials for single-slip tasks with varying coefficients of friction and single-trip tasks with varying obstacle heights.
Introduction R OUGH terrain occurs not only in natural environments but also in environments that have been constructed or modified for human use. Currently, most legged robots lack the control techniques that would allow them to behave robustly on such relatively simple rough terrain as stairs, curbs, grass, and slopes. Even smooth terrain becomes difficult to traverse if it includes small obstacles, loose particles, and slippery areas. Many control systems for bipedal robots have assumed steady-state running over smooth surfaces, but some have explored control techniques for rough terrain. Statically stable robots, which always maintain their balance over at least three legs, have used controllers with foot-placement algorithms to insure viable footholds. However, for dynamically stable robots, which run with a ballistic flight phase, constraints on timing and foot placement increase the difficulty of designing controllers that can anticipate rough terrain or react to er- rors. This paper demonstrates the effectiveness of preprogrammed high-level responses to errors during locomotion in a complex dynamic environment. A suite of responses allows a simulated, three-dimensional, bipedal robot to recover from slipping on low friction surfaces and tripping over small obstacles (Figure 1). Many ground contact errors would be avoided if the control system could guide the robot around slippery areas and obstacles. However, the approximate nature of sensor information obtained at a distance means that it is not always possible to sense the surface properties of terrain before making contact. For example, small holes, bumps, debris, and sticky or slippery areas are difficult to detect from a distance with current technology. If the robot cannot detect and avoid or prepare for surface features in ad- vance, then robust locomotion on rough terrain requires that the robot respond to unexpected features after the contact error has occured and before the robot crashes. For College of Computing, Georgia Institute of Technology, Atlanta, GA 30332-0280. [gboonejjkh]@cc.gatech.edu. Submitted to Autonomous Robots. Slip Sequence Trip Sequence Fig. 1. Examples of a Slip and Trip. Without the addition of reflexes for recovering from slips and trips, the simulated robot does not respond successfully to slippery areas or contact with an obstacle. Foot Foot 3 degree of freedom hip joint for each leg 1 degree of freedom telescoping leg joint for each leg Rearward Leg Body Direction of Travel Forward Leg Fig. 2. Biped Structure. The simulated bipedal robot consists of a body and two telescoping legs. Each leg has three degrees of freedom at the hip and a fourth degree of freedom for the length of the leg. dynamically stable robots, the time available for modeling the surface and planning an appropriate reaction is severely limited. In the case of the dynamically stable bipedal robot shown in Figure 2, the controller may have less than a few hundredths of a second in which to choose or plan an appropriate recovery. We define reflexes as responses with limited sensing and no explicit modeling. That is, the robot can detect a slip or a trip, but makes no attempt to estimate the properties of the surface or obstacle or to calculate a corresponding recovery plan. Instead, the slipping and tripping sensors trigger fixed responses. These reflexes are defined at a high level, such as reconfigurations of the leg positions, and at a low level, such as modifications of servo gains. Just as animal motor programs can be considered both open-loop and closed-loop[1], several low-level feedback control laws operate during the primarily open-loop reflex responses. For example, a reflex may reconfigure the leg position, but sensing is used to determine transitions in the leg controller state machine during the recovery step. Fig. 3. Physical Biped Slip. Planar two-legged robot running across an oily spot on the floor. Footage from the MIT Leg Labora- tory. [Frames: 0, 35, 70, 91, 105, 140] During experimentation with a physical, planar biped, the robot sometimes slipped on hydraulic oil or tripped on cables in its path. Because the robot had no responses customized for these error conditions, it almost always immediately crashed. This paper reports a set of fixed reflexes that enable robust recoveries for a simulated three-dimensional robot in tasks involving a single slip or trip. In the next section, we describe previous approaches to legged locomotion in rough terrain. In Section III, we consider biological reflexes. Section IV describes the simulated bipedal robot and its control system. The slipping prob- lem, slipping reflexes, and simulation results are presented in Section V, followed by the tripping problem, tripping re- flexes, and results in Section VI. The reflex approach and results are discussed in Section VII. II. Locomotion on Rough Terrain A suitable foothold is one that allows a legged system to maintain balance and continue walking or running. For statically stable locomotion, the difficulty is not in placing the robot's feet on footholds, but in deciding which locations on the terrain provide suitable footholds. Successful locomotion on rough terrain was demonstrated by the Adaptive Suspension Vehicle[2] and by the Ambler[3], [4], [5]. These large, statically stable machines traversed grassy slopes, muddy cornfields, and surfaces that included railroad ties and large rocks. Static stability allowed these robots to emphasize detection at a distance and avoidance of obstacles and uncertain footholds. Klein and Kittivatcharapong[6] proposed algorithms for insuring that foot forces remain within the friction cone and identifying situations in which these constraints, or the desired body forces and torques, could not be achieved. Their work addressed prevention of slipping and did not consider sensor noise or responses to unmodeled surfaces. For dynamically stable robots, the control of step length for locomotion on rough terrain interacts with the control of balance. Hodgins and Raibert[7] implemented three methods for controlling step length of a running bipedal robot. Each method adjusted one parameter of the run-0.5hip altitude forward speed (m/s) pitch -2020leg angle foot position time Fig. 4. Slipping Data of the Physical Robot. The physical planar robot slipped on oil during a laboratory experiment at the point indicated by the vertical dotted lines. The top three graphs show the height, forward speed, and orientation of the body. The bottom two graphs show the angles of each leg and the position of each foot on the ground. For each step but the last, the foot is stationary while it is on the ground. ning cycle: forward running speed, running height, or duration of ground contact. In laboratory demonstrations, a biped running machine used these methods for adjusting step length to place its feet on targets, leap over obstacles, and run up and down a short flight of stairs. However, unlike the tasks described below, the size and location of the objects were known to the controller in advance. developed algorithms for running on terrain that was known to be slippery. By running slowly, the robot generated nearly vertical foot forces. His controller used a priori knowledge or estimation of friction coefficients to prevent slipping by confining control forces and torques to slip-free regions. Kajita and Tani[9] used an ultrasonics sensor to construct a ground profile of terrain that consisted of horizontal surfaces at varying heights. Yamaguchi et al have built a bipedal robot that uses feet to sense ground inclinations and plan appropriately[10], although it was not able to react to slips or trips. III. Reflexive Responses to Errors Biological systems use many different reflexes in locomotion and manipulation. Reflexes help to restore balance when perturbations occur during walking or stand- ing[11], [12], [13]. The role of reflexes in walking is com- plex: the same stimulus elicits a different response in the stance phase than in the swing phase[14], [15], [16]. Touching the foot of a cat or human during a swing phase, for example, will cause the leg to flex, raising the foot. If an obstacle caused the stimulus, this response might lift the foot over the obstacle and allow walking to continue. During the stance phase, a stimulus delivered to the foot causes the leg to push down harder, resulting in a shorter stance phase. Although these actions are opposite, both facilitate the continuation of locomotion. Robotics has adopted the term "reflex" from the biological literature, but in both biology and robotics the precise definition of the term varies from study to study. Most researchers in robotics use the term to mean a quick response initiated by sensory input. Some require reflexes to be open-loop and to proceed independently of subsequent sensory input[17], [18]; others apply the term more loosely to describe actions that are performed with feedback until a terminating sensory event occurs[19]. In some cases, reflexes refer to general purpose actions[20], [21] and in others only to actions taken to correct errors or to compensate for disturbances[19] or transitions[22]. Brooks's subsumption architecture[21] combined several simple reflex-like actions to produce complex behaviors such as six-legged walking. A global gait generator specified the order of leg use while inhibitory connections between the legs prevented conflicting reflexes from acting simultaneously. Other hexapod robot researchers have designed subsumption controllers for rough terrain[23] and have integrated reactive leg control with gait planning for rough terrain[24]. Hirose[20] built and controlled a statically stable quadruped that used a reflexive probing action to climb over objects and to walk up and down steps without visual input or a map of the terrain. Wong and Orin[19] implemented two reflex responses for a prototype leg of the Adaptive Suspension Vehicle. Using velocity and hydraulic pressure information from sensors at the joints, they were able to detect foot contact and slippage. A foot contact reflex reduced the peak forces at touchdown. A foot slippage reflex was used to detect and halt slipping. Reflex responses have also been used in manipulation. Tomovic and Boni[17] used a reflex response to implement grasping for the Belgrade prosthetic hand. Bekey and To- movic[18] continued the exploration of prosthetic control systems with a rule-based technique that relied on sensory data and fixed response patterns. IV. Dynamic Bipedal Robots The simulated robot used in our research is based on a planar bipedal robot constructed by Raibert and col- leagues[25], [26]. The simulated robot is three-dimensional and has three controlled degrees of freedom at each hip and one for the length of each leg (Figure 2). In the physical robot, the leg contains a hydraulic actuator in series with an air spring. The simulation models the spring and actuator as a linear spring with a controllable rest length. In Fig. 5. Simulated Biped Slip. The dark circle represents an area of the floor with a reduced coefficient of friction. Without slipping reflexes, the simulated robot is unable to complete a step on a slippery surface. The first leg slips, almost immediately becoming airborne as it accelerates forward. As the body falls, the second leg hits the surface and also slips. The second leg continues to accelerate forward. [Friction coefficient: 0.04. Times (s): 0.0, 0.06, 0.09, 0.11, 0.12, 0.13] experiments with the physical robots, hydraulic fluid occasionally created slippery spots that caused the robot to fall Figures 3 and 4). A simulation of a similar fall is plotted in Figures 5 and 6. The physical robot was also able to climb stairs and jump over boxes[26]; however, the positions of the obstacles were known in advance. The current research extends the controller to handle unexpected slips and unanticipated collisions with a box. The controller achieves dynamically stable, steady-state running by decomposing the control problem into three largely decoupled subtasks: hopping height, forward ve- locity, and body attitude. Hopping height is maintained by adding enough energy to the spring in the leg during stance to account for the system's dissipative losses. Forward velocity is maintained by choosing a leg angle at touchdown that provides symmetric deceleration and acceleration as the leg compresses and extends. The attitude of the body (pitch, roll, and yaw) is maintained with proportional-derivative servos that apply torques between the body and the leg while the foot is on the ground. The robot control system is implemented as a state machine that sequences through the flight and stance phases for each leg, applying the control laws that are appropriate for each state. As shown in Figure 7, flight is followed by a stance phase of four states. During loading , the foot makes contact with the ground and begins to bear the weight of the robot. During compression, the leg spring is compressed by the downward velocity of the robot. After the spring has stopped the vertical deceleration of the body, the body begins to rebound during thrust . As the leg reaches maximum extension during unloading , it ceases to bear weight. After liftoff, the roles of the legs are reversed and the second leg is positioned forward in anticipation of touchdown. For further details on the control system, see [26] and [25]. The control system's state machine depends on measurements of leg length to determine state transitions dur-0.20.61.0 hip altitude forward speed (m/s) -0.3 -0.2 -0.0pitch (rad) leg angle (rad) foot position time Fig. 6. Slipping Data of the Simulated Robot. After taking five steps on a surface with a friction coefficient of 1.0, the simulated robot steps on a region with a coefficient of 0.20 and slips. Because no slipping recovery strategies are active, the robot falls. The top three graphs show the height, forward speed, and orientation of the body. The bottom two graphs show the leg angles of both legs and the position of each foot on the ground. When the foot slips (vertical dotted line), it leaves the ground and the other foot soon impacts. Loading Compression Thrust Unloading Direction of Travel Flight Fig. 7. Control States. Running is achieved by dividing each step into several states and applying the appropriate control laws during each part of the running step. ing steps. Slips may interfere with control by altering leg lengths unexpectedly. The transition from loading to com- pression, for example, occurs when the leg has shortened by a small amount. After a slip, the leg may lengthen. Not only must slipping reactions prevent these errors, but they must minimize interference with normal control, such as the adjustment of body attitude. V. Slipping The impact of the foot on the ground, the weight of the robot, and the forces and torques generated by the hip and leg servos create a force on the ground during a step (Fig- F Direction of Travel Fig. 8. Foot Forces and the Friction Cone. During a step, the foot produces forces on the ground, F , with horizontal and vertical components,F h and Fv . Slipping occurs when the angle of the impact force is outside the friction cone. ure 8). Slipping occurs when the horizontal component of the force of the foot on the ground, F h , exceeds the maximum force of static friction generated by the ground. A simple model of this interaction is that the maximum force of static friction is directly proportional to the normal force of the ground on the foot, F v . Under this model, slipping will occur when the horizontal component of F exceeds the vertical component times the coefficient of static friction: where - s is the coefficient of static friction. When slipping occurs, the horizontal force returned by the ground is given by where - d is the coefficient of dynamic friction and the sign of F h should remain unchanged. These relationships define a friction cone, illustrated in Figure 8. When the force of the foot on the ground lies within the friction cone, the foot does not slip. The angle of the cone is given by Note that this cone is defined for foot forces, not leg angles. The motion of the leg prior to impact affects the direction of the foot's force on the ground, as do the control torques applied to the hip joint and the leg spring. Foot forces are most likely to exceed the friction cone at the beginning or end of a step, when the angle of the force vector is greatest. Slips at the beginning of a step are more likely than slips at liftoff because the foot is moving with respect to the ground at touchdown. In contrast, the foot is stationary at liftoff. Slips during liftoff are often less critical because the step is nearly complete; the controller has already executed corrections during the step. Our simulations assumed minimal sensory information: the properties of the surface and the extent of the slipping area were not available to the control system. The controller could not adjust the leg configuration prior to touchdown or try to position the foot outside the slippery area to find a secure foothold. Neither the forces on the feet nor the coefficients of friction were available to the control system. However, the control system could detect slips. In the simulation, slips were detected when a foot moved while in contact with the ground. The control system of a Increase Leg Force Increase Hip Torque Direction of Travel Fig. 9. Same-Step Reactions. When a slip has been detected, a torque can be applied at the hip to reduce the horizontal force on the ground or the leg can be extended to increase the vertical force. physical robot can detect slips indirectly by measuring joint angles and velocities or structural forces. Direct methods include encoder wheels and micro-slip detectors. When the control system has detected a slip, it can attempt to continue the step or abandon that step and pull the leg off the ground. In the first case, hip torques or leg forces can be applied to increase the vertical component of the foot force while decreasing the horizontal component, thus returning the force vector to within the friction cone. If the step is abandoned, one of the legs can be positioned during the next flight phase so that the leg angle at the next touchdown will be near vertical or both legs can be moved to a triangular configuration. In the simulations described here, we defined a response to be successful if the robot was able to continue running after slipping and taking a recovery step in the slippery region, then taking subsequent steps on a non-slippery surface. Changes in velocity or hopping height were not considered failures provided that the control system was able to maintain balance and return to steady-state running. A. Same-Step Response Strategies Reacting to a slip requires careful management of the horizontal and vertical components of the forces generated by the impact of the foot on the ground. Initial responses to a slip can attempt to alter the force vector immediately by generating a torque at the hip or a force axial to the leg Figure 9). The first reaction responds to a slip by increasing the hip torque by a fixed amount. In most cases, this action increases the vertical component of the foot's force on the ground. After the foot stops slipping, the hip controller reverts to its normal task of correcting pitch errors. This strategy may have undesirable consequences because a torque applied at the hip also increases the forward velocity of the body thus increasing the likelihood of a slip on a subsequent step. Applying a torque at the hip also interferes with the correction of body attitude during stance and tends to increase the pitch of the body. The second reaction responds to a slip by compressing the leg spring a fixed amount to increase the vertical force at the foot and regain a foothold. In a normal running step, the leg spring stores energy during the stance phase and causes the body mass to have approximately equal and opposite vertical velocities at liftoff and touchdown. To maintain the duration of flight, the control system length- l d l d l l Ground Contact Maximum Compression Detected Maximum Compression Fig. 10. Forcing the Foot into the Ground. In a normal step (top), energy is added into the leg spring at the moment of maximum compression. In the forced step (bottom), the loading on the leg is increased just after touchdown, forcing the foot into the ground and shortening the step duration. l d is the desired leg length. \Deltal is the change in desired leg length that returns the robot to the desired hopping height. Rebound Compress Detect Slip Set Front Leg Front Foot Reposition Detect Slip Rebound Compress Set Rear Leg Rear Foot Reposition Detect Slip Rebound Compress Triangle Stable Triangle Fig. 11. Repositioning Strategies. After a slip has been detected, the initial step is abandoned and one or both legs are repositioned for the next step. The leg angle at touchdown on the next step will be closer to vertical, keeping the impact force vector within the friction cone. ens the leg to add energy equivalent to that lost due to internal mechanical losses and to the impact of the un- sprung mass of the lower leg with the ground. In a normal step, thrust occurs at the moment of maximum compression of the spring (Figure 10). In responding to a slip, the control system may alter this sequence by extending the leg as soon as the slip is detected. If the leg is close to vertical, this extension increases the vertical component of the foot's force on the ground and may stop the slip. The extension also adds energy into the leg spring. The extra energy is removed later in the step by lengthening the leg spring when the leg is vertical, leaving the hopping height unchanged (Figure 10). One effect of this reaction is to slow the robot, a desirable 6effect when the surface is slippery. However, the foot forcing reflex may lead to a crash if the leg geometry and velocity is such that extending the leg increases the horizontal forces on the foot more than the vertical forces. Thus, the foot forcing reflex may not be sufficient in itself to recover from slips. The foot forcing reaction shortens the period of time during which the spring is passively compressed, leading to a shorter stance phase and a style of running that utilizes quick hops rather than long strides. We have observed that this quick-stepping behavior is a useful method for running briefly on slippery surfaces because the leg angle at touchdown is near vertical. However, the shorter stance phase also reduces the available time for correcting the body attitude and makes steady-state running difficult to achieve. B. Repositioning Strategies The step on which the initial slip occured may be abandoned by immediately lifting the foot; the resulting flight phase provides a brief opportunity to prepare for another landing on the slippery surface. By reconfiguring the legs during the flight phase following the initial slip, the control system can attempt to keep the foot forces within the friction cone. Because the coefficient of friction is not known, the size of the friction cone is unknown. Therefore, the best place for the foot at the next touchdown is directly under the body, making the leg vertical at touchdown. Figure 11 diagrams the strategies that reposition the legs. Figures 12, 13, and 14, contain sequences showing the repositioning strategies involved in recovering from a slip. After a slip has been detected, both legs may be used in the recovery by configuring them in a narrow fixed triangle vertically centered under the body. The control system attempts to hold this triangle throughout the subsequent step and does not apply the normal pitch, roll, and yaw adjustments. Instead, the robot bounces, letting the geometric configuration provide stability rather than using active control. The leg angles in normal running are nearly symmetric during the flight phase of steady-state running; the control system only has to equalize the leg lengths to create a symmetric triangle. Because the extent of the friction cone is unknown, the triangle is narrowed so the legs are close to vertical. When both feet contact the ground, foot forcing is applied to each to reduce the time of stance. After both feet have lifted off the ground, the control returns to a normal flight state. C. Slipping Results The slipping strategies were tested in simulation by varying the initial velocity of the robot and the coefficient of friction to produce multiple runs. For each trial, a circular slippery area was simulated at the location of the first footfall. During successful runs, the robot stepped once in the slippery area and then five additional times on a non- slippery surface. The initial velocity was 2.5 \Sigma 0.25 m/s. The size of the slippery area for each reaction strategy was large enough to prevent a foot from sliding to the edge, a situation that allowed an easy recovery. The slippery area was small enough that subsequent footfalls were located outside of it. Twenty friction coefficients between 0.025 and 0.5 were used. Both static and dynamic coefficients were set to the same value for each trial of 20 simulations with different initial velocities. The robot was judged able to recover from a slip at a given coefficient of friction if at least half of the trials were completed successfully. For the successful trials, we computed a measure of the error at touchdown of the step after the recovery step that followed the slip. The error measure was the summed absolute values of differences between the actual and desired angles for the body yaw, ff, pitch, fi, and roll, fl: The error calculation was designed to measure how well the slip recovery strategy had positioned the robot after the slip step, the recovery step, and the subsequent ballistic flight. The errors for the successful trials were averaged to compute the data shown in Figure 15. This graph illustrates the tradeoff between the two types of strategies. With no active reflexes, the controller is able to negotiate friction coefficients as low as 0.28. Upon contact, the foot slides; as it is loaded, the vertical and horizontal forces increase, pushing the foot back under the body. Eventually the forces on the foot reenter the friction cone, slipping ceases, and a normal step ensues. The foot forcing strategy causes the foot to slide further out from under the body, leading to fewer recoveries at lower coefficients of friction than the steady-state control system. We observed this effect for several running speeds and heights. However, it may be a consequence of the geometry of the robot design; foot forcing may be useful for slow moving robots or those with other gait patterns. The hip torque reflex succeeds at pulling the leg back and enables recoveries as low as 0.22. Note that hip torque does indeed increase the body pitch, producing increased errors shown in the graph. The repositioning strategies delay error correction while the legs are reconfigured. As a result, the repositioning strategies produce larger errors upon return to normal running than the foot forcing and hip torque reflexes. However, the repositioning strategies are able to recover from slips on surfaces with smaller coefficients of friction. By lifting the leg and repositioning it within the friction cone, the front and rear repositioning reflexes are able to recover from surfaces with coefficients as low as 0.07 and 0.15, respectively. The front repositioning strategy is more successful than the rear repositioning strategy because it more effectively reduces the relative speed of the foot over the ground before impact. Because the robot is moving forward while the foot is airborne, bringing the rear leg forward increases the relative speed between the foot and the ground. The front repositioning strategy brings the front leg back, reducing the relative speed. On impact, the foot with the lower relative speed is subjected to smaller horizontal forces and is less likely to slip. The robot experiences increased slipping as the coefficient of friction decreases, but it often recovers because Fig. 12. Front Leg Repositioning. The front leg is lifted and repositioned for a more vertical impact. [Friction coefficient: 0.20. Times Fig. 13. Rear Leg Repositioning. The rear leg is brought under the slipping robot to arrest the fall. The newly planted leg slips upon takeoff, but the step is successful because the body attitude is not disturbed significantly. The robot is able to continue running. [Friction coefficient: 0.20. Time (s): 0.0, Fig. 14. Stable Triangle Recovery. After detecting a slip, the robot forms a stable triangle. Although the legs slip just prior to liftoff, the control system is able to recover because the slip is symmetric and occurs at the end of the step. [Friction coefficient: 0.02. Times (s): Friction Coefficient0.10.30.50.70.9Error None Foot Force Hip Torque Front Reposition Rear Reposition Stable Triangle Fig. 15. Touchdown Errors. If the robot recovers from a slip, it starts the next step with some error. This graph illustrates the tradeoff between smooth running and slip recovery. Lower curves indicate smaller errors in body and leg angle. Longer curves indicate that a greater range of friction coefficients can be tolerated. the slips occur at the end of the recovery step. Figure 13 shows a normal ground contact and rebound followed by a slip upon takeoff. Because the hopping height, forward speed, and body attitude control algorithms have already been applied, the slip has little effect on the configuration of the robot. Figure 14 shows slipping upon takeoff for the stable triangle strategy, which applies no attitude correction during the recovery step. However, as Figure 14 shows, both legs slip symmetrically, cancelling the effect of their torque on the body. Thus, the stable triangle reflex is capable of recovering from surfaces with coefficients as low as 0.025. VI. Tripping For steady-state running, the control system detects expected events, such as foot contact or initial leg spring compression, and uses these signals to transition between control states. During each state, it applies the appropriate collection of control laws. Tripping occurs when the robot feet or legs encounter unexpected obstacles, causing the controller to execute inappropriate servo commands (Fig- ure 16). To explore reflexive responses to tripping, we considered the task of returning the robot to steady-state running after Fig. 16. Simulated Trip. The front foot contacts the vertical face of a box and slides down the surface. With no response, the robot is unable to continue running and crashes. [Times (s): 0.0, 0.05, 0.09, a collision with a box. The existing controller allowed the robot to continue running for some unexpected contacts. For example, foot contacts on the top surfaces of boxes, though premature in the flight phase, allowed a normal step to occur. Oblique contacts, such as brushing the side of the box, also did not usually prevent running from continuing. Other contacts, such as a foot or leg contacting the vertical face of a box, resulted in crashes. A. Tripping Responses As in the case of slipping, the sensing requirements were minimal. The controller detected only that a contact with a foot or leg had occurred. It did not detect where on the leg the contact had occurred. These conditions could be determined on a physical robot with contact sensors on the legs or via the existing joint angle sensors. When a leg or foot hits the front surface of a box, a foot must be repositioned to find a foothold on or beyond the box. If the forward foot hits the box, either the forward or the rear foot can be retracted and repositioned to contact the top surface of the box, where good footholds are avail- able. We call these strategies the "front lift" and "rear lift" reflexes, depending on which leg is lifted to the top surface of the box. If the rear leg hits a box, the leg can be pulled back, allowing it to pass over the box without contact. We refer to this strategy as "rear pull." These reflexes are diagrammed in Figure 17 and shown in Figures 18, 19 and 20. B. Tripping Results To test the tripping reactions, boxes of varying heights were placed in the path of a robot running in steady state. For the front lift and rear lift reflexes, the vertical face of each box was divided into 20 impact heights and the robot was released with the front foot 2 cm from the box at each height. For the rear pull reflex, the robot was placed straddling boxes of varying heights with the forward foot making an initial ground contact in a normal running step. As the box height increased, the rear leg eventually contacted the box as it swung forward. In all simulations, Compress Rebound Detect Trip Set Front Leg Front Lift Trip Response Rebound Detect Trip Compress Set Rear Leg Rear Lift Trip Response Rebound Detect Trip Compress Pull Rear Leg Rear Pull Trip Response Fig. 17. Trip Recovery Strategies. After a trip has been detected, one of the legs is repositioned in an attempt to contact the top surface of the obstacle or avoid it entirely. the initial forward speed of the robot was varied by a small random factor. With no reflex responses, the robot was unable to continue running following a trip. The front lift and rear lift response curves show that as the box heights increase, the tripping reflexes are less likely to produce a recovery (Fig- ures 21 and 22). The number of crashes increases as the box height increases. This increase in crashes is due to the increasing distances to the box top as the height increases. If the foot hits the box near the top, there may be sufficient time to lift it to the top of the box. However, as the box height increases, fewer potential contact points are near the top edge of the box. To measure the disturbance to normal running, we computed the same error measure as was used in the slipping trials. The error measure was the sum of the absolute values of the errors between actual and desired yaw, ff, pitch, fi, and roll, fl: The bottom graphs in Figures 21 and 22 show that if the robot is able to recover, it does so with approximately the same error independent of box height. The front lift reflex causes less touchdown error than does the rear lift reflex. To recover with the front foot, the foot must lift over the box edge, whereas a recovery with the rear foot must move the rear foot from its position behind the robot to the box. The rear lift reflex accumulates more errors during the additional flight time. With no reflex responses, the robot is unable to recover when the rear leg hits a box of any height. However, Figure 23 shows that pulling the leg back after the initial con- Fig. 18. Front Lift Trip Response. The front leg is lifted and repositioned to achieve a better foothold. [Times (s): 0.0, Fig. 19. Rear Lift Trip Response. The rear leg is lifted and repositioned to achieve a better foothold. [Time (s): 0.0, 0.07, 0.09, 0.11, Fig. 20. Rear Pull Trip Response. When a leg hits an obstacle while swinging forward, it is pulled back to allow it to clear the obstacle. [Times 0.25515Number of Crashes Front Lift Response Error Curves at Touchdown (rad.) Fig. 21. Front Lift Results. The top graph shows the number of crashes as the obstacle height increases. The bottom graph shows the average error in body attitude at the start of the next step after recovering from a trip. As the box height increases, trips more often lead to crashes. Note however, that the errors remain relatively constant for those trials where the robot is able to recover and continue running. There were 20 runs per box height. Box heights below 5 cm did not cause trips; box heights above 28:75 cm did not allow recovery. tact allows the robot to pass the leg over boxes as high as 23 cm without crashes. For boxes between 23 cm and cm, the leg, though pulled back, hits the box again, but may still be able to recover. Above 25 cm, the boxes are too high for the retracted leg to pass over, increasing the number of crashes. VII. Discussion and Conclusions We have considered the problem of creating reflexes for slipping and tripping given only the information that a slip or a trip has occurred. We evaluated two kinds of responses to slipping, one-step strategies and two-step strate- gies, depending on whether the correction was applied in the slip step or in the following step. Responses that continue the slipping step produce smoother recoveries but only for higher friction coefficients. Responses that abandon the slipping step are capable of negotiating surfaces with a larger range of friction coefficients but accumulate larger errors. Our slipping simulations focused on traversing a patch in which one footfall slipped; however, some observations can be made regarding running on a slippery surface. For higher coefficients of friction, the strategy with the smallest errors, the hip torque reaction, is most likely to succeed. The repositioning strategies are limited because continual 0.25515Number of Crashes Rear Lift Response Error Curves at Touchdown (rad.) Fig. 22. Rear Lift Results. Taller boxes are more likely to cause a crash. However, if the robot does recover, it does so with a relatively constant error. The rear lift reflex recovers about as often as the front lift reflex (Figure 21), but with higher resulting errors. There were runs per box height. 0.25515Number of Crashes Rear Pull Response Error Curves at Touchdown (rad.) Fig. 23. Rear Pull Results. Pulling the tripping foot back so it passes over the box allows the robot to continue running, but with some additional attitude error. For box heights below 13:75 cm, the rear foot passes over the box without tripping due to the retraction of the leg during running. There were 20 runs per box height with variation in the initial velocity of the robot. slipping would cause them to abandon every other step. However, all of the reflexive strategies except the hip torque strategy reduce the forward velocity during slip recovery, thus making the foot forces more vertical on subsequent steps. Preliminary results indicate that only a few slipping reactions may be required to achieve steady running on a slippery surface without slipping. If the foot is moving with respect to the ground at touch- down, the horizontal force on the ground is increased in the direction of motion, thereby increasing the danger of slip- ping. Strategies for running on slippery surfaces should try to reduce the relative motion of the foot between the ground prior to impact. This principle, commonly called ground-speed matching, is useful in slip prevention. It also reduces the impact of ground contact and is used by animals and human runners. We evaluated several reflexes that repositioned the foot after a trip to find a viable foothold or to avoid the box. For trips in which the forward foot struck the vertical face of the lifting either the front or rear foot allowed recoveries. However, lifting the front foot produced the smallest errors at the start of the subsequent step. For trips in which the rear leg hit the box, pulling the leg back to let it pass over the box allowed the robot to continue running, but with some additional error in body attitude. The slipping and tripping reflexes have been validated for single slip or trip tasks. The next task is to integrate the reflexes to enable running through general rough terrain with arbitrary obstacles and slippery areas. Additional controllers may be used to select among the applicable reflexes based on sensing or modeling of the environment. Finally, within the time constraints of the rapidly evolving dynamic system, limited replanning may be used to aid recovery. These slipping and tripping reflexes are robust despite their minimal sensing requirements. Without determining friction or obstacle properties, without modeling the surface, and without online planning, the reflexes enable the robot to continue running under many circumstances. Even if more sensing and computational resources are available for foot placement, surface modeling, and replanning, reflexes such as these will remain necessary due to sensing and modeling errors. Slipping and tripping reflexes are fundamental to many rough terrain problems. Slopes, uneven surfaces, and small obstacles create oblique impact angles that can cause slips and trips. Reflexive responses will facilitate the successful traversal of these terrains in combination with other reflexive strategies for foothold errors such as adhesions, bounces, and loss of firm footing. VIII. Acknowledgments This project was supported in part by NSF Grant No. IRI-9309189 and funding from the Advanced Research Projects Agency. --R Motor Control and Learning "The adaptive suspension vehicle" "Configuration of autonomous walkers for extreme terrain" "Terrain mapping for a walking planetary rover" "Perception, planning, and control for autonomous walking with the ambler planetary rover" "Optimal force distribution for the legs of a walking machine with friction cone con- 1straints" "Adjusting step length for rough terrain locomotion" "Realistic animation of legged running on rough ter- rain" "Adaptive gait control of a biped robot based on realtime sensing of the ground profile" "De- velopment of a dynamic biped walking system for humanoid: Development of a biped walking robot adapting to the humans' living floor" "Adapting reflexes controlling the human pos- ture" "Fixed patterns of rapid postural responses among leg muscles during stance" "Balance adjustments of humans perturbed while walking" "Stumbling correct reaction: A phase-dependent compensatory reaction during locomotion" "Phasic control of reflexes during locomotion in vertebrates" "Corrective responses to perturbation applied during walking in humans" "An adaptive artificial hand" "Robot control by reflex ac- tions" "Reflex control of the prototype leg during contact and slippage" "A study of design and control of a quadrupedwalking vehicle" "A robot that walks: Emergent behaviors from a carefully evolved network" "Robot impact control inspired by human reflex" "Control of a six-legged robot walking on abrupt terrain" "Developing planning and reactive control for a hexapodrobot" Legged Robots That Balance "Running experiments with a planar biped" --TR --CTR Christophe Sabourin , Olivier Bruneau , Gabriel Buche, Control Strategy for the Robust Dynamic Walk of a Biped Robot, International Journal of Robotics Research, v.25 n.9, p.843-860, September 2006 Tao Geng , Bernd Porr , Bernd Florentinwrgtter, A Reflexive Neural Network for Dynamic Biped Walking Control, Neural Computation, v.18 n.5, p.1156-1196, May 2006
rough terrain;tripping;slipping;reactive control;reflexes;biped locomotion
591464
Globally Consistent Range Scan Alignment for Environment Mapping.
A robot exploring an unknown environment may need to build a world model from sensor measurements. In order to integrate all the frames of sensor data, it is essential to align the data properly. An incremental approach has been typically used in the past, in which each local frame of data is aligned to a cumulative global model, and then merged to the model. Because different parts of the model are updated independently while there are errors in the registration, such an approach may result in an inconsistent model.In this paper, we study the problem of consistent registration of multiple frames of measurements (range scans), together with the related issues of representation and manipulation of spatial uncertainties. Our approach is to maintain all the local frames of data as well as the relative spatial relationships between local frames. These spatial relationships are modeled as random variables and are derived from matching pairwise scans or from odometry. Then we formulate a procedure based on the maximum likelihood criterion to optimally combine all the spatial relations. Consistency is achieved by using all the spatial relations as constraints to solve for the data frame poses simultaneously. Experiments with both simulated and real data will be presented.
Introduction 1.1 Problem Definition The general problem we want to solve is to let a mobile robot explore an unknown environment using range sensing and build a map of the environment from sensor data. In this paper, we address the issue of consistent alignment of data frames so that they can be integrated to form a world model. However, the issue of building a high-level model from registered sensor data is beyond the scope of this paper. A horizontal range scan is a collection of range measurements taken from a single robot position. In previous robot navigation systems, range scans have often been used for robot self-localization in known environments [3]. Using range measurements (sonar or laser) for modeling an unknown environment has also been studied in the past [11, 4, 8]. A range scan represents a partial view of the world. By merging many such scans taken at different locations, a more complete description of the world can be obtained. Figure 1 gives an example of a single range scan and a world model consisting of many scans. a b Figure 1: Building world model from range scans. (a) One range scan in a simulated world; (b) model consisting of many scans. The small circles show the poses at which the scans are taken. The essential issue here is to align the scans properly so that they can be merged. But the difficulty is that odometry information alone is usually inadequate for determining the relative scan poses (because of odometry errors that accumulate). On the other hand, we are unable to use pre-mapped external landmarks to correct pose errors because the environment is unknown. A generally employed approach of building a world model is to incrementally integrate new data to the model. When each frame of sensor data is obtained, it is aligned to a previous frame or to a cumulative global model. Then the new frame of data is integrated into the global model by averaging the data or using a Kalman filter [1, 10, 11, 4, 8]. A major problem with this approach is that the resulting world model may eventually become inconsistent as different parts of the model are updated independently. Moreover, it may be difficult to resolve such inconsistency if the data frames have already been permanently integrated. To be able to resolve inconsistency once it is detected at a later stage, we need to maintain the local frames of data together with their estimated poses. In addition, we need a systematic method to propagate pose corrections to all related frames. Consider an example as shown in Fig. 2(a). The robot starts at P 1 and returns to a nearby location P n after visiting along the way. By registering the scan taken at P n against scan P n\Gamma1 , the pose of P n can be estimated. However since P n is close to P 1 , it is also possible to derive pose P n based on P 1 by matching these two scans. Because of errors, the two estimates of could be conflicting. If a weighted average of the two is used as the estimate of P n , the pose of P should also be updated as otherwise the relation P will be inconsistent with its previous estimate. This inconsistency could be significant if the looped path is long. Similarly, other poses along the path should also be updated. In general, the result of matching pairwise scans is a complex, and possibly conflicting, network of pose relations. We need a uniform framework to integrate all these relations and resolve the conflicts. In this paper, we present such a framework for consistently registering multiple range scans. The idea of our approach is to maintain all the local frames of data as well as a network of spatial relations among the local frames. Here each local frame is defined as the collection of sensor data measured from a single robot pose. The robot pose, in some global reference frame, is also used to define the local coordinate system of the data frame. Spatial relations between local frames are derived from matching pairs of scans or from odometry measurements. We treat the history of robot poses in a global coordinate system (which define all the local frame positions) as variables. Our goal is to estimate all these pose variables using the network of constraints, and register the scans based on the solved poses. Consistency among the local frames is ensured as all the spatial relations are taken into account simultaneously. Figure 2 shows an example of consistently aligning a set of simulated scans. Part (a) shows the original scans badly misaligned due to accumulated pose errors. Part (b) shows the result of aligning these scans based on a network of relative pose constraints (with edges indicated by line segments). a Pn Figure 2: An example of consistently aligning a set of simulated scans. (a) Original scans badly misaligned due to accumulated pose errors; (b) the result of aligning these scans based on a network of relative pose constraints. The constraints are indicated by line segments connecting pairs of poses. Two types of constraints are used: those derived from aligning a pair of scans (marked by both solid and dotted lines), and those from odometry measurements (marked by solid lines). 1.2 Related Work The first project that systematically studied the consistency issue in dynamic world modeling is the HILARE project [2]. In this system, range signals are segmented into objects which are associated with local object frames. Each local frame is referenced in an absolute global frame along with the uncertainty on the robot's pose at which the object frame is constructed. New sensor data are matched to the current model of individual object frames. If some object which has been discovered earlier is observed again, its object frame pose is updated (by averaging). In circumstances that the uncertainty of some object frame is less than the uncertainty of the current robot pose, as it happens when the object frame is created earlier, and later the robot sees the object again, the robot's pose may be corrected with respect to that object frame. After correcting the current robot pose, the correction is propagated backwards with a "fading" effect to correct the previous poses. Although the HILARE system addressed the issue of resolving model inconsistency, its solution has the following potential problems. First of all, the system associates local frames with "objects". But if the results of segmenting sensor data or matching the data with model are imperfect, the "objects" and therefore the local frames may not be defined or maintained consistently. When a previously recorded object is detected again, the system only attempts to update the poses (and the associated frames) along the path between the two instances of detecting this object, while the global consistency among all frames in the model may not be maintained. HILARE uses a scalar random variable to represent the uncertainty of a three-degree-of-freedom pose, therefore it can not model the confidences in the individual pose components. Moutarlier and Chatila presented a theoretical framework for fusing uncertain measurements for environment modeling [14]. They first discussed two types of representations: relation-based and location-based. In relation-based representation, an object is related to another by the uncertain transform between their reference frames. A network of relationships is maintained as the world model. When new observations are made, all the relationships need to be updated to preserve consistency. In location-based representation, the global references of individual object frames are maintained together with their uncertainties. When objects are re-observed, these object frames and other related frames are updated with respect to the global reference frame. After comparing these two approaches, Moutarlier and Chatila choose to use the location-based approach. They treat the object and robot locations as state variables and maintain all the object variance/covariance matrices as state information. A stochastic-based formulation for fusing new measurements and updating the state variables is introduced. In addition to a global updating approach, they also introduced a relocation-fusion approach which first updates the robot position based on the new observations and then updates the object frames. The relocation-fusion approach reduces the influence of sensor bias in the estimation, although the algorithm is suboptimal. In a series of work by Durrant-Whyte [5, 6, 7], the problem of maintaining consistency in a network of spatial relations was studied thoroughly. In their formulation, the environment model is represented by a set of spatial relations between objects. A probabilistic fusion algorithm similar to the Kalman filter is employed to integrate new measurements to the a priori model. When some relations are updated as a result of new observations, the consistency among all relations are enforced by using explicit constraints on the loops of the network. The updating procedure is formulated as constrained optimization and it allows new observations to be propagated through the network while consistency between prior constraints and observed information is maintained. In another similar approach, Tang and Lee [17] formulated a geometric feature relation graph for consistent sensor data fusion. They proposed a two-step procedure for resolving inconsistency in a network of measurements of relations. In the first step, a compromise between the conflicting measurements of relations is achieved by the fusion of these measurements. Then in the second step, corrections are propagated to other relations in the network. The difficulty in maintaining model consistency in a relation-based representation is that the relations are not independent variables. Therefore additional constraints are needed in formulating an updating procedure. The constrained optimization approach seems very complicated and difficult to apply in practice. In view of the previous methods, we present a new approach which has the following distinctive characteristics: 1. We use an unambiguous definition of an object frame as the collection of sensor measurements observed from a single robot position. Thus we avoid the difficult task of segmenting and recognizing objects (which the previous methods rely on in order to create and update object frames). It is also important to note that we use a robot pose to define the reference for an object frame. In a local frame, the relative object positions with respect to the robot pose are fixed (whose uncertainty is no more than bounded sensing errors). During the estimation process, when the robot position in the global reference frame is updated, effectively the global coordinates of all objects in the current frame are updated accordingly. Therefore by maintaining the history of robot poses, we also maintain the spatial relationships among the object frames. 2. Our approach uses a combination of relation-based and location-based representations. We treat relations as primitives, but treat locations as free variables. This is different from the pure relation-based approach in that we do not directly update the existing relations in the network when new observations are made. We simply add new relations to the network. All the relations are used as constraints to solve for the location variables which, in turn, define a set of updated and consistent relations. On the other hand, our approach is different from the location-base approach by Moutarlier and Chatila [14] in that we do not assume the covariance matrices between the object frames as known. Our state information is the entire set of raw relations. We derive the covariance matrices at the same time as we solve for the position variables. 3. We obtain direct spatial relations between object frames. Because our object frames are tied to robot poses, odometry measurements directly provide spatial relations between the frames. More importantly, we may align two overlapping frames of data (in our case range scans) to derive more accurate relations between frames. In previous approaches, the robot typically relies on odometry to first determine its new pose. Then the detection of objects allows the robot pose as well as the object locations to be updated. Since the relations between object frames are updated rather indirectly through the robot pose, biases in odometry measurements may lead to divergence in the estimation of object positions, as reported in [14]. Moutarlier and Chatila propose an algorithm that is supposed to address the divergence problem at the expense of a sub-optimal solution. Our formulation does not have this problem, as we obtain direct spatial relations between object frames by aligning the data, and therefore we are less sensitive to odometry biases. 2 Overview of Approach We formulate our approach to multiple scan registration as one of estimating the global poses of the scans, by using all the pose relations as constraints. Here the scan poses are considered as variables. A pose relation is an estimated spatial relation between the two poses which can be derived from matching two range scans. We also obtain pose relations from odometry mea- surements. Finally, we estimate all the poses by solving an optimization problem. The issues involved in this approach are discussed in the following subsections. 2.1 Deriving Pose Relations Since we use a robot pose to define the local coordinate system of a scan, pose relations between scans can be directly obtained from odometry which measures the relative movement of the robot. In section 4.2, we will discuss the representation of odometry pose constraint and its uncertainty. More accurate relations between scan poses are derived from aligning pairwise scans of points. Here any pairwise scan matching algorithm can be used. One possible choice is the extension to Cox's algorithm [3] where line segments are first fit to one scan and then points in another scan are matched to the derived line segments. In our previous studies, we proposed another scan matching algorithm which is based on direct point to point matching [12, 13]. In either case, the scan matching algorithm takes two scans and a rough initial estimate of their relative pose (for example from odometry information) as input. The output is a much improved estimate of the relative pose. After aligning two scans, we can record a set of corresponding points on the two scans. This correspondence set will form a constraint between the two poses. In section 4.3, we will formulate this type of constraint and its uncertainty as used in the estimation algorithm. When we match two scans, we first project one scan to the local coordinate of the other scan, and discard the points which are likely not visible from the second pose. The amount of overlap between two scans is estimated empirically from the spatial extent of the matching parts between the two scans. A pose relation is only derived when the overlap is significant enough (larger than a given threshold). 2.2 Constructing a Network of Pose Relations Given the pairwise pose relations, we can form a network. Formally, the network of constraints is defined as a set of nodes and a set of links between pairs of nodes. A node of the network is a pose of the robot on its trajectory at which a range scan is taken. Here a pose is defined as a three dimensional vector (x; consisting of a 2D robot position and the home orientation of the rotating range sensor. We then define two types of links between a pair of nodes. First, if two poses are adjacent along the robot path, we say that there is a weak link between the two nodes which is the odometry measurement of the relative pose. Second, if the range scans taken at two poses have a sufficient overlap, we say that there is a strong link between the two nodes. To decide whether there is sufficient overlap between scans, we use an empirical measure. The spatial extent in the overlapping part of two scans should be larger than a fixed percentage of the spatial extent covered by both scans. For each strong link, a constraint on the relative pose is determined by the set of corresponding points on the two scans given by the matching algorithm. It is possible to have multiple links between two nodes. Figure 3 shows an environment and the constructed network of pose relations. 2.3 Combining Pose Relations in a Network The pose relations in a network can be potentially inconsistent because they are not independent variables (the number of relations may be more than the degrees of freedom in the network), while the individually estimated relations are prone to errors. Our task is to combine all the pose relations and resolve any inconsistency. This problem is formulated as one of optimally Figure 3: Example of constructing a network of pose relations from matching pairwise scans. (a) A simulated environment where the scan poses are labeled by circles; (b) the network of pose relations constructed from matching overlapping scans. estimating the global poses of nodes in the network. We do not deal with the relations directly. Rather, we first solve for the nodes which constitute a set of free variables. Then a consistent set of relations which represents a compromise of all a priori relations is defined by the poses on the nodes. An optimization problem is defined as follows. We construct an objective function from the network with all the pose coordinates as variables (except one pose which defines our reference coordinate system). Every link in the network is translated into a term in the objective function which can be conceived as a spring connecting two nodes. The spring achieves minimum energy when the relative pose between the two nodes equals the measured value (either from matching two scans or from odometry). Then the objective function represents the total energy in the network. We finally solve for all the pose variables at once by minimizing this total energy function. 2.4 The Three-Node Example Using the 3-node example, we illustrate the difference of our formulation from previous approaches Assume that the network consists of three nodes: relations . When there is new measurement - the algorithm by Durrant-Whyte [6] updates the three relations to T 0 3 based on an optimization criterion which is subject to the constraint T 0 In our approach, we pool together all the relations T T 1 to form an optimization problem and solve for a new estimate for the nodes: P 0 3 . These node positions define a consistent set of relations: T 0 1 . Note that the node positions are so we do not need to solve a complex constrained system. Moutarlier and Chatila [14] also treat the node positions as variables when updating the network with new measurements. But they assume the knowledge of covariance matrices among the a priori estimates of However, we only require the variances of individual measurement errors on the relations T are directly available from sensor models. The rest of the paper is organized as follows. In section 3, we present the optimization criterion by considering a generic optimal estimation problem. We derive a closed-form solution in a linear special case. In section 4, we formulate the pose relations as well as the objective function in the context of range scan registration. The closed-form solution derived in section 3 is applied to solve for the scan poses. In section 5, we present experimental results. Optimal Estimation from a Network of Relations In this section, we formulate a generic optimal estimation algorithm which combines a set of relations in a network. This algorithm will later be applied in section 4 in the context of robot pose estimation and scan data registration. 3.1 Definition of the Estimation Problem We consider the following generic optimal estimation problem. Assume that we are given a net-work of uncertain measurements about n+1 nodes X Here each node X i represents a d-dimensional position vector. A link D ij between two nodes X i and X j represents a measurable difference of the two positions. Generally, D ij is a (possibly nonlinear) function of X i and and we refer to this function as the measurement equation. Especially interesting to us is the simple linear case where the measurement equation is We model an observation of D ij as - a random Gaussian error with zero mean and known covariance matrix C ij . Given a set of measurements - pairs of nodes and the covariance C ij , our goal is to derive the optimal estimate of the position by combining all the measurements. Moreover, we want to derive the covariance matrices of the estimated X i 's based on the covariance matrices of the measurements. Our criterion of optimal estimation is based on the maximum likelihood or minimum variance concept. The node position X i 's (and hence the position difference D ij 's) are determined in such a way that the conditional joint probability of the derived D ij 's, given their observations ij 's, is maximized. If we assume that all the observation errors are Gaussian and mutually independent, the criterion is equivalent to minimizing the following Mahalanobis distance (where the summation is over all the given measurements): (D Even if the observation errors are not independent, a similar distance function can still be formed. However, it will involve the correlation matrices of the measurements. The assumption of independence is actually not necessary in our formulation. The assumption makes practical sense as the covariances of errors are difficult to estimate. A typical application of the optimal estimation problem is in mobile robot navigation, where we want to estimate the robot pose and its uncertainty in three degrees of freedom (x; '). The observations are the relative robot poses from odometry, and also possible from matching sensor measurements. We want to utilize all the available measurements to derive the optimal estimate of the robot poses. Note that in this application, the measurement equation is non-linear because of the ' component in the robot pose. Our approach above differs from the one typically used within a Kalman filter formulation, in which only the current pose is estimated, while the history of previous poses and associated measurements is collapsed into the current state of the Kalman filter. Our objective, however, is not simply getting from A to B safely and accurately, but also building a map of the environment. It is, therefore, meaningful to use all the measurements obtained so far in the mapping process. The old poses themselves are not particularly useful. But we are using the poses to define local object frames. Thus maintaining the history of robot poses is equivalent to maintaining the structure of the environment model. The advantage of using a pose to define a data frame is that it is unambiguous and it avoids the difficult segmentation and object identification problem present in other work. Next, we study the case when the measurement equation is linear and we derive closed-form solutions for the optimal estimates of the nodes and their covariances. Later, we will solve the non-linear robot pose estimation problem by approximately forming linear measurement equations. 3.2 Solution of Optimal Linear Estimation We consider the special case where the measurement equation has the simple linear are the nodes in the network which are d-dimensional vectors and the D ij 's are the links of the network. Without loss of generality, we assume that there is a link D ij between every pair of nodes . For each D ij , we have an observation - which is assumed to have Gaussian distribution with mean value D ij and known covariance C ij . In case the actual link D ij is missing, we can simply let the corresponding C ij be 0. Then the criterion for the optimal estimation is to minimize the following Mahalanobis distance: 0-i!j-n Note that W is a function of all the position X i 's. Since we can only solve for relative positions given the relative measurements, we choose one node X 0 as a reference and consider its coordinate as constant. Without loss of generality, we let X relative positions from X 0 . We can express the measurement equations in a matrix form as where X is the nd-dimensional vector which is the concatenation of is the concatenation of all the position differences of the form D H is the incidence matrix with all entries being 1, \Gamma1, or 0. Then the function W can be represented in matrix form as: D is the concatenation of all the observations - ij for the corresponding D ij and C is the covariance of - D which is a square matrix consists of C ij 's as sub-matrices. Then the solution for X which minimizes W is given by The covariance of X is If the measurement errors are independent, C will be block-diagonal and the solution can be simplified. Denote the nd \Theta nd matrix H t C \Gamma1 H by G and expand the matrix multiplications. We can obtain the d \Theta d sub-matrices of G as Denote the nd-dimensional vector H t C D by B. Its d-dimensional sub-vectors are the following (let - Then the position estimates and covariance can be written as The above algorithm requires to be invertible. If the network is fully connected and the individual error covariances are normally behaved, we believe it is possible to prove that G is invertible. Note the dimension of G (number of free nodes) is less than or equal to the dimension of C (number of edges) in a fully connected graph. 3.3 Special Networks (b) (a) Figure 4: (a) Serial connection; (b) parallel connection. We will apply the formula in Eq. 9 to two interesting special cases as in Figure 4. First, if the network consists of two serially connected links, D 01 and D 12 , the derived estimate of X 2 and its covariance matrix are Another case to consider is the network which consists of two parallel links D 0 and D 00 between two nodes X 0 and X 1 . If the covariance of the two links are C 0 and C 00 , the estimate of X 1 and its covariance are given by The solution is equivalent to the Kalman filter formulation. The above two cases correspond to the compounding and merging operations given by Smith and Cheeseman [16], which are used to reduce a complex network to a single relation. Smith and Cheeseman's algorithm has a limitation that it only applies to networks formed by serial and parallel connections. Figure 5: A Wheatstone bridge network. Consider the network in the form of a Wheatstone bridge (Fig. 5). The estimate of X 3 can not be obtained through compounding and merging operations. Therefore, the method by Smith and Cheeseman can not be directly applied to simplify this network 1 , while in our method, the variables can be solved from the linear system G =B @ \GammaC \Gamma1- The covariance matrix for the estimated position X 3 has a nice symmetric form (derived by expanding G \Gamma1 1 It is possible to first convert a triangle in the network to an equivalent Y-shaped connection and then the network becomes one with serial and parallel links. However, this Delta-to-Y conversion still can not turn every network into a combination of serial and parallel connections. 4 Derivation of Pose Relations In this section, we will apply the optimal estimation algorithm, as derived in section 3, to the robot pose estimation and scan data registration problem. To do this, we need to derive linearized measurement equations for the pose relations. In the following subsections, we study a constraint on pose difference given by matched scans or odometry measurements. For each constraint, we formulate a term in the form of Mahalanobis distance. For convenience in discussions of pose measurements, we will first define a pose compounding operation. 4.1 Pose Compounding Operation Assume that the robot starts at a pose its pose by ending up at a new pose V a = (x a ; y a ; ' a ) t . Then we say that pose V a is the compounding of V b and D. We denote it as: The coordinates of the poses are related by: y This is the same compounding operation as defined by Smith and Cheeseman [16]. If we consider that an absolute pose defines a coordinate system (the xy coordinates of the origin and the direction of one axis), and a relative pose defines a change of coordinate system (a translation followed by a rotation), then the compounding operation gives the pose which defines the new coordinate system after the transformation. The compounding operation is not commutative, but it is associative. We can thus define the compounding of a series of poses. It is also useful to define the inverse of compounding which takes two poses and gives the relative pose: The coordinates are related by the following equations: If D ab is the relative pose V a \Psi V b , the reversed relative pose D a can be obtained from D ab by a unary operation: We can verify that (\PsiD) \Phi We also want to define a compounding operation between a full 3D pose 2D position vector . The result is another 2D vector u We still denote the operation as The coordinates for u 0 are given by the first two equations of the full 3D pose compounding (Eq. 18,19). This 2D compounding operation is useful for transforming an non-oriented point (typically from a range sensor) from its local sensor coordinate system to the global coordinate system. 4.2 Pose Relations from Matched Scans Let V a and V b be two nodes in the network and assume there is a strong link connecting the two poses. From the pairwise scan matching algorithm, we get a set of pairs of corresponding points: u a where the 2D non-oriented points u a k are from scan S a and S b , respectively. Each pair (u a corresponds to the same physical point in the robot's environment while they are represented in different local coordinate systems. If we ignore any sensing or matching errors, two corresponding points are related by: If we take the random observation errors into account, we can regard \DeltaZ k as a random variable with zero mean and some unknown covariance C Z k . From the correspondence pairs, we can form a constraint on the pose difference by minimizing the following distance function: F ab (V a k(V a \Phi u a If we notice that a pose change is a rigid transformation under which the squared Euclidean distance k \Delta k 2 is invariant, we can rewrite the function in an equivalent form: F ab (V a k((V a \Psi V b ) \Phi u a Thus F ab is a function of D . The solution of D 0 which minimizes F ab can be derived in closed-form (see [12]). The relation D is the measurement equation. In order to reduce F ab into the Mahalanobis distance form, we linearize each term \DeltaZ k . Let close estimates of V a and V b . Denote \DeltaV and \DeltaV (the global coordinates of a pair of matching points). Then for small \DeltaV a and \DeltaV b , we can derive from Taylor expansion: V a \Phi u a \DeltaV a \Gamma \DeltaV b V a \Phi u a H a \DeltaV a \Gamma - where H a =B @ y a We can rewrite Eq. where V a \Phi u a H a \DeltaV a \Gamma - Thus we can now regard D in Eq. 35 as the pose difference measurement equation to replace . For the m correspondence pairs, we can form m equations as in Eq. 32. If we concatenate the - Z k 's to form a 2m \Theta 1 vector Z, and stack the M k 's to form a 2m \Theta 3 matrix M, then F ab can be rewritten as a quadratic function of D: F ab We can then solve for the D which minimizes F ab as The criterion of minimizing F ab (D) constitutes a least-squares linear regression. In Eq. 32, M k is known and - Z k is observed with an error \DeltaZ k having zero mean and unknown covariance C Z k . If we assume that all the errors are independent variables having the same Gaussian distribution, and further assume that the error covariance matrices have the form: then the least squares solution - D has the Gaussian distribution whose mean value is the true underlying value and whose estimated covariance matrix is given by is the unbiased estimate of oe D) D) Moreover, we notice that Eq. 37 can be rewritten as F ab (D) - We can define the energy term W ab corresponding to the pose relation which is equivalent to a Mahalanobis distance: where is the estimated covariance of - D. Note that D (as given in Eq. 35) is the linearized pose difference measurement equation. In deriving the covariance matrix CD , we made assumptions that the matrix is diagonal and the individual components of errors are zero mean Gaussian. It is probably difficult to justify these assumption. However, we believe that they are reasonable ones in practice. If any other estimates of the covariance matrices are available, they can certainly also be incorporated in our global estimation formulation. 4.3 Pose Relations from Odometry We also form an energy term in the objective function for each weak link. Suppose odometry gives a measurement - D 0 of the relative pose D 0 as the robot travels from pose V b to pose V a . The measurement equation is: We define the energy term in the objective function as follows: where C 0 is the covariance of the odometry error in the measurement - The covariance of measurement error is estimated as follows. Consider that a cycle of pose change consists of: (1) the robot platform rotation by an angle ff to face towards the new target position; (2) the robot translation by a distance L to arrive at the new position; (3) the sensor rotation by a total cumulative angle fi (usually 360 ffi ) to take a scan of measurements while the platform is stationary. We model the deviations oe ff , oe L , oe fi , of the errors in the variables ff, L, and fi as proportional to their corresponding values, while the constant ratios are determined empirically. The 3D pose change D derived as: Then the covariance C 0 of the pose change D 0 can be approximated as: fiC A J t (48) where J is the Jacobian matrix consisting of the partial derivatives of (x; with respect to \GammaL sin ff cos ff 0 We would also like to linearize and transform the measurement equation of D 0 to make the pose difference representation for odometry measurements consistent with that for matched sensing data. Consider the observation error \DeltaD of odometry. Let - close estimates of V a and V b . Denote \DeltaV Then through Taylor expansion, the observation error \DeltaD 0 becomes: V a \Psi - b (\DeltaV a \Gamma - H ab \DeltaV b ) (52) where sin - H ab =B @ Notice that - a H a and - H b are defined in Eq. 31. If we define a new observation error H a then we can rewrite Eq. 52 as H a \DeltaV a \Gamma - where we denote H a V a \Psi - H a \DeltaV a \Gamma - Notice that now we are dealing with the measurement equation for D which is consistent with that for matched sensing data. - D can be considered as an observation of D. The covariance C of - D can be computed from the covariance C 0 of - as: H a The energy term in the objective function now becomes: ab - 4.4 Optimal Pose Estimation Once we have uniformly formulated the two types of measurements, we can apply the estimate algorithm in section 3 to solve for the pose variables. Denote the robot poses as The total energy function from all the measurements is : is the linearized pose difference between V j and and - ij is an observation of D ij ( - is derived from the true observations, either range data or odometry measurements). The covariance C ij is also known. By regarding X as the state vector corresponding to a node of the network as in Section 3.2, we can directly apply the closed-form linear solution to solve for the X i 's as well as their covariance C X . The formulas are in Eq. 5 to Eq. 9. Then the pose V i and its covariance C i can be updated as: Note that the pose estimate V i and the covariance C i is given based on the assumption that the reference pose is non-zero, the solution should be transformed to where sin 4.5 Sequential Estimation The estimation algorithm we previously discussed is a one-step procedure which solves for all the pose variables at the same time. The algorithm is to be applied only after collecting all the measurements. Yet it will be more practically useful if we have a sequential algorithm which continuously provides estimates about the current or past pose variables after getting each new measurement. Here we will describe such a sequential procedure. Our sequential algorithm maintains the current best estimate about the poses of previously visited places. At each step, a new location is visited and measurements about the new location as well as the previous locations are gathered. By using these new measurements, the current pose can be estimated while the previous poses can be updated. be the pose vectors which we have previously estimated and let X n be the current new pose which we are about to measure. Let X represent the concatenation of X . Assume that we currently have an estimate X 0 of X whose inverse covariance matrix is C Because we have no knowledge about X n yet, the X n component in X 0 contains an arbitrary value and the matrix C has all zeros in the last d rows and d columns, where consider the addition of a set of new measurements relating X n to some of the past pose vari- ables. Let the measurement equation, in matrix form, be is a constant matrix). Assume that the set of measurements is - D which is an unbiased observation of D whose error has Gaussian distribution with covariance matrix CD . The updated estimate of X after using the new measurements is the one which minimizes the following function, using the maximum likelihood criterion, and assuming independent errors: The solution can be derived as D) (65) and the new covariance of X is A convenient way of updating X and CX is to maintain a matrix (the summation is over different sets of measurements). Then at each step, the updating algorithm is the following: First increase the dimensions of G and B to include the new pose X n . Update G and B as Then the new X and CX are given by One potential problem with the above sequential updating procedure is that the state variable expanding as it is augmented by a new state at each step. In case the robot path is very long, the variable size may become too large, causing storage or performance problems. A possible solution is to delete some of the old variables while adding the new ones. We propose a strategy of reducing the number of state variables as follows. In order to choose a pose to be deleted, we check all pairs of poses and find a pair (X the correlation between the two poses is the strongest. We then force the relative pose between X i and X j to be fixed as a constant. Then X i can be deleted from the state variables as it can be obtained from When deleting the node X i from the network, we transform any link (X link from X j to X k . Note that the covariance matrix CX contains all the pairwise covariance between any two poses. A correlation ratio between two poses can be computed from the covariance and variance. By only fixing some relative poses, the basic structure in the network is still maintained. Thus we are still able to consistently update all the pose variables once given new measurements. This strategy is more flexible than the simple strategy of fixing selected absolute poses as constants. Another approach to reducing the size of the system is to decompose the large network into smaller components. The estimation algorithm is to be applied to each sub-network. The relative pose between two nodes in different sub-networks can be obtained through pose compounding. If there is a single link connecting two parts of a network, the poses in two parts can be estimated separately and then combined with compounding, without loss of information. If, however, the network is strongly connected that there are two or more links between any two nodes, then a decomposition could give a sub-optimal estimation. 5 Implementation and Experiments 5.1 Implementation of Estimation Procedure The implementation of the estimation algorithm is as follows. After building the network, we obtain the initial pose estimates - by compounding the odometry measurements. Then for each link, we compute a measurement vector - ij and a covariance matrix C ij according to Eq. 38, 44 or Eq. 55, 57. Finally, we form a large linear system explained in Section 3.2 and solve for the pose variables X. The components needed to build G and B are C \Gamma1 ij . In the case of a strong link (from matching a pair of scans), these components can be readily computed as C \Gamma1 which can be expanded into simple summations by noting the regularity in the matrix M. In the case of a weak link (from odometry), these components can be computed by multiplications of small matrices (3 \Theta 3). The most expensive operation in the estimation process is to compute the inverse of a 3n \Theta 3n matrix G which gives the covariance of X. The network is stored as a list of links and a list of nodes. Each link contains the following information: type of link, labels of the two nodes, the computed measurement (relative pose), and the covariance matrix of the measurement. Each node contains a range scan. Note that we made linear approximations in the measurement equations in formulating the optimization criterion. The first order approximation error is proportional to the error in the initial pose estimate. Clearly, if we employ the newly derived pose estimate to formulate the linear algorithm again, a even more accurate pose estimate can be obtained. The iterative strategy based on this observation converges very fast. Typically, the first iteration corrects over 90% of the total pose error correctable by iterating the process. It usually takes four or five iterations to converge to the limit of machine accuracy. 5.2 Experiments with Simulated and Real Scan Data We now present experiments of applying our algorithm to register simulated and real range scan data. We first show an example using a simulated environment and measurements. This is useful because ground truth is known. Then an example using real data is presented. In the first example, we simulate a rectangular environment with a width of 10 units. The robot travels around a central object and forms a loop in the path. There are 13 poses along the path at a Figure Global registration of multiple scans using simulated scan data. (a) scans recorded in a global coordinate system where the pose of each scan is obtained from compounding odometry measurements. The scans align poorly because of accumulation of odometry error. (b) the result of correcting pose errors. Both the dashed lines and solid lines show the constraints from matching scan pairs. The solid lines also give the robot path and odometry constraints. which simulated range scans are generated (with random measurement errors). We also simulate a random odometry error (which is the difference between a pose change the robot thinks it made and the actual pose change) at each leg of the trajectory. The magnitude of the accumulated odometry error is typically in the range of 0.5 units. We apply our iterative global pose estimation algorithm to correct the pose errors. In Fig. 6(a), we show all the scans recorded in the initial coordinate system where the pose of each scan is obtained by compounding odometry measurements. Due to the accumulation of odometry error the scan data are aligned poorly. In Fig. 6(b), we show the result of correcting the pose errors and realigning the scan data. Each line segment (either dashed or solid) in the figure represents a strong link obtained from matching two scans. In addition, the solid lines show the robot path which corresponds to the weak links. A plot of orientational and positional errors of the poses along the path, both before and after the correction, is given in Fig. 7. Pose errors are accumulated along the path while the corrected pose errors are bounded. For comparison, we also apply a local registration procedure which matches one scan only to the previous scan. The pose errors along the path after this local correction are also shown in Fig. 7. Although pose errors are also significantly reduced after local corrections, they can still potentially grow without bound. In this example, global registration produces more accurate results than local correction.0.010.030.050.070.090 2 4 Radian Pose Number Magnitude of Orientational Errors along the Path Before correction After local correction After global correction0.10.30.50 Unit Pose Number Magnitude of Positional Errors along the Path Before correction After local correction After global correction a b Figure 7: Pose errors along the path, before correction, after local correction, and after global correction. (a) Orientational errors; (b) positional errors. Then we present the experiment using real range scans and odometry data. The testing environment is the cafeteria and nearby corridor in FAW, Ulm, Germany. The robot travels through the environment following a given path. A sequence of 30 scans which were taken by the robot with an interval of about 2 meters between scan poses were obtained. The laser sensor is a Ladar 2D IBEO Lasertechnik which is mounted on the AMOS robot. This laser sensor has a maximum Figure 8: Consistent global registration of real range scans which are collected by a robot at FAW, Ulm, Germany. (a) unregistered scans whose poses are subject to large odometry errors. (b) registered scans after correcting the pose errors. The robot path estimated from odometry is shown in dashed lines. The corrected path is shown in solid lines. Figure 9: Mapping of a Hallway using the RWI Pioneer platform and a SICK laser range scanner (a) Raw laser range scans (b) Aligned laser range scans. viewing angle of 220 degrees. Thus having only the 2D positions of two poses close together does not necessarily ensure a sufficient overlap between the scans taken at the two poses; we also need the sensor heading directions to be similar. Among the links from matching overlapping scan pairs are constructed. Some of these pairwise scan matching results have been shown in [12]. In Fig. 8, we show (a) the unregistered scans and (b) the globally registered scans in part (b). Further experimental results with a variant of our algorithm are reported in [9]. Figure 9 contains experimental results which are obtained using our global registration procedure together with a modified version of Cox's pairwise scan matching algorithm 2 . The laser data are collected on the RWI Pioneer platform using the SICK laser ranging device [15]. The Pioneer is a low-cost platform with odometry error significantly higher than the much more expensive platforms used in our other experiments. The hallway environment shown in Figure 9 is poor in features that allow localization of the robot along the hallway. The data was collected by a robot that went up and down the hallway several times. A large rotation error was introduced by the large turns at the ends of the hallway. 6 Discussion In this paper, we formulated the problem of consistent range data registration as one of optimal pose estimation from a network of relations. The main ideas are as follows. We associate a robot pose to a range scan to define an unambiguous object frame. By consistently maintaining the history of robot poses, we effectively allow all object frames to be consistently registered in the global reference frame. We use a combination of relation-based and location-based approach to represent the world model. It can be viewed as a two-step procedure. First, spatial relations between object frames are directly derived from odometry measurements and matching pairwise frames. These relations, along with their uncertainties, constitute all the information in the model. In the second step, the relations are converted to object frame locations based on an optimization criterion. This formulation avoids the use of complex constrained optimization. Furthermore, it does not require the assumption of known a priori covariance between object frames. We also derived measurement equations compatible with the formulation. It allows practical implementation of the algorithm. We have experimentally demonstrated the effectiveness of our estimation procedure in maintaining consistency among multiple range scans. The most 2 We are grateful to Steffen Gutmann of the AI Laboratory at the Albert-Ludwigs-Universit?t Freiburg for providing us with these experimental results. expensive operation, besides pairwise scan matching, is to compute the inverse of an 3n \Theta 3n matrix. Although the number of poses n may be large for a long robot path, there are ways to limit this size to speed up the computation. The sequential procedure enables the robot to continuously maintain the optimal registration result. Our approach assumes that the robot stops to collect a complete range scan at its current position. An alternative would be to perform continuous scanning as the robot moves. Continuous scanning would generate large amounts of data that would have to be sampled. In addition, the problem of associating measurements with the correct robot position arises, as different parts of a scan will have been obtained from different robot positions. Solving this problem would require an accurate model of the robots motion. A possible solution to the problem of excessive amounts of data is to partition the continuous scan data and transform each part to one pose on the path, based on the odometry model. These are both worthwhile problems, which we consider outside the scope of this paper. Although we develop our method for mapping a 2D environment using 2D range scans, our formulation is general and it can be applied to the 3D case as well, by generalizing pose composition and linearization [12]. Acknowledgement Funding for this work was provided by NSERC Canada and by the ARK project which receives its funding from PRECARN Associates Inc., the Department of Industry, Science and Technology, NRC Canada, the Ontario Technology Fund, Ontario Hydro, and AECL. The authors would like to thank Steffen Gutmann, Joerg Illmann, Thomas Kaempke, Manfred Knick, Erwin Prassler, and Christian Schlegel from FAW, Ulm for collecting range scans and making the data available for our experiments. We thank Dr. Ingemar Cox, and the anonymous reviewers for many constructive comments. --R Maintaining representations of the environment of a mobile robot. Position referencing and consistent world modeling for mobile robots. Blanche: An experiment in guidance and navigation of an autonomous robot vehicle. World modeling and position estimation for a mobile robot using ultrasonic ranging. Consistent integration and propagation of disparate sensor observa- tions Integration, coordination and control of multisensor robot systems. Uncertain geometry in robotics. Map building for a mobile robot equipped with a 2D laser rangefinder. AMOS: Comparison of scan matching approaches for self-localization in indoor environments Stereo vision and navigation in buildings for mobile robots. Dynamic map building for an autonomous mobile robot. Shape Registration Using Optimization for Mobile Robot Navigation. Robot pose estimation in unknown environments by matching 2D range scans. Stochastic multisensory data fusion for mobile robot location and environment modelling. SICK Laser range scanner. On the representation and estimation of spatial uncertainty. A geometric feature relation graph formulation for consistent sensor fusion. --TR --CTR Reid Simmons , Dani Goldberg , Adam Goode , Michael Montemerlo , Nicholas Roy , Brennan Sellner , Chris Urmson , Alan Schultz , Myriam Abramson , William Adams , Amin Atrash , Magda Bugajska , Michael Coblenz , Matt MacMahon , Dennis Perzanowski , Ian Horswill , Robert Zubek , David Kortenkamp , Bryn Wolfe , Tod Milam , Bruce Maxwell, GRACE: an autonomous robot for the AAAI Robot challenge, AI Magazine, v.24 n.2, p.51-72, Summer Michael Montemerlo , Sebastian Thrun , Daphne Koller , Ben Wegbreit, FastSLAM: a factored solution to the simultaneous localization and mapping problem, Eighteenth national conference on Artificial intelligence, p.593-598, July 28-August 01, 2002, Edmonton, Alberta, Canada Tom Duckett , Stephen Marsland , Jonathan Shapiro, Fast, On-Line Learning of Globally Consistent Maps, Autonomous Robots, v.12 n.3, p.287-300, May 2002 Andrew Howard , Lynne E. Parker , Gaurav S. Sukhatme, Experiments with a Large Heterogeneous Mobile Robot Team: Exploration, Mapping, Deployment and Detection, International Journal of Robotics Research, v.25 n.5-6, p.431-447, May-June 2006 Andrew Howard, Multi-robot Simultaneous Localization and Mapping using Particle Filters, International Journal of Robotics Research, v.25 n.12, p.1243-1256, December 2006 Giorgio Grisetti , Gian Diego Tipaldi , Cyrill Stachniss , Wolfram Burgard , Daniele Nardi, Fast and accurate SLAM with Rao-Blackwellized particle filters, Robotics and Autonomous Systems, v.55 n.1, p.30-38, January, 2007 Sebastian Thrun , Wolfram Burgard , Dieter Fox, A Probabilistic Approach to Concurrent Mapping and Localization for Mobile Robots, Autonomous Robots, v.5 n.3-4, p.253-271, July-August 1998 Denis F. Wolf , Gaurav S. Sukhatme, Mobile Robot Simultaneous Localization and Mapping in Dynamic Environments, Autonomous Robots, v.19 n.1, p.53-65, July 2005 Sebastian Thrun , Wolfram Burgard , Dieter Fox, A Probabilistic Approach to Concurrent Mapping and Localization for Mobile Robots, Machine Learning, v.31 n.1-3, p.29-53, April/May/June 1998 Wolfram Burgard , Panos Trahanias , Dirk Hhnel , Mark Moors , Dirk Schulz , Haris Baltzakis , Antonis Argyros, Tele-Presence in Populated Exhibitions Through Web-Operated Mobile Robots, Autonomous Robots, v.15 n.3, p.299-316, November Wesley H. Huang , Kristopher R. Beevers, Topological Map Merging, International Journal of Robotics Research, v.24 n.8, p.601-613, August 2005 Cheng Chen , Han Wang, Appearance-Based Topological Bayesian Inference for Loop-Closing Detection in a Cross-Country Environment, International Journal of Robotics Research, v.25 n.10, p.953-983, October 2006 Y. L. Ip , A. B. Rad, Incorporation of Feature Tracking into Simultaneous Localization and Map Building via Sonar Data, Journal of Intelligent and Robotic Systems, v.39 n.2, p.149-172, February 2004 Wolfgang Hbner , Hanspeter A. Mallot, Metric embedding of view-graphs, Autonomous Robots, v.23 n.3, p.183-196, October 2007 Christian Frh , Avideh Zakhor, An Automated Method for Large-Scale, Ground-Based City Model Acquisition, International Journal of Computer Vision, v.60 n.1, p.5-24, October 2004 Maarja Kruusmaa, Global Level Path Planning for Mobile Robots in Dynamic Environments, Journal of Intelligent and Robotic Systems, v.38 n.1, p.55-83, September Udo Frese, A Discussion of Simultaneous Localization and Mapping, Autonomous Robots, v.20 n.1, p.25-42, January 2006 Sebastian Thrun , Michael Montemerlo, The Graph SLAM Algorithm with Applications to Large-Scale Mapping of Urban Structures, International Journal of Robotics Research, v.25 n.5-6, p.403-429, May-June 2006 Ioannis Rekleitis , Gregory Dudek , Evangelos Milios, Multi-robot collaboration for robust exploration, Annals of Mathematics and Artificial Intelligence, v.31 n.1-4, p.7-40, 2001 Dieter Fox , Wolfram Burgard , Hannes Kruppa , Sebastian Thrun, A Probabilistic Approach to Collaborative Multi-Robot Localization, Autonomous Robots, v.8 n.3, p.325-344, June 2000 Frank Dellaert , Michael Kaess, Square Root SAM: Simultaneous Localization and Mapping via Square Root Information Smoothing, International Journal of Robotics Research, v.25 n.12, p.1181-1203, December 2006 Sebastian Thrun, Robotic mapping: a survey, Exploring artificial intelligence in the new millennium, Morgan Kaufmann Publishers Inc., San Francisco, CA,
range scan alignment;mapping;range scan registration;sensor-based mobile robotics;laser range scanning
591512
Sensor-Based Control Architecture for a Car-Like Vehicle.
