# QSDPNAL: a two-phase augmented Lagrangian method for convex quadratic semidefinite programming

## Abstract

In this paper, we present a two-phase augmented Lagrangian method, called QSDPNAL, for solving convex quadratic semidefinite programming (QSDP) problems with constraints consisting of a large number of linear equality and inequality constraints, a simple convex polyhedral set constraint, and a positive semidefinite cone constraint. A first order algorithm which relies on the inexact Schur complement based decomposition technique is developed in QSDPNAL-Phase I with the aim of solving a QSDP problem to moderate accuracy or using it to generate a reasonably good initial point for the second phase. In QSDPNAL-Phase II, we design an augmented Lagrangian method (ALM) wherein the inner subproblem in each iteration is solved via inexact semismooth Newton based algorithms. Simple and implementable stopping criteria are designed for the ALM. Moreover, under mild conditions, we are able to establish the rate of convergence of the proposed algorithm and prove the R-(super)linear convergence of the KKT residual. In the implementation of QSDPNAL, we also develop efficient techniques for solving large scale linear systems of equations under certain subspace constraints. More specifically, simpler and yet better conditioned linear systems are carefully designed to replace the original linear systems and novel shadow sequences are constructed to alleviate the numerical difficulties brought about by the crucial subspace constraints. Extensive numerical results for various large scale QSDPs show that our two-phase algorithm is highly efficient and robust in obtaining accurate solutions. The software reviewed as part of this submission was given the DOI (Digital Object Identifier) https://doi.org/10.5281/zenodo.1206980.

## Keywords

Quadratic semidefinite programming Schur complement Augmented Lagrangian Inexact semismooth Newton method## Mathematics Subject Classification

90C06 90C20 90C22 90C25 65F10## 1 Introduction

**P**) and its dual. We are particularly interested in the case where the dimensions

*n*and/or

*m*are large, and it may be impossible to explicitly store or compute the matrix representation of \(\mathcal{Q}\). For example, if \(\mathcal{Q}= H\otimes H\) is the Kronecker product of a dense matrix \(H\in \mathcal{S}^n_+\) with itself, then it would be extremely expensive to store the matrix representation of \(\mathcal{Q}\) explicitly when

*n*is larger than, say, 500. As far as we are aware of, the best solvers currently available for solving (

**P**) are based on inexact primal-dual interior-point methods [31]. However, they are highly inefficient for solving large scale problems as interior-point methods have severe inherent ill-conditioning limitations which would make the convergence of a Krylov subspace iterative solver employed to compute the search directions to be extremely slow. While sophisticated preconditioners have been constructed in [31] to alleviate the ill-conditioning, the improvement is however not dramatic enough for the algorithm to handle large scale problems comfortably. On the other hand, an interior-point method which employs a direct solver to compute the search directions is prohibitively expensive for solving (

**P**) since the cost is at least \(O( (m+n^2)^3)\) arithmetic operations per iteration. It is safe to say that there is currently no solver which can efficiently handle large scale QSDP problems of the form (

**P**) and our paper precisely aims to provide an efficient and robust solver for (

**P**).

**P**) (in its equivalent minimization form):

Due to its great potential in applications and mathematical elegance, QSDP has been studied quite actively both from the theoretical and numerical aspects [1, 11, 14, 15, 19, 23, 31, 32]. For the recent theoretical developments, one may refer to [7, 10, 22, 30] and the references therein. Here we focus on the numerical aspect and we will next briefly review some of the methods available for solving QSDP problems. Toh et al. [32] and Toh [31] proposed inexact primal-dual path-following interior-point methods to solve the special class of convex QSDP without the constraint in \(\mathcal{K}\). In theory, these methods can be used to solve QSDP problems with inequality constraints and constraint in \(\mathcal{K}\) by reformulating the problems into the required standard form. However, as already mentioned above, in practice interior-point methods are not efficient for solving QSDP problems beyond moderate scales either due to the extremely high computational cost per iteration or the inherent ill-conditioning of the linear systems governing the search directions. In [34], Zhao designed a semismooth Newton-CG augmented Lagrangian (NAL) method and analyzed its convergence for solving the primal QSDP problem (**P**). However, the NAL algorithm often encounters numerical difficulty (due to singular or nearly singular generalized Hessian) when the polyhedral set constraint \(X\in \mathcal{K}\) is present. Subsequently, Jiang et al. [13] proposed an inexact accelerated proximal gradient method for least squares semidefinite programming with only equality constraints where the objective function in (**P**) is expressed explicitly in the form of \(\Vert \mathcal{B}X - d \Vert ^2\) for some given linear map \(\mathcal{B}\).

More recently, inspired by the successes achieved in [28, 33] for solving linear SDP problems with nonnegative constraints, Li et al. [18] proposed a first-order algorithm, known as the Schur complement based semi-proximal alternating direction method of multipliers (SCB-sPADMM), for solving the dual form (**D**) of QSDP. As far as we aware of, [18] is the first paper to advocate using the dual approach for solving QSDP problems even though the dual problem (**D**) looks a lot more complicated than the primal problem (**P**), especially with the presence of the subspace constraint involving \(\mathcal{W}\). By leveraging on the Schur complement based decomposition technique developed in [17, 18], Chen et al. [6] also employed the dual approach by proposing an efficient inexact ADMM-type first-order method (which we name as SCB-isPADMM) for solving problem (**D**). Promising numerical results have been obtained by the dual based first-order algorithms in solving various classes of QSDP problems to moderate accuracy [6, 18]. Naturally one may hope to also rely on the ADMM scheme to compute highly accurate solutions. However, as one will observe from the numerical experiments presented later in Sect. 6, ADMM-type methods are incapable of finding accurate solutions for difficult QSDP problems due to their slow local convergence or stagnation. On the other hand, recent studies on the convergence rate of the augmented Lagrangian method (ALM) for solving convex semidefinite programming with multiple solutions [7] show that comparing to ADMM-type methods, the ALM can enjoy a faster convergence rate (in fact asymptotically superlinear) under milder conditions. These recent advances thus strongly indicate that one should be able to design a highly efficient algorithm based on the ALM for (**D**) for solving QSDP problems to high accuracy. More specifically, we will propose a two-phase augmented Lagrangian based algorithm with Phase I to generate a reasonably good initial point to warm start the Phase II algorithm so as to compute accurate solutions efficiently. We call this new method Qsdpnal since it extends the ideas of SDPNAL [35] and SDPNAL+ [33] for linear SDP problems to QSDP problems. Although the aforementioned two-phase framework has already been demonstrated to be highly efficient for solving linear SDP problems [33, 35], it remains to be seen whether we can achieve comparable or even more impressive performance on various QSDP problems.

In recent years, it has become fashionable to design first-order algorithms for solving convex optimization problems, with some even claiming their efficacy in solving various challenging classes of matrix conic optimization problems based on limited performance evaluations. However, based on our extensive numerical experience in solving large scale linear SDPs [28, 33, 35], we have observed that while first-order methods can be rather effective in solving easy problems which are well-posed and nondegenerate, they are typically powerless in solving difficult instances which are ill-posed or degenerate. Even for a well designed first-order algorithm with guaranteed convergence and highly optimized implementations, such as the ADMM+ algorithm in [28], a first-order method may still fail on slightly more challenging problems. For example, the ADMM+ algorithm designed in [33] can encounter varying degrees of difficulties in solving linear SDPs arising from rank-one tensor approximation problems. On the other hand, the SDPNAL algorithm in [35] (which exploits second-order information) is able to solve those problems very efficiently to high accuracy. We believe that in order to design an efficient and robust algorithm to solve the highly challenging class of matrix conic optimization problems including QSDPs, one must fully combine the advantages offered by both the first and second order algorithms, rather than just solely relying on first-order algorithms even though they may appear to be easier to implement.

