Abstract
When using interior methods for solving semidefinite programming (SDP), one needs to solve a system of linear equations at each iteration. For problems of large size, solving the system of linear equations can be very expensive. In this paper, based on a semismooth equation reformulation using Fischer's function, we propose a filter method with trust region for solving large-scale SDP problems. At each iteration we perform a number of conjugate gradient iterations, but do not need to solve a system of linear equations. Under mild assumptions, the convergence of this algorithm is established. Numerical examples are given to illustrate the convergence results obtained.
1. Introduction
Semidefinite programming (SDP) is convex programming over positive semidefinite matrices. For early application, SDP has been widely used in control theory and combinatorial optimization (see, e.g., [1โ3]). Since some algorithms for linear optimization can be extended to many general SDP problems, that aroused much interest in SDP. In the past decade, many algorithms have been proposed for solving SDP, including interior-point methods (IPMs) [4โ7], augmented methods [8โ10], new Newton-type methods [11], modified barrier methods [12], and regularization approaches [13].
For small and medium sized SDP problems, IPMs are generally efficient. But for large-scale SDP problems, IPMs become very slow. In order to improve this shortcoming, [9, 14] proposed inexact IPMs using an iterative solver to compute a search direction at each iteration. More recently, [13] applied regularization approaches to solve SDP problems. All of these methods are first-order based on a gradient, or inexact second-order based on an approximation of Hessian matrix methods [15].
In this paper, we will extend filter-trust-region methods for solving linear (or nonlinear) programming [16] to large-scale SDP problems and use Lipschitz continuity. Furthermore, the accuracy of this method is controlled by a forcing parameter. It is shown that, under mild assumptions, this algorithm is convergent.
The paper is organized as follows. Some preliminaries are introduced in Section 2. In Section 3, we propose a filter-trust-region method for solving SDP problems, and we study the convergence of this method in Section 4. In Section 5, some numerical examples are presented to demonstrate the convergence results obtained in this paper. Finally, we give some conclusions in Section 6.
In this paper, we use the following common notation for SDP problems: and denote the space of real symmetric matrices and the space of vectors with dimensions, respectively; denotes that is positive semidefinite (positive definite), and is used to indicate that is negative semidefinite (negative definite). A superscript represents transposes of matrices or vectors. For , the standard scalar product on the space of is defined by If and , we denote that is the Frobenius norm of , that is, and is the 2-norm of , that is, , respectively. Let be a matrix. Then we denote by a vector made of columns of stacked one by one, and the operator is the inverse of , that is, . We also denote that is identity matrix.
2. Preliminaries
We consider a SDP problem of the form where , , , and are given dates; is a linear map from to given by The dual to the problem (2.1) is given by where is an adjoint operator of given by Obviously, and are the primal and dual variables, respectively.
It is easily verified that the SDP problem (2.1) is convex. When (2.1) and (2.3) have strictly feasible points, then strong duality holds, see [5, 12]. In this case, a point is optimal for SDP problems (2.1) and (2.3) if and only if In the sense that solves SDP problems (2.1) and (2.3) if and only if solves (2.5) when both SDP problems (2.1) and (2.3) have strictly feasible points.
We now introduce some lemmas which will be used in the sequel.
Lemma 2.1 (see [17]). Let and let . Then if and only if .
For , we define a mapping given by which is attributed by Fischer to Burmeister (see [18, 19]). This function is nondifferentiable and has a basic property.
Lemma 2.2 (see [20, Lemma 6.1]). Let be the Fischer-Burmeister function defined in (2.6). Then
In addition, for and , we define a mapping by which is differentiable and has following results.
Lemma 2.3 (see [11, Propositionโโ2.3]). Let be any positive number and let be defined by (2.8). Then
Lemma 2.4. Let be any positive number, and let be defined by (2.8). If , we would have
Proof. The proof can be obtained from Lemmas 2.2 and 2.3.