This paper presents a control architecture endowing a car-like vehicle moving in a dynamic and partially known environment with autonomous motion capabilities. Like most recent control architectures for autonomous robot systems, it combines three functional components: a set of basic real-time skills, a reactive execution mechanism and a decision module. The main novelty of the architecture proposed lies in the introduction of a fourth component akin to a meta-level of skills: the sensor-based manoeuvers, i.e., general templates that encode high-level expert human knowledge and heuristics about how a specific motion task is to be performed. The concept of sensor-based manoeuvers permit to reduce the planning effort required to address a given motion task, thus improving the overall response-time of the system, while retaining the good properties of a skill-based architecture, i.e., robustness, flexibility and reactivity. The paper focuses on the trajectory planning function (which is an important part of the decision module) and two types of sensor-based manoeuvers, trajectory following and parallel parking, that have been implemented and successfully tested on a real automatic car-like vehicle placed in different situations.
Introduction Autonomy in general and motion autonomy in particular has been a long standing issue in Robotics. In the late sixties-early seventies, Shakey (Nilsson, 1984) was one of the first robots able to move and perform simple tasks au- tonomously. Ever since, many authors have proposed control architectures to endow robot systems with various autonomous capabilities. Some of these architectures are reviewed in x7 and compared to the one presented in this paper. These Institut National de Recherche en Informatique et en Automatique approaches differ in several ways, however it is clear that the control structure of an autonomous robot placed in a dynamic and partially known environment must have both deliberative and reactive capabilities. In other words, the robot should be able to decide which actions to carry out according to its goal and current situation; it should also be able to take into account events (expected or not) in a timely manner. The control architecture presented in this paper aims at meeting these two requirements. It is designed to endow a car-like vehicle moving on the road network with motion autonomy and was developed in the framework of the French Prax- it'ele programme aimed at the development of a Laugier et al. new urban transportation system based on a fleet of electric vehicles with autonomous motion capabilities (Parent and Daviet, 1996). The road network is a complex environment, it is partially known and highly dynamic with moving obstacles (other vehicles, pedestrians, etc.) whose future behaviour is not known in advance. However the road network is a structured environment with motion rules (the highway code) and it is possible to take advantage of these features in order to design a control architecture that is efficient, robust and flexible. The control architecture is presented in this paper as follows: in the next section, the rationale of the architecture and its main features are overviewed. It introduces the key concept of sensor-based manoeuvres, i.e. general templates that encode the knowledge of how a specific motion task is to be performed. The model of the car-like vehicle that is used throughout the paper is then described (x3). One important component of the architecture is the trajectory planner whose purpose is to determine the trajectory leading the vehicle to its goal. Trajectory planning for car- Skills library Mission Description Sensor Data Execution Report data Commands Motion Controller Mission Monitor Skills instantiation and execution PMP generation and update SBMs library execution Trajectory Planner World Model Fig. 1. The overall control architecture. Sensor-Based Control Architecture 3 like vehicles in dynamic environments remains an open problem and a practical solution to this intricate problem is presented in x4. Afterwards the concept of sensor-based manoeuvres is explored in x5 and two types of manoeuvres are presented in detail. These two manoeuvres have been implemented and successfully tested on an experimental vehicle, the results of these experiments are finally presented in x6. 2. Overview of the Control Architecture The control architecture is depicted in Fig. 1. It relies upon the concept of sensor-based manoeuvres (SBM) which is derived from the Artificial Intelligence concept of script (Rich and Knight, 1983). A script is a general template that encodes procedural knowledge of how a specific type of task is to be performed. A script is fitted to a specific task through the instantiation of variable parametres in the template; these parameters can come from a variety of sources (a priori knowl- edge, sensor data, output of other modules, etc. Script parametres fill in the details of the script steps and permit to deal easily with the current task conditions. The introduction of SBM was motivated by the observation that the kind of motion task that a vehicle has to perform can usually be described as a series of simple steps (a script). A SBM is a script, it combines control and sensing skills. Skills are elementary functions with real-time abil- ities: sensing skills are functions processing sensor data whereas control skills are control programs (open or closed loop) that generate the appropriate commands for the vehicle. Control skills may use data provided directly by the sensors or by the sensing skills. The idea of combining basic real-time skills to build a plan in order to perform a given task can be found in other control architectures (cf. x7); they permit to obtain robust, flexible and reactive behaviours. SBMs can be seen as "meta-skills", their novelty is that they permit to encapsulate high-level expert human knowledge and heuristics about how to perform a specific motion task (cf. x5). Accordingly they permit to reduce the planning effort required to address a given motion task, thus improving the overall response-time of the system, while retaining the good properties of a skill-based architecture, i.e. robustness, flexibility and reactivity. The control architecture features two main com- ponents, the mission monitor and the motion con- troller, that are described afterwards. 2.1. Mission Monitor When given a mission description, e.g. "go park at location l", the mission monitor (MN) generates a parameterized motion plan (PMP) which is a set of generic sensor-based manoeuvres (SBM) possibly completed with nominal trajectories. The SBMs are selected from a SBM library. A SBM may require a nominal trajectory (it is the case of the "Follow Trajectory" SBM). A nominal trajectory is a continuous time-ordered sequence of (position, velocity) of the vehicle that represents a theoretically safe and executable trajectory, i.e. a collision-free trajectory which satisfies the kinematic and dynamic constraints of the vehicle. Such trajectories are computed by the trajectory planner by using: ffl an a priori known or acquired model of the vehicle environment, ffl the current sensor data, e.g. position and velocity of the moving obstacles, and ffl a world prediction that gives the most likely behaviours of the moving obstacles. Trajectory planning is detailed in x4. The current SBM with its nominal trajectory is passed to the motion controller for its reactive execution. 2.2. Motion Controller The goal of the Motion Controller (MC) is to execute in a reactive way the current SBM of the PMP. For that purpose, the current SBM is instantiated according to the current execution con- text, i.e. the variable parametres of the SBM are set by using the a priori known or sensed information available at the time, e.g. road curvature, available lateral and longitudinal space, velocity and acceleration bounds, distance to an obstacle, etc. As mentioned above, a SBM combines control and sensing skills that are either control pro- 4 Laugier et al. grams or sensor data processing functions. It is up to MC to control and coordinate the execution of the different skills required. The sequence of control skills that is executed for a given SBM is determined by the events detected by the sensor skills. When an event that cannot be handled by the current SBM happens, MC reports a failure to MN which updates PMP either by applying a replanning procedure (time permitting), or by selecting in real-time a SBM adapted to the new situation. 3. Model of the Vehicle A car-like vehicle is modelled as a rigid body moving on the plane. It is supported by four wheels making point contact with the ground, it has two rear wheels and two directional front wheels. The model of a car-like vehicle that is used is depicted in Fig. 2. The configuration, i.e. the position and orientation of the vehicle, are characterized by the triple and are the coordinates of the rear axle midpoint and the orientation of the vehi- cle, i.e. the angle between the x axis and the main axis of the vehicle. The motion of the vehicle is described by the following equations:! sin OE (1) is the steering angle, i.e. the average orientation of the two front wheels of the vehicle. is the locomotion velocity of the front axle midpoint and L is the wheelbase. (OE; v), the steering angle and locomotion veloc- ity, are the two control commands of the vehicle. Since the steering angle of a car is mechanically limited, the following constraint also holds (max- imum curvature constraint): Eqs. (1) correspond to a system with non-holonomic kinematic constraints because they involve the derivatives of the coordinates of the vehicle and are non-integrable (Latombe, 1991). They are valid for a vehicle moving on flat ground with perfect rolling assumption (no slippage between the wheels and the ground) at relatively low speed. f y x Fig. 2. Model of a car-like vehicle. For high-speed motions, the dynamics of the vehicle must also be considered. In the current implementation of the architecture, only velocity and acceleration bounds are taken into account. 4. Trajectory Planning As mentioned earlier, trajectory planning is an important function in the control architecture pro- posed. Its purpose is to compute a nominal trajectory leading the vehicle to its goal. A trajectory is a continuous time-ordered sequence of states, i.e. (configuration, velocity) pairs, between the current state of the vehicle and its goal. A trajectory must be collision-free and satisfy the kinematic and dynamic constraints of the vehicle. In order to plan a trajectory that avoids the moving obstacles of the environment, the knowledge of their future behaviour is required. In most cases, this information is not a priori known. An estimation of the most likely behaviour of the moving obstacles is provided by a prediction func- tion. The prediction function can be very simple (assuming that the moving obstacles keep a constant velocity) or more sophisticated (using models of human driver behaviour for instance). The quality of the prediction determines the quality of the nominal trajectory. However keep in mind that the trajectory planned is nominal: if the world does not 'behave' according to the predic- tion, the motion controller will deal with the prediction error and react accordingly. On the other Sensor-Based Control Architecture 5 hand, if the prediction is correct then the vehicle will follow a trajectory that has been planned so as to be optimal in time. Trajectory planning for car-like vehicles in dynamic environments remains an open problem and a practical solution to this intricate problem is presented in this section. 4.1. Outline of the Approach The motion of a vehicle is subject to several types of constraints and the nominal trajectory has to respect them. These constraints are: ffl Kinematic constraints: a wheeled car-like vehicle is subject to kinematic constraints, called non-holonomic, that restricts the geometric shape of its motion. Such a vehicle can move only in a direction which is perpendicular to its rear wheel axle (non-steering wheels) and its turning radius is lower-bounded. ffl Dynamic constraints: these constraints arise because of the dynamics of the vehicle and the capabilities of its actuators (engine power, braking force, ground-wheel interaction, etc. They restrict the accelerations and velocities of the vehicle. constraints: collision with stationary and moving obstacles of the environment are forbidden. A trajectory is a time-ordered sequence of states q). It can be represented also by a geometric path and a velocity profile along this path. Because of the intrinsic complexity of trajectory planning (cf. (Latombe, 1991) for complexity is- sues), the trajectory planner addresses the problem at hand in two complementary steps of lesser complexity: 1. Path planning: a geometric path leading the vehicle to its goal is computed. It is collision-free with the stationary obstacles of the environment and it respects the non-holonomic kinematic constraints of the vehicle. 2. Velocity planning: the velocity profile of the vehicle along its path is computed; this profile respects the dynamic constraints of the vehicle and yields no collisions between the vehicle and the moving obstacles of the environment. Path planning is illustrated in the left-hand side of Fig. 3. It depicts an example path between two configurations. This collision-free path is a curve whose curvature is continuous and upper-bounded so as to respect the kinematic constraints of a car-like vehicle. Velocity planning is illustrated in the right - hand side of Fig. 3. Recall that it requires the knowledge of the future behaviour of the moving obstacles (this information is provided by the prediction function). In the current implementa- tion, a simple prediction function that assumes constant velocity for the moving obstacles is used. The right-hand side of Fig. 3 depicts a space-time diagram (the horizontal axis being the position along the path and the vertical one the time di- mension). The curve represents the motion of the vehicle through time whereas the thick black lines are the traces left by moving obstacles when they cross the path of the vehicle. The next two sections respectively present the path planning and the velocity planning steps. 4.2. Path Planning As mentioned earlier, a car-like vehicle is subject to non-holonomic kinematic constraints: it can move only along a direction perpendicular to its rear wheels axle (continuous tangent direction), and its turning radius is lower-bounded (maxi- mum curvature). In the past ten years, numerous works, e.g. (Barraquand and Latombe, 1989; Laumond et al., 1994; - Svestka and Overmars, 1995), have tackled the problem of computing feasible paths for this type of vehicle. Almost all of them compute paths made up of circular arcs connected with tangential line segments. The key reason for that is that these paths are the shortest ones that respect the non-holonomic kinematic constraints of such a vehicle (Dubins, 1957; Reeds and Shepp, 1990). However their curvature profile is not continuous. Accordingly a vehicle following such a path has to stop at each curvature discon- tinuity, i.e. at each transition between a segment and an arc, in order to reorient its front wheels. This is hardly acceptable for a vehicle driving on the road. A solution to this problem is therefore to plan paths with a continuous curvature profile. In addition, a constraint on the curvature deriva- 6 Laugier et al. x s y Fig. 3. (a) Path planning and (b) velocity planning. Fig. 4. Examples of continuous curvature paths. tive is introduced; it is upper-bounded so as to reflect the fact that the vehicle can only reorient its front wheels with a finite velocity. Addressing a similar problem (but without the maximum curvature constraint), (Boissonnat et al., 1994) proves that the shortest path between two vehicle's configurations is made up of line segments and clothoids 1 of maximum curvature derivative. Unfortunately, (Kostov and later proves that these shortest paths are, in the general case, made up of an infinity of clothoids. These results also apply to the problem including the maximum curvature constraint. Therefore, in order to come up with a practical solution to the problem at hand, a set of paths that contain at most eight parts, each part being either a line segment, a circular arc, or a clothoid, has been defined. It is shown in (Scheuer and Laugier, 1998) that such paths are sub-optimal in length. They are used to design a local path planner, i.e. a non-complete collision-free path planner, which in turn is embedded in a global path planning scheme. The result is the first path planner for a car-like vehicle that generates collision-free paths with continuous curvature and upper-bounded curvature and curvature derivative. The reader is referred to (Scheuer and Fraichard, 1997) for a complete presentation of the continuous curvature path planner. Various experimental results are depicted in Fig. 4. Sensor-Based Control Architecture 7 s search graph moving obstacles trajectory Fig. 5. An example of velocity planning. 4.3. Velocity Planning Given the nominal path generated by the path planner, the problem is to determine the trajectory of the vehicle along this path, i.e. its velocity profile; this profile must respect the dynamic constraints of the vehicle and yields no collision between the vehicle and the moving obstacles of the environment. To address these two issues, i.e. moving obstacles and dynamic constraints, the concept of state-time space, has been introduced. It stems from two concepts that have been used before in order to deal respectively with moving obstacles and dynamic constraints, namely the concepts of configuration-time space (Erdmann and Lozano- Perez, 1987), and state space, i.e. the space of the configuration parameters and their deriva- tives. Merging these two concepts leads naturally to state-time space, i.e. the state space augmented of the time dimension (Fraichard, 1993). In this framework, the constraints imposed by both the moving obstacles and the dynamic constraints are represented by static forbidden regions of statetime space. Besides a trajectory maps to a curve in state-time space hence trajectory planning in dynamic workspaces simply consists in finding a curve in state-time space, i.e. a continuous sequence of state-times between the current state of the vehicle and a goal state. Such a curve must obviously respect additional constraints due to the fact that time is irreversible and that velocity and acceleration constraints translate to geometric constraints on the slope and the curvature along the time dimension. However it is possible to extend previous methods for path planning in configuration space in order to solve the problem at hand. In particular, a method derived from the one originally presented in (Canny et al., 1988) has been designed to solve the problem at hand. It follows the paradigm of near-time-optimization: the search for the solution trajectory is performed over a restricted set of canonical trajectories hence the near-time-optimality of the solution. These canonical trajectories are defined as having piece-wise constant acceleration that change its value at given times. Besides the acceleration is selected so as to be either minimum, null or maximum (bang controls). Under these assumptions, it is possible to transform the problem of finding the time-optimal canonical trajectory to finding the shortest path in a directed search graph embedded in the state-time space. An example of velocity planning is depicted in Fig. 5. There are two windows: a trace window showing the part of the search graph which has been explored and a result window displaying the final trajectory. Any such window represents the s\Thetat plane (the position axis is horizontal while the time axis is vertical; the frame origin is at the upper-left corner). The thick black 8 Laugier et al. segments represent the trails left by the moving obstacles and the little dots are nodes of the underlying state-time search graph. The obstacles are assumed to keep a constant velocity. The vehicle starts from position 0 (upper-left corner) with a null velocity, it is to reach position 1 (right border) with a null velocity. The reader is referred to (Fraichard, 1993) and (Fraichard and Scheuer, 1994) for more details about velocity planning. 5. Sensor-Based Manoeuvres Recall that the control architecture proposed relies upon the concept of sensor-based manoeuvres (SBM). At a given time instant, the vehicle is carrying out a particular SBM that has been instantiated to fit the current execution context (see x2). SBMs are general templates encoding the knowledge of how a given motion task is to be performed. They combine real-time functions, control and sensing skills, that are either control programs or sensor data processing functions. This section describes the two SBMs that have been developed and integrated in the control architecture proposed: trajectory following and parallel parking. These two manoeuvres have been implemented and successfully tested on a real automatic vehicle, the results of these experiments are presented in x6. The Orccad tool (Simon et al., 1993) has been selected to implement both SBMs and skills. "Robot procedures" (in the Or- ccad are used to encode SBMs while "robot-tasks" encode skills. Robot procedures and robot tasks can both be represented as finite automata or transition diagrams. The "trajectory and "parallel parking" SBMs are depicted in Fig. 6 as transition diagrams. The control skills are represented by square boxes, e.g. "find parking place", whereas the sensing skills appear as predicates attached to the arcs of the diagram, e.g. "parking place detected", or conditional statements, e.g. "obstacle overtaken?". The next two sections describe how the two manoeuvres illustrated in Fig. 6 operates. 5.1. Trajectory Following The purpose of the trajectory following SBM is to allow the vehicle to follow a given nominal trajectory as closely as possible, while reacting appropriately to any unforeseen obstacle obstructing the way of the vehicle. Whenever such an obstacle is detected, the nominal trajectory is locally modified in real time, in order to avoid the colli- sion. This local modification of the trajectory is done, in order to satisfy a set of different motion constraints: collision avoidance, time constraints, kinematic and dynamic constraints of the vehicle. In a previous approach, a fuzzy controller combining different basic behaviours (trajectory tracking, obstacle avoidance, etc.) was used to perform trajectory following (Garnier and Fraichard, 1996). However this approach proved unsatisfactory: it yields oscillating behaviours, and does not guarantee that all the aforementioned constraints are always satisfied. The trajectory following SBM makes use of local trajectories to avoid the detected obstacles. These local trajectories allow the vehicle to move away from the obstructed nominal trajectory, and to catch up this nominal trajectory when the (sta- tionary or moving) obstacle has been overtaken. All the local trajectories verify the motion con- straints. This SBM relies upon two control skills, trajectory tracking and lane changing (cf. Fig. 6), that are detailed now. 5.1.1. Trajectory Tracking The purpose of this control skill is to issue the control commands that will allow the vehicle to track a given nominal trajectory. Several control methods for non-holonomic robots have been proposed in the lit- erature. The method described in (Kanayama et al., 1991) ensures stable tracking of a feasible trajectory by a car-like robot. It has been selected for its simplicity and efficiency. The vehicle's control commands are of the following where represents the error between the reference configuration q ref and the current configuration q of the vehicle (q ' ref and v R;ref are the reference velocities, is the rear axle midpoint velocity, k x , Sensor-Based Control Architecture 9 Updating Goal reached ? obstacle detection Free-lane reached no no yes no Trajectory tracking Lane changing Obstacle avoidance possible Nominal Trajectory Trajectory tracking Lane changing reached ? Nominal trajectory yes no yes no Nominal trajectory to reach backward & forward motions Parallel Parking SUCCESS SUCCESS Success ? Generation-execution of appropriate Find a parking place building local map Reaching an appropriate start location Free parking place detected Free-space parameters Free-lane description yes yes Trajectory Following Fig. 6. The "parallel parking" and "trajectory following" SBMs. are positive constants (the reader is referred to (Kanayama et al., 1991) for full details about this control scheme). 5.1.2. Lane Changing This control skill is applied to execute a lane changing manoeuvre. The lane changing is carried out by generating and traffic lane nominal trajectory obstacle traffic lane Fig. 7. Generation of smooth local trajectories for avoiding an obstacle. tracking an appropriate local trajectory. Let T be the nominal trajectory to track, d T be the distance between T and the middle line of the free lane to reach, s T be the curvilinear distance along T between the vehicle and the obstacle (or the selected end point for the lane change), and be the curvilinear abscissa along T since the starting point of the lane change (cf. Fig. 7). A feasible smooth trajectory for executing a lane change can be obtained using the following quintic polynomial (cf. (Nelson, 1989)): s s s In this approach, the distance d T is supposed to be known beforehand. Then the minimal value required for s T can be estimated as follows: Laugier et al. where C max stands for the maximum allowed curvature is the maximum allowed lateral acceleration, is an empirical constant, e.g. in our experiments. At each time t from the starting time T 0 , the reference position p ref is translated along the vector represents the unit normal vector to the nominal velocity vector along T ; the reference orientation ' ref is converted into @d @s , and the reference velocity v R;ref is obtained using the following equation: \Deltat where dist stands for the Euclidean distance. As shown in Fig. 6, this type of control skill can also be used to avoid a stationary obstacle, or to overtake another vehicle. As soon as the obstacle has been detected by the vehicle, a value s T;min is computed according to (6) and compared with the distance between the vehicle and the obstacle. The result of this computation is used to decide which behaviour to apply: avoid the obstacle, slow down or stop. In this approach, an obstacle avoidance or overtaking manoeuvre consists of lane changing manoeuvre towards a collision-free "virtual" parallel trajectory(see Fig. 7). The lane changing skill operates the following way: 1. Generate a smooth local trajectory - 1 which connects T with a collision-free local trajec- B22 parking place traffic lane parking lane border of the parking lane traffic direction Fig. 8. Situation at the beginning of a parallel parking manoeuvre. tory - 2 "parallel" to T (- 2 is obtained by translating appropriately the involved piece of 2. Track - 1 and - 2 until the obstacle has been overtaken. 3. Generate a smooth local trajectory - 3 which connects - 2 with T , and track - 3 . 5.2. Parallel Parking Parallel parking comprises three main steps (cf. Fig. 6): localizing a free parking place, reaching an appropriate start location with respect to the parking place, and performing the parallel parking manoeuvre using iterative backward and forward motions until the vehicle is parked. During the first step, the vehicle moves slowly along the traffic lane and uses its range sensors to build a local map of the environment and detect obsta- cles. The local map is used to determine whether parking space is available to park the vehicle. A typical situation at the beginning of a parallel parking manoeuvre is depicted in Fig. 8. The autonomous vehicle A1 is in the traffic lane. The parking lane with parked vehicles B1, B2 and a parking place between them is on the right-hand side of A1. L1 and L2 are respectively the length and width of A1, and D1 and D2 are the distances available for longitudinal and lateral displacements of A1 within the place. D3 and D4 are the longitudinal and lateral displacements of the corner A13 of A1 relative to the corner B24 of B2. Distances D1, D2, D3 and D4 are computed from data obtained by the sensor systems. The length (D1 \Gamma D3) and wide (D2 \Gamma D4) of the free parking place are compared with the length L1 and width L2 of A1 in order to determine whether the parking place is sufficiently large. During parallel parking, iterative low-speed backward and forward motions with coordinated control of the steering angle and locomotion velocity are performed to produce a lateral displacement of the vehicle into the parking place. The number of such motions depends on the distances and the necessary parking depth (that depends on the width L2 of the vehicle A1). The start and end orientations of the vehicle are the same for each iterative motion. Sensor-Based Control Architecture 11 For the i-th iterative motion (but omitting the index "i"), let the start coordinates of the vehicle be x 0 and the end coordinates be x where T is duration of the motion. The "parallel parking" condition means that where admissible error in orientation of the vehicle. The following control commands of the steering angle OE and locomotion velocity v provide the parallel parking manoeuvre (Paromtchik and Laugier, 1996b): are the admissible magnitudes of the steering angle and locomotion velocity respectively, k corresponds to a right side (+1) or left side (-1) parking place relative to the traffic lane, k corresponds to forward (+1) or backward (-1) motion, 4-t . The shape of the type of paths that corresponds to the controls (12) and is shown in Fig. 9. The commands (10) and (11) are open-loop in the (x; ')-coordinates. The steering wheel servo-system and locomotion servo-system must execute the commands (10) and (11), in order to provide the desired (x; y)-path and orientation ' of the vehicle. The resulting accuracy of the motion in the (x; ')-coordinates depends on the accuracy of these servo-systems. Possible errors are compensated by subsequent iterative motions. For each pair of successive motions the coefficient k v in (11) has to satisfy the equation that alternates between forward and backward directions. Between successive motions, when the velocity is null, the steering wheels turn to the opposite side in order to obtain a suitable steering angle OE max or \GammaOE max to start the next iterative motion. In this way, the form of the commands (10) and (11) is defined by (12) and (13) respectively. In order to evaluate (10)-(13) for the parallel parking manoeuvre, the durations T and T , the magnitudes OE max and v max must be known. The value of T is lower-bounded by the kinematic and dynamic constraints of the steering wheel servo-system. When the control command is applied, the lower bound of T is s OE OE max are the maximal admissible steering rate and acceleration respectively for the steering wheel servo-system. The value of min gives duration of the full turn of the steering wheels from \GammaOE max to OE max or vice versa, i.e. one can choose T min . The value of T is lower-bounded by the constraints on the velocity v max and acceleration v max and by the condition T ! T . When the control command (11) is applied, the lower bound of T is ae oe tion, serves to provide a smooth motion of the vehicle when the available distance D1 is small. The computation of T and OE max aims to obtain the maximal values such that the following "longi- tudinal" and "lateral" conditions are still satisfied: Fig. 9. Shape of a parallel forward/backward motion. 12 Laugier et al. Using the maximal values of T and OE max assures that the longitudinal and, especially, lateral displacement of the vehicle is maximal within the available free parking space. The computation is carried out on the basis of the model (1) when the commands (10) and (11) are applied. In this computation, the value of v max must correspond to a safety requirement for parking manoeuvres, e.g. empirically. At each iteration i the parallel parking algorithm is summarized as follows: 1. Obtain available longitudinal and lateral displacements respectively by processing the sensor data. 2. Search for maximal values T and OE max by evaluating the model (1) with controls (10), so that conditions (16), (17) are still satisfied 3. Steer the vehicle by controls (10), (11) while processing the range data for collision avoidance 4. Obtain the vehicle's location relative to environmental objects at the parking place. If the "parked" location is reached, stop; else, go to step 1. When the vehicle A1 moves backwards into the parking place from the start location shown in Fig. 8, the corner A12 (front right corner of the vehicle) must not collide with the corner B24 (front left corner of the place). The start location must ensure that the subsequent motions will be collision-free with objects limiting the parking place. To obtain a convenient start location, the vehicle has to stop at a distance D3 that will ensure a desired minimal safety distance D5 between the vehicle and the nearest corner of the parking place during the subsequent backward mo- tion. The relation between the distances D1, D2, D3, D4 and D5 is described by a function This function can not be expressed in closed form, but it can be estimated for a given type of vehicle by using the model (1) when the commands (10) and (11) are applied. The computations are carried out off-line and the results are stored in a look-up table which is used on-line, to obtain an estimate of D3 corresponding to a desired minimal safety distance D5 for given D1, D2 and D4 (Paromtchik and Laugier, 1996a). When the necessary parking "depth" has been reached, clearance between the vehicle and the parked ones is provided, i.e. the vehicle moves forwards or backwards so as to be in the middle of the parking place between the two parked vehicles. 6. Experimental Results The approach described in the paper has been implemented and tested on our experimental automatic vehicle (a modified Ligier electric car). This vehicle is equipped with the following capabilities: 1. a sensor unit to measure relative distances between the vehicle and environmental objects, 2. a servo unit to control the steering angle and the locomotion velocity and 3. a control unit that processes data from the sensor and servo units in order to "drive" the vehicle by issuing appropriate servo commands This vehicle can either be manually driven, or it can move autonomously using the control unit based on a Motorola VME162-CPU board and a transputer net. A VxWorks real-time operating system is used. The sensor unit of the vehicle makes use of a belt of ultrasonic range sensors (Polaroid 9000) and of a linear CCD- camera. The servo unit consists of a steering wheel servo-system, a locomotion servo-system for forward and backward motions, and a braking servo-system to slow down and stop the vehicle. The steering wheel servo-system is equipped with a direct current motor and an optical encoder to measure the steering angle. The locomotion servo-system of the vehicle is equipped with a asynchronous motor and two optical encoders located onto the rear wheels (for odometry data). The vehicle has an hydraulic braking servo- system. The Motion Controller monitors the current steering angle, locomotion velocity, travelled distance, coordinates of the vehicle and range data from the environment, calculates an appropriate local trajectory and issues the required servo com- mands. The Motion Controller has been implemented using the Orccad software tools (Simon et al., 1993) running on a Sun workstation. The Sensor-Based Control Architecture 13 compiled code is transmitted via Ethernet to the VME162-CPU board. The experimental car is equipped with 14 ultrasonic range sensors (Polaroid 9000), eight of them (a minimal configuration) are used for the current version of the automatic parking system: three ultrasonic sensors are at the front of the car (looking in the forward direction), two sensors are situated on each side of the car and one ultrasonic sensor is at the rear of the car (looking in the backward direction). The measurement range is 10:0m, the sampling rate is 60ms. The sensors are activated sequentially (four sensors are emitting/receiving signals at each instant (one for each side of the car). This sensor system is intended to test the control algorithms only and for low-speed motion only. Certainly, a more complex sensor system, e.g. a combination of vision and ultrasonic sensors, must be use to ensure reliable operation in a dynamic environment. An experimental run of the "follow trajectory" SBM with obstacle avoidance on circular road (roundabout) is shown in Fig. 10. In this experi- ment, the Ligier vehicle follows a nominal trajectory along the curved traffic lane, and it finds on its way another vehicle moving at a lower velocity (see Fig. 10a). When the moving obstacle is de- tected, a local trajectory for a right lane change is generated by the system, and the Ligier performs the lane changing manoeuvre, as illustrated in Fig.10b. Afterwards, the Ligier moves along a trajectory parallel to its nominal trajectory, and a left lane change is performed as soon as the obstacle has been overtaken (Fig. 10c). Finally the Ligier catches up its nominal trajectory, as illustrated in Fig. 10d. The corresponding motion of the vehicle is depicted in Fig. 11a. The steering and velocity controls applied during this manoeuvre are shown in Fig. 11b and Fig. 11c. It can be noticed in this example that the velocity of the vehicle has increased when moving along the local "parallel" trajectory (Fig. 11c); this is due to the fact that the vehicle has to satisfy the time constraints associated to its nominal trajectory. An experimental run of the parallel parking SBM in a street is shown in Fig. 12. This manoeuvre can be carried out in environments including moving obstacles, e.g. pedestrians or some other vehicles (cf. the video (Paromtchik and Laugier, 1997)). In this experiment, the Ligier was manually driven to a position near the parking place, the driver started the autonomous parking mode and left the vehicle. Then, the Ligier moved forward autonomously in order to localize the parking place, obtained a convenient start location, and performed a parallel parking ma- noeuvre. When, during this motion a pedestrian crosses the street in a dangerous proximity to the vehicle, as shown in Fig. 12a, this moving obstacle is detected, the Ligier slows down and stops to avoid the collision. When the way is free, the Ligier continues its forward motion. Range data is used to detect the parking bay. A decision to carry out the parking maneuver is made and a convenient start position for the initial backward movement is obtained, as shown in Fig. 12b. Then, the Ligier moves backwards into the bay, as shown in Fig. 12c. During this backward motion, the front human-driven vehicle starts to move back- wards, reducing the length of the bay. The change in the environment is detected and taken into ac- count. The range data shows that the necessary "depth" in the bay has not been reached, so further iterative motions are carried out until it has been reached. Then, the Ligier moves to the middle between the rear and front vehicles, as shown in Fig. 12d. The parallel parking maneuver is completed. The corresponding motion of the vehicle is depicted in Fig. 13a where the motion of the corners of the vehicle and the midpoint of the rear wheel axle are plotted. The control commands (10) and for parallel parking into a parking place situated at the right side of the vehicle are shown in Fig. 13b and Fig. 13c respectively. The length of the vehicle is 2:5m, the width is 1:4m, and the wheelbase is 1:785m. The available distances are relative to the start location of the vehicle. The lateral distance measured by the sensor unit. The longitudinal distance was estimated so as to ensure the minimal safety distance 0:2m. In this case, five iterative motions are performed to park the vehicle. As seen in Fig. 13, the durations T of the iterative motions, magnitudes of the steering angle OE max and locomotion velocity v max correspond to the available displacements D1 and D2 within the 14 Laugier et al. a b c d Fig. 10. Snapshots of trajectory following with obstacle avoidance in a roundabout: (a) following the nominal trajectory, (b) lane changing to the right and overtaking, (c) lane changing to the left, (d) catching up with the nominal trajectory. a y x [m] motion direction nominal trajectory local trajectory -0.4 angle time velocity time [s] Fig. 11. Motion and control commands in the "roundabout" scenario: (a) motion, (b) steering angle and (c) velocity controls applied. a b c d Fig. 12. Snapshots of a parallel parking: (a) localizing a free parking place, (b) selecting an appropriate start location, (c) performing a backward parking motion; (d) completing the parallel parking. a y x [m] start location location -0.4 angle time -0.4 velocity time [s] Fig. 13. Motion and control commands in the parallel parking scenario: (a) motion, (b) steering angle and (c) velocity controls applied. parking place (e.g. the values of T , OE max and v max differ for the first and last iterative motion). 7. Related Works As mentioned in x1, motion autonomy has been a long standing issue in Robotics hence the important number of works presenting control architectures for robot systems. All these architectures are not reviewed here, the main trends are indicated instead. Three main functions are to be found in any control architecture: perception, decision and action (hence the 'perception-decision-action' paradigm). After a careful examination of the existing control architectures, it appears that, to some extent, the difference between them lies in the decision function. Two types of approaches of completely opposite philosophy have appeared: Sensor-Based Control Architecture 15 deliberative approaches: in this type of ap- proach, complex models of the environment of the robot are built from sensory data or a priori knowledge. These models are then used to perform high-level reasoning, i.e. planning, in order to determine which action to under- take. Maintaining these models and reasoning about them is, in most cases, a time-consuming process that makes these methods unable to deal with dynamic and uncertain environments. (Moravec, 1983; Nilsson, 1984) and (Waxman et al., 1985) are good examples of this type of control architectures. reactive approaches: the philosophy of this type of approach is just the opposite: they favor reactivity. The decision function is reduced to a minimum. Action follows perception closely, almost like a reflex. This type of approach is most appropriate to dynamic and uncertain environments since unexpected events can be dealt with as soon as they are detected by the sensors of the robot. One drawback however, high-level reasoning is very difficult to achieve (if not impossible). (Brooks, 1990) is the canonical sensor-based control architecture; other examples are given in (Khatib and Chatila, 1995) or (Zapata et al., 1990). In an attempt to combine the advantages of both deliberative and reactive approaches, several authors have tried to combine high and low-level reasoning functions within a single control architec- ture. This idea permits to obtain hybrid control architectures with both high-level reasoning capabilities and reactivity. The first hybrid architectures were obtained by simply putting together a deliberative and a re-active component. For instance, (Arkin, 1987) integrates a simple motion planner to a reactive architecture whereas (Gat et al., 1990) sends the output of a task planner to a simple reactive execution controller: when a problem is detected at execution time, a reflex action is performed and the task planner is reinvoked. The performance of these approaches in terms of robustness, flexibility and reactivity are far from satisfactory. Better architectures have been proposed since, e.g. (Alami et al., 1998; Gat, 1997) or (Simmons, 1994), they all combine three functional components: ffl A set of elementary real-time functions (con- trol loops, sensor data processing functions, etc. A task is performed through the activation of such functions. ffl A reactive execution mechanism that control and coordinates the execution of the real-time functions. ffl A decision module that produces the task plan and supervises its execution. It may react to events from the execution function. The control architecture presented in this paper clearly falls into this class of hybrid architectures. Skills are the real-time functions, the motion controller is the execution mechanism while the mission monitor is the decision module. With regard to these architectures, the main novelty of the approach proposed lies in the introduction of a meta-level of real-time functions, the sensor-based ma- noeuvres, that encapsulate high-level expert human knowledge and heuristics about the motion tasks to be performed, that permit to reduce the planning effort required to address a given motion task and thus to improve the overall response-time of the system. 8. Conclusion This paper has presented an integrated control architecture endowing a car-like vehicle moving in a dynamic and partially known environment (the road network) with autonomous motion ca- pabilities. Like most recent control architectures for autonomous robot systems, it combines three functional components: a set of basic real-time skills, a reactive execution mechanism and a decision module. The main novelty of the architecture proposed lies in the introduction of a fourth component akin to a meta-level of skills: the sensor-based manoeuvres, i.e. general templates that encode high-level expert human knowledge and heuristics about how a specific motion task is to be performed. The concept of sensor-based manoeuvres permit to reduce the planning effort required to address a given motion task, thus improving the overall response-time of the system, while retaining the good properties of a skill-based architecture, i.e. robustness, flexibility and reactivity Laugier et al. After a general overview of the architecture pro- posed, the paper has covered in more details the trajectory planning function (which is an important part of the decision module) and two types of sensor-based manoeuvres: trajectory following and parallel parking. Experimental results with a real automatic car-like vehicle in different situations have been reported to demonstrate the efficiency of the approach. Future works will include the development and testing of other types of sensor-based manoeuvres. Acknowledgements This work was partially supported by the Inria- Inrets 2 Praxit'ele programme on urban public transport [1994-1997], and the Inco-Copernicus project "Multi-agent robot systems for industrial applications in the transport domain" [1997-1999]. The authors would like to thank E. Gauthier for his valuable contribution to the final version of the paper. Notes 1. A clothoid is a curve whose curvature is a linear function of its arc length. 2. Institut National de Recherche sur les Transports et leur S'ecurit'e. --R An architecture for autonomy. Journal of Robotics Research 17(4) Motor schema based navigation for a mobile robot. Revue d'Intelligence Artificielle 3(2) A note on shortest paths in the plane subject to a constraint on the derivative of the curvature. A robust layered control system for a mobile robot. On the complexity of kynodynamic planning. of the IEEE Symp. On curves of minimal length with a constraint on average curvature On multiple moving objects. Dynamic trajectory planning with dynamic constraints: a 'state-time space' approach. San Diego A fuzzy motion controller for a car-like vehicle On three-layer architectures Path planning and execution monitoring for a planetary rover. A Stable Tracking Control Method for a Non-Holonomic Mobile Robot An Extended Potential Field Approach For Mobile Robot Sensor-Based Mo- tions Some properties of clothoids. de Recherche en Informatique et en Automatique. Robot Motion Planning. A motion planner for non-holonomic mobile robots The stanford cart and the CMU rover. Continuous curvature paths for autonomous vehicles. Shakey the robot. AI Center Automated urban ve- hicles: towards a dual mode PRT (Personal Rapid Transit) Automatic parallel car parking. Optimal paths for a car that goes both forwards and backwards. Artificial Intelligence. Planning sub-optimal and continuous-curvature paths for car-like robots Structured control for autonomous robots. Nagoya (JP). Visual Navigation of Roadways. Christian Laugier received the M. "Mo- tion planning for a non-holonomic mobile in a dynamic workspace" Philippe Garnier received the B. His research interests include motion control for autonomous car-like vehicles in dynamic and structured environments Igor Paromtchik received the M. Computer Systems and Robotics of the university of Karlsruhe Alexis Scheuer entered the Ecole Normale Sup'erieure de Lyon --TR
motion autonomy;control architecture;car-like vehicle
591532
A Robotic Excavator for Autonomous Truck Loading.
Excavators are used for the rapid removal of soil and other materials in mines, quarries, and construction sites. The automation of these machines offers promise for increasing productivity and improving safety. To date, most research in this area has focussed on selected parts of the problem. In this paper, we present a system that completely automates the truck loading task. The excavator uses two scanning laser rangefinders to recognize and localize the truck, measure the soil face, and detect obstacles. The excavators software decides where to dig in the soil, where to dump in the truck, and how to quickly move between these points while detecting and stopping for obstacles. The system was fully implemented and was demonstrated to load trucks as fast as human operators.
Introduction The surface mining of metals, quarrying of rock, and construction of highways require the rapid removal and handling of massive quantities of soil, ore, and rock. Typically, explosive or mechanical techniques are used to pulverize the material, and digging machines such as excavators load the material into trucks for haulage to landfills, storage ar- eas, or processing plants. As shown in Figure 1, an excavator sits atop a bench and loads material into trucks that queue up to its side. The operator is responsible for designating where the truck should park, digging material from the face and depositing it in the truck bed, and stopping for people and obstacles in the loading zone. Figure 1. Excavator loading a truck with soil in a typical mass excavation work scenario. The opportunities for automation are immense. Typical- ly, loading a truck requires several passes, each of which takes 15 to 20 seconds. Reducing the time of each loading pass by even a second translates into an enormous gain across the entire job. The operator's performance peaks early in the work shift and degrades as the shift wears on. Scheduled idle times, such as lunch and other breaks, also diminish average production across a shift. All of these factors are areas where automation can improve productivity. Safety is another opportunity. Excavator operators are most likely to be injured when mounting and dismounting the machine. Operators tend to focus on the task at hand and may fail to notice other site personnel or equipment entering the loading zone. Automation can improve safety by removing the operator from the machine and by providing complete sensor coverage to watch for potential hazards entering the work area. Numerous researchers have addressed aspects of automated earthmoving (Singh, 1997). The lowest and most common level of automation has been teleoperation. Typi- cally, the operator is removed from the scene for reasons of safety. Teleoperated excavators are used in applications that pose a danger to humans, such as the uncovering of buried ordnance (Nease and Alexander, 1993) and waste (Burks et al., 1992; Wohlford et al., 1990), or excavation around buried utilities. A higher level of autonomy is achieved by systems that share control of the excavation cycle with a human operator. Typically, these systems (Bradley et al., 1993; Bullock and Oppenheim, 1989; Huang and Bernold, 1994; Lever et al., 1994; Rocke, 1994; Sakai and Cho, 1988; Salcudean et al., 1997; Sameshima and Tozawa, 1992; Seward et al., 1992) concentrate on the process of digging. An operator chooses the starting location for the excavator's bucket and a control system takes over the process of filling the bucket using force and/or joint position feedback to accomplish the task. At the next level of autonomy are systems that automatically select where to dig. Such systems measure the topology of the terrain using ranging sensors (Feng et al., 1992; Singh, 1995; Takahashi et al., 1995) and compute dig trajectories that maximize excavated volume. At the highest level of autonomy are systems that sequence digging operations over a long period (Bullock et al., 1990; Romero-Lois et al., 1989; Singh, 1998). The prior work addresses many subproblems important for autonomous truck loading, however in order to field a fully automated system that performs at the level of its manually operated equivalent, a much broader set of problems must be solved than just digging. Sensors are needed to sense the dig face, recognize and localize the truck, and detect obstacles in the workspace. Perception algorithms are needed to process the sensor data and provide information about the work environment to the planning algo- rithms. Planning and control algorithms are needed to decide how to work the dig face, deposit material in the truck, and move the bucket between the two. We have developed a complete system for loading trucks fully autonomously with soft materials such as soil. The Autonomous Loading System (ALS) was implemented and demonstrated on a 25-ton hydraulic excavator and succeeded in loading trucks as fast as an expert human operator. The rest of the paper describes the ALS and presents results from experimental trials. 2. System Overview The Autonomous Loading System uses two scanning laser rangefinders that are mounted on either side of the boom (see Figure to sense the dig face, truck, and obstacles in the workspace. Two scanners are needed for full coverage of the workspace and to enable concurrent sensing opera- tions. Each sensor has a sample rate of 12 kHz, and a motorized mirror sweeps the beam circularly in a vertical plane. Additionally, each scanner can pan at a rate of up to degrees per second, enabling this circle to be rotated about the azimuth, as shown in Figure 3. The scanner positioned over the operator's cab is called the ``left scanner'', and it is responsible for sensing the workspace on the left hand side of the excavator. The "right scanner", which is located at a symmetric position on the right side of the boom, is responsible for sensing the workspace on the right hand side of the excavator. Figure 2. Sensors mounted on excavator. Figure 3. Two axis scanning sensor configuration. The excavator uses its scanners in the following fashion when loading a truck (Figure 4). While the excavator digs its first bucket, the left scanner pans left from the dig face across the truck both to detect obstacles and to recognize, localize, and measure the dimensions of the truck. Using this information, a desired location in the truck to dump the material is planned, and the bucket swings toward the truck. During this swing motion, the right scanner pans left across the dig face to measure its new surface, and the next location to dig is calculated. The right scanner continues to pan toward the truck. After the soil is dumped into the truck, the right scanner pans back across the dig face to detect obstacles in the way of the implements. The excavator swings back to the next dig point. During this swing, the left scanner pans across the truck to measure the soil distribution in the truck bed, and the next desired dump location is calculated. This process repeats for each subsequent loading pass until the truck is full, with the exception that truck recognition is only necessary for the first pass for each new truck. Typically, six passes are needed to load our twenty-ton truck with our excavator testbed. Left scanner Right scanner Pan axis Vehicle cab Rotating reflector Distance Scan axis Obstacle or terrain Scan plane sensor measuring Figure 4. Top view of sensor configuration. Information from the scanners is processed using an on-board array of four MIPs processors. The software architecture is shown in Figure 5. The boxes are software modules that can run on one of the system processors. Circles are hardware components such as sensors. Lines represent communication channels. The sensor interfaces receive data from the two scanners and control the panning motion of the devices. Sensor data from the interfaces are passed to scanline processors, where they are converted from spher- ical, sensor coordinates to Cartesian, world coordinates using corresponding data from the position system. These three-dimensional range points are then made available to whatever perception software modules require them. One consumer of this processed sensor data is the truck recognizer, which recognizes the truck and measures both its dimensions and location. Two others are the dig point planner, which plans a sequence of dig points for eroding the dig face, and the dump point planner, which plans a sequence of dump points for loading soil into the truck bed. The digging motion planner controls the excavator during digging at the specified location. The dumping motion planner dumps the bucket of soil into the truck and returns to the dig face. The sensor motion planner controls the panning for both scanners to coordinate scanner and excavator motion, following the scenario described above. The obstacle detector processes sensor data from the scanner that is sweeping in advance of excavator's motion and stops the machine if an obstacle is detected in its path. The machine controller interface communicates commands to the low level machine joint controller, which executes the commands and sends excavator state information back to the planning modules. Figure 5. ALS software architecture. 3. Hardware Subsystem The ALS hardware subsystem consists of the servo-controlled excavator, on-board computing system, perception sensors, and associated electronics. In this paper we focus on the perceptual sensors which provide the data from which the truck is identified, the dig location is chosen, obstacles are avoided, and ultimately the mass excavation process is achieved. With the target application of earthmoving, we focussed on developing a laser based scanning system that would be able to penetrate a reasonable amount of dust and smoke in the air. The laser itself would need to be able to accurately measure range from a variety of target materials (e.g., met- als, wood, dirt, rock, snow, ice, and water), colors and tex- tures. We also needed a system that would be robust to dust and dirt accumulating on the protective "exit" window (glass or plastic which protects the laser and optics from weather and dirt, though permits the beam to pass). Over the past decade, a variety of laser based scanners have been produced. With the exceptions of the Dornier (Shulz, 1997) and Schwartz (Schwartz) scanners, most have either been research devices or limited to indoor us- age. None that we know of addresses the problems of dust penetration or a partially occluded (i.e., dirty) exit window. We have developed two different time-of-flight scanning ladar systems that are impervious to ambient dust condi- tions. The first uses a "last-pulse" technique that observes the waveform of the returned light and rejects early returns that can arise from internal reflections off of a dirty exit Left sensor Right sensor Dig face Truck scan plane scan plane Right scanner Left scanner left sensor interface right sensor interface left scanline processor right scanline processor left sensor right sensor position system truck recognizer dig point planner dump point planner digging motion planner dumping motion planner sensor motion planner machine controller interface obstacle detector excavator states commands position data sensor data sensor data sensor data sensor data dig pt. dump pt. truck info. sensor sensor commands states sensor data to obst. det. window, or from a dust cloud obscuring the target (see Figure 6). In general, the next-to-last pulse returns are due to dust in the scene and are indicative of what a normal "first pulse" rangefinder would see. For instance, in Figure 6, a first pulse rangefinder would detect the dirty exit window and would be unable to "see" the target. Even if the window were clean, the first pulse unit would still "see" the dust cloud instead of the target. Since reflections off the exit window are rejected with the last pulse technique, the unit can be environmentally sealed using an inexpensive transparent cover that does not have to be optically perfect or clean. Another advantage is that the laser system can also report when multiple returns occur, giving a warning that dust is present. This is important because overall ranging reliability and accuracy is decreased in dusty condi- tions, so an autonomous machine might need to adopt a slower, more conservative motion strategy. Figure 6. Last pulse detection concept. Figure 7. Trailing edge detection of target when target is obscured in dust cloud. There is, however, a limitation to last-pulse rangefind- ing. When the target is within the dust cloud, the receiver electronics can have difficulty separating the dust and target returns (see Figure 7). We have built a second dust penetrating scanner system that identifies that target by locating the "trailing edge" of the last return signal as is shown in Figure 7. Like the last pulse system, this device is also robust to occlusions on the exit window making it ideal for construction and mining environments. Though the trailing edge detection technique forgoes some range accu- racy, we believe it is a superior approach for environments where the dust may frequently surround the target. Figure 8. Last pulse vs. trailing edge detection when target is within dust cloud. The television monitor pictured in Figure 8 shows range points plotted from a single scanline for both the last pulse and trailing edge scanners. Range increases from the left to the right of the monitor. The top monitor screen shows scans of the rear of a dump truck. The bottom screen shows scans of the same truck but shrouded in a heavy dust cloud. Note that the last pulse device is unable to separate the dust cloud from the truck and reports the front of the cloud. The trailing edge device correctly reports range to the truck regardless of the presence of dust. It is important to note that both dust penetrating techniques are physically limited by very heavy dust levels that attenuate the return target signal below the point of detectability 4. Software Subsystem The software subsystem consists of several software modules that process sensor data, recognize the truck, select digging and dumping locations, move the excavator's joints, and guard against collision. In this section, the algorithms employed by key software modules in the software architecture are described. 4.1. Truck Recognition In order to properly load a truck, an excavator operator must verify that it is a loadable vehicle, determine its loca- tion, and determine its dimensions. This information is essential for calculating a loading strategy and for planning the sequence of joint motions that implements this strategy. Dirty exit window Dust cloud Target Laser source Threshold (using last pulse) Return signal Target range Dirty exit window Dust cloud Target Laser source Return signal Threshold (w/trailing edge) Target range (w/last pulse) In some scenarios, such as surface mining, the loaders are serviced by a mine-owned fleet of haulage trucks. An automated system could acquire this information by equipping each truck with a global positioning system (GPS) sensor and an identification transponder. However, in other scenarios such as highway construction, the loaders are serviced by a variety of independently-owned, on-highway trucks of varying dimensions, so equipping each and every truck with such sensors could be infeasible. For such sce- narios, an automated system could acquire the necessary information using rangefinder data. Figure 9. Raw range data of a truck. Figure 10. Truck model fit to segmented data. The truck recognizer uses sensor data to automatically recognize, localize, and dimension haulage trucks. As the excavator digs its first bucket of soil, the left scanner pans across the truck, which is assumed to be parked to the ex- cavator's side. The raw sensor data are shown in Figure 9. Each rotation of the mirror returns one vertical scanline of data, created by intersecting a vertical plane with the truck. Each scanline is processed into line segments which are grouped with coplanar line segments from other scanlines to form planar regions. Using an interpretation tree approach (Grimson, 1990), the simple model for a truck bed, shown in Figure 10, is matched to the segmented data region by region. Depth-first search is used to hypothesize model-to-scene region matches. At each level in the tree, constraints are used to prune the search and to check for consistency with previously hypothesized matches. The interpretation that matches most of the model regions and survives the verification stage is selected as the correct one. In order for the truck recognizer to recognize a class of truck models rather than just a single model, the model in Figure 10 uses parameter ranges rather than single parameter values. Ranges are used on the sizes of the planar regions in the model, the locations of their centroids relative to each other, and the angles between the planes. These parameter ranges are checked for consistency at every level in the interpretation process to prune the search. This specification allows the truck recognizer to identify trucks of varying sizes and truck bed shapes. For each complete interpretation (i.e. an attempt to match all model regions to scene regions), the truck recognizer performs a verification. The verification consists of finer-grained consistency checking of truck parameters, and the identification of the four "corner points" in the sensor data that define the opening of the truck bed. For the selected in- terpretation, the corner points are used to calculate the position and orientation of the truck bed. This information is passed to other modules in the system for producing a dumping strategy. Figure 10 shows the model matched to the planar regions segmented from the raw sensor data. 4.2. Coarse-to-Fine Dig Point Planning Automated earthmoving operations such as leveling a mound of soil are distinguished from typical planning problems in two important ways. First, soil is diffuse and therefore a unique description of the world requires a very large number of variables. Second, the interaction between the robot and the world is very complex and only approximate models that are also computationally tractable are available. The large state space and complex robot-world interaction imply that only locally optimal planners (i.e. per dig) can be created. In order to deal with the practical issues of excavating large volumes of earth in applications, we have developed a multi-resolution planning and execution scheme. At the highest level is a coarse planning scheme that uses the geometry of the site and the goal configuration of the terrain to plan a sequence of "dig-regions." In turn, each dig region is searched for the "best" dig that can be executed in that region. Finally, the selected dig is executed by a force based closed loop control scheme (Rocke, 1994). Treatment of the problem at three levels meets different ob- jectives. The coarse planner ensures even performance over a large number of digs. The refined planner chooses digs that meet geometric constraints (reachability and colli- sions) and which locally optimize a cost function (e.g. vol- ume, energy, time). At the lowest level is a control scheme that is robust to errors in sensing the geometry of the terrain Figure 11 shows the process of coarse to fine planning for the excavator. Figure 11. Coarse to fine planning strategy. Figure 12. Coarse plan for an excavator. The coarse planner takes as input processed sensor data which it places in a terrain map (a 2-D grid of height val- ues). The output is a sequence of dig regions, each of which is in turn sent to a refined planner. Figure 12 shows a strategy for removing material that was recommended by an expert excavator operator. Each box indicates a region, and the number within the box indicates the order in which the region is provided to the refined planner. In this strategy, material is removed from left to right, and from the top of the face to the bottom. There are several reasons for choosing this strategy. In most cases, the truck is parked on the operator's left hand side so that the operator has an unobstructed view of it. By digging from left to right, the implements do not need to be raised as high to clear material when swinging to the truck. In digging from top to bottom, less force is required from the implements because it is not necessary then to work against the weight of the material up above. In addition, clearing material away from the top minimizes the range shadows cast on the face of the terrain given a scanning range sensor that is mounted on the cab. The refined planner operates on an abstract representation of an atomic action (i.e. a single dig). Rather than searching for a bucket trajectory, the refined planner searches through compact task parameters within the bounds specified by the coarse planner. In order to select the best digging action, the refined planner evaluates candidates through the use of a forward model that simulates the result of choosing an action (in our case the starting location of the bucket). An evaluation function scores the trajectory resulting from each action, and the action that meets all constraints and optimizes the cost function is chosen. This process is shown in Figure 13. Figure 13. Operation of refined planner. 4.3. Template Based Dump Planning The truck must be loaded evenly and completely. Because of uncertainty in soil settlement, the dumping strategy may need to be revised for each successive bucket load. The dump point planner applies a template-based approach to robustly find the low regions of soil distribution in the truck bed. Sensor data are gathered after each bucket of soil has been dumped in the truck as the excavator is swinging back to the dig face. Like the dig point planner, the sensor data are placed in a 2-D terrain map. The dump point planner also requires information about the location of the truck, provided by the truck recognizer module, so it can filter out any irrelevant sensor data that are outside of the truck bed. The terrain map is then smoothed using a simple Gaussian filter to eliminate any sensor noise. The current grid cell resolution of the truck bed terrain map is 15 cm, with a typical map containing on the order of 500 cells. Occlusion of the deposited soil by the truck bed walls is a serious problem. Rather than assuming that nothing is in the unseen regions of the truck bed, the dump point planner fills in any unknown grid cells with the average elevation of the known grid cells. This results in some slight inaccuracies in the perceived soil distribution at first, but they diminish as more soil is placed in the truck bed. Finally, a specific terrain shape template is convolved Coarse Planner Refined Planner Closed Loop Executor Region Goal Candidate for each dig: (starting Constraints (kinematics, coarse plan Digs bucket angle, location) Forward Dig Model Evaluation swept volume, trajectory, over the entire truck bed terrain map to produce a score for each grid cell. This small 5x5 or 7x7 grid cell template looks for a certain profile of the material in the truck bed, such as a slope or a hole. Simple templates of constant elevations can be used to find the lowest elevation in the truck bed terrain map as well. The convolution operator produces a score which represents how well the template matched the particular region in the truck bed, and the location of the cell with the best score is returned as the desired dump location 4.4. Script Based Motion Planning The motion planning software coordinates the motions of the excavator's joints for each loading pass, beginning immediately after digging a bucket of soil and ending when the bucket has returned to the next dig point. The main objectives of the motion planner are to plan motions which place the soil at the desired dump location, avoid all known obstacles in the workspace such as the truck, and execute each loading cycle as quickly as possible. Figure 14. Truck loading script for an excavator. Because of power constraints and joint coupling effects of the excavator's hydraulic system, as well as the difficulty in accurately modeling the dynamics of such a machine, more traditional optimal trajectory generation schemes do not work well. Instead, recognizing the fact that the exca- vator's motions are highly repetitive and very similar from loading cycle to loading cycle, and that it operates in a relatively small portion of its total workspace, a script based approach to motion planning was adopted (Rowe and Stentz, 1997). A script is a set of rules which define the general motions of the excavator's joints for a certain task, in this case loading trucks. These rules contain a number of variables, known as script parameters, which get instantiated on every different loading pass. The rules of script were designed with the input of an expert human excavator operator and implicitly constrain what the excavator is and is not allowed to do. For example, if it was advised that moving two particular joints simultaneously was a bad idea, then the rules of the script make that motion impossible. The left hand side of the rules are functions of the excavator's state, and the right hand side of the rules are the commands which the planner sends to the excavator's low level joint controllers. Thus, when the left hand side of a particular rule evaluates to true, its corresponding command gets sent to the excavator. The rules get re-evaluated at a fixed rate, 10 Hz for example, during the execution of the excavator's motion. Figure 14 shows the script rules for the truck loading task. The numbers in boldface are one example set of script parameters, which will be described in more detail below. The q's are the excavator's state, in this case the angular positions of the joints. The commands are desired angular joint positions. Notice that each joint has its own separate script. Therefore, only one rule per joint may be active at a time. The script parameters are computed before each loading pass starts using the information about the truck's location and the desired dig and dump points. There are two types of script parameters, those which appear in the left hand side of the script rules and affect which commands are sent by the planner, and the joint commands themselves which appear on the right hand side of the rules. The command script parameters in the right hand side of the rules are primarily computed by geometric and kinematic means. For example, consider the command of from step 1 of the boom joint's script in Figure 14. That is the boom angle which is required for the excavator's bucket to safely clear the top of the truck, and is a kinematic function of the height and location of the truck relative to the excavator. Similarly, the stick joint commands are computed using knowledge about the radial distance of the truck from the excavator, and the swing joint's commands are found from the desired dig and dump points. The script parameters in the left hand side of the rules are found through a combination of simple excavator dynamic models and heuristics. These simple dynamic models capture first order effects of the excavator's closed loop behavior when given desired angular position commands. These models provide information about the velocities, accelera- tions, and command latencies for each joint, which are used to intelligently coordinate the different joint motions, resulting in faster loading times. As an example, consider the case when the excavator has finished digging, and the bucket is raising up out of the ground. The excavator does not need to wait until the bucket has raised to its full clearance height before swinging to the truck. Instead, it can begin swinging at some earlier point as the bucket is still Joint 1: Swing digging finishes, wait swing to truck swing to dig Joint 3: Stick digging finishes, wait move to spill point move to dump point move to dig Command Command Joint 2: Boom digging finishes, raise lower to dig Joint 4: Bucket digging finishes, curl move to dig Command Command raising, but it must have the knowledge provided by the dynamic models about how much time it will take to swing to the truck and to raise the bucket so it can safely couple the two motions to avoid a collision. 4.5. Obstacle Detection A major requirement for automated loading is detecting and stopping for people and other obstacles which pose a threat for collision. Obstacle detection software has been developed which uses sensor data to perceive objects in the excavator's workspace, and simple dynamic models to predict where the excavator's linkage will be for a short time in the future as the excavator swings back and forth between the dig face and the truck. The predicted excavator linkage locations are compared to the sensor data, and if there is an intersection, the excavator is immediately commanded to stop. It is crucial that the sensors scan far enough ahead of the excavator's motion, and the prediction is far enough in the future, for the excavator to have enough time and space to come to a complete stop and avoid hitting the obstacle. This look-ahead distance is a function of the swing joint's maximum velocity and was found through experimentation to be between 40 to 50 in front of the ex- cavator's swing joint. The prediction of the excavator's location is done using the simplified models of the excavator's closed loop dynamic behavior. Not only is the obstacle detection algorithm predicting what the excavator itself will do, it must also simulate what the motion planner will do using the predicted excavator state. It performs this prediction at the same rate of the script rule base update, 10 Hz for instance. The final result is a list of predicted excavator linkage states for some amount of time. The look-ahead time was found empirically to be between 2 - 3 seconds. Figure 15. Depiction of the points that are calculated on the underside of the linkage. For each predicted linkage state, the coordinates of points on the envelope underneath the linkage are comput- ed. This is done using the forward kinematics of the excavator and simple linear models of the shapes of the linkages. This is shown in Figure 15. Each point on the underside of the linkage for each predicted linkage state is then compared to the 2-D elevation map of sensor data. If any point on the underside of the linkage is lower than the elevation of the grid cell that coincides with it, then a predicted collision is reported and the excavator is commanded to stop. Figure 16. Typical dig for truck loading. 5. Results Figure shows the excavator after digging a bucket of soil, and Figure 17 shows the truck after it has been loaded with six buckets of soil. To date, we have autonomously loaded our truck hundreds of times. The typical loading times are 15 to 20 seconds per pass, with six passes needed to load the truck. This rate is very close to the loading times logged by an expert operator manually loading trucks in the same configuration using the same excavator. Figure 17. Truck is loaded after six passes. 6. Conclusion We have demonstrated an autonomous loading system for excavators which is capable of loading trucks with soft ma- underside linkage envelope terial at the speed of expert human operators. The system uses two scanning laser rangefinders to recognize the truck, measure the soil on the dig face and in the truck, and to detect obstacles in the workspace. The system modifies both its digging and dumping plans based on settlement of soil as detected by its sensors. Expert operator knowledge is encoded into templates called scripts which are adjusted using simple kinematic and dynamics rules to generate very fast machine motions. We believe ours to be the first fully autonomous system to load trucks for mass excavation. Acknowledgments This paper summarizes the work of the Autonomous Loading System team. This team consists of Stephannie Behrens, Scott Boehmke, Howard Cannon, Lonnie Devier, Frazier, Tim Hegadorn, Herman Herman, Alonzo Kelly, Murali Krishna, Keith Lay, Chris Leger, Oscar Lu- engo, Bob McCall, Ryan Miller, Richard Moore, Jorgen Pedersen, Chris Ravotta, Les Rosenberg, Wenfan Shi, and Hitesh Soneji in addition to the authors. --R Artificial intelligence in the control and operation of construction plant-the autonomous robot excavator A Laboratory Study of Force-Cognitive Excavation Force and Geometry Constraints in Robot Excavation. Remote excavation using the telerobotic small emplacement excava- tor Research on Control Method of Planning Level for Excavation Robot. Object Recognition by Computer: The Role of Geometric Constraints Control Model for Robotic Backhoe Excavation and Obstacle Handling. Robotics for Challenging Environments. Intelligent Excavator Control for a Lunar Mining System. A Strategic Planner for Robot Excavation. Operation System for Hydraulic Excavator for Deep Trench Works. Impedance control of a teleoperated mini excavator of the 8th IEEE International Conference on Advanced Robotics (ICAR) Development of Auto Digging Controller for Construction Machine by Fuzzy Logic Control. Schwartz Electro-optics Inc Imaging Ladar Camera for Washing Robots. Synthesis of Tactical Plans for Robotic Excava- tion State of the Art in Automation of Earthmoving Autonomous shoveling of rocks by using image vision system on LHD. New capability for remotely controlled excavation. --TR --CTR Joaqun Gutirrez , Dimitrios Apostolopoulos , Jos Luis Gordillo, Numerical comparison of steering geometries for robotic vehicles by modeling positioning error, Autonomous Robots, v.23 n.2, p.147-159, August 2007
laser rangefinder;manipulator;dig planning;autonomous excavation;software architecture;robotic excavator;integrated robotic system
591811
Grounded Symbolic Communication between Heterogeneous Cooperating Robots.