**D**):

**D**) is defined by

**D**) where the inner subproblem in each iteration is solved via an inexact semismooth Newton based algorithm. Given \(\sigma _0 >0\), \((Z^0,W^0,S^0,y^0,X^0)\in \mathcal{Z}\times \mathcal{S}^n\), the \((k+1)\)th iteration of the ALM consists of the following steps:where \(\sigma _k\in (0,+\infty )\). The first issue in the above ALM is the choice of the subspace \(\mathcal{W}\). The obvious choice \(\mathcal{W}= \mathcal{S}^n\) can lead to various difficulties in the implementation of the above algorithm. For example, since \(\mathcal{Q}:\mathcal{S}^n\rightarrow \mathcal{S}^n\) is only assumed to be positive semidefinite, the Newton systems corresponding to the inner subproblems may be singular and the sequence \(\{W^{k}\}\) generated by the ALM can be unbounded. As a result, it will be extremely difficult to analyze the convergence of the inner algorithm for solving the ALM subproblems. The second issue is that one needs to design easy-to-check stopping criteria for the inner subproblems, and to ensure the fast convergence of the ALM under reasonable conditions imposed on the QSDP problems. Concerning the first issue, we propose to choose \(\mathcal{W}= Ran (\mathcal{Q})\), although such a choice also leads to obstacles which we will overcome in Section 4. Indeed, by restricting \(W\in Ran (\mathcal{Q})\), the difficulties in analyzing the convergence and the superlinear (quadratic) convergence of the Newton-CG algorithm are circumvented as the possibilities of singularity and unboundedness are removed. For the second issue, under the restriction that \(\mathcal{W}= Ran (\mathcal{Q})\), thanks to the recent advances in [7], we are able to design checkable stopping criteria for solving the inner subproblems inexactly while establishing the global convergence of the above ALM. Moreover, we are able to establish the R-(super)linear convergence rate of the KKT residual. At the first glance, the restriction that \(W\in Ran (\mathcal{Q})\) appears to introduce severe numerical difficulties when we need to solve a linear system under this restriction. Fortunately, by carefully examining our algorithm and devising novel numerical techniques, we are able to overcome these difficulties as we shall see in Sect. 4. Our extensive evaluations of Qsdpnal have demonstrated that our algorithm is capable of solving large scale general QSDP problems of the form (

**P**) to high accuracy very efficiently and robustly. For example, we are able to solve an elementwise weighted nearest correlation matrix estimation problem with matrix dimension \(n=10{,}000\) in less than 11 h to the relative accuracy of less than \(10^{-6}\) in the KKT residual. Such a numerical performance has not been attained in the past.

As the readers may have already observed, even though our goal in developing algorithms for solving convex optimization problems such as (**P**) and (**D**) is to design those with desirable theoretical properties such as asymptotic superlinear convergence, it is our belief that it is equally if not even more important for the algorithms designed to be practically implementable and able to achieve realistic numerical efficiency. It is obvious that our proposed two-phase augmented Lagrangian based algorithm for solving (**P**) and (**D**) is designed based on such a belief.

The remaining parts of this paper are organized as follows. The next section is devoted to our main algorithm Qsdpnal, which is a two-phase augmented Lagrangian based algorithm whose Phase I is used to generate a reasonably good initial point to warm-start the Phase II algorithm so as to obtain accurate solutions efficiently. In Sect. 3, we propose to solve the inner minimization subproblems of the ALM by semismooth Newton based algorithms and study their global and local superlinear (quadratic) convergence. In Sect. 4, we discuss critical numerical issues concerning the efficient implementation of Qsdpnal. In Sect. 5.1, we discuss the special case of applying Qsdpnal to solve least squares semidefinite programming problems. The extension of Qsdpnal for solving QSDP problems with unstructured inequality constraints is discussed in Sect. 5.2. In Sect. 6, we conduct numerical experiments to evaluate the performance of Qsdpnal in solving various QSDP problems and their extensions. We conclude our paper in the final section.

Below we list several notation and definitions to be used in the paper. For a given closed proper convex function \(\theta :\mathcal{X}\rightarrow (-\infty ,\infty ]\), where \(\mathcal{X}\) is a finite-dimensional real inner product space, its Fenchel conjugate function is denoted by \(\theta ^*:\mathcal{X}\rightarrow (-\infty , +\infty ]\) . For a given closed convex set \(D\subseteq \mathcal{X}\) and \(x\in \mathcal{X}\), we define by \(\Pi _{D}(x)\) the metric projector of *x* onto *D* and \(\mathrm{dist}(x,D): = \inf _{d\in D}\Vert x - d \Vert = \Vert x - \Pi _{D}(x) \Vert .\) For any \(X\in \mathcal{S}^n\), we use \(\lambda _{\max }(X)\) and \(\lambda _{\min }(X)\) to denote the largest and smallest eigenvalues of *X*, respectively. Similar notation is used when *X* is replaced by the linear operator \(\mathcal{Q}\).

## 2 A two-phase augmented Lagrangian method

In this section, we shall present our two-phase algorithm Qsdpnal for solving the QSDP problems (**D**) and (**P**). For the convergence analysis of Algorithm Qsdpnal, we need to make the following standard assumption for (**P**). Such an assumption is analogous to the Slater’s condition in the context of nonlinear programming in \(\mathfrak {R}^m\).

### Assumption 1

### 2.1 Phase I: An SCB based inexact semi-proximal ADMM

In Phase I, we propose a new variant of the Schur complement based inexact semi-proximal ADMM (SCB-isPADMM) developed in [6] to solve (**D**). Recall the augmented Lagrangian function associated with problem (**D**) defined in (1).

**D**) are given as follows.

### Remark 2.1

### Remark 2.2

In contrast to Aglorithm SCB-isPADMM in [6], our Algorithm Qsdpnal-Phase I requires the subspace constraint \(W\in Ran (\mathcal{Q})\) explicitly in the subproblems (2) and (3). Note that due to the presence of the subspace constraint \(W\in Ran (\mathcal{Q})\), there is no need to add extra proximal terms in the subproblems corresponding to *W* to satisfy the positive definiteness requirement needed in applying the inexact Schur compliment based decomposition technique developed in [17, 18]. This is certainly more elegant than the indirect reformulation strategy considered in [6, 18].

The convergence of the above algorithm follows from [6, Theorem 1] without much difficulty, and its proof is omitted.

### Theorem 2.1

Suppose that the solution set of (**P**) is nonempty and Assumption 1 holds. Let \(\{(Z^k,W^k,S^k,y^k,X^k)\}\) be the sequence generated by Algorithm Qsdpnal-Phase I. If \(\tau \in (0,(1+\sqrt{5}\,)/2)\), then the sequence \(\{(Z^k,W^k,S^k,y^k)\}\) converges to an optimal solution of (**D**) and \(\{X^k\}\) converges to an optimal solution of (**P**).

