Stochastic methods based on VU-decomposition methods
Transcription
Stochastic methods based on VU-decomposition methods
Hindawi Publishing Corporation Mathematical Problems in Engineering Volume 2014, Article ID 894248, 5 pages http://dx.doi.org/10.1155/2014/894248 Research Article Stochastic Methods Based on VU-Decomposition Methods for Stochastic Convex Minimax Problems Yuan Lu,1 Wei Wang,2 Shuang Chen,3 and Ming Huang3 1 Normal College, Shenyang University, Shenyang 110044, China School of Mathematics, Liaoning Normal University, Dalian 116029, China 3 School of Mathematical Sciences, Dalian University of Technology, Dalian 116024, China 2 Correspondence should be addressed to Wei Wang; wei [email protected] Received 6 August 2014; Revised 29 November 2014; Accepted 29 November 2014; Published 4 December 2014 Academic Editor: Hamid R. Karimi Copyright © 2014 Yuan Lu et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. This paper applies sample average approximation (SAA) method based on VU-space decomposition theory to solve stochastic convex minimax problems. Under some moderate conditions, the SAA solution converges to its true counterpart with probability approaching one and convergence is exponentially fast with the increase of sample size. Based on the VU-theory, a superlinear convergent VU-algorithm frame is designed to solve the SAA problem. 1. Introduction In this paper, the following stochastic convex minimax problem (SCMP) is considered: minπ π (π₯) , (1) π (π₯) = max {πΈ [ππ (π₯, π)] : π = 0, . . . , π} , (2) π₯βπ where and the functions ππ (π₯, π) : π π β π , π = 0, . . . , π, are convex and πΆ2 , π : Ξ© β Ξ β π π is a random vector defined on probability space (Ξ©, Ξ₯, π); πΈ denotes the mathematical expectation with respect to the distribution of π. SCMP is a natural extension of deterministic convex minimax problems (CMP for short). The CMP has a number of important applications in operations research, engineering problems, and economic problems. While many practical problems only involve deterministic data, there are some important instances where problems data contains some uncertainties and consequently SCMP models are proposed to reflect the uncertainties. A blanket assumption is made that, for every π₯ β π π , πΈ[ππ (π₯, π)], π = 0, . . . , π, are well defined. Let π1 , . . . , ππ be a sampling of π. A well-known approach based on the sampling is the so-called SAA method, that is, using sample average value of ππ (π₯, π) to approximate its expected value because the classical law of large number for random functions ensures that the sample average value of ππ (π₯, π) converges with probability 1 to πΈ[ππ (π₯, π)] when the sampling is independent and identically distributed (idd for short). Specifically, we can write down the SAA of our SCMP (1) as follows: minπ πΜπ (π₯) , π₯βπ (3) where πΜπ (π₯) = max {πΜππ (π₯) : π = 0, . . . , π} , 1 π πΜππ (π₯) := β ππ (π₯, ππ ) . π π=1 (4) The problem (3) is called the SAA problem and (1) the true problem. The SAA method has been a hot topic of research in stochastic optimization. Pagnoncelli et al. [1] present the SAA method for chance constrained programming. Shapiro