In this paper, we describe the implementation of a heterogeneous cooperative multi-robot system that was designed with a goal of engineering a grounded symbolic representation in a bottom-up fashion. The system comprises two autonomous mobile robots that perform cooperative cleaning. Experiments demonstrate successful purposive navigation, map building and the symbolic communication of locations in a behavior-based system. We also examine the perceived shortcomings of the system in detail and attempt to understand them in terms of contemporary knowledge of human representation and symbolic communication. From this understanding, we propose the Adaptive Symbol Grounding Hypothesis as a conception for how symbolic systems can be envisioned.
Introduction The Behavior-based approach to robotics has proven that it is possible to build systems that can achieve tasks robustly, react in real-time and operate reliably. The sophistication of applications implemented ranges from simple reactivity to tasks involving topological map building and navigation. Conversely, the classical AI approach to robotics has attempted to construct symbolic representational systems based on token manipulation. There has been some success in this endeavor also. While more powerful, these systems are generally slow, brittle, unreliable and do not scale well - as their 'symbols' are ungrounded. In this paper, we present an approach for engineering grounded symbolic communication between heterogeneous cooperating robots. It involves designing behavior that develops shared groundings between them. We demonstrate a situated, embodied, behavior-based multi-robot system that implements a cooperative cleaning task using two autonomous mobile robots. They develop shared groundings that allow them to ground a symbolic relationship between positions consistently. We show that this enables symbolic communication of locations between them. The subsequent part of the paper critically examines the system and its limitations. The new understanding of the system we come to shows that our approach will not scale to complex symbolic systems. We argue that it is impossible for complex symbolic representational systems to be responsible for appropriate behavior in situated agents. We propose the Adaptive Symbol Grounding Hypothesis as a conception of how systems that communication symbolically can be envisioned. Before presenting the system we have developed, the first section briefly discusses cooperation and communication generally and looks at some instances of biological cooperation in particular. From this, we determine the necessary attributes of symbolic systems. 2. Cooperation and Communication Cooperation and communication are closely tied. Communication is an inherent part of the agent interactions underlying cooperative behavior, whether implicit or explicit. If we are to implement a concrete cooperative task that requires symbolic level communication, we must first identify the relationship between communication and cooperative behavior. This section introduces a framework for classifying communication and uses it to examine some examples of cooperative behavior in biological systems. From this examination, we draw conclusions about the mechanisms necessary to support symbolic communication. These mechanisms are utilized in our implementation of the cooperative cleaning system, as described in the subsequent section. It is important to realize that 'cooperation' is a word - a label for a human concept. In this case, the concept refers to a category of human and possibly animal behavior. It does not follow that this behavior is necessarily beneficial to the agents involved. Since evolution selects behavioral traits that promote the genes that encourage them, it will be beneficial to the genes but not necessarily the organism or species. Human cooperative behaviour, for example, is a conglomerate of various behavioural tendencies selected for different reasons (and influenced cultural knowledge). Because the design of cooperative robot systems is in a different context altogether, we need to understand which aspects are peculiar to biological systems. 2.1 Some Characteristics of Communication Many authors have proposed classifications for the types of communication found in biological and artificial systems (e.g. see Arkin and Hobbs, 1992b; Balch and Arkin, 1994; Cao et al., 1995; Dudek et al., 1993; Kube and Zhang,# 1997a). Like any classification, these divide the continuous space of communication characteristics into discrete classes in a specific way - and hence are only useful within the context for which they are created. We find it necessary to introduce another classification here. A communicative act is an interaction whereby a signal is generated by an emitter and 'interpreted' by a receiver. We view communication in terms of the following four characteristics. . Interaction distance - This is the distance between the agents during the communicative interaction. It can range from direct physical contact, to visual range, hearing range, or long range. . Interaction simultaneity - The period between the signal emission and reception. It can be immediate in the case of direct contact, or possibly a long time in the case of scent markers, for example. . Signaling explicitness - This is an indication of the explicitness of the emitter's signaling behavior. The signaling may be a side effect of an existing behavior (implicit), or an existing behavior may have been modified slightly to enhance the signal through evolution or learning. The signaling may also be the result of sophisticated behavior that was specifically evolved, learnt, or in the case of a robot, designed, for it. . Sophistication of interpretation - This can be applied to either the emitter or the receiver. It is an indication of the complexity of the interpretation process that gives meaning to the signal. For example, a chemical signal may invoke a relatively simple chain of chemical events in a receiving bacterium. It is possible that a signal has a very different meaning to the emitter and receiver - the signal may have no meaning at all to its emitter. Conversely, the process of interpretation of human language is the most complex example known. 2.2 Representation It is not possible to measure the sophistication of the interpretive process by observing the signal alone. Access to the mechanics of the process within an agent would be necessary. Unfortunately, our current understanding of the brain mechanisms underlying communication in most animals is poor, at best. Therefore, our only approach is to examine the structure of the communicated signals. Luckily, there appears to be some correlation between the structural complexity of communicated signals and the sophistication of their interpretive processes. Insect mating calls are simple in structure and we posit a simple interpretive process. At the other end of the spectrum, human language has a complex structure and we consider its interpretation amongst the most sophisticated processes known. Bird song and human music are possible exceptions, as they are often complex in structure, yet have a relatively simple interpretation. This is due to other evolutionary selection pressures, since song also provides fitness information about its emitter to prospective mates and, in the case of birds, serves to distinguish between members of different species. Science, through the discipline of linguistics, has learned much about the structure of the signals generated by humans that we call language (see Robins, 1997). We utilize a small part of that here by describing a conception of the observed structure of language. Using Deacon's terms we define three types of reference, or levels of representation: - iconic, indexical and symbolic (Deacon, 1997). Iconic representation is by physical similarity to what it represents. The medium may be physically external to the agent - for example, as an orange disc painted on a cave wall may represent the sun. Alternatively, it may be part of the agent, such as some repeatable configuration of sensory neurons, or "internal analog transforms of the projections of distal objects on our sensory surfaces" (Shepard and Cooper, 1982). Indexical reference represents a correlation or association between icons. All animals are capable of iconic and indexical representation to varying degrees. For example, an animal may learn to correlate the icon for smoke with that for fire. Hence, smoke will come to be an index for fire. Even insects probably have limited indexical capabilities. Empirical demonstrations are a mechanism for creating indexical references in others. Pointing, for example, creates an association between the icon for the physical item being indicated and the object of a sentence. The second part of this paper describes how we have used empirical demonstration to enable the communication of locations between robots. Indexical Iconic Symbolic Figure of representation The third level of representation is symbolic. A symbol is a relationship between icons, indices and other symbols. It is the representation of a higher-level pattern underlying sets of relationships. It is hypothesized that language can be represented as a symbolic hierarchy (Newell and Simon, 1972). We will use the term sub-symbolic to refer to representations that need only iconic and indexical references for their interpretation. If the interpretation of a symbol requires following references that all eventually lead to icons, the symbol is said to be grounded. That is, the symbols at the top of Figure 1 ultimately refer to relationships between the icons at the bottom. A symbol's grounding is the set of icons, indices and other symbols necessary to it. The problems associated with trying to synthesize intelligence from ungrounded symbol systems - the classical AI approach - have been well documented in the literature. One such problem is termed the frame problem (see Ford and Hayes, 1991; Pylyshym, 1987). The importance of situated and embodied agents has been actively espoused by members of the behavior-based robotics community, in recognition of these problems, for many years (see Brooks,# 1992a; Pfeifer, 1995; Steels, 1996 for a selection). This hypothesis is called the physical grounding hypothesis (Brooks, 1990). Consequently, we adopted the behavior-based approach for our implementation. 2.3 Cooperative biological systems In this subsection we describe five selected biological cooperative systems and classify the communication each employs using the scheme introduced above. From these, in the following subsection, we identify the necessary mechanisms for symbolic communication, which were transferred to the implementation of the cooperative multi-robot cleaning system. 2.3.1 Bacteria Cooperation between simple organisms on earth is almost as old as life on earth itself. Over a billion years ago bacteria existed similar to contemporary bacteria recently observed to exhibit primitive cooperation. Biologists have long understood that bacteria live in colonies. Only recently has it become evident that most bacteria communicate using a number of sophisticated chemical signals and engage in altruistic behavior (Kaiser and Losick, 1993). For example, Mycobacteria assemble into multi-cellular structures known as fruiting bodies. These structures are assembled in a number of stages each mediated by different chemical signal systems. In these cases the bacteria emit and react to chemicals in a genetically determined way that evolved explicitly for cooperation. Hence, we classify the signaling as explicit. The interaction distance is moderate compared to the size of a bacterium, and the simultaneity is determined by the speed of chemical propagation. The mechanism for interpretation is necessarily simple in bacteria. We can consider this communication without meaning preservation - the meaning of the signal is different for emitter and receiver. The emitter generates a signal without interpreting it at all (hence, it has no meaning to the emitter). The receiver interprets it iconically 1 . This type of communication has been implemented and studied in multi-robot systems. For example, Balch and Arkin have implemented a collective multi-robot system, both in simulation and with real robots, to investigate to what extent communication can 1 The stereotypical way a particular chemical receptor on the bacteria's surface triggers a chain of chemical events within, is an icon for the presence of the external chemical signal. increase their capabilities (Balch and Arkin, 1994). The tasks they implemented were based on eusocial insect tasks, such as forage, consume, and graze. One scheme employed was the explicit signaling of the emitter's state to the receiver. They showed that this improves performance, as we might expect. Specifically it provides the greatest benefit when the receiver cannot easily sense the emitter's state implicitly. This finding was also observed by Parker in the implementation of a puck moving task where each robot broadcast its state periodically (Parker, 1995); and by Kube and Zhang with their collective box pushing system (Kube and Zhang, 1994). The second part of this paper will demonstrate that the result also holds for our system. 2.3.2 Ants Of the social insect societies, the most thoroughly studied are those of ants, termites, bees and wasps (Wilson, 1971; Wilson, 1975; Crespi and Choe, 1997). Ants display a large array of cooperative behaviors. For example, as described in detail by Pasteels et al. (Pasteels et al., 1987), upon discovering a new food source, a worker ant leaves a pheromone trail during its return to the nest. Recruited ants will follow this trail to the food source with some variation while laying their own pheromones down. Any chance variations that result in a shorter trail to the food will be reinforced at a slightly faster rate, as the traversal time back and forth is less. Hence, it has been shown that a near optimal shortest path is quickly established as an emergent consequence of simple trail following with random variation. In this case, the interaction distance is local - the receiver senses the pheromone at the location it was emitted. As the signal persists in the environment for long periods, there may be significant delay between emission and reception. The signaling mechanism is likely to be explicit and the interpretation, while more complex than for bacteria, is still relatively simple. The ants also communicate by signaling directly from antennae to antennae. Since both emitter and receiver can interpret the signal in the same way, we consider it communication with meaning preservation. The crucial element being that both agents share the same grounding for the signal. In this case, the grounding is probably genetically determined - through identical sensors and neural processes. This mechanism can also be applied to multi-robot systems. For example, if two robots shared identical sensors, they could simply signal their sensor values. This constitutes an iconic representation, and it is grounded directly in the environment for both robots identically. Nothing special needs to be done to ensure a shared grounding for the signal. 2.3.3 Wolves A social mammal of the Canine family, wolves are carnivores that form packs with strict social hierarchies and mating systems (Stains, 1984). Wolves are territorial. Territory marking occurs through repeated urination on objects on the periphery of and within the territories. This is a communication scheme reminiscent of our ants and their chemical trails. Wolves also communicate with pheromones excreted via glands near the dorsal surface of the tail. Wolves hunt in packs. During a pack hunt, individuals cooperate by closely observing the actions of each other and, in particular, the dominant male who directs the hunt to some extent. Each wolf knows all the pack members and can identify them individually, both visually and by smell. Communication can be directed at particular individuals and consists of a combination of specific postures and vocalizations. The interaction distance in this case is the visual or auditory range respectively, and the emission and reception is effectively simultaneous. The signals may be implicit, in the case of observing locomotory behavior, for example; or more explicit in the case of posturing, vocalizing and scent marking. It seems likely that the signals in each of these cases are interpreted similarly by the emitter and receiver. Again, this is an instance of communication with meaning preservation. A significant difference is that the shared grounding enabling the uniform interpretation of some signals (e.g. vocalizations and postures) is not wholly genetically determined. Instead, a specific mechanism exists such that the grounding is partially learnt during development - in a social environment sufficiently similar to both that a shared meaning is ensured. 2.3.4 Non-human primates Primates display sophisticated cooperative behavior. The majority of interactions involve passive observation of collaborators via visual and auditory cues, which are interpreted as actions and intentions. As Bond writes in reference to Vervet monkeys, "They are acutely and sensitively aware of the status and identity of other monkeys, as well as their temperaments and current dispositional states" (Bond, 1996). Higher primates are able to represent the internal goals, plans, dispositions and intentions of others and to construct collaborative plans jointly through acting socially (Cheney and Seyfarth, 1990). In this case, the interaction is simultaneous and occurs within visual or auditory range. The signaling is implicit but the sophistication of interpretation for the receiver is considerable. Some explicit posing and gesturing is also utilized, which is used to establish and control ongoing cooperative interactions. As with the Wolves, we observe communication with meaning preservation through a shared grounding that is developed through a developmental process. In this case, the groundings are more sophisticated, as is the developmental process required to attain them. 2.3.5 Humans In addition to the heritage of our primate ancestors, humans make extensive use of communication, both written and spoken, that is explicitly evolved or learnt. There is almost certainly some a priori physiological support for language learning in the developing human brain (Bruner, 1982). Humans cooperate in many and varied ways. We display a basic level of altruism toward all humans and sometimes animals. We enter into cooperative relationships - symbolic contracts - with mates, kin, friends, organizations, and societies whereby we exchange resources for mutual benefit. In many cases, we provide resources with no reward except the promise that the other party, by honoring the contract, will provide resources when we need them, if possible. We are able to keep track of all the transactions and the reliability with which others honor contracts (see Deacon, 1997 for a discussion). Humans also use many types of signaling for communication. Like our primate cousins, we make extensive use of implicit communication, such as posturing (body language). We also use explicit gesturing - pointing, for example. Facial expressions are a form of explicit signaling that has evolved from existing expressions to enhance the signaling reliability and repertoire. Posturing, gesturing and speaking all involve simultaneous interaction. However, with the advent of symbolic communication we learned to utilize longer-term interactions. A physically realized icon, such as a picture, a ring or body decoration, is more permanent. The ultimate extension of this is written language. The coming of telephones, radios and the Internet have obviously extended the interaction distances considerably. While symbolic communication requires considerable sophistication of interpretation, humans also use signals that can be interpreted more simply. For example, laughter has the same meaning to all humans, but not to other animals. We can make the necessary connection with the emotional state since we can hear and observe others and ourselves laughing - we share the same innate involuntary laugh behavior. The developmental process that provides the shared groundings for human symbolic communication - cultural language learning - can be seen as an extension of the processes present in our non-human primate ancestors (Hendriks-Jansen, 1996). The major differences being in the complexity due to the sheer number of groundings we need to learn and the intrinsic power of symbolic representation over exclusively indexical and iconic representation. Symbolic representations derive their power because they provide a degree of independence from the symbolic, indexical and iconic references that generated the relationship represented. New symbols can be learnt using language metaphor (Lakoff and Johnson, 1980; Johnson, 1991). 2.4 Symbolic communication and its prerequisites Evolution does not have the luxury of being able to make simultaneous independent changes to the design of an organism and also ensure their mutual consistency (in terms of the viability of the organism). For this reason, once a particular mechanism has been evolved, it is built upon rather than significantly re-designed to effect a new mechanism. It is only when selection pressures change enough to render things a liability that they may be discarded. This is why layering is observed in natural systems (e.g. Mallot, 1995). In the examples above, we can perceive a layering of communication mechanisms that are built up as we look at each in turn - from bacteria to humans. Each leveraging the mechanism developed in the previous layer. The ant's use of chemical pheromone trails to implement longer duration interactions is supported by direct-contact chemical communication, pioneered by their distant bacterial ancestors. Wolves also employ this type of communication, which provides an environment that supports the developmental process for learning other shared groundings. The sophistication of such developmental processes is greater in non-humans primates and significantly so in humans. However, even for humans, these processes still leverage the simpler processes that provide the scaffolding of shared iconic and indexical groundings (see Thelen and Smith, 1994; Hendriks-Jansen, 1996). We believe such layering is integral to the general robustness of biological systems. If a more sophisticated mechanism fails to perform, the lesser ones will still operate. We emulate the layering in the implementation of our system for this reason. From our examination, the following seem to be necessary for symbolic communication between two agents. . Some iconic representations in common (e.g. by possessing some physically identical sensory-motor apparatus). . Either a shared grounding for some indexical representations, a common process that develops shared indexical groundings, or a combination of both (e.g. a mechanism for learning the correlation between icons - such as correlating 'smoke' with `fire'). . A common process that develops shared groundings (e.g. mother and infant 'innate' behavior that scaffolds language development - turn-taking, intentional interpretation, mimicking etc.) Additionally, unless the symbol repertoire is to be fixed with specific processes for acquiring each symbol, it seems necessary to have: . A mechanism for learning new symbols by communicating known ones (e.g. interpretation and learning through metaphor). The implementation of this last necessity in a robot system is currently beyond the state-of-the-art. However, the first three are implemented in the cooperative cleaning system, as described in the following section. 3. The System Our research involved the development of an architecture for behavior-based agents that supports cooperation (Jung, 1998; Jung and Zelinsky, 1999) 2 . To validate the architecture we implemented a cooperative cleaning task using the two Yamabico mobile robots pictured in Figure (Yuta et al., 1991). The task is to clean our laboratory floor space. Our laboratory is a cluttered environment, so the system must be capable of dealing with movable obstacles, people and other hazards. 3.1 The Robots As we are interested in heterogeneous cooperation, we built each robot with a different set of sensors and actuators, and devised the cleaning task such that it cannot be accomplished by either robot alone. One of the robots, 'Joh', has a vacuum cleaner that can be turned on and off via software. Joh's task is to vacuum piles of 'litter' from the laboratory floor. As our aim was not to design a high performance cleaning system per se, chopped Styrofoam serves as 'litter'. Joh cannot vacuum close to walls or furniture, as the vacuum is mounted between the drive wheels. It has the capability to 'see' piles of litter using a CCD camera and a video transmitter that sends video to a Fujitsu MEP tracking vision system. The vision system uses template correlation, and can match about 100 templates at frame rate. The vision system can communicate with the robot, via a UNIX host, over a radio modem. Visual obstacle-avoidance behavior has been demonstrated at speeds of up to 600mm/sec (Cheng and Zelinsky, 1996). Figure - The two Yamabicos 'Flo' and `Joh' The other robot, 'Flo', has a brush tool that is dragged over the floor to sweep distributed litter into larger piles for Joh to pick-up. It navigates around the perimeter of the laboratory where Joh cannot vacuum and deposits the litter in open floor space. Sensing is primarily using four specifically developed passive tactile 'whiskers' (Jung and Zelinsky, 1996a). The whiskers provide values proportional to their angle of deflection. Both robots are also fitted with ultrasonic range sensors and wheel encoders. 3.2 A layered solution We implemented the cleaning task by layering solutions involving more complex behavior over simpler solutions. This provides a robust final solution, reduces the complexity of implementation and allows us to compare the system performance at intermediate stages of development. The first layer involves all the basic behavior required to clean the floor, but does not include any capacity to purposefully navigate, explicitly communicate or cooperate. Flo sweeps up litter and periodically deposits it into piles where it is accessible by Joh. Joh uses the vision to detect the piles and vacuum them up. Therefore, the signaling - depositing litter piles - is implicit in this case, as it is normal cleaning behavior. The interaction is not simultaneous, as Joh doesn't necessarily see the piles as soon as they are deposited. The interaction distance ranges over the size of the laboratory. Flo doesn't interpret the piles of litter as a signal at all - and in fact has no way of sensing them. Joh has a simple interpretation - the visual iconic representation of the pile acts as a releaser to vacuum over it. no awarness of each other implicit visual communication of likely litter position Layer 3 explicit communication of litter relative positions Layer 4 communication of litter locations Figure solution Figure visually tracking Flo (no vacuum attached) The second layer gives Joh an awareness of Flo. We added the capability for Joh to visually detect and track the motion of Flo. This is another communication mechanism that provides state information about Flo to Joh. In this case, the signaling is again implicit, the interaction distance is visual range and the interaction is simultaneous. Joh uses the visual iconic representation of Flo to ground an indexical reference for the likely location of the pile of litter deposited. Figure 4 shows Joh visually observing Flo via a distinctive pattern. Details of the implementation the visual behavior we employed can be found in (Jung et al., 1998a). A top view of typical trajectories of the robot is shown in Figure 5. The third layer introduces explicit communication. Specifically, upon depositing a pile of litter, Flo signals via radio the position (distance and orientation) of the pile relative to its body and the relative positions of the last few piles deposited. Flo and Joh both have identical wheel encoders, so we are ensured of a shared grounding for the interpretation of the communicated relative distance and orientation to piles. Although odometry has a cumulative error, this can be ignored over such short distances. The catch is that the positions are relative to Flo. Hence, Joh must transform them to egocentric positions based on the observed location of Flo. If Flo is not currently in view, the information is ignored. A typical set of trajectories is shown in Figure 6. Joh Flo Figure trajectories when Joh can observe Flo depositing litter (Layer 2) Joh Flo Litter Figure trajectories when explicit communication is utilized (Layer The fourth and final layer involves communication of litter locations by Flo to Joh even when Flo cannot be seen. This is accomplished by using a symbolic interpretation for a specific geometric relationship of positions to each other. What is communicated to convey a location is analogous to 'litter position is <specific-geometric-relation-between> <position-A> <position-B> '. The positions are indexical references that are themselves grounded through a shared process, a location-labeling behavior, described below. The distance and direction are in fact raw encoder data, hence an iconic reference, relying on the shared wheel encoders. There is no signal communicated for the symbolic relation itself (like a word), since there is only one symbol in the system, it is unambiguous. Obviously, if more symbols were known, or a mechanism for leaning new symbols available, labels for the symbols would need to be generated and signaled (and perhaps syntax established). First, we describe the action selection scheme employed, as it is the basis for the navigation and map building mechanism, which in turn is the basis for the location-labeling behavior. 3.3 Action Selection We needed to design an action selection mechanism that is distributed, grounded in the environment, and employs a uniform action selection mechanism over all behavior components. Because the design was undertaken in the context of cooperative cleaning, we also required the mechanism to be capable of cooperative behavior and communication, in addition to navigation. Each of these requires some ability to plan. This implies that the selection of which action to perform next must be made in the context of which actions may follow - that is, within the context of an ongoing plan. In order to be reactive, flexible and opportunistic, however, a plan cannot be a rigid sequence of pre-defined actions to be carried out. Instead, a plan must include alternatives, have flexible sub-plans and each action must be contingent on a number of factors. Each action in a planned sequence must be contingent on internal and external circumstances including the anticipated effects of the successful completion of previous actions. Other important properties are that the agent should not stop behaving while planning occurs and should learn from experience. There were no action selection mechanisms in the literature capable of fulfilling all our requirements. As our research is more concerned with cooperation than action selection per se, we adopted Maes' spreading activation algorithm and modified it to suit our needs. Her theory "models action selection as an emergent property of an activation/inhibition dynamics among the actions the agent can select and between the actions and the environment" (Maes, 1990a). 3.3.1 Components and Interconnections The behavior of a system is expressed as a network that consists of two types of nodes - Competence Modules and Feature Detectors. Competence modules (CMs) are the smallest units of behavior selectable, and feature detectors information about the external or internal environment. A CM implements a component behavior that links sensors with actuators in some arbitrarily complex way. Only one CM can be executing at any given time - a winner-take-all scheme. A CM is not limited to information supplied by FDs - the FDs are only separate entities in the architecture to make explicit the information involved in the action selection calculation. FD FD CM CM CM FD Key: (sucessor, predecessor or conflictor) +ve Correlation -ve Correlation Activation Link Precondition Figure Network components and interconnections The graphical notation is shown above where rectangles represent CMs and rounded rectangles represent FDs. Although there can be much exchange of information between CMs and FDs the interconnections shown in this notation only represent the logical organization of the network for the purpose of action selection. Each FD provides a single Condition with a confidence [0.1] that is continuously updated from the environment (sensors or internal states). Each CM has an associated Activation and the CM selected for execution has the highest activation from all Ready CMs whose activations are over the current global threshold. A CM is Ready if all of its preconditions are satisfied. The activations are continuously updated by a spreading activation algorithm. The system behavior is designed by creating CMs and FDs and connecting them with precondition links. These are shown in the diagram above as solid lines from a FD to a CM ending with a white square. It is possible to have negative preconditions, which must be false before the CM can be Ready. There also exist correlation links, dotted lines in the figure, from a CM to a FD. The correlations can take the values [-1.1] and are updated at run-time according to a learning algorithm. A positive correlation implies the execution of the CM causes, somehow, a change in the environment that makes the FD condition true. A negative correlation implies the condition becomes false. The designer usually initializes some correlation links to bootstrap learning. Together these two types of links, the precondition links and the correlation links, completely determine how activation spreads thought the network. The other activation links that are shown in Figure 7 are determined by these two and exist to better describe and understand the network and the activation spreading patterns. The activation links dictate how activation spreads and are determined as follows. . There exists a successor link from CM p to CM s for every FD condition in s's preconditions list that is positively correlated with the activity of p. . There exists a predecessor link in the opposite direction of every successor link. . There exists a conflictor link from CM x to CM y for every FD condition in y's preconditions list that is negatively correlated with the activity of x. The successor, predecessor and conflictor links resulting from the preconditions and correlations are shown in Figure 7. In summary, a CM s has a predecessor CM p, if p's execution is likely to make one of s's preconditions true. A CM x has a conflictor CM y, if y's execution is likely to make one of x's preconditions false. 3.3.2 The Spreading of Activation A rigorous description of the spreading activation algorithm is beyond the scope of this paper. The algorithm has been detailed in previous publications (Jung, 1998; Jung and Zelinsky, 1999). The activation rules can be more concisely described in terms of the activation links. The main spreading activation rules can be simply stated: . Unready CMs increase the activation of predecessors and decrease the activation of conflictors, and . Ready CMs increase the activation of successors. In addition, these special rules change the activation of the network from outside in response to goals and the current situation: . Goals increase the activation of CMs that can satisfy them and decrease the activation of those that conflict with them, and . FDs increase the activation of CMs for which they satisfy a precondition. To get a feel for how it works, we describe part of a network that implements the cleaning task for Flo, as shown in Figure 8. With some of the components shown, a crude perimeter-following behavior is possible. The rectangles are basic behaviors (CMs), the ovals feature detectors (FDs), and only the correlation and precondition links are shown (the small circles indicate negation of a precondition). The goal is Cleaning. This occurs when Flo roughly follows the perimeter of the room by using Follow to follow walls and ReverseTurn to reverse and turn away from the perimeter when an obstacle obstructs the path. Periodically the litter that has accumulated in the sweeper is deposited away from the perimeter by DumpLitter. The spreading activation algorithm 'injects' activation into the network CMs via goals and via FDs that meet a precondition. Therefore, the Cleaning goal causes an increase in the activation of Follow, DumpLitter and ReverseTurn. Suppose Flo is in a situation where its left whiskers are against a wall (ObstacleOnLeft is true) and there are no obstacles in front (ObstacleAhead and FrontHit both false). In this case, the activation of Follow will be increased by all the FDs in its precondition set (including Timer which is false before being triggered). Being the only CM ready, it is scheduled for execution until the situation changes. Once the Timer FD becomes true, Follow is no longer ready, but DumpLitter becomes ready and is executed. Follow and DumpLitter also decrease each other's activation as they conflict - each is correlated with the opposite state of Timer. Although, the selection of CMs in this example depends mainly on the FD states, when the selection of CMs depends more on the activation spread from other CMs, the networks can exhibit 'planning' - as Maes has shown. This is the basis for action planning in our networks, and gives rise to path planning as will be described below. Figure Partial network for Flo (produced by our GUI). From the rules we can imagine activation spreading backward through a network, from the goals, through CMs with unsatisfied preconditions via the precondition links until a ready CM is encountered. Activation will tend to accumulate at the ready CM, as it is feeding activation forward while its successor is feeding it backward. Eventually it may be selected for execution, after which its activation is reset to zero. If its execution was successful, the precondition of its successor will have been satisfied and the successor may be executed (if it has no further unsatisfied preconditions). We can imagine multiple routes through the network, activation building up faster via shorter paths. These paths of higher activation represent 'plans' within the network. The goals act like a 'homing signal' filtering out through the network and arriving at the current 'situation'. One important difference between our and Maes' networks is that in ours the flow of activation is weighted according to the correlations - which are updated continuously at run-time according to previous experience. The mechanism for adjusting the correlation between a given CM-FD pair is simple. Each time the CM becomes active, the value of the FD's condition is recorded. When the CM is subsequently deactivated, the current value of the condition is compared with the recorded value. It is classified as one of: Became True, Became False, Remained True or Remained False. A count of these cases is maintained (B t , B f , R t , R f ). The correlation is then: corr Where the total samples N B B R R To keep the network plastic, the counts are decayed so recent samples have a greater effect than historic ones. 3.4 Navigation and map building 3.4.1 Spatial and Topological path planning There are two main approaches to navigational path planning. One method utilizes a geometric representation of the robot environment, perhaps implemented using a tree structure. Usually a classical path planner is used to find shortest routes through the environment. The distance transform method falls into this category (Zelinsky et al., 1993). These geometric modeling approaches do not fit with the behavior-based philosophy of only using categorizations of the robot- environment system that are natural for its description, rather than anthropocentric ones. Hence, numerous behavior-based systems use a topological representation of the environment in terms only of the robot's behavior and sensing (e.g. see # 1992). While these approaches are more robust than the geometric modeling approach, they suffer from non-optimal performance for shortest path planning. This is because the robot has no concept of space directly, and often has to discover the adjacency of locations. Consider the example below, where the robot in (a) has a geometric map and its planner can directly calculate the path of least Cartesian distance, directly from A to D. However, the robot in (b) has a topological map with nodes representing the points A, B, C and D, connected by a follow-wall behavior. Since it has never previously traversed directly from A to D, the least path through its map is A-B-C-D. Figure 9 - (a) Geometric vs (b) Topological Path Planning Consequently, our aim was to combine the benefits of geometric and topological map representations in a behavior-based system using our architecture. 3.4.2 A self-organizing map In keeping with the behavior-based philosophy, we found no need to explicitly specify a representation for a map or a specific mechanism for path planning. Instead, by introducing the key notion of location feature detectors (location FDs), the correlation learning and action selection naturally gave rise to map building and path planning - for 'free'. A location feature detector is a component of our architecture specialized to respond when the robot is in a particular location (the detector's characteristic location). We employ many detectors and the locations to which they respond are non-uniformly distributed over the laboratory floor space. Each location FD contains a vector v, whose components are elements of the robot state vector: non-location FD values The variable g contains global Cartesian coordinates and orientation estimated from wheel encoders and a model of the locomotion controller. The sensors include ultrasonic range readings and in Flo's case, tactile whisker values. The fds component contains the condition values of all FDs in the system, except for the location FDs themselves. For example, in Joh's case this includes visual landmark FDs. The condition confidence value of each location FD is updated by comparing it to the current state of the robot's sensors and other non-location FDs. A weighted Euclidean norm N w is used - with the (x,y) coordinate weights dominating. Hence, the vector of the location FD whose condition is true with highest confidence is considered to represent the 'current location' of the robot. The detectors are iconic representations of locations (see Figure 10). The location FD vectors v are initialized such that the (x,y) components are distributed as a regular grid over the laboratory floor space, and the other components are randomly distributed over the vector space. During operation of the system, the location FD vectors are updated using Kohonen's self-organizing map (SOM) algorithm (Kohonen, 1990). This causes the spatial distribution of the location FD vectors to approximate the frequency distribution of the robot's state vector over time. Figure shows how the detectors have organized themselves to represent one of our laboratories. One useful property of a SOM is that it preserves topology - nodes that are adjacent in the representation are neighboring locations in the vector space. Since the location FD vectors v are continuously matched with the robot state vector x, in which the (x,y) coordinates are estimated via odometry, there is a major drawback. The odometry error in (x,y) is cumulative. We remedy this by updating the robot state vector coordinates. Specifically, the system has feature detectors for various landmark types that are automatically correlated with the location FDs by the correlation learning described above. If it should happen that a landmark FD becomes true with high confidence that is strongly correlated with a location FD neighboring the location FD for the 'current location', then the state vector (x,y) component is updated. The coordinates are simply moved closer to the coordinates of the location FD to which the landmark is correlated. Assuming the landmarks don't move over moderate periods, this serves to keep the location FD (x,y) components registered with the physical floor space. Now Flo Location feature detectors (iconic refererences to position) Indexical reference to current location Figure location detector SOM and current location index The system also maintains an indexical reference that represents the robot's current location. Recall that an indexical reference is a correlation between icons. The robots each have a sense of time - in terms of the ordering relation between sensed events (which is shared to the extent that the ordering of external events is perceived to be the same by both robots). Hence, the current location index is an association between the most active location detector and the current time. It is clear this mechanism fulfills our requirement for spatial mapping. The topological mapping derives again from the correlation learning in the architecture. Specifically, the system learns by experience that a particular behavior can take the robot from one state to another - for example by changing the current location index in a consistent way. Over time, behavior such as becomes correlated with the start and end locations of a wall segment. The spreading activation will cause the behavior to be activated when the system needs to 'plan' a sub-path from the start to the end. Similarly, simple motion behavior becomes correlated with moving the robot from one location to one of its neighbors. 3.4.3 Navigation Once we have feature detectors that respond to specific locations, it is straightforward to add spatial and topological navigation. Each time a behavior (a CM) is activated, the identity of the current location FD before and after its execution is recorded. A new instance of the CM is created, and initialized with the 'source' location FD as a precondition and the 'destination' as a positive correlate. Hence, the system remembers which behavior can take it from one specific location to another. If the CM does not consistently do this, its correlation with the destination location FD will soon fall. If it falls to zero, the CM is removed from the network. Changes in the environment also cause correlations to change, thus allowing the system to adapt. With this mechanism, the system learns topological adjacency of locations in terms of behavior. For example, if the activation of the Follow CM consistently takes the robot from the location FD corresponding to the start of a wall, to the end of the wall, then the links shown below will be created. FL FL Figure Behavioral adjacency of locations via Follow The spreading activation algorithm for action selection is able to plan a sequence of CM activations to achieve navigation between any arbitrary locations. Spatial navigation is achieved by initializing the network so that a simple Forward behavior links each location FD with its eight neighbors in both directions. Hence, initially the system 'thinks' it can move in a straight line between any locations that are neighbors in the SOM. If presence of an obstacle blocks the straight-line path from one location to its neighbor, then this will be learnt through a loss of correlation between the corresponding Forward CM and the 'destination' FD. The mechanisms described here for map building and navigation are presented in detail in (Jung, 1998; Jung and Zelinsky, 1999a). 3.5 A shared grounding for locations For layer 4 of the implementation, we wanted to add the capability for Flo to communicate the locations of litter piles in a more general way. In such a way that it would be useful to Joh if Flo were not in view or even in another room. In the system as described thus far, Flo and Joh do not share any representations except the iconic representations of their shared sensors (odometry and ultrasonic). The location feature detectors may be correlated with visual landmarks in Joh's map, and whisker landmarks in Flo's (among other information). Hence, before we can communicate Flo's representation for location we need a procedure to establish a shared grounding with Joh. For this purpose, we have implemented a location labeling procedure. Location labeling is essentially behavior whereby Flo teaches Joh a location by empirical demonstration. It proceeds as follows. If Joh is tracking Flo in its visual field at a particular time and there are no previously labeled locations near by, then Joh signals Flo indicating that Flo's current location should be labeled. Although an arbitrary signal could be generated and communicated to serve as a common labeling icon for the location, in this specific case no signal is necessary. Because there are only two robots, the time ordering of the labeling procedures is identical to each. Hence, a time ordered sequence number maintained by each serves as the labeling icon with a shared grounding. The first location is labeled '1 st Label', the next `2 nd Label', etc. If Joh receives a confirmation signal from Flo, it associates the label icon with Flo's current location. Joh calculates Flo's location based on its own location and a calculation of Flo's range from visual tracking information. Flo also labels its own location index in the same way. This procedure creates an indexical representation of specific locations that are associations between a location detector icon and the label icon (the shared sequence number). Although the locations themselves are not represented using the same icons by both Flo and Joh, they represent the same physical location. Figure 12 shows the situation after the labeling procedure has occurred four times (the symbol is explained below). 3.6 A symbol for a relationship between locations The next step is to endow both Joh and Flo with the ability to represent an arbitrary location in relationship to already known locations. Recall that a symbol is defined as a relationship between other symbolic, indexical and iconic references. Ideally, symbols should be learnt, as in biological systems. The relationship a symbol represents is a generalization from a set of observed 'exemplars' - specific relationships between other symbols, indices and icons. How this can be accomplished is still an open research area. For this reason, and because we only need a single symbol that will not be referenced by higher-level symbols, we chose to simply provide the necessary relationship. We can consider the symbol a 'first-level as it is not dependent on any other symbols, but grounded directly to iconic and indexical representations. As symbol systems go, ours is as impoverished as it can be. The relationship represented by the symbol is between two known location indices and a distance and orientation in terms of wheel encoder data. The two known locations define a line segment that provides an origin for position and orientation. The wheel encoder data then provides a distance and orientation relative to this - which together defines a unique location (see Figure 13). For example, a pile could be specified as being approximately 5m away from the 2 nd labeled location at an angle of relative to the direction of the st labeled location from the 2 nd . The top of Figure 12 shows the symbol in the context of the overall system. Indexical references to shared labeled locations Represented position (wheel encoder data) iconic distance and orientation Figure - Schematic of the <specific-geometric-relation- between> symbol used to communication locations 3.7 Symbolic communication Finally, we are in a position to see how a location can be symbolically communicated from Flo to Joh. With a particular pile location in mind, Flo first calculates the representation for it using the symbolic relationship above. It selects the two closest locations, previously labeled, as the indexical references and computes the corresponding iconic wheel encoder data that will yield the desired pile location. This information is then signaled to Joh by signaling the labels for each of the known locations in turn, followed by the raw encoder data. This signal is grounded in both robots, as the 1st label 2nd label 3rd label 4th label Indexical references to shared labeled locations Now Indexical Iconic Symbolic Encoder data Symbol (represents relationship for describing a location index in relation to two known location indices and iconic encoder data) Flo current location Figure references that represent sensory data, Indexical references that associate pairs of icons (a label with a location) and a symbol (see text). The fine lines between location feature detectors show their adjacency in the SOM; the pairs of arrow headed lines from indexical references define which two icons they associate; and the two sets of arrow headed lines from the symbol designate two 'exemplars' (see text). labels were grounded through the location labeling procedure, and the wheel encoders are a shared sense. Hence, the meaning is preserved. Joh can recover the location by re-grounding the labels and reversing the computation. 3.8 Results The typical trajectories in Figure 14 show that Joh is able to successfully vacuum the litter in the pile to the left. This occurs after the location of the pile has been communicated symbolically by Flo. The pile was initially obscured by the cardboard box, but Joh was able to correctly compute its location and plan a path around the box using its map. This can be contrasted with the layer 3 solution shown in Figure 6, where no symbolic communication or map was utilized. If the box were blocking the straight-line path to the litter pile in that case, Joh would not have been able to navigate to within visual range to locate it. As the system was not designed as a floor cleaning system per-se, rigorous experiments to record its cleaning performance were not conducted. However, we did run experiments that seem to show that the addition of symbolic communication does improve cleaning performance. We expect this intuitively, as the governing factor in vacuuming performance is the path length between litter piles. The ability to navigate purposively from one known litter pile location to the next, instead of having to rely on an obstacle free path, or chance discovery of the pile locations, shortens the average path length. Joh Flo Litter Figure 14 - Typical trajectories during cooperation We also ran experiments utilizing each layer in turn (including the lower ones on which it builds). We recorded the percentage of the floor cleaned every two minutes from 3-15 minutes. It was difficult to run all of the experiments consistently for more than 15 minutes due to problems with hardware reliability. The results are plotted in Figure 15. Initially, about 30% of the 'litter' was distributed around the perimeter and the remainder scattered approximately uniformly over the rest of the floor. The percentage cleaned was estimated by dividing the floor into a grid and counting how many tiles had been cleaned.1030507090 Time (mins) Cleaned Layer 1 Layers 1&2 Layers 1-3 Layers 1-4 Figure Performance of layered cleaning solutions Clearly, the addition of each layer improves the cleaning performance. In particular, layer 4, utilizing initially falls behind as some time is used to perform location labeling rather than cleaning. This starts to pay off later after a number of locations have been labeled. This experiment also shows the robustness gained by layering the solution. The implementation of layer 1 is robust due to its simplicity. If any of the mechanisms employed in the subsequent layers were to fail, we have demonstrated that the system will continue to perform the cleaning task, although not as quickly. 4. A Critical Examination 4.1 The limitations of our system The are two obvious limitations to the approach we have described for developing grounded symbolic communication between robots. The first is that the common process by which a shared symbol grounding is developed is the design process. That is, the shared grounding was established by identical design and implementation of the mechanism for its interpretation. This is an impractical way to develop sophisticated systems, as the mechanism for the interpretation of each symbol must be designed in turn. Is this just a practicality problem, or it is impossible in principle? When we designed the system, we believed that it was possible, if impractical, to build general symbol systems in this way - by explicitly designing the process of interpretation for each symbol. We hypothesized that all that was missing was a mechanism to learn the symbolic representations - to effectively automate the process. However, we argue below that is it in fact impossible in principle (for all but the simplest systems - like the one presented). The second obvious limitation is a related one. The approach doesn't include a mechanism for learning new symbols, even if it had an existing symbol repertoire designed in. 4.2 Symbols revisited Our definition of grounded from section 2.2 contained a hidden assumption. We defined a symbol to be grounded if its interpretation required following references that all eventually lead to icons. Recall that, symbols and the structure of their relationships to each other and to indices and icons, is a linguistic one. It is the empirically observed structure of the signals that humans generate and interpret. This grammatical structure of spoken and written language is a relatively persistent one (ignoring the fact that languages change slowly over time). The hidden assumption, which we now believe to be incorrect, was that this somehow implies that a similarly persistent analogous structure must be present within the mind of the humans that generate signals conforming to the structure. That is, just because there is a relatively persistent symbolic system present in human cultural artifacts - such as books, paintings, buildings, music, etc. - this does not imply that any symbol system persists within the human mind. It was with this invalid assumption that we proceeded to construct just such a system within the robots, by representing and relating icons, indices and symbols. We believe there is ample evidence that no persistent symbolic structure within the human mind that mirrors the structure of human language exists - but this remains to be seen. Dennett has argued strongly against the idea of a Cartesian theater - a place in the mind where all the distributed information is integrated for a central decision-maker (Dennett, 1993). It seems that distributed information about the external world (possibly contradictory) need not be integrated unless a particular discrimination is necessary for performance (for example to speak or behave). Even then, only the information necessary for the discrimination need be integrated. Even if humans don't use the equivalent of a persistent cognitive grammar to reason about the world, why can't robots use one? 4.3 Symbolic representation is not situated A symbol represents a discrete category in the continuous space of sensory-motor experience. Hence it defines a boundary such that points in the space lie either within the category or outside of it - there are no gray areas. Therefore, a symbol system is a way of characterizing sensory-motor experience in terms of membership of the categories it defines. Symbols derive their power by conferring a degree of independence from the context dependent, dynamic and situated experiences from which they are learnt. This allows symbolic communication to preserve its meaning when the interaction is extended in time (e.g. the period between these words being written and you reading them). Suppose we build a robot for a particular task that necessitates symbolic communication, and endow it with a symbolic representation system according to the approach we have outlined, whereby static symbol groundings are designed in. The robot is situated in the sense that the task for which is it designed provides a context for its interaction with the environment (from the theory of situated action - Mills, 1940; Suchman, 1987). The robot is an embodied agent and has a grounded symbol system. It satisfies the criteria of the physical grounding hypothesis (Brooks, 1990). We argue that this approach to building a robot will not necessarily work, except in the simplest cases. The task in which the robot is situated dictates the discriminations it must make in order to behave appropriately - it must behave in terms of its affordances (Gibson, 1986). Since the discriminations it can make are determined by the categories defined by its symbol system, which is necessarily static, it will only work if the task very specific - ensuring the appropriate discriminations don't change. This is precisely the situation in which our system operates - in the situated context defined by a statically specified cleaning task. A robot capable of operating flexibly in a dynamic situated context must continually adapt the discriminations it makes. If using a symbolic representation system, this implies the categories defined by the symbols, and hence the meaning of the symbols themselves, must change 3 . However, a dynamic symbol system looses its power for communication - one of the main reasons for endowing the robot with a symbol system in the first place. Consequently, a robot that utilizes a static symbolic representation system (like the one we presented) cannot be situated if its task is to behave flexibly in a dynamic context. Hence, our approach of designing in the robot's groundings does not scale from systems designed to achieve simple specific tasks, to more general flexible behavior. We also see a more pragmatic way in which larger systems built via our approach can become unsituated. In order to manage complexity in the design process, we often structure a system by categorizing and apply linguistic labels to design components (i.e. we need to name elements of our designs). Although this activity is logically independent from the way the system 3 It may be possible in principle for an agent to use a static symbol system that covers all possible categorizations and hence accommodates any possible discrimination needed for appropriate behavior in any situated context. However, we dismiss this as impossible in practice due to computation intractability. functions, the anthropocentric groundings we use in our interpretation of the linguistic labels inevitably effect the design. For example, by naming a behavior component WallFollowing, we may accidentally allow hidden assumptions from our understanding of 'walls' to come into play, despite being aware of this pitfall. If the robot possesses anything that could be called a concept for a 'wall', it is surely impoverished compared to our human understanding of 'walls'. We contend that avoiding this pitfall becomes harder, to the point of practical impossibility, as the symbol systems become more complex and the discrepancy between our labels and the robot's representations grow. 4.4 Adaptive Symbol Grounding Hypothesis There is increasing evidence that humans do not reason about the world and behave using symbolic representations (Hendriks-Jansen, 1996 provides a thorough argument). Instead, like other biological systems, we represent 4 the world in terms of changing affordances - dictated by our situatedness. We make only the discriminations necessary to behave appropriately. The symbols we use to communicate seem to be generated during language production and interpretation by a dynamic process that grounds them in our adaptive internal representations while preserving their static, public, statistically persistent meaning. Hence, the symbols we generate are influenced by our situated representations during production and they have the power to influence them during interpretation. The representations themselves are only transient. We refer to this conception as the Adaptive Symbol Grounding Hypothesis. By this conception, we envisage the process of learning new concepts as follows. A process within the emitter wishing to communicate a new concept dynamically generates a transient symbolic representation that best approximates it by matching the internal representation with learnt static linguistic relationships. This structure is reflected in the signal. The interpretation process within the receiver causes a similar transient symbolic structure to emerge. Again, an approximate match is made between the symbolic structure and the internal representation - which influences the representations. In this case, the influence causes a new concept to be discovered. The structure provides the scaffolding necessary to get the receiver thinking in the right way to discover the new concept. 4 We do not mean to imply that biological agents represent the world to themselves. Of course any observations of the internal states of an agent can be said to represent something - if we as scientific observers interpret it, it represents something to us. So the essential points of the Adaptive Symbol Grounding Hypothesis can be summarized as follows. . The persistent relationships between icons, indices and symbols that comprise the hierarchical structure of language (e.g. grammar) are only observed in the communicated signals. . Agents engaging in symbolic communication do not need to maintain an explicit representation analogous to the symbolic structure of the language. . Symbol grounding is transient and adaptive. Explicit symbolic representations and their situated groundings only persist during the generation and interpretation of the signals of communication. The specific groundings with which icons for particular symbols are associated depend upon a history of use. The mapping adapts both to the immediate context and to track long-term common usage within a community of language users. 4.5 Implication for cooperative robotics In the future, we will require increasingly complex tasks to be carried out by multi-robot teams. Hence, the behavioral sophistication of the individual robots will be greater. If we wish to engineer multi-robot systems that can cooperate in complex ways, they will eventually require symbolic communication. The Adaptive Symbol Grounding Gypothesis implies that all symbols are learnt. Hence, we advocate the ubiquitous use of learning in engineering all robotic systems. Without it, we don't believe symbolic communication of significance is possible. Multi-robot systems are usually classified as either homogeneous or heterogeneous. This is usually based upon physical attributes, such as sensors and actuators; but can be equally applied to the computational and behavioral ability of the robots. A robot system is classified as heterogeneous if one or more agents are different from the others. Balch proposes a metric to measure the diversity in multi-robot systems he calls social entropy - which also recognizes physically identical robots that differ only in their behavioral repertoire (Balch, 1997). If robots are engineered with an emphasis on learning and are consequently more a product of their experience, as we suggest above, then even physically homogeneous teams will have significant social entropy. The teams will necessarily be heterogeneous in terms of their representation of the world and hence behavior. Therefore, we don't envisage homogeneous multi-robot systems playing a large role in the cooperative robotics domain in the long term. 5. Summary In the first part of the paper, we defined what we mean by grounded and provided a framework for talking about symbols in terms of indexical and iconic references. We also introduced the classification scheme for communication involving the characteristics interaction distance, interaction simultaneity, signaling explicitness and sophistication of interpretation. We discussed cooperation and communication in bacteria, ants, wolves, primates and humans in these terms to deduce some prerequisites for symbolic communication. If we are not interested in preserving the meaning of a signal between emitter and receiver, then the implementation is straightforward. If we wish to preserve meaning, then we have to ensure a shared grounding between the agents. In the case of iconic representations, as they are essentially grounded directly in sensory information, this can only be ensured if the sensors are identical between the agents. In the case of indexical and symbolic representations, a specific mechanism for establishing a shared grounding is needed. For indexical representations, an empirical demonstration can serve to ground them to appropriate icons. The location labeling procedure we implemented on our robots takes this form. We described the implementation of the cooperative cleaning system, including the spreading activation action-selection mechanism and purposive navigation in order to provide an understanding for the communication mechanism. The symbolic communication relies on: . the shared grounding of icons through common sensors, . the shared grounding for locations, developed through a specific process - the location labeling behavior, and . the shared grounding for the symbol representing a specific relationship between locations - provided by design. In the final part of the paper, we critically examined the system and its limitations. Specifically, one obvious limitation is that the system only contains a single symbol, and it was provided at design time - with no mechanism for learning further symbols. By looking again at the notion of a symbol we were able to understand that this approach cannot scale to larger systems. We argued that situated, embodied agents cannot use symbolic representations of the world to interactively behave in it. The Adaptive Symbol Grounding Hypothesis was introduced as an alternative conception for how symbol system might be used in situated agents. Finally, we concluded that symbol grounding must be learnt. Consequently, we advocate the ubiquitous use of learning in heterogeneous multi-robot systems, because without it symbolic communication is not possible. We believe this would be a severe limitation to the sophistication of cooperation in the future. --R Dimensions of Communication and Social Organization in Multi-Agent Robotic Systems Simulation of Adaptive Behavior 92 Communication in Reactive Multiagent Robotic Systems Social Entropy: a New Metric for Learning Multi-robot Teams An Architectural Model of the Primate Brain of Computer Science Elephants Don't Play Chess Intelligence Without Reason "The Analysis of Action" Cooperative Mobile Robotics: Antecedents and Directions How Monkeys see the world The Evolution of Social Behaviour in Insects and Arachnids The Symbolic Species: The co-evolution of language and the human brain Consciousness Explained A taxonomy for swarm robots Reasoning Agents in a Dynamic World: The Frame Problem The Ecological Approach to Visual Perception Catching Ourselves in the Act Knowing through the body Range and Pose Estimation for Visual Servoing on a Mobile Robotic Target An architecture for distributed cooperative planning in a behaviour-based multi-robot system Integrating Spatial and Topological Navigation in a Behavior-Based Multi-Robot Application How and Why Bacteria Talk to Each Other The self-organising map Collective Robotics: From Social Insects to Robots Metaphors we Live By Situated Agents Can Have Goals. Layered Computation in Neural Networks Situated actions and vocabularies of motive Human problem solving The Effect of Action Recognition and Robot Awareness in Cooperative Robotic Teams "From individual to collective behavior in social insects" The Robot's Dilemma. A Short History of Linguistics Mental images and their transformations "Orders and Families of Recent Mammals of the World" The origins of intelligence Plans and Situated Actions: The Problem of Human-Machine Communication A Dynamic Systems Approach to the Development of Cognition and Action The Insect Societies: Their Origin and Evolution Sociobiology: The New Synthesis Implementation of a small size experimental self-contained autonomous robot - sensors LAAS/CNRS. A Qualitative Approach to Achieving Robust Performance by a Mobile Agent --TR --CTR David Hurt , Paul Tarau, An empirical evaluation of communication effectiveness in autonomous reactive multiagent systems, Proceedings of the 2005 ACM symposium on Applied computing, March 13-17, 2005, Santa Fe, New Mexico Ariel Felner , Yaron Shoshani , Yaniv Altshuler , Alfred M. Bruckstein, Multi-agent Physical A* with Large Pheromones, Autonomous Agents and Multi-Agent Systems, v.12 n.1, p.3-34, January 2006 Luca Iocchi , Daniele Nardi , Maurizio Piaggio , Antonio Sgorbissa, Distributed Coordination in Heterogeneous Multi-Robot Systems, Autonomous Robots, v.15 n.2, p.155-168, September Peter Stone , Manuela Veloso, Multiagent Systems: A Survey from a Machine Learning Perspective, Autonomous Robots, v.8 n.3, p.345-383, June 2000
learning;cooperative robotics;heterogeneous systems;symbolic communication;representation;mobile robots;symbol grounding;behavior-based
591902
Theory of Mind for a Humanoid Robot.
If we are to build human-like robots that can interact naturally with people, our robots must know not only about the properties of objects but also the properties of animate agents in the world. One of the fundamental social skills for humans is the attribution of beliefs, goals, and desires to other people. This set of skills has often been called a theory of mind. This paper presents the theories of Leslie (1994) and Baron-Cohen (1995) on the development of theory of mind in human children and discusses the potential application of both of these theories to building robots with similar capabilities. Initial implementation details and basic skills (such as finding faces and eyes and distinguishing animate from inanimate stimuli) are introduced. I further speculate on the usefulness of a robotic implementation in evaluating and comparing these two models.