### Remark 2.3

Under some error bound conditions on the limit point of \(\{(Z^k,W^k,S^k,y^k,X^k)\}\), one can derive the linear rate of convergence of the exact version of Algorithm Qsdpnal-Phase I. For a recent study on this topic, see [10] and the references therein. Here we will not address this issue as our Phase II algorithm enjoys a better rate of convergence under weaker conditions.

### 2.2 Phase II: An augmented Lagrangian algorithm

**D**). The purpose of this phase is to obtain high accuracy solutions efficiently after being warm-started by our Phase I algorithm. The Phase II of our algorithm has the following template.

**P**) be denoted as \(\mathcal{F}:=\{X\in \mathcal{S}^n\mid \mathcal{A}X = b, \, X\in \mathcal{S}^n_+\cap \mathcal{K}\}\). Define the feasibility residual function \(\gamma :\mathcal{S}^n\rightarrow \mathfrak {R}\) for the primal problem (

**P**) by

*X*. Similar to [7, Proposition 4.2], we can use this feasibility residual function to derive an upper bound on the distance of a given point to the feasible set \(\mathcal{F}\) in the next lemma. Its proof can be obtained without much difficulty by applying Hoffman’s error bound [9, Lemma 3.2.3] to the nonempty polyhedral convex set \(\{X\in \mathcal{S}^n\mid \mathcal{A}X = b,\, X\in \mathcal{K}\}\), e.g., see [2, Theorem 7].

### Lemma 2.1

**D**), numerically it is difficult to execute the criteria \((A'')\) and \((\mathrm{B}_1'')\) proposed in [26]. Fortunately, Lemma 2.1 and recent advances in the analysis of the ALM [7] allow us to design easy-to-verify stopping criteria for the subproblems in Qsdpnal-Phase II. For any \(k\ge 0\), denote

### Lemma 2.2

**P**). Then one can always find \((Z^{k+1},W^{k+1},S^{k+1},y^{k+1})\) and \(X^{k+1} = X^k + \sigma _k(Z^{k+1} - \mathcal{Q}W^{k+1} + S^{k+1} + \mathcal{A}^*y^{k+1} - C)\) satisfying both (A) and (B). Moreover, (A) implies that

### Proof

With the help of Lemma 2.1, one can establish the assertion in the same fashion as in [7, Proposition 4.2, Proposition 4.3]. \(\square \)

**P**), which is given by

**P**).

**P**), the second order growth condition holds at an optimal solution \({\overline{X}}\in \Omega \) with respect to the set \(\Omega \) if there exist \(\kappa >0 \) and a neighborhood

*U*of \({\overline{X}}\) such that

**D**) be given as follows:Now, with Lemma 2.2, we can prove the global and local (super)linear convergence of Algorithm Qsdpnal-Phase II by adapting the proofs in [26, Theorem 4] and [7, Theorem 4.2]. It shows that, for most QSDP problems, one can always expect the KKT residual of the sequence generated by Qsdpnal-Phase II to converge at least R-(super)linearly.

### Theorem 2.2

**P**), is nonempty and Assumption 1 holds. Then the sequence \(\{(Z^k,W^k,S^k,y^k,X^k)\}\) generated by Algorithm Qsdpnal-Phase II under the stopping criterion \((A )\) for all \(k\ge 0\) is bounded, and \(\{X^k\}\) converges to an optimal solution \(X^{\infty }\) of (

**P**), and \(\{(Z^k,W^k,S^k,y^k)\}\) converges to an optimal solution of (

**D**). Moreover, for all \(k\ge 0\), it holds that

**P**), the second order growth condition (7) holds at \(X^{\infty }\) with respect to the set \(\Omega \), i.e., there exists a constant \(\kappa > 0\) and a neighborhood

*U*of \(X^{\infty }\) such that

*k*sufficiently large, it holds that

Next we give a few comments on the convergence rates and assumptions made in Theorem 2.2.

### Remark 2.4

Under the assumptions of Theorem 2.2, we have proven that the KKT residual, corresponding to (**P**) and (**D**), along the sequence \(\{(Z^{k},W^k,S^k,y^k,X^k)\}\) converges at least R-(super)linearly. Indeed, under stopping criteria (A), (B) and from (8),(9) and (10), we know that the primal feasibility, the dual feasibility and the duality gap all converge at least R-(super)linearly.

### Remark 2.5

The assumption that the second order growth condition (7) holds for (**P**) is quite mild. Indeed, it holds when any optimal solution \({\overline{X}}\) of (**P**), together with any of its multiplier \({\overline{S}} \in \mathcal{S}^n_+\) corresponding only to the semidefinite constraint, satisfies the strict complementarity condition [7, Corollary 3.1]. It is also valid when the “no-gap” second order sufficient condition holds at the optimal solution^{1} to (**P**) [4, Theorem 3.137].

## 3 Inexact semismooth Newton based algorithms for solving the inner subproblems (4) in ALM

### 3.1 A semismooth Newton-CG algorithm for (11) with \(\mathcal{K}= \mathcal{S}^n\)

Note that in quite a number of applications, the polyhedral convex set \(\mathcal{K}\) is actually the whole space \(\mathcal{S}^n\). Therefore, we shall first study how the inner problems (11) in Algorithm ALM can be solved efficiently when \(\mathcal{K}= \mathcal{S}^n\). Under this setting, *Z* is vacuous, i.e., \(Z=0\).

*W*,

*y*), i.e., the Clarke subdifferential of \(\nabla \varphi \) at (

*W*,

*y*).

The convergence results for the above SNCG algorithm are stated in the next theorem.

### Theorem 3.1

Suppose that Assumption 1 holds. Then Algorithm SNCG generates a bounded sequence \(\{(W^j,y^j)\}\) and any accumulation point \(({\overline{W}}, {\bar{y}}) \in Ran (\mathcal{Q})\times \mathfrak {R}^m\) is an optimal solution to problem (12).

The following proposition is the key ingredient in our subsequent convergence analysis.

### Proposition 3.1

### Proof

Base on the above proposition, under the constraint nondegeneracy condition for (**P**), we shall show in the next theorem that one can still ensure the positive definiteness of the coefficient matrix in the semismooth Newton system at the solution point.

### Theorem 3.2

- (i)The constraint nondegeneracy condition,holds at \({\overline{Y}}\), where \(\mathrm{lin}(\mathcal{T}_{\mathcal{S}^n_+}(\overline{Y}))\) denotes the lineality space of the tangent cone of \(\mathcal{S}^n_+\) at \(\overline{Y}\).$$\begin{aligned} \mathcal{A}\,\mathrm{lin}(\mathcal{T}_{\mathcal{S}^n_+}(\overline{Y})) = \mathfrak {R}^m, \end{aligned}$$(18)
- (ii)Every element inis self-adjoint and positive definite on \(Ran (\mathcal{Q})\times \mathfrak {R}^m.\)$$\begin{aligned} \left[ \begin{array}{cc} \mathcal{Q}&{}\quad \\ &{}\quad 0 \end{array} \right] +\sigma \left[ \begin{array}{c} \mathcal{Q}\\ -\mathcal{A}\end{array} \right] \partial \Pi _{\mathcal{S}^n_+}(\mathcal{A}^*{\bar{y}} - \mathcal{Q}{\overline{W}} - {\widehat{C}}) [\mathcal{Q}\, -\mathcal{A}^*] \end{aligned}$$