et al. [2] consider the stochastic generalized equation by using the SAA method. Xu [3] raises the SAA method for a class of stochastic variational inequality problems. Liu et al. [4] 2 Mathematical Problems in Engineering give the penalized SAA methods for stochastic mathematical programs with complementarity constraints. Chen et al. [5] discuss the SAA methods based on Newton method to the stochastic variational inequality problem with constraint conditions. Since the objective functions of the SAA problems in the references talking above are smooth, then they can be solved by using Newton method. More recently, new conceptual schemes have been developed, which are based on the VU-theory introduced in [6]; see else [7β11]. The idea is to decompose π π into two orthogonal subspaces V and U at a point π₯, where the nonsmoothness of π is concentrated essentially on V and the smoothness of π appears on the U subspace. More precisely, for a given π β ππ(π₯), where ππ(π₯) denotes the subdifferential of π at π₯ in the sense of convex analysis, then π π can be decomposed into direct sum of two orthogonal subspaces, that is, π π = U β V, where V = lin(ππ(π₯) β π), and U = Vβ₯ . As a result an algorithm frame can be designed for the SAA problem that makes a step in the V space, followed by a UNewton step in order to obtain superlinear convergence. A VU-space decomposition method for solving a constrained nonsmooth convex program is presented in [12]. A decomposition algorithm based on proximal bundle-type method with inexact data is presented for minimizing an unconstrained nonsmooth convex function in [13]. In this paper, the objective function in (1) is nonsmooth, but it has the structure which has the connection with VUspace decomposition. Based on the VU-theory, a superlinear convergent VU-algorithm frame is designed to solve the SAA problem. The rest of the paper is organized as follows. In the next section, the SCMP is transformed to the nonsmooth problem and the proof of the approximation solution set converges to the true solution set in the sense that Hausdorff distance is obtained. In Section 3, the VU-theory of the SAA problem is given. In the final section, the VU-decomposition algorithm frame of the SAA problem is designed. (d) The moment-generating function ππ (π‘) = πΈ[ππ‘π (π) ] of π (π) is finite-valued for all π‘ in a neighborhood of zero, where ππ (π‘) = πΈ[ππ‘(π(π ,π)βπΊ(π )) ] is the moment-generating function of the random variable π(π , π) β πΊ(π ). 2. Convergence Analysis of SAA Problem This shows that π(π₯Μπ, πβ ) < π, which implies π·(π₯Μπ, π β ) < π. In this section, we discuss the convergence of (3) to (1) as π increases. Specifically, we investigate the fact that the solution of the SAA problem (3) converges to its true counterpart as π β β. Firstly, we make the basic assumptions for SAA method. In the following, we give the basic assumptions for SAA method. Assumption 1. (a) Letting π be a set, for π = 1, . . . , π, the limits ππΉπ (π‘) := lim ππΉππ (π‘) πββ (5) exist for every π₯ β π. (b) For every π β π, the moment-generating function ππ (π‘) is finite-valued for all π‘ in a neighborhood of