Introduction Human social dynamics rely upon the ability to correctly attribute beliefs, goals, and percepts to other people. This set of metarepresentational abilities, which have been collectively called a "theory of mind" or the ability to "mentalize", allows us to understand the actions and expressions of others within an intentional or goal-directed framework (what Dennett [15] has called the intentional stance). The recognition that other individuals have knowl- edge, perceptions, and intentions that differ from our own is a critical step in a child's development and is believed to be instrumental in self-recognition, in providing a perceptual grounding during language learning, and possibly in the development of imaginative and creative play [9]. These abilities are also central to what defines human interactions. Normal social interactions depend upon the recognition of other points of view, the understanding of other mental states, and the recognition of complex non-verbal signals of attention and emotional state. Research from many different disciplines have focused on theory of mind. Students of philosophy have been interested in the understanding of other minds and the representation of knowledge in others. Most recently, Dennett [15] has focused on how organisms naturally adopt an "intentional stance" and interpret the behaviors of others as if they possess goals, intents, and beliefs. Ethologists have also focused on the issues of theory of mind. Studies of the social skills present in primates and other animals have revolved around the extent to which other species are able to interpret the behavior of conspecifics and influence that behavior through deception (e.g. Premack [33], Povinelli and Preuss [32], and Cheney and Seyfarth [12]). Research on the development of social skills in children have focused on characterizing the developmental progression of social abilities (e.g. Fodor [17], Wimmer and Perner [37], and Frith and Frith [18]) and on how these skills result in conceptual changes and the representational capacities of infants (e.g. Carey [10], and Gelman [19]). Furthermore, research on pervasive developmental disorders such as autism have focused on the selective impairment of these social skills (e.g. Perner and Lang [31], Karmiloff-Smith et. al. [24], and Mundy and Sigman [29]). Researchers studying the development of social skills in normal children, the presence of social skills in primates and other vertebrates, and certain pervasive developmental disorders have all focused on attempting to decompose the idea of a central "theory of mind" into sets of precursor skills and developmental modules. In this paper, I will review two of the most popular and influential models which attempt to link together multi-disciplinary research into a coherent developmental explanation, one from Baron-Cohen [2] and one from Leslie [27]. Section 4 will discuss the implications of these models to the construction of humanoid robots that engage in natural human social dynamics and highlight some of the issues involved in implementing the structures that these models propose. Finally, Section 5 will describe some of the precursor components that have already been implemented by the author on a humanoid robot at the MIT Artificial Intelligence lab. Leslie's Model of Theory of Mind Leslie's [26] theory treats the representation of causal events as a central organizing principle to theories of object mechanics and theories of other minds much in the same way that the notion of number may be central to object Brian Scassellati (a) (b) (c) (d) Fig. 1. Film sequences used by Leslie [25] to study perception of causality in infants based on similar tests in adults performed by Michotte [28]. The following six events were studied: (a) direct launching - the light blue brick moves off immediately after impact with the dark red brick; (b) delayed reaction - spatially identical to (a), but a 0.5 second delay is introduced between the time of impact and the movement of the light blue brick; (c) launching without collision - identical temporal structure but without physical contact; (d) collision with no launching - identical result but without causation; (e) no contact, no launching another plausible alternative. Both adults and infants older than six months interpret events (a) and (e) as different from the class of events that violate simple mechanical laws (b-d). Infants that have been habituated to a non-causal event will selectively dishabituate to a causal event but not to other non-causal events. Adapted from Leslie [25]. representation. According to Leslie, the world is naturally decomposed into three classes of events based upon their causal structure; one class for mechanical agency, one for actional agency, and one for attitudinal agency. Leslie argues that evolution has produced independent domain-specific modules to deal with each of these classes of event. The Theory of Body module (ToBY) deals with events that are best described by mechanical agency, that is, they can be explained by the rules of mechanics. The second module is system 1 of the Theory of Mind module which explains events in terms of the intent and goals of agents, that is, their actions. The third module is system 2 of the Theory of Mind module (ToMM-2) which explains events in terms of the attitudes and beliefs of agents. The Theory of Body mechanism (ToBY) embodies the infant's understanding of physical objects. ToBY is a domain-specific module that deals with the understanding of physical causality in a mechanical sense. ToBY's goal is to describe the world in terms of the mechanics of physical objects and the events they enter into. ToBY in humans is believed to operate on two types of visual input: a three-dimensional object-centered representation from high level cognitive and visual systems and a simpler motion-based system. This motion-based system accounts for the causal explanations that adults give (and the causal expectations of children) to the "billiard ball" type launching displays pioneered by Michotte [28] (see figure 1). Leslie proposed that this sensitivity to the spatio-temporal properties of events is innate, but more recent work from Cohen and Amsel [13] may show that it develops extremely rapidly in the first few months and is fully developed by 6-7 months. ToBY is followed developmentally by the emergence of a Theory of Mind Mechanism (ToMM) which develops in two phases, which Leslie calls system-1 and system-2 but which I will refer to as ToMM-1 and ToMM-2 after Baron-Cohen [2]. Just as ToBY deals with the physical laws that govern objects, ToMM deals with the psychological laws that govern agents. ToMM-1 is concerned with actional agency; it deals with agents and the goal-directed actions that they produce. The primitive representations of actions such as approach, avoidance, and escape are constructed by ToMM-1. This system of detecting goals and actions begins to emerge at around 6 months of age, and is most often characterized by attention to eye gaze. Leslie leaves open the issue of whether ToMM-1 is innate or acquired. ToMM-2 is concerned with attitudinal agency; it deals with the representations of beliefs and how mental states can drive behavior relative to a goal. This system develops gradually, with the first signs of development beginning between months of age and completing sometime near 48 months. ToMM- Intentionality Detector (ID) Eye Direction Detector (EDD) Shared Attention Mechanism (SAM) Theory of Mind Mechanism (ToMM) Stimuli with self-propulsion and direction Eye-like stimuli Dyadic representations (desire, goal) Dyadic representations (sees) Triadic representations Full range of mental state concepts, expressed in M-representations Knowledge of the mental, stored and used as a theory Fig. 2. Block diagram of Baron-Cohen's model of the development of theory of mind. See text for description. Adapted from [2]. 2 employs the M-representation, a meta-representation which allows truth properties of a statement to be based on mental states rather than observable stimuli. ToMM-2 is a required system for understanding that others hold beliefs that differ from our own knowledge or from the observable world, for understanding different perceptual perspectives, and for understanding pretense and pretending. 3 Baron-Cohen's Model of Theory of Mind Baron-Cohen's model assumes two forms of perceptual information are available as input. The first percept describes all stimuli in the visual, auditory, and tactile perceptual spheres that have self-propelled motion. The second percept describes all visual stimuli that have eye-like shapes. Baron-Cohen proposes that the set of precursors to a theory of mind, which he calls the "mindreading system," can be decomposed into four distinct modules. The first module interprets self-propelled motion of stimuli in terms of the primitive volitional mental states of goal and desire. This module, called the intentionality detector (ID) produces dyadic representations that describe the basic movements of approach and avoidance. For example, ID can produce representations such as "he wants the food" or "she wants to go over there". This module only operates on stimuli that have self-propelled motion, and thus pass a criteria for distinguishing stimuli that are potentially animate (agents) from those that are not (objects). Baron-Cohen speculates that ID is a part of the innate endowment that infants are born with. The second module processes visual stimuli that are eye-like to determine the direction of gaze. This module, called the eye direction detector (EDD), has three basic functions. First, it detects the presence of eye-like stimuli in the visual field. Human infants have a preference to look at human faces, and spend more time gazing at the eyes than at other parts of the face. Second, EDD computes whether the eyes are looking at it or at something else. Baron-Cohen proposes that having someone else make eye contact is a natural psychological releaser that produces pleasure in human infants (but may produce more negative arousal in other animals). Third, EDD interprets gaze direction as a perceptual state, that is, EDD codes dyadic representational states of the form "agent sees me" and "agent looking-at not-me". The third module, the shared attention mechanism (SAM), takes the dyadic representations from ID and EDD and produces triadic representations of the form "John sees (I see the girl)". Embedded within this representation is a specification that the external agent and the self are both attending to the same perceptual object or event. This shared attentional state results from an embedding of one dyadic representation within another. SAM additionally can make the output of ID available to EDD, allowing the interpretation of eye direction as a goal state. By allowing the agent to interpret the gaze of others as intentions, SAM provides a mechanism for creating nested representations of the form "John sees (I want the toy)". 4 Brian Scassellati The last module, the theory of mind mechanism (ToMM), provides a way of representing epistemic mental states in other agents and a mechanism for tying together our knowledge of mental states into a coherent whole as a usable theory. ToMM first allows the construction of representations of the form "John believes (it is raining)". ToMM allows the suspension of the normal truth relations of propositions (referrential opacity), which provides a means for representing knowledge states that are neither necessarily true nor match the knowledge of the organism, such as "John thinks (Elvis is alive)". Baron-Cohen proposes that the triadic representations of SAM are converted through experience into the M-representations of ToMM. Baron-Cohen's modules match a developmental progression that is observed in infants. For normal children, ID and the basic functions of EDD are available to infants in the first 9 months of life. SAM develops between 9 and months, and ToMM develops from months to 48 months. However, the most attractive aspects of this model are the ways in which it has been applied both to the abnormal development of social skills in autism and to the social capabilities of non-human primates and other vertebrates. Autism is a pervasive developmental disorder of unknown etiology that is diagnosed by a checklist of behavioral criteria. Baron-Cohen has proposed that the range of deficiencies in autism can be characterized by his model. In all cases, EDD and ID are present. In some cases of autism, SAM and ToMM are impaired, while in others only ToMM is impaired. This can be contrasted with other developmental disorders (such as Down's syndrome) or specific linguistic disorders in which evidence of all four modules can be seen. Furthermore, Baron-Cohen attempts to provide an evolutionary description of these modules by identifying partial abilities in other primates and vertebrates. This phylogenetic description ranges from the abilities of hog- nosed snakes to detect direct eye contact to the sensitivities of chimpanzees to intentional acts. Roughly speaking, the abilities of EDD seem to be the most basic and can be found in part in snakes, avians, and most other vertebrates as a sensitivity to predators (or prey) looking at the animal. ID seems to be present in many primates, but the capabilities of SAM seem to be present only partially in the great apes. The evidence on ToMM is less clear, but it appears that no other primates readily infer mental states of belief and knowledge. 4 Implications of these Models to Humanoid Robots A robotic system that possessed a theory of mind would allow for social interactions between the robot and humans that have previously not been possible. The robot would be capable of learning from an observer using normal social signals in the same way that human infants learn; no specialized training of the observer would be necessary. The robot would also be capable of expressing its internal state (emotions, desires, goals, etc.) through social interactions without relying upon an artificial vocabulary. Further, a robot that can recognize the goals and desires of others will allow for systems that can more accurately react to the emotional, attentional, and cognitive states of the observer, can learn to anticipate the reactions of the observer, and can modify its own behavior accordingly. The construction of these systems may also provide a new tool for investigating the predictive power and validity of the models from natural systems that serve as the basis. An implemented model can be tested in ways that are not possible to test on humans, using alternate developmental conditions, alternate experiences, and alternate educational and intervention approaches. The difficulty, of course, is that even the initial components of these models require the coordination of a large number of perceptual, sensory-motor, attentional, and cognitive processes. In this section, I will outline the advantages and disadvantages of Leslie's model and Baron-Cohen's model with respect to implementation. In the following section, I will describe some of the components that have already been constructed and some which are currently designed but still being implemented. The most interesting part of these models is that they attempt to describe the perceptual and motor skills that serve as precursors to the more complex theory of mind capabilities. These decompositions serve as an inspiration and a guideline for how to build robotic systems that can engage in complex social interactions; they provide a much-needed division of a rather ambiguous ability into a set of observable, testable predictions about behavior. While it cannot be claimed with certainty that following the outlines that these models provide will produce a robot that has the same abilities, the evolutionary and developmental evidence of sub-skills does give us hope that these abilities are critical elements of the larger goal. Additionally, the grounding of high-level perceptual abilities to observable sensory and motor capabilities provides an evaluation mechanism for measuring the amount of progress that is being made. From a robotics standpoint, the most salient differences between the two models are in the ways in which they divide perceptual tasks. Leslie cleanly divides the perceptual world into animate and inanimate spheres, and allows for further processing to occur specifically to each type of stimulus. Baron-Cohen does not divide the perceptual world quite so cleanly, but does provide more detail on limiting the specific perceptual inputs that each Fig. 3. Cog, an upper-torso humanoid robot with twenty-one degrees of freedom and sensory systems that include visual, auditory, tactile, vestibular, and kinesthetic systems. module requires. In practice, both models require remarkably similar perceptual systems (which is not surprising, since the behavioral data is not under debate). However, each perspective is useful in its own way in building a robotic implementation. At one level, the robot must distinguish between object stimuli that are to be interpreted according to physical laws and agent stimuli that are to be interpreted according to psychological laws. However, the specifications that Baron-Cohen provides will be necessary for building visual routines that have limited scope. The implementation of the higher-level scope of each of these models also has implications to robotics. Leslie's model has a very elegant decomposition into three distinct areas of influence, but the interactions between these levels are not well specified. Connections between modules in Baron-Cohen's model are better specified, but they are still less than ideal for a robotics implementation. Issues on how stimuli are to be divided between the competencies of different modules must be resolved for both models. On the positive side, the representations that are constructed by components in both models are well specified. 5 Implementing a Robotic Theory of Mind Taking both Baron-Cohen's model and Leslie's model, we can begin to specify the specific perceptual and cognitive abilities that our robots must employ. Our initial systems concentrate on two abilities: distinguishing between animate and inanimate motion and identifying gaze direction. To maintain engineering constraints, we must focus on systems that can be performed with limited computational resources, at interactive rates in real time, and on noisy and incomplete data. To maintain biological plausibility, we focus on building systems that match the available data on infant perceptual abilities. Our research group has constructed an upper-torso humanoid robot with a pair of six degree-of-freedom arms, a three degree-of-freedom torso, and a seven degree of freedom head and neck. The robot, named Cog, has a visual system consisting of four color CCD cameras (two cameras per eye, one with a wide field of view and one with a narrow field of view at higher acuity), an auditory system consisting of two microphones, a vestibular system consisting of a three axis inertial package, and an assortment of kinesthetic sensing from encoders, potentiometers, and strain gauges. (For additional information on the robotic system, see [7]. For additional information on the reasons for building Cog, see [1, 6].) In addition to the behaviors that are presented in this section, there are also a variety of behavioral and cognitive skills that are not integral parts of the theory of mind models, but are nonetheless necessary to implement the desired functionality. We have implemented a variety of perceptual feature detectors (such as color saliency detectors, motion detectors, skin color filters, and rough disparity detectors) that match the perceptual abilities of young in- fants. We have constructed a model of human visual search and attention that was proposed by Wolfe [38]. We have also implemented motor control schemes for visual motor behaviors (including saccades, smooth-pursuit tracking, 6 Brian Scassellati and a vestibular-occular reflex), orientation movements of the head and neck, and primitive reaching movements for a six degree-of-freedom arm. We will briefly describe the relevant aspects of each of these components so that their place within the larger integrated system can be made clear. 5.1 Pre-attentive visual routines Human infants show a preference for stimuli that exhibit certain low-level feature properties. For example, a four- month-old infant is more likely to look at a moving object than a static one, or a face-like object than one that has similar, but jumbled, features [16]. To mimic the preferences of human infants, Cog's perceptual system combines three basic feature detectors: color saliency analysis, motion detection, and skin color detection. These low-level features are then filtered through an attentional mechanism before more complex post-attentive processing (such as face detection) occurs. All of these systems operate at speeds that are amenable to social interaction (30Hz). Color content is computed using an opponent-process model that identifies saturated areas of red, green, blue, and yellow [4]. Our models of color saliency are drawn from the complementary work on visual search and attention from Itti, Koch, and Niebur [22]. The incoming video stream contains three 8-bit color channels (r, g, and b) which are transformed into four color-opponency channels (r 0 , Each input color channel is first normalized by the luminance l (a weighted average of the three input color channels): r l l l (1) These normalized color channels are then used to produce four opponent-color channels: nb n kr n g n k (5) The four opponent-color channels are thresholded and smoothed to produce the output color saliency feature map. This smoothing serves both to eliminate pixel-level noise and to provide a neighborhood of influence to the output map, as proposed by Wolfe [38]. In parallel with the color saliency computations, The motion detection module uses temporal differencing and region growing to obtain bounding boxes of moving objects [5]. The incoming image is converted to grayscale and placed into a ring of frame buffers. A raw motion map is computed by passing the absolute difference between consecutive images through a threshold function This raw motion map is then smoothed to minimize point noise sources. The third pre-attentive feature detector identifies regions that have color values that are within the range of skin tones [3]. Incoming images are first filtered by a mask that identifies candidate areas as those that satisfy the following criteria on the red, green, and blue pixel components: The final weighting of each region is determined by a learned classification function that was trained on hand- classified image regions. The output is again median filtered with a small support area to minimize noise. 5.2 Visual attention Low-level perceptual inputs are combined with high-level influences from motivations and habituation effects by the attention system (see Figure 4). This system is based upon models of adult human visual search and attention [38], and has been reported previously [4]. The attention process constructs a linear combination of the input feature detectors and a time-decayed Gaussian field which represents habituation effects. High areas of activation in this composite generate a saccade to that location and compensatory neck movement. The weights of the feature detectors can be influenced by the motivational and emotional state of the robot to preferentially bias certain stimuli. For example, if the robot is searching for a playmate, the weight of the skin detector can be increased to cause the robot to show a preference for attending to faces. Frame Grabber Eye Motor Control inhibit reset Motivations, Drives and Emotions Color Detector Motion Detector Habituation Skin Detector Attention Process Fig. 4. Low-level feature detectors for skin finding, motion detection, and color saliency analysis are combined with top-down motivational influences and habituation effects by the attentional system to direct eye and neck movements. In these images, the robot has identified three salient objects: a face, a hand, and a colorful toy block. 5.3 Finding eyes and faces The first shared attention behaviors that infants engage in involve maintaining eye contact. To enable our robot to recognize and maintain eye contact, we have implemented a perceptual system capable of finding faces and eyes [35]. Our face detection techniques are designed to identify locations that are likely to contain a face, not to verify with certainty that a face is present in the image. Potential face locations are identified by the attention system as locations that have skin color and/or movement. These locations are then screened using a template-based algorithm called "ratio templates" developed by Sinha [36]. The ratio template algorithm was designed to detect frontal views of faces under varying lighting conditions, and is an extension of classical template approaches [36]. Ratio templates also offer multiple levels of biological plausibility; templates can be either hand-coded or learned adaptively from qualitative image invariants [36]. A ratio template is composed of regions and relations, as shown in Figure 5. For each target location in the grayscale peripheral image, a template comparison is performed using a special set of comparison rules. The set of regions is convolved with an image patch around a pixel location to give the average grayscale value for that region. Relations are comparisons between region values, for example, between the "left forehead" region and the "left temple" region. The relation is satisfied if the ratio of the first region to the second region exceeds a constant value (in our case, 1.1). The number of satisfied relations serves as the match score for a particular location; the more relations that are satisfied the more likely that a face is located there. In Figure 5, each arrow indicates a relation, with the head of the arrow denoting the second region (the denominator of the ratio). The ratio template algorithm has been shown to be reasonably invariant to changes in illumination and slight rotational changes [35]. Locations that pass the screening process are classified as faces and cause the robot to saccade to that target using a learned visual-motor behavior. The location of the face in peripheral image coordinates is then mapped into foveal image coordinates using a second learned mapping. The location of the face within the peripheral image can then be used to extract the sub-image containing the eye for further processing (see Figure 6). This technique has been successful at locating and extracting sub-images that contain eyes under a variety of conditions and from many different individuals. These functions match the first function of Baron-Cohen's EDD and begin to approach 8 Brian Scassellati Fig. 5. A ratio template for face detection. The template is composed of 16 regions (the gray boxes) and 23 relations (shown by arrows). Darker arrows are statistically more important in making the classification and are computed first to allow real-time rates. Fig. 6. A selection of faces and eyes identified by the robot.Faces are located in the wide-angle peripheral image. The robot then saccades to the target to obtain a high-resolution image of the eye from the narrow field-of-view camera. the second and third functions as well. We are currently extending the functionality to include interpolation of gaze direction using the decomposition proposed by Butterworth [8] (see section 6 below). 5.4 Discriminating animate from inanimate We are currently implementing a system that distinguishes between animate and inanimate visual stimuli based on the presence of self-generated motion. Similar to the findings of Leslie [25] and Cohen and Amsel [13] on the classification performed by infants, our system operates at two developmental stages. Both stages form trajectories from stimuli in consecutive image frames and attempt to maximize the path coherency. The differences between the two developmental states lies in the type of features used in tracking. At the first stage, only spatio-temporal features (resulting from object size and motion) are used as cues for tracking. In the second stage, more complex object features such as color, texture, and shape are employed. With a system for distinguishing animate from inanimate stimuli, we can begin to provide the distinctions implicit in Leslie's differences between ToBY and ToMM and the assumptions that Baron-Cohen requires for ID. Computational techniques for multi-target tracking have been used extensively in signal processing and detection domains. Our approach is based on the multiple hypothesis tracking algorithm proposed by Reid [34] and implemented by Cox and Hingorani [14]. The output of the motion detection module produces regions of motion Humanoids2000 9 and their respective centroids. These centroid locations form a stream of target locations fP 1 t g with k targets present in each frame t. The objective is to produce a labeled trajectory which consists of a set of points, one from each frame, which identify a single object in the world as it moves through the field of view: However, because the number of targets in each frame is never constant and because the existence of a target from one frame to the next is uncertain, we must introduce a mechanism to compensate for objects that enter and leave the field of view and to compensate for irregularities in the earlier processing modules. To address these problems, we introduce phantom points that have undefined locations within the image plane but which can be used to complete trajectories for objects that enter, exit, or are occluded within the visual field. As each new point is introduced, a set of hypotheses linking that point to prior trajectories are generated. These hypotheses include representations for false alarms, non-detection events, extensions of prior trajectories, and beginnings of new trajectories. The set of all hypotheses are pruned at each time step based on statistical models of the system noise levels and based on the similarity between detected targets. This similarity measurement is based either purely on distances between points in the visual field (a condition that represents the first developmental stage described above) or on similarities of object features such as color content, size, visual moments, or rough spatial distribution (a condition that reflects a sensitivity to object properties characteristic of the second developmental stage). At any point, the system maintains a small set of overlapping hypotheses so that future data may be used to disambiguate the scene. Of course, the system can also produce the set of non-overlapping hypotheses that are statstically most likely. We are currently developing metrics for evaluating these trajectories in order to classify the stimulus as either animate or inanimate using the descriptions of Michotte's [28] observations of adults and Leslie's [25] observations of infants. The general form of these observations indicate that self-generated movement is attributed to stimuli whose velocity profiles change in a non-constant manner, that is, animate objects can change their directions and speed while inanimate objects tend to follow a single acceleration unless acted upon by another object. 6 Ongoing Work The systems that have been implemented so far have only begun to address the issues raised by Leslie's and Baron- Cohen's models of theory of mind. In this section, three current research directions are discussed: the implementation of gaze following; the extensions of gaze following to deictic gestures; and the extension of animate-inanimate distinctions to more complex spatio-temporal relations such as support and self-recognition. 6.1 Implementing gaze following Once a system is capable of detecting eye contact, three additional subskills are required for gaze following: extracting the angle of gaze, extrapolating the angle of gaze to a distal object, and motor routines for alternating between the distal object and the caregiver. Extracting angle of gaze is a generalization of detecting someone gazing at you, but requires additional competencies. By a geometric analysis of this task, we would need to determine not only the angle of gaze, but also the degree of vergence of the observer's eyes to find the distal object. However, the ontogeny of gaze following in human children demonstrates a simpler strategy. Butterworth [8] has shown that at approximately 6 months, infants will begin to follow a caregiver's gaze to the correct side of the body, that is, the child can distinguish between the caregiver looking to the left and the caregiver looking to the right (see Figure 7). Over the next three months, their accuracy increases so that they can roughly determine the angle of gaze. At 9 months, the child will track from the caregiver's eyes along the angle of gaze until a salient object is encountered. Even if the actual object of attention is further along the angle of gaze, the child is somehow "stuck" on the first object encountered along that path. Butterworth labels this the "ecological" mechanism of joint visual attention, since it is the nature of the environment itself that completes the action. It is not until 12 months that the child will reliably attend to the distal object regardless of its order in the scan path. This "geometric" stage indicates that the infant can successfully determine not only the angle of gaze but also the vergence. However, even at this stage, infants will only exhibit gaze following if the distal object is within their field of view. They will not turn to look behind them, even if the angle of gaze from the caregiver would warrant such an action. Around months, the infant begins to enter a "representational" stage in which it will follow gaze angles outside its own field of view, that is, it somehow represents the angle of gaze and the presence of objects outside its own view. months: Representational stage 6 months: Sensitivity to field 9 months: Ecological stage months: Geometric stage Fig. 7. Proposed developmental progression of gaze following adapted from Butterworth (1991). At 6 months, infants show sensitivity only to the side that the caregiver is gazing. At 9 months, infants show a particular strategy of scanning along the line of gaze for salient objects. By one year, the child can recognize the vergence of the caregiver's eyes to localize the distal target, but will not orient if that object is outside the field of view until months of age. Implementing this progression for a robotic system provides a simple means of bootstrapping behaviors. The capabilities used in detecting and maintaining eye contact can be extended to provide a rough angle of gaze. By tracking along this angle of gaze, and watching for objects that have salient color, intensity, or motion, we can mimic the ecological strategy. From an ecological mechanism, we can refine the algorithms for determining gaze and add mechanisms for determining vergence. Once the robot and the caregiver are attending to the same object, the robot can observe both the vergence of its own eyes (to achieve a sense of distance to the caregiver and to the target) and the pupil locations (and thus the vergence) of the caregiver's eyes. A rough geometric strategy can then be implemented, and later refined through feedback from the caregiver. A representational strategy will require the ability to maintain information on salient objects that are outside of the field of view including information on their appearance, location, size, and salient properties. 6.2 Extensions of gaze following to deictic gestures Although Baron-Cohen's model focuses on the social aspects of gaze (primarily since they are the first to develop in children), there are other gestural cues that serve as shared attention mechanisms. After gaze following, the next most obvious is the development of imperative and declarative pointing. Imperative pointing is a gesture used to obtain an object that is out of reach by pointing at that object. This behavior is first seen in human children at about nine months of age, and occurs in many monkeys [11]. However, there is nothing particular to the infant's behavior that is different from a simple reach - the infant is initially as likely to perform imperative pointing when the caregiver is attending to the infant as when the caregiver is looking in the other direction or when the caregiver is not present. The caregiver's interpretation of infant's gesture provides the shared meaning. Over time, the infant learns when the gesture is appropriate. One can imagine the child learning this behavior through simple reinforcement. The reaching motion of the infant is interpreted by the adult as a request for a specific object, which the adult then acquires and provides to the child. The acquisition of the desired object serves as positive reinforcement for the contextual setting that preceded the reward (the reaching action in the presence of the attentive caregiver). Generation of this behavior is then a simple extension of a primitive reaching behavior. Declarative pointing is characterized by an extended arm and index finger designed to draw attention to a distal object. Unlike imperative pointing, it is not necessarily a request for an object; children often use declarative pointing to draw attention to objects that are clearly outside their reach, such as the sun or an airplane passing overhead. Declarative pointing also only occurs under specific social conditions; children do not point unless there is someone to observe their action. I propose that imitation is a critical factor in the ontogeny of declarative pointing. This is an appealing speculation from both an ontological and a phylogenetic standpoint. From an ontological perspective, declarative pointing begins to emerge at approximately 12 months in human infants, which is also the same time that other complex imitative behaviors such as pretend play begin to emerge. From the phylogenetic Humanoids2000 11 perspective, declarative pointing has not been identified in any non-human primate [33]. This also corresponds to the phylogeny of imitation; no non-human primate has ever been documented to display imitative behavior under general conditions [21]. I propose that the child first learns to recognize the declarative pointing gestures of the adult and then imitates those gestures in order to produce declarative pointing. The recognition of pointing gestures builds upon the competencies of gaze following and imperative pointing; the infrastructure for extrapolation from a body cue is already present from gaze following, it need only be applied to a new domain. The generation of declarative pointing gestures requires the same motor capabilities as imperative pointing, but it must be utilized in specific social circumstances. By imitating the successful pointing gestures of other individuals, the child can learn to make use of similar gestures. 6.3 Extensions of animate-inanimate distinctions The simple spatio-temporal criteria for distinguishing animate from inanimate has many obvious flaws. We are currently attempting to outline potential extensions for this model. One necessary extension is the consideration of tracking over longer time scales (on the order of tens of minutes) to allow processing of a more continuous object identity. This will also allow for processing to remove a subset of repetitively moving objects that are currently incorrectly classified as animate (such as would be caused by a tree moving in the wind). A second set of extensions would be to learn more complex forms of causal structures for physical objects, such as the understanding of gravity and support relationships. This developmental advance may be strongly tied to the object concept and the physical laws of spatial occupancy [20]. Finally, more complex object properties such as shape features and color should be used to add a level of robustness to the multi-target tracking. Kalman filters have been used to track complex features that gradually change over time [14]. 7 Conclusion While theory of mind studies have been more in the realm of philosophy than the realm of robotics, the requirements of humanoid robotics for building systems that can interact socially with people will require a focus on the issues that theory of mind research has addressed. Both Baron-Cohen and Leslie have provided models of how more complex social skills can be developmentally constructed from simpler sensory-motor skill sets. While neither model is exactly suited for a robotic implementation, they do show promise for providing the basis of such an implementation. I have presented one initial attempt at building a framework of these precursors to a theory of mind, but certainly much more work is required. However, the possibility of a robotic implementation also raises the questions of the use of such an implementation as a tool for evaluating the predictive power and validity of those models. Having an implementation of a developmental model on a robot would allow detailed and controlled manipulations of the model while maintaining the same testing environment and methodology used on human subjects. Internal model parameters could be varied systematically as the effects of different environmental conditions on each step in development are evaluated. Because the robot brings the model into the same environment as a human subject, similar evaluation criteria can be used (whether subjective measurements from observers or quantitative measurements such as reaction time or accuracy). Further, a robotic model can also be subjected to controversial testing that is potentially hazardous, costly, or unethical to conduct on humans. While this possibility does raise a host of new questions and issues, it is a possibility worthy of further consideration. --R The Cog project. Mindblindness. MIT Press Social constraints on animate vision. A context-dependent attention system for a social robot The Cog project: Building a humanoid robot. The ontogeny and phylogeny of joint visual attention. Machiavellian Intelligence: Social Expertise and the Evolution of Intellect in Monkeys Sources of conceptual change. How Monkeys See the World. Reading minds or reading behavior? Precursors to infants' perception of the causality of a simple event. An efficient implementation of Reid's multiple hypothesis tracking algorithm and its evaluation for the purpose of visual tracking. The Intentional Stance. Infants' recognition of invariant features of faces. A theory of the child's theory of mind. Interacting minds - a biological basis First principles organize attention to and learning about relevant data: number and the animate-inanimate distinction as examples Building a cognitive creature from a set of primitives: Evolutionary and developmental insights. Evolution of Communication. A model of saliency-based visual attention for rapid scene analysis Machine Vision. Is there a social module? The perception of causality in infants. Spatiotemporal continuity and the perception of causality in infants. The perception of causality. The theoretical implications of joint attention deficits in autism. The role of features in preattentive vision: Comparison of orientation Development of theory of mind and executive control. Theory of mind: evolutionary history of a cognitive specialization. "Does the chimpanzee have a theory of mind?" An algorithm for tracking multiple targets. Finding eyes and faces with a foveated vision system. PhD thesis Beliefs about beliefs: Representation and constraining function of wrong beliefs in young children's understanding of deception. Guided search 2.0: A revised model of visual search. --TR --CTR Jos M. Buenaposada , Luis Baumela, A computer vision based human-robot interface, Autonomous robotic systems: soft computing and hard computing methodologies and applications, Physica-Verlag GmbH, Heidelberg, Germany, Deb Roy , Kai-Yuh Hsiao , Nikolaos Mavridis, Conversational robots: building blocks for grounding word meaning, Proceedings of the HLT-NAACL workshop on Learning word meaning from non-linguistic data, p.70-77, May 31, Christof Teuscher , Jochen Triesch, To each his own: The caregiver's role in a computational model of gaze following, Neurocomputing, v.70 n.13-15, p.2166-2180, August, 2007 Jochen Triesch , Hector Jasso , Gedeon O. Dek, Emergence of Mirror Neurons in a Model of Gaze Following, Adaptive Behavior - Animals, Animats, Software Agents, Robots, Adaptive Systems, v.15 n.2, p.149-165, June 2007 Cynthia Breazeal , Daphna Buchsbaum , Jesse Gray , David Gatenby , Bruce Blumberg, Learning From and About Others: Towards Using Imitation to Bootstrap the Social Understanding of Others by Robots, Artificial Life, v.11 n.1, p.31-62, January 2005
social interaction;humanoid robots;visual perception
591962
Searching for a Global Search Algorithm.
We report on a case study to assess the use of an advanced knowledge-based software design technique with programmers who have not participated in the techniques development. We use the KIDS approach to algorithm design to construct two global search algorithms that route baggage through a transportation net. Construction of the second algorithm involves extending the KIDS knowledge base. Experience with the case study leads us to integrate the approach with the spiral and prototyping models of software engineering, and to discuss ways to deal with incomplete design knowledge.
Introduction Advanced techniques to support software construction will only be widely accepted by practitioners if they can be successfully used by software engineers who were not involved in their development and did not get on-site training by their inventors. Experience has to be gained how knowledge-based methods can be integrated into the practical software engineering process. We report on experience made with the application of the approach to algorithm design underlying the Kestrel Interactive Development System (KIDS) [14] to the construction of control software for a simplified baggage transportation system at an airport. In this paper, we use the term "KIDS approach" to denote the concepts that have been implemented in the system The KIDS approach has been applied to a number of case studies at Kestrel Institute. In par- ticular, it has been used in the design of a transportation scheduling algorithm with impressive performance [15, 16]. We wished to find out if we were able to use this method based on the available publi- We did not use the implemented system KIDS in the case study. cations and produce satisfactory results with reason-able effort. A second goal of this work has been to study how a knowledge-based approach can be integrated into the overall software engineering process. As a case study we chose a non-trivial abstraction of a practically relevant problem to make our experience transferable to realistic applications. In the following, we elaborate on two issues: a process model we found useful to support application of the KIDS approach, and the merits and shortcomings we encountered when we explored alternative solutions to the transportation scheduling problem. We have integrated the spiral and prototyping models of software engineering [2] with the KIDS approach. We developed the first formal specification and a prototype implementation in parallel. The prototype served to validate the specification and to improve understanding of the problem domain. In the KIDS approach, global search algorithms are constructed by specializing global search theories that abstractly describe the shape of the search tree set up by the algorithm. For the case study, we wished to explore two alternative search ideas. While we found a theory suitable for the first one in the literature, the second one could not be realized with the documented design knowledge. This lead us to develop a new global search theory that needs a slightly modified specialization procedure. In Section 2, we introduce the baggage transportation problem. Section 3 provides a brief review of the global search theory and the KIDS approach. We present its integration into a process model in Section 4. The design of two transportation schedulers is described in Section 5. Optimizations are sketched in Section 6, where we also discuss the resulting algo- rithms. We summarize our experience with the approach in Section 7. Baggage Transportation Scheduling We wish to develop a controller for the transportation of baggage at an airport. Pieces of baggage are transported from check-in counters to gates, from gates to other gates, or from gates to baggage delivery points. The controller must schedule the baggage through the network in such a way that each piece arrives at its destination in due time. To simplify the problem, we do not consider on-line scheduling of a continuous flow of baggage fed into the system at the check-in counters, but schedule all baggage checked-in at a particular point in time. 2.1 Domain Model We model the transportation net as a directed graph as shown in Figure 1. Check-in counters and baggage delivery counters, gates and switches are represented by nodes. We classify these in three kinds: input nodes, transportation nodes and output nodes. Check-in counters correspond to input nodes, switches to transportation nodes and baggage delivery counters to output nodes. Since gates serve to load and unload airplanes, we represent them by an input and an output node. The edges of the graph model conveyor belts. The capacity of a belt is the maximum amount of baggage ("total weight") it can carry at a time. Its velocity is the time it takes to carry a piece from the start to the node. gate G2 gate G1 gate G3 gate G4 delivery D1 delivery D2 checkin C1 checkin C2 ceckin C3 checkin C4 checkin C5 Figure 1: Transportation network A piece of baggage is described by its weight, source and destination nodes, and its due time. Source and destination are input and output nodes, respectively. Weight and due time are positive natural numbers. Due times are specified relative to the beginning of the transportation process. 2.2 Problem Specification Our task is to assign, to each piece of baggage, a route through the network leading from its source to its destination node in due time. To keep things sim- ple, we require an acyclic network without depots at Function transport plan( g : graph where acyclic(g)- returns dom(q):feasible path(g; b; snd(q(b)))g)) Figure 2: Problem Specification the transportation nodes. Thus, the only way to re-solve scheduling conflicts that arise if capacities of conveyor belts are exceeded is to delay baggage at source nodes. A route therefore is a pair of a delay and a path through the network. A plan maps pieces of baggage to routes. Attempting to find a plan for a particular set of baggage makes sense only if there exists a feasible path for each piece. This is a path p through the network g leading from the source to the destination nodes of a piece of baggage b. feasible path(g; b; p) () is path(g; p) - (1) Note that we do not require a punctual schedule to exist. We wish to find a plan in every case, punctual and without delaying baggage at input nodes if possible. Thus, we define a cost function based on the criteria if all baggage is delivered in time and if baggage is delayed at input nodes. cost all baggage in time no delays 1 yes no no yes 3 no no For example, imagine we have a suitcase b 1 at gate and another one b 2 at C 4 in Figure 1. Both have weight 1. They are checked in for the same flight, starting at gate G 2 . Let the transportation time of each belt be one time unit and its capacity also be one unit. As we have to avoid exceeding the capacity of the belt leading to G 2 , a solution is to delay b 1 by one time unit. This gives us the transportation plan: Now we can set up the problem specification as shown in Figure 2. For each acyclic graph g and set of baggage bs with feasible paths through the graph, we are interested in plans q such that all baggage is scheduled dom(q)), the total amount of baggage on a belt at any time does not exceed its capacity (capacity bounded), and all assigned paths are feasi- ble. From this set, we select a plan p with minimal cost. 3 Design of Global Search algorithms The basic idea of algorithm design in the KIDS approach [12, 14] is to represent design knowledge in design theories. Such a theory is a logical characterization of problems that can be solved by an algorithmic paradigm like "divide and conquer" or "global search". Algorithm design consists of showing that a given problem is an instance of some design theory. In the following, we briefly summarize how global search algorithms are developed in the KIDS approach. For a full account, we refer the reader to [13, 14]. 3.1 Design Theory A problem specification is a quadruple hD; R; I; Oi where D is the input domain and R is the output range of the function f to synthesize. The predicate I : D ! Bool describes the admissible inputs and O describes the input/output behavior of f . Hence, f : D ! R is a solution to P design theory extends a problem specification by additional functions. It states properties of these functions sufficient to formulate a schematic algorithm that correctly solves the problem. The global search theory of Figure 3 describes the search for an optimal solution with respect to a cost function c. The basic idea is to split search spaces containing candidate solutions into "smaller" ones until solutions are directly extractable. The R is the type of search space descriptors, - I defines legal descriptors. For an input x, - r 0 and Split describe the search tree for solutions z with O(x; z): its root is - r 0 (x), the initial search space; a descendent relation on nodes is given by Split: if - s is a (direct) subspace of - r for an input x. The solutions obtainable by looking at a single node - r of the search tree are described by Extract(z; - r). By axiom GS3, Satisfies(z; - r) describes the solutions z contained in a search space - r that can be found with finite effort. There must exist a finite path in the search tree from - r to a search space - s from which z can be extracted. Split is defined by Since we wish to find a globally optimal solution, GS2 requires that all feasible solutions are contained in Sorts D;R; - Operations R Satisfies R \Theta - Extract Axioms GSC: - is total ordering on C Function where I(x) returns Function returns Figure 3: Global search theory and algorithm schema the initial search space. Axioms GS0 and GS1 ensure that all considered search spaces are legal. The program shown on the bottom of Figure 3 provides a schematic algorithm consistent with the global search theory. The function F computes an optimal solution z for a given input x by initiating a global search in the initial search space - (x). The actual search algorithm is implemented in f gs. It minimizes over the solutions that are directly extractable from the input search space - r and the ones found by recursive search in spaces - s obtained by splitting - r. Necessary filters provide the basis to optimize the code gained by instantiating the algorithm schema. A necessary filter \Phi is used to prune branches of the search tree that cannot contain solutions. It satisfies the implication r) (2) 3.2 Algorithm Design How can we find a global search algorithm for a given problem specification? We have to find a search space description - R and operations - and Extract such that the global search axioms are ful- filled. In the KIDS approach this is done by referring to knowledge about search strategies on concrete data structures that is formalized in a library of general global search theories 2 . Examples are theories to enumerate all sequences over a finite set and to enumerate all mappings between finite sets. A global search theory for a given problem is constructed by specializing a theory from the library. A problem theory specializes to a problem theory Constructively verifying the existential quantifier in (3) gives us a systematic way to find a global search theory for the problem at hand. In this way, the structure of the search tree expressed in A is adapted to B and gives us an algorithm for B that is an instance of the algorithm schema in Figure 3. In general this algorithm is very inefficient, but has a high potential for optimization which can be exploited by deriving necessary filters, by program transformation, and by data structure refinement. 4 A Process Model Our presentation of the application domain theory and problem specification in Section 2 only describes the final result of the specification effort. To develop the domain theory is one of the major tasks if not the most complex and time consuming one in the KIDS approach. Much of its complexity stems from two requirements we demand of the domain theory: it must not only make precise the informal - usually incomplete and sometimes inconsistent - ideas about the nature and context of the problem, but it must also be formulated so as to aid and not impede the subsequent design process. As a consequence, it is very unlikely that a satisfying domain theory can be built from scratch. This observation led us to integrate the KIDS approach with the prototyping and spiral models of software engineering [2]. One cycle of development, as sketched in Figure 4, has three phases. The first is concerned with establishing or enhancing the domain theory, the second produces code, and in the third phase code and theory are tested and validated. We found it useful to build the first draft of the domain theory in parallel with a prototype. In this early phase, shaded gray in Figure 4, the domain theory is not rich enough to apply algorithm design knowledge from design theories. Building a prototype enables us to get a deeper understanding of the problem and the We used Appendix A of [13] code prototype validation / test domain theory algorithm design optimization specification validation theory validation domain application test Figure 4: Process model essential properties of the application area. It helps us to build a complete domain theory and to avoid dead-end developments. The way in which the domain model is expressed, the data structures used, and the properties stated, can have much influence on the ease with which algorithm design can be carried out. Thus, what seems to be one cycle of design in Figure 4, may in practice require several rounds of refining the domain theory until the formalized notions smoothly fit with the design theory we wish to use. One example from the baggage transportation case study is the way we modeled delays in routes (cf. Section 2.2). In an early version of the domain theory, we described them by repetitions of the input nodes in paths, each occurrence of the input node denoting a delay of one time unit. This forced us to introduce predicates to characterize legal routes, and we could not use an acyclic graph model. When we decided to reformulate the theory and make delays explicit the theory became much more elegant and further design was much easier. The process of theory refinement perpetuates as we derive filters and optimize code. The validation and test phases also serve us to validate the code with respect to properties that are not captured by the design knowledge put to our disposal in the KIDS approach. 5 Two Ways to Find Transportation Plans Looking at the sort of transportation plans, nat \Theta seq(vertex)) suggests two strategies to search for solutions to our scheduling problem. 1. Domain Extension. Start with the initially empty map and successively extend it by assigning possibly delayed feasible paths to baggage. 2. Image Modification. Start with the map that assigns their source nodes and no delay to all bag- successively modify the assigned routes by extending paths or increasing delays. Both strategies enumerate all feasible transportation plans. In the KIDS approach, search strategies are provided in a library of general global search the- ories. Algorithm design proceeds by specializing one of these to the problem at hand. The first condition in (3), RB ' RA , suggests to match the output domains of the problem specification with the ones of the library theories to find candidates to specialize. When we began algorithm design for the transportation problem, our initial idea was to use the image modification strategy, but there is no general global search theory documented in the KIDS library [13] that models image modification. Instead, we found a theory that describes domain extension. This motivated us to explore both approaches. 5.1 Domain Extension The global search theory gs finite mappings enumerates all maps from a finite set U to a finite set . F 7! gs finite mappings R 7! map(ff; fi) I 7! -hU; V i:jU O 7! -hU; V R 7! set(ff) \Theta set(ff) \Theta map(ff; fi) I 7! -hU; V Satisfies 7! Extract 7! A search space is described by a partition of U into two sets S and T , and by a map M from S to V . M can be completed to a map from U to V by assigning elements of V to all elements of T . Split performs one of these assignments: it picks arbitrary elements a and b of T and V , respectively, and extends M by a 7! b. We find instantiations for the type variables in gs finite mappings by unifying its output domain with the one of transport plan. ff 7! baggage route We can specialize gs finite mappings to transport plan if we can constructively verify (3) for these theories, i.e. we must find expressions in and bs for U and V to prove dom(M):feasible path(g; b; snd(M(b))) The predicate Map is defined by Comparing the right hand side of this definition with and the definition (1) of feasible paths suggests to use the sets U 7! bs to specialize gs finite mappings. The use of the upper bound md(g; bs) on delays makes the set of routes finite, and thus ensures termination of the algorithm. We discuss termination further in Section 7. Since there are feasible paths for all baggage in bs (cf. the precondition of transport plan in Figure 2), we can assign to md(g; bs) the sum of the times needed to traverse a feasible path for each piece of baggage. Applying the substitution for ff, fi, U , and V to gs finite mappings gives us a global search theory for transport plan. R 7! set(baggage) \Theta set(baggage) \Theta plan I 7!-hg; Satisfies 7! Extract 7! The resulting search strategy assigns complete routes to one piece of baggage after another. Without further optimization, Split assigns arbitrary routes to pieces of baggage and only when a complete plan can be extracted is tested whether the assigned routes are feasible. An obvious way to prevent infeasible assignments in the first place is to develop a necessary filter. Instantiation of (2) and the fact that capacity bounded is monotonic in domain extensions of M gives us 5.2 Image Modification There is no global search theory documented in [11, 13, 14] that supports searching for maps by image modification. So we developed a new theory for this purpose. Abstracting from the concrete scheduling problem, the image modification strategy can be sketched as follows: The images of a given map (the initial sched- ule) are increased along the various degrees of freedom that are given by the range type of the map. A suitable successor relation on the elements of the range type can be used to describe the "direction" in which to increase the images of the map. This idea is formalized in gs parallel mappings 3 . F 7! gs parallel mappings D 7! map(ff; fi) \Theta set(fi \Theta fi) R 7! map(ff; fi) I 7!-hM;Si: O 7!-hM;Si; N: R 7! map(ff; fi) \Theta set(fi \Theta fi) I Satisfies 7! -N; hM;Si:(8x 2 dom(M):M(x) S N(x)) Extract 7! -N; The inputs are a map M and a successor relation S. The search enumerates all maps with the same domain as M and images that are extended "along" S, i.e. they are in the reflexive and transitive closure S of S. The domain of M must be finite, and S must be non-dense to ensure that GS3 holds. The relation S is anti-reflexive to ensure progress in the search. Note that the search spaces (of sort - R) do not only contain the map built so far but also the relation S. This is necessary to describe (in Satisfies) if a solution is contained in a search space. 5.2.1 Data Type Driven Specialization Having captured our search idea in gs parallel mappings, we want to specialize this theory to the transportation planning problem. It turns out that the corresponding instance of (3) does not help much in systematically finding M and S, because we cannot refer to the structure of the 3 x S y is a notation for hx; yi 2S. range type fi in gs parallel mappings. Still, it is not desirable to develop a global search theory for the specific structure of the range type of the transport schedules because the newly developed theory should be sufficiently abstract so as to be applicable to a broad range of problems. To solve this dilemma, we propose to specialize gs parallel mappings in two steps. The first step determines a suitable successor relation S while the second step finds a substitution for M . To determine S, we first find substitutions for the type variables ff and fi. For the transportation problem we get ff 7! baggage Now we can analyze the range type nat \Theta seq(vertex) to find a canonical successor relation on its elements based the basic types it is composed of. We know the usual successor function on natural numbers, and a canonical way to extend sequences is to append an element. In analogy to lexicographical orderings on pairs, we construct a successor relation by extending either element of a pair. Thus, we define S by hn; We substitute (5) for S in (3). After simplification we get capacity As in Section 5.1, we can now easily determine a substitution for M and get the map that assigns to each b in bs the non-delayed path only consisting of the source node of b. Like the algorithm of Section 5.1, the one working with image modification has a high potential for optimization and its development proceeds by filter construction and optimization. 6 Optimization and Results The algorithms resulting from the instantiation of the scheme in Figure 3 are so inefficient that optimization is absolutely necessary. The efficiency of the resulting code heavily depends on the adequate choice of filters and program transformations. Important optimizations on the resulting algorithm are the introduction of a priority queue on search spaces and simplifications on often used predicates, such as capacity bounded. We have also eliminated common subexpressions and introduced an analysis of the transportation net with respect to the input baggage to eliminate vertices that do not lie on feasible paths. The ordering on search spaces used in the priority queue encodes a heuristic to determine which node of the search tree to consider next. We have used one based on cost, total delay, and length of the paths in a plan. Test runs of the algorithms show that, despite the optimizations, performance of both is still poor. They also reveal that the algorithm based on image modification is about a factor of two faster than the one based on domain extension. The size of theories and programs are summarized in Table 1. Document Lines Library of Specifications 490 Domain Theory 240 Algorithm Theory 350 Implementation 920 Table 1: Size of theories and programs All specifications are written in the specification language SPECTRUM [3]. We use an existing library of basic spectrum specifications such as natural numbers, sets, sequences, directed graphs and others. Based on these modules, the domain theory comprises about 240 lines. The algorithms extend the domain theory by 110 lines. Both schedulers are implemented in the functional language OPAL [10]. The translation of specifications into executable code and its optimization produces about 920 lines of code for the OPAL program. The code is well-structured and highly reusable. Both implementations share most of the code which facilitates exploring alternatives. The case study required an effort of approximately 9 person months. We spent about one third of that time to learn the KIDS approach. Approximately 75% of the remaining time were devoted to building the domain theory. There exists a number of approaches to algorithm design and program synthesis, e.g. [1, 4, 8, 9]. We have chosen the KIDS approach for our case study because it provides design steps that reflect significant design decisions of programmers and are described precisely by logical theories. In this way, they are good reference points for software engineers who wish to learn and use them. However, we did not use the KIDS system because we wanted to have full control over the design process and adapt it to our needs if necessary. Transportation scheduling. Our case study relates to the research on design of transportation schedulers at Kestrel [15, 16]. They study schedulers that assign trips to resources like planes, ships, and trucks to meet movement requirements. In this setting, trips fully occupy resources for an interval of time, i.e. the load of a resource cannot be extended during a trip. Furthermore, a trip changes the availability of a re- source: the destination of one trip becomes the source of the next one. In baggage transportation, how- ever, load of resources can continually change as baggage flows through the net, but source and destination points of a resource remain fixed in time. Another difference lies in the focus of our work. For several years, a highly specialized theory on transportation scheduling has been developed at Kestrel with the aim to produce extremely efficient schedulers. Recently, this has even led to a refinement of the abstract global search theory [16]. The purpose of our case study, in contrast, has been to study in how far the KIDS approach as documented in the literature can support programmers who have no particular experience with the approach, to design algorithms for a non-trivial problem. Process model. The steps in designing a global search algorithm: specializing a theory, deriving fil- ters, and applying optimizing program transforma- tions, provide a clear separation of concerns during design. Specialization determines the basic structure of the search, necessary filters exploit properties of the application domain, and only the final program transformations and data type refinements eliminate redundancies in the code and "fuse" filters with the basic search structure to gain efficiency. Each of these tasks corresponds to one cycle in the process model that we introduced in Section 4. Thus the model helps programmers to focus activities on a particular task and to avoid introducing certain design ideas at the "wrong" time into the development. In early attempts to design the algorithm of Section 5.2, we tried to introduce optimizations too early - trying to generate delayed routes only if necessary - which totally messed up our design. The first phase of the development, before a sufficiently complete application domain theory is avail- able, is the most complex part of the process. We found prototyping useful to understand the problem domain, but techniques to guide theory development remain to be established. With a domain theory at hand, the KIDS approach is well suited to construct prototypes in little time. This supports exploring alternative designs. Termination. In the global search theory we have used, the issue of termination of the constructed algorithms is not addressed. This lead us to the somewhat unnatural introduction of the upper bound md(g; bs) on delays (cf. Section 5). Termination of global search algorithms can be spoiled in two ways. There may be branches of the search tree with infinite length, or there may be nodes with infinitely many children. In [14], a well-founded ordering is introduced into the abstract global search theory to prevent infinite chains of Split-operations. Kreitz [7] has formalized global search in the Nuprl type theory [4]. He prevents infinite branchings of the search tree by using only finite sets in his for- malization. He introduces wf-filters to prune infinite branches and proposes to provide a collection of wf- filters for each theory. There are methodological reasons not to require termination of all global search theories. Both theories used in Section 5 enumerate an infinite number of maps. Still, we would appreciate a systematic way that relieves programmers of dealing with termination on-the-fly. Dealing with incomplete design knowledge. As we have seen in Section 5.2, it is not unlikely that the knowledge expressed in design theories fails to support a particular design idea. Although we are not aware of systematic support for constructing new theories in KIDS, it is still worthwhile to stick to the approach and develop a new design theory that describes the desired search strategy in an abstract way. In [5], we decided to construct the problem specific algorithm theory of Section 5.2 in one step and to manually verify it against the abstract global search theory. This decision was mainly due to lack of experience and increased the complexity of the task considerably. More- over, it led to a less efficient algorithm. It seems to be unlikely to find "practically com- plete" knowledge bases for software construction sys- tems. Such systems should be designed to ease routine extension of their knowledge bases. In [6], a generic system architecture based on the notion of strategies is proposed. Strategy modules have a clearly defined interface to the system kernel, so new ones can be integrated into the system in a routine way. The system Specware [17] under development at Kestrel also seems to allow for a modularized and easily extendible knowledge base. Constructing a new global search theory is a non-trivial task and deserves support if the approach shall be applied routinely. A starting point may be the observation that search strategies often seem to derive from the structure of the output domain R. Acknowledgement . We would like to thank David Basin, Maritta Heisel and Burkhart Wolff for fruitful discussions. Klaus Didrich and Maritta provided comments on a draft of this paper. --R Artificial Intelligence A spiral model of software development and enhancement. The requirement and design specification language Spectrum. Implementing Mathematics with the Nuprl Proof Development System. Eine Fallstudie zur Entwicklung korrek- ter Software: Steuerung einer Gep-ackf-orderanlage Tool support for formal software development: A generic ar- chitecture Deriving Programs that Develop Programs. Automating Software Design. A deductive approach to program synthesis. The programming language OPAL. Kestrel Interactive Development System Algorithm theories and design tactics. Structure and design of global search al- gorithms KIDS: A semiautomatic program development system. Transformational approach to transportation scheduling. Synthesis of high-performance transportation schedulers Specware: formal support for composing software. --TR
program synthesis;KIDS;scheduling;formal methods
592010
Extending Design Environments to Software Architecture Design.
Designing a complex software system is a cognitively challenging task; thus, designers need cognitive support to create good designs. Domain-oriented design environments are cooperative problem-solving systems that support designers in complex design tasks. In this paper we present the architecture and facilities of Argo, a domain-oriented design environment for software architecture. Argos own architecture is motivated by the desire to achieve reuse and extensibility of the design environment. It separates domain-neutral code from domain-oriented code, which is distributed among active design materials as opposed to being centralized in the design environment. Argos facilities are motivated by the observed cognitive needs of designers. These facilities extend previous work in design environments by enhancing support for reflection-in-action, and adding new support for opportunistic design and comprehension and problem solving.