### Proof

In the same fashion as in [35, Proposition 3.2], we can prove that \( \mathcal{A}\mathcal{U}\mathcal{A}^* \) is positive definite for all \(\mathcal{U}\in \partial \Pi _{\mathcal{S}^n_+}(\mathcal{A}^*{\bar{y}} - \mathcal{Q}{\overline{W}} - {\widehat{C}})\) if only if (i) holds. Then, by Proposition 3.1, we readily obtain the desired results. \(\square \)

### Theorem 3.3

### Proof

From Theorem 3.2, we know that under the constraint nondegeneracy condition (18), every \(V \in {\hat{\partial }}^{2} \varphi ({\overline{W}},{\bar{y}})\) is self-adjoint and positive definite on \( Ran (\mathcal{Q})\times \mathfrak {R}^n\). Hence one can obtain the desired results from [35, Theorem 3.5] by further noting the strong semismoothness of \(\Pi _{\mathcal{S}^n_+}(\cdot )\). \(\square \)

We note that the convergence results obtained in this subsection depend critically on the restriction that \(W\in \mathcal{W}= \mathrm{Ran}(\mathcal{Q})\). Without this restriction, the possible singularity of the Newton systems (15) and the unboundedness of \(\{W^j\}\) will make the convergence analysis highly challenging, if possible at all.

### 3.2 Semismooth Newton based inexact ABCD algorithms for (11) when \(\mathcal{K}\not =\mathcal{S}^n\)

When \(\mathcal{K}\ne \mathcal{S}^n\), we will adapt the recently developed inexact accelerated block coordinate descent (ABCD) algorithm [29] to solve the inner subproblems (11) in the augmented Lagrangian method.

*Z*,

*W*,

*S*,

*y*) is decomposed into two groups, namely

*Z*and (

*W*,

*S*,

*y*). In this case, (

*W*,

*S*,

*y*) is regarded as a single block and the corresponding subproblem in the ABCD algorithm can only be solved by an iterative method inexactly. Here, we propose to develop a semismooth Newton-CG method to solve the corresponding subproblem.

Note that in order to meet the convergence requirement of the inexact ABCD algorithm, a proximal term involving the positive parameter \(\eta \) is added in (19) to ensure the strong convexity of the objective function in the subproblem. For computational efficiency, one can always take \(\eta \) to be a small number, say \(10^{-6}\). For the subproblem (19), it can be solved by a semismooth Newton-CG algorithm similar to the one developed in Sect. 3.1. Since \(\eta >0\), the superlinear convergence of such a semismooth Newton-CG algorithm can also be proven based on the strong semismoothness of \(\Pi _{\mathcal{S}^n_+}(\cdot )\) and the symmetric positive definiteness of the corresponding generalized Hessian.

The convergence results for the above Algorithm ABCD are stated in the next theorem, whose proof essentially follows from that in [29, Theorem 3.1]. Here, we omit the proof for brevity.

### Theorem 3.4

## 4 Numerical issues in Qsdpnal

*k*th iteration, we need to solve the following linear system of equations

*n*, say \(n= 500\). Under the high dimensional setting which we are particularly interested in, the matrix representation of \(\mathcal{Q}\) is generally not available or too expensive to be stored explicitly. Thus (20) can only be solved inexactly by an iterative method. However when \(\mathcal{Q}\) is singular (and hence \(Ran (\mathcal{Q})\ne \mathcal{S}^n\)), due to the presence of the subspace constraint \(W\in Ran (\mathcal{Q})\), it is extremely difficult to apply preconditioning to (20) while ensuring that the approximate solution is contained in \(Ran (\mathcal{Q})\). Fortunately, as shown in the next proposition, instead of solving (20) directly, we can solve a simpler and yet better conditioned linear system to overcome this difficulty.

### Proposition 4.1

### Proof

By Proposition 4.1, in order to obtain \(\widehat{W}_{\mathcal{Q}}\), we can first apply an iterative method such as the preconditioned conjugate gradient (PCG) method to solve (22) to obtain \(\widehat{W}\) and then perform the projection step. However, by carefully analysing the steps in Qsdpnal-Phase I, we are surprised to observe that instead of explicitly computing \(\widehat{W}_{\mathcal{Q}}\), we can update the iterates in the algorithm by using only \(\mathcal{Q}\widehat{W}_{\mathcal{Q}}=\mathcal{Q}\widehat{W}\). Thus, we only need to compute \(\mathcal{Q}{\widehat{W}}\) and the potentially expensive projection step to compute \(\widehat{W}_\mathcal{Q}\) can be avoided completely.

It is important for us to emphasize the computational advantage of solving the linear system (22) over (20). First, the former only requires one evaluation of \(\mathcal{Q}(\cdot )\) whereas the latter requires two such evaluations in each PCG iteration. Second, the coefficient matrix in the former system is typically much more well-conditioned than the coefficient matrix in the latter system. More precisely, when \(\mathcal{Q}\) is positive definite, then \(\mathcal{I}+\sigma \mathcal{Q}\) is clearly better conditioned than \(\mathcal{Q}+\sigma \mathcal{Q}^2\) by a factor of \(\lambda _{\max }(\mathcal{Q})/\lambda _{\min }(\mathcal{Q})\). When \(\mathcal{Q}\) is singular, with its smallest positive eigenvalue denoted as \(\lambda _{+}(\mathcal{Q})\), then \(\mathcal{I}+\sigma \mathcal{Q}\) is better conditioned when \(\lambda _{\max }(\mathcal{Q}) \ge \lambda _+(\mathcal{Q})(1+\sigma \lambda _+(\mathcal{Q}))\). The previous inequality would obviously hold when \(\lambda _+ \le (\sqrt{4\sigma \lambda _{\max }(\mathcal{Q})+1}-1)/(2\sigma )\).

*dW*,

*dy*) to solve the following linear system

### Proposition 4.2

### Proof

## 5 Adaption of QSDPNAL for least squares SDP and inequality constrained QSDP

Here we discuss how our algorithm Qsdpnal can be modified and adapted for solving least squares semidefinite programming as well as general QSDP problems with additional unstructured inequality constraints which are not captured by the polyhedral set \(\mathcal{K}.\)

### 5.1 The case for least squares semidefinite programming

In this subsection, we show that for least squares semidefinite programming problems, Qsdpnal can be used in a more efficient way to avoid the difficulty of handling the subspace constraint \(W\in Ran (\mathcal{Q})\).

### Remark 5.1

When the polyhedral constraint \(X\in \mathcal{K}\) in (26) is absent, i.e., the polyhedral convex set \(\mathcal{K}= \mathcal{S}^n\), Jiang, Sun and Toh in [14] have proposed a partial proximal point algorithm for solving the least squares semidefinite programming problem (26). Here our Algorithm Qsdpnal is built to solve the much more general class of convex composite QSDP problems.

### 5.2 Extension to QSDP problems with inequality constraints

*x*, we can equivalently rewrite (31) into the following standard form:

*x*. The dual of (32) is given by

**D**) as follows:

## 6 Computational experiments

In our implementation of Qsdpnal, we always run Qsdpnal-Phase I first to generate a reasonably good starting point to warm start our Phase II algorithm. We terminate the Phase I algorithm and switch to the Phase II algorithm if a solution with a moderate accuracy (say a solution with \(\eta _{\mathrm{qsdp}} < 10^{-4}\)) is obtained or if the Phase I algorithm reaches the maximum number of iterations (say 1000 iterations). If the underlying problems contain inequality or polyhedral constraints, we further employ a restarting strategy similar to the one in [33], i.e., when the progress of Qsdpnal-Phase II is not satisfactory, we will restart the whole Qsdpnal algorithm by using the most recently computed \((Z,W,S,y,X,\sigma )\) as the initial point. In addition, we also adopt a dynamic tuning strategy to adjust the penalty parameter \(\sigma \) appropriately based on the progress of the primal and dual feasibilities of the computed iterates.