Lemma 2.5 (see [20, pages 170โ171]). For any , define the linear operator by Then is strictly monotone and so has an inverse .
Lemma 2.6 (see [21, Lemmaโโ2]). Let , and let be defined by (2.8). For any , we have that is Frรฉchet-differentiable and where .
Lemma 2.7 (see [22, Corollary 2.7]). Let be a map from to . If is locally Lipschitzian on , then is almost everywhere Frรฉchet-differentiable on .
3. The Algorithm
In this section, we will present a filter-trust-region method for solving SDP problems (2.1) and (2.3). Firstly, for a parameter , we construct a function: where .
According to Lemmas 2.1, 2.3 and 2.4, the following theorem is obvious.
Theorem 3.1. Let and let be defined by (3.1). If SDP problems (2.1) and (2.3) have strictly feasible points, then
In what follows, we will study properties of the function . For simplicity, in the remaining sections of this paper, we denote , and .
Theorem 3.2. Let be defined by (3.1). For any and , then is Frรฉchet-differentiable and where and .
Proof. For any , since and are linear functions and continuous differentiable, it follows that they are also locally Lipschitz continuous. Then, from Lemma 2.7, and are Frรฉchet-differentiable. Furthermore, is Frรฉchet-differentiable from Lemma 2.6. Thus, is Frรฉchet-differentiable and has the form of (3.3). We complete the proof.
We endow the variable with the following norm: In addition, we set where We also define the function and the vector with the following norm:
Now, for any , we define the merit function by
Lemma 3.3. For any and , if and are nonsingular, then is locally Lipschitz continuous and twice Frรฉchet-differentiable at every .
Proof. For any , since is convex and continuously differentiable, it follows that is also locally Lipschitz continuous.
In addition, for any , from [20, pages 173โ175], is twice Frรฉchet-differentiable. Furthermore, , , and are continuous at every when , which, together with Lemma 2.7, is twice Frรฉchet-differentiable. The proof is completed.
Lemma 3.4. Let and be defined by (3.1) and (3.8), respectively. For any , we have
Proof. The proof can be immediately obtained from the definition of and .
We follow the classical method for solving , which consists some norm of the residual. For any , we consider where . Thus, for any , we want to find a minimizer of . Furthermore, if , then is also a solution of .
In order to state our method for solving (3.10), we consider using a filter mechanism to accept a new point. Just as [16, pages 19โ20], the notation of filter is based on that of dominance.
Definition 3.5. For any and any , a point dominates a point if and only if
Thus, if iterate dominates iterate , the latter is of no real interest to us since is at least as good as for each of the components of . All we need to do is remember iterates that are no dominated by other iterates by using a structure called a filter.
Definition 3.6. Let be a set of 4-tuples of the following form: We define as a filter if and belong to , when , then
Definition 3.7. A new point is acceptable for the filter if and only if where is a small constant.
Now, we formally present our trust region algorithm by using filter techniques.
Algorithm 3.8. The Filter-Trust-Region Algorithm
Step 0. Choose an initial point , , and . The constants , , , , , , and are also given and satisfy
Compute , set , and only in the filter .
Step 1. If , stop.
Step 2. Compute by solving the following problem:
where
If , stop.
Otherwise, computer the trial point .
Step 3. Compute and define the following ratio:
Step 4. If , set .
If but satisfies (3.14), then add to the filter and remove all points from dominated by . At the same time, set .
Else, set .
Step 5. Update by choosing
and update trust-region radius by choosing
Step 6. Set and go to Step 1.
Remark 3.9. Algorithm 3.8 can be started any . In fact, in order to increase the convergent speed greatly, we always choose . In addition, in this algorithm, we fix at first, then search for to update . At last we update and repeat.
The following lemma is a generalized case of Proposition 3.1 in [23].
Lemma 3.10. Algorithm 3.8 is well defined, that is, the inner iteration (Step 2) terminates finitely.
For the purpose of our analysis, in the sequence of points generated by Algorithm 3.8, we denote , and . It is clear that, .