Figure 1.Design environment facilities of Janus, adapted from Fischer, 1994 . nizational guidelines, and the opinions of fellow project stakeholders and domain experts. Design environments suchasFramer Lemke and Fischer, 1990 , Janus Fischer, 1994 , Hydra Fischer et al., 1993 , and VDDE Voice Dialog Design Environment Sumner, Bonnardel, and Kallak, 1997 support re ection-in-action. Figure 1 shows facilities of this family of design environments. The domain-oriented construction kit facility allows users to visualize and manipulate a design. The construction analyzer facility critiques the design to give design feedback that is linked to hypertext argumentation. The goal speci cation facility helps to keep critics relevanttothe designer's objectives. Re ection-in-action is also supported by simulation facilities that allow what-if analysis as a further design evaluation. A catalog of example designs can be accessed via the catalog explorer facility. Designers will gain the most from design feedback that is both timely and relevant to their current design task. Design environments can address timeliness by linking critics to a model of the design process. For instance, Framer uses a checklist to model the process of designing a user interface window. At a given time the designer works on one checklist item and only critics relevant to that item are active. Design environments can address relevance by linking critics to speci cations of design goals. For instance, Janus and Hydra allow the designer to specify goals for kitchen oorplans, and thus activate only those critics relevant to stated design goals. Furthermore, Hydra uses critiquing perspectives i.e., explicit critiquing modes to activate critics relevanttoany given set of design issues and deactivate irrelevant critics. 2.2. Architectural Styles Work on software architecture Perry and Wolf, 1992 has focused on representing systems as composed of software components and connectors Garlan and Shaw, 1993 . Architectural styles constrain and inform architectural design by de ning the types of components and connectors available and the ways in which they may be combined Abowd, Allen, and Garlan, 1993 . Styles can be expressed as a set of style rules. A simple architectural style is pipe-and- lter which de nes components to be batch processes with standard input and output streams, and connectors to be data pipes. One pipe-and- lter style rule is that the architecture should contain no cycles. Argo supports the C2 architectural style Taylor et al., 1996 . C2 is a component- and message-based style designed to model applications that have a graphical user interface. The style emphasizes reuse of UI components such as dialogs, structured graphics models, and constraint managers Medvidovic, Oreizy, and Taylor, 1997 . The C2 style can be informally summarized as a layered network of concurrent components that communicate via message broadcast buses. Components may only send messages requesting operations upward, and noti cations of state changes downward. Buses broadcast messages sent from one component to all components in the next higher or lower layer. Each component has a top and bottom interface. The top interface of a component speci es the noti cations that it handles, and the requests it emits upward. The bottom interface of a component speci es the noti cations that it emits downward, and the requests it handles. 2.3. Software Architecture Design Tools Design tools in the domain of software architecture have tended to emphasize analysis of well-formedness and code generation. The Aesop system allows for a style- speci c design tool to be generated from a speci cation of the style Garlan, Allen, and Ockerbloom, 1994 . The DaTE system allows for construction of a running system from an architectural description and a set of reusable software components Batory and O'Malley, 1992 . Although not a software architecture tool, AMPHION is similar in that it allows users to enter a graphical speci cation from which the system can generate a running program Lowry et al., 1994 . Eachof these systems provides support for design representation, manipulation, transforma- tion, and analysis, but none of them explicitly supports architects' cognitive needs. Argo can generate code to combine software components into a running system; however, the main contribution of Argo to the software architecture communityis its emphasis on cognitive needs. KBSA ADM Benner, 1996 is a software design environment that embodies the results of many research projects stemming from a seminal vision of knowledge-based software development support Green et al., 1983 . KBSA ADM has many features in common with Argo, including critics, a to do" list, multiple coordinated models of the system under design, and process modeling. KBSA ADM is intended to package previous research results into a full-featured software developmentenvi- ronment. In contrast, Argo is intended to explore possible features that explicitly support identi ed cognitive needs. Support for cognitive needs in both KBSA ADM and Argo is inspired by previous work in design environments, however we believe that Argo has a more integrated, reusable, and scalable infrastructure that yields better cognitive support. 6 ROBBINS, HILBERT, AND REDMILES Critics with Design Knowledge Feedback Control Situated Analysis Internal Design Architect Representation Perspectives Design Interactions Figure 2.Design environment facilities of Argo. 3. Overview of Argo To Do Decision Model Process Model Figure 2 provides an overview of selected facilities of the Argo software architecture design environment. The architect uses multiple, coordinated design perspectives Figure 3 to view and manipulate Argo's internal representation of the architecture which is stored as an annotated, connected graph. Critics monitor the partial architecture as it is manipulated, placing their feedback in the architect's to do" list Figure 4 . Argo's process model Figure 5 serves the architect as a resource in carrying out an architecture design process, while the decision model lists issues that the architect is currently considering. Criticism control mechanisms use that decision model to ensure the relevance and timeliness of feedback from critics. For comparison, Figure 1 shows facilities of the Janus family of design environ- ments. Like Janus, Argo provides a diverse set of facilities to support re ection-in- action including construction and critiquing mechanisms. Argo, however, extends these facilities byintegrating them with a exible process model and to do" list to explicitly support opportunistic design, and multiple, coordinated design perspectives to aid in comprehension and problem solving. Each of these cognitive theories and the facilities that support them are discussed in Section 5. The subsections below describe each of Argo's facilities. The last subsection provides a usage scenario that describes howanarchitect mightinteract with Argo. 3.1. Critics Critics support decision making by continuously and pessimistically analyzing partial architectures and delivering design feedback. Each critic performs its analysis independently of others, checking one predicate, and delivering one piece of design feedback. Critics provide domain knowledge of a varietyoftypes. Correctness critics detect syntactic and semantic aws. Completeness critics remind the architect of incomplete design tasks. Consistency critics point out contradictions within Figure 3. Three architecture design perspectives: the Component Component perspective top shows conceptual component communication, the Classes perspective middle shows modular structure, and the Resource Component perspective bottom shows machine and operating system resource allocation. The small window in the lower left shows the running KLAX game, represented by this architecture. 8 ROBBINS, HILBERT, AND REDMILES the design. Optimization critics suggest better values for design parameters. Alternative critics present the architect with alternatives to a given design decision. Evolvability critics consider issues, such as modularization, that a ect the e ort needed to change the design over time. Presentation critics look for awkward use of notation that reduces readability. Tool critics inform the architect of other available design tools at the times when those tools are useful. Experiential critics provide reminders of past experiences with similar designs or design elements. Organizational critics express the interests of other stakeholders in the development organization. These types serve to describe and aggregate critics so that they may be understood and controlled as groups. Some critics maybeofmultiple types, and new types may be de ned, as appropriate, for a given application domain. Altogether, we have authored over fty critics, including examples of eachtype. Some examples of architecture critics are given in Table 1. We expect critics to be authored by project stakeholders for various reasons. An initial set of critics is developed by a domain engineer when constructing a domain-oriented design environment. Practicing architects may de ne critics to capture their experience in building systems and distribute those critics to other architects in their organization, or keep them for their own use in the future. A similar authoring activitywas observed by Gantt and Nardi who found that groups of CAD tool users often had members they called gardeners" that assumed the role of codifying solutions to local problems Gantt and Nardi, 1992 . Practicing architects may also re ne existing critics by adding special cases to their predicates or by modifying their feedback. For example, one way for an architect to resolve criticism is to suggest a modi cation to the critic that raised the issue. Researchers may also de ne critics to support an architectural style. Existing literature on architectural styles and system design is a rich source of advice that can be made active via critics. Many organizations already have design guidelines that currently require designers to manually check their design. Software componentvendors may de ne critics to add value to the components that they sell and to reduce support costs. For example, a critic supplied with an ASCII spell checking component might suggest upgrading to a Unicode version if the architect declares that internationalization is a goal. Interactions among stakeholders in the design community can guide the evolution of critics. If the architect does not understand a particular critic's feedbackor believes it to be incorrect, he or she may send structured email through Argo to the author of that critic. This opens a dialog between knowledge providers i.e., domain experts and consumers i.e., practicing architects so that the critics may be revised to be more relevant and timely. In this way critics can be thoughtofas pro-active answers" in an organizational memory Terveen, Selfridge, and Long, 1993; Ackerman and McDonald, 1996 . Possible improvements to Argo's support for organizational memory include associating multiple experts with each critic, prioritizing experts based on organizational distance, and tracking email dialogs so that requests for changes are not forgotten. Table 1. Selected Argo Architectural Critics Name of Critic Critic Type Decision Category Missing Memory Rqmts Completeness Machine Resources Component Choice Alternative Component Selection Too Many Components Evolvability Topology Hard Combination to Test Organizational Component Selection Generator Limitation Tool Component Selection Not Enough Reusable Components Consistency Reuse Avoid Overlapping Nodes Presentation Readability Questionable Experiential Portability Explanation of Problem The memory required to run this component has not been speci ed. There are other components that could t" in place of what you have: list of components. There are too many components at the same level of decomposition to be easily understood. If you need to use these components together, please make arrangements with the testing manager. The default code generator cannot make full use of this component. The fraction of components marked as being reusable is belowyour stated goals. Overlapping nodes does not haveany meaning in this notation and obscures node labels. Your colleague, name of person, had di culty using this component under name of OS. 3.2. Criticism Control Mechanisms Formalizing the analyses and rules of thumb used by practicing software architects could produce hundreds of critics. To provide the architect with a usable amountof information, a subset of applicable critics must be selected for execution at any given time. Critics must be controlled so as to make e cient use of machine resources, but our primary focus is on e ectiveinteraction with the architect. Criticism control mechanisms are predicates used to limit execution of critics to when they are relevant and timely to design decisions being considered by the ar- chitect. For example, critics related to readability should not be active when the architect is trying to concentrate on machine resource utilization. Computing relevance and timeliness separately from critic predicates allows critics to focus entirely on identifying problematic conditions in the product i.e., the partial architecture while leaving cognitive design process issues to the criticism control mechanisms. This separation of concerns also makes it possible to add value to existing critics by de ning new control mechanisms. 3.3. The To Do" List Design feedback from large numbers of critics must be managed so as not to overwhelm or distract the architect. The to do" list user interface Figure 4 presents Figure 4.The architect's to do" list. design feedback to the architect in a non-disruptiveway. When a to do" item is added to the list, the architect may act on it immediately,ormay continue manipulating the design uninterrupted. To do" items come from several sources: critics post items presenting their analyses, the process model posts items to remind the architect to nish tasks that are in progress, and the architect may post items as reminders to return to deferred design tasks. Architects may address items in any order. Tabs on the to do" list lter items into categories. Each to do" item is tied into the design context in whichitwas generated. That context includes the state of the design, background knowledge about the domain, and experts to contact within the design community. When the architect selects an item from the upper pane of the To Do List" window, the lower pane displays details about the open design issue and possible resolutions. Double-clicking on an item highlights the associated or architectural elements in all visible design perspectives. Once an item is selected, the architect may manipulate the critic that produced that item, send email to its author, or followhyperlinks to background information. 3.4. Design Perspectives A design perspective de nes a projection or subgraph of the design materials and relationships that represent a software architecture. Perspectives are chosen to present only architectural elements relevant to a limited set of related design issues. Figure 3 shows three perspectives on a system modeled in Argo. The systemshown is a simple video game called KLAX in which falling, colored tiles must be arranged in rows and columns. In the Component Component perspective, nodes represent software components and connectors, while arcs represent communication pathways. Small circles on the components represent the communication ports of each component. The Resource Component perspective hierarchically groups modules into operating system threads and processes. The Classes perspective maps conceptual components to classes in the hierarchy of programming language classes that implement them. 3.5. Process Model Argo uses a process model to support architects in carrying out design processes. Design processes are di cult to state prescriptively because they are exploratory, tend to be driven by exceptions, and often change when new requirements, con- straints, or opportunities are uncovered Cugola et al., 1995 . Rather than address traditional process modeling concerns e.g., scheduling and enactment , our approach focuses on cognitive issues of the design process by annotating each task with the types of decisions that the architect must consider during that task. We use a simpli ed version of the IDEF0 process notation IFIP, 1993 that models dependencies between tasks without prescribing a temporal ordering. To support cognitive needs, Argo must maintain a model of some aspects of the architect's state of mind. Speci cally, Argo's decision model lists decision types that the architect is currently considering. This information is used to control critics so that they are relevant and timely to the tasks at hand. The primary source of information used to determine the state of the decision model is decision type annotations on tasks in the process model. The architect may edit the decision model directly. Design manipulations performed by the architect can also indicate which decisions are currently being considered. Figure 5 shows an example coarse-grained architecture design process model. Two of the tasks are to choose machine resources Choose Rsrcs and to choose reusable components Choose Comps . The second task is annotated with the decision type Reuse. When the architect indicates that he or she is working on choosing reusable components, these annotations cause Argo to enable critics that support reuse decisions. The design process model shown in Figure 5 is a fairly simple one, partly because the C2 style does not impose any explicit process constraints, and partly because this example does not consider issues of organizational policy. In practice, the process would be more complex. 3.6. Usage Scenario In this section we describe howanarchitect mightinteract with Argo while working through several steps in transforming the basic KLAX game shown in Figure 3 Figure 5.A model of the design process. into a multi-player spelling game. The basic KLAX game uses sixteen separate components, including components that generate colored tiles, display those tiles, and determine when the player has aligned matching tiles. The spelling game variation will use the same basic architecture with new components to generate and display letters and to determine when the player has aligned letters to spell a word. While working on the architecture of the basic KLAX game, the architect places the TileArtist component in the architecture. Shortly thereafter, an alternative critic posts a to do" item indicating that another component from the company's library, LetterArtist, de nes the same interface and should be considered as an alternative. The architect knows that LetterArtist is not appropriate for basic KLAX and takes no action, but the suggestion inspires the idea of building a spelling variation, so he or she leaves the item on the to do" list. Later, when basic KLAX is completed, the architect reviews any remaining to do" items and is reminded to investigate the spelling variation. He or she replaces TileArtist with LetterArtist and de nes new components for NextLetter and Spelling to replace NextTile and Matching, respectively. While the architect is replacing these components the architecture is temporarily in an inconsistent state. Critics that check for consistency between componentinterfaces may post to do" items describing these interface mismatches, but those items are automatically removed when the new components are connected and their interfaces are fully de ned. Adding two new components to the architecture may cause a consistency critic to re if the current percentage of reused components falls below stated reuse goals. Satis ed with the choice of components and their communication relationships, the architect uses Argo's process model to decide what to do next. The process model contains a task for choosing reusable components and a task for allocating machine resources which depends on its output. At this point the architect decides to work on machine resource allocation and marks that task as in progress. Doing so enables critics that support design decisions related to machine resource allocation, and three new to do" items are posted indicating that the three new components Figure 6.Architect's workspace after modifying KLAX. have not been allocated to any host or operating system process. The architect then turns to the Resource Component design perspective and nds that the nodes representing TileArtist, NextTile, and Matching have been removed and new nodes for LetterArtist, NextLetter, and Spelling have been added but not connected to any process or host. The architect connects the new components as the old ones were connected. At this point it occurs to the architect that a server-side Spelling component might be too slow in a future multi-player product, so he or she connects Spelling to the game client process instead. By viewing the Component Component perspective and Resource Component perspective the architect is able to understand interactions between two aspects of the architecture. Figure 6 shows what the architect would see at this point: two design perspectives are open and several new potential problems have been reported by critics. The selected to do" item arose because the Spelling component requires more memory than is available. In this usage scenario the architect has engaged in a constructive dialog with design critics: critics prompted the architect with new possibilities and pointed out inconsistencies. The architect used Argo's process model to help decide which design task to address next, and used two design perspectives to visualize and manipulate aspects of the architecture relevanttotwo distinct design issues. These 14 ROBBINS, HILBERT, AND REDMILES later two aspects of the scenario highlight new facilities of Argo that are not found in previous work on DODEs. 4. Implementation This section discusses the implementation of Argo. We base our discussion on two prototypes: an initial prototype implemented in Smalltalk and the currentversion implemented in Java. First we discuss the core elements of our criticism control mechanisms, perspectives, and processes. We then describe Argo's own architecture and the representation of architectures being designed with Argo. 4.1. Critics In Argo, a critic is implemented as a combination of 1 an analysis predicate, type and decision category attributes for determining relevance, and 3 a to do" list item to be given as design feedback. The stored to do" list item contains a headline, a description of the issue at hand, contact information for the critic's author, and a hyperlink to more information. We encode critics as programming language predicates. Determining which languages are best suited for expressing critics is a topic for future research. Each critic is associated with one type of design material and is applied to all instances of that type. Critics may access the attributes of the design materials they are applied to, and traverse relationships to other design materials. Critic predicates are written pessimistically: unspeci ed design attributes are assumed to havevalues that cause the critic to re. Table 2 presents one critic in detail. Argo provides a critic run-time system that executes critics in a background thread of control. Critics may be run periodically or be triggered by speci c architecture manipulations. During execution a critic applies its analysis predicate to evaluate the design and posts a copy of its to do" item, if appropriate. Another thread of control periodically examines each item on the architect's to do" list and removes items that are no longer applicable. 4.2. Criticism Control Mechanisms Criticism control mechanisms are implemented as predicates that determine if each critic should be enabled. Argo uses several criticism control mechanisms, any one of which can disable a critic. In each of the following examples, criticism control mechanisms decide which critics should be enabled by comparing information provided by the architect to attributes on the critics. Architects may hush" individual critics, rendering them temporarily disabled, if their feedback is felt to be inappropriate or too intrusive. This allows architects to defer the issues raised bya particular critic without risk of leaving the critic disabled inde nitely. Argo's user Table 2. Details of the Invalid Connection critic Attribute Value Name Invalid Connection Design Material Component Types f Correctness g Decision Categories f Component Selection, Message Flows g Hushed False Smalltalk Predicate :comp invalidServices inputs , comp outputs select: :s s isSatisfied not . invalidServices isEmpty not. Feedback This component needs the following messages be sentor received, but they are not present: a list of messages Author jrobbins ics.uci.edu MoreInfo http: www.ics.uci.edu pub arch argo v05 docs . interface allows groups of critics to be enabled or disabled bytype. This allows the architect to control groups of critics easily. Another control mechanism checks the critic's decision types against those listed in the decision model. This keeps critics relevant to the tasks at hand. Criticism control mechanisms normally enhance relevance and timeliness. How- ever, relevance and timeliness can be reduced if criticism control mechanisms use incorrect information. For example, if the architect mistakenly indicates that a given issue is not of interest, then the architect will see no feedback related to that issue and might mistakenly assume that the architecture has no problems. This situation can be avoided byhushing critics instead of disabling them and by using awell designed process that reminds the architect to review all the issues. Argo advises the architect to check the decision model when the to do" list becomes overly full or if too many to do" items are being suppressed. The number of suppressed to do" items is computed by occasionally running disabled critics without presenting their feedback. 4.3. Design Perspectives In Argo, perspectives are objects that de ne a subgraph of the design materials in the current design. Twotypes of perspectives are de ned in Argo: predicate and ad- hoc. Predicate perspectives contain a predicate that selects a subgraph of the design. Ad-hoc perspectives contain an explicit list of design materials and relationships. This latter mechanism allows for manual construction of perspectives via a diagram editor. When a new design material instance is added to the design, predicate perspectives automatically include it if appropriate, whereas ad-hoc perspectives will only contain the new material if it is explicitly added to that perspective. 4.4. Process Model Argo's process modeling plug-in provides a simpli ed process modeling notation based on IDEF0 Figure 5 . The design process is modeled as a task network, where each task in the design process works on input produced by upstream tasks and produces output for use bydownstream tasks. No control model is mandated: tasks can be performed in any order provided needed inputs are available ; tasks can be repeated; and anynumber of tasks can be in progress at a given moment. Each task is marked with a status: future, in progress,or nished. Each task is also marked with a list of decision types. Status information is shown graphically via color in the process diagram. These attributes are used to update the decision model. When the architect indicates that a task is considered nished, the design environment can use this cue to generate additional criticism, perhaps marking the task as still in progress if there are high priority to do" items pending. The process of de ning and evolving the process referred to as the meta-process is itself a complex, evolutionary task for which architects may need support. The process model in Argo is rst-class: it is represented as a connected graph of active design materials and the architect may de ne and modify the process model via the same facilities used to work on architectures. Multiple perspectives may be de ned to view the process. Critics may operate on the process model to check that it is well-formed and guide its construction and modi cation, e.g., the output of this task should be used by another task. The same techniques used to control architecture critics can be used on process critics, including modeling the meta-process so that process critics will be relevant and timely. While the abilitytochange the process gives exibility to individual architects, process critics can communicate or enforce external process constraints. 4.5. Design Environment Architecture Figures 1 and 2 indicate what facilities are available to architects, but they give little indication of how the design environment is implemented. Janus and similar systems have tended to have one major software component for each facility. Those components form a knowledge-rich design environment with tight user interface, data, and control integration Thomas and Nejmeh, 1992 . Our interest in software architecture motivated us to seek a more exible and extensible architecture, while retaining a fairly high level of integration. Figure 7 presents Argo's architecture as a virtual machine. The lowest layer provides domain-neutral infrastructure and user interface components including support for representing connected graphs, multiple perspectives, the critic run-time system, to do" list, and logging facilities. Domain-speci c plug-ins are built Shared SoftArch Document User's Active Design Document Shared Process Document SoftArch Plug-in: adds support for code generation, simulation, . Process Plug-in: adds control over decision model Domain-Neutral Kernel: connected graphs, perspectives, rationale log, critic run-time, "to do" list, decision model, reusable interface elements Active design documents store architectures or reusable design templates with palettes of active design materials, critics, code generation templates, simulation parameters, etc. Figure 7.Argo's architecture presented as a virtual machine. on top of that infrastructure if needed. These plug-ins typically provide pervasive functionality that cannot be built into any particular design material. For example, code generation support is useful for all design materials in the software architecture domain. Most domain-oriented artifacts are stored in active documents" in the top layer. These documents are active in that they contain design materials e.g., software components that carry their own domain knowledge and behavior in the form of critics, simulation routines, and code generation templates. Documents may contain palettes of design materials, designs, reusable design templates, process fragments, or other supporting artifacts. One advantage of this architecture is that artifacts from various supporting domains may be used. Here, a domain is a coherent body of concepts and relationships found in a given problem area, and a supporting domain is a domain for a problem area of secondary concern to the designer, but is useful in completing the design task. For example, a software architect's primary design domain is the construction of systems from software components, whereas recording and managing design rational is a domain of concern that is important to architects but secondary to con- struction. In Argo, plug-ins for software architecture, process modeling, and design rationale may all be available simultaneously, providing software architects with rst-class supporting artifacts for process and rationale. Each supporting artifact may be manipulated, visualized, and critiqued. In designing this architecture we shift away from a monolithic, knowledge-rich design environment that manipulates passive design materials to a modular, domain-neutral infrastructure that allows the architect to interact with knowledge-rich, active design materials. The same trend toward distributing knowledge and behavior to the objects of interest can be observed in the general rise of object-oriented and component-based approaches to software design. Active design materials can be thought of as rst-class objects with local attributes and methods. The analysis predicates of critics can be thought of as methods. Critics that cannot easily be associated with any one design material may be associated with one or more design perspectives. For simplicity, Figure 2 presents critics as looking down on the design from above; a more literal presentation would show critics associated with each node, looking around at their neighbors. The advantages of this shift include increased extensibility, scalability, and separation of concerns in the design environment, and stronger encapsulation of design materials. Encapsulation is enhanced because attributes needed for analysis can be made local, or private, to the design materials, thus supporting local name spaces and data typing conventions. This increases extensibility because each design material may be packaged with its own analyses, and thus de ne its own semantics, which need not be anticipated by the design environment builder. Scalability in the number of critics is increased because there is no central repository of critics critics simply travel with design materials. Concerns are separated because the design environment only provides infrastructure to support analyses packaged as critics and need not perform any analysis itself. All of these advantages support the evolution of architectures, design environments, and software architecture communities over time. ective support for diverse design decisions depends on the architect's ability to obtain and manage large numbers of critics. In the Javaversion of Argo, design materials and critics may be dynamically loaded over the Internet. For example, in a software component marketplace, an architect mightdownload several component design materials, try them in the current architecture, consider the resulting design feedback, and make an informed component selection. 5. Cognitive Theories and Extensions to the DODE Approach Our extensions to previous design environment facilities are motivated by theories of designers' cognitive needs. Speci cally,we extend previous design environment facilities by enhancing support for re ection-in-action and adding new support for cognitive needs identi ed in the theories of opportunistic design and comprehension and problem solving. These theories identify the cognitive needs of designers and serve to de ne requirements on design environments. In the subsections below we describe how Argo addresses these requirements. Table 3 summarizes Argo's support for cognitive needs. 5.1. Re ection-In-Action 5.1.1. Theory As discussed in Section 2.1, Schoen's theory of re ection-in-action indicates that designers iteratively construct, re ect on, and revise eachintermediate, partial design. Guindon, Krasner, and Curtis note the same e ect as part of a study of software developers Guindon, Krasner, and Curtis, 1987 . Calling it serendipitous design," they noted that as the developers worked hands-on with the design, their mental model of the problem situation improved, hence improving their design. Software architectures are evolutionary artifacts in that they are constructed incrementally as the result of manyinterrelated design decisions made over extended periods of time. We visualize design as a process in which a path is traced through a space of branching design alternatives. A particular software architecture can be Table 3. Argo features and the cognitive theories that they support. Critics ProLPMcroeuwslsteibnmpatlseror,dfieoervlsdebtrolacapkuptihnogrship CRProeoemnscteinsstduaeoptriueossrnstoapcneordcintipeicvsesesslsfimistic Kept relevent and timely Produce feedback with links ToAPCdrulolsocltiweosmtsiczhraiobtilces DesPirgoncepsesrespdietcintigves Reflection-in-action Diversity of knowledge X Evaluation during design XX Providing missing knowledge X X Opportunistic design Timliness X Reminding Process flexibility X XX Process guidance X X Process visibility X X Comp. & problem solving Dividing complexity X Multiple perspectives that match multiple mental models thought of as a product of one of the possible paths through this space. Choices at any point can critically a ect alternatives available later, and every decision has the potential of requiring earlier decisions to be reconsidered. Traditional approaches to software architecture analysis require architects to make numerous design decisions before feedback is provided. Such analyses evaluate the products of relatively complete paths through design space, without providing much guidance at individual decision points. As a result, substantial e ort may be wasted building on poor decisions before feedbackisavailable to indicate the existence of problems, and fewer design alternatives can be explored. Furthermore, when analysis is performed only after extended design episodes, it may be di cult to identify where exactly in the decision path the architect initially went wrong. Diverse analyses are required to support architects in addressing diverse design issues, such as performance, security, fault-tolerance, and extensibility. Researchto date has produced a diverse set of architectural analysis techniques. They include static techniques, such as determining deadlock based on communication protocols between components Allen and Garlan, 1994 and checking consistency between architectural re nements Moriconi, Qian, and Riemenschneider, 1995 , as well as dynamic techniques, such as architecture simulation Luckham and Vera, 1995 . The need for diversity in analysis is further driven by the diversity in project stakeholders and the potentially con icting opinions of experts in the software architecture eld itself Garlan, 1995 . Curtis, Krasner, and Iscoe note con icting requirements and thus design evaluation criteria as a major problem for software development in general Curtis, Krasner, and Iscoe, 1988 . Con ict will naturally arise in architecture design, and analysis techniques should be capable of accommodating it. Accommodating con ict in analysis yields more complete support, whereas forbidding con ict essentially prevents architects from being presented with multiple sides of a design issue. Consider architectural styles, which provide design guidance by suggesting constraints on component and connector topology: a given architecture may satisfy the rules of several diverse styles simultaneously.Feedback items related to each of those styles can be useful, even if they contain con icting advice. 5.1.2. Support in Argo Argo supports re ection-in-action with critics and the to do" list. Critics deliver knowledge needed to evaluate design decisions. The to do" list serves as a knowledge in-box" by presenting knowledge from various sources. We will soon add visual indicators to draw the architect's attention to design materials with pending criticism Silverman and Mezher, 1992; Terveen, Stolze, Hill, 1995 . The to do" list and informative assumption described below together support decision making by allowing the architect to browse potential design problems, guideline violations, and expert opinions. Existing software analysis techniques are extremely powerful for detecting well- de ned properties of completed systems, such as memory utilization and perfor- mance. These approaches adhere to what we call the authoritative assumption: they support architectural evaluation by proving the presence or absence of well- de ned properties. This allows them to give de nitive feedback to the architect, but may limit their application to late in the design process, after the architect has committed substantial e ort building on unanalyzed decisions. Such approaches also tend to use an interaction model that places a substantial cognitive burden on architects. For example, architects are usually required to know of the availability of analysis tools, recognize their relevance to particular design decisions, explicitly invoke them, and relate their output back to the architecture. This model of interaction draws the architect's attention away from immediate design goals and toward the steps required to get analytical feedback. Explicit invocation of external tools scales well in terms of machine resources, but not in terms of human cognitive ability.We believe the cognitive burden of interacting with external tools may be enough to prevent their e ective use. Argo follows the DODE tradition in using what we call the informative assump- tion: architects are ultimately responsible for making design decisions, and analysis is used to support architects by informing them of potential problems and pending decisions. Critics are pessimistic: they need not go so far as to prove the presence of problems; in fact, formal proofs are often not possible, or even meaningful, on partial architectures. Heuristic analyses can identify problems involving design details that may not be explicitly represented in the architecture, either because the model is too abstract, or because the architecture is only partially speci ed. Critics can pessimistically predict problems before they are evident in the partial design, and positively detect problems very quickly after they are evident in the partial design, typically within seconds of the design manipulation that introduces the problem. Criticism control mechanisms help trade early detection for relevance to current goals and concerns. In cases where all relevant design details are speci ed, critics can produce authoritative feedback. Unfortunately, for most design issues, there are inherit trade-o s that prevent achieving both informative and authoritative feedback. There will always be a gap between the making of a decision and the analysis of that decision. That gap allows the passing of time, expenditure of e ort, and loss of cognitive context. When one decision is analyzed in isolation, the gap may be small, but the feedbackisat best informative because that decision interacts with others that have not yet been made. When analysis is deferred until groups of interrelated decision have all been made, the gap is necessarily larger, but the feedbackmay be more authoritative because more interactions are known. However, there is a compromise for the informative vs authoritative tradeo : existing analysis tools can be modi ed to make pessimistic assumptions in cases where partial architectures lack information needed for authoritative analysis; and existing critics can be controlled so as to achieve a certain degree of con dence before providing feedback. Alternatively, external batch analysis tools can be paired with tool critics that remind the architect when those tools would be useful. For example, a tool critic could watch for modi cations that a ect the result of the batch analysis and check that the architecture is in a state that can be analyzed i.e., it has no syntax errors that would prevent that particular analysis , then re-run the batch tool, and parse its output into to do" items with links back to the design context. In this case the critic's knowledge is of tools available in the development environment and when they are applicable, whereas the tools themselves provide architectural or domain knowledge. Reusing existing analysis tools is one way to produce new critics, but we expect most critics to be written and modi ed by domain engineers, domain experts, ven- dors, or practicing architects. Argo's approach helps to ease critic authoring in that critics are pessimistic, critic authors need not coordinate their activities with 22 ROBBINS, HILBERT, AND REDMILES other authors to avoid giving con icting advice, and critics need not consider relevance and timeliness. Argo's infrastructure eases critic authoring by providing a framework for implementing critics, a user interface for managing critics and their feedback, and templates for critics and their More Info" web pages. 5.2. Opportunistic Design 5.2.1. Theory It is customary to think of solutions to design problems in terms of a hierarchical plan. Hierarchical decomposition is a common strategy to cope with complex design situations. However, in practice, designers have been observed to perform tasks in an opportunistic order Hayes-Roth and Hayes-Roth, 1979; Guindon, Krasner, and Curtis, 1987; Visser, 1990 . The cognitive theory of opportunistic design explains that although designers plan and describe their work in an ordered, hierarchical fashion, in actuality, they choose successive tasks based on the criteria of cognitive cost. Simply stated, designers do not followeven their own plans in order, but choose steps that are mentally least expensive among alternatives. The cognitive cost of a task depends on the background knowledge of designers, accessibility of pertinent information, and complexity of the task. Designers' background knowledge includes their design strategies or schemas Soloway et al., 1988 . If they are lacking knowledge about how to structure a solution or proceed with a particular task, they are likely to delay this task. Accessibility of information may also cause a deviation in planned order. If designers must search for information needed to complete a task, that task might be deferred. Complexity of a task roughly corresponds to the number of smaller tasks that comprise it. Priority or importance of a step is the primary factor that supersedes the least cost criteria. Priority or importance may be set by external forces, e.g., an organizational goal or a contract. Designers may also set their own priorities. In some observations, designers placed a high priorityonoverlooked steps or errors Visser, 1990 . Thus, the theory of opportunistic design outlines a natural" design process in which designers choose their next steps to minimize cognitive cost. However, there are inherent dangers in this natural" design process. Mental context switches occur when designers change from one task to another. When starting a new step or revisiting a former one, designers must recall schemas and information needed for the task that were not kept in mind during the immediately preceding task. Inconsistencies can evolve undetected. Some requirements maybeoverlooked or forgotten as the designer focuses on more engaging ones. E ciency is lost because of many context switches. Guindon, Krasner, and Curtis observed the following di culties. The main breakdowns observed are: 1 lack of specialized design schemas; 2 lack of a meta-schema about the design process leading to poor allocation of resources to the various design activities; 3 poor prioritization of issues leading to poor selection of alternative solutions; 4 di culty in considering all the stated or inferred constraints in de ning a solution; 5 di cultyin performing mental simulations with many steps or test cases; 6 di culty in keeping track and returning to subproblems whose solution has been post- poned; and 7 di culty in expanding or merging solutions from individual subproblems to form a complete solution. Guindon, Krasner, and Curtis,One implication is that designers would bene t from the use of process modeling. Common process models support stakeholders in carrying out prescribed activities, e.g., resolving a bug report. Software process research has focused on developing process notations and enactment tools that help ensure repeatable execution of prescribed processes. However, in their focus on repeatable processes, process tools have tended to be restrictive in their enforcement of process steps. Design environments can allow the bene ts of both an opportunistic and a prescribed design process. They should allow, and where possible augment, human designers' abilities to choose the next design task to be performed. They can help designers by providing information so they do not make a context switch. Process support should exhibit the following characteristics to accommodate the real design process as described by the theory of opportunistic design and address the problems ed by Guindon, Krasner, and Curtis. Visibility helps designers orient themselves in the process, thus supporting the designer in following a prescribed process while indicating opportunities for choice. The design process model should be able to represent what has been done so far and what is possible to do next. Visibility enables designers to take a series of excursions into the design space and re-orient themselves afterwards to continue the design process. Flexibility allows designers to deviate from a prescribed sequence and to choose which goal or problem is most e ective for them to work on. Designers must be able to add new goals or otherwise alter the design process as their understanding of the design situation improves. The process model should serve primarily as a resource to designers' cognitive design processes and only secondarily as a constrainton them. Allowing exibility increases the need for guidance and reminding. Guidance suggests which of the many possible tasks the designer should perform next. Opportunistic design indicates that cognitive costs are lower when tasks are ordered so as to minimize mental context switching. Guidance sensitive to priorities e.g., schedule constraints must also be considered. Guidance can include simple suggestions and criticisms. It may also include elaborate help, such as explanations of potential design strategies or arguments about design alternatives. Reminding helps designers revisit incomplete tasks or overlooked alternatives. Reminding is most needed when design alternatives are many and when design processes are complex or driven by exceptions. Timeliness applies to the delivery of information to designers. If information and design strategies can be provided to designers in a timely fashion, some plan deviations and context switches maybeavoided. Achieving timeliness depends on anticipating designers' needs. Even an approximate representation of designers' planned steps can aid in achieving timeliness. 5.2.2. Support in Argo Motivated by the theory of opportunistic design, wehave attempted to move from prede ned processes that force a certain ordering of design decisions to exible process models with the properties outlined above. We extend previous work in design environments byintroducing an explicit model of the design process with progress information and a more exible to do" list user interface for presenting design feedback. Argo's process model supports visibilityby displaying the process and the archi- tect's progress in it. Visibility is further supported by the availabilityofmultiple perspectives on the process. For example, an architect maychoose a perspective that shows only parts of the process that lead to a certain goal. Furthermore, the to do" list presents a list of issues that the architect may consider next. Several authors have noted that traditional, sequential work- ow systems do not adequately support exibility and proposed the use of constraint-based process models Dourish et al., 1996; Glance, Pagani, and Pereschi, 1996 . In Argo, exibility is allowed by the simple fact that Argo does not use the process model to constrain the architect's actions: the architect may address any to do" item or perform any design manipulation at any time. Furthermore, exibilityissupported by the architect's ability to modify the process model to better represent their mental model of the design process. Process critics, process perspectives, and a meta-process all support the architect in devising a good design in the process domain. In the currentversion of Argo, guidance is provided only implicitly by the layout of the process model and the prioritization of the to do" items. However, the theory of opportunistic design suggests that guidance should be based, in part, on the mental context required to perform each task. Pending to do" items could be prioritized based on a rough estimate of the cognitive cost of addressing them. The to do" list and process model together support reminding by showing the issues that are yet to be addressed. The to do" list reminds the architect of issues that can be addressed immediately while the process model shows tasks that must be addressed eventually. Critics and to do" items remind the architect of issues that need to be reconsidered as problems arise. Beyond the knowledge contained in the critics and the process model, the architect can also create to do" items that contain arbitrary text and links as personal reminders. The continuous application of critics enables Argo to provide timely feedback. Criticism control mechanisms help make continuous critiquing practical and reduce distractions i.e., unneeded context switches due to irrelevant feedback. In addition to improving design decisions, timely feedback helps the architect make timely process decisions, e.g., is this design excursion complete?" and does a past decision need reconsideration?" 5.3. Comprehension and Problem Solving 5.3.1. Theory The theory of comprehension and problem solving observes that designers must bridge a gap between their mental model of the problem or situation and the formal model of a solution or system Kintsch and Greeno, 1985; Fischer, 1987 . The situation model consists of designers' background knowledge and problem-solving strategies related to the current problem or design situation. The system model consists of designers' knowledge of an appropriate formal description. Problem solving or design proceeds through successive re nements of the mapping between elements in the design situation and elements in the formal description. Successive re nements are equated with increased comprehension, hence the name of the theory. In the domain of software, designers must map a problem design situation onto a formal speci cation or programming language Pennington, 1987; Soloway and Ehrlich, 1984 . In this domain, the situation model consists of knowledge of the application domain and programming plans or design strategies for mapping appropriate elements of the domain into a formal description. The system model consists of knowledge of the speci cation or programming language's syntax and semantics. Programming plans or design strategies enable designers to successively decompose the design situation, identify essential elements and relationships, and compose these elements and relationships into elements of a solution. At successive steps, designers can acquire new information about the situation model or about the system model. Pennington observed that programmers bene ted from multiple representations of their problem and iterative solutions Pennington, 1987 . Namely multiple representations such as program module decomposition, state, and data ow, enabled programmers to better identify elements and relationships in the problem and solution and, thus, more readily to create a mapping between their situation models and working system models. Kintsch and Greeno's research indicated that familiar aspects of a situation model improved designers' abilities to formulate solutions Kintsch and Greeno, 1985 . These two results were applied and extended in Red- miles' research on programmers' behavior, where again multiple representations supported programmers' comprehension and problem solving when working from examples Redmiles, 1993 . Dividing the complexity of the design into multiple perspectives allows each perspective to be simpler than the overall design. Moreover, separating concerns into perspectives allows information relevant to certain related issues to be presented together in an appropriate notation Robbins et al., 1996 . Design perspectives mayoverlap: individual design materials may appear in multiple perspectives. Co-ordination among design perspectives ensures that materials and relationships presented in multiple perspectives may be consistently viewed and manipulated in any of those perspectives. Overlapping, coordinated perspectives aid understanding of 26 ROBBINS, HILBERT, AND REDMILES new perspectives because new design materials are shown in relationship to familiar ones Redmiles, 1993 . Good designs usually have organizing structures that allow designers to locate design details. However, in complex designs the expectation of a single unifying structure is a naive one. In fact, complex software system development is driven bya multitude of forces: human stakeholders in the process and product, functional and non-functional requirements, and low-level implementation constraints. Alternative decompositions of the same complex design can support the organizing structures that arise from these forces and the di erent mental models of stakeholders with di ering backgrounds and interests. Using diverse organizing structures supports communication among stakeholders with diverse backgrounds and mental models whichiskey to developing complex systems that are robust and useful. It is our contention that no xed set of perspectives is appropriate for every possible design; instead perspective views should emphasize what is currently important in the project. When new issues arise in the design, it may be appropriate to use a new perspective on the design to address them. While we emphasize the evolutionary character of design perspectives, an initial set of useful, domain-oriented perspectives can often be identi ed ahead of time Fischer et al., 1994 . 5.3.2. Support in Argo Multiple, overlapping design perspectives in Argo allow for improved comprehension and problem solving through the decomposition of complexity, the leveraging of the familiar to comprehend the unfamiliar, and the use of notations appropriate to multiple stakeholders' interests. Supporting the mental models of a particular domain must be done by domain engineers, practicing architects, and other stake-holders who apply Argo to a speci c domain. Architects and other stakeholders may de ne their own perspectives in the course of design. Presentation and evolvability critics advise architects in de ning and using perspectives. Soni, Nord, and Hofmeister identify four architectural views: 1 conceptual software architecture describes major design elements and their relationships; 2 modular architecture describes the decomposition of the system into programming language modules; 3 execution architecture describes the dynamic structure of the system; and 4 code architecture describes the way that source code and other artifacts are organized in the developmentenvironment Soni, Nord, and Hofmeister, 1995 . Their experience indicates that separating out the concerns of each view leads to an overall architecture that is more understandable and reusable. The 4+1 View Model Kruchten, 1995 consists of four main views: 1 the logical view describes key abstractions classes and their relationships, e.g., is a and instantiates; 2 the process view describes software components, how they are grouped into operating system processes, and how those processes communicate; 3 the development view describes source code modules and their dependencies; 4 the physical view describes how the software will be distributed among processors during execution. These four views are supplemented with scenarios and use cases that describe essential requirements and help relate elements of the various views to each other. The views provide a well-de ned model of the system, but more importantly they identify and separate major concerns in software develop- ment. The Uni ed Modeling Language UML also uses multiple perspectives to visualize various aspects of a design Fowler and Scott, 1997 . In demonstrating Argo, wechose perspectives similar to those described in these approaches; how- ever, we believe that the choice of perspectives depends on the type of software being built and the tasks and concerns of design stakeholders. Argo supports multiple, coordinated perspectives with customization. In addition to the perspectives described in this paper, Argo allows for the construction of new perspectives and their integration with existing perspectives. Architects who are given a xed set of formal notations often revert to informal drawings when those notations are not applicable Soni, Nord, and Hofmeister, 1995 . One goal of Argo is to allow for the evolution of new notations as new needs are rec- ognized. In addition to the structured graphics representing the architecture and process, we allow architects to annotate perspectives with arbitrary, unstructured graphics as demonstrated in Figure 3 . Customizable presentation graphics are needed because the unifying structures of the system under construction must be communicated convincingly to other architects and system implementors. Tobe convincing, the style of presentation must t the professional norms of the development organization: it should look like a presentation, not an architect's scratch pad. Furthermore, ad-hoc annotations that are found to be useful can be incrementally formalized and incorporated into the notations of future designs Shipman and McCall, 1994 . We expect that Argo's low barrier to customization will encourage evolution from unstructured notations to structured ones as recurring formalization needs are identi ed. 6. Evaluation The preceding section has provided theoretical evaluation of our extensions to the DODE approach. Also, the implementation of Argo described in Section 4 provides a proof-of-concept that many of the desired features for Argo can be realized. This section outlines our plans to further evaluate Argo as a working tool. Argo's architecture and infrastructure can be evaluated in terms of howwell they support construction of design environments in new domains. Argo's support for cognitive needs can be evaluated by measuring qualities of design processes and products. 6.1. Application to New Domains The process of applying Argo to a new domain consists of de ning new design materials with critics, a design process, and design perspectives. Belowwe describe our experience in carrying out these tasks for three domains. 28 ROBBINS, HILBERT, AND REDMILES In the domain of C2-style software architectures, there are two basic design ma- terials: software components and connectors. The basic relationship between these materials describes how they communicate. This basic model was extended to include design materials for operating system threads, operating system processes, and source code modules. We rapidly authored approximately twenty critics that check for completeness and consistency of the representation and adherence to published C2 style guidelines Taylor et al., 1996 . The number of needed critics is small because the C2 style addresses only system topology and simple communication patterns. The C2 design process started with tasks to create each of the design material types, and was re ned by splitting activities based on possible design material attributes, e.g., reused components vs. new components. We started with two perspectives discussed in previous work on C2, conceptual and implemen- tation, and later included a perspective to visualize relationships between software components and the program modules that implement them. Wehave also adapted the Argo infrastructure to implement a design environment for object-oriented design that supports a subset of the Object Modeling Technique Rumbaugh et al., 1991 . Since this domain is well de ned and described in a single book, it was straightforward to identify the design materials, relationships, graphical notations, and perspectives. Our OMT subset includes the object model, behavioral model, and information model, but excludes the more advanced features of each. A core set of ten critics that address correctness and completeness of the design was also straightforward to implement, e.g., an abstract class with no subclasses indicates an incomplete design. Additional critics were inspired by a book on OO design heuristics Riel, 1996 , e.g., a base class should not make direct references to its subclasses because that means that adding new subclasses requires modi cations to the base class. Some of these heuristics were more di cult to specify as critics because they rely on information not present in the represen- tation, e.g., semantically related data and behavior should be kept together. In this example, an authoritative answer cannot be given because the OMT design representation does not contain enough semantic information; however, critics may apply pessimistic heuristics to identify when this issue might be a problem. The provided process and perspectives collected various process fragments described in these two books. Wehave begun to apply Argo to software requirements speci cations using the CoRE methodology Faulk et al., 1992 in the avionics application domain. CoRE is based on the SCR requirements methodology Henninger, 1980 with extensions that deal with the modular decomposition of the requirements document. As with OMT, existing documents describe the design materials, standard notations, and analyses. Existing tools cover essentially all analyses that can be performed on the requirement speci cation without considering the application domain, e.g., identifying non-deterministic transitions in a mode-transition table. In implementing a CoRE design environmentwe will demonstrate added value over existing tools byintegrating analysis more tightly with the cognitive process of devising a speci cation, and by providing heuristics to support modularization of requirements documents in the avionics domain, e.g., autopilot control modes are largely inde- pendentofcockpit display modes and should be speci ed in separate requirements modules, however there should be certain constraints between the two. To date wehave implemented twenty critics that check correctness and completeness of CoRE requirements speci cations and integrated them into an independently developed requirements editing tool. Doing so has given us additional con dence in our critiquing infrastructure. Argo's architecture and infrastructure have provided satisfactory support for the initial implementations of domain-oriented design environments in three domains. We plan to evaluate howwell our infrastructure extends in three dimensions: larger domains with more critics and more complex designs and processes, 2 addition of new domain-oriented plug-ins e.g., design rationale , and 3 use of the infrastructure by people outside of our research group e.g., an avionics software development group . 6.2. Evaluating Cognitive Support Toevaluate Argo's support for the cognitive needs of designers, user testing will focus on how Argo a ects the productivity of the designer and the qualityofthe resulting product. Our support for re ection-in-action should increase productiv- ityby decreasing time spent reworking design decisions, lead to better designs in cases where critics provide knowledge that the designer lacks, shorten the lifespan of errors, reduce the number of missing design attributes, and strengthen the de- signer's con dence in the nal design because more issues will have been raised and addressed. We expect our support for opportunistic design will allow designers to rely less on mental or paper notes, and to make better process choices. Comprehension of a sample design should increase when the designer's mental models match one or more design perspectives. Some experimental data can be automatically col- lected, e.g., the lifespan of errors, while others will rely on human observation and interviews. Experimental subjects will use Argo with all features enabled, while control subjects will use Argo with some features disabled. We plan to evaluate the resulting designs with the help of blind judges, as described by Murray Murray, 1991 . Tests that measure our hypotheses will indicate the degree to which identied cognitive needs are supported by Argo's features, and thereby suggest weights for the associations in Table 3. A related task is devising a methodology for on-going evaluation of the qualityof the knowledge provided by critics, the guidance contained in process models, and the mental models suggested by perspectives. This methodology should support on-going maintenance of the design environment and periodic reorganization and reseeding" of domain knowledge Fischer et al., 1994 . Structured email between designers and knowledge providers is one source of data for this evaluation. We are also investigating event monitoring techniques that capture data to help evaluate howwell provided knowledge impacts actual usage. Examples of quantities that could be automatically collected include the number of critics that re, how many to do" items the designer views, and how many errors are xed as a result of viewing feedback from critics. A recentevaluation of VDDE Voice Dialog Design Environment raised several questions about the character of the impact of design critics Sumner, Bonnardel, and Kallak, 1997 . The study found that designers preempted critical feedbackby anticipating criticisms and avoiding errors that the critics could identify. Designers assessed the relevance of each criticism before taking action, and in cases where experienced designers disagreed with criticism they usually added design rationale describing their decision. Sumner, Bonnardel, and Kallak suggest that evaluation of critiquing systems should explicitly consider designers of di ering skill levels. They further suggest that future critiquing systems use alternativeinterface metaphors that users will perceive as cooperative rather than adversarial. In our own experiments we plan to group subjects by experience and watch closely for anticipation of criticism. 7. Conclusions and Future Work Designing a complex system is a cognitively challenging task; thus, designers need cognitive support to create good designs. In this paper wehave presented the architecture and facilities of Argo, our software architecture design environment. Argo's architecture is motivated by the desire for reuse and extensibility. Argo's facilities are motivated by the observed cognitive needs of designers. The architecture separates domain-neutral code from domain-oriented code and active design materials. The facilities extend previous work in design environments by enhancing support for re ection-in-action, and adding new support for opportunistic design, and comprehension and problem solving. In future work, we will continue exploring the relationship between cognitive theories and tool support. Further identi cation of the cognitive needs of designers will lead to new design environment facilities to support those needs. Also, we will seek ways to better support the needs that wehave identi ed in this paper, e.g., a process model that approximates the cognitive cost of switching design tasks. Fur- thermore, we will investigate ways of better supporting and using design rationale. For example, the architect's interactions with the to do" list is a potentially rich source of data for design rationale: items are placed on the list to identify open issues, and removed from the list as those issues are resolved. Design rationale is an important part of design context and to do" items should reference relevant past items when possible. Our current prototype of Argo is robust enough for experimental usage. It is our goal to develop and distribute a reusable design environment infrastructure that others may apply to new application domains. Successful use of our infrastructure by others will serve to inform and evaluate our approach. A Javaversion of Argo with documentation, source code, and examples is available from the authors. Acknowledgments The authors would like to thank Gerhard Fischer CU Boulder , David Morley Rockwell International , and Peyman Oreizy, Nenad Medvidovic, the other members of the Chiron research team at UCI, and the anonymous reviewers. sponsored by the Defense Advanced Research Projects Agency, and Air Force Research Laboratory, Air Force Materiel Command, USAF, under agree- mentnumber F30602-97-2-0021 and F30602-94-C-0218, and by the National Science Foundation under Contract Number CCR-9624846. Additional support is provided byRockwell International. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the policies or endorsements, either expressed or implied, of the Defense Advanced Re-search Projects Agency, Air Force Research Laboratory or the U.S. Government. Approved for Public Release Distribution Unlimited. Notes 1. KLAX is a trademark of Atari Games. --R Answer Garden 2: Merging Organizational Memory with Collaborative Help. Using Style to Understand Descriptions of Software Architecture. Beyond De nition Use: Architectural Interconnection. The Design and Implementation of Hierarchical Software Systems with Reusable Components. Addressing Complexity How to Deal with Deviations During Process Model Enactment. A Field Study of the Software Design Process for Large Systems. Free ow: Mediating Between Representations and Action in Work ow Systems. A Conceptual Framework for the Augmentation of Man's Intellect. Cognitive View of Reuse and Redesign. Supporting Software Designers with Integrated Domain-Oriented Design Environments Construction Kits and Design Environments: Steps Toward Human Problem-Domain Communication Embedding Computer-Based Critics in the Contexts of Design UML Distilled: Applying the Standard Object Modeling Language. Gardeners and gurus: patterns of cooperation among CAD users. Exploiting Style in Architectural Design Envi- ronments An Introduction to SoftwareArchitecture: Advances in Software Engineering and Knowledge Engineering Generalized Process Structure Grammars GPSG for Flexible RepresentationsofWork. Report on a Knowledge-Based Software Assistant Requirements and Design of DesignVision Breakdown and Processes During Early Activities of Software Design by Professionals. A Cognitive Model of Planning. Specifying Software Requirements for Complex Systems: New Techniques and Their Application. Understanding and Solving Word Arithmetic Problems. A Cooperative Problem Solving System for User Interface Design. An Event-Based Architecture De nition Language Reuse of O Correct Architecture Re nement. KI: A Tool for Knowledge Integration. Program transformation systems. Stimulus Structures and Mental Representations in Expert Comprehension of Computer Programs. Foundations for the study of software architecture. Reducing the Variability of Programmers' Performance Through Explained Examples. Cooperative Software. Visual Language Features Supporting Human-Human and Human-Computer Communication The Re ective Practitioner: How Professionals Think in Action. Designing as Re ective Conversation with the Materials of a Design Situation. Supporting Knowledge-Base Evolution with IncrementalFor- malization Expert critics in engineering design: lessons learned and research needs. Empirical Studies of Programming Knowledge. Designing Documentation to Compensate for Delocalized Plans. Software Architecture in Industrial Applications. The Cognitive Ergonomics of Knowledge-Based Design Support Systems " to Living Design Memory." " to Magic World" De nitions of Tool Integration for Environments. Received Date Accepted Date Final Manuscript Date --TR --CTR Shuping Cao , John Grundy , John Hosking , Hermann Stoeckle , Ewan Tempero , Nianping Zhu, Generating web-based user interfaces for diagramming tools, Proceedings of the Sixth Australasian conference on User interface, p.63-72, January 30-February 03, 2005, Newcastle, Australia David M. Hilbert , David F. Redmiles, Agents for collecting application usage data over the Internet, Proceedings of the second international conference on Autonomous agents, p.149-156, May 10-13, 1998, Minneapolis, Minnesota, United States John Grundy , Yuhong Cai , Anna Liu, SoftArch/MTE: Generating Distributed System Test-Beds from High-Level Software Architecture Descriptions, Automated Software Engineering, v.12 n.1, p.5-39, January 2005
human-computer interaction;domain-oriented design environments;evolutionary design;software architecture;human cognitive skills
592026
Efficient Specification-Based Component Retrieval.
In this paper we present a mechanism for making specification-based component retrieval more efficient by limiting the amount of theorem proving required at query time. This is done by using a classification scheme to reduce the number of specification matching proofs that are required to process a query. Components are classified by assigning features that correspond to necessary conditions implied by the component specifications. We show how this method of feature assignment can be used to approximate reusability relationships between queries and library components. The set of possible classification features are formally defined, permitting automation of the classification process. The classification process itself is made efficient by using a specialized theorem proving tactic to prove feature implication. The retrieval mechanism was implemented and evaluated experimentally using a library of list manipulation components. The results indicate a better response time than existing formal approaches. The approach provides higher levels of consistency and automation than informal methods, with comparable retrieval performance.
Introduction The concept of component reuse is fundamental to all engineering disciplines. Components provide levels of abstraction used to effectively construct increasingly complex systems. Software Engineering is no exception, where a main focus has been providing languages and methodologies to help software designers create useful and reusable abstractions. In fact, software reuse was recently described by Mili et al. as the "only realistic approach" to meet the needs of the software industry [23]. The potential for software reuse is not limited to source code, but includes algorithms, architectures, domain models, design decisions, program transformations, documentation - virtually every possible aspect of a software system. In addition, the benefits of software reuse extend beyond the design phase to the analysis and maintenance phases of development. While this article focuses on functional components, the methodology presented is compatible with any software reuse artifacts where a refinement ordering exists between specifications. This includes source code, high-level components [1], software architectures [9, 26], and program transformations [4]. Note also that component retrieval is not confined to library reuse; it can also be considered in the context of integrating components into generic software architectures, either statically [1] or dynamically [27]. While the potential benefits of software reuse are far reaching, in practice software reuse has not flourished. There are both managerial and technical reasons for this. One major technical barrier has been providing tools to automate the reuse process. To understand the reason for the technical difficulty in automating reuse, it is helpful to decompose the library reuse process into three activities: Retrieval - Specifying a query and locating potential reuse candidates within a software library. Evaluation - Determining the relationship between a retrieved component and the specification of the desired component. Adaptation - Making changes to a component to meet reuse requirements. Ideally, a reuse tool should provide automated assistance for all three reuse activities. These activities are interdependent: high reuse potential (from retrieval) should indicate low reuse effort (during evaluation) with respect to known adaptation methods. However, automation is difficult because each activity requires different information about a component. Retrieval benefits from an abstract classification of component function that supports efficient comparison. For automated evaluation to be useful, it must provide a designer with both a precise relationship and a high level of assurance. Therefore, evaluation depends on a precise and detailed description of component behavior. Adaptation requires knowledge about the structure of a component and the functions of its sub-components. Therefore, the choice of a component representation scheme determines what software reuse activities can be automated effectively. Source code reuse has been the main focus of most software reuse efforts. However, automation of the reuse processes requires understanding aspects of a component that are not described in source code. Source code provides a description of how a component performs its function. For purposes of retrieval, we are interested in precisely what this function is. The difference between what and how represents a semantic gap that makes it difficult to understand the function of the code and recognize a potentially reusable component. Source code is not a sufficient representation to support automation of the reuse processes. Therefore, a method is required that represents the additional information necessary to automate reuse. 1.1. Specification-Based Reuse Formal specification languages provide the expressiveness and precision necessary to capture what the function of a component is. Specification matching [35, 45] applies theorem proving to evaluate relationships between specifications. Given a formal definition of reusability, specification matching can be used to evaluate the reusability of a component with respect to a requirements specification. In addition, automated reasoning can be used to determine what changes are necessary to reuse a component and guide component adaptation [29, 39]. To illustrate the potential of using formal methods to assist software reuse, we present a sample scenario of a designer interacting with a specification-based retrieval system. First the designer must develop a specification for the desired component that will be used as a query to the library. The specification defines the domain, range, precondition and postcondition of the desired component. The specification can be abstract and need not be complete. For example, if a designer is looking for list decomposition methods in a library of list manipulation components, the following query might be used: PRE true POST FORALL (x:E) The result of submitting a query to the component retrieval system is a list of components, a matching, and any necessary port substitutions. Matching conditions are formal relationships between the component and query specifications representing conditions under which a component may be reused. The matching conditions may be associated with a set of rules or automated mechanisms for adapting a component. In the case of the example, the retrieval system gives the following results: Component Match Substitutions removeFirst Weak Post rest 7! output removeLast Weak Post rest 7! output split Weak Post left 7! output split Weak Post right 7! output The Weak Post match, defined formally as I C -OC ) OP , indicates that a component provides a solution for part (but not all) of a problem's domain. The designer would like to find a component that covers the entire domain. However, in this case it is not obvious what should happen in the case of an empty input list. The designer can strengthen the precondition of the query to eliminate this case: In response to the new query, the system provides the following results: Component Match Substitutions removeFirst Weak Plug-in rest 7! output removeLast Weak Plug-in rest 7! output split Weak-post left 7! output split Weak-post right 7! output The Weak Plug-in Match is one of several matches that gives a formal guarantee that a component provides a valid solution for the legal inputs of a problem. Therefore both the removeFirst and removeLast components can be safely reused to solve the second query. 1.2. Technical Overview Due to the overhead of automated theorem proving, specification matching is too computationally expensive to test a large number of components in an acceptable amount of time [7, 21, 43]. To attain practical response times, specification matching must be limited to evaluating a small number of components retrieved by a separate mechanism. An alternative to limiting specification matching is to limit the expressibility of the specification language, making retrieval more efficient [31]. However, this lowers the level of assurance during evaluation because it takes into account fewer aspects of a component's behavior. Approaches that do not distinguish retrieval from evaluation will either have inefficient retrieval or weak evaluation. To effectively separate retrieval from evaluation, it is necessary to maintain consistency between the evaluation criteria and the retrieval goal. Specifically, the classification scheme used for retrieval should identify components that will match the query specification with respect to a reusability relationship. Because evaluation is based on component semantics, classification schemes based upon syntactical measures are not guaranteed to be consistent with the evaluation criteria. This paper presents a method for making specification-based retrieval efficient by using the semantic information provided by the specifications. Figure 1 depicts the flow of design information within the retrieval system. In the diagram, boxes represent data structures, ovals represent computations and arrows represent data flow into computations and references between data structures. In the system, formal specifications are used to model the problem requirements and the function of the library components. The specifications are based on abstract data types defined in an algebraic domain theory [14, 42]. In Section 2 we present formal component specifications. Section 3 explains how specification matching is used to determine component reusability. Component retrieval is made efficient by layering a classification scheme above the specifications. The classification scheme consists of a collection of formal definitions representing possible component features in the domain. These definitions control the classification process in place of a human domain expert. Formalizing the scheme permits automated classification of the specifications. The semantic classification scheme is presented in Section 4. The output of the classification phase is a set of features that are used as a query to the component library. The library retrieval mechanism returns components that have feature sets similar to the query. Feature-based retrieval is discussed in Section 5. The similar components are passed on to a more detailed evaluation that uses specification matching to determine each component's precise relationship to the requirements specification. The retrieval system was implemented using the ML programming language [25] and the HOL [12] theorem proving system. In Section 6 we evaluate the response time and retrieval performance experimentally and compare it with other retrieval methods. Section 7 compares the experimental results with other published results. Section 8 discusses the performance of semantic classification in terms of the effects on the retrieval system. We then discuss related work and conclude. 2. Component Specifications Software systems can be represented at the architectural level as a collection of interconnected components [38]. The components of a system are its constituent subsystems that encapsulate part of the system's Classification Scheme Specification Matching Classification Retrieval Theory Domain Requirements Specification Component Library Component Retrieval System Similar Components Feature Matching Components Domain Modeler Designer Figure 1. Overview of Specification-Based Retrieval with Semantic Classification RecordList includes Record, List(Rec) introduces defined asserts 8 l:List[Rec], r:Rec, k:Key 9 r (contains(l,r) - k=r.key); r.key ? first(l).key assumes TotalOrder(Key) introduces asserts Rec partitioned by .key prepend first rest last contains asserts List[E] partitioned by empty, first, rest append(-,e) == prepend(e,-); Figure 2. Larch Shared Language specification of a list of records functionality. Each sub-system is, in turn, composed of a collection of sub-systems resulting in a hierarchical system structure. Decomposition continues in this manner until reaching components that are not implemented by component composition but by actual source code. Formal specifications provide benefits over informal specifications and source code descriptions. First, because the specifications are formal, they provide a precise, unambiguous description of the component's function. Second, the description is declarative as opposed to operational in nature, meaning that it describes what a component does without reference to how it does it. This is important issue for reuse because dependency upon unspecified or unintended functionality can cause problems in the context of evolving requirements and implementations. A formal specification can be broken into two levels of abstraction: domain theories and interface specifications [14]. The domain theory defines the vocabulary used in the specification by providing models of the data types and operations used in the domain of interest. Interface specifications define the behavior of system components in terms of the domain theory. 2.1. Domain Theories We use algebraic specifications to build domain theories. An algebraic specification defines a set of abstract data types and operations. For example, Figure 2 contains a Larch Shared Language [14] specification for a list of records. The RecordList specification (called traits in Larch) includes both the Record specification and a parameterized List specification. In the List specification, operators are defined for appending, COMPONENT search IS IMPORT RecordList; END search; Figure 3. Example Interface Specification prepending, concatenating, and computing list membership and list length. The generated by clause defines the operators that can be used to recursively construct all values of the type. The partitioned by clause identifies a set of operators that can be evaluated to deduce the equivalence of two values of the type. Once a domain theory has been developed for a certain problem domain its data types and operations can be referenced by interface specifications. 2.2. Interface Specifications An interface specification defines the behavior of a component in terms of a domain, range, precondition and postcondition. For example, Figure 3 shows an interface specification for a search component. The domain and range define the input and output types of the component, respectively. The precondition specifies the set of inputs that the component's operation is defined over, called the legal inputs. The postcondition defines the relationship that must hold between an input and a valid output. If the precondition is true when the component begins executing, the component is guaranteed to terminate in a state where the postcondition is true [13]. There are no restrictions or guarantees on the behavior of the component when the precondition does not hold. Formally, a component specification, P , can be translated into the following predicate logic axiom: where DP and RP are the domain and range of the component and I P (x) and OP (x; z) are the precondition and postcondition, respectively. 3. Reusability Using formal specifications to evaluate reusability requires a formal definition of the relationship that must exist for a component to be reused to solve a new problem. We stress this point because existing literature on specification-based component retrieval is not consistent in the choice of evaluation conditions. We first consider the case where a component completely satisfies a problem or query specification. Then we discuss other relationships that may indicate components that can be adapted and reused. A component completely solves a problem if it results in one of the problem's valid outputs for each of the problem's legal inputs. Formally, component specification C satisfies problem specification P if the following condition holds: The first conjunct states that the component will accept all legal inputs to the problem. The second conjunct states that all valid outputs of the component for a legal problem input are valid outputs of the problem. The behavior of a component outside of the legal problem inputs is of no concern in determining its ability to solve the problem. We assume ignoring potential subtype substitutions. Subtypes can be supported implicitly using predicates, such as Integer(x) ) Real(x), or explicitly for more efficient reasoning [36]. Plug-in Weak Plug-in \Gamma\Psi Plug-in Post @ @R Satisfies \Gamma\Psi Weak Post I C - OC ) OP @ @R \Gamma\Psi Stronger Weaker Figure 4. A lattice of specification matches used for evaluating reusability It is not always the case that a component completely satisfying the problem will exist in the library. Therefore, it is desirable for a query to match components that can be adapted or combined to solve the problem. Zaremski and Wing [43, 45] have identified a collection of specification matches that can be useful in comparing specifications. Figure 4 shows a subset of these matches that we think are of interest in determining reusability, with the addition of Satisfies match. 1 Following Zaremski and Wing, matches are arranged in a lattice, where an arrow between two matches indicates that the match at the base of the arrow is stronger than (logically implies) the match at the head of the arrow. The formal notation is abbreviated by dropping the quantifiers and variable arguments for the predicates. The three matches on the left-most path all require that the precondition of the problem be stronger than the precondition of the component. They differ in the set of inputs that require a valid output from the postcondition relation of the component: Plug-in match checks the whole domain, Weak Plug-in restricts the check to the legal component inputs, and Satisfies further restricts this to the legal inputs of the problem. Because of the logical relationship between these matches, any components matching under Plug-in or Weak Plug-in will match under Satisfies. However, Plug-in and Weak Plug-in will cause the disregard of useful components if used as a retrieval condition. The Plug-in Post and Weak Post matches differ from the others by not requiring all legal problem inputs to be legal component inputs. If a component matches a problem in one of these ways, there could be problem inputs that cause unspecified behavior in the component. However, we do know that for any legal problem input that is also a legal component input, the component provides a valid problem output. Therefore, components that match in these ways can be used as a partial solution to the problem. A collection of such components can be composed to provide a complete solution to the problem. These matches are also useful for finding components with non-trivial preconditions without having to specify a precondition in the query [43]. Because the specification matches are formally defined, they can be checked using an automated theorem prover. However, due to the complexity of automated theorem proving, specification matching is too computationally expensive to test a large number of components [7, 21, 43]. Practical specification-based retrieval requires an efficient way to identify components that will match the query specification with respect to a reusability relationship. The next section presents a method for doing this by classifying components using the semantics of their specifications. 4. Semantic Classification The efficiency required for specification-based retrieval can be achieved using a feature-based classification scheme. Feature-based retrieval is efficient because it relies only on the syntactic matching of attribute-value pairs, or features. The similarity of two components is measured by the number of features they have in common. When applying feature-based classification by hand, library components are assigned a set of features by a domain expert. To retrieve a set of potentially useful components, the designer translates (classifies) the problem requirements into a set of features and the corresponding class of components is retrieved from the library. Queries can be generalized by relaxing how the feature sets are compared. A potential problem with feature-based classification is maintaining the consistency of the classification process [24, 33]. Effective retrieval requires the domain expert and the developers to have a common understanding of the intended semantics of the features. It would be desirable to systematize and automate classification to increase the confidence that classification is consistent among the domain experts and the developers. Automatic indexing based on semantics 2 is not practical from a code reuse standpoint because it would require a massive re-engineering effort. However, formal component specifications provide an explicit semantic representation that can be used as a foundation for component classification. Automation of feature-based classification requires answering two questions [17]: how do we automatically generate features from a specification and what are the possible features that a component can have? We answer the first question by describing a framework for assigning features to components in a way that assists the search for reusable components. We then discuss how we define the set of possible features used to classify components. 4.1. Feature Set Generation The classification scheme used for retrieval should identify components that will match the query specification with respect to the reuse matches identified in Section 3. Inspection of these specification matches reveals a general pattern: in each case part of one specification logically implies part of another specification, such as I P . The search for these situations can be guided using necessary conditions. A necessary condition for a predicate P is another predicate \Phi logically weaker than P , i.e. P ) \Phi. For two predicates, P and Q, such that P ) Q, every necessary condition of Q will be a necessary condition of Therefore, we can commonly describe P and Q by the fact that they both logically imply \Phi. Necessary conditions can be used in this way to assist the search for reusable components. More specifically, they can identify components that cannot match a specification and therefore should not undergo detailed analysis. Both the component and query specifications are classified based on a given set of necessary conditions. A classification scheme associates a feature with each necessary condition. A feature is assigned to a component if its associated necessary condition is logically implied by the component's specification: The general result of this semantic classification method is that if a component has a feature set similar to that of the query then there is the potential for a reuse match to hold between the component and query. Conversely, components that do not have features in common are less likely to be reusable. Therefore, syntactic comparison of feature sets can be used to efficiently approximate the semantic relationships between components. In general, semantic classification cannot be guaranteed to eliminate only non-matching components. The behavior of the system depends upon the set of necessary conditions used in the classification scheme and which reuse matches are being approximated. In effect, approximate reasoning based on feature sets is unsound and incomplete. However, this should not be considered as a critical flaw until it is clear what its effects are on the practical performance of the system [5]. The impact of unsound and incomplete classification is evaluated by experimentally measuring its effect on retrieval performance (Section 6). 4.2. Feature Definitions The classification process is controlled by a collection of feature definitions that determine the set of necessary condition/feature pairs used. The feature definitions capture the knowledge a domain expert would use to classify components by hand. By formally defining the classification features and the feature assignment process, classification can be fully automated. Filter(T Figure 5. Sample feature definitions for a data-flow abstraction To characterize aspects of interface specifications, the features must represent abstract relationships between component inputs and outputs. The feature definitions link the feature names and values to logical predicates which specify the associated concept. Feature definitions have the following format: The feature name provides a syntactic label for a concept. The feature value (Type1; Type2) represents data type parameters that are instantiated based on the domain and range of the specification. If two components perform a similar function over different data types, they will have a feature with the same name, but different values. The variables x and y are actually metavariables that range over the set of input or output variables for a component. The isIO() predicates, either isInput() or isOutput(), determine which of these sets a metavariable ranges over and associates the type of the variable with the feature value parameters. The Condition() predicate is the necessary condition associated with the feature. A feature definition is instantiated by substituting in all combinations of input and output variables from the component being classified. Instantiation is restricted by type checking any operators which are used when specifying Condition(). The set of features used to classify a reuse library is domain dependent; the necessary conditions in the feature definitions are specified in terms of the formal domain theory. This allows specific abstractions to be made about operators that provide similar functionality over different types in the domain. For example, the containment operator (contains()) is used to describe containment in many different situations. By referencing this operator in a feature definition, we can draw similarities between components whose function is specified in terms of containment. Figure 5 shows a subset of the features that are currently defined that provide a data-flow style abstraction for the list processing domain. For example, the Select feature represents the case where an output is an element of an input variable. Because the definition is parameterized on the datatypes, Select is a possible feature whenever the containment operator exists between an input and output type. This set of feature definitions is used in the following example. 4.3. Classification Example Figure 6 shows the classification of the search component from Figure 3 based on the first two feature definitions from Figure 5. First, the domain and range of the interface specification are substituted into the feature definitions to create the set of feature/necessary conditions used for classification. For example, the Domain Theory Instantiate Feature/Goal Pairs Definitions Feature Proof Tactic Specialized Feature Assignment Requirements Specification Figure 6. Example: feature-based classification of a search component Select and NonMember feature definitions: are instantiated with the input variables and types from the domain (input : (Rec) of the specification, giving: The second and fourth instantiated definitions are eliminated due to type checking constraints, because the Bool is not defined in the domain theory. Next, the specialized proof tactic is used to check that the necessary conditions are implied by the specifi- cation. For example, the following proof obligation is generated for the first instantiated definition: In this example, the proof for the term contains(input; item) succeeds. Therefore, its associated feature, Select(List[Rec]; Rec), is assigned to the component. 5. Retrieval Given the feature set representation for a problem, we wish to retrieve components that match the problem in terms of the formal specification matches. Because the features are assigned to the components based on necessary conditions, the less features that a component has in common with a query, the less likely it is that one of the reusability matches holds between the two. Therefore, we are interested in components that have features in common with the query. Because identical specifications will have identical feature sets, the initial query is for components having an identical feature set to the problem. If there are no such components, the query is generalized by loosening the feature comparison constraints to include a larger class of components. In the prototype, query generalization is automated and continues until the number of retrieved components reaches a user-supplied threshold. The first step in generalization is to weaken the requirements for feature value equivalence. Because the feature values represent types that the component operates over, this identifies components performing similar operations on different types. The second step is to reduce the number of the problem features that the component must contain. This allows retrieval of components that may be partial solution to the problem. Finally, the two types of generalization are combined to find components having any features in common with the query. The following sample retrieval session provides an example of each type of query generalization. Figure 7 shows two problem specifications and a set of component specifications, all with their associated feature sets. The specifications have all been classified using the feature definitions in Figure 5 and the resulting feature sets are listed below each specification. Components are displayed beneath the problem specifications for which they were retrieved. To exhibit the relationship between the feature sets and the specification matches, the strongest match that exists between the problem and a component is listed under the component specification. In the first example, we wish to find the record with a specified key within a list. The component is expected to perform correctly only if there is such a record in the list. Two components are retrieved that have identical feature sets to find. The first, search, matches under Plug-in (and therefore Satisfies) meaning it can be used to solve the problem. The second, binarySearch, only matches under Plug-in Post because it requires the input to be sorted. It is possible that the input list is known to be sorted, but the designer did not include that information in the query. Therefore, binarySearch is of possible use in solving the problem. A third component, treeSearch can be retrieved by weakening the constraint that feature values must be equal. The treeSearch component has both the Select and Build features, however with different type values. This component may be useful after substituting List[Rec] for Tree and Rec for Bucket. The second example problem is specified more abstractly than the first. The designer is looking for components that take a list and return a smaller list composed of elements from the first list. The manner of selecting the elements for the new list is unspecified. This query is useful for finding the existing options for decomposing lists. The first two components that are returned, removeFirst and removeLast, both match the query under Weak Plug-in. This is because the precondition of the components, NOT empty(input), is required to be true to ensure that the output list is smaller than the input list. Technically, for these two components and this query, Weak Plug-in and Satisfies are equivalent matches because the preconditions are logically equivalent. The third component, split, only matches under Weak Post because the component precondition is stronger than query precondition. The split component provides a valid solution to the problem, except in the case where length(input) = 1. 6. Empirical Evaluation Component retrieval based on brute force specification matching attempts to match a query with every component in the library. Therefore, it involves many individual proof attempts, most of which will fail [37]. The goal of semantic classification is to eliminate components that will lead to unsuccessful proof attempts during evaluation, saving time and effort. The number of (matching and/or non-matching) components eliminated depends upon the performance of the retrieval system. There were several experiments performed to evaluate the retrieval system performance. 