All our computational results are obtained from a workstation running on 64-bit Windows Operating System having 16 cores with 32 Intel Xeon E5-2650 processors at 2.60 GHz and 64 GB memory. We have implemented Qsdpnal in Matlab version 7.13.

### 6.1 Evaluation of Qsdpnal on the nearest correlation matrix problems

*E*is a randomly generated symmetric matrix with entries uniformly distributed in \([-\,1,1]\) except for its diagonal elements which are all set to 1. The weight matrix

*H*is generated from a weight matrix \(H_0\) used by a hedge fund company. The matrix \(H_0\) is a \(93 \times 93\) symmetric matrix with all positive entries. It has about \(24\%\) of the entries equal to \(10^{-5}\) and the rest are distributed in the interval \([2, 1.28\times 10^3].\) The Matlab code for generating the matrix

*H*is given byThe reason for using such a weight matrix is because the resulting problems generated are more challenging to solve as opposed to a randomly generated weight matrix. We also test four more instances, namely PDidx2000, PDidx3000, PDidx5000 and PDidx10000, where the raw correlation matrix \({\widehat{G}}\) is generated from the probability of default (PD) data obtained from the RMI Credit Research Initiative

^{2}at the National University of Singapore. We consider two choices of \(\mathcal{K}\), i.e., case (i): \(\mathcal{K}= \mathcal{S}^n\) and case (ii): \(\mathcal{K}= \{X\in \mathcal{S}^n\mid \, X_{ij} \ge -\,0.5, \;\forall \; i,j=1,\ldots ,n\}\).

The performance of Qsdpnal and Qsdpnal-Phase I on H-weighted NCM problems (dual of (36)) (accuracy \(= 10^{-6}\)). In the table, “a” stands for Qsdpnal and “b” stands for Qsdpnal-Phase I, respectively. The computation time is in the format of “hours:minutes:seconds”

Problem | | \(\alpha \) | iter.a | iter.b | \(\eta _{qsdp }\) | \(\eta _gap \) | Time |
---|---|---|---|---|---|---|---|

it (subs) | itSCB | a|b | a|b | a|b | ||||

\(\mathcal{K}= \mathcal{S}^n\) | |||||||

Lymph | 587 | 0.10 | 12 (40) | 52 | 251 | 9.1–7 | 9.1–7 | 8.2–7 | − 3.9–7 | 13 | 23 |

Lymph | 587 | 0.05 | 11 (32) | 38 | 205 | 9.5–7 | 9.9–7 | 7.5–7 | − 4.1–7 | 09 | 19 |

ER | 692 | 0.10 | 12 (41) | 54 | 250 | 9.8–7 | 9.9–7 | 5.4–7 | − 4.8–7 | 17 | 33 |

ER | 692 | 0.05 | 12 (38) | 43 | 218 | 7.3–7 | 9.7–7 | 2.5–7 | − 4.4–7 | 14 | 28 |

Arabidopsis | 834 | 0.10 | 12 (42) | 56 | 285 | 8.5–7 | 9.9–7 | 2.8–7 | − 5.3–7 | 27 | 57 |

Arabidopsis | 834 | 0.05 | 12 (41) | 44 | 230 | 8.0–7 | 9.5–7 | − 6.8–8 | − 4.5–7 | 24 | 46 |

Leukemia | 1255 | 0.10 | 12 (41) | 62 | 340 | 8.4–7 | 9.9–7 | 3.1–7 | − 5.4–7 | 1:08 | 2:48 |

Leukemia | 1255 | 0.05 | 12 (38) | 49 | 248 | 7.6–7 | 8.7–7 | − 1.3–7 | − 4.5–7 | 58 | 2:06 |

hereditarybc | 1869 | 0.10 | 13 (47) | 76 | 393 | 6.4–7 | 9.9–7 | − 2.2–7 | − 9.8–7 | 3:01 | 7:10 |

hereditarybc | 1869 | 0.05 | 13 (45) | 60 | 311 | 8.6–7 | 9.9–7 | − 4.7–7 | \({-}{} { 1.0}{-}{} { 6}\) | 2:39 | 5:44 |

PDidx2000 | 2000 | 0.10 | 13 (51) | 131 | 590 | 9.5–7 | 9.9–7 | 2.4–7 | − 8.5–7 | 5:04 | 11:43 |

PDidx2000 | 2000 | 0.05 | 14 (58) | 139 | 626 | 7.5–7 | 9.9–7 | − 5.6–8 | − 9.5–7 | 5:52 | 12:41 |

PDidx3000 | 3000 | 0.10 | 14 (55) | 145 | 1201 | 8.1–7 | 9.9–7 | − 2.8–7 | \(\textit{2.1}{-}{} \textit{6}\) | 14:59 | 1:15:01 |

PDidx3000 | 3000 | 0.05 | 14 (58) | 136 | 1263 | 6.8–7 | 9.7–7 | − 2.6–7 | \(\textit{2.0}{-}{} \textit{6}\) | 14:50 | 1:19:27 |

PDidx5000 | 5000 | 0.10 | 15 (63) | 189 | 1031 | 8.0–7 | 9.9–7 | − 1.9–7 | \(\textit{1.8}{-}{} \textit{6}\) | 1:17:47 | 4:17:10 |

PDidx5000 | 5000 | 0.05 | 14 (59) | 164 | 1699 | 9.2–7 | 9.9–7 | − 3.3–7 | − 1.3–7 | 1:11:46 | 6:18:29 |

PDidx10000 | 10,000 | 0.10 | 16 (71) | 200 | 2572 | 7.1–7 | 9.9–7 | 1.6–7 | − 1.5–7 | 9:57:18 | 60:07:08 |

PDidx10000 | 10,000 | 0.05 | 16 (73) | 200 | 2532 | 9.5–7 | 9.9–7 | 4.7–8 | 1.4–7 | 10:34:31 | 59:34:13 |

\(\mathcal{K}= \{X\in \mathcal{S}^n\mid \, X_{ij}\ge -\,0.5\;\forall \; i,j=1,\ldots ,n\}\) | |||||||

Lymph | 587 | 0.10 | 5 (14) | 129 | 244 | 9.8–7 | 9.9–7 | − 1.0–7 | − 4.4–7 | 18 | 30 |

Lymph | 587 | 0.05 | 5 (12) | 120 | 257 | 9.9–7 | 9.9–7 | − 3.4–7 | − 4.2–7 | 15 | 28 |

ER | 692 | 0.10 | 5 (14) | 126 | 266 | 9.9–7 | 9.9–7 | − 1.5–7 | − 5.1–7 | 22 | 40 |

ER | 692 | 0.05 | 5 (14) | 117 | 217 | 8.4–7 | 9.9–7 | − 2.7–7 | − 4.4–7 | 21 | 32 |

Arabidopsis | 834 | 0.10 | 6 (16) | 240 | 472 | 9.9–7 | 9.9–7 | − 5.4–7 | − 6.0–7 | 1:03 | 1:56 |