Remark 3.11. Lemma 3.3 implies that there exists a constant such that for all and . The second of above inequalities ensures that the constant can also be chosen such that
4. Convergence of Analysis
In this section, we present a proof of global convergence of Algorithm 3.8. First, we make the following assumptions.
Some lemmas will be presented to be used in the subsequent analysis.โ(S1), where is a solution of (3.16). (S2)โ The iterations generated by Algorithm 3.8 remain in a close, bounded domain.
Lemma 4.1 (see [24]). Let assumptions (S1) and (S2) hold. If there exists such that for all ; then there exists such that .
Lemma 4.2. Let be the infinite sequence generated by the Algorithm 3.8. Then
Proof. Since , from Steps 4 and 5 of Algorithm 3.8, and . Therefore, . Moreover, for , which completes the proof.
Theorem 4.3. Let , assumptions (S1) and (S2) hold. Then there exists such that
Proof. Suppose that for all . Then there exists such that
From Lemma 4.1, there exists such that
On the other hand, , let be the last successful iteration, then are unsuccessful iterations. From Steps 4 and 5 of Algorithm 3.8, , for sufficiently large , we have
which contradicts (4.5). The proof is completed.
We now consider what happens if the set is infinite in the course of Algorithm 3.8.
Theorem 4.4. Suppose that , assumptions (S1) and (S2) hold. For any and , if and are nonsingular, then each accumulation point of the infinite sequences generated by Algorithm 3.8 is a stationary point of .
Proof. The proof is by contradiction. Suppose that is an infinite sequence generated by Algorithm 3.8, and any accumulation point of is not a stationary point of . Suppose furthermore that and are the accumulation points of and , respectively. Since is not a stationary point of , then
and there exists such that
For some , let be a neighborhood of . From (4.8), there exists such that
where .
For , because
we obtain that
From (4.10), we know that is monotone decreasing and bounded below, which implies that for . Thus,
As a result, we have
By the update rule of , there exists an infinite subsequence , and we have that
which contradicts . This completes the proof.
In what follows, we investigate the case where the number of iterations added to the filter in the course of Algorithm 3.8 is infinite.
Theorem 4.5. Suppose that but , SDP problems (2.1) and (2.3) have strictly feasible points. Suppose furthermore that assumptions (S1) and (S2) hold. For any and , if and are nonsingular, then
Proof. First let be the sequence generated by Algorithm 3.8. From Lemma 4.2, we have which, together with assumption (S2), the desired result follows from [16, Lemmaโโ3.1].
5. Numerical Experiments
In this section, we describe the results of some numerical experiments with the Algorithm 3.8 for the random sparse SDP considered in [13]. All programs are written in Matlab code and all computations are tested under Matlab 7.1 on Pentium 4.
In addition, in the computations, the following values are assigned to the parameters in the Algorithm: , , , , , , and . We also use the stopping criteria is being of .
In the following Table 1, the first two columns give the size of the matrix and the dimension of the variable . In the middle columns, โ-timeโ denotes the computing time (in seconds), โ-it.โ denotes the numbers iteration, and โ-obj.โ defines the value of when our stopping criteria is satisfied. Some numerical results of [13] are shown in the last two columns.
As shown in Table 1, all test problems have been solved just few iterations compared with [13]. Furthermore, this algorithm is less sensitive to the size of SDP problems. Comparatively speaking, our method is attractive and suitable for solving large-scale SDP problems.
6. Conclusions
In this paper, we have proposed a filter-trust-region method for SDP problems. Such a method offers a trade-off between the accuracy of solving the subproblems and the amount of work for solving them. Furthermore, numerical results show that our algorithm is attractive for large-scale SDP problems.
Acknowledgments
The authors would like to thank Professor Florian Jarre for his advice and guidance, Thomas David and Li Luo for their grateful help, and also the referees for their helpful comments. This work is supported by National Natural Science Foundation of China 10971162.