zero. (c) There exists a measurable function π : Ξ© β π + such that σ΅© σ΅¨ σ΅© σ΅¨σ΅¨ σ΅¨σ΅¨π (π σΈ , π) β π (π , π)σ΅¨σ΅¨σ΅¨ β€ π (π) σ΅©σ΅©σ΅©π σΈ β π σ΅©σ΅©σ΅© (6) σ΅© σ΅¨ σ΅© σ΅¨ for all π β Ξ© and all π σΈ , π β π. Theorem 2. Let πβ and ππ denote the solution sets of (1) and (3). Assuming that both πβ and ππ are nonempty, then, for any π > 0, one has π·(ππ, πβ ) < π, where π·(ππ, πβ ) = supπ₯βππ π(π₯, πβ ). Proof. For any points π₯Μ β ππ and π₯ β π π , we have Μ = max {πΜππ (π₯) Μ , π = 0, . . . , π} πΜπ (π₯) } {1 π Μ ππ ) , π = 0, . . . , π} = max { β ππ (π₯, π } { π=1 (7) } {1 π β€ max { β ππ (π₯, ππ ) , π = 0, . . . , π} . π } { π=1 From Assumption 1, we know that, for any π > 0, there exist π; if π > π, π = 0, . . . , π, then σ΅¨σ΅¨ σ΅¨σ΅¨ π σ΅¨ σ΅¨σ΅¨ 1 σ΅¨σ΅¨ βπ (π₯, ππ ) β πΈ [π (π₯, ππ )]σ΅¨σ΅¨σ΅¨ < π. σ΅¨σ΅¨ σ΅¨σ΅¨ π π σ΅¨σ΅¨ σ΅¨σ΅¨ π π=1 σ΅¨ σ΅¨ (8) By letting π > π, we obtain } {1 π Μ β€ max { βππ (π₯, ππ ) , π = 0, . . . , π} πΜπ (π₯) π } { π=1 β€ max {πΈ [ππ (π₯, ππ )] + π, π = 0, . . . , π} (9) = π (π₯) + π. We now move on to discuss the exponential rate of convergence of SAA problem (3) to the true problem (1) as sample increases. Theorem 3. Let π₯π be a solution to the SAA problem (3) and πβ is the solution set of the true problem (1). Suppose Assumption 1 holds. Then, for every π > 0, there exist positive constants π(π) and π(π), such that Prob {π (π₯π, πβ ) β₯ π} β€ π (π) expβππ(π) (10) for π sufficiently large. Proof. Let π > 0 be any small positive number. By Theorem 2 and σ΅¨σ΅¨σ΅¨ π σ΅¨σ΅¨σ΅¨ σ΅¨σ΅¨ 1 σ΅¨ σ΅¨σ΅¨ βππ (π₯, ππ ) β πΈ [ππ (π₯, ππ )]σ΅¨σ΅¨σ΅¨ < π, (11) σ΅¨σ΅¨ π σ΅¨σ΅¨ σ΅¨σ΅¨ π=1 σ΅¨σ΅¨ Mathematical Problems in Engineering 3 we have π(π₯π, π) < π. Therefore, by Assumption 1, we have Prob {π (π₯π, πβ ) β₯ π} σ΅¨σ΅¨ σ΅¨ σ΅¨ } {σ΅¨σ΅¨σ΅¨σ΅¨ 1 π π π σ΅¨σ΅¨ β€ Prob {σ΅¨σ΅¨σ΅¨ β ππ (π₯, π ) β πΈ [ππ (π₯, π )]σ΅¨σ΅¨σ΅¨ β₯ πΏ} σ΅¨σ΅¨ σ΅¨π σ΅¨ } {σ΅¨σ΅¨ π=1 Theorem 5. Suppose Assumption 4 holds. Then π π can be decomposition at π₯ : π π = U β V, where , V = lin {βπΜππ (π₯) β βπΜ0π (π₯)}0=πβπΌ(π₯) ΜΈ (12) β€ π (π) expβππ(π) . The proof is complete. 3. The VU-Theory of the SAA Problem In the following sections, we give the VU-theory, VUdecomposition algorithm frame, and convergence analysis of the SAA problem. The subdifferential of πΜπ at a point π₯ β π π can be computed in terms of the gradients of the function that are active at π₯. More precisely, ππΜπ (π₯) β©} = 0. U = {π β π π | β¨π, {βπΜππ (π₯) β βπΜ0π (π₯)}0=πβπΌ(π₯) ΜΈ (19) Proof. The proof can be directly obtained by using Assumption 4 and the definition of the spaces V and U. Given a subgradient π β ππΜπ with V-component πV = π π π, the U-Lagrangian of πΜπ, depending on πV , is defined by π dim U β π’ σ³¨σ³¨β πΏ π’ (π’; πV ) := min {πΜπ (π₯ + ππ’ + πV) β β¨πV , Vβ©V } . Vβπ dim V (20) The associated set of V-space minimizers is defined by { πΌ π = Conv {π β π π | π = β π ββππ (π₯, ππ ) : π π=1 πβπΌ(π₯) { π (π’; πV ) (13) } πΌ = (πΌπ )πβπΌ(π₯) β Ξ |πΌ(π₯)| } , } (14) is the set of active indices at π₯, and π (15) π=1 Let π₯ β π π be a solution of (3). By continuity of the structure functions, there exists a ball π΅π (π₯) β π π such that βπ₯ β π΅π (π₯) , πΌ (π₯) β πΌ (π₯) . (16) For convenience, we assume that the cardinality of πΌ(π₯) is π1 + 1 (π1 β€ π) and reorder the structure functions, so that πΌ(π₯) = {0, . . . , π1 }. From now on, we consider that βπ₯ β π΅π (π₯) , πΜπ (π₯) = πΜππ (π₯) , π β πΌ (π₯) . (17) The following assumption will be used in the rest of this paper. Assumption 4. The set {βπΜππ (π₯) β βπΜ0π (π₯)}0=πβπΌ(π₯) ΜΈ is linearly independent. (i) the nonlinear system, with variable V and the parameter π’, πΜππ (π₯ + ππ’ + πV) β πΜ0π (π₯ + ππ’ + πV) = 0, Ξ π = {πΌ β π π | πΌπ β₯ 0, βπΌπ = 1} . (21) Theorem 6. Suppose Assumption 4 holds. Let π(π’) = π₯ + π’ β V(π’) be a trajectory leading to π₯ and let π» := β2 πΏ π’ (0, 0). Then for all π’ sufficiently small the following hold: where πΌ (π₯) = {π β πΌ | πΜπ (π₯) = πΜππ (π₯)} := {V : πΏ π’ (π’; πV ) = πΜπ (π₯ + ππ’ + πV) β β¨πV , Vβ©V } . (18) 0 =ΜΈ π β πΌ (π₯) (22) has a unique solution V = V(π’) and V : π dim U β π dim V is a πΆ2 function; (ii) π(π’) is a πΆ2 -function with π½π(π’) = π + ππ½V(π’); (iii) πΏ π’ (π’; 0) = πΜππ(π₯ + π’ β V(π’)) = πΜπ(π₯ + π’ β V(π’)) = πΜπ(π₯) + (1/2)π’π π»π’ + π(|π’|2 ); π (iv) βπΏ π’ (π’; 0) = π»π’ + π(|π’|); (v) πΜπ(π(π’)) = πΜπ(π(π’)), π β πΌ(π₯). π Proof. Item (i) follows from the assumption that ππ are πΆ2 and applying a Second-Order Implicit Function Theorem (see [14], Theorem 2.1). Since V(π’) is πΆ2 , π(π’) is πΆ2 and the Jacobians π½V(π’) exist and are continuous. Differentiating the primal track with respect to π’, we obtain the expression of π½π(π’) and item (ii) follows. (iii) By the definition of πΏ π’ (π’; πV ) and π(π’; πV ), we have πΏ π’ (π’; 0) = πΜππ (π₯ + π’ β V (π’)) = πΜπ (π₯ + π’ β V (π’)) . (23) 4 Mathematical Problems in Engineering According to the second-order expansion of πΏ π’ , we obtain πΏ π’ (π’; 0) = πΏ π’ (0; 0) where β πΌπ (π’) βπΜπ π (π₯(π) ) = πΜ(π) β ππΜπ (π₯(π) ) π πβπΌ(π₯) 1 + β¨βπΏ π’ (0; 0) , π’β© + π’πβ2 πΏ π’ (0; 0) π’ + π (|π’|) . 2 (24) Since πΏ π’ (0; 0) = πΜππ(π₯), π β πΌ(π₯), βπΏ π’ (0; 0) = 0, and π» = β πΏ π’ (0; 0), 2 1 πΏ π’ (π’; 0) = πΜππ (π₯) + π’π π»π’ + π (|π’|2 ) . 2 Similar to (iii), we get (iv): (25) βπΏ π’ (π’; 0) = βπΏ π’ (0; 0) + β¨π’π β2 πΏ π’ (0; 0) , π’β© + π (|π’|2 ) = π» + π (|π’|) . (26) The conclusion of (v) can be obtained in terms of (i) and the definition of π(π’). 4. Algorithm and Convergence Analysis Supposing 0 β ππΜπ(π₯), we give an algorithm frame which can solve (3). This algorithm makes a step in the V-subspace, followed by a U-Newton step in order to obtain superlinear convergence rate. π (π) β0 = is such that π πΜ(π) = 0. Compute π₯(π+1) = π₯Μ(π) + πΏU (π) (π) (π) π₯ + πΏU β πΏV . Step 6. Update: set π = π + 1 and return to Step 1. Theorem 8. Suppose the starting point π₯(0) is close to π₯ enough Μ and 0 β riππ(π₯), β2 πΏ π’ (0; 0) β» 0. Then the iteration points (π) β {π₯ }π=1 generated by Algorithm 7 converge and satisfy σ΅© σ΅©σ΅© (π+1) σ΅© σ΅© σ΅©σ΅©π₯ β π₯σ΅©σ΅©σ΅©σ΅© = π (σ΅©σ΅©σ΅©σ΅©π₯(π) β π₯σ΅©σ΅©σ΅©σ΅©) . σ΅© Step 0. Initialization: given π > 0, choose a starting point π₯(0) close to π₯ enough and a subgradient πΜ(0) β ππΜπ(π₯(0) ) and set π = 0. Step 1. Stop if σ΅©σ΅© (π) σ΅©σ΅© σ΅©σ΅©πΜ σ΅©σ΅© β€ π. σ΅© σ΅© (27) σ΅© σ΅© σ΅© σ΅©σ΅© (π+1) σ΅©σ΅©(π₯ β π₯)V σ΅©σ΅©σ΅©σ΅© = σ΅©σ΅©σ΅©σ΅©(π₯(π) β π₯)V σ΅©σ΅©σ΅©σ΅© σ΅© σ΅© σ΅© σ΅© σ΅© = π σ΅©σ΅©σ΅©σ΅©(π₯(π) β π₯)U σ΅©σ΅©σ΅©σ΅© = π σ΅©σ΅©σ΅©σ΅©(π₯(π) β π₯)σ΅©σ΅©σ΅©σ΅© . Step 3. Construct VU-decomposition at π₯; that is, π π = V β U. Compute π β2 πΏ π’ (0; 0) = π π (0) π, (28) π (0) = β πΌπ β2 πΜππ (π₯) . (29) (33) Since β2 πΏ π’ (0; 0) exists and βπΏ π’ (0; 0) = 0, we have from the definition of U-Hessian matrix that π βπΏ π’ (π’(π) ; 0) = π πΜ(π) (34) (π) By virtue of (30), we have β2 πΏ π’ (0; 0)(π’(π) + πΏU ) = π(βπ’(π) βU ). 2 It follows from the hypothesis β πΏ π’ (0; 0) β» 0 that β2 πΏ π’ (0; 0) (π) is invertible and hence βπ’(π) + πΏU β = π(βπ’(π) βU ). In consequence, one has (π₯(π+1) β π₯)U = (π₯(π+1) β π₯(π) )U + (π₯(π) β π₯(π) )U + (π₯(π) β π₯)U Step 2. Find the active index set πΌ(π₯). (32) (π) Proof. Let π’(π) = (π₯(π) β π₯)U and V(π) = (π₯(π) β π₯)V + πΏV . It follows from Theorem 6(i) that σ΅© σ΅© = 0 + β2 πΏ π’ (0; 0) π’(π) + π (σ΅©σ΅©σ΅©σ΅©π’(π) σ΅©σ΅©σ΅©σ΅©U ) . Algorithm 7 (algorithm frame). (31) σ΅© σ΅© (π) = π’(π) + πΏU = π (σ΅©σ΅©σ΅©σ΅©π’(π) σ΅©σ΅©σ΅©σ΅©U ) (35) σ΅© σ΅© = π (σ΅©σ΅©σ΅©σ΅©π₯(π) β π₯σ΅©σ΅©σ΅©σ΅©) . The proof is completed by combining (33) and (35). where πβπΌ(π₯) Conflict of Interests (π) Step 4. Perform V-step. Compute πΏV which denotes V(π’) in (π) (π) (π) (22) and set π₯ = π₯ + 0 β πΏV . (π) from the system Step 5. Perform U-step. Compute πΏU π π π π (0) ππΏU + π πΜ(π) = 0, (30) The authors declare that there is no conflict of interests regarding the publication of this paper. Acknowledgments The research is supported by the National Natural Science Foundation of China under Project nos. 11301347, 11171138, and 11171049 and General Project of the Education Department of Liaoning Province no. L201242. Mathematical Problems in Engineering References [1] B. K. Pagnoncelli, S. Ahmed, and A. Shapiro, βSample average approximation method for chance constrained programming: theory and applications,β Journal of Optimization Theory and Applications, vol. 142, no. 2, pp. 399β416, 2009. [2] A. Shapiro, D. Dentcheva, and A. Ruszczynski, Lecture on Stochastic Programming: Modelling and Theory, SIAM, Philadelphia, Pa, USA, 2009. [3] H. Xu, βSample average approximation methods for a class of stochastic variational inequality problems,β Asia-Pacific Journal of Operational Research, vol. 27, no. 1, pp. 103β119, 2010. [4] Y. Liu, H. Xu, and J. J. Ye, βPenalized sample average approximation methods for stochastic mathematical programs with complementarity constraints,β Mathematics of Operations Research, vol. 36, no. 4, pp. 670β694, 2011. [5] S. Chen, L.-P. Pang, F.-F. Guo, and Z.