- Home
- Documents
*AN INEXACT ACCELERATED PROXIMAL GRADIENT METHOD AN INEXACT ACCELERATED PROXIMAL GRADIENT METHOD...*

prev

next

out of 23

View

0Download

0

Embed Size (px)

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.

SIAM J. OPTIM. c© 2012 Society for Industrial and Applied Mathematics Vol. 22, No. 3, pp. 1042–1064

AN INEXACT ACCELERATED PROXIMAL GRADIENT METHOD FOR LARGE SCALE LINEARLY CONSTRAINED CONVEX SDP∗

KAIFENG JIANG† , DEFENG SUN‡ , AND KIM-CHUAN TOH§

Abstract. The accelerated proximal gradient (APG) method, first proposed by Nesterov for minimizing smooth convex functions, later extended by Beck and Teboulle to composite convex objective functions, and studied in a unifying manner by Tseng, has proven to be highly efficient in solving some classes of large scale structured convex optimization (possibly nonsmooth) problems, including nuclear norm minimization problems in matrix completion and l1 minimization problems in compressed sensing. The method has superior worst-case iteration complexity over the classical projected gradient method and usually has good practical performance on problems with appropriate structures. In this paper, we extend the APG method to the inexact setting, where the subproblem in each iteration is solved only approximately, and show that it enjoys the same worst-case iteration complexity as the exact counterpart if the subproblems are progressively solved to sufficient accuracy. We apply our inexact APG method to solve large scale convex quadratic semidefinite programming (QSDP) problems of the form min{ 1

2 〈x, Q(x)〉 + 〈c, x〉 | A(x) = b, x � 0}, where Q,A are given

linear maps and b, c are given data. The subproblem in each iteration is solved by a semismooth Newton-CG (SSNCG) method with warm-start using the iterate from the previous iteration. Our APG-SSNCG method is demonstrated to be efficient for QSDP problems whose positive semidefinite linear maps Q are highly ill-conditioned or rank deficient.

Key words. inexact accelerated proximal gradient, convex quadratic SDP, semismooth Newton- CG, structured convex optimization

AMS subject classifications. 90C06, 90C22, 90C25, 65F10

DOI. 10.1137/110847081

1. Introduction. Let Sn be the space of n×n real symmetric matrices endowed with the standard trace inner product 〈·, ·〉 and Frobenius norm ‖·‖, and let Sn+ (Sn++) be the set of positive semidefinite (definite) matrices in Sn. We consider the following linearly constrained convex semidefinite programming (SDP) problem:

(P ) min{f(x) : A(x) = b, x � 0, x ∈ Sn},

where f is a smooth convex function on Sn+, A : Sn → Rm is a linear map, b ∈ Rm, and x � 0 means that x ∈ Sn+. Let A∗ be the adjoint of A. The dual problem associated with (P ) is given by

(D) max{f(x)− 〈∇f(x), x〉 + 〈b, p〉 : ∇f(x)−A∗p− z = 0, p ∈ Rm, z � 0, x � 0}.

We assume that the linear map A is surjective and that strong duality holds for (P ) and (D). Let x∗ be an optimal solution of (P ) and (x∗, p∗, z∗) be an optimal solution

∗Received by the editors September 6, 2011; accepted for publication (in revised form) June 15, 2012; published electronically September 11, 2012.

http://www.siam.org/journals/siopt/22-3/84708.html †Department of Mathematics, National University of Singapore, 119076 Singapore

(kaifengjiang@nus.edu.sg). ‡Department of Mathematics and Risk Management Institute, National University of Singapore,

119076 Singapore (matsundf@nus.edu.sg). §Department of Mathematics, National University of Singapore, 119076 Singapore, and Singapore-

MIT Alliance, Singapore 117576 (mattohkc@nus.edu.sg).

1042

D ow

nl oa

de d

09 /0

7/ 14

to 1

37 .1

32 .1

23 .6

9. R

ed is

tr ib

ut io

n su

bj ec

t t o

SI A

M li

ce ns

e or

c op

yr ig

ht ; s

ee h

ttp ://

w w

w .s

ia m

.o rg

/jo ur

na ls

/o js

a. ph

p

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.

AN INEXACT APG FOR LINEARLY CONSTRAINED CONVEX SDP 1043

of (D). Then, as a consequence of strong duality, they must satisfy the following KKT conditions:

A(x) = b, ∇f(x)−A∗p− z = 0, 〈x, z〉 = 0, x, z � 0. The problem (P ) contains the following important special case of convex quadratic

semidefinite programming (QSDP):

(1) min {1 2 〈x, Q(x)〉+ 〈c, x〉 : A(x) = b, x � 0

} ,

where Q : Sn → Sn is a given self-adjoint positive semidefinite linear operator and c ∈ Sn. Note that the Lagrangian dual problem of (1) is given by

(2) max { − 1

2 〈x, Q(x)〉 + 〈b, p〉 : A∗(p)−Q(x) + z = c, z � 0

} .