Arabidopsis | 834 | 0.05 | 6 (15) | 240 | 442 | 8.5–7 | 9.9–7 | − 4.4–7 | − 5.6–7 | 1:02 | 1:46 |

Leukemia | 1255 | 0.10 | 7 (22) | 188 | 333 | 9.9–7 | 9.9–7 | − 4.4–7 | − 5.5–7 | 2:10 | 3:06 |

Leukemia | 1255 | 0.05 | 7 (19) | 159 | 253 | 9.9–7 | 9.9–7 | − 5.4–7 | − 5.3–7 | 1:46 | 2:18 |

hereditarybc | 1869 | 0.10 | 8 (22) | 397 | 577 | 9.3–7 | 9.9–7 | − 8.0–7 | − 8.9–7 | 10:28 | 12:59 |

hereditarybc | 1869 | 0.05 | 8 (22) | 361 | 472 | 9.6–7 | 9.9–7 | − 8.1–7 | − 8.6–7 | 9:39 | 10:04 |

PDidx2000 | 2000 | 0.10 | 20 (52) | 672 | 716 | 9.9–7 | 9.9–7 | − 6.8–7 | − 7.9–7 | 21:32 | 17:42 |

PDidx2000 | 2000 | 0.05 | 22 (60) | 756 | 1333 | 9.6–7 | 5.8–7 | − 6.3–7 | − 4.0–7 | 25:20 | 39:34 |

PDidx3000 | 3000 | 0.10 | 34 (101) | 659 | 1647 | 9.9–7 | 9.9–7 | − 7.0–7 | − 9.4–7 | 1:14:15 | 1:53:13 |

PDidx3000 | 3000 | 0.05 | 41 (117) | 728 | 1538 | 9.9–7 | 9.9–7 | − 6.2–7 | \({-}{} \textit{1.2}{-}{} \textit{6}\) | 1:21:13 | 1:50:47 |

PDidx5000 | 5000 | 0.10 | 29 (79) | 829 | 1484 | 9.3–7 | 8.4–7 | − 5.5–7 | 6.4–7 | 5:00:35 | 7:25:19 |

PDidx5000 | 5000 | 0.05 | 33 (107) | 1081 | 1722 | 9.9–7 | 9.9–7 | − 6.4–7 | − 1.5–7 | 6:30:35 | 7:16:08 |

PDidx10000 | 10,000 | 0.10 | 42 (136) | 1289 | 2190 | 9.9–7 | 9.9–7 | − 7.1–7 | 2.6–7 | 58:44:49 | 64:17:14 |

PDidx10000 | 10,000 | 0.05 | 40 (122) | 1519 | 3320 | 9.9–7 | 4.5–7 | − 6.6–7 | − 1.7–8 | 65:13:19 | 94:53:10 |

In Table 1, we report the numerical results obtained by Qsdpnal and Qsdpnal-Phase I in solving various instances of the H-weighted NCM problem (36). In the table, “it (subs)” denotes the number of outer iterations with subs in the parenthesis indicating the number of inner iterations of Qsdpnal-Phase II and “itSCB” stands for the total number of iterations used in Qsdpnal-Phase I. We can see from Table 1 that Qsdpnal is more efficient than the purely first-order algorithm Qsdpnal-Phase I. In particular, for the instance PDidx10000 where the matrix dimension \(n=10{,}000\), we are able to solve the problem in about 11 h while the purely first-order method Qsdpnal-Phase I needs about 60 h.

### 6.2 Evaluation of Qsdpnal on instances generated from BIQ problems

*Q*and

*c*are taken from the Biq Mac Library maintained by Wiegele, which is available at http://biqmac.uni-klu.ac.at/biqmaclib.html.

Same as Table 1 but for QSDP-BIQ problems

| \(m_E;m_I\) | | iter.a | iter.b | \(\eta _{qsdp }\) | \(\eta _gap \) | Time |
---|---|---|---|---|---|---|

it (subs)|itSCB | a|b | a|b | a|b | |||

be200.3.1 | 201 ; 59,700 | 201 | 66 (135) | 3894 | 4701 | 7.8–7 | 9.8–7 | − 3.5–7 | − 7.2–7 | 3:37 | 3:57 |

be200.3.2 | 201 ; 59,700 | 201 | 37 (74) | 2969 | 13,202 | 9.7–7 | 9.9–7 | − 2.1–7 | − 6.7–8 | 2:42 | 12:20 |

be200.3.3 | 201 ; 59,700 | 201 | 51 (107) | 5220 | 10,375 | 8.1–7 | 9.9–7 | − 1.1–7 | − 6.5–7 | 5:00 | 8:52 |

be200.3.4 | 201 ; 59,700 | 201 | 36 (72) | 3484 | 4966 | 9.8–7 | 9.9–7 | − 1.6–7 | − 4.1–7 | 3:15 | 4:14 |

be200.3.5 | 201 ; 59,700 | 201 | 22 (44) | 2046 | 3976 | 9.8–7 | 9.9–7 | − 5.9–8 | − 3.0–7 | 1:53 | 3:28 |

be250.1 | 251 ; 93,375 | 251 | 98 (196) | 6931 | 12,220 | 9.9–7 | 9.9–7 | 3.2–7 | 3.5–8 | 8:11 | 14:07 |

be250.2 | 251 ; 93,375 | 251 | 81 (169) | 6967 | 16,421 | 9.3–7 | 9.9–7 | 3.2–7 | − 5.7–7 | 8:35 | 20:01 |

be250.3 | 251 ; 93,375 | 251 | 123 (250) | 7453 | 9231 | 9.3–7 | 9.8–7 | − 1.7–7 | − 5.1–7 | 9:27 | 10:25 |

be250.4 | 251 ; 93,375 | 251 | 36 (72) | 3583 | 4542 | 9.9–7 | 9.9–7 | 5.2–8 | − 2.1–7 | 4:31 | 5:06 |

be250.5 | 251 ; 93,375 | 251 | 99 (198) | 5004 | 12,956 | 8.3–7 | 9.9–7 | 1.8–7 | − 1.8–7 | 6:38 | 15:52 |

bqp500-1 | 501 ; 374,250 | 501 | 62 (131) | 5220 | 11,890 | 9.9–7 | 9.9–7 | − 7.1–7 | − 8.2–8 | 37:56 | 1:23:58 |

bqp500-2 | 501 ; 374,250 | 501 | 41 (84) | 3610 | 8159 | 5.5–7 | 9.9–7 | − 3.8–7 | − 8.7–8 | 24:01 | 55:14 |

bqp500-3 | 501 ; 374,250 | 501 | 89 (200) | 5877 | 6402 | 9.9–7 | 8.6–7 | 5.4–7 | − 1.9–7 | 40:29 | 41:51 |

bqp500-4 | 501 ; 374,250 | 501 | 95 (256) | 7480 | 11,393 | 6.3–7 | 9.9–7 | − 1.5–7 | − 1.1–7 | 56:12 | 1:17:56 |

bqp500-5 | 501 ; 374,250 | 501 | 107 (247) | 6976 | 8823 | 5.1–7 | 9.9–7 | 6.2–7 | − 1.0–7 | 52:24 | 59:11 |

bqp500-6 | 501 ; 374,250 | 501 | 159 (412) | 10,461 | 9587 | 8.3–7 | 9.9–7 | − 6.2–7 | − 1.3–7 | 1:18:11 | 1:04:41 |

bqp500-7 | 501 ; 374,250 | 501 | 92 (223) | 8585 | 9066 | 8.1–7 | 9.9–7 | 4.7–8 | − 1.1–7 | 1:00:52 | 1:00:35 |

bqp500-8 | 501 ; 374,250 | 501 | 68 (140) | 5828 | 7604 | 6.7–7 | 9.9–7 | − 4.7–8 | − 1.1–7 | 40:56 | 51:58 |

bqp500-9 | 501 ; 374,250 | 501 | 50 (108) | 4704 | 11,613 | 9.5–7 | 9.9–7 | − 3.7–7 | − 9.8–8 | 34:05 | 1:21:17 |

bqp500-10 | 501 ; 374,250 | 501 | 71 (163) | 6462 | 8474 | 8.7–7 | 9.9–7 | − 6.2–7 | − 8.7–8 | 48:07 | 57:33 |

gka1e | 201 ; 59,700 | 201 | 74 (163) | 5352 | 9071 | 9.2–7 | 9.9–7 | − 3.0–7 | − 2.9–7 | 7:59 | 9:35 |

gka2e | 201 ; 59,700 | 201 | 49 (98) | 4008 | 6659 | 9.2–7 | 9.9–7 | 5.0–8 | − 1.7–7 | 4:17 | 6:29 |

gka3e | 201 ; 59,700 | 201 | 35 (71) | 2731 | 4103 | 8.3–7 | 9.7–7 | 2.3–7 | − 2.2–8 | 2:59 | 4:14 |

gka4e | 201 ; 59,700 | 201 | 34 (68) | 2999 | 3430 | 9.9–7 | 9.9–7 | − 1.7–7 | − 4.6–7 | 3:20 | 3:21 |

gka5e | 201 ; 59,700 | 201 | 43 (90) | 3367 | 2712 | 9.9–7 | 9.9–7 | − 4.9–8 | − 6.5–8 | 3:54 | 2:47 |

### 6.3 Evaluation of Qsdpnal on instances generated from QAP problems

*i*,

*j*)-th block of

*X*when it is partitioned uniformly into an \(l\times l\) block matrix with each block having dimension \(l\times l\). The convex set \(\mathcal{K}= \{X \in \mathcal{S}^{n} \mid X \ge 0 \}\),

*E*is the matrix of ones, and \(\delta _{ij} = 1\) if \(i=j\), and 0 otherwise. Note that here we use the same self-adjoint positive semidefinite linear operator \(\mathcal{Q}:\mathcal{S}^n\rightarrow \mathcal{S}^n\) constructed in (37). In our numerical experiments, the test instances \((A_1,A_2)\) are taken from the QAP Library [5].

Same as Table 1 but for QSDP-QAP problems

Problem | \(m_E\) | | iter.a | iter.b | \(\eta _{qsdp }\) | \(\eta _gap \) | Time |
---|---|---|---|---|---|---|

it (subs)|itSCB | a|b | a|b | a|b | |||

chr12a | 232 ; 144 | 45 (239) | 1969 | 50,000 | 9.9–7 | \(\textit{2.2}{-}{} \textit{6}\) | \({-}{} \textit{6.0}{-}{} \textit{6} | {-}{} \textit{2.5}{-}{} \textit{5}\) | 41 | 6:34 |

chr12b | 232 ; 144 | 56 (324) | 2428 | 50,000 | 9.9–7 | \(\textit{3.7}{-}{} \textit{6}\) | \({-}{} \textit{2.0}{-}{} \textit{5} | {-}{} \textit{6.0}{-}{} \textit{5}\) | 50 | 6:27 |

chr12c | 232 ; 144 | 56 (358) | 2201 | 50,000 | 9.9–7 | \(\textit{4.5}{-}{} \textit{6}\) | \({-}{} \textit{1.6}{-}{} \textit{5} | {-}{} \textit{6.1}{-}{} \textit{5}\) | 46 | 6:27 |

chr15a | 358 ; 225 | 84 (648) | 2866 | 50,000 | 9.9–7 | \(\textit{5.7}{-}{} \textit{6}\) | \({-}{} \textit{1.6}{-}{} \textit{5} | {-}{} { 1.0{-}4}\) | 2:25 | 12:23 |

chr15b | 358 ; 225 | 90 (584) | 4700 | 50,000 | 9.9–7 | \(\textit{7.3}{-}{} \textit{6}\) | \({-}{} \textit{1.3}{-}{} \textit{5} | {-}{ 1.3{-}4}\) | 3:07 | 12:31 |

chr15c | 358 ; 225 | 65 (425) | 2990 | 50,000 | 9.9–7 | \(\textit{6.8}{-}{} \textit{6}\) | \({-}{} \textit{2.4}{-}{} \textit{5} | {-}{} \textit{9.7}{-}{} \textit{5}\) | 1:58 | 12:40 |

chr18a | 511 ; 324 | 256 (1957) | 6003 | 50,000 | 7.3–7 | \(\textit{6.1}{-}{} \textit{6}\) | \({-}{} \textit{1.8}{-}{} \textit{5} | {-}{} { 1.3{-}4}\) | 14:34 | 22:26 |

chr18b | 511 ; 324 | 86 (565) | 3907 | 50,000 | 9.9–7 | \(\textit{8.4}{-}{} \textit{6}\) | \({-}{} \textit{1.8}{-}{} \textit{5} | {-}{ 1.6{-}4}\) | 5:26 | 22:19 |

chr20a | 628 ; 400 | 39 (274) | 1751 | 4133 | 9.5–7 | \(\textit{9.7}{-}{} \textit{7}\) | \({-}{} \textit{3.3}{-}{} \textit{5} | {-}{} \textit{3.4}{-}{} \textit{5}\) | 4:50 | 5:46 |

chr20b | 628 ; 400 | 72 (490) | 4044 | 50,000 | 9.6–7 | \(\textit{9.1}{-}{} \textit{6}\) | \({-}{} \textit{3.7}{-}{} \textit{5} | {-}{ 1.4{-}4}\) | 12:30 | 58:57 |

chr20c | 628 ; 400 | 144 (981) | 5242 | 50,000 | 9.9–7 | \(\textit{1.4}{-}{} \textit{5}\) | \({-}{} \textit{3.1}{-}{} \textit{5} | {-}{ 3.1{-}4}\) | 21:56 | 55:41 |

chr22a | 757 ; 484 | 67 (473) | 2804 | 50,000 | 9.9–7 | \(\textit{5.0}{-}{} \textit{6}\) | \({-}{} \textit{1.0}{-}{} \textit{5} | {-}{} \textit{7.6}{-}{} \textit{5}\) | 13:49 | 1:21:01 |

chr22b | 757 ; 484 | 69 (505) | 3581 | 50,000 | 9.9–7 | \(\textit{6.5}{-}{} \textit{6}\) | \({-}{} \textit{1.2}{-}{} \textit{5} | {-}{ 1.1{-}4}\) | 17:12 | 1:19:48 |

els19 | 568 ; 361 | 43 (403) | 2437 | 50,000 | 9.8–7 | \(\textit{1.2}{-}{} \textit{6}\) | \({-}{} \textit{4.2}{-}{} \textit{6} | {-}{} \textit{8.6}{-}{} \textit{6}\) | 4:39 | 1:04:08 |