6.1. Implementation The semantic classification system was implemented using the the ML programming language [25] and HOL [12] theorem proving system. Several precautions were taken to reduce the overhead of automated reasoning during classification [28]. First, the feature sets for the library components are calculated beforehand and stored in an index. Second, a special purpose proof tactic was constructed in HOL to solve theorems in the form of feature implication proofs. The proof tactic is parameterized on the set of domain axioms it applies, making it domain independent. Finally to speed up inference, inductive proofs were eliminated by burying the induction into proofs of lemmas and adding the lemmas to the domain theory. These precau- 2Problem Specifications: COMPONENT find IS IMPORT RecordList; END find; Features: Select(List[Rec],Rec),Build(Key,Rec) Component Specifications: COMPONENT subSet IS POST FORALL (x:Rec) IMPORT RecordList; END subSet; Features: Filter(List[Rec],Rec) COMPONENT search IS IMPORT RecordList; END search; Features: Select(List[Rec],Rec),Build(Key,Rec) COMPONENT binarySearch IS IMPORT RecordList; END find; Features: Select(List[Rec],Rec),Build(Key,Rec) COMPONENT treeSearch IS IMPORT BucketTree; END find; Features: COMPONENT removeFirst IS first : Rec; IMPORT RecordList; END removeFirst; Features: Select(List[Rec],Rec),Filter(List[Rec],Rec) COMPONENT removeLast IS last IMPORT RecordList; END removeLast; Features: Select(List[Rec],Rec),Filter(List[Rec],Rec) COMPONENT split IS POST NOT isEmpty(left) AND NOT isEmpty(right) IMPORT RecordList; END split; Features: Split(List[Rec],Rec),Filter(List[Rec],Rec) Figure 7. Example Problem Specification and Component Specifications with Feature Sets tions result in an incomplete proof procedure. One goal of the experiments was evaluate the impact of this incompleteness on retrieval performance. 6.2. The Library The component retrieval evaluation was done using a library of specification for list manipulation components. This library has been used in experiments with other specification-based component retrieval systems [37], providing an opportunity for direct comparison of results. The library was designed to test whether the specification-based retrieval can handle variation in the way that components are specified and the way that queries are posed to the system. For example, there are 3 different specifications for a head component that takes a list and returns a list containing only the first element of the original list. 6.3. Evaluation Method The two traditional measures of component retrieval performance are recall and precision [23]. Recall is the ratio of the number of relevant items retrieved to the total number of relevant items in the library. High recall indicates that relatively few relevant components were overlooked. Precision is the ratio of the number relevant items retrieved to the total number of items retrieved. High precision means that relatively few irrelevant components were retrieved. In general, there is a tradeoff between precision and retrieval. The goal is to find a practical balance between the two. The relevance condition is fundamental to the evaluation of a retrieval system. As discussed below, experiments were conducted with two different relevance conditions. It was also informative to observe the number of components retrieved by the system. This number can help estimate the load that would be placed on the designer to interpret the results of a query in an interactive system, or similarly, the search space that would be faced by an adaptation system when considering component compositions [28]. The response time of the system was also measured to determine the practicality of the method. For each measured quantity, the minimum, maximum and median was calculated for each of the scenarios in the experiment. 6.4. Design of the Experiments The experiments were designed to compare the way that several factors affected the performance of the retrieval system. The first and foremost was the ability of the automated classification system to derive classification features. Second, we were interested in the performance of retrieval in the context of both exact retrieval and approximate retrieval. Finally, the nature of the library raised questions about determining an appropriate query set for experimentation. Therefore, two different query sets were used. 6.4.1. Feature Generation The retrieval system as a whole can be separated into the classification scheme (as defined by the feature definitions) and the classification mechanism (the specialized proof tactic and domain theorems). In a sense, the mechanism attempts to implement the scheme. Both aspects of the retrieval system can affect precision and recall. The classification scheme affects precision by the size and consistency of component clusters. It affects recall because a scheme may not always contain a feature that can be inferred from a relevant components. The incompleteness of the classification mechanism (the specialized proof tactic is incomplete) could cause it to generate fewer features and subsequently retrieve fewer components than intended. If the missing components are relevant it will lower both recall and precision. The classification mechanism is sound (it is implemented in terms of sound constructs in HOL) so it will only derive implied features. In each experiment three scenarios were tested: 1. Signature Matching [44]: retrieval based on component signatures 2. Expected Features: retrieval based on features that would be assigned by a complete proof procedure. 3. Derived Features: retrieval based on features assigned by the implemented system. Signature matching retrieves components with identical signatures. For this library, all of the components have identical signatures (list ! list). Therefore, the performance of signature matching provides a profile of the composition of the library in terms of relevant components. The expected features determine the performance that the classification scheme, if implemented perfectly, would allow. Expected features were determined by inspection with the aid of the lattice of specifications for the library. The implemented retrieval mechanism (i.e., the domain theory axioms together with the feature derivation tactics) was used to generate the derived features. The results were evaluated to see how close it comes to implementing the classification scheme. 6.4.2. Relevance Conditions The choice of a relevance condition is fundamental in determining the significance of precision and recall measures [22]. In our experiment, we evaluated the performance of the system with respect to two relevance conditions. First, we use Satisfies match to be consistent with standard specification-based retrieval experiments [21, 37]. Second, we consider a relevant component to be one that matches the query specification with respect to any of the reuse matches identified in Section 3. This relevance condition is important in the context of the adaptation, where a relevant component is one that can (potentially) be adapted by the system [28, 29]. 6.4.3. Query Set Following the experiments done by Schumann and Fischer [37], the library components themselves are used as the set of queries to test the performance of the system. Using the components as the query set makes two assumptions: (1) the component specifications represent a good sample of queries that may be asked and (2) these queries all have the same probability of being posed to the system. To get results that would predict the performance of the tool in a realistic setting, it would be necessary to have a distribution of queries that represents how the tool would be used. There is no study in the component retrieval literature that would provide this information. The library used in the experiment has several groups of functionally equivalent specifications that would, in practice, all point to a one component. This raises a question about what should define a unique query in the experiment: a specification or a component with potentially many specifications. Therefore, the experiment was run once with duplicate specifications in the query set, and once without. The difference between the results for the two cases was negligible, indicating that the system does not favor the components with multiple specifications [28]. We present only the results from the experiment including equivalent queries here. 6.5. Experimental Results The experiment involved a library of list manipulation components all having the signature list 7! list. The library contained 63 specifications for 45 functionally unique components. The experiment was divided into two parts. For the first part, the classification scheme of Figure 5 was used. Due to the limited signatures, only the following features applied: Filter(T For the second part of the experiment, the classification scheme was extended. The extension of the scheme was guided by placing the library components into a lattice based on the Satisfies matching condition. The domain theory theorems that were used as parameters to the classification mechanism are shown in Figure 8. These theorems were selected by observing the failed proofs of expected features and determining Normalization Rules: Normalization Implications: Rewrite Rules: Expansion Rules: Figure 8. Domain Theory for List Library Experiment Table 1. Retrieval for Satisfies Match Using the Initial Classification Scheme Scenario Retrieved Precision Recall Signature Match 63.00 (63 - 63) 0.11 (0.02 - 0.64) 1.00 (1.00 - 1.00) Expected Features Exact Match 22.71 (3 - 31) 0.20 (0.03 - 0.74) 0.84 (0.05 - 1.00) Relaxed Match Derived Features Exact Match 22.27 (3 - 31) Relaxed Match 43.57 (12 - 51) 0.12 (0.02 - 0.65) 0.86 (0.10 - 1.00) the theorems necessary to make the proofs succeed. The theorems mainly deal with reasoning about containment (which is used to define the feature definitions) in terms of the list type constructors CONS, APPEND and []. 6.5.1. Satisfies Match The results of the initial part of the experiment for Satisfies match are shown in Table 1. The table entries denote the average value with the minimum and maximum in parenthesis. The retrieval mechanism comes very close to implementing the classification scheme: the expected feature sets were derived for 61 of the 63 specifications. The failed classification was due to the use of a three way conditional in the specification that was not supported directly in the domain theory. The domain theory could be extended to support these conditionals, however an effort was made to not over-specialize the domain theory toward supporting classification application. Therefore, this extension was not made. The distribution of expected feature sets for the library is show in Table 2. This shows that the classification scheme does a questionable job of clustering components in the library; nearly half of the components are only assigned the Filter feature. In fact, the 3 features are not independent but represent a generalization hierarchy: Route ) Permute ) Filter. While it is useful to have features that specialize other features, it would also be useful to have other orthogonal features to provide better coverage of the library. Table 2. Distribution of feature sets for initial classification scheme. Feature Set No. Specifications No. Components fFilterg some_total last_total1,2 tail1,3 swap id_segment id_front run_max_eq segment_ne_total id segment_front id_single segment swap_outer swap_outer_total swap_total perm_r1,2 run_max_bracket run_eq1,2,3 run_bracket1,2 lead_total1,2 lead segment_rear segment_ne perm_lr1,2 id_nil elim_dup_lr rot_r1,2 rot_r_total1,2,3 rot_l_total1,2,3 rotate_total rotate rot_l1,2 tail_total1,2 elim_dup_unique_l elim_dup_r last head1,3 some id_rear elim_dup_unique_r perm_l1,2 elim_dup_l no_dup elim_dup_unique_lr FILTER Figure 9. Lattice of Specifications for the List Component Library The partial-order lattice induced by the Satisfies Match on the library is a useful aid for discovering potential features. This lattice is shown in Figure 9 with the areas covered by the Filter, Permute and Route features shown with dashed lines. As hoped, the features tend to group components that are related in the lattice. The lattice can be used to identify groups of components that are closely related, and then their specifications may be inspected to identify a logical feature that they share. 3 For example, the some component is the root of a tree containing last, head, some total, last total and head total. The some component has the following specification: POST EXISTS(i:Rec) mem1 mem2 . This specification is very similar to the definition of the Select feature, only Select looks for an element as an output, not a singleton list. Therefore, if the definition of Select is modified to identify a singleton list, it should be an expected feature of all of these components. Other useful feature can be identified using the lattice. For example, associating a feature with the id nil component: and its descendants provides coverage for the id segment tree an and also divides the components with the Filter feature roughly in half in a fairly orthogonal manner. As a complementary feature to IdNil, a feature NoNil could be defined stating that a component does not accept an empty input, with the intention that a component could not imply both of these features. These three features were formalized as follows: They were added to the feature definitions and the experiment was rerun. On inspection of the derived feature sets, it was immediately obvious that there were two problems: (1) the Some feature was never derived and (2) many components were assigned both IdNil and NoNil, which was not the intention of the scheme. The source of the first problem was that the proof tactic failed to handle existential goals. This problem was solved by extending the tactic to attempt to solve a goal by substituting in free variables from the goal for the existential variable. Additionally, the expected results were wrong: as defined, the Some feature will not hold for some total and its descendants. The components named * total are total functions that map an empty input list to an empty output list. In this case, the output does not contain an element and therefore, these specifications will not imply the Some feature. The second problem with the extended scheme was that IdNil and NoNil were not complementary as thought. In fact, NoNil implies IdNil because, in the case where the input is not empty, assuming it is empty (which is be the first step in proving IdNil) allows the proof of IdNil to succeed trivially. This results in the coverage of the IdNil feature to be nearly identical to the Filter feature, making it a useless extension. Taken together, these experiences indicate the need for tools to support the construction of classification schemes, as discussed in the future work section. Table 3. Distribution of Feature Sets for Extended Classification Scheme. Feature Set No. Specifications No. Components Table 4. Retrieval for Satisfies Match Using the Extended Scheme Scenario Retrieved Precision Recall Signature Match 63.00 (63 - 63) 0.11 (0.02 - 0.64) 1.00 (1.00 - 1.00) Expected Features Exact Match 13.64 (1 - 23) Relaxed Match Derived Features Exact Match 14.05 (3 - 23) 0.29 (0.04 - 1.00) 0.69 (0.05 - 1.00) Relaxed Match 43.57 (12 - 51) 0.12 (0.02 - 0.65) 0.86 (0.10 - 1.00) Table 5. Approximate Retrieval Using the Initial Classification Scheme Scenario Retrieved Precision Recall Signature Match 63.00 (63 - 63) 0.14 (0.02 - 0.81) 1.00 (1.00 - 1.00) Expected Features Exact Match 22.71 (3 - 31) 0.27 (0.03 - 1.00) 0.84 (0.06 - 1.00) Relaxed Match Derived Features Exact Match 22.27 (3 - 31) 0.28 (0.03 - 1.00) 0.81 (0.06 - 1.00) Relaxed Match 43.57 (12 - 51) 0.16 (0.02 - 0.96) 0.85 (0.07 - 1.00) The experiment was run again, this time without the IdNil feature and with the expected results more throughly evaluated. The distribution of feature sets is shown in Table 3. The extended scheme does a better job of breaking the components into clusters. Nearly 1/3 of the components that were initially assigned only the Filter feature are now assigned additional features. The results of evaluating the system with the extended classification scheme are shown in Table 4. Compared to the results of the previous classification scheme, there was a noticeable drop in the average number of components retrieved along with an increase in precision. This was accompanied by a slight decrease in recall. Once again, the classification mechanism came very close to implementing the classification scheme. 6.5.2. Approximate Retrieval The experiments were repeated while considering a relevant component to be one that matches the query specification with respect to any of the reuse matches identified in Section 3. Because this relevance condition is logically weaker than the Satisfies match (it is the disjunction of Satisfies and Weak Post) the relevant components for a query will be a superset of those relevant to Satisfies. Table 6. Approximate Retrieval Using the Extended Scheme Scenario Retrieved Precision Recall Signature Match 63.00 (63 - 63) 0.14 (0.02 - 0.83) 1.00 (1.00 - 1.00) Expected Features Exact Match 14.97 (1 - 25) 0.32 (0.04 - 1.00) 0.57 (0.04 - 1.00) Relaxed Match Derived Features Exact Match 14.05 (3 - 23) 0.31 (0.04 - 1.00) 0.54 (0.06 - 1.00) Relaxed Match 43.57 (12 - 51) 0.16 (0.02 - 0.96) 0.84 (0.07 - 1.00) The results of the experiment using the initial classification scheme are shown in Table 5. The precision of this experiment was higher, because more of the retrieved components are relevant. Recall remained the same indicating that the same percentage of new relevant components is retrieved. The implementation of the classification scheme continues to come close to the performance of a complete and consistent implementation. The results of the experiment using the extended classification scheme are shown in Table 6. For the exact match there was an increase in precision over the initial scheme comparable with the increase seen in the Satisfies match experiment. However, there was a larger drop in recall. This is caused by the exact match not retrieving many of the new relevant components. This means that the exact feature match using the extended scheme does not approximate the relevance condition very well in terms of recall. The relaxed match results remain consistent with the other experiments. The results indicate that relaxed feature match is more appropriate for approximate retrieval than exact feature match. 6.5.3. Response Time The response time of the classification system during the experiments ranged from 0.15 to 0.66 seconds with an average of 0.35 seconds on a 200MHz PentiumPro processor running Linux. Database access accounted for only 0.025 seconds, on average, therefore the bulk of the time was spent classifying queries. This is well within the acceptable response time for an interactive system. 7. Comparison of Results 7.1. Specification-Based Retrieval Most specification-based component retrieval systems are in the "proof of concept" stage and therefore have not been evaluated over a sizable component library. A notable exception is the NORA/HAMMR system of Fischer and Schumann [37]. The NORA/HAMMR retrieval system is set up as a chain of filters. The initial filter is signature matching and they become more restrictive as they progress. The final filter in the chain is full scale specification matching. The NORA/HAMMR system was evaluated using the same library as our experiments using all of the specifications (with duplicates) as the query set. As an intermediate filter, they use the MACE model checker in several configurations to select a subset of components from the library to undergo specification matching. Using a 20 second time limit for model checking computation, they observed average recall rates between 74.7% and 81.3% with precision between 18.5% and 16.5%. This is comparable to the results achieved with the initial classification scheme in our experiments, however, our response time was 0.66 seconds in the worst case. They have experimented with several automated theorem provers to do specification matching as the final stage of retrieval. For example, using the SETHEO prover and a time limit of 20 seconds, the results were a recall rate of 61.2% and precision of 100%. The high precision is due to the fact that SETHEO's proof procedure is sound. The loss in recall is due to a lack of completeness that comes from a technique for approximating induction, restricting the set of axioms available for inference and the 20 second time limit. Lowering the time limit to 1 second causes the recall to drop below 50% [6]. Using semantic classification as a filter prior to specification matching could reduce the load on specification matching and allow this time limit to be raised, increasing recall. 7.2. Information Retrieval Methods The majority of component retrieval tools used in practice are based on information retrieval methods. Frakes and Pole conducted an extensive study of representation methods for reusable components [8]. They did comparison of the retrieval performance of attribute-value, enumerated, faceted and keyword based representations. The relevance determination was made by two domain experts. There is no mention of the (manual) method used to classify the components. The recall and precision measurements were in the 30-40% and 50-100% ranges, respectively for all of the methods. Statistical analysis showed no significant difference among the methods. Our retrieval results are consistent with these numbers. The benefit that semantic classification provides over these methods is consistency and automation of the classification process. While the uses have to be familiar with the specification language, they do not have to be familiar with the organization of the component library. In addition, specification-based retrieval provides a precise relationship that exists between a retrieved component and a query, increasing the utility of the retrieval results. All of this is achieved with similar performance results. Girardi and Ibrahim [10, 11] evaluate a method based on syntactic and semantic analysis of natural language descriptions (ROSA). They use normalized versions of the precision and recall formulas because the system returns ranked results. They used a library of 418 general purpose Unix commands and 20 queries derived from user descriptions of frequently used commands. They report recall values in the 99-100% range with precision values in the 90-92% range. However they do not state their relevance condition, making the results rather non-informative. Their selection of the query set is also important because it does not evaluate the method in the less frequent cases. 8. Discussion 8.1. System Wide Effects The retrieval performance of semantic classification must be considered in the context of its role in the retrieval system. The results of retrieval are passed on to an evaluation phase (based on specification matching) that has perfect precision and recall, assuming a sound and complete proof procedure is used. The point is that relevant components cannot be added, but only removed by specification matching. This means semantic classification acts as a filter that sets an upper bound on the recall of the combined retrieval/evaluation system. Limiting recall will shrink the design space that can searched, potentially lowering the quality of the designs created by the system. In contrast, the precision of the retrieval phase has no effect on the precision of the combined system. The precision effects the number of proofs that must be attempted during specification matching, and therefore has an effect on the system response time. Therefore, the recall/precision tradeoff in the feature-based retrieval phase translates into a design quality/response time tradeoff in the context of the entire retrieval system. 8.2. Limitations In the experiments, the major reason for feature generation failure was due to specification that was broken into cases based on conditions that were not supported by rules in the domain theory. In these specifications the postcondition has the form: To prove that a feature is implied by a specification of this form requires a proof by cases approach. To facilitate this, there must be an axiom in the domain theory of the form: If there are no rules in the domain theory to support the specific case decomposition used in the specification, the feature proofs cannot succeed. This can be fixed by either guiding the user during specification to use a set of conditions that are supported by the domain theory, or by allowing conditional features that are assigned if a feature holds under any (rather than all) of the conditions specified. Feature derivation runs into a similar problem in the case of partial specifications, where component behavior is not defined for all legal inputs. A feature is only assigned to a component if it can be derived in terms of the behavior of the component for every legal input value. Partial specifications are too logically weak to allow a feature to be derived in the case of the undefined behavior. This can be fixed by (1) disallowing partial specifications, (2) strengthening the postcondition for the purposes of feature generation, or (3) allowing conditional features as described above. There was also a problem where two components we specified in such a way that their specifications cause non-termination of rewriting. For example, a head component with input l and output m can be specified as: If this statement is used as a left-to-right rewrite rule, the rewriting system will not terminate. One possible solution is to build a timeout option into the rewriting system, similar to that used by Schuman and Fischer [37]. Finally, it should be noted that, in general, specification-based component retrieval is susceptible to loss of recall due to the semantic gap between a component and a specification [22]. A component is associated with a specification if the component correctly implements a specification. However, there is a gap that a query may fall into: it is possible that a component may satisfy a query that its specification does not satisfy. The effects of this situation cannot be evaluated in the experiments because we are working only with specifications. However, it should be noted as a potential limitation of the method. 8.3. Building Classification Schemes Discovering features is not an easy process. Formalizing an abstract concept that is shared among several components is difficult and picking a collection of useful ones is even harder. Several times during this investigation, the intuitive concept of a feature was disproved by the system. Recognizing the utility of hierarchical and complementary features is a good start toward building better schemes. However, tool support would be necessary to scale the method to larger libraries and more sophisticated classification schemes. There are several ways that the formalized classification framework provides a foundation for automated tools. The formal definitions of the necessary conditions can support analysis of the scheme. For example, a set of features can be proven to be mutually exclusive while their disjunction is a tautology. Therefore, every component would be assigned exactly one of these features. These features would provide coverage similar to the "facets" in faceted classification [34]. It is also possible to provide support for extending classification schemes within the framework. For example, if two distinguishable components are classified identically, it may be possible to identify the parts of the specifications that are distinguishable and automatically derive a new feature that represents this difference. 8.4. Signatures vs. Semantics In most specification-based component retrieval systems, the first filter used to reduce the library is signature matching [7, 43, 44]. Signature matching uses the types from a component's interface to determine its compatibility with a query. The guiding assumption is that if the types do not match there is no reason to further examine a component's behavior. We believe there are two cases where semantic classification has potential for greater recall than signature-based approaches. The retrieval performance of signature matching degrades when considering relevant components that could be combined or adapted to satisfy a query. Standard signature matching will eliminate relevant components with partial matches to an interface. Allowing partial matches would allow too many component to match, greatly decreasing the retrieval precision of signature matching. Because feature definitions are not required to constrain all inputs and outputs in a component, semantic classification does not have this problem. In fact, components are retrieved that provide slices of the appropriate behavior, independent of the type used. The type information (the feature values) is used to identify type substitutions before evaluation. Using signatures has been suggested to discover combinations of components in a library that match a query signature [15, 23]. We take a semantic approach to the problem of combining components [28, 29]. classification assists this approach by retrieving components that provide pieces of the appropriate behavior. Therefore, it can locate partial solutions even when the remainder of the solution is not in the library. 9. Related Work The use of formal specifications to assist software component retrieval has been widely proposed projects [3, 37, 16, 21]. 4 Zaremski and Wing [45] provided a foundation for studying the more general activity of specification matching, the verification of logical relationships between specifications. There are many approaches to making specification-based component retrieval more efficient. However, only a few of these methods make use of the semantics provided by formal specifications. The NORA/HAMMR deductive retrieval tool built by Fischer and Schumann uses a series of filters to identify reusable components [37]. This tool is novel in its use of model checking as one of the search filters. Before running the theorem prover to check the match condition, the conditions are checked by searching for a model in a small part of the specification theory. Since it is necessary for such a model to exist for the conditions to hold, only those components that pass the model checking stage need to be checked in the entire theory. They also use various techniques to improve the performance of the theorem prover during specification matching, such as reducing the set of axioms used and parallel proof attempts. Evaluation of prototype implementations on libraries of list manipulation functions showed very encouraging results. However, the recall (retrieved components/useful components) of the prototype was limited by signature matching. The notion of using filters to restrict the search space is consistent with our use of necessary conditions to eliminate non-matching component. Because a semantic-based classification scheme can provide better recall than signature matching, it could be used as a preliminary filter and potentially increase recall. The Inquire retrieval mechanism [32] within the Inscape [31] environment supports retrieval based on component specifications. Preconditions and postconditions for components are formulated in terms of a given set of formally defined logical predicates. An inference mechanism is used during the retrieval process to retrieve components that provide the various predicates. The formal predicate definitions are useful as unambiguous descriptions of the predicate vocabulary. The prototype implementation is reported to work very well, but with a restricted specification language. The restricted language reduces the inference power necessary for component retrieval. The methodology presented here uses a reversed approach from Inscape. Interface specifications are defined in full first-order logic. The formally defined features (predicates) are then assigned to the specifications if they are logically implied by the specification. In this case, the predicates are not the complete specification of the component, but represent various aspects of the component's function. Because feature predicates do not have to be useful as specification predicates, they can be more abstract, allowing more flexibility in the types of similarity that can be represented. In addition, the more expressive specification language allows precise evaluation of reusability. Deductive program synthesis [20, 40] also makes use of formal methods to automate software reuse. For example, the Amphion system [18, 41] successfully uses deductive synthesis to construct software from a subroutine library for solar system geometry. In these systems, components (language primitives or sub- routines) are represented as mathematical functions and their behavior specified via axioms. A program is synthesized by proving that, for any valid input, there exists an output that satisfies the specification. The occurrence of a primitive function in the proof constructed during deductive synthesis corresponds to a call to the associated subroutine. Therefore, a component is effectively "retrieved" when its corresponding axioms are used during the proof process. This means that the domain theory axioms and the tactics used in decomposing proofs will determine which components are used. Ongoing research is exploring the integration of architectural decomposition tactics with our current component retrieval system to support automated component adaptation and integration [28, 29, 30]. 10. Conclusion Software reuse and formal specification are two methodologies that show high potential impact on software productivity and reliability. Used together, they permit increased automation and assurance in the reuse process. We presented how component reusability and similarity can be described formally as logical relationships between the problem specification and a component specification. However, it is too computationally expensive to formally verify these relationships in the quantities required for practical component retrieval. Therefore, specification-based retrieval would benefit from a method to approximate these relationships and identify a subset of the library to undergo verification. In this paper, we described a method for classifying components based on their formal specifications. Features are assigned to components based on specific necessary conditions that are implied by the component specifications. The logical form of the specification matches determining reusability ensures that components with similar feature sets are more likely to match. The collection of necessary conditions that controls the classification scheme is defined formally to allow automated classification. The theorem proving required during classification is applied in such a way that complexity of component classification is much less than that of applying multiple specification matches. Once classified, components are retrieved via syntactic comparison of feature sets. The results of empirical evaluation on a library of list components show that the method can provide retrieval performance comparable to existing methods. The benefits are a faster response time than other formal approaches. The method improves upon informal methods by providing higher levels of consistency and automation. Our future work will focus on integrating specification-based component retrieval with support for automated component adaptation and integration [28]. A long-term goal is to develop support for run-time component integration in high-assurance component-based systems [19]. We are also investigating tools to support development and maintenance of formal classification schemes. Acknowledgments We would like to thank Bernd Fischer, Gary Leavens, Amy Moormann Zaremski, Ali Mili, Santos Lazzeri and Dale Martin for helpful comments on the development and evaluation of this research. We also thank the anonymous reviewers of the current and earlier versions of this work for their suggestions for the presentation and evaluation of the work. Support for this work was provided in part by the Advanced Research Projects Agency and monitored by Wright Labs under the RAASP Technology Program contract number F33615- 93-C-1316 and the CEENSS Technology Program contract number F33615-93-C-4304. Notes 1. We do not use Zaremski and Wing's method for identifying reuse matches based on syntactic patterns. We select matches based on formalization of intuitive notions of reusability, and their utility in component retrieval. 2. As opposed to free text indexing of source code and/or comments, which would not satisfy the high level of assurance required in this application. 3. It should be noted that distances in the lattice are meaningless; the lattice was arranged by hand to minimize the crossing of links. 4. For a general overview of component retrieval methods for assisting software reuse, see the survey by Mili et al [23] --R Validating component compositions in software system generators. Perlis, editors. Software Reusability - Concepts and Models Program development as a formal activity. Two theses of knowledge representation: languages restrictions NORA/HAMMR: Making deduction-based software component retrieval practical An empirical study of representation methods for reusable software components. Formal Foundations for the Specification of Software Architecture. Automatic indexing of software artifacts. Using english to retrieve software. HOL: A proof generating system for higher-order logic The Science of Programming. Languages and Tools for Formal Specification. Generalized behavior-based retrieval Using formal methods to construct a software library. A formal approach to domain-oriented software design environments Fundamentals of deductive program synthesis. A refinement based system. A survey of software reuse libraries. Reusing software: Issues and research directions. Another nail to the coffin of facated controlled-vocabulary component classification and retrieval The Definition of Standard ML. Correct architecture refinement. Automated Component Retrieval and Adaptation Using Formal Specifications. Toward automated component adaptation. Declarative specification of software architectures. The Inscape environment. Rub'en Prieto Rub'en Prieto Specifications as search keys for software libraries. Automated deduction and formal methods. NORA/HAMMR: Making deduction-based software component retrieval practical Software Architecture: Perspectives on an Emerging Discipline. Derived preconditions and their use in program synthesis. Deductive composition of astronomical software from subroutine libraries. Algebraic Specifications in Software Engineering: An Introduction. Signature and Specification Matching. Signature matching Specification matching of software components. --TR --CTR David Hemer, Semi-Automated Component-Based Development of Formally Verified Software, Electronic Notes in Theoretical Computer Science (ENTCS), 187, p.173-188, July, 2007 Robert G. Bartholet , David C. Brogan , Paul F. Reynolds, Jr., The computational complexity of component selection in simulation reuse, Proceedings of the 37th conference on Winter simulation, December 04-07, 2005, Orlando, Florida Sofien Khemakhem , Khalil Drira , Mohamed Jmaiel, SEC: a search engine for component based software development, Proceedings of the 2006 ACM symposium on Applied computing, April 23-27, 2006, Dijon, France Brandon Morel , Perry Alexander, SPARTACAS Automating Component Reuse and Adaptation, IEEE Transactions on Software Engineering, v.30 n.9, p.587-600, September 2004
component retrieval;formal specification;software reuse
592038
Efficient Implementations of Software Architectures via Partial Evaluation.
The notion of flexibility (that is, the ability to adapt to changing requirements or execution contexts) is recognized as a key concern in structuring software, and many architectures have been designed to that effect. However, the corresponding implementations often come with performance and code size overheads. The source of inefficiency can be identified to be in the loose integration of components, because flexibility is often present not only at the design level but also in the implementation. To solve this flexibility vs. efficiency dilemma, we advocate the use of partial evaluation, which is an automated technique to produce efficient, specialized instances of generic programs. As supporting case studies, we consider several flexible mechanisms commonly found in software architectures: selective broadcast, pattern matching, interpreters, software layers, and generic libraries. Using Tempo, our specializer for C, we show how partial evaluation can safely optimize implementations of those mechanisms. Because this optimization is automatic, it preserves the original genericity and extensibility of the implementation.
Introduction What is partial evaluation? Partial evaluation is a technique to partially execute a program, when only some of its input data are available. Consider a program p requiring two inputs, x 1 and x 2 . When specific values d 1 and d 2 are given for the two inputs, we can run the program, producing a result. When only one input value d 1 is given, we cannot run p, but can partially evaluate it, producing a version p d1 of p specialized for the case where x Partial evaluation is an instance of program specialization, and the specialized version p d1 of p is called a residual program. For an example, consider the following C function power(n, x), which computes x raised to the n'th power. int n,x; f int p; while (n ? if (n % 2 == else Given values 7, we can compute power(5,7), obtaining the result exploits that x even integers n). Suppose we need to compute power(n, x) for a great many different values of x. Then we can partially evaluate the function for obtaining the following residual function: int x; f int p; We can compute power 5(7) to obtain the result 7 In fact, for any input x, computing power 5(x) will produce the same result as computing Department of Computer Science, University of Copenhagen, Universitetsparken 2 Department of Mathematics and Physics, Royal Veterinary and Agricultural University, Thorvaldsensvej 40, DK-1871 Frederiksberg C, Denmark power(5,x). Since the value of variable n is available for partial evaluation, we say that n is static; conversely, the variable x is dynamic because its value is unavailable. This example shows the strengths of partial evaluation: in the residual program power 5, all tests and all arithmetic operations involving n have been eliminated. The flow of control (that is, the conditions in the while and if statements) in the original program was completely determined by the static variable n. Now suppose we needed to compute power(n,7) for many different values of n. This is the opposite problem of the above: now n is dynamic (unknown) and x is static (known). There is little we can do in this case, since the flow of control is determined by the dynamic variable n. One could imagine creating a table of precomputed values of 7 n for some values of n, but how are we to know which values are relevant? In many cases some of the control flow is determined by static variables, and in these cases substantial speed-ups can be achieved by partial evaluation. 1.1 Notation We can consider a program in two ways: as a function transforming inputs to outputs, and also as a data object, being input to or output from other programs. We need to distinguish the function computed by a program from the program text itself. Writing p for the program text, we write for the function computed by or when we want to make explicit the language L in which p is written. Consequently, denotes the result of running program p with input d on an L-machine. Now we can assert that power 5 is a correct residual program (in C) for power and given input 5: 1.2 Interpreters and compilers An interpreter Sint for language S, written in language L, satisfies for any S- program s and input data d: That is, running s with input d on an S-machine gives the same result as using the interpreter Sint to run s on an L-machine. This includes possible nontermination of both sides. A compiler STcomp for source language S, generating code in target language T , and written in language L, satisfies That is, p can be compiled to a target program p 0 such that running p 0 on a T-machine with input d gives the same result as running p with input d on an S-machine. Though the equation doesn't specify this, we normally assume compilation to always produce a target program. Partial evaluators 2.1 What is a partial evaluator? A partial evaluator is a program which performs partial evaluation. That is, it can produce a residual program by specializing a given program with respect to part of its input. Let p be an L-program requiring two inputs x 1 and x 2 as above. A residual program for p with respect to x is a program p d1 such that for all values d 2 of the remaining input, A partial evaluator is a program peval which, given a program p and a part d 1 of its input, produces a residual program p d1 . In other words, a partial evaluator peval must satisfy: peval This is the so-called partial evaluation equation, which reads as follows: If partial evaluation of p with respect to d 1 produces a residual program p d1 , then running d1 with input d 2 gives the same result as running program p with input [d 1 As for compilers, the equation does not guarantee termination of the left-hand side of the implication. In contrast to compilers we will, however, not always assume partial evaluation to succeed. While it is desirable for partial evaluation to always terminate, this is not guaranteed by a large number of existing partial evaluators. See sections 2.2 and 5.5 for more about the termination issue. Above we have not specified the language L in which the partial evaluator is written, the language S of the source programs it accepts, or the language T of the residual programs it produces. These languages may be all different, but for notational simplicity we assume they are the same, . Note that opens the possibility of applying the partial evaluator to itself (see below). For an instance of the partial evaluation equation, consider 5, then from must follow that power(5,7) 2.2 What is achieved by partial evaluation? The definition of a partial evaluator does not stipulate that the specialized program must be any better than the original program. Indeed, it is easy to write a program peval which satisfies the partial evaluation equation in a trivial way, by prepending a new 'specialized' function power 5 to the original program. The new function simply calls the original one with the given argument: int n,x; f int p; while (n ? if (n % 2 == else int x; f return(power(5, x)); g While this program is a correct residual program, it is no faster than the original program, and quite possibly slower. Even so, the construction above can be used to prove existence of partial evaluators, a proof similar to Kleene's (1952) proof of the s-m-n theorem [63], a theorem that essentially stipulates the existence of partial evaluators in recursive function theory. But, as the example in the introduction demonstrated, it is sometimes possible to obtain residual programs that are arguably faster than the original pro- gram. The amount of improvement depends both on the partial evaluator and the program being specialized. Some programs do not lend themselves to spe- cialization, as no computation can be done before all input is known. Sometimes choosing a different algorithm may help, but in other cases the problem itself is ill-suited for specialization. An example is specializing the power function to a known value of x, as discussed in the introduction. Looking at the definition of power, one would think that specialization with respect to a value of x would give a good result: the assignments, do not involve n, and as such can be executed during specialization. The loop is, however, controlled by n. Since the termination condition is not known, we cannot fully eliminate the loop. But x and p will have different values in different iterations of the loop, so we cannot replace them by constants. Hence, we find that we cannot perform the computations on x and p anyway. We could force unfolding of the loop to keep the values of x and p known, but since there is no bound on the number of different values x and p can obtain, no finite amount of unfolding can eliminate x and p from the program. This conflict between termination of specialization and quality of residual program is common. The partial evaluator must try to find a balance that ensures termination often enough to be interesting (preferably always) while yielding sufficient speed-up to be worthwhile. Due to the undecidability of the halting problem, no perfect strategy exists, so a suitable compromise must be found. See Section 5.5 for more on this subject. 3 Another approach to program specialization A generating extension of a two-input program p is a program p gen which, given a value d 1 for the first input of p, produces a residual program p d1 for p with respect to d 1 . In other words, The generating extension takes a given value d 1 of the first input parameter x 1 and constructs a version of p specialized for x As an example, we show below a generating extension of the power program from the introduction: int n; f printf("int x;"n"); printf("f int p;"n"); while (n ? if (n % 2 == else f printf(" printf(" return(p);"n"); Note that power-gen closely resembles power: those parts of power that depend only on the static input n are copied directly into power-gen, and the parts that also depend on x are made into strings, which are printed as part of the residual program. Running power-gen with input yields the following residual program: int x; f int p; This is almost the same as the one shown in the introduction. The difference is because we have now made an a priori distinction between static variables (n) and dynamic variables (x, p). Since p is dynamic, all assignments to it are made part of the residual program, even was executed at specialization time in the example shown in the introduction. Later we shall see that a generating extension can be constructed by applying a sufficiently powerful partial evaluator to itself. One can even construct a generator of generating extensions that way. 4 Partial evaluation, interpreters, and compila- tion 4.1 Compilation using a partial evaluator In Section 1.2 we defined an interpreter as a program taking two inputs: a program to be interpreted and input to that program: We often expect to run the same program repeatedly on different inputs. Hence, it is natural to partially evaluate the interpreter with respect to a fixed, known program and unknown input to that program. Using the partial evaluation equation we get [s; d] for all d Using the definition of the interpreter we get The residual program is thus equivalent to the source program. The difference is the language in which the residual program is written. If the input and output languages of the partial evaluator are identical, then the residual program is written in the same language L as the interpreter Sint. Hence we have compiled s from S, the language that the interpreter interprets, to L, the language in which it is written. 4.2 Compiler generation using a self-applicable partial eval- uator We have seen that we can compile programs by partially evaluating an in- terpreter. Typically we will want to compile many different programs, which amounts to partially evaluating the same interpreter repeatedly with respect to different programs. This situation calls for optimization by yet another application of partial evaluation. Hence we use a partial evaluator to specialize a partial evaluator peval with respect to a program Sint, but without the argument s of Sint. Using the partial evaluation equation we get: peval ]][peval; peval Sint Using the results from above, we get peval Sint for which we have d for all d We recall the definition of a compiler from Section 1.2: We see that peval Sint fulfills the requirements for being a compiler from S to T . In the case where the input and output languages of the partial evaluator are identical, the language in which the compiler is written and the target language of the compiler are both the same as the language L in which the interpreter is written. Note that we have no guarantee that partial evaluation terminates, neither when producing the compiler nor when using it. Experience has shown that while this may be a problem, it is often the case that if compilation terminates for a few general programs then it terminates for all. Note that the compiler peval Sint is a generating extension of the interpreter Sint , according to the definition shown in section 3. This generalizes to any program, not just interpreters: partially evaluating a partial evaluator peval with respect to a program p yields a generating extension p program. 4.3 Compiler generator generation Having seen that it is interesting to partially evaluate a partial evaluator, we may want to do this repeatedly: to partially evaluate a partial evaluator with respect to a range of different programs (e.g., interpreters). Again, we may exploit partial evaluation: peval peval Since which is a generating extension of p, we can see that peval peval is a generator of generating extensions. The program peval peval is itself a generating extension of the partial evaluator: peval peval peval . In the case where p is an interpreter, the generating extension p gen is a compiler. Hence, peval gen is a compiler generator, capable of producing a compiler from an interpreter. 4.4 Summary : The Futamura projections Instances of the partial evaluation equation applied to interpreters, directly or through self-application of a partial evaluator, are collectively called the Futamura projections. The three Futamura projections are: The first Futamura projection: compilation peval The second Futamura projection: compiler generation peval ]][peval; The third Futamura projection: compiler generator generation The first and second equations were devised by Futamura in 1971 [40], and the latter independently by Beckman et al. [12] and Turchin et al. [94] around 1975. 5 Techniques for partial evaluation 5.1 Polyvariant specialization Polyvariant specialization is a technique for partial evaluation which works for a range of languages. A program is thought of as a collection of program points , connected by control-flow edges . In a flow-chart language, program points and control-flow edges are, respectively, labelled basic blocks and jumps ; in a functional language, they are defined functions and function calls ; in a logic language, they are predicates and predicate applications (atoms). Polyvariant specialization constructs a residual program by creating zero or more specialized variants of each program point, and connecting them by residual control-flow edges. 5.1.1 The exponentiation example revisited To illustrate polyvariant specialization, consider the power function from Section 1 in flowchart form, with explicitly labelled basic blocks: lab1: if (n != if (n % 2 != goto lab1 goto lab1 lab3: return(p) A program on this form is specialized to given values of the static variables by specializing the basic blocks. For each basic block, and for each set of static variable values with which it may be executed, one creates a specialized basic block in the residual program. This is polyvariant specialization [24, 57]. For instance, the basic block labelled lab1 may be executed with static Hence one creates a specialized basic block, whose label lab1 fn=5;p=1g consists of the original label and bindings for the static variables. The body of the specialized basic block consists of the specialized residual commands from the original basic block. Naturally, the specialized version of a jump goto lab is itself a jump goto lab f:::g to a specialized (decorated) version of lab. To see how this works, let us specialize the above program with the known value and an unknown value for x. First we get to lab1 with Using this information to specialize that basic block, we perform the conditional ifs statically, find that (n != 0) is false and (n % 2 != 0) is true, and so must jump to lab2, still with We create a residual goto command, and a new specialized label lab2 fn=5;p=1g goto lab2 fn=5;p=1g The corresponding specialized basic block is the block at lab2 specialized with respect to 1. The assignment specializes to the residual since x is dynamic. This means that p is no longer static. The assignment can be executed because n is static. The new static environment has 4. Hence the goto lab1 specializes to goto lab1 fn=4g and we get: goto lab1 fn=4g Note that we had to generate a constant expression 1 to represent the static value 1 of p in the residual program. We say that the static value of p has been lifted to appear in the residual program. Next we must specialize the basic block at lab1 with respect to 4, and so on. This process continues until specialized basic blocks have been created for all specialized labels occurring in the residual program. In total, the following residual program is obtained: goto lab2 fn=5;p=1g goto lab1 fn=4g goto lab1 fn=2g goto lab1 fn=1g goto lab2 fn=1g goto lab1 fn=0g goto lab3 fn=0g return(p) This can be simplified by replacing jumps (gotos) with the code they jump to; this is called transition compression or unfolding . The result is almost as in Section 1: return(p) The technique of polyvariant specialization turns out to work for other languages too; this is demonstrated in later sections for functional languages and logic languages. The specialization process builds a graph whose nodes are specialized program points (labels), and whose edges are residual control-flow edges (jumps). This may be done by maintaining a set pending of the specialized program points still to be created, and a mapping out from specialized program points to specialized program code fragments (basic blocks). One repeatedly chooses and removes a program point pp from pending, constructs the corresponding specialized program code fragment code pp , and extends the mapping out with [pp 7! code pp ]. Moreover, one extends the set pending by any new specialized labels reachable from code pp . More precisely, pending is extended with the set is the set of program points pp 0 to which there is a jump goto pp 0 from code pp . To begin with, pending contains just the program's entry point together with the initial values of its static variables, and out is empty. The procedure terminates if and when pending becomes empty, in which case out contains the residual program. This process may fail to terminate, as discussed in Section 5.5 below. 5.2 Online versus offline partial evaluation There are two types of partial evaluators. An online partial evaluator is a kind of generalized interpreter, which needs no a priori division of variables into static and dynamic. During partial evaluation, the environment maps static variables to concrete values, and dynamic variables to symbolic expressions. When processing an expression e, the partial evaluator makes an online decision whether to evaluate it (giving a concrete value), or to residualize it (giving a residual expression), based on the current bindings of the variables appearing in e. An offline partial evaluator, by constrast, works in two phases. The first phase is a binding-time analysis , which classifies the program's variables into (definitely) static and (possibly) dynamic, and similarly classifies all operations. The second phase is the specialization proper. This phase simply uses the static/dynamic classification of variables and operations when processing an ex- pression; all evaluate/residualize decisions have been made offline. It never uses the actual value of a variable or expression, unless the binding-time analysis guarantees that it is static and hence indeed is a concrete value. Since offline partial evaluators rely on a program analysis, they are usually more conservative than online partial evaluators, missing some opportunities for specialization. On the other hand, offline specializers have a simpler structure, and may exploit the global knowledge about the program gained by the binding-time analysis. Experience shows that it is harder to construct self-applicable online specializers than offline ones. Hybrids of online and offline specializers have been constructed. For instance, one may use a three-valued binding-time analysis, which classifies variables and expressions as 'definitely static', `definitely dynamic', or 'undecided' [92]. The specialization phase will just obey the static and dynamic annotations, but use the actual (specialization time) value of variables to decide whether to evaluate or residualize. A generator of generating extensions is similar to an offline partial evaluator, since a generating extension embodies an a priori distinction between early (static) inputs and late (dynamic) inputs. A generator of generating extension usually includes a binding-time analysis. Online partial evaluation has been studied for Scheme by Ruf and Weise [82, 100] and by many researchers in the logic programming community. 5.3 Binding-time analysis The classification of variables into static and dynamic is called a division. The division must be congruent : if the value of some dynamic expression e may be assigned to a variable y, then y must be made dynamic. The expression e is static if it contains no dynamic variables. Considering again the flow-chart version of the power function in Section 5.1, we see that if n is static and x is dynamic from the outset, then p must be classified as dynamic because p x is assigned to p, whereas n remains static: n is never assigned a dynamic value. A simple binding time analysis may be performed by means of an abstract interpretation in which each variable and expression takes one of the abstract values S (for static) or D (for dynamic). One builds an initial division in which all variables are S, except for the dynamic input parameters. Now all assignments in the program are abstractly executed, possibly reclassifying variables as dynamic to satisfy the congruence requirement, until no more variables need to be reclassified as dynamic. Alternatively, binding-time analysis may be done by type inference with sub- types, where S is considered a subtype of D, meaning that S may be coerced to D (corresponding to the lifting of a static value). This kind of binding-time analysis may be implemented efficiently by constraint solving [53]. When composite data structures (tuples, records, lists) are considered, a data structure may be partially static. For instance, the value of a variable may be a whose first component is static, and whose second component is dynamic. This may be described by the binding-time S \Theta D. Similarly, a list of such pairs may be described by the binding-time (S \Theta D) list. The type inference approach to binding-time analysis is especially useful for handling partially static data structures in strongly typed languages, such as Standard ML, Pascal, or C. Latently typed languages, such as Scheme, are handled essentially by considering dynamic expressions to be untyped. When a division has been computed by the binding-time analysis, one must decide for each operation in the program whether it must be evaluated or re- sidualized (producing residual code) during partial evaluation. An arithmetic operation must be residualized unless all its operands are static. An if statement must be residualized unless the condition is static. We shall assume that an assignment will be residualized unless the assigned variable is static. We also assume that all gotos are residualized (any excess gotos may be removed by subsequent transition compression). To visualize the classification of operations, we annotate the dynamic operations by underlining. For the power function, the annotation would be: lab1: if (n != if (n % 2 != goto lab1 goto lab1 lab3: return(p) Doing polyvariant specialization of this program with blindly following the annotations, we obtain (after transition compression): return(p) which is just the result obtained by the generating extension in Section 3. This is because the generator of generating extensions presuppose the a priori distinction between the static (n) and dynamic (x; p) variables. 5.4 Residual programs containing loops The residual program generated above contains no loops; all conditionals were statically decidable and all transitions could be compressed. However, the machinery in Section 5.1 suffices for creating residual programs containing loops. Consider the following contrived example: while (n ? else Written as a flow-chart, the program is lab1: if (n != goto lab3 lab2: goto lab3 lab3: goto lab1 lab4: return(sum) Let us specialize it with respect to static dynamic sum and n. Specializing the basic block at lab1 with respect to must create a residual version of the first conditional, because n is dynamic, whereas the second conditional can be reduced, because k is static and non-zero, giving a residual jump to lab2 fk=3g if (n != goto lab2 fk=3g Next we specialize the code at label lab2 with respect to goto lab3 fk=3g Continuing in this manner, we obtain this residual program: if (n != goto lab2 fk=3g goto lab3 fk=3g goto lab1 fk=3g return(sum) After transition compression, we get: if (n != goto lab1 fk=3g return(sum) The decorated labels lab1 fk=3g and lab4 fk=3g may be replaced by simple ones, such as lab1r and lab4r. Then we see that partial evaluation has eliminated the tests on k inside the loop; effectively, they were found to be loop-invariant. The loop is recreated in the residual program simply because the jump at lab3 fk=3g goes back to the specialized program point lab1 fk=3g at the beginning of the program. 5.5 Termination of partial evaluation Transition compression should be applied with care in the program just shown. An attempt to (repeatedly) unfold all remaining occurrences of goto lab1 fk=3g would never terminate. Infinite looping due to transition compression is avoided fairly easily; either by unfolding a jump to a (residual) label only if there is exactly one way to reach that label [21], or by ascertaining that unfolding must stop due to some descending chain condition [87]. Termination problems caused by infinite specialization are harder to deal with. For illustration, consider again the power program in Section 5.1, but now with static straightforward application of polyvariant specialization will attempt to produce an infinite residual program: if (n != if (n % 2 != goto lab1 fx=49;p=1g goto lab1 fx=7;p=7g if (n != if (n % 2 != goto lab1 fx=2401;p=1g goto lab1 fx=49;p=49g This program is incomplete, and it cannot be completed using a finite number of program points, if we insist on keeping x and p static. In an online partial evaluatior, one may recognize that the configuration is 'similar' to the previously encountered and that the two program points should therefore be merged into a single more general one, e.g. by making x dynamic (which eventually forces p to be dynamic also). This process is called generalization. In an offline partial evaluator, one may recognize after binding-time analysis that the tests are dynamic (not decided by the static variables), and that static data are constructed under dynamic control. This is a sign of danger, indicating that x and p should be made dynamic too, making specialization completely trivial (but safe). Holst developed a finiteness analysis and used it to ensure termination of polyvariant specialization [54]. 5.6 Generalized partial evaluation One more lesson may be learnt from the (partially constructed) residual program just shown. The basic block labelled lab3 fx=49;p=1g is superfluous. Reaching it would require the tests (n != 0) and (n % 2 != 0) to fail and the test ((n/2 != to succeed, which is impossible for integral n; the former two imply that 2. The superfluous basic block is created because the static environment (as outlined above) takes into account only the values of static variables (x and p), not the outcome of previously encountered dynamic tests (on n). Polyvariant specialization may be enhanced to do so, giving generalized partial evaluation. Then a theorem prover is required to decide static conditionals and to decide whether two static environments are equivalent [41]. In certain data domains and applications, less powerful methods may suffice [46]. 6 Partial evaluation for other languages 6.1 Functional languages 6.1.1 First-order languages Partial evaluation of a first-order functional language may be done by polyvariant specialization as described in Section 5.1 above. The notions of label , basic block , and global variable must be replaced by the notions of function name, function definition, and function parameter. Henceforth a specialized program point is a specialized function name, and a residual program is a collection of specialized function definitions. For illustration, consider a functional version of the power program from Section 1, here using Standard ML syntax: else if n mod else x * power(n-1, x) Specializing the function power with respect to static dynamic x, we obtain fun power fn=5g and power fn=4g and power fn=2g and power fn=1g and power fn=0g Note that a specialized function name power fn=5g consists of an original function name power together with a binding for the static parameters, here just n. The residual program may be simplified by unfolding trivial function calls (and reducing the subexpression (x * x) * 1 arising from this unfolding): fun power fn=5g and power fn=2g This residual program is equivalent to that generated for the C version of power in Section 1. Binding-time analysis may proceed as for a flow-chart language. For each application (f e) where e may be dynamic, reclassify the formal parameter of f to dynamic. Since the language is first-order, f must be a known function. As for flow-chart languages, a partial evaluator may either be offline or online. An offline partial evaluator will perform a binding-time analysis of the program, to classify all parameters as either static or dynamic, before embarking on the specialization phase proper. A complete description of a simple offline self- applicable partial evaluator for a first-order functional language may be found in [58, Chapter 5 and Appendix A]. 6.1.2 Higher-order functional languages Polyvariant specialization can be applied to higher-order functional languages (in which functions may be passed around as values) as well. The main new challenges are: how to represent static functional values during partial evalu- ation, how to lift functional values from static to dynamic, how to specialize with respect to functional values, and how to do binding-time analysis. A functional value may be represented by a closure (g; vs) consisting of a function name g together with the values vs of the static free variables in g's body. Lifting of a (partially) static functional value to a dynamic value is complicated and is usually avoided in offline partial evaluators, by requiring that every (partially) static functional value must be applied to an argument. Any functional value occurring in a dynamic context will be reclassified as dynamic by the binding-time analysis. Specializing a function f with respect to a fully static functional closure (g; vs) is simple; just specialize with respect to the function name g and the values vs of the (static) free variables. Specializing f with respect to a partially static (g; vs) is more involved, since the body of g may have dynamic free variables. These variables may be free also in the residual expression resulting from applying (g; vs). Hence the dynamic must be lifted out of g's body at specialization time, and must be passed as extra parameters to the residual function f (g;vs) . A higher-order functional program may contain applications evaluates to some function. A closure analysis can provide an approximation to the set of functions that e 1 may evaluate to; using this information, binding-time analysis may proceed as for a first-order language [58, Chapter 15]. Alternatively, the binding-time analysis may be based on type inference [53]; this is preferable if one wants to permit partially static data structures also. Self-applicable partial evaluators exist for realistic higher-order functional languages such as Scheme [20, 21, 28, 29, 31] and Standard ML [70] as well as for the call-by-value lambda calculus with some extensions [51], and for the pure lambda calculus [74, 75]. For further information, see e.g. [58, Chapter 10]; for full details, see the above-mentioned papers. 6.2 Logic programming languages (Prolog) A distinguishing feature of Prolog and other logic programming languages is the ability to run with incomplete input. While this seems similar to partial evaluation, there are a number of differences: ffl The result of running a Prolog program with incomplete input is a (pos- sibly) infinite list of instantiations of both input and output variables. Though this can be considered a list of facts, and hence a restricted form of program, we generally want a partial evaluator to be able to produce non-trivial residual programs, possibly containing loops. ffl Prolog has some non-logical features that means that running a program with incomplete input is not a generalization of running with complete input. As an example, calling the predicate defined by with partially instantiated input p(A,a) returns the result running with complete input p(a,a) would fail. Most of the research in partial evaluation (or partial deduction, as it is often called) of logic languages has tended to avoid the second issue by working with pure logic languages [43, 68]. Some systems, however, deal with non-logical features of Prolog [76, 84]. Partial evaluation of logic languages is typically done using the same basic techniques as for functional languages: call unfolding and polyvariant specializ- ation, program points being predicates. A major source of speed-up in partially evaluating logic programs is the ability to detect failing computations at specialization time, and cut these away in the residual program. This way, not only static computations but also dynamic computations in failing branches can be eliminated by partial evaluation. This makes the potential speed-up by partial evaluation greater in logic languages than in functional or imperative languages. Online specialization has been the preferred technique in the logic language community, usually combined with powerful techniques for avoiding non-termination [23, 71]. In logic languages, online specialization presents more opportunities for specialization than offline specialization, because unification will often instantiate otherwise dynamic variables. When self-application has been a major goal, offline specialization has been used also [52, 76]. An example of Prolog specialization is shown below. It specializes a program for regular expression matching accepts(R,[]) :- nullable(R). accepts(R,[C-S]) :- first(R,C), next(R,C,R1), accepts(R1,S). The program takes a regular expression and a string as arguments. If the string is empty, the regular expression is tested for nullability (acceptance of empty string). If the string starts with a character C, it is tested whether this is among the first set of the regular expression. If this is the case, a new regular expression for matching the rest of the string is produced by next. The predicates nullable, first and next are not shown, but note that the set of Cs for which first succeeds is determined by R. Hence, partial evaluation of first with respect to a known R and unknown C will yield a number of instantiations of C. Specializing the program above with respect to R being the regular expression (ajb) aba yields the following residual program: accepts-3([]). Since nullable depends only on static values, it is completely eliminated, only visible as failed or true cases for the empty string. The call to first has instantiated C with a or b. This instantiation has made it possible to fully evaluate next, which has yielded a total of four different regular expressions, each giving rise to a specialized version of accepts. For accepts-0, the regular expression is (ajb) aba, for accepts-1 it is (ajb) abajba, for accepts-2 it is (ajb) abaja, and for accepts-3 it is (ajb) abajbajffl. 6.3 A full imperative language (C) We have seen that polyvariant specialization suffices for partial evaluation of flow-chart languages, and hence for simple imperative languages. A realistic imperative language, such as C, includes composite data structures (records and indexed arrays), pointers and dynamic data structures, functions which may have side effects on global variables, etc. An offline partial evaluator for C needs a sophisticated binding-time analysis to deal with pointers and composite data structures. For instance, a pointer variable p may be dynamic, or the pointer may be static but point to a dynamic object, or both the pointer and the pointed-to object may be static. The binding-time analysis may require programs to be 'well-behaved'. Assume that a is an array, and that the program contains an assignment of the form e, where e is dynamic. Then in principle any variable in the program may become dynamic as a result of this assignment, in case the address a[n] is outside the allocated array a. This would be too conservative, making partial evaluation trivial. Instead, one should require programs to be well-behaved, so that any such address is indeed inside a. For a taste of the difficulties caused by the combination of non-local side effects and (recursive) functions, consider a function which has a side effect on a static global variable, but where the side effect is controlled by some dynamic expression dyn: int global; int stmts else After the call to foo, the value of global may be either 1 or -1, but we cannot know which one at partial evaluation time, because dyn is dynamic. The simplest solution is to reclassify global as dynamic, but this wastes static information which might be useful when partially evaluating stmts. Another solution is to unfold the call to foo, giving a residual program of this form: int global; int f if dyn f stmts fglobal=1g ; g else f stmts fglobal=\Gamma1g ; g Here stmts fglobal=1g is a specialized version of stmts. However, when function foo is recursive, such unfolding is impossible. A third solution is to introduce a (dynamic) continuation variable cont, which will be assigned a different value in each branch of the residual version foo' of foo. In function main, after the call to foo', there will be a switch on cont: int global; int switch (cont) f case 1: stmts fglobal=1g ; break; case 2: stmts fglobal=\Gamma1g ; break; However, this will not work when foo is recursive, and the recursion is under dynamic control, since the number of paths through foo will not be statically bounded in that case. Hence for recursive procedures, the only feasible option may be to reclassify global as dynamic. These and many other problems were studied by Lars Ole Andersen, who constructed two systems for specialization of C programs. The first one is a self-applicable partial evaluator for a C subset, including procedures as well as pointers and arrays [6, 8]. The second one is a generator of generating extensions for all of ANSI C [9]; the latter system can be licensed from the University of Copenhagen. The techniques for C should carry over to e.g. Ada, Modula, or Pascal with little modification, but to our knowledge this has not been done. 7 Partial evaluation in perspective 7.1 Program specialization without a partial evaluator So far we have focused mainly on specialization using a partial evaluator. But the ideas and methods presented here can be, and indeed have been, used without using a partial evaluator. Specialization by hand It is quite common for programmers to hand-tune code for particular cases. Often this amounts to doing partial evaluation by hand. As an example, here is a quote from a paper [80] about the programming of a video-game: How Nevryon manages to keep up its speed Basically there are two ways to write a routine: It can be one complex multi-purpose routine that does everything, but not quickly. For example, a sprite routine that can handle any size and flip the sprites horizontally and vertically in the same piece of code. Or you can have many simple routines each doing one thing. Using the sprite routine example, a routine to plot the sprite one way, another to plot it flipped vertically and so on. The second method means more code is required but the speed advantage is dramatic. Nevryon was written in this way and had about separate sprite routines, each of which plotted sprites in slightly different ways. specialization is used. But it is doubtful that a general purpose partial evaluator was used to do the specialization. Instead the specialization has been performed by hand, possibly without ever explicitly writing down the general purpose routine that forms the basis for the specialized routines. Using hand-written generating extensions We saw in Section 3 how a generating extension for the power function was easily produced from the original code, using knowledge about which variables contained values known at specialization time. While it is not always quite so simple as in this example, it is often not particularly difficult to write generating extensions of small to medium sized procedures or programs. In situations where no partial evaluator is available, this is often a viable way to obtain specialized programs. Using a generating extension instead of writing the specialized versions by hand is useful when either a large number of variants must be generated, or when it is not known in advance what values the program will be specialized with respect to. A common use of hand-written generating extensions is for run-time code generation, where a piece of specialized code is generated and executed, all at run-time. As in the sprite example above, one often generates specialized code for each plot operation when large bitmaps are involved. The typical situation is that a general purpose routine is used for plotting small bitmaps, but special code is generated for large bitmaps. The specialized routines can exploit knowledge about the alignment of the source bitmap and the destination area with respect to word boundaries, as well as clipping of the source bitmap. Other aspects such as scaling, differences in colour depth etc. have also been targets for run-time specialization of bitmap-plotting code. Hand-written generating extensions have been used for optimizing parsers by specializing with respect to particular tables [78], and for converting interpreters into compilers [77]. Handwritten generating extension generators In recent years, it has become popular to write a generating extension generator instead of a partial evaluator [9, 16, 55], but the approach itself is quite old [12]. A generating extension generator can be used instead of a traditional partial evaluator as follows. To specialize a program p with respect to data d, first produce a generating extension p gen , then apply p gen to d to produce a specialized program p d . Conversely, a self-applicable partial evaluator can produce a generating extension generator (cf. the third Futamura projection), so the two approaches seem equally powerful. So why write a generating extension generator instead of a self-applicable partial evaluator? Some reasons are: ffl The generating extension generator can be written in another (higher level) language than the language it handles, whereas a self-applicable partial evaluator must be able to handle its own text. ffl For this reason, among others, it may be easier to write a generating extension generator than a self-applicable partial evaluator. ffl A partial evaluator must contain an interpreter, which may be problematic for typed languages, as explained below. Neither the generating extension generator, nor the generating extensions, need to contain an interpreter. When writing an interpreter for a strongly typed language, one must use a single type in the interpreter to represent an unbounded number of types used in the programs that are interpreted. The same is true for a partial evaluator: a single universal type must be used for the static input to the program that will be specialized. Hence, the static input must be coded. This means that the partial evaluation equation must be modified to take this coding into account: peval where overlining means that a value is coded, e.g. d 1 is the coding of the value of d 1 . When self-applying the partial evaluator, the static input is a program. The program is normally represented in a special data type that represents program text. This data type must now be coded in the universal type: peval ]][peval; This double encoding is space- and time-consuming, and has been reported to make self-application intractable, unless special attention is paid to make the encoding compact [67]. A generating extension produced by self-application must also use the universal type to represent static input, even though this will always be of the same type. This observation leads to the idea of making generating extensions accept uncoded static input. To achieve this, the generating extension generator simply copies the type declarations of the original program into the generating extension. The generating extension generator takes a single input: a program, and need not deal with arbitrarily typed data. A generating extension handles values from a single program, the types of which are known when the generating extension is constructed. Hence, neither the generator of generating extensions, nor the generating extensions themselves, need to handle arbitrarily typed values. The equation for specialization using a generating extension generator is shown below. Note the absence of coding. We will usually expect generator generation to terminate, but, as for normal partial evaluation, allow the construction of the residual program (performed by gen ) to loop. 7.2 When is partial evaluation worthwhile? In Section 2.2 we saw that we cannot always expect speed-up from partial eval- uation. Sometimes no significant computations depend on the known input only, so virtually all the work is postponed until the residual program is executed. Even if computations appear to depend on the known input only, evaluating these during specialization may require infinite unfolding (as seen in Section 2.2) or just so much unfolding that the residual programs become intractably large. On the other hand, the example in Section 1 manages to perform a significant part of the computation at specialization time. Even so, partial evaluation will only pay off if the residual program is executed often enough to amortize the cost of specialization. So, we must have two conditions before we can expect any benefit from partial evaluation: There are computations that depend only on static data. These are executed repeatedly, either by repeated execution of the program as a whole, or by repetition (looping or recursion) within a single execution of the program. The static (known) data can be obtained in several ways: it may be constants appearing in the program text or it can be part of the input. It is quite common that library functions are called with some constant para- meters, such as format strings, so in some cases partial evaluation may speed up programs even when no input is given. In such cases the partial evaluator works as a kind of optimizer, often achieving speed-up when most optimizing compilers would not. On the other hand, most partial evaluators may loop or create an excessive amount of code while trying to optimize programs, and hence are ill-suited as default optimizers. Specialization with respect to partial input is the most common situation. Here there are often more opportunities for speed-up than just exploiting constant parameters. In some cases (e.g., when specializing interpreters) most of the computation can be done during partial evaluation, sometimes yielding speed-ups by an order of magnitude or more, similar to the speed difference between interpreted and compiled programs. When you have a choice between running a program interpreted or compiled, you will choose the former if the program is only executed a few times and contains no significant repetition, whereas you will want to compile it if it is run many times or involves much repetition. The same principle carries over to specialization. Partial evaluation often gets most of its benefit from replication: loops are unrolled and the index variables exploited in constant folding, or functions are specialized with respect to several different static parameters. In some cases this replication can result in enormous residual programs, which may be undesirable even if much computation is saved. In the example in Section 1 the amount of unrolling and hence the size of the residual program is proportional to the logarithm of n, the static input. This expansion is small enough that it doesn't become a problem. If the expansion were linear in n, it would be acceptable for small values of n. Specialization of interpreters typically yield residual programs that are proportional to the size of the source program, which is reasonable. On the other hand, quadratic or exponential expansion is hardly ever acceptable. It may be hard to predict the amount of replication caused by a partial evaluator. In fact, seemingly innocent changes to a program can dramatically change the expansion done by partial evaluation, or even make the difference between termination or nontermination of the specialization process. Similarly, small changes can make a large difference in the amount of computation that is performed during specialization and hence the speed-up obtained. This is similar to the way parallelizing compilers are sensitive to the way programs are writ- ten. Hence, specialization of off-the-shelf programs often require some (usually minor) modification to get optimal benefit from partial evaluation. Ideally, the programmer should write his program with partial evaluation in mind, avoiding the structures that can cause problems, just like programs for parallel machines are best written with the limitations of the compiler in mind. 7.3 Partial evaluation, optimizing compilers, and modern machines Many compilers perform transformations such as constant folding and inlining (of small functions) to improve target programs. These transformations are similar to some of those performed by a partial evaluator. However, in contrast to a partial evaluator, a compiler rarely produces more than one specialized version of a given piece of code (except possibly by inlining). This kind of specialization is essential in partial evaluators, and must be handled correctly also in the presence of loops and recursive procedures. With the complex memory hierarchies of modern computer hardware it is hard to know when a program modification actually achieves a speed-up. Exploiting the memory hierarchy well (data registers and instruction pipeline, two or more levels of cache, main memory, and virtual memory) is crucial for the performance of modern machines. Hence it may be detrimental to unroll a loop so that it does not fit in the cache, but beneficial to inline a procedure if this replaces indirect jumps by linear code sequences. How much unrolling, inlining, or replication to perform is machine dependent, and you often see optimizations that improve performance on one machine but degrade it on others. With the increasing degree of micro-parallelism in modern microprocessors, one may even get no benefit from eliminating the static computations, as they may not be part of the critical path and hence may be executed in parallel with the dynamic computations. On the other hand, the elimination of variables by specialization reduces register pressure, and unrolling of loops and inlining of functions increase basic block size, giving more opportunities for low-level optimization. This means that it is hard to predict the amount of speed-up obtained by partial evaluation. Examples exist where a residual program is twice as fast as the original program on one machine is, and slower than the original on another machine. The speed-up is also affected by the optimizations performed when compiling the residual programs. 8 Applications of partial evaluation We saw in Section 4 that partial evaluation can be used to compile programs and to generate compilers. This has been one of the main practical uses of partial evaluation. Not for making compilers for C or similar languages, but for rapidly obtaining implementations of acceptable performance for experimental or special-purpose languages. Since the output of the partial evaluator typically is in a high-level language, a traditional compiler is used as a back-end for the compiler generated by partial evaluation [1, 14, 25, 27, 30, 33, 61]. In some cases, the compilation is from a language to itself. In this case the purpose is to make certain computation strategies explicit (e.g., continuation passing style) or to add extra information (e.g., for debugging) to the program [20, 42, 83, 93]. Many types of programs, e.g. scanners and parsers, use a table or other data structure to control the program. It is often possible to achieve speed-up by partially evaluating the table-driven program with respect to a particular table [7, 78]. However, this may produce very large residual programs, as tables (unless sparse) often represent the information more compactly than does code. These are examples of converting structural knowledge representation to procedural knowledge representation. The choice between these two types of representation has usually been determined by the idea that structural information is compact and easy to modify but slow to use while procedural information is fast to use but hard to modify and less compact. Automatically converting structural knowledge to procedural knowledge can overcome the disadvantage of difficult modifiability of procedural knowledge, but retains the disadvantage of large space usage. Section 7.1 mentioned a few applications of specialization to computer graph- ics. This has been one of the areas that have seen most applications of partial evaluation. An early example is [49], where an extended form of partial evaluation is used to specialize a renderer used in a flight simulator. In a flight simulator the same landscape is viewed repeatedly from different angles. Though the occlusion of surfaces depend on the angle of view, it is often the case that the knowledge that a particular surface occludes (or not) another can decide the occlusion question of other pairs of surfaces. Hence, the partial evaluator simulates the sorting of surfaces and when it cannot decide which of two surfaces must be plotted first, it leaves that test in the residual program. Furthermore, it uses the inequalities of the occlusion test as positive and negative constraints in the branches of the conditional it generates. These constraints are then used to decide later occlusion tests (by attempting to solve the constraints by the Simplex method). Each time a test cannot be decided more information is added to the constraint set (which effectively constrains the view-angle), allowing more later tests to be decided. Goad reports that for a typical landscape with 1135 surfaces (forming a triangulation of the landscape) the typical depth of paths in the residual decision tree was 27, compared to the more than 10000 comparisons needed for a full sort [49]. This rather extreme speed-up is due to the nature of landscapes: many surfaces are almost parallel, and hence can occlude each other only in very narrow viewing angles. Another graphics application has been ray-tracing. In ray-tracing, a scene is rendered by tracing rays (lines) from each pixel on the screen into an imaginary world behind the scene, testing which objects these rays hit. The process is repeated for all rays using the same fixed scene. Since there may be millions of pixels (and hence rays) in a typical ray-tracing application, specialization with respect to a fixed scene but unknown ray can give speed-up even for rendering single pictures. Speed-ups of more than 6 have been reported for a simple ray-tracer [73]. For a more realistic ray-tracer, speed-ups in the range 1.5 to 3 have been reported [10]. The speed-up is gained from several sources: the ray/object intersection routine is specialized for each object and the (highly parametrized) shading (colouring) function is specialized for each object. Furthermore, the representation of the scene is converted to procedural form. Figure 1 is an example of a ray-traced picture made by the ray-tracer from [73]. The picture shows a 3D diagram of the process of partial evaluation: A program P and one of its inputs x are fed to the partial evaluator PE yielding a residual program P x . Partial evaluation has also been applied to numerical computation, in particular simulation programs. In such programs, part of the model will be constant during the simulation while other parts will change. By specializing with respect to the fixed parts of the model, some speed-up can be obtained. An example is the N-body problem, simulating the interaction of moving objects through gravitational forces. In this simulation, the masses of the objects are constant, whereas their position and velocity change. Specializing with respect to the mass of the objects can speed up the simulation. Berlin reports speed-ups of more than 30 for this problem [15]. However, the residual program is written in C whereas the original one was in Scheme, which may account for part of the speed-up. In another experiment, specialization of some standard numerical algorithms gave speed-ups ranging from none at all to about 5 [47]. When neural networks are trained, they are usually run several thousand times on a number of test cases. During this training, various parameters will be fixed, e.g. the topology of the net, the learning rate and the momentum. By specializing the trainer to these parameters, speed-ups of 25 to 50% are reported [56]. This list of applications is not exhaustive, but should give an impression of the range of possibilities. 9 Further reading Here we first sketch the history from 1952 to 1984, then give a number of pointers to the literature on partial evaluation and some related topics. The book by Jones, Gomard, and Sestoft [58] includes more material on the subjects mentioned above, and a large bibliography; the updated source text for that bibliography is available for anonymous ftp from ftp.diku.dk as file pub/diku/dists/jones-book/partial-eval.bib.Z. 9.1 History Kleene's s-m-n theorem (1952) asserts the feasibility of partial evaluation [63], and his constructive proof provides the design for a partial evaluator. This design did not, and was not intended to, provide any improvement of the specialized program. Such improvement, by symbolic reductions or similar, has been the goal in all subsequent work in partial evaluation. Lombardi is probably the first one to use the term 'partial evaluation' [69]. Futamura is the first researcher to consider a partial evaluator as a program as well as a transformer, and thus to consider the application of the partial evaluator to itself [40]. Futamura's paper gives the equations for compilation and compiler generation by partial evaluation, but not for compiler generator generation. The Figure 1: Partial evaluation in action three equations were called the Futamura projections by Andrei Ershov [38]. Futamura's early ideas were not implemented. Around 1975, Beckman, Haraldsson, Oskarsson, and Sandewall developed a partial evaluator called Redfun for a substantial subset of Lisp [12], and described the possibility of compiler generator generation by double self-application. Turchin and his group also worked with partial evaluation in the early 1970s, in the context of the functional language Refal, and gave a description of self-application and double self-application [94]. The history of that work is briefly summarized in English in [95]. Andrei Ershov worked with imperative languages, and used the term mixed computation to mean roughly the same as partial evaluation [34, 35]. In 1984, Jones, Sestoft, and S-ndergaard constructed a self-applicable partial evaluator for a simple first-order functional language [59, 60, 86]; until then neither single nor double self-application had been carried out in practice. At the same time the interest in partial evaluation in logic programming and other areas was increasing. This was the background for the 1987 Workshop on Partial Evaluation and Mixed Computation [19, 39]. Subsequent proceedings on partial evaluation may be found in [2, 3, 4, 5, 32, 88]. 9.2 Partial evaluators Imperative languages: Early papers on partial evaluation for imperative languages include [34, 36, 37]. Bulyonkov and Ershov reported a self-applicable partial evaluator for a flow chart language [25]; so did Gomard and Jones [50]. Gl-uck et al. created a (non-self-applicable) specializer for numeral algorithms in Fortran [11, 47]. Andersen [6, 8, 9] developed two systems for specialization of C programs; see Section 6.3. Lisp and Scheme: The first major partial evaluator for Lisp was Redfun, reported by Beckman et al. [12]. Weise et al. constructed a fully automatic online partial evaluator for a subset of Scheme [100]. Jones et al. constructed a self- applicable partial evaluator for a first-order functional language [59, 60]; Romanenko improved it in various respects [81]. Consel constructed the self-applicable partial evaluator Schism for a Scheme subset, handling partially static structures and polyvariant binding times [28, 29, 31]. Bondorf and Danvy constructed the self-applicable partial evaluator Similix for a subset of Scheme [20, 21]. Standard ML: Danvy, Heintze, and Malmkjaer developed the partial evaluator Pell-Mell [70]. Birkedal and Welinder created a generator of generating extensions [17, 18]. Refal and supercompilation: Turchin created the Refal language and developed the program transformation techniques of driving and supercompilation, which generalize partial evaluation [95, 96, 97]. A number of recent surveys on driving and supercompilation exist [48, 89, 90, 91]. Prolog partial evaluation was pioneered by Komorowski [64, 65]; subsequent work on Prolog includes [13, 44, 45, 66, 93, 98, 99]. Sahlin constructed a practical but non-self-applicable partial evaluator for full Prolog [84, 85]. Bondorf and Mogensen [76] constructed a self-applicable partial evaluator for a Prolog subset, Gurr one for the logic language G-odel [52]. J-rgensen and Leuschel created a generator of generating extensions for Prolog [62]. 9.3 Related topics McCarthy used program transformation rules in calculational proofs for recursive functional programs [72]. Boyer and Moore automated some proofs of this kind [22]. Burstall and Darlington viewed 'manual' program transformation as the application of a few types of meaning-preserving program rewritings: definition, instantiation, unfolding, folding, abstraction, and laws [26]. Partial evaluation specializes a program forwards by using knowledge about available input. Conversely, program slicing specializes a program backwards, using knowledge about the demand for output [79]. --R A compiler based on partial evalu- ation New Haven Partial evaluation of C and automatic compiler generation (extended abstract). Program Analysis and Specialization for the C Programming Language. Partial evaluation applied to ray tracing. Partial evaluation of numerical programs in Fortran. A partial evaluator A partial evaluation procedure for logic pro- grams Compiling scientific code using partial evaluation. Partial evaluation applied to numerical computation. Partial evaluation of Standard ML. Partial Evaluation and Mixed Computation. Automatic autoprojection of higher order recursive equations. Proving theorems about Lisp functions. A general criterion for avoiding infinite unfolding. Polyvariant mixed computation for analyzer programs. How do ad-hoc compiler constructs appear in universal mixed computation processes? In A transformation system for developing recursive programs. Compiling or-parallelism into and-parallelism New insights into partial evaluation: The Schism experiment. Binding time analysis for higher order untyped functional lan- guages The Schism Manual Dagstuhl Seminar on Partial Evaluation On compiling embedded languages in Lisp. On the partial computation principle. Mixed computation in the class of recursive program schemata. On the essence of compilation. Mixed computation: Potential applications and problems for study. On Futamura projections. Special Issue: Selected Papers from the Workshop on Partial Evaluation and Mixed Computation Partial evaluation of computation process - an approach to a compiler-compiler Generalized partial computation. Transforming logic programs by specialising interpreters. Tutorial on specialisation of logic programs. Some low-level source transformations for logic programs Specialisation of Prolog and FCP programs using abstract interpretation. Generating optimizing specializers. Application of metasystem transition to function inversion and transformation. Automatic construction of special purpose programs. Compiler generation by partial evaluation. A partial evaluator for the untyped lambda- calculus Efficient type inference for higher-order binding-time analysis Finiteness analysis. Handwriting cogen to avoid problems with static typing. Speeding up the back-propagation algorithm by partial evaluation Automatic program specialization: A re-examination from basic principles Partial Evaluation and Automatic Program Generation. An experiment in partial evaluation: The generation of a compiler generator. Generating a compiler for a lazy language by partial eval- uation Efficiently generating efficient generating extensions in Prolog. Introduction to Metamathematics. A Specification of an Abstract Prolog Machine and Its Application to Partial Evaluation. Partial evaluation as a means for inferencing data structures in an applicative language: A theory and implementation in the case of Prolog. A Prolog partial evaluation system. A strongly-typed self-applicable partial evaluator Partial evaluation in logic program- ming Lisp as the language for an incremental computer. ML partial evaluation using set-based analysis Ensuring global termination of partial deduction while allowing flexible polyvariance. A basis for a mathematical theory of computation. The application of partial evaluation to ray-tracing Converting interpreters into compilers. Comparative efficiency of general and residual parsers. Program specialization via program slicing. The realm of Nevryon. A compiler generator produced by a self-applicable specializer can have a surprisingly natural and understandable structure On the specialization of online program specializers. Meta interpreters for real. The Mixtus approach to automatic partial evaluation of full Prolog. An Automatic Partial Evaluator for Full Prolog. The structure of a self-applicable partial evaluator Automatic call unfolding in a partial evaluator. Special Issue on Partial Evaluation and Turchin's supercompiler revisited. An algorithm of generalization in positive supercompilation. Towards unifying partial eval- uation How to have your cake and eat it Partial evaluation of Prolog programs and its application to meta programming. Basic Refal and Its Implementation on Computers. A supercompiler system based on the language Refal. The concept of a supercompiler. Program transformation with metasystem transitions. A Prolog meta-interpreter for partial evaluation and its application to source to source transformation and query-optimisation A partial evaluation system for Prolog: Some practical considerations. Automatic online partial evaluation. --TR --CTR Mohan Rajagopalan , Saumya K. Debray , Matti A. Hiltunen , Richard D. Schlichting, Profile-directed optimization of event-based programs, ACM SIGPLAN Notices, v.37 n.5, May 2002 Arvind S. Krishna , Aniruddha Gokhale , Douglas C. Schmidt , Venkatesh Prasad Ranganath , John Hatcliff, Towards highly optimized real-time middleware for software product-line architectures, ACM SIGBED Review, v.3 n.1, p.13-16, January 2006 Arvind S. Krishna , Aniruddha S. Gokhale , Douglas C. Schmidt, Context-specific middleware specialization techniques for optimizing software product-line architectures, ACM SIGOPS Operating Systems Review, v.40 n.4, October 2006 Yasushi Shinjo , Calton Pu, Achieving Efficiency and Portability in Systems Software: A Case Study on POSIX-Compliant Multithreaded Programs, IEEE Transactions on Software Engineering, v.31 n.9, p.785-800, September 2005 Anne-Franoise Le Meur , Julia L. Lawall , Charles Consel, Towards bridging the gap between programming languages and partial evaluation, ACM SIGPLAN Notices, v.37 n.3, p.9-18, March 2002 Jacques Carette, Gaussian elimination: a case study in efficient genericity with MetaOCaml, Science of Computer Programming, v.62 n.1, p.3-24, September 2006 Anne-Franoise Le Meur , Julia L. Lawall , Charles Consel, Specialization Scenarios: A Pragmatic Approach to Declaring Program Specialization, Higher-Order and Symbolic Computation, v.17 n.1-2, p.47-92, March-June 2004
interpreters;software architectures;program specialization;genericity;adaptability;selective broadcast;partial evaluation;software layers;extensibility;pattern matching
592047
The Model-Composition Problem in User-Interface Generation.