A typical example of QSDP is the nearest correlation matrix problem, where given a symmetric matrix u ∈ Sn and a linear map L : Sn → Rn×n, one intends to solve

(3) min {1 2 ‖L(x− u)‖2 : diag(x) = e, x � 0

} ,

where e ∈ Rn is the vector of all ones and u ∈ Sn is given. If we let Q = L∗L and c = −L∗L(u) in (3), then we get the QSDP problem (1). A well-studied special case of (3) is the W -weighted nearest correlation matrix problem, where L = W 1/2�W 1/2 for a given W ∈ Sn++ and Q = W � W . Note that for U ∈ Rn×r, V ∈ Rn×s, U � V : Rr×s → Sn is the symmetrized Kronecker product linear map defined by U � V (M) = (UMV T + VMTUT )/2.

There are several methods available for solving this special case of (3), which in- clude the alternating projection method [4], the quasi-Newton method [6], the inexact semismooth Newton-CG (SSNCG) method [10], and the inexact interior-point method [13]. All these methods, excluding the inexact interior-point method, rely critically on the fact that the projection of a given matrix x ∈ Sn onto Sn+ has an analyti- cal formula with respect to the norm ‖W 1/2(·)W 1/2‖. However, all above-mentioned techniques cannot be extended to efficiently solve theH-weighted case [4] of (3), where L(x) = H ◦x for some H ∈ Sn with nonnegative entries and Q(x) = (H ◦H)◦x, with “◦” denoting the Hadamard product of two matrices defined by (A ◦ B)ij = AijBij . The aforementioned methods are not well suited for the H-weighted case of (3) be- cause there is no explicitly computable formula for the following problem:

(4) min {1 2 ‖H ◦ (x− u)‖2 : x � 0

} ,

where u ∈ Sn is a given matrix. To tackle the H-weighted case of (3), Toh [12] proposed an inexact interior-point method for a general convex QSDP including the H-weighted nearest correlation matrix problem. Recently, Qi and Sun [11] introduced an augmented Lagrangian dual method for solving the H-weighted version of (3), where the inner subproblem was solved by an SSNCG method. The augmented La- grangian dual method avoids solving (4) directly, and it can be much faster than the inexact interior-point method [12]. However, if the weight matrix H is very sparse or ill-conditioned, the conjugate gradient (CG) method would have great difficulty in solving the linear system of equations in the semismooth Newton method, and the augmented Lagrangian method would not be efficient or would even fail.

D ow

nl oa

de d

09 /0

7/ 14

to 1

37 .1

32 .1

23 .6

9. R

ed is

tr ib

ut io

n su

bj ec

t t o

SI A

M li

ce ns

e or

c op

yr ig

ht ; s

ee h

ttp ://

w w

w .s

ia m

.o rg

/jo ur

na ls

/o js

a. ph

p

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.

1044 KAIFENG JIANG, DEFENG SUN, AND KIM-CHUAN TOH

Another example of QSDP comes from the civil engineering problem of estimating a positive semidefinite stiffness matrix for a stable elastic structure from r measure- ments of its displacements {u1, . . . , ur} ⊂ Rn in response to a set of static loads {f1, . . . , fr} ⊂ Rn [15]. In this application, one is interested in the QSDP problem min{‖f − L(x)‖2 | x ∈ Sn+}, where L : Sn → Rn×r is defined by L(x) = xu, and f = [f1, . . . , fr], u = [u1, . . . , ur]. In this case, the corresponding map Q = L∗L is given by Q(x) = (xB +Bx)/2 with B = uuT .

The main purpose of this paper is to design an efficient algorithm to solve the problem (P ). The algorithm we propose here is based on the accelerated proximal gradient (APG) method of Beck and Teboulle [1] (the method is called FISTA in [1]), where in the kth iteration with iterate x̄k, a subproblem of the following form must be solved:

min { 〈∇f(x̄k), x− x̄k〉+ 1

2 〈x− x̄k, Hk(x − x̄k)〉 : A(x) = b, x � 0

} ,(5)

where Hk : Sn → Sn is a given self-adjoint positive definite linear operator. In FISTA [1], Hk is restricted to LI, where I denotes the identity map and L is a Lipschitz constant for ∇f . More significantly, for FISTA in [1], the subproblem (5) must be solved exactly to generate the next iterate xk+1. In this paper, we design an inexact APG method which overcomes the two limitations just mentioned. Specifically, in our inexact algorithm, the subproblem (5) is solved only approximately and Hk is not restricted to be a scalar multiple of I. In addition, we are able to show that if the subproblem (5) is progressively solved with sufficient accuracy, then the number of iterations needed to achieve ε-optimality (in terms of the function value) is also proportional to 1/

√ ε,