-Q. Xia, βStochastic methods based on Newton method to the stochastic variational inequality problem with constraint conditions,β Mathematical and Computer Modelling, vol. 55, no. 3-4, pp. 779β784, 2012. [6] C. Lemarechal, F. Oustry, and C. Sagastizabal, βThe ULagrangian of a convex function,β Transactions of the American Mathematical Society, vol. 352, no. 2, pp. 711β729, 2000. [7] R. Mifflin and C. Sagastiz´abal, βVU-decomposition derivatives for convex max-functions,β in Ill-Posed Variational Problems and Regularization Techniques, M. Th´era and R. Tichatschke, Eds., vol. 477 of Lecture Notes in Economics and Mathematical Systems, pp. 167β186, Springer, Berlin, Germany, 1999. [8] C. Lemar´echal and C. Sagastiz´abel, βMore than first-order developments of convex functions: primal-dual relations,β Journal of Convex Analysis, vol. 3, no. 2, pp. 255β268, 1996. [9] R. Mifflin and C. Sagastizabal, βOn VU-theory for functions with primal-dual gradient strcture,β SIAM Journal on Optimization, vol. 11, no. 2, pp. 547β571, 2000. [10] R. Mifflin and C. Sagastiz´abal, βFunctions with primal-dual gradient structure and U-Hessians,β in Nonlinear Optimization and Related Topics, G. Pillo and F. Giannessi, Eds., vol. 36 of Applied Optimization, pp. 219β233, Kluwer Academic, 2000. [11] R. Mifflin and C. Sagastiz´abal, βPrimal-dual gradient structured functions: second-order results; links to EPI-derivatives and partly smooth functions,β SIAM Journal on Optimization, vol. 13, no. 4, pp. 1174β1194, 2003. [12] Y. Lu, L. P. Pang, F. F. Guo, and Z. Q. Xia, βA superlinear space decomposition algorithm for constrained nonsmooth convex program,β Journal of Computational and Applied Mathematics, vol. 234, no. 1, pp. 224β232, 2010. [13] Y. Lu, L.-P. Pang, J. Shen, and X.-J. Liang, βA decomposition algorithm for convex nondifferentiable minimization with errors,β Journal of Applied Mathematics, vol. 2012, Article ID 215160, 15 pages, 2012. [14] S. Lang, Real and Functional Analysis, vol. 142 of Graduate Texts in Mathematics, Springer, New York, NY, USA, 3rd edition, 1993. 5 Advances in Operations Research Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Advances in Decision Sciences Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Journal of Applied Mathematics Algebra Hindawi Publishing Corporation http://www.hindawi.com Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Journal of Probability and Statistics Volume 2014 The Scientific World Journal Hindawi Publishing Corporation http://www.hindawi.com Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 International Journal of Differential Equations Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Volume 2014 Submit your manuscripts at http://www.hindawi.com International Journal of Advances in Combinatorics Hindawi Publishing Corporation http://www.hindawi.com Mathematical Physics Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Journal of Complex Analysis Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 International Journal of Mathematics and Mathematical Sciences Mathematical Problems in Engineering Journal of Mathematics Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Discrete Mathematics Journal of Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Discrete Dynamics in Nature and Society Journal of Function Spaces Hindawi Publishing Corporation http://www.hindawi.com Abstract and Applied Analysis Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 International Journal of Journal of Stochastic Analysis Optimization Hindawi Publishing Corporation http://www.hindawi.com Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Volume 2014