Automated user-interface generation environments have been criticized for their failure to deliver rich and powerful interactive applications. To specify more powerful systems, designers require multiple specialized modeling notations. The model-composition problem is concerned with automatically synthesizing powerful, correct, and efficient user interfaces from multiple models specified in different notations. Solutions to the model-composition problem must balance the advantages of separating code generation into specialized code generators each able to take advantage of deep, model-specific knowledge against the correctness and efficiency obstacles that result from such separation. We present a correct and efficient solution that maximizes the advantage of separation by using run-time composition mechanisms.
Introduction Building user interfaces (UIs) is time consuming and costly. In systems with graphical UIs (GUIs), nearly 50% of source code lines and development time can be attributed to the UI [14]. GUIs are usually built from a fixed set of modules composed in regular ways. Hence, GUI construction is a natural target for automa- tion. Automated tools have been successful in supporting the presentation aspect of GUI functionality, but they provide only limited support for specifying behavior and the interaction of the UI with underlying application functionality. The model-based approach to interactive system development addresses this deficiency by decomposing UI design into the construction of separate models, each of which is declaratively specified [5]. Once specified, automated tools integrate the models and generate an efficient system from them. The model-composition problem is the need to efficiently implement and automatically integrate interactive software specified in separate, declarative models. This paper introduces the model-composition problem and presents a solution. A model is a declarative specification of some single coherent aspect of a user interface, such as its appearance or how it interfaces to and interacts with the underlying application functionality. By focusing attention on a single aspect of a user interface, a model can be expressed in a highly-specialized notation. This property makes systems developed using the model-based approach easier to build and maintain than systems produced using other approaches [23]. e s R e e UI Synchronization Toolkit Module Model Model Model Dialogue Application Presentation Presentation Dialogue Application Module Module Module Figure 1. Model-based code generation The Mastermind project [5, 15] is concerned with the automatic generation of user interfaces from three kinds of models: Presentation models represent the appearance of user interfaces in terms of their widgets and how the widgets behave; Application models represent which parts (functions and data) of applications are accessible from the user interface; and Dialog models represent end-user interac- tions, how they are ordered, and how they affect the presentation and the applica- tion. A dialog model acts as the glue between presentation and application models by expressing constraints on the sequencing of behavior in those models. Model-specific compilers generate modules of code from each model, and these resulting modules are composed into a complete user interface (Figure 1). A distinguishing characteristic of Mastermind is that the model-specific code generators work independently of one another. Composing code generated from multiple models is difficult. A model, by de- sign, represents a single aspect of a system and is neutral with respect to others [3]. Inevitably, however, functionality described in one model overlaps with or is dependent upon functionality described in another. A button, for example, is specified in a presentation model, but the behavior of the button influences behavior in other models, such as when pressing the button causes other widgets to be enabled or dis- abled. Such effects are described in a dialog model. The effect of pressing a button can also cause an application method to be invoked. Such effects are described in an application model. When code generated from multiple models must cooperate, these redundancies and dependencies can be difficult to resolve. Resolving them automatically means that behavior in different models must be correctly unified, and the mechanism for this unification must be implemented efficiently. The model-composition problem is concerned with automatically synthesizing powerful, correct, and efficient user-interfaces from separate presentation, dialog, and application models. We present a two-fold solution. First, we formalize the three models as concurrent agents, which synchronize on common events (Sec- tion 3). Second, we present a runtime architecture that supports the composition of modules generated from independent model compilers (Section 4). We present MODEL-COMPOSITION PROBLEM 3 the results of this approach on two examples and give evidence to show that it scales up (Section 5). 2. Background Model-based approaches to user-interface generation use models that are specified in diverse and often incompatible notations. This characteristic complicates model composition because the composition mechanisms in one model may not exist in another (Section 2.1). Prior research on the architecture of user-interfaces suggests using communicating agents to structure user-interface code (Section 2.2). Formal models of communicating agents provide a technique called conjunction, which is useful for composing partial specifications of a system (Section 2.5). The contribution of this paper is an extension of conjunction as a specification-composition operator into a runtime-composition mechanism. 2.1. Model-based generation The model-based approach to interactive system development expresses system analysis, design, and implementation in terms of an integrated collection of mod- els. Unlike conventional software engineering, in which designers compose software documentation whose meaning and relevance can diverge from that of the delivered code, in the model-based approach, designers build models of critical system attributes and then analyze, refine, and synthesize these models into running sys- tems. Model-based UI generation works on the premise that development and support environments may be built around declarative models of a system. Developers using this approach build interfaces by specifying models that describe the desired interface, rather than writing a program that exhibits the behavior [21]. One characteristic of model-based approaches is that, by restricting the focus of a model to a single aspect of a system, modeling notations can be specialized and highly declarative. The Mastermind Presentation Model [6], for example, combines concepts and terminology from graphic design with mechanisms for describing complex presentations using functional constraints. The Mastermind Dialog Model [19] uses state and event constructs to describe the user-computer conversation; the composition features include state hierarchy, concurrency, and communication. The Mastermind Application Model combines concepts and terminology from object-oriented design techniques [18] with mechanisms for composing complex behavior based on method invocation. Figure 2 compares the Mastermind models in terms of their domains of dis- course, communication mechanisms, runtime components, and how they are com- posed. Composition mechanisms in one model may not exist in another model. single one of these intra-model mechanisms is sufficient for composing all three Mastermind models. The model-composition problem can be restated as the need to unify behavior in multiple models without violating the rules of intra-model composition and while generating efficient code. The model-composition problem is a declarative instance of the problem of constructing a software system where the ma- Module Process Implementation Action Implementation Intra-module Composition Application Abstract Method Invocation Subclassing Aggregation Presentation Amulet Objects Constraints, Commands Instantiation, Aggregation Dialog State Machines Synchronous Message passing Orthogonal Composition Figure 2. Multi-paradigm action implementations jor components are expressed with programming languages from different families or paradigms. Zave has called this the multi-paradigm programming problem [24]. 2.2. Multi-agent user-interface architectures The Mastermind approach to model composition builds on prior work in multi-agent user-interface architectures, which provide design heuristics for structuring interactive systems. These architectures describe interactive systems as collections of communicating agents, which are independent computational units with identity and behavior. Two general frameworks-Model-View-Controller (MVC) [11] and agent roles and provide guidance on how agents should be connected. MVC prescribes how SmallTalk simulations can be composed by instantiating instances of three types of agents: models (not to be confused with the Masterimind models) describing application state, views providing presentations of models, and controllers allowing users to affect simulation behavior. A view registers interest in one or more attributes of a model. When an attribute changes, all registered views are notified so that they can recompute their display if necessary. The PAC framework more closely matches Mastermind than does MVC. In PAC, a presentation agent maintains the state of the display and accepts input from the user, an abstraction agent maintains a representation of the underlying application state, and a controller agent ensures that presentation and abstraction remain synchronized. The Mastermind Presentation, Application, and Dialog models are descriptions of the roles played by PAC's presentation, abstraction, and controller agents. Since Mastermind models describe PAC agents, we chose to make Mastermind models compose in the same manner that PAC agents compose. Specifically, the presentation and application models define actions, which are ordered by temporal constraints in the dialog model. To make these ideas more formal, we built upon prior work on formal definitions of agent composition. MODEL-COMPOSITION PROBLEM 5 2.3. Formal models of agents The PAC framework provides heuristic definitions of user-interface agent roles and connections. PAC agents are concurrent, and they compose by communicating control and data messages among themselves. To generate code from the models of these agents, we need to formalize the building blocks of agents and agent compo- sition. We chose the terminology and definitions that have been adopted by the various process algebras, specifically Lotos [4]. Process algebras formalize concurrency and communication, and they have proved particularly useful for describing UI software as a collection of agents [1, 2]. Other notations, such as StateCharts [8] and Petri nets [16], have also been explored for modeling UI agents, as these alternative notations also provide definitions of concurrency and communication. We chose Lotos because composition in Lotos resembles conjunction [25], which is a useful paradigm for composing partial specifications (Section 2.5). We model the behavior of an agent using a Lotos abstraction called a process, which is a computational entity whose internal structure can only be discovered by observing how it interacts with its environment. Processes perform internal (unob- servable) computations and interact with other, concurrently executing, processes. The interaction between processes is synchronous: If one process tries to communicate with a process that is not ready to communicate, the former process blocks until the latter is ready. Thus, the act of communicating synchronizes concurrent processes. A process represents the state of an agent as a procedure for performing future actions. An action is an atomic computational step taken by an individual pro- cess. Actions of a process can be observed through the events in which the actions participate. An event is an observable unit of multi-process communication. Multiple processes participate in an event by simultaneously performing actions over the same gate. A gate is a primitive synchronization device used to observe the occurrence of an action in a process. Each action is associated with a single gate. The gates of a composite agent are the union of the gates of its constituents. If two or more constituents name the same gate, then any actions over that gate proceed simultaneously. That is, the processes associated with the constituent agents synchronize actions that share the same gate name. Thus, gates also represent a class of possible inter-process synchronization events. During such an event, an action can offer one or more data values that can be observed by actions in other processes that are participating in the same event. A complete agent is modeled by a process that represents the initial state of the agent. A multi-agent system is modeled by a collection of concurrent, communicating processes. When composing a system of multiple agents, the designer must decide how to coordinate actions in the various processes that model the agents. are coordinated by synchronizing actions labeled with identically named gates. 6 STIREWALT AND RUGABER 2.4. Lotos Lotos is a rich language for specifying the partial ordering of actions within a process and the structure of multi-process interactions. Complex processes may be expressed by either combining sub-processes using an ordering operator (e.g., process P is the sequential composition of sub-processes P 1 and P 2 ) or by conjoining sub-processes so that they run independently but synchronize actions with gates. An event allows values to flow between participating actions. Lotos also describes the semantics of value passing with respect to synchronization. Actions in Lotos have the following structure: action ::= gate (inputjoutput) input output gate ::= identifier Each action names a gate and zero or more inputs and outputs. An input names a variable in which to record a value that is offered by an action in another process. An output is an expression for computing a value to offer to actions in other processes. Actions concisely represent the occurrence of many possible events. Like actions, events are associated with a particular gate. Unlike actions, events have no concept of input or output; rather they represent unique values that flow between actions. Events have the following structure: event ::= gate (value) value Note that the values are always constants because events are unique assignments of values during a synchronization. In Lotos, the gates over which two conjoined processes are required to synchronize must be specified between the vertical lines that symbolize the conjunction operator (k). For example, given the following Lotos process definitions: process process process R [ Process R behaves like P on gate g 3 and Q on gate g 4 , but R must behave like P and Q in synchrony on gates g 1 and g 2 . For processes with many gates, the Lotos notation quickly becomes unreadable. In this paper, we abbreviate the conjunction operator using notational conventions similar to those used in CSP [9]. In our abbreviated notation, we write the conjunction of P and Q as P k Q with the understanding that P and Q must synchronize on gates that are common to the agents whose states P and Q respectively represent. Suppose the behavior of an agent can be described by a Lotos process B. If the agent can perform an action by synchronizing on event e (denoted B(e)), then its MODEL-COMPOSITION PROBLEM 7 behavior from that point on is defined by another process B(e). The systems under study are deterministic, which means that B(e) is always unique. Moreover, when a system is defined by conjoining sub-processes, the compositional structure is preserved throughout the lifetime of the system. That is, if then where: occurs over a gate of agent i Any event that can be observed of a process P can also be observed of any conjunction of P with other processes. This fact will be important when we define the observer function (Section 3.4). 2.5. Conjunction as composition Alexander uses conjunction to compose separately defined application and presentation agents [2]. Abowd uses agent-based separation to illuminate usability properties of interactive systems [1]. Both of these approaches rely on the use of conjunction to compose agents that are defined separately but interact. In fact, conjunction is a general operator for composing partial specifications of a system [25]. The idea is that each partial specification imposes constraints upon variables (or, in the case of agents, events) that are mentioned in other partial specifications. When these specifications are conjoined, the common variables must satisfy all constraints. We define the behavior of a system generated from Mastermind models to be any behavior that is consistent with the conjunction of constraints imposed by the dialog, presentation, and application models. We then extend conjunction from a specification tool into a mechanism for composing runtime modules. 2.6. Summary Three issues must be addressed to solve the model-composition problem: The solution must generate user-interfaces with rich dynamic behavior, the correctness of module composition must be demonstrated, and the generated modules must co-operate efficiently. In Mastermind, the rich expressive power is achieved through special-purpose modeling notations [15, 5]. The remainder of this paper addresses the generation of correct implementations with maximal efficiency while preserving the expressive power of Mastermind models. 3. Model-composition theory Recall from Figure 1 that each class of model has a code generator that synthesizes runtime modules for models in that class. The modules are generated without detailed knowledge of the other models. At run time, however, modules must cooperate as prescribed by the conjunction of the models that generated them. This section describes the relationship between model composition and the mechanism by which the associated modules cooperate at runtime. 3.1. Notation The subject of this paper is the automatic generation and composition of runtime modules from design-time models. A module is a unit of code generated from a single model. We use a third class of construct-the Lotos process-to define the correctness of model and module composition. In formal arguments, we need to refer to all three types of constructs; thus we distinguish the constructs by using different fonts. We also need special functions that map models and modules into comparable domains. We represent the classes of Mastermind models using German letters. The symbols D, and A represent respectively the classes of Mastermind presentation, dialog, and application models. We use the italic font to represent Lotos processes and the semantic models of these processes. The set P rocess represents the set of Lotos processes. Specific processes are written in capital italic letters (e.g., P , D, and A, respectively). The set T raceSets defines the set of event traces over the alphabet of gates and the space of values that can be offered and observed by Lotos actions. The function raceSets maps a Lotos process to the set of all event traces that can be observed of that process. We represent runtime entities using the Sans serif font. The set Component represents the class of all runtime components. A component is a block of code that provides gates for observing the actions of the component. By defining components as runtime code that provides gates for observing behavior, we can define the function raceSets that maps a component to the set of event traces that can be observed through the gates that the component provides. There are two categories of component in the Mastermind architecture: the generated modules and the synchronous composition of these modules. Instances of the generated modules are written Pres, Dialog, and Appl, respectively. We also think of the modules in synchronous composition as a component, which is attained by connecting the generated modules using some synchronization infra-structure (de- fined in Section 4). This composite component is written Synch[Pres; Dialog; Appl]. The name Synch suggests that the component is the synchronization of the three generated modules; the brackets suggest that the generated modules fit into the larger system and that Synch by itself is not a component. 3.2. Inter-model composition Model-based code generators construct runtime modules from design-time mod- els. The code generation strategy is model-specific, reflecting the specialization of models to a particular aspect of a system. At run time, however, modules must co- operate, and the cooperative behavior must not violate any correctness constraints imposed by the models. There is an inherent distinction between behavior that MODEL-COMPOSITION PROBLEM 9 is limited to the confines of a given model and behavior that affects or is affected by other models. Inter-model composition is concerned with managing this latter inter-model behavior. Some behavior is highly model specific and neither influences nor is affected by behavior specified in other models. As Figure 2 illustrates, in a Mastermind presentation model, graphical objects are implemented using primitives from the Amulet toolkit [13], and attribute relations are implemented as declarative formulas that, at runtime, eagerly propagate attribute changes to dependent attributes. As long as changes in these attributes do not trigger behavior in dialog or application models, these aspects can be ignored when considering model composition. In an application model, object specifications are compiled into abstract classes under the assumption that the designer will later extend these into subclasses and provide implementations for the abstract methods. As long as the details of these extensions do not trigger behavior in dialog or presentation models, this application behavior may also be ignored when defining model composition. Within a module, entities compose according to a model-specific policy. In a presentation model, for example, objects compose by part-whole aggregation, and attributes compose by formula evaluation over dependent attributes. In an application model, objects compose using a combination of subclassing, aggregation, and polymorphism. When considering how models compose, some details of intra-model composition can be abstracted away, but not all of them. Models impose temporal sequencing constraints on the occurrence of inter-model actions, and models contribute to the values computed by the entire system. These constraints and contributions must be captured in some form and used to reason about model composition. We map this inter-model behavior into a semantic domain that is common across all of the models. This domain is described by the Lotos notation, which specifies temporal constraints on actions and data values. We assume that Lotos processes can be derived from the text of a model specification (Section 3.4). Designers may, for example, need to designate actions of interest to other models. Lotos processes do not capture all of the behavior of models in composition, but they do express the essential inter-model constraining behavior. 3.3. Example We now present an example of inter-model behavior expressed as a Lotos process. The dialog model being considered is for a Print/Save widget similar to those found in the user interfaces of drawing tools, web browsers, and word processors (See Figure 3). These widgets allow the user to format a document for printing either to a physical printer or to a file on disk; we call the former task printing and the latter task saving. Options specific to printing, such as print orientation (e.g., portrait vs. landscape), and to saving, such as the name of the file into which to save, are typically enabled and disabled depending upon the user's choice of task. These ordering dependencies are reflected in the dialog model for this widget shown by the Lotos process in Figure 4. Figure 3. Screen shot of the Print/Save dialog box The process P rintSave can synchronize on any of the gates that follow it in square brackets. In this example, the gates print, save, go, cancel, layout, and kbd (line 1 in the figure) define points for synchronizing with the presentation; whereas the gates lpr and write define points for synchronizing with the underlying application. The process parameters lpdhost and f ilename (line 2) store the name of the default printer and the user-selected filename, respectively. The parameter doc represents the document to be printed or saved, and the parameter port represents the print orientation (portrait if true, landscape if false). The widget in Figure 3 is specified by a separate presentation model (not shown). This model defines a pair of radio buttons labeled File and Printer and two buttons labeled OK, and CANCEL. When these buttons are pressed, they offer the events save, print, go, and cancel respectively. The presentation model also contains a pair of radio buttons that specify paper orientation. These buttons display graphics of a page in either portrait or landscape mode and, when selected, offer the event port with a value of true if the choice is for portrait orientation and false for landscape orientation. Finally, there is a text entry box in which the user can type in a file name. As the user edits this name, the text box responds by offering the contents of the string typed so far as part of the kbd event. Note that the actual being pressed are not returned, as editing functionality is best handled in a text widget and is not considered inter-model behavior. A separate application model (not shown) defines procedures for issuing a print request and saving a file to disk. These procedures are responsive to the events lpr and write respectively. Actions that synchronize on these events offer a number of values including printer name (lpdhost) and filename (f ilename). MODEL-COMPOSITION PROBLEM 11 1. process PrintSave[ print, save, go, cancel, layout, kbd, lpr, write 2. 3. P[ go, lpr, write, layout, kbd 4. where 5. process P[ go, lpr, write, layout, kbd 7. endproc 8. process F[ go, lpr, write, layout, kbd 9. Edit[ go, write, kbd 10. endproc 11. process Layout[ go, lpr, layout 12. 13. [] ( go; lpr ! lpdhost 14. endproc 15. process Edit[ go, write, kbd 17. 19. endproc Figure 4. Print/Save dialog process. The temporal structure of dialog, presentation, and application model composition is given in the behavior specification (line 3). The behavior of P rintSave is the behavior of the process P (defined on lines 5 through 7) with the caveat that it may be disabled (terminated) at any time by the observation of the cancel event. Disabling is shown with the [? operator. Process P represents which interactions and application invocations must happen in order to send a document to a printer. Most of this functionality is actually expressed in the sub-process Layout (defined on lines 11 through 14). P behaves like Layout in the normal case, but it can be disabled if the save event is observed. Recall that the save event is offered whenever the user presses the Save to File button in the presentation model. The process F (defined on lines 8 through 10) likewise behaves like the process Edit (defined on lines 15 through 18) in the normal case, but is disabled if the event print is observed. Note that F and P can disable each other, which means that the user can switch back and forth between printing and saving as many times as he or she likes before hitting the Go button. \Gamma\Gamma\Gamma\Gamma! P rocess CD y Component Obs \Gamma\Gamma\Gamma\Gamma! T raceSets Figure 5. Dialog compiler correctness. 3.4. Models, modules, and processes like those shown in Figure 4 are useful for understanding the relationship between models and modules. This relationship is complex, and so we describe it first for a single model and then for the three models in composition. We now formalize correctness conditions for the Mastermind dialog model. A similar formalization exists for the other Mastermind models. Figure 5 shows the relationship between dialog models (members of the set D), runtime modules generated by dialog models (members of the set Dialog), and the inter-model behavior of dialog models (members of the set P rocess). The relationships between these sets are defined as functions that map members of one set into members of another. The function CD : D ! Dialog maps dialog models to runtime modules. Think of CD as an abstract description of the dialog-model compiler. The rocess maps dialog models into Lotos processes describing their inter-model behavior. Think of AD as an abstract interpretation of the dialog model expressing its semantics in Lotos. These sets and functions are related by the commutative diagram of Figure 5. Externally observable model behavior is mapped into a Lotos process by AD , and the set of traces of a module's externally observable events is recorded by Obs. We say that a dialog model d 2 D is consistent with the module CD (d) if every trace is in the set T r(AD (d)) and if there are no sequences such that & 62 Obs(CD (d)). That is, the inter-model behavioral interpretation of d agrees exactly with the observable behavior of the runtime module generated from d. Commutativity of the diagram requires this property for any dialog model in the set D. 3.5. Model-based synthesis The correctness relationship between models and modules (Figure 5) can be extended to specify the correctness of module composition. We now have functions AP , AD , and AA that map models into Lotos processes. These processes should compose by conjunction. We also have a runtime component Synch that combines modules Pres, Dialog, and Appl into a single component whose actions are observable by the Obs function. Figure 6 shows the constraints on the behavior of these entities. Let p 2 P, d 2 D, and a 2 A. Then the code generated from these MODEL-COMPOSITION PROBLEM 13 Figure 6. Module-composition correctness. models is correct if and only if, for any observable behavior oe, oe is a legal trace in the conjunction of the models. This equation defines the conditions necessary for correct module composition without assuming any model-specific interpretation of these actions. It serves, therefore, as a specification of design requirements. In the next section, we present an implementation that satisfies these requirements. 4. Module-composition runtime architecture We now turn to the designs of the run-time synchronization module and model-specific compilers of Figure 1. The essential design problem is how to make the generated modules compose while retaining the independence of the model-specific compilers. The conditions of Figure 6 impose constraints on these designs. For- tunately, these constraints do not require model-specific knowledge (e.g., graphical concepts in the presentation model or data layout in the application model). Thus, module-composition logic can be separated from the model-specific functionality within a module. This separation is the key to making model-based synthesis independent without sacrificing the correctness of module integration. The Mastermind runtime library contains efficient primitive classes that enable independent module synthesis and correct composition by conjunction. This library provides a great deal of generality and flexibility for code generation. In this paper, we describe only those aspects of the library that are relevant for supporting independent synthesis. First, we introduce the mechanism for composing generated modules (Section 4.1). We then describe how this mechanism implements conjunction without sacrificing the independence of model synthesis (Section 4.2) and demonstrate its operation through an example (Section 4.3). 4.1. Design structures to support conjunction To facilitate the independence of model synthesis, we designed a mechanism that enables a module to compose with other modules without directly referencing them. As Figure 1 suggests, generated modules compose through the aid of a special synchronization component, called Synch. We designed the Synch interface to simplify the generation of modules. This section describes the interface and the process of model compilation and integration. Figure 7 illustrates the interface between the generated modules and the Synch component. Modules contain Action objects that link (explicitly refer to) Gate 14 STIREWALT AND RUGABER Action Action Action Action Action Action Gate Gate Gate Synch Dialog Pres Appl Figure 7. Structural depiction of composition according to Synch[Pres, Dialog, Appl]. objects in the Synch component. As the names suggest, an Action object reifies a Lotos action, and a Gate object reifies a Lotos gate. At runtime, Actions implement a unit of observable behavior in a module, and Gates implement the synchronization of Actions by conjunction. The mathematical connection between Lotos actions and gates is reified using explicit links between Action and Gate objects. These links constitute the mechanism for composing generated modules with the Synch component: A module "plugs in" to the architecture by linking its Action objects to appropriate Gate objects in the Synch component. The dashed lines in Figure 7 illustrate some (of many possible) links. This architecture enables model synthesis to be treated separately from module integration, similar to the way compilation is treated separately from linking in traditional programming. This separation allows a module to be synthesized from a single model, independent of the synthesis of the other models. During synthesis, model-based compilers independently generate modules. Any behavior that must be observed by other modules must be packaged into an instance of the class Action. When emitting the code that creates this instance, the compiler also writes out the name of the associated gate to an auxiliary file. Consequently the output of a model compiler is a module and an auxiliary file listing the names of dependent gates. During module integration, a module integrator reads in these auxiliary files, creates the Synch component, and combines it with the generated modules to produce an executable image. Going back to our running example, consider the compilation of the presentation model for the Print/Save dialog box (Figure 3). As the model is processed, the compiler emits Action objects that interface directly with UI toolkit widgets. After compilation, the Pres module will contain an Action for each widget in the dialog box. For example, there will be a distinct Action object paired with the OK and CANCEL buttons, each of the radio buttons, and Filename text-entry widget. To integrate the Pres module with the other modules, each of these Actions must link to Gate objects in the Synch component. Note that when the Actions are being emitted, the corresponding Gate object will not yet exist, as the Gate is created by the module integrator. Thus, the link between an Action and its corresponding Gate cannot be established at compile time. Instead, an Action object is instantiated with the name of the gate over MODEL-COMPOSITION PROBLEM 15 void enable(); void disable(); Action {abstract} ModuleSource void register(Appl*); void unregister(Appl*); DialAppGate void register(Pres*); void unregister(Pres*); PresDialGate Gate {abstract} void confirm(Listener*); void synchronize(); PresDialAppGate void unregister(Dial*); void register(Dial*); DialGate void execute(); Command {abstract} Listener {abstract} void listen(); void ignore(); generalization (disjoint subclasses) generalization (overlapping subclasses) Legend synchronizes ActionRole {abstract} Figure 8. Detailed design of action and gate classes. which it must synchronize. At runtime, the Action uses this name to locate the corresponding Gate. Because the module integrator creates a Gate for each named gate, the Action object can assume that the gate will exist at runtime. This design greatly simplifies model compilation: The presentation-model compiler need not concern itself with locating an object in another component. Rather, the compiler simply creates a module using Action objects and writes out the names of gates to an auxiliary file. 4.2. Behavior of the design structures The synthesis of one Mastermind model can proceed independently of the synthesis of other models because the generated modules only refer to each other indirectly, through Gate objects. The Gate objects are responsible for determining when a synchronization should occur and dispatching control the associated Action objects in an appropriate order once the synchronization constraints are satisfied. Conse- quently, Action objects need not be concerned with these issues. Rather, Actions are concerned with implementing model-specific functionality. This separation is crucial to supporting the independence of model synthesis. Figure 8 describes the design of classes Action and Gate. Class Gate is designed to internalize information about the modules whose actions are required to synchronize at the gate. Henceforth, we shall refer to this information as the synchronization constraint of a Gate. The rules of conjunction (Figure number of possible variations of this constraint. At runtime, a Gate determine whether or not to synchronize by checking whether or not this constraint is satisfied. To make this determination, a Gate must infer the location (module) of each Action that wishes to synchronize over the Gate. We call this process of inference tabulation. Tabulation occurs when an Action announces its readiness to synchronize. Such announcements are made by an Action registering itself with its Gate; an Action registers itself by passing itself to an invocation of the register operation on its Gate. When a Gate determines that its constraint is satisfied, it invokes the synchronize operation, which dispatches control to the registered Actions so that they may execute. For a Gate to tabulate the modules that request activity, the Gate must be able to infer the module of every Action that registers. This means that an Action must know the module in which it exists. Class Action has a subclass, called ModuleSource, which further specializes into three subclasses, Pres, Dial,and Appl (not shown in the figure). The concrete class of every Action must inherit from one of these three subclasses. We implemented tabulation by specializing the register operation so that it dispatches based on these subclasses. The subclasses of Gate contain module variations of the register function. These subclasses embody each of the three possible synchronization constraints that arise in Mastermind. The constraint associated with class PresDialGate requires Pres and Dial actions to be present at the Gate. Similarly, the constraint associated with class DialApplGate requires Dial and Appl actions to be present at the Gate, and the constraint associated with class PresDialogApplGate requires all actions from all three modules to be present at the Gate. These are the only three types of synchronization constraints required of Mastermind-generated user interfaces. The next issue concerns dispatching control to registered actions once a Gate's synchronization constraint is satisfied. Mastermind supports two different action- control mechanisms (generalized by ActionRole). One mechanism is a generic interface for executing a model-specific operation (class Command). The other mechanism is a generic interface for reactively observing an asynchronous event, such as a user interaction with a graphical widget (class Listener). What happens when a Gate's synchronization constraint is satisfied depends upon the control mechanisms used by the registered Actions. For example, if two Commands are waiting at a Gate, and they satisfy the synchronization constraint for the Gate, then the execute method for both Commands are invoked. If, instead, one of these actions is a Listener and the other is a Command, then the Command is not invoked until the Listener receives an event. Because Listeners are reactive, they need to be able to announce the reception of an event to the Gate. This is accomplished by invoking the confirm operation on the Gate. A module requests the performance of an Action by invoking the operation enable. Enabling causes an Action to register itself with its Gate. Our design abstracts the logic for requesting the performance of an Action into the enable and disable methods, which correctly cooperate with the corresponding Gate irrespective of the particular synchronization constraints. Thus, the logic can be completely encapsulated in the abstract class Action, which a model-compiler writer need never modify. Moreover, model-compiler writers can package model-specific functionality using one of two quite different control policies, Command and Listener. One consequence of this design is that the module integrator must determine the type of Gate to emit. This is a simple task, however, given the information written to the MODEL-COMPOSITION PROBLEM 17 generalization Legend pointer to operation pseudocode void listen(); void ignore(); Listener {abstract} widget Am_Text_Input_Widget void Do(); TextFieldAction void listen(); void ignore(); Pres Figure 9. Example of use. auxiliary files by the model compilers. For example, the gate cancel that is used in the Print/Save dialog is used in both the presentation model, where it observes the pressing of the CANCEL button, and in the dialog model, where it observes the completion of the dialog. Because modules compose by conjunction, the Gate associated with cancel always synchronizes an action from the Pres module with an action from the Dialog module. To implement this behavior, the module integrator emits an instance of PresDialogGate, which is returned when the associated Actions link to the named gate. 4.3. Example We now demonstrate how these features work in the context of the Print/Save dialog. Recall from Figure 3 the text entry field that allows a user to enter file name in which to save a document. In the dialog model (Figure 4), the entry of the file name is modeled as an atomic action over the gate kbd. To connect this dialog action to the text entry widget that ultimately witnesses the action, we need a presentation Action that knows how to attach to the text entry widget, and we need a Gate object to represent the kbd gate. Figure 9 illustrates how a reusable action that listens for text entry can be created from the primitives introduced in Section 4.2. The presentation-model compiler emits instances of this class to implement text-entry boxes. In the fig- ure, we rendered the primitive classes in grey to distinguish them from new objects and classes that the model-compiler writer creates. The new class is called TextFieldAction. It inherits from class Pres because its instances will be emitted into the Pres module. It inherits from class Listener because it is concerned with monitoring and controlling the text-input widget. The class is associated with an Am Text Input Widget object by an association called widget. This object is prede- fined in the Amulet toolkit [13], which the current version of Mastermind uses for presentation support. The TextFieldAction controls the Amulet object by invoking the Start and Stop operations on the object, which instruct the widget to enable and disable keyboard input. The invocations of these methods form the implementation of listen and ignore respectively. We also need a way for the widget to signal the Action object with the event. This is accomplished by overriding the Do method of the widget to go find the Gate associated with the Action and invoke the confirm operation on this Gate to signify the occurrence of the event. The Do method can be thought of as a callback function that Amulet invokes to deliver an event (in this case, the event is a keyboard return). The example serves to illustrate the sequence of behaviors that are enacted by the Mastermind library primitives. Suppose an object of class TextFieldAction is registered at the Gate associated with kbd. If the synchronization constraint for this Gate is satisfied, the Gate invokes the listen method of the TextFieldAction. This invocation in turn causes the Start method of the Am Text Input Widget to be invoked, which enables user input at the widget. If the synchronization context changes so that the constraint is no longer satisfied-either because the Pres module disables the TextFieldAction or because another module disables an Action that is waiting at the Gate, then the Gate invokes the ignore operation. This causes the TextFieldAction to invoke the Stop method of the Am Text Input Widget, which disables text input. If, on the other hand, the user enters a string and hits the return key, the Do method of the widget is invoked. This causes the invocation of the confirm method on the Gate, and the Gate proceeds to execute any Commands that are waiting. 4.4. Summary Our design enables independent code generation because the Actions in a generated module are insulated from Actions in other modules by the gate objects. We compose modules by creating Gate objects that embody the synchronization requirements of the models and by linking Actions to their Gates. The independence that is afforded by this approach allows model-based code generators to apply deep model-specific knowledge to the synthesis of code. 5. Results and status We evaluated our solution to the model-composition problem with respect to power, correctness, and efficiency. Multi-paradigm actions have proved easy to specialize to accommodate features from disparate implementation toolkits and architectures. For example, we have specialized Actions to represent actions in: the Amulet object system [13], the C++ object system, and a special-purpose state-machine language. Figure 2 summarizes the different applications and results. MODEL-COMPOSITION PROBLEM 19 5.1. Power We were able to express user interfaces using our modeling notations in several case studies. We tested the quality of user interfaces on two specific examples: the Print/Save widget described in Section 3.3 and an airspace-and-runway executive that supports an air-traffic controller (ATC) [19]. The former demonstrates the ability to generate common, highly reusable code for standard graphical user inter- faces. The latter demonstrates the ability to support a complex application using a direct-manipulation interface. The ATC example testifies to the power of our approach. When a flight number is keyed into a text-entry box, an airplane graphic, annotated with the flight number, appears in the airspace. As more planes come into the airspace, the controller keys their flight number into the text-entry box. When the controller decides to change the position of a plane, she does so by dragging the airplane icon to a new location on the screen. As soon as she presses and holds the mouse button, a feedback object shaped like an airplane appears and follows the mouse to the new location. When the mouse is released, the plane icon moves to the newly selected location. The presentation model of the ATC example is quite rich. It specifies gridding so that airplane graphics are always uniformly placed within the lanes, and it specifies feedback objects that present controllers information about the planes during operation. In a real deployment, the locations of flights change in response to asynchronous application signals from special hardware monitors. In such a deployment, these signals would be connected to Listener actions and would fit into the frame-work without change. For more details on this case study and the Print/Save dialog, see Stirewalt's dissertation [19]. 5.2. Correctness In addition to being able to generate and manage powerful user interfaces, the composition of our modules is correct. Two aspects of our approach require justification on these grounds. First is the design of runtime action synchronization. Second is the synthesis of runtime dialog components (members of the set D) from dialog models. This paper addresses the theoretical issues involved in the design of runtime action synchronization. The Gate and Action classes are traceable refinements of the corresponding concepts in Lotos. In practice, we found this design to be quite robust. One aspect of synchronization correctness, which we do not address in this paper, is how to show that a model-specific specialization of Action does not violate the delicate callback protocol that underlies the system. For example, say that an Appl, which when modeled in Lotos observes a value x and offers a value y, is to be implemented using a method invocation. The method should bind x to its parameter and bind its return value to y. Since value offerings are evaluated in sequence, how can we be sure that the ordering of evaluation does not interfere with the invocation of the method or vice versa? Currently, we check this by inspection, but we are investigating ways of packaging this problem so that a model checker (e.g., SPIN [10], SMV [12]) can detect such anomalies. Stirewalt used the SMV model checker to validate the inter-operation of Action and Gate objects [19]. As we mentioned earlier, the Mastermind Dialog model notation is a syntactic sugaring for a subset of Lotos. This language is described in greater detail in [20]. We implemented a prototype dialog model code generator whose correctness was validated as described in Stirewalt [19]. This code generator compiles dialog models without reference to other models. 5.3. Efficiency We measured efficiency empirically by applying our code generator on the ATC example. We generated dialog modules and connected these with hand-coded presentation and application modules. On the examples we tried, we observed no time delays between interactions. We quantified these results by instrumenting the source code to measure the use of computation resources and wall-clock time. The maximum time taken during any interaction was 0:04 seconds. This compares well to the de facto HCI benchmark of response time, 0:1 seconds. We believe that more heavyweight, middle-ware solutions, such as implementing synchronization through object-request brokers, are not competitive with these results. 5.4. Future work We are currently completing a more robust dialog code generator. This new code generator incorporates state-space reduction technology and will improve interaction time, which in the prototype is a function of the depth of a dialog expression, with interaction that executes in constant time. 6. Conclusions How to generate code for a specialized modeling notation is a well understood problem. Integrating code generated from multiple models is not. Integration is much more complicated than merely linking compiled object modules. For models to be declarative, they must assume that entities named in one model have behavior that is elaborated in another model. Designers want to treat presentation, temporal ordering, and effect separately because each aspect in isolation can be expressed in a highly specialized language that would be less clear if it were required to express the other aspects as well. For interactive systems, composition by conjunction is essential to separating complex specifications into manageable pieces. Unfortunately, programming languages like C++ and Java do not provide a conjunction operator. Such an operator is difficult to implement correctly and effi- ciently, and, in fact, we did not try to implement it. Rather, by casting model composition into a formal framework that includes conjunction, we are able to express a correct solution and then refine the correct solution into an efficient de- MODEL-COMPOSITION PROBLEM 21 sign. This is a key difference between our approach and middle-ware solutions that implement object composition by general event registry and callback. Our results contribute to the body of automated software engineering research in two ways. First, our framework is a practical solution that helps to automate the engineering of interactive systems. Second, our use of formal methods to identify design constraints and the subsequent refinement of these constraints into an object-oriented design may serve as a model for other researchers trying to deal with model composition in the context of code generation. The formality of the approach allows us to minimize design constraints and is the key to arriving at a powerful, correct, and efficient solution. --R Formal Aspects of Human-Computer Interaction Structuring dialogues using CSP. Developing Software for the User Interface. Introduction to the ISO specification language Lotos. Using declarative descriptions to model user interfaces with Mastermind. Declarative models of presentation. PAC, an object-oriented model for dialog design On visual formalisms. Communicating Sequential Processes. The model checker spin. A cookbook for using the model view controller user interface paradigm in smalltalk. Symbolic Model Checking: An Approach to the State Explosion Problem. The Amulet environment: New models for effective user-interface software development Survey on user interface programming. Knowledgeable development environments using shared design models. Validating interactive system design through the verification of formal task and system models. The Mecano project: Comprehensive and integrated support for model-based user interface development Automatic Generation of Interactive Systems from Declarative Models. Design and implementation of mdl: The mastermind dialogue language. Declarative models for user-interface construction tools: the Mastermind approach Beyond interface builders: Model-based interface tools Beyond hacking: A model based approach to user interface design. A compositional approach to multiparadigm programming. Conjunction as composition. --TR
model-based;user interface;multi-paradigm;code generation
592049
Explanation-Based Scenario Generation for Reactive System Models.
Reactive systems control many useful and complex real-world devices. Tool-supported specification modeling helps software engineers design such systems correctly. One such tool, a scenario generator, constructs an input event sequence for the spec model that reaches a state satisfying given criteria. It can uncover counterexamples to desired safety properties, explain feature interactions in concrete terms to requirements analysts, and even provide online help to end users learning how to use a system. However, while exhaustive search algorithms such as model checkers work in limited cases, the problem is highly intractable for the functionally rich models that correspond naturally to complex systems engineers wish to design. This paper describes a novel heuristic approach to the problem that is applicable to a large class of infinite state reactive systems. The key idea is to piece together scenarios that achieve subgoals into a single scenario achieving the conjunction of the subgoals. The scenarios are mined from a library captured independently during requirements acquisition. Explanation-based generalization then abstracts them so they may be coinstantiated and interleaved. The approach is implemented, and I present the results of applying the tool to 63 scenario generation problems arising from a case study of telephony feature validation.
Introduction Reactive systems control many useful and complex real-world devices, such as telephone switches, air and space craft, and software agents. Such feature-rich systems are difficult to design correctly, particularly when distinct functional features are designed by different people at different times over the lifecycle of a product family. Specification modeling[11, 16] allows engineers to apply relatively sophisticated validation tools such as simulation, coverage analysis, model checking[17, 5], or theorem proving[12, 20], to relatively abstract models of the system's behavior in order to find design errors before implementation. It is the abstractness of the models that makes many of the reasoning techniques tractable. The validated spec model can be used as a starting point for code genera- tion, as documentation of the behavior of the system, and in support of maintenance and evolution[11]. A spec modeling tool suite benefits significantly from a scenario generator, which constructs an input event sequence for the spec model that reaches a state satisfying given criteria. Such a tool can uncover counterexamples to desired safety properties, explain feature interactions in concrete terms to requirements analysts, increase test coverage, and even function as documentation, showing end users how to achieve their goals while still learning how to use a system. However, while some model checkers[17, 4] are capable of generating scenarios for certain limited classes of reactive systems, such as finite state machines with small (or highly symmetric) state spaces, the problem is intractable for functionally rich models that arise as natural abstractions of systems engineers wish to design. For example, in addition to requiring search in an infinite state space, models incorporating arithmetic operators can require the scenario generator to find satisfying instances of arbitrary arithmetic constraints, which is undecidable. This paper describes a novel heuristic approach, called ization") which is applicable to a large class of infinite state reactive systems. The key idea is to instantiate and piece together abstracted scenarios that achieve subsets of the conjuncts of a goal predicate into a single scenario achieving the conjunction of the sub- sets. The scenarios are mined from a library of concrete scenarios captured independently during requirements acquisition. Critically, they are then abstracted via explanation-based generalization. The approach is sound, but incomplete, so it will not succeed in finding scenarios in all cases of satisfiable goal predicates; however, it is intended to be fast, even in failing cases, so that it can be a practical interactive tool. Moreover, the approach's power can be increased by adding more scenarios to the library, so, as more requirements are uncovered and specified, the power of the tool grows naturally. Even an incomplete generator is quite use- ful. Typically, an engineer will discover (e.g. via static analysis or proof attempts) descriptions of states in which spec inconsistencies may arise, or correctness properties may be violated; the scenario generator is then run on these descriptions. Whenever the generator is successful, a definite design flaw has been found, so the engineer can focus attention there first. The other cases, which may not even be satisfiable, can be put off to later in the design process, after the known problems are fixed. Fixing these first problems may either alter or eliminate the other ones anyway. When the generator fails, putting out a scenario coming as close to the goal as possible can be helpful as well. This paper can be summed up in three key ideas: ffl Current limited-domain exhaustive search approaches (such as model checking[17]) to scenario generation are not enough; we need a usable scenario generator that accommodates more expressive logics and large state spaces, even though the problem is highly intractable; ffl The heuristic SGEN 2 approach, based on mining and abstracting requirements knowledge using explanation-based generalization, applies to richly expressive logics and large state spaces; ffl A moderate sized case study involving feature interactions in telephony gives initial empirical evidence that SGEN 2 is practical and useful. After Section 2 defines terms and describes the tool suite in which SGEN 2 is implemented, the next three sections make these key points. I conclude with discussion of related work, limitations, and future work. Modeling This work is performed within the Interactive Specification Acquisition Tools (ISAT) framework. ISAT[11, 12, 13] is a prototype tool suite for reactive system design that is intended to support full-lifecycle spec modeling as well as code generation. A reactive system is a (not necessarily finite-) state machine that reacts to parameterized input events by changing its state and by performing acts, which can be thought as output events. ISAT is based on two hypotheses: ffl Functional requirements are most reliably elicited from and validated by requirers as concrete, formal behavior scenarios; and ffl Specifications must be executable and amenable to automatic analysis. A designer constructs a reactive system model in the executable spec language, while a requirer specifies functional requirements as concrete scenarios. The latter are interleaved sequences of input events and act or state observations required to be true. Thus, crucially to SGEN 2 , a natural part of the design lifecycle is the acquisition of a library of validated concrete scenarios describing the system's behavior. 2.1 Model Formalism and Backpropagation An ISAT spec model consists of a theory definition together with a set of event handlers. The theory defines the types, functions, and semantic axioms of a pure computational logic, as well as the signatures of the state relations, events, and acts that make up the system. In order to support model simulation (execution), all primitive function declarations in the model's theory must include a total operational function capable of computing the value of the function on inputs in its domain and some non- error, type-compatible output value on inputs outside its declared domain. (ISAT model theories are somewhat similar to computational logic as described in [3].) Thus, ISAT supports arbitrary functional rich- ness, bounded only by the user's willingness and ability to code implementations for the functions and provide the logical axioms supporting the other reasoning tools (see below). For example, models can operate on arbitrary data structures. I have used this richness to good advantage in my work on applying ISAT to the specification and implementation of the Email Channels system[13]: the ISAT model operates on message data structures, lists of users and messages, and even database relation objects. Event handlers are expressed in a limited procedural language P-EBF ("procedural event-based formal- ism"), which is semantically related to the rule-based EBF I described in [11]. The details of P-EBF are not crucial to this paper, except that it supports a predicate backpropagation operator BackProp. Note that P-EBF need not be the input language seen by the designer; many domain-appropriate front-end formalisms (e.g. domain-specific languages or graphical programming environments) may be compiled into P- EBF. Such formalisms are beyond the scope of the present paper, however. Formally, the state of an ISAT model is represented as a collection of parameterized (partial) functional relations and each D i are data domains (types). For example, the relation Address 7! Call stores for an address (i.e., phone number) the object representing its ongoing call (if any). State values are referred to within P- EBF expressions via the LOOKUP operator; for example, (LOOKUP CALL "1234") returns the current call in which extension "1234" is involved, if any. A state predicate is a Boolean-typed ISAT expression. Predicates may be parameterized by typed formal parameters. Here is a state predicate of one address parameter, usr: (and (member? usr (lookup known-addresses)) (equal IDLE (lookup mode usr)) (not (equal NO-CALL (lookup call usr)))) This predicate represents all states in which there is an idle address that nevertheless still has a valid call object. It is the negation of a desirable state invariant; thus, a generated scenario reaching such a state proves the existence of a design error. The BackProp Operator. Formally, ISAT's BackProp takes six arguments, and returns three values a). P 0 is a state predicate and a 0 is a list of actual (concrete) parameters for P 0 such that P 0 is true when evaluated in model M 's state s 0 ; s is a state for model M such that applying the concrete input event e to M in s results in the new state s 0 . Pictorially, true in s 0 . The return value E is an event schema for (variablization of) the concrete event e, defining fresh formal parameters. P is a state predicate taking the same arguments as P 0 plus the formals of E, and a is a list of actuals for P such that P (a) is true in s. Moreover, we specify that a) if and only if for all states S and actual parameters applying E(AE ) to the model M in S results in a state S 0 in which P 0 (AP 0 To clarify, the formals of P are just the union of the formals of P 0 and those of the event schema E. Thus, the actual list A will have values both for the formals of E and the formals of P 0 . Intuitively, BackProp computes a sufficient (not necessarily necessary) condition on event E and the state prior to applying E, such that P 0 is true afterward. BackProp applies explanation-based generalization [11, 8] to the P-EBF formalism. Others have described similar operators, such as Dijkstra's predicate transformers, or Igerashi et al's verification condition generators[18]. It is beyond the scope of this paper to explain the algorithm in detail, but here is an exam- ple. In state 1, user "1234" has MODE IDLE. The event results in state 2 in which the MODE of "1234" is DIALING. Then BackProp applied to the 1-parameter predicate (EQUAL DIALING (LOOKUP MODE returns the event schema (OFFHOOK ?y) and predicate (AND (EQUAL :IDLE (LOOKUP MODE ?x)) (EQUAL ?x ?y)). (The actuals lists bind both ?x and ?y to "1234".) Intuitively, this means that if we offhook any idle user, that user will move to the dialing mode. BackProp*. Note that if we have a succeeding scenario trace involving a sequence of input events, we can iteratively apply BackProp to get an entire generalized scenario, where the initial predicate will not depend on the state at all (because ISAT scenarios are defined never to succeed if they depend on uninitialized state values). The rest of the paper will refer to this operation as BackProp*; it takes in a model, a scenario trace, and a predicate to be backpropagated together with its satisfying actuals list, and returns this fully backpropagated predicate, its actuals list, and the list of event schemas making up the generalized scenario. 2.2 ISAT Tools Overview ISAT exploits the two hypotheses above to provide a suite of analysis tools to help the designer produce a specification that meets the true needs of the requirer. ISAT includes the following tools: ffl Scenario simulation takes a scenario and a model and executes the model to determine whether the scenario represents correct behavior of the model. Thus, requirements scenarios can be directly validated ffl Coverage analysis reports states never reached by, and statements of the model that are not executed by, any of the requirement scenarios. This helps the designer elicit adequate requirements from the requirer. ffl Layered theorem proving[12, 20] is a technique for proving arbitrary correctness properties, such as state invariants and pseudo-state diagrams[12]. ffl Conflict detection[14] returns predicates describing states in which the model, if it reaches them, will derive an inconsistent next state (potentially causing either a crash of the simulator or, worse, the implemented system). Inconsistencies can result from setting state relations to two inconsistent values or raising conflicting output events, such as playing both the ringback tone and the busy tone at the same time to the same phone. Coverage analysis, conflict detection, and proof attempts produce state predicates to which we can apply a scenario generator in order to discover whether they represent reachable states of the model. 3 The Scenario Generation Problem Formally, the scenario generation problem is to take a model M and state predicate P 0 and find a sequence L of concrete input events and a list of actual parameters executing L in M starting from the undefined initial state results in a state s 0 satisfying I have concentrated on conjunctive state predicates, i.e. those whose expression consists of the logical AND of a collection of predicates. The method can be applied to disjunctions of conjunctive state predicates by applying it concurrently to each of the disjuncts, but that requires engineering for efficiency that is beyond the scope of this paper. Sections 1 and 2 discussed some ways a tool suite can benefit from solving this problem. Why Rich Formalisms? Model checkers[17] and symbolic model checkers[5] guarantee that when they find a property not valid in a model, they return a concrete counterexample (scenario) illustrating the vi- olation. Thus, we should explore under what circumstances these tools solve the scenario generation problem before inventing different ways to solve it. Model checkers exhaustively search the state space of the system, testing the property in each state. Thus, they are limited by the size of the state space they can handle. Some model checkers exploit limited forms of state space symmetry to handle systems with larger spaces, but all eventually run into this "state explosion problem". And while symbolic model checkers have checked properties in impressively large (10 120 spaces, it is not clear if the technique can be extended to handle nonboolean logics. For a survey of model checking and its relation to theorem proving for verification, see[6]. Should we simply avoid models with large state spaces? I believe the answer is "no." Several common types of design problems are only manifest in more complex (large or unbounded state space) models of a system. For example, complex systems are frequently designed in a modular fashion by designing functional "features" independently and then combining feature sets to meet customization or market needs. Telephone switching systems are a good example of this approach, yet many other systems are built this way. The problem is that even though individual features are valid in isolation, their combination may lead to undesirable interactions that lead to faulty behavior. The only way for a tool to discover these interactions is to model the feature combinations; it follows that the more features a system has, the more complex must its model be in order to detect interactions. Another reason limited-space approaches are not the final answer is that it is difficult both to do enough abstraction to make the problem tractable and yet to retain enough detail to manifest the problems of interest. In particular, each property to be checked may require a different, hand-constructed model ab- straction. And since designers don't know in advance which problems the system has, there could be a lot of wasted effort and/or false confidence in results. By dealing with more complex models, the abstraction can be relatively straight-forward, and a single one can be used for all properties. Finally, another reason to prefer a single, easily produced abstraction that is clearly faithful to the system, is that there is the possibility of generating implementations directly from the models, either through code synthesis or by direct manual implementation. Of- ten, abstractions that are necessary for tractability are missing too much detail to allow any direct mapping to implementations. For example, Alur et al[1] report on a model checking effort for a phone switch in which it was necessary to model queue data structures by 7 bit integers (representing the number of items in the queue). An implementation must supply all details of queue implementation, as well as any system behavior depending on the actual contents of the queues. Why is scenario generation hard? As soon as our representation language allows event and state parameterization and functions, we have added an uncomputable constraint satisfaction problem to the problem of combinatorial search in large state spaces. For example, designers commonly need models with arithmetic, lists and other data structures, text manipulation functions such as pattern matching, etc. But then it is possible to define systems and properties that are only satisfied when the system reaches a state satisfying an arbitrary sentence of this rich theory. Proving such a state reachable is undecidable, by G-odel's Incompleteness Theorem; generating a scenario that actually reaches it is even harder because of the combinatorial search. Thus, in summary, we want to be able to apply scenario generation to complex modeling formalisms, and yet the problem goes from merely search to un- computable. Our only hope in these cases is to find an approach that can solve the problem in usefully many cases, and not take too long doing it. We also require that whenever the tool returns a scenario, it actually satisfies the goal predicate (soundness). These are the goals of the SGEN 2 approach. 4 The SGEN 2 Approach Let us term the overall conjunctive state predicate the "goal predicate" and the individual conjuncts making it up the "conjunct predicates" or simply the "conjuncts." There are two key insights behind the algorithm. First, the library of requirement scenarios, while unlikely to have a scenario which reaches a state satisfying the goal predicate, nevertheless is likely to have scenarios that reach states satisfying sets of the conjuncts. Thus, we might find such scenarios and somehow paste them together into a single scenario that achieves the full conjunction. typically two such scenarios will operate on different data items; for example, scenario 1 may achieve set 1 of the conjuncts for address "1234", while scenario 2 achieves set 2 for address "5678". Thus, these two concrete scenarios cannot be interleaved to form a scenario that achieves the union of the sets for a single address. However, the second key insight is that we can solve this subproblem by abstracting the two scenarios, using BackProp*, and finding a common instantiation of them (binding of their variables to data values) such that the union of the two predicate subsets is satisfied. Once such a common instantiation is found, a heuristic search merges the two event sequences into one, achieving the union of the conjunct sets. Appendix A gives a precise high-level pseudocode description of the SGEN 2 algorithm. The following illustrative example is taken from the case study. Consider the goal predicate (and (member ?y (lookup known-addresses)) (lookup fpr-active ?y) (equal dialing (lookup mode ?x)) (lookup tcs-active ?y) (member ?x (lookup tcs-screened-list ?y))) This describes states in which known address ?y has two features, "fpr" and "tcs" both active, with ?x on its tcs-screened-list, and in which ?x is dialing. Initialization. SGEN 2 first mines its library and discovers scenario (init) (init-address "1234") (activate-tcs "1234" "1234") (offhook "1234") which results in a state satisfying 4 of the 5 conjuncts: (and (member ?y (lookup known-addresses)) (equal dialing (lookup mode ?x)) (lookup tcs-active ?y) (member ?x (lookup tcs-screened-list ?y))) when we bind both ?x and ?y to "1234". Since it is unlikely we will find another scenario that fortuitously achieves the rest of the goal for the constant "1234", we apply BackProp* to the above predicate and the trace of scenario S 1 to get the generalized scenario (init) (init-address ?x) (activate-tcs ?y ?x) (offhook ?x) subject to the backpropagated condition (equal ?x ?y). SGEN 2 also records the actual bindings recursive step. SGEN 2 -REC continues by searching the mined library information for satisfiers of the remaining conjunct(s) of the goal. In this case, it discovers (among others) that the scenario S 2 (activate-fpr achieves the remaining conjunct (lookup fpr-active ?y) when ?y is bound to "5678". Note that since S 1 and S 2 operate on different constants, they cannot be directly interleaved to get a scenario reaching the desired conjunction. Applying BackProp* to the remaining conjunct and the trace of S 2 , we get the generalized (activate-fpr ?y ?t1 ?t2 ?w) subject to no constraints (other than implicit type constraints), with actual bindings: "1357"g. calls the Coinstantiate routine which attempts to find a common instantiation of G 1 and G 2 obeying both sets of constraints. In this case, since the constraint set for G 2 is empty, Coinstantiate quickly finds that the common instantiation sets. finally calls MergeScenarios on the two scenarios G 1 (I) and G 2 (I), which denote the instances of G 1 and G 2 obtained by applying I . MergeScenarios also takes the two predicates are satisfied by G 1 (I) and respectively, so that it can check whether its result satisfies both simultaneously. In the case above, MergeScenarios finds the following interleaving which does, indeed, satisfy the conjunct sets. (init) (init-address "1234") (activate-tcs "1234" "1234") (activate-fpr (offhook "1234") If at this point there were still unsatisfied conjuncts of the goal, SGEN 2 -REC would call BackProp* to generalize this result scenario and then recur to search for yet another scenario to satisfy the next subset. If Coinstantiate or MergeScenarios fails, then move on to the next candidates in the search (cf Appendix A). 4.1 Library Mining The first step of SGEN 2 is to search the library of execution traces of requirement scenarios for states in which sets of conjuncts are satisfied. The subroutine MineLibrary accomplishes this as follows. For each scenario in the requirements library, it first generates an execution trace by calling the simulator. It then extracts from the trace sets of data values (grouped by type) appearing in the trace. Then, for each possible well-typed assignment of these data values to the formal parameters of the goal predicate, it searches the states of the execution trace for those in which a conjunct first becomes true (for that parameter as- signment). It creates a predicate group satisfier (pgs) for that state, which records the assignment and which set of conjuncts are satisfied. This set of satisfied conjuncts is termed the satset of the pgs. MineLibrary returns the entire collection of PGSs found in this way in all traces. It sorts the list in decreasing order of the size of the satset so that SGEN 2 will consider earlier those PGSs that satisfy the most predicates at once. MineLibrary is linear in the total number of states in all traces in the library. More importantly, however, it is proportional to the number of parameter assignments, which is exponential in the number of goal predicate parameters. While the current implementation seems to work adequately fast on the case study examples (- 5 parameters each), it may be necessary to limit the number of assignments considered when the goal predicate has many parameters. 4.2 Coinstantiation Coinstantiate heuristically attacks the (in gen- eral) uncomputable problem of coinstantiation by simply trying out all possible well-typed assignments of constants to the parameters of G 1 and G 2 , where the constant pool is simply the union of all constants in the actual-bindings of the PGSs from which G 1 and G 2 were generalized. This has proven effective in the case study, and takes negligible time (see statistics below). If necessary, Coinstantiate can be made to consider larger constant pools, such as those in all scenarios. 4.3 Scenario Merging MergeScenarios takes in two scenario/predicate pairs, where each scenario results in a state satisfying its predicate. The goal is to return an interleaving of the two scenarios that satisfies both predicates. MergeScenarios does not attempt to check all possible interleavings, as this would require checking exponentially many (in the sum of the lengths of the two input scenarios) interleavings in the worst case. And note that the worst case occurs any time no interleaving exists, so it is fairly common. Designate the input scenario/predicate pairs as the "left" scenario and predicate and the "right" scenario and predicate. Our approach is to sequentially select the front event off of either the left or right scenario and add it to the end of the result scenario until both left and right are empty. Doing this in all possible ways, waiting until left and right are empty before checking the predicates, would result in the exponential worst case mentioned above. Instead, MergeScenarios heuristically limits the search as follows. Each time it selects an event e l from the left scenario, it checks to see whether, if the result scenario were extended from that point with the remainder of the right scenario, the right predicate would be satisfied. If not, e l is vetoed; otherwise, it proceeds to the next choice. (By induction, one can show that if instead we extended the result with the remainder of the left scenario, the left predicate would also be satisfied.) The dual check is done when the event is selected from the right. When the front events on left and right are identical, the algorithm also attempts the third option of adding one event and discarding the other. Note that since there can be interleavings that satisfy both predicates at the end but which contain intermediate points at which the check would fail, this approach is less powerful than brute force search; how- ever, in the case study, MergeScenarios only failed Total Scenario Satisfiable/ Not Attempts Generated No scenario Satisfiable Table 1: SGEN 2 success on case study once when a brute force search would have succeeded, and yet was as much as 12 times faster (average: 2x). 5 Case Study I ran SGEN 2 on 63 distinct scenario generation problems that arose in a larger case study of feature interactions in a telephone switch specification. (The study actually produced 66 problems, but three were duplicates, so were discarded for this paper.) The larger study is actually a tool contest associated with the 1998 Feature Interactions Workshop[7]. The system being modeled is a telephone switch implementing Plain Old Telephone Service (POTS), plus 12 functional features such as Call Forwarding (CF), Terminating Call Screening (TCS), FreePhone Routing and nine others. This SGEN 2 case study was performed before four of the twelve were modeled, so only POTS and eight features are included here. In a related paper[14], I explain how I used the ISAT tool set to model these specs and to detect various types of feature interactions among them, many of which are predicates describing states in which undesired things may happen, such as feature inconsistencies becoming manifest (conflicts) or feature correctness properties being violated. In the absence of a scenario genera- tor, it is left to the user to determine whether those state predicates describe reachable states of the model. Thus, these 63 problems provide a moderately complex test of the power and usefulness of a scenario generator, and are representative of the problems that may be encountered by such a tool. The full data is available at [15]. Results. The 63 predicates averaged 1.72 parameters and 5.98 conjuncts each. Table 1 shows the results of running the generator. "Scenario Generated" refers to trials in which SGEN 2 succeeded in finding a scenario; "Satisfiable/No Scenario" refers to the cases when it failed to find a scenario, even though the predicate is satisfiable; and "Not satisfiable" refers to those cases determined (through external means) to be unsatisfiable and, hence, there exists no scenario to generate. Scen.Gen No Scen.Gen All Only Only Total 8938 603 8335 Library Mining 658 510 148 BackProp 1766 Coinstantiation Merge Table 2: SGEN 2 aggregate run times (rounded to nearest second). Table shows run time statistics for the 63 tri- als. All times are measured on a 225 MHz Macintosh clone (144 MB memory) running the ISAT system under Macintosh Common Lisp 4.2. For this table, the "no.scen.gen only" condition includes all cases where the tool did not find a scenario, whether or not the goal predicate was satisfiable (since to the user these are equivalent when waiting for the tool to finish). Discussion. Of the 60 cases in which it was possible to generate a scenario, SGEN 2 succeeded 40% of the time. Thus, the user can be sure that at least those cases illustrate real design errors and therefore concentrate first on fixing them. Note that one error can cause scenarios to fail (due to conflicts) that would otherwise succeed far enough to reach a second error state. Thus, fixing an error can cause SGEN 2 to succeed when it failed previously. I know of one definite case (and some others suspected) where this sort of error interference occurred in the case study. When I first ran the study, a few cases failed because individual conjuncts were not covered by the scenario library. Of course, if there is no known way to satisfy a single conjunct, the goal predicate won't be satisfied either. Fortunately, it is relatively easy to discover a scenario covering a single conjunct, such as (member ?x (lookup tcs-screened-list ?x)). I easily created three scenarios to cover these cases, resulting in one more success and several failures. The results above reflect these additional scenarios. Turning to time, we see that the average time per trial is 142 seconds overall, with succeeding cases taking 25 seconds on average (101 sec maximum) and failing cases requiring 214 seconds (1054 sec maxi- mum). Note that the distribution of time is radically different between succeeding and failing cases, with MergeScenarios dominating for failing cases and MineLibrary dominating for succeeding cases. Coinstantiate was never significant, suggesting that there is room to improve its power (by checking larger constant pools, for example) without significantly harming the overall run time. On the other hand, we must be extremely careful in increasing the power of MergeScenarios since that is the bottle-neck in failing cases. These results are only intended to be suggestive of future algorithmic improvements; I believe they can be significantly reduced by a careful re-engineering effort. (The current ISAT system is an exploratory prototype.) Note also that these results depend on the model and scenario library as well. In summary, it seems that at least for validation purposes an imperfect scenario generator can still be quite useful as long as it doesn't take too long. Of course, we can always hope for a better success av- erage, and future work will go into improving the heuristics. However, it is desirable to keep the times relatively low in all cases, including failure cases, so the tool is still usable. Thus, we must engineer the power/speed tradeoff carefully. 6 Related Work Having discussed model checkers above, I will only summarize here. Model checkers are useful solutions to the problem of scenario generation as long as one can effectively generate models in the limited formalism necessary to run the tool tractably. However, there is reason to believe that we need to handle the more complex formalisms addressed in this work for at least the reasons discussed in Section 3. In addi- tion, we may wish to use scenario generation in ways beyond validation, such as online help systems. For comparison, it is amusing to estimate the state space size necessary to model the telephony case study specs in a finite state formalism. If we model all 12 features for n users, I estimate there are at least reachable states (logs are to base 2). If we consider call-waiting and similar features, we need at least 3 users, but if we add forwarding and other multi-user features, one can easily imagine properties referring to 6 or more users, leading to which would challenge even the best model checkers. Note that even infinite-state model checkers, such as that of Bultan et al[4], are highly restrictive. That system is restricted to state spaces that are the cross product of a boolean state space and one representing integer inequalities (higher dimensional polyhedra). While increasing model checking power by adding specialized constraint reasoners shows promise, it is not even clear that most reactive systems people design will be expressible within such restricted formalisms, due to the common occurrence of functions mixing arguments of several different types. Another class of approaches to the problem that may seem applicable are AI-based planners, such as STRIPS[9] or Prodigy[19]. The problem with applying these systems, where the spec model provides the planning operators, is that planning operators must explicitly list their consequences; for example, STRIPS operators must have ADD and DELETE lists. Simi- larly, the macro operators learned by the EBG-based PRODIGY system must explicitly include the goal(s) they achieve. This is too limiting, because users of scenario generation may give any goal statement they wish in terms of functions defined in the logic. Any planning operator derived from the spec model potentially achieves too many (infinitely many, in fact) different goals; far too many to be stored explicitly even if we could bound the vocabulary. SGEN 2 avoids this problem by doing its abstraction and reasoning on the fly in MineLibrary and BackProp*. The only knowledge stored is the "raw" scenario traces, unadorned with any goal information. There is also work in the traditional testing literature on generating test inputs to cover a given path in a program. For example, Gotlieb et al[10] describe a constraint-based approach, which essentially reduces to trying to find a satisfying assignment for a boolean functional expression which is, of course, uncomputable once we enrich the formalism to include (e.g.) arithmetic. However, the constraint based approach may prove useful in improving Coinstantiate and MineLibrary; it does not address the state space search needed to handle reactive systems. Finally, there are other spec modeling tool suites providing some of the same (and many contrasting) tools as ISAT, such as the SCR tool suite[16]. Such environments may incorporate model checking, but none capable of dealing with rich formalisms have scenario generators, to my knowledge. 7 Limitations and Future Work The most basic limitation of SGEN 2 is that it is fundamentally a hill-climbing algorithm. In particu- lar, there are examples in the case study which are easily solved by merging two scenarios from the li- brary, but which SGEN 2 cannot find. As an example, one scenario achieves a particular conjunct set halfway through its event sequence, but then has several more steps that are removed by BackProp* as irrelevant to achieving that set. It turns out they are necessary, however, if one wishes to later merge it with a second scenario achieving the rest of the goal. (These "extra" steps are things like hanging up a phone after activating a feature, because a subsequent scenario must start from the idle state.) power comes from the richness of the scenario library; it is therefore likely to be more useful in development processes and environments that encourage the formalization of such scenarios. SGEN 2 pro- vides, perhaps, a new argument in favor of integrating formal scenarios into the software process. SGEN 2 is still in its early youth, and there are many ways it can be improved. For example, in its search, SGEN 2 only considers the first PGS having a given sat-set. A better, but more expensive, approach is to try a PGS if and only if its BackProp*- generalization is not isomorphic to one seen previously. The effect of this on run-time must be monitored, how- ever. MergeScenarios, being the time bottleneck on failing cases, may profit from more work on limiting its search. MineLibrary needs to search fewer cases when the predicate takes many parameters. Of course, results from one case study are not con- clusive, so future work should investigate SGEN 2 's effectiveness on other domains and systems. Conclusions A scenario generation tool can be useful in a specification modeling tool suite, in focusing attention on design errors demonstrably present, in helping communicate errors in the requirements, and even in implementing online help systems. Exhaustive-search approaches, such as model checking, while useful, are not tractable in rich formalisms allowing more direct system models to be expressed. SGEN 2 is a heuristic approach to a highly uncomputable problem, based on the simple idea of piecing together partially satisfying scenarios from the requirements library, using explanation-based generalization to abstract them in order to be able to coinstantiate them. Results from the case study are encouraging; SGEN 2 seems to succeed often enough to be useful and yet be efficient enough to be engineered into an interactive tool. While the work needs further empirical validation, it seems promising and should be pursued. --R Model checking of real-time systems: a telecommunications application A computational logic hand- book Verifying systems with integer constraints and boolean predicates: a composite approach states and beyond Formal methods: state of the art and future directions Feature Interaction Detection Tool Contest Learning and executing generalized robot plans Automatic test data generation using constraint solving techniques Systematic incremental validation of reactive systems via sound scenario generalization Reactive system validation using automated reasoning over a fragment library How to avoid unwanted email Feature combination and interaction detection via foreground/background models Complete case study data for this paper Automated consistency checking of requirements specifi- cations Design and validation of computer protocols Automatic program verification I: a logical basis and its imple- mentation Quantitative results concerning the utility of explanation-based learning Seven layers of knowledge representation and reasoning in support of software development --TR
scenario generation;validation;reactive system;explanation-based generalization
592050
Specification-Based Browsing of Software Component Libraries.
Specification-based retrieval provides exact content-oriented access to component libraries but requires too much deductive power. Specification-based browsing evades this bottleneck by moving any deduction into an off-line indexing phase. In this paper, we show how match relations are used to build an appropriate index and how formal concept analysis is used to build a suitable navigation structure. This structure has the single-focus property (i.e., any sensible subset of a library is represented by a single node) and supports attribute-based (via explicit component properties) and object-based (via implicit component similarities) navigation styles. It thus combines the exact semantics of formal methods with the interactive navigation possibilities of informal methods. Experiments show that current theorem provers can solve enough of the emerging proof problems to make browsing feasible. The navigation structure also indicates situations where additional abstractions are required to build a better index and thus helps to understand and to re-engineer component libraries.
Introduction Large software libraries represent valuable assets but the larger they grow, the harder it becomes to capitalize them for reuse purposes. The main problems are to keep the overview over the library and to extract appropriate components. This requires better library organizations and retrieval algorithms than a linear search through a at list of components. Libraries are thus often structured by syntactic means, e.g., inheritance hierarchies. But this is misleading because it need not to express any semantic relation between components. Information science oers semantic methods for library organization and component retrieval e.g., [17, 24], but these methods are informal because they rely only on the meaning conveyed by words. As a more exact alternative, the application of formal specication methods to software libraries has been investigated, starting with [10, 23, 25]. The general idea is quite simple. Each component is indexed with a formal specication which captures its relevant behavior. Any desired relation between two components (e.g., renement or matching) is expressed by a logical formula composed This work is supported by the DFG within the Schwerpunkt \Deduktion", grant Sn11/2-3. from the indices. An automated theorem prover is used to check the validity of the formula. If (and only if) the prover succeeds the relation is considered to be established. The most ambitious of these approaches is specication-based retrieval [21, 22, 18, 26, 6]. It allows arbitrary specications as search keys and retrieves all components from a library whose indexes satisfy a given match relation with respect to the key. However, in spite of all research eorts (cf. [19] for a detailed survey), it is still far away from being practicable. Notwithstanding all progress in automated deduction, the required theorem proving capabilities remain the main bottleneck. Here, we investigate a more practical approach, specication- based browsing of component libraries. Its crucial success factor is that any time-consuming deduction can be moved into an o-line indexing phase (\pre- processing") and can thus be separated from navigation. The user works only on the pre-processed, xed navigation structure which re ects the semantic properties of the components with respect to the index. We show that dierent match relations must be used to build an appropriate index and how formal concept analysis can be used to build a concept lattice which serves as navigation structure. Both techniques|specication-based library organization [9, 18] and concept-based browsing [7, 12]|have been proposed before, but their combination is new and unique to this research. It thus combines the exact semantics of formal methods with the interactive navigation of informal methods. Experiments show that this approach is feasible. Apart from writing the specications in the rst place, indexing can be fully automated. Current theorem provers can solve enough of the emerging problems, even with modest timeouts. Calculation of the concept lattice is fast enough and navigation works without delay. Specication-based browsing is not only useful for reuse but also for analyz- ing, understanding, and re-engineering component libraries. Although browsing is dened via specications, they are not actually required for navigation. Instead, symbolic names can be used which \hide" the actual formulas. An intelligent choice of such abstractions can thus speed-up and improve under- standing. The lattice even indicates situations where additional abstractions are required to build a better index. Browsing vs. Retrieval Library browsing and retrieval are closely related but following [19] a clear distinction can be made. Retrieval consists in extracting components which satisfy a predened matching criterion. Its main operation is thus the satisfaction check or matching. The criterion is usually given by an arbitrary user-dened search or query which is matched against the candidates' indices. Browsing consists in inspecting candidates for possible extraction, but without a predened criterion. Its main operation is thus navigation which determines in what order the components are visited and whether they are visited at all. Browsing usually works stepwise and we denote the set of all components which can be visited in the next step as the focus. In contrast to retrieval, it requires no search key but works on a pre-processed, usually hierarchical navigation struc- ture. The obvious although not optimal way to compute such a structure is to order the components by inclusion on their retrieval results using their own index as query. In the specication-based case, these dierences prove to be crucial for the greater practicability of browsing. The pre-processing of the navigation structure allows us to resort to o-line proving and thus to evade the deductive bottleneck. Less obvious but equally important, the construction of the hierarchy via a cross-match of the component library against itself benets the proof problems. Since no arbitrary user specications are involved, the spec- ications are much more uniform in style. This allows some obvious prover tuning; however, the real gain comes from the absence of data mismatches. Consider for example a graph library where the graphs are represented as map from nodes to node sets and a query using a representation as a list of node pairs. Then, the prover must repeatedly, for each candidate, show that both data representations are equivalent. Although signature matching can mitigate the data mismatch problem [6], it is still the major source of complexity in deduction-based retrieval. Renement Lattices Reconsidered Formal specications can be used to order components and hence to organize libraries hierarchically. These hierarchies can then be exploited to optimize retrieval or to compute a navigation structure. The obvious question is how to order the components and the obvious answer is by renement or plug-in- compatibility [21, 6]. Given two components G and S with respective axiomatic specications (pre G ; post G ) and (pre S ; post S ), S is said to rene G (or to be more specic than G; S w G, or G to subsume S), i (pre G ) pre holds. 1 Intuitively, (1) expresses the fact that S can be plugged into any place where G is used because it has a wider domain and produces more specic results than G. Using a relational view (i.e., specications are considered as sets of valid (input, output)-pairs), [18] show that (1) denes a partial order which induces a lattice-like structure on the set of all specications. This structure is generally known as the renement lattice although it is strictly spoken no lattice. Turning the renement lattice into a navigation structure for library browsing exposes, however, some unexpected problems. First of all, libraries do not oer enough structure, i.e., the renement hierarchies they induce are too shal- low. While this is a good thing from a design point of view|it simply says that the library contains only little redundancy|it is a bad thing for browsing. It can be overcome by the introduction of meta-nodes or abstractions. Such specications do not represent real, existing components but just factor out 1 For the sake of brevity, we shall omit the quantication over the respective argument and return variables and their identication via type compatibility predicates. For arguments ~x and result variables ~y, the full form is pre S (~x S post G (~x G ~y G ))). similarities between some of them. As an example, consider the specication of an abstract element lter: 2 lter some (l : list) r : list pre l post exists l1, lter some species only that a singleton element is removed from the list (hence it cannot be empty) but not which one. It is thus via (1) rened by both components tail and lead: list lead (l list pre l pre l post exists post exists However, a nave introduction of meta-nodes yields unexpected results. If we introduce another meta-node segment segment (l list pre true post exists l1, l2 : list & to capture the property that both components return continuous sublists of their argument, this does not work: neither tail nor lead rene segment. The reason for this at rst glance counterintuitive behavior is that segment is specied as a total function (pre segment both tail and lead are partial. And while we can x this particular aw by setting segment's precondition also to becomes increasingly infeasible. If the library also contains components which work on sorted lists only, we have to integrate this property into the precondition, too. In eect, if we want an abstraction which captures all segment-like components we have to adjoin all occurring preconditions con- junctively. If, however, two of them are contradictory the result becomes false and segment subsumes the entire library. The solution to this dilemma is easy. While we can use renement to index components with abstractions, we additionally need a second relation to model the above situation. Since we are only interested in the eect of the calculation (i.e., the postcondition post G ) we can drop pre G . We want post G to hold on the appropriate domain only, hence pre post G (2) which is also known as conditional compatibility [6] or weak post match [21] in deduction-based retrieval. We can then consider G as derived attribute or feature [22] of holds whenever the execution of S was legal (pre S holds) and terminated (post S holds.) In our example, segment is a feature of tail and lead, as expected. It is easy to verify that features are inherited along with the renement relation, i.e., if R renes S and G is a feature of S, then G is a feature of R, too. We use VDM-SL for our example specications. Here, ^ means concatenation of lists, [ the empty list, [i] a singleton list with item i. & reads as \such that". A similar problem arises when we want to consider preconditions only. While we can use the abstraction total list pre true post true to subsume all total functions, it is much harder to index partial functions properly. The meta-node requires non empty (l list pre l post true correctly subsumes all functions which work on non-empty lists only but it is not really appropriate: it also subsumes all total functions and is thus not discriminative. Hence, we need a third relation. Since we are now only interested in the properties of the legal domains, we can drop the postconditions. But in contrast to renement we want the domain of S now to be more restricted, hence pre S ) pre G (3) Again, G is a derived attribute of S|it is a requisite, S w r G|and using (3), requires non empty now works as index. Requisites are also compatible with renement but in contrast to features their absence is propagated. If R renes S and G is no requisite of S, then G cannot be a requisite of R. top run lead tail duplicate rst requires non empty segment front segment works on empty total lter some top run w w f w f w w lead tail w w r w f w duplicate rst w w r Figure 1: Example index Figure 1 shows the index for the examples in this paper. The components are represented as rows, the attributes as columns; the symbols indicate which relations have been used to index the components with the respective attributes. We also see that the library is indeed shallow: each component indexes only itself. However, (1-3) are not the only sensible relations we could use. Instead of indexing a component S with its requisites, we could also index S with all requisites it does not require, i.e., with all its valid border conditions. In terms of preconditions, this is formalized by :pre G ) pre S (4) and denoted by S w r G: G is not a requisite for S, or S also works on G. Hence, we have of course tail 6w r requires non empty but for a topological sort function top list pre acyclic(l) post top we have top sort w r requires non empty as expected. However, in principle, (4) is not necessary. We can achieve the same eect using a modied version works on empty (l list pre post true of requires non empty and renement: top sort w works on empty. 3 However, this hides the fact that requires non empty and works on empty are complementary to each other. We now use (1-3) 4 to compute an appropriately modied version of the renement lattice but even this variant is not yet adequate for browsing. It still lacks the single-focus property, i.e., it does not contain enough structure to represent the focus by a single node. Consider for example lead and tail. Apart from further renements, they are the only two components which have the feature segment and are subsumed by lter some at the same time. 5 Yet there is no meta-node to represent this and a user has to keep his focus on both distinguishing properties to capture the conceptual similarity of the components. The deeper reason for this is that even the modied renement lattice has lattice-like properties only on the set of all possible specications, not on arbitrary subsets or libraries. True lattices, on the other hand, have the single-focus property by denition and we will show how to embed the renement lattice into a true lattice using formal concept analysis. 4 Concept Lattices 4.1 Formal Concept Analysis Formal concept analysis [30, 8, 3] applies lattice-theoretic methods to investigate abstract relations between objects and their attributes. A concept lattice is a structure with strong mathematical properties which reveals hidden structural and hierarchical properties of the original relation. It can be computed automatically from any given relation. 3 Notice that this relies on the fact that post works on empty = true|otherwise, the postcondition part of (1) would not be valid. 4 We still need renement to represent all information of interest. E.g., we cannot split total into a requisite and a feature which have both the value true because both of them index the library. 5 By lter some we have to remove an element, but by segment we are not allowed to split the list. Hence, there are only the two choices to remove the element at either end of the list. Denition 1 A formal context is a triple (O; A; R) where O and A are sets of objects and attributes, respectively, and R O A is an arbitrary relation. Contexts can be imagined as cross tables where the rows are objects and the columns are attributes. Hence, the index shown in Figure 1 can also be considered as a formal context, provided that the dierent relations (i.e., w; w r and w f ) are merged. Denition 2 Let (O; A; R) be a context, O O and A A. The common attributes of O are dened by (O) def Rg, the common objects of A by !(A) def Rg. Objects from a context share a set of common attributes and vice versa. Concepts are pairs of objects and attributes which are synonymous and thus characterize each other. Denition 3 Let C be a context. is called a concept of C A and O and A (c) def are called c's extent and intent, respectively. The set of all concepts of C is denoted by B(C). Concepts can be imagined as maximal rectangles (modulo permutation of rows and columns) in the context table, e.g., (flead, tailg, fsegment, requires non empty, lter someg). They are partially ordered by inclusion of extents (and intents) such that a concept's extent includes the extent of all of its subconcepts (and its intent includes the intent of all of its superconcepts). Denition 4 Let C be a context, c 1 and c 2 are ordered by the subconcept relation, c 1 c 2 , i O 1 O 2 . The structure of B and is denoted by B(C). The intent-part follows by duality. As an immediate consequence of the preceding denitions we get that the strict order corresponds to strict inclusion of extents and intents, i.e., c 1 and A 1 The following basic theorem of formal concept analysis states that the structure induced by the concepts of a formal context and their ordering is always a complete lattice. (Cf. Figure 2 for an example lattice.) Theorem 5 ([30]) Let C be a context. Then B(C) is a complete lattice, the concept lattice of C. Its inmum and supremum operation are given by i2I i2I O i2I _ i2I i2I O i )); i2I Each attribute and object has a uniquely determined dening concept in the lattice. This can be calculated from the attribute or object, respectively, alone. Denition 6 Let B(O; A; R) be a concept lattice. The dening concept of an attribute a 2 A (object is the greatest (smallest) concept c such that a 2 A (c) (o 2 O (c)) holds. It is denoted by (a) ((o)). Theorem 7 ([3]) In any concept lattice we have and 4.2 From Renement Lattices to Concept Lattices [12] has shown that keyword-indexed components can be considered as a formal context with the components as objects and the (informal) keywords as attributes. We now lift this idea to formal specications. Denition 8 Let be a formally specied library with components L, requisites R, features F , and abstractions A. Its induced context is dened by Again, we consider the components as objects, and, of course, the keywords are replaced by (the names of) the specications 6 but the context table is slightly more complicated. To prevent dierent components from \collapsing" into a single concept if the index is insu-cient, the component specications L double as objects and attributes. The relations, however, remain original. works on empty (i) segment @ @ @ @ @ (ii) requires non empty total front segment (iii) lter some top sort top sort run run (iv) lead lead tail tail duplicate rst duplicate rst Figure 2: Example lattice We then calculate the concept lattice from this context. Figure 2 shows the result for the example context. Each bullet represents a concept. The labels 6 Wlog. we assume that L; R; F , and A are pairwise disjoint. over the bullet are the attributes dened at this concept. E.g., the concept (iii) denes the attribute lter some. However, since attributes in this representation are inherited downwards, its intent A is the set fsegment, requires non empty, lter someg. None of the attributes are equivalent in the sense that they index the same set of components. Hence, each concept introduces only one attribute. The labels under the bullet denote the objects dened at this concept, e.g., lead at (iv). Since none of the actual components subsumes an other, each concept introduces at most one object and is atomic if it introduces an object at all. The concept lattice is not an \extension" of the renement lattice: for two attributes a 1 ; a 2 with (a 1 ) (a 2 ) it is possible to be completely unrelated, i.e., neither of the relations (1-3) holds. However, for two reasons, it is an adequate representation of the index. First, subconcepts preserve renement of the original components. Second, a superconcept can be distinguished from any subconcept by an attribute which is not valid for at least one component in the extent of the superconcept but is valid for all components in the extent of the subconcept. Formally: Theorem 9 Let B(CL ) be the concept lattice of the context CL induced by a library L and c 1 ) such that either 1. 9m 2 O 2. 9 a 2 R a 2 A (c 1 9 a 2 F a 2 A (c 1 9 a 2 A a 2 A (c 1 This theorem, which follows from denitions 3 and 4, makes the concept lattice already suitable for specication-based navigation. However, we can impose even more structure if we double R and use w r in addition to dene the induced context. Then, Theorem 9 holds appropriately and, additionally, we get Theorem 10 Let B(CL ) be the concept lattice of the context CL induced by a library L. Then, for any two complementary requisites a; a 2 R we have a and consequently (a) u Hence, the dening concepts of two complementary requisites are complementary to each other in the lattice. Moreover, their extents divide the entire library into two partitions which is not the case for two arbitrary complementary nodes in the lattice. 5 Navigation in Concept Lattices [12] has also shown how concept lattices can be used as navigation structure for interactive and incremental retrieval (i.e., browsing in our terminology). The focus is represented by (the extent of) a concept. Narrowing the focus is a downward movement in the lattice and is done in two steps: 1. The user selects an additional attribute. As a consequence of the lattice structure, the system can support this selection by calculating all attributes which actually narrow the focus but do not sweep it entirely. It can thus prevent navigation into dead ends (i.e., an empty focus.) 2. The system calculates the new focus in the lattice as the meet (which exists due to Theorem 5) of the actual focus and the dening concept of the selected attribute (obtained by Theorem 7.) Similarly, the focus can also be widened again by de-selecting an attribute. The system then calculates the new focus using the join operation. In the specication-based case, navigation works quite similar. We use the derived properties (i.e., R; F , and A) as navigation attributes. Since the property sets are pairwise disjoint, we can even split the set of navigation attributes into three dimensions. These dimensions are not independent of each other but can be selected independently because all interdependencies are contained in the concepts of the lattice. If we use the modied context (i.e., double R and use (1-4)), we get a fourth dimension. This is still independent but due to Theorem 10, independent selection from R and R is not benecial. Instead, we can toggle between them, in addition to selection/de-selection. Initially, all attributes are de-selected and the focus concept is >: the focus is the entire library. Now, for an example, assume that we select segment. This reduces the focus to O tailg. Further renement is possible by attributes whose dening concepts have a strictly smaller but non-bottom meet with the current focus concept. Thus, for (i), any navigation attribute is possible. If we select requires non empty, the new focus concept is (i) u the choice of requires non empty eliminates run from the focus. Moreover, it leaves front segment as the only possible further renement. This navigation style is attribute-based : the focus is essentially a function of the selected attributes. Due to their dual nature, concept-lattices also allow object-based navigation. Here, the user selects or de-selects a single component and the system calculates the new focus similarly. However, selecting an additional component widens the focus and is thus realized by the join operation. While attribute-based navigation depends on the explicit and learned choice of functional properties and thus is more suited for reuse purposes, object-based navigation exposes implicit conceptual similarities of components: the intent of the focus concept contains all properties which are common to all selected components; its extent also contains all other components which share these properties, even if they have not been selected explicitly. Hence, it is more appropriate for library understanding and re-engineering. 6 Practical Aspects We made a series of experiments to support the claim that browsing is more practical in the specication-based case than retrieval. For these, we used a variant of the list processing library which we also used in our retrieval experiments [6]. It comprises 5 requisites, 31 features, and 86 components and abstractions. All example specications in this paper are taken from that library. 6.1 Calculation of the Renement Lattice Even if the calculation of the renement lattice is done in advance and is thus not time-criticle in principle, it is not obvious that it is feasible at all. Two questions are of main concern: 1. How high is the computational eort in practice? 2. How di-cult are the proof problems in practice? Are current theorem provers powerful enough? The answer to both questions depends on the number and structure of the arising proof problems. At rst glance, it seems that we have to check each requisite, feature, ab- straction, and component against each other to calculate the modied rene- ment lattice. However, in practice this can be optimized due to three obser- vations. First, we do not need to compare the components and abstractions pairwise but can use recursive comparison as in [9] because renement is tran- sitive. Then, we do not need to check requisites and features against each other but only against the components and abstractions. Finally, since the former are compatible with renement, we can \sink them in" once we have the renement lattice on the other nodes ready. In the worst case, the number of problems is thus O(jR [ F j jA [ Lj 2 ). Nevertheless, still too many problems arise to be handled manually. As in other software engineering applications, a fully automated system is required which feeds and controls the prover. However, the sheer numbers become a problem only because most of the proof problems (ap- proximately 85% in our experiments) are logically invalid and thus not provable at all. But theorem provers do usually not check for unprovability and are thus stopped by time-out only. Hence, dedicated disproving lters are required. Nevertheless, the computation is practically feasible. Using techniques from [6] we generated the full set of more than 14.000 proof tasks (i.e., \ready-to- run" versions of the problems which also contain appropriate axioms and prover control information) and ltered out approximately 6.600 as unprovable. This took approximately 8 hours on a Sun SparcStation 20. For simplicity, we did not use the optimizations explained above. This would have reduced the original number of tasks to about 11.000. We then used the automated theorem prover SPASS [29] on a network of PCs to check the surviving tasks. With a time-out of 60 seconds, SPASS was able to solve 1.250 tasks. For the remaining 6.250 problems, we re-generated the tasks, using a dierent subset of the axioms. After a third iteration, SPASS had solved a total of 1.460 or almost 80% of the valid problems. This required a total of approximately 340 hours runtime, or equivalently, a weekend of real time. 6.2 Calculation of the Concept Lattice Concept lattices can grow exponentially in the number of attributes and objects. In practice, however, the worst case rarely occurs and a polynomial behavior is usual. [12] contains more experimental evidence for this. For our example library, the concept lattices from the full (i.e., manually computed) and the approximated (i.e., computed using SPASS) renement lattices contained 153 and 180 concepts, respectively. Their computation took approximately a second and is thus negligible compared to the time required for proving. 6.3 Navigation During our experiments it became quickly obvious that neither the modied renement lattice nor the concept lattice are suitable for presentation because they are too big and complex. [12] makes the same observation and describes a simple text-based interface which works on the attribute and object names only. We are currently adapting this system. The navigation process itself, however, is very fast: the system responds without noticeable delay, even for much larger concept lattices than we are currently investigating. 6.4 Knowledge Acquisition The formal specications of the library components and some initial abstractions 7 must be supplied. Once this seed is available, specication-based browsing can already support further knowledge acquisition. Consider for example a seed comprising lter some, segment, tail, lead, and list pre true post exists l1 : list & list which computes the longest ordered initial subsegment (i.e., run) of a list. From this seed, an initial concept lattice is calculated. Object-based navigation conrms that both tail and lead have a common superconcept, which has the attributes lter some and segment, and, as expected, no other objects. But it also reveals that there is no concept which has the extent of lead and run only| selecting both also causes tail to appear. To disambiguate tail, a feature front segment (l list pre true post exists l1 : list & must be introduced which factors out the common property of lead and run. 7 Related Work Most work on applying specication-based techniques to software libraries examines retrieval only. Relevant for browsing are the investigation of dierent match relations [21] and their eect on software reuse [5, 6]. [22] introduced features as indexes to speed up retrieval. 7 Initial requisites and features can be derived automatically by splitting of the supplied specications. Any resulting indiscriminate attributes are merged into a single concept by construction of the lattice. [9] builds a two-tiered hierarchy from the library. The lower level is based on a modied denition of subsumption which works modulo arbitrary user-dened congruences on literals and is thus unsound in general. The upper level uses a similarity metric derived from the normal forms of the specications. This hierarchy is then visualized to support browsing. [18] only use subsumption to build a hierarchical representation of a library and exploit that only to optimize retrieval. In programming language research, [15] and [16] apply formal methods to the specication and verication of object-oriented class libraries. There, behavioral subtyping corresponds to subsumption. Concept lattices or Galois lattices have been developed as a means to structure arbitrary observations. They have already been applied to various problems in software engineering, e.g., inference of conguration structures [11] or identication of modules [14, 28] and objects [27] in legacy programs. Their application to software component libraries, however, seems to be obvious only in retrospect, and there is only little related work. [7] also uses concept lattices for navigation but presents the entire lattice to the user and oers only a subset of all possible attributes for selection. As far as navigation is concerned, [12] is thus most closely related to our own work. But there, object-based navigation, which is instrumental in knowledge acquisition, is not supported. Conclusions Only specication-based methods can provide exact content-oriented access to software components. Retrieval, however, still requires more deductive power than current theorem provers and hardware can oer. Browsing can evade this bottleneck by moving any time-consuming deduction into an o-line indexing phase. In this paper, we have shown that dierent match relations must be used to index a library properly and how this index is turned into a navigation structure using formal concept analysis. Experiments show that it is feasible to calculate an approximation of the index which is accurate enough for browsing purposes, using current theorem provers and hardware (e.g., SPASS on a small network of PCs.) The computational eort, however, is still high. The concept lattice reveals the implicit structure of a library as it follows from the index. It can even indicate situations where a ner index is required. Due to its dual nature, the lattice allows two complementary navigation styles which are based either on attributes or on objects. Due to the lattice nature, both navigation styles automatically have the single-focus property and refrain the user from dead ends. In our approach, theorem provers are used to derive formally dened properties of components. For navigation, these formal denitions are still available but not actually required|symbolic property names su-ce. However, since informally dened and derived properties (e.g., reliability) are usually also represented by symbolic names (e.g., trusthworty), concept-based browsing allows a smooth integration of formal and informal attributes and thus refutes a conjecture of [1] that formal and informal methods are incompatible. Future work especially concerns scale-up. We expect the fraction of non- theorems to grow further with increasing library size; dedicated disproving techniques are thus one area of interest. Since the remaining tasks are homogeneous in style, learning theorem provers [4, 2] can be expected to perform well on them. Acknowledgments Christian Lindig's work on concept-based retrieval also triggered this research; discussions with him greatly improved my own understanding of formal concept analysis. Comments by Christian, Jens Krinke, and Gregor Snelting improved the presentation of this paper. Christoph Weidenbach did the actual theorem proving at the MPII. --R "Semantic-Based Software Retrieval to Support Rapid Prototyping" "DISCOUNT: A Distributed and Learning Equational Prover" Introduction to Lattices and Order. "Learning Domain Knowledge to Improve Theorem Proving" "Reuse by Contract" "Deduction-Based Software Component Retrieval" "Design of a Browsing Interface for Information Retrieval" Formale Begri "Using formal methods to construct a software component library" "PARIS: A System for Reusing Partially Interpreted Schemas" "On The Inference of Con guration Structures from Source Code" "Concept-Based Component Retrieval" "Assessing modular structure of legacy code based on mathematical concept analysis" "A Behavioral Notion of Subtyping" "Speci cation and Veri cation of Object-Oriented Programs Using Supertype Abstraction" "An Information Retrieval Approach For Automatically Constructing Software Libraries" "Storing and Retrieving Software Components: A Re nement-Based System" "A Survey of Software Reuse Li- braries" "Speci cation Matching of Software Components" "Classi cation and Retrieval of Reusable Components Using Semantic Features" "The Inscape Environment" "Implementing Faceted Classi cation for Software Reuse" "Speci cations as Search Keys for Software Libraries" "NORA/HAMMR: Making Deduction- Based Software Component Retrieval Practical" "Applying Concept Formation Methods to Object caton in Procedural Code" "Identifying Modules Via Concept Analysis" "Spass and Flotter version 0.42" "Restructuring Lattice Theory: An Approach Based on Hierarchies of Concepts" --TR --CTR Katsuro Inoue , Reishi Yokomori , Tetsuo Yamamoto , Makoto Matsushita , Shinji Kusumoto, Ranking Significance of Software Components Based on Use Relations, IEEE Transactions on Software Engineering, v.31 n.3, p.213-225, March 2005 Benedikt Kratz , Ralf Reussner , Willem-Jan van den Heuvel, Empirical Research Similarity Metrics For Software Component Interfaces, Journal of Integrated Design & Process Science, v.8 n.4, p.1-17, December 2004 Ge Li , Lu Zhang , Yan Li , Bing Xie , Weizhong Shao, Shortening retrieval sequences in browsing-based component retrieval using information entropy, Journal of Systems and Software, v.79 n.2, p.216-230, February 2006 Balaji Padmanabhan , Alexander Tuzhilin, On the Use of Optimization for Data Mining: Theoretical Interactions and eCRM Opportunities, Management Science, v.49 n.10, p.1327-1343, October
browsing;software reuse;formal specifications;retrieval;software component libraries
592961
QoS and Contention-Aware Multi-Resource Reservation.
To provide Quality of Service (QoS) guarantee in distributed services, it is necessary to reserve multiple computing and communication resources for each service session. Meanwhile, techniques have been available for the reservation and enforcement of various types of resources. Therefore, there is a need to create an integrated framework for coordinated multi-resource reservation. One challenge in creating such a framework is the complex relation between the end-to-end application-level QoS and the corresponding end-to-end resource requirement. Furthermore, the goals of (1) providing the best end-to-end QoS for each distributed service session and (2) increasing the overall reservation success rate of all service sessions are in conflict with each other. In this paper, we present a QoS and contention-aware framework of end-to-end multi-resource reservation for distributed services. The framework assumes a reservation-enabled environment, where each type of resource can be reserved. The framework consists of (1) a component-based QoS-Resource Model, (2) a runtime system architecture for coordinated reservation, and (3) a runtime algorithm for the computation of end-to-end multi-resource reservation plans. The algorithm provides a solution to alleviating the conflict between the QoS of an individual service session and the success rate of all service sessions. More specifically, for each service session, the algorithm computes an end-to-end reservation plan, such that it guarantees the highest possible end-to-end QoS level under the current end-to-end resource availability, and requires the lowest percentage of bottleneck resource(s) among all feasible reservation plans. Our simulation results show excellent performance of this algorithm.
Introduction With the advances in resource reservation and scheduling techniques, it is possible to provide end-to-end Quality of Service (QoS) guarantees for distributed applications and services. Various resource reservation and scheduling frameworks have been proposed for individual system resources such as CPU [7, 5], network bandwidth [11], disk I/O bandwidth [9], and memory [7]. Now it becomes necessary to create an environment where all these resources can be reserved and scheduled in an integrated manner. In such an environment, an end-to-end multi-resource reservation will be performed for each client requesting a distributed This work was supported by the National Science Foundation under contract number 9870736, the Air Force Grant under contract number F30602-97-2-0121, National Science Foundation Career Grant under contract number NSF CCR 96-23867, NSF PACI grant under contract number Infrastructure grant under contract number NSF EIA 99-72884, NSF CISE Infrastructure grant under contract number NSF CDA 96-24396, and NASA grant under contract number NASA NAG 2-1250. service, so that it can be guaranteed a certain level of end- to-end QoS. A key question in creating such an environment is: for a distributed service, how to determine the best end- to-end QoS level and the corresponding multi-resource re- quirement, under the constraint of current end-to-end multi-resource availability. One difficulty in answering the above question is: the relation between an end-to-end QoS level and the corresponding end-to-end resource requirement can be very complex. Both the QoS level and the multi-resource requirement are generally expressed as partial-ordered multi-dimensional vectors. Every resource contributes to the end-to-end QoS, and there may exist trade-offs between different resources for the same end-to-end QoS level. In this case, the multi-resource requirement can not be determined by looking at these resources separately. Instead, it must be determined by a coordinating entity placed on top of these resources. Another difficulty is: even in a reservation-based envi- ronment, there is still resource contention. Different applications and services need to reserve from the same pool of resources, inevitably causing the reservations for some ap- plications/services to fail. In fact, the goals of (1) increasing the overall success rate of multi-resource reservations for different service requests, and (2) achieving the best end-to- end QoS for each service request, are in conflict with each other. In this paper, we propose a solution to the difficulties discussed above. We present a QoS and contention- aware multi-resource reservation algorithm for distributed and component-based services. The algorithm computes an end-to-end multi-resource reservation plan, which achieves the highest possible end-to-end QoS level under the constraint of current resource availability. In the meantime, the multi-resource reservation plan tends to cause low bottle-neck resource contention among all feasible resource reservation plans which lead to the same level of end-to-end QoS. The rest of the paper is organized as follows. In Section 2, we describe an enabling system architecture for multi-resource reservation, and a QoS-resource model on which our algorithm is based. In Section 3, we present the QoS and contention aware multi-resource reservation algorithm. In Section 4, we show the performance of this algorithm by simulation. Finally, we conclude this paper in Section 6. System Architecture and QoS-Resource Model 2.1 Distributed and Component-Based Services The distributed services studied in this paper are component-based. With distributed object programming techniques, a distributed service can be implemented as a set of collaborating service components. A service component is a functional unit participating in the service delivery. For example, in a distributed video streaming service with object tracking functionality, besides streaming a video to a client, the service can also track an object of interest in the video for the client. The client host will be able to playback the video, and there will be a rectangle around the object being tracked. In this service, the service components include a VideoSender service component running on a video server, an ObjectTracking service component running on a tracking server, and a VideoPlayer service component running on each client host. Each service component in a distributed service is able to achieve one or more levels of service qual- ity, depending on the amounts of resources reserved for this component. The service quality achieved by each individual service component finally leads to the end-to-end QoS provided for the client. 2.2 An Architecture for Multi-Resource Reserva- tion In order to deploy such a distributed and component-based service in a reservation-enabled environment, we introduce an enabling system architecture. The architecture is shown in Figure 1. It involves the following entities: Resource Brokers (RBs), QoSProxies, and service compo- nents. On each host in the environment, there is one or more RBs managing individual resources, one for each type of resource. A QoSProxy runs on each host, coordinating the reservation activities of local RBs. The multi-resource reservation algorithm will be executed by the QoSProxies on the hosts involved in a distributed service. A RB is responsible for the reservation and scheduling of a resource 1 . The basic functions of a RB includes: (1) reporting current resource availability, (2) making 1 For end-to-end network bandwidth, we look at it as one resource - a pipe from the sender to the receiver. To be compatible with RSVP, we assume that the network RB of the receiver is always responsible for initiating an end-to-end bandwidth reservation. component Service component Service QoSProxy component Service QoSProxy Translation function QoSProxy Figure 1. An Architecture for Multi-Resource Reservation and enforcing reservations, (3) releasing reservations, and (4) reporting possible reservation degradations. A QoSProxy is responsible for coordinating the reservation activities of individual RBs on the same host. Its basic functions include: (1) collecting resource availability information from individual RBs, (2) executing the multi-resource reservation algorithm and dispatching the resultant reservation plan to individual RBs, and (3) starting the service component on the same host when the multi-resource reservation is completed. A QoSProxy has to understand the relation between the QoS levels and the corresponding resource requirements of a service component, in order to compute a resource reservation plan. However, this relation is highly application- specific. For this reason, our architecture allows service developers to provide translation functions as plug-ins for the QoSProxies, as shown in Figure 1. Each translation function the relation between the multiple QoS levels and their resource requirements of a service component (the formal definition of a translation function will be given in Section 2.3). Therefore, a QoSProxy can call the translation function during the execution of the multi-resource reservation algorithm. 2.3 QoS-Resource Model To express the relation between a service component's QoS and its resource requirement, we adopt a QoS-resource model. This model was originally proposed in [8]. Each service component c is associated with both the input quality Q in and the output quality Q out . Q in and Q out are both represented as vectors of multiple QoS parameters. Their QoS parameter sets may not be identical. For simplicity, we assume that each parameter takes discrete values. There- fore, Q in and Q out of each service component are enumer- able. To compare two QoS vectors, they must have the same parameter set. For two QoS vectors Q a and Q b , Q a Q b holds if and only if for each QoS parameter, the corresponding value of Q a is not larger than that of Q b . For a service component c, the resource requirement to achieve a certain output quality Q out , given an input quality Q in , is computed by the translation function T c R). The resource requirement is formally represented as a resource requirement vector R. Therefore, given a pair (Q in ; Q out ), we have: The resource requirement vector M) is the required amount of the mth re- source. To compare two resource requirement vectors, they must have the same set of resources. For two resource requirement vectors R a and R b , R a R b holds if and only if for each type of resource, the corresponding value of R a is not larger than that of R b . For a distributed service, the participating service components organize themselves into a dependency graph. In general, the dependency graph is a Direct Acyclic Graph (DAG). Nodes of a dependency graph represent service components. Edges of a dependency graph represent the dependencies among the service components. Figure 2 shows an example of the dependency graph. An edge from service components c to c 0 indicates that the output of c is the input of c 0 ; and the Q out of c is equivalent to the Q in of c 0 . In addition, for a node with no in-coming edges (for example, c 1 in Figure 2), its Q in represents the original quality of the source data involved in this service (for example, a stored video clip or a live video source involved in a multimedia service); for a node with no out-going edges (for example, c 3 in Figure 2), its Q out represents the resultant end-to-end QoS. out in out in in Figure 2. Example Dependency Graph of a Distributed Service 3 Multi-Resource Reservation Algorithm After introducing the system architecture for multi-resource reservation and the QoS-resource model, we now present the QoS and contention-aware multi-resource reservation algorithm. Given a service request, the algorithm computes an end-to-end resource reservation plan for service components participating in this distributed service, so that the best end-to-end QoS can be delivered to the client, under the constraint of current end-to-end resource availability observed by the client. The goals of our algorithm involve both QoS-awareness and contention-awareness: QoS-awareness Each service component may accept multiple levels of Q in , and achieve multiple levels of Q out . The algorithm must compute a resource reservation plan by selecting appropriate levels of Q in and Q out for each service component, so that it will lead to the best possible end-to-end QoS for the client-side service component according to the dependency graph. Contention-awareness Resources may be shared by other services and applications on a competitive basis. Therefore, resource contention may exist during resource reservation. The degree of resource contention varies from time to time, from resource to resource. It may affect the overall success rate of resource reservations in the environment 2 . The algorithm must find a resource reservation plan among all possible reservation plans, such that it will reserve only the minimum amount of bottleneck resource(s). Therefore, if every multi-resource reservation is disciplined by this algo- rithm, the overall resource contention in the environment will be alleviated. In the following subsections, we first define QoS- Resource Graph (QRG) - a key data structure to study the multi-resource reservation problem. We then study the special case in which the dependency graph of a distributed service is a chain. Finally, we extend the algorithm to deal with the general case in which the dependency graph of a distributed service is a DAG. 3.1 QoS-Resource Graph We formally define the multi-resource reservation problem using a QoS-Resource Graph (QRG). For a distributed service, a QRG is generated for each service request at run- time, based on the dependency graph of the requested ser- vice. However, the definition of a QRG is different from that of a dependency graph. A node in a QRG represents a QoS level for the Q in or Q out of a service component c. An edge in a QRG from a node Q in to a node Q out represents the corresponding resource requirement vector computed by the translation function T c . However, such an edge exists if and only if the current resource availability (also represented as a vector) is no less than the resource requirement vector. Figure 3 shows an example QRG generated from the dependency graph in Figure 2. The dotted rectangles in the QRG represent the corresponding service components in the dependency graph. We assume that a multi-resource reservation is not successful, if at least one resource can not be reserved. For simplicity (without lowering the problem's complex- ity) , we further assume that the original quality of the source data associated with a service request has a single QoS level. We define the node representing this QoS level as the source node of a QRG (for example, Q a in Figure 3).000.50.60.50.6 Figure 3. Example QRG (weights of the edges are shown) For the client-side service component, whose Q out nodes represent the end-to-end QoS levels (for example, service component c 3 in Figure 3), we define its Q out nodes as the sink nodes of a QRG (for example, Q l and Qm in Figure 3). We also assume that the sink nodes (i.e. the end-to-end QoS levels) can be ranked in a linear order. The linear ranking can be determined by a client's preferences, and may be subjective. For example, when two end-to-end QoS levels are not comparable, the client requesting the distributed service can arbitrate that the QoS level with a smaller delay parameter value is better than the one with a larger value. We now define the weight of each edge in a QRG. The weight will reflect the degree of resource contention caused by the resource requirement represented by the edge. For an edge from a node Q in to a node Q out , let R [r req M ] be the corresponding resource requirement vector (computed by calling T c (Q in ; Q out )). On the other hand, let M ] be the current resource availability vector (collected by querying the RBs). We first define a contention index i to evaluate how 'competitive' it is to reserve r req i amount of resource r i , under the constraint of availability r avail i . In this paper, we choose a simple definition of as follows: r req r avail Intuitively, the larger the percentage of a resource one tries to reserve under the current availability constraint, the less likely the reservation will succeed 3 . Now, we can further 3 In fact, there are other definitions for which also exhibit this charac- define the weight for the edge as: r req r avail For an edge from a node Q out to a node Q in , it only represents their equivalence - the output quality of a service component is the input quality of its dependent service com- ponent. Therefore, the weight of such an edges is defined as zero (shown in Figure 3). 3.2 Algorithm: the Chain Case After defining the QRG, we are now ready to present the algorithm. We first consider the special case in which the dependency graph of a distributed service is a chain. In a QRG, an edge with non-zero weight exists if and only if the corresponding R req R avail , i.e. the reservation of resources according to R req is feasible. There- fore, we have the following observation: every path from the source node to one of the sink nodes represents a feasible end-to-end resource reservation plan - in other words, if we reserve resources according to the resource requirement vectors represented by the non-zero-weight edges on the path, the end-to-end QoS represented by the sink node will be guaranteed. Furthermore, the best achievable end-to-end QoS under the current resource availability is represented by the sink node which has the highest ranking among all 'reachable' sink nodes from the source node. For example, in Figure 3, if we assume that Q l ranks higher than Qm , then Q l is the best achievable end-to-end QoS level. However, there are multiple paths from Q a to Q l , i.e. there exist more than one feasible resource reservation plans to achieve Q l . To minimize resource contention, our algorithm will select a path such that the value of P is the smallest among these paths - P is defined as: (each edge e on path P ) e (4) By definition, it is easy to see that P represents the contention index of the bottleneck resource on the path (note that the bottleneck resource on each path may be different). To find a path from Q a to Q l whose bottleneck resource has the smallest contention index, our algorithm finds the shortest path from Q a to Q l , with operator '+' re-defined as 'max'. This is done by running Dijkstra's algorithm on the QRG. Figure 4 shows such a shortest path (shown by the thicker edges). The value inside each node is generated during the execution of Dijkstra's algorithm. The computation complexity of the reservation algorithm in the chain case is O(KQ 2 ). K is the number of service components in the dependency graph of a distributed teristics. Fortunately, it is easy for our algorithm to adopt a different (and more accurate) definition in the future. f Qm000.50.60.50.6 Figure 4. The Shortest Path from Q a to Q l (representing the end-to-end reservation plan computed by the algorithm) service. Q is the maximum number of Q out levels (nodes) among the service components. For example, in Figure 4, (which is the number of Q out nodes of Fortunately, K and Q usually have fairly small values in practice (for example, K, Q 10). Therefore, scalability is not a major concern for the multi-resource reservation algorithm. The number of Q out levels of a service component is set by the service developer in the translating function of the service component. Our experience shows that this number can be effectively controlled by limiting the number of possible values for each QoS parameter. 3.3 Algorithm: the DAG Case We now consider the more general case in which the dependency graph of a distributed service is a DAG. For a DAG dependency graph, we first extend the definition of a service component's QoS levels: For a service component with more than one out-going edge, its Q out will become the Q in of each service component on the other end of the out-going edge (as shown in Figure 5). We call the service component with more than one out-going edges a fan-out service component. For example, c 2 in Figure 5 is a fan-out service component. For a service component with more than one in-coming edge, its Q in is defined as the concatenation of the Q out of service components on the other end of the in-coming edges. We call the service component with more than one in-coming edges a fan-in service component. For example, c 5 in Figure 5 is a fan-in service component. An example QRG generated from this DAG dependency graph is shown in Figure 6. A feasible end-to-end reser- out in C3 out out in out in out q 33, . ] [q q . ] out 41, 43, q 33, . C1in Figure 5. Example DAG Dependency Graph of a Distributed Service vation plan is now represented by an embedded graph in the QRG such that: (1) for each service component, among the edges from a Q in node to a Q out node, there is one - and only one edge that belongs to the embedded graph; (2) within the embedded graph, the sink node (representing the end-to-end QoS level achieved by this reservation plan) is reachable from each node in the embedded graph. The goal of our reservation algorithm is to compute a feasible end-to- reservation plan, represented by an embedded graph G, such that (1) the sink node in the embedded graph has the highest QoS ranking; and (2) the value of G is the smallest both among all feasible end-to-end reservation plans. G is defined as follows: (each edge e in G) e (5)00.530.30.20.60000000.60.80.70.40.450.5Q a c cc c c Figure 6. Example QRG Based on a DAG Dependency Graph It can be shown that such a problem is NP-complete. Therefore, we focus on providing an efficient and effective heuristics to compute an end-to-end reservation plan that achieves the best end-to-end QoS, while trying to maintain a low value of G . Our heuristics is based on the following two-pass procedure on the QRG. Pass I on the QRG is similar to the reservation algorithm in the case of chain dependency graph. It also runs Dijkstra's algorithm to explore the 'shortest path' from the source node to the sink nodes of the QRG. As an example, Figure 7 shows the result of pass I on the QRG in Figure 6. Notice that when generating the value in a Q in node of a fan-in service component (for example, node Q r of c 5 ), we set the value as the maximum of the values in Q out nodes on the other end of the in-coming edges (for exam- ple, nodes Q n and Q p ). This is different from Dijkstra's algorithm. By our definition, Q r is the concatenation of Q n and Q p . Therefore, the resource contention to reach Q r is the maximum of resource contention to reach Q n and Q p , respectively. s c c c Figure 7. Running the Heuristics: Result of Pass I Pass II on the QRG proceeds in the reversed direction of pass I. Starting from the reachable sink node with the highest QoS ranking (for example, Q v ), we backtrack the edges toward the source node, according to the result of pass I. This is to determine the embedded graph that represents the resultant end-to-end reservation plan. However, the back-tracking may encounter the following problem: when arriving at a fan-out service component, the backtracked edges do not converge at the same Q out node. For example, in Figure 7, the backtracked edges (the thicker ones in the Figure) lead to different Q out nodes Q h and Q i . In our heuristics, we use the following method to resolve this non-convergence locally: For the service components dependent on the fan-out component (for example, c 3 and c 4 ), fix their Q out nodes that have been backtracked (for example, nodes Q n and select such a Q out node of the fan-out service component: it causes the lowest resource contention to reach the fixed Q out nodes of the dependent service com- ponents. For example, in Figure 7, Q i will be selected (in- stead of Q h ), because for Q i to reach Q n and Q p , the resource contention is 0.6, while for Q h to reach Q n and Q p , the resource contention is 0.7. e c c c c c Figure 8. Running the Heuristics: the Embedded Graph Representing the Resultant Reservation Plan By using this two-pass heuristics, an end-to-end reservation plan can be computed for the QRG in Figure 7. The embedded graph representing the reservation plan is shown in Figure 8. The limitation of this heuristics is: for a sink node of the QRG which is reachable in pass I, the heuristics may not necessarily find a feasible reservation plan in pass II to guarantee the end-to-end QoS level represented by the sink node. Furthermore, due to the local (instead of global) nature of the non-convergence resolution in pass II, the reservation plan computed by the heuristics may not incur the lowest bottleneck resource contention among all feasible reservation plans. 4 Simulation Results In this section, we evaluate the success rate of multi-resource reservations achieved by the proposed reservation algorithm. The results in this section are initial, and obtained by simulation. We simulate a simple scenario: there is a distributed service which involves three service components In our simulated environment, c 1 runs on one host. c 2 runs on another host. c 3 is the client-side service component, and runs on each client host. Resource contention exists on the hosts where c 1 and c 2 ex- ecute. In addition, we introduce background computation task on each client host, so that resource contention also exists between the execution of c 3 and the background task. For simplicity, we assume that each service component only requires one type of resource. The QoS levels and the corresponding resource requirements are shown in figure 9. Each value in the brackets denotes the required amount of resource for the corresponding (Q in ; Q out ) pair. We also assume that Q l has a higher QoS ranking than Qm . Notice that Figure 9 is not a QRG. The total amount of resource on the host where c 1 executes is 800 units. The total amount of resource on the host where c 2 executes is 400 units. The total amount of resource on each client host where c 3 executes is 1 unit. We assume that at the beginning of the simulation, all these resources are free. We also assume that for each client host, right before it makes a service request, a background computation task will begin to run with a 0.5 probability, and the amount of resource it consumes is uniformly distributed between 0.25 and 0.75 unit. [1.5] [2.5] [1.0] [0.3] [0.3] [x]: Amount of Resource Required Figure 9. QoS Levels and Resource Requirements of the Simulated Service Components In the first experiment, we simulate multi-resource reservations made for 16000 service requests spreading over 400 minutes. The duration (i.e. resource holding time) of each service session varies uniformly between 5 and 50 minutes. Service requests from different clients arrive at an average rate of 40 requests per minute. The success rates of multi-resource reservations are shown in Figure 10 - each point represents the success rate in a 5-minute interval. Here, we compare our algorithm with a random algorithm, which randomly selects a feasible multi-resource reservation plan represented by a path from Q a to Q l in Figure 9. During the 400-minute period, the overall success rate using our algorithm is 96.33%, while the random algorithm achieves an overall success rate of only 78.06%. In the second experiment, we simulate different average request arrival rates for the same distributed service. For10305070900 50 100 150 200 250 300 350 400 Success Rate of Multi-Resource Reservations (%, Time (minutes) 'Our Algorithm' 'Random Algorithm' Figure 10. Multi-Resource Reservation Success Rate over the 400-Minute Period each average arrival rate, we measure the overall success rate over a 400 minute period using our algorithm and using the random algorithm. Figure 11 shows the overall multi-resource reservation success rate under different service request arrival rates. The results show that our algorithm constantly achieves higher overall success rate than the random algorithm. 5 Related Work The problem of multi-resource reservation has been addressed from different angles. In [3] and [4], a resource co-allocation architecture and its mechanisms for alloca- tion, configuration, monitoring, and control are presented. It is suggested that resource co-allocation should be an integral part of the resource management architecture for Grid environments. In addition, an advance reservation mechanism is also proposed. One of our next steps is to extend our algorithm to accommodate advance reservation. In [6], the problem of apportioning multiple finite resources to satisfy the QoS needs of multiple applications along multiple QoS dimensions is studied. However, their solution is based on a static set of applications to be executed at the same time, and they do not consider the dynamic arrival and completion of applications. Therefore, their solution is not contention-aware. In the Darwin Project [1], a hierarchical service and resource brokerage architecture is Success Rate of Multi-Resource Reservations (%, overminutes) Arrival Rate of Multi-Resource Reservation Requests (requests/min) 'Our Algorithm' 'Random Algorithm' Figure 11. Multi-Resource Reservation Success Rate under Different Request Arrival Rates introduced. In order to compose value-added services, allocation of multiple resources is needed. The signaling protocol during multi-resource allocation is the Beagle signaling protocol [2]. However, this protocol is not contention-aware either. In our earlier work of Qualman system [7], different QoS-aware resource brokers are proposed. They are responsible for the reservation and enforcement of CPU, network bandwidth, and memory resources, respectively. However, there is no coordination among these resource brokers, and no algorithm is proposed to compute multi-resource reservation plans to guarantee end-to-end application level QoS. Finally, in [10], we study the multi-resource reservation problem only in the case of chain dependency graph. In this paper, we extend our solution to deal with the more general case of DAG dependency graph. 6 Conclusion In a reservation-based environment where every type of resource can be reserved, we need system support to compute end-to-end multi-resource reservation plans and to make corresponding reservations in an integrated and systematic manner. In this paper, we first propose a system architecture that enables such an integrated multi-resource reservation for distributed and component-based services. We then present a QoS and contention-aware multi-resource reservation algorithm that computes a reservation plan for each distributed service request, such that (1) it achieves the highest level of end-to-end QoS under the constraint of current resource availability, and (2) it causes the least bottleneck resource contention in the case of chain dependency while it tends to cause low bottleneck resource contention in the case of DAG dependency graph. Our future work includes the extension to support advance reservation, and the study of reservation fairness among service requests with highly heterogeneous resource requirements and service durations. --R Resource management for value-added customizable network service A signaling protocol for structured resource allocation. Resource co-allocation in computational grids A distributed resource management architecture that supports advance reservations and co-allocation CPU reservation and time constraints: efficient A scalable solution to the multi-resource QoS problem A disk scheduling framework for next generation operating systems. Multimedia service configuration and reservation in heterogeneous environ- ments RSVP: A new resource reservation protocol. --TR --CTR Ashish M. Mehta , Jay Smith , H. J. Siegel , Anthony A. Maciejewski , Arun Jayaseelan , Bin Ye, Dynamic resource allocation heuristics that manage tradeoff between makespan and robustness, The Journal of Supercomputing, v.42 n.1, p.33-58, October 2007 Jong-Kook Kim , Sameer Shivle , Howard Jay Siegel , Anthony A. Maciejewski , Tracy D. Braun , Myron Schneider , Sonja Tideman , Ramakrishna Chitta , Raheleh B. Dilmaghani , Rohit Joshi , Aditya Kaul , Ashish Sharma , Siddhartha Sripada , Praveen Vangari , Siva Sankar Yellampalli, Dynamically mapping tasks with priorities and multiple deadlines in a heterogeneous environment, Journal of Parallel and Distributed Computing, v.67 n.2, p.154-169, February, 2007
resource contention;distributed service;resource reservation