esc16a | 406 ; 256 | 86 (506) | 5446 | 50,000 | 9.9–7 | \(\textit{8.3}{-}{} \textit{6}\) | \({-}{} \textit{2.5}{-}{} \textit{5} | {-}{ 1.3{-}4}\) | 4:47 | 16:54 |

esc16b | 406 ; 256 | 157 (1425) | 9222 | 50,000 | 9.9–7 | \(\textit{1.3}{-}{} \textit{5}\) | \({-}{} \textit{3.7}{-}{} \textit{5} | {-}{ 2.7{-}4}\) | 11:18 | 16:56 |

esc16c | 406 ; 256 | 188 (1404) | 13,806 | 50,000 | 9.9–7 | \(\textit{1.2}{-}{} \textit{5}\) | \({-}{} \textit{4.6}{-}{} \textit{5} | {-}{ 3.5{-}4}\) | 14:29 | 16:57 |

esc16d | 406 ; 256 | 101 (603) | 8043 | 50,000 | 9.9–7 | \(\textit{4.8}{-}{} \textit{6}\) | \({-}{} \textit{1.1}{-}{} \textit{5} | {-}{} \textit{7.4}{-}{} \textit{5}\) | 6:13 | 16:55 |

esc16e | 406 ; 256 | 110 (847) | 4286 | 50,000 | 9.9–7 | \(\textit{4.8}{-}{} \textit{6}\) | \({-}{} \textit{1.4}{-}{} \textit{5} | {-}{} \textit{5.6}{-}{} \textit{5}\) | 5:50 | 16:35 |

esc16g | 406 ; 256 | 85 (581) | 3818 | 50,000 | 9.9–7 | \(\textit{7.2}{-}{} \textit{6}\) | \({-}{} \textit{2.2}{-}{} \textit{5} | {-}{} \textit{9.4}{-}{} \textit{5}\) | 4:23 | 16:44 |

esc16h | 406 ; 256 | 228 (1732) | 11,733 | 50,000 | 8.6–7 | \(\textit{8.6}{-}{} \textit{6}\) | \({-}{} \textit{9.3}{-}{} \textit{6} | {-}{} \textit{8.7}{-}{} \textit{5}\) | 13:58 | 16:21 |

esc16i | 406 ; 256 | 41 (307) | 3165 | 50,000 | 9.6–7 | \(\textit{4.6}-\textit{6}\) | \({-}{} \textit{2.0}{-}{} \textit{5} | {-}{} \textit{6.0}{-}{} \textit{5}\) | 2:28 | 16:54 |

esc16j | 406 ; 256 | 163 (1179) | 5603 | 50,000 | 9.9–7 | \({ 6.6}{-}{} { 6}\) | \({-}{} { 2.0}{-}{} { 5} | {-}{} { 1.0{-}4}\) | 8:03 | 16:24 |

esc32b | 1582 ; 1024 | 80 (456) | 5026 | 50,000 | 9.9–7 | \({ 1.5{-}4}\) | \({-}{} { 2.9}{-}{} { 5} | {-}{} { 5.3{-}4}\) | 1:53:02 | 7:42:25 |

esc32c | 1582 ; 1024 | 105 (667) | 4203 | 50,000 | 8.9–7 | \({ 7.2}{-}{} { 6}\) | \({-}{} { 1.0}{-}{} { 5} | {-}{} { 7.3}{-}{} { 5}\) | 2:40:09 | 7:36:49 |

esc32d | 1582 ; 1024 | 141 (909) | 4852 | 50,000 | 9.9–7 | \({ 5.7}{-}{} { 6}\) | \({-}{} { 8.6}{-}{} { 6} | {-}{} { 6.4}{-}{} { 5}\) | 3:09:45 | 7:19:17 |

had12 | 232 ; 144 | 53 (320) | 2903 | 50,000 | 9.3–7 | \({ 3.3}{-}{} { 6}\) | \({-}{} { 7.2}{-}{} { 6} | {-}{} { 2.7}{-}{} { 5}\) | 55 | 6:52 |

had14 | 313 ; 196 | 60 (443) | 3634 | 50,000 | 9.9–7 | \({ 4.5}{-}{} { 6}\) | \({-}{} { 6.3}{-}{} { 6} | {-}{} { 2.8}{-}{} { 5}\) | 1:50 | 11:21 |

had16 | 406 ; 256 | 208 (1616) | 7604 | 50,000 | 8.1–7 | \(\textit{9.9}{-}{} \textit{6}\) | \({-}{} { 4.7}{-}{} { 6} | {-}{} { 9.1}{-}{} { 5}\) | 10:49 | 16:54 |

had18 | 511 ; 324 | 82 (537) | 4367 | 50,000 | 9.9–7 | \({ 9.2}{-}{} { 6}\) | \({-}{} { 1.5}{-}{} { 5} | {-}{} { 7.3}{-}{} { 5}\) | 6:10 | 24:46 |

had20 | 628 ; 400 | 121 (848) | 5024 | 50,000 | 9.5–7 | \(\textit{1.0}{-}{} \textit{5}\) | \({-}{} { 1.2}{-}{} { 5} | {-}{} { 1.0{-}4}\) | 20:30 | 51:08 |

kra30a | 1393 ; 900 | 107 (674) | 4665 | 50,000 | 9.5–7 | \({ 6.5}{-}{} { 6}\) | \({-}{} { 6.7}{-}{} { 5} | {-}{} { 1.7{-}4}\) | 1:51:20 | 7:23:56 |

kra30b | 1393 ; 900 | 107 (674) | 4853 | 50,000 | 9.9–7 | \({ 6.5}{-}{} { 6}\) | \({-}{} { 5.7}{-}{} { 5} | {-}{} { 1.7{-}4}\) | 2:00:03 | 8:02:08 |

kra32 | 1582 ; 1024 | 106 (636) | 6875 | 50,000 | 9.9–7 | \({ 7.4}{-}{} { 6}\) | \({-}{} { 3.6}{-}{} { 5} | {-}{} { 1.6{-}4}\) | 2:47:33 | 10:28:26 |

lipa30a | 1393 ; 900 | 64 (451) | 2924 | 50,000 | 9.9–7 | \({ 5.6}{-}{} { 6}\) | \({-}{} { 6.8}{-}{} { 6} | {-}{} { 3.1}{-}{} { 5}\) | 1:10:12 | 6:52:34 |

lipa30b | 1393 ; 900 | 257 (1918) | 7507 | 50,000 | 9.9–7 | \({ 7.0}{-}{} { 6}\) | \({-}{} { 1.9}{-}{} { 6} | {-}{} { 1.9{-}4}\) | 4:28:38 | 7:18:10 |

lipa40a | 2458 ; 1600 | 51 (349) | 2193 | 50,000 | 7.7–7 | \({ 4.2}{-}{} { 6}\) | \({-}{} { 2.3}{-}{} { 6} | {-}{} { 2.0}{-}{} { 5}\) | 3:03:58 | 23:53:41 |

lipa40b | 2458 ; 1600 | 156 (1339) | 4750 | 50,000 | 9.1–7 | \({ 3.9}{-}{} { 6}\) | \({ 6.0}{-}{} { 6} | {-}{} { 8.9}{-}{} { 5}\) | 9:36:10 | 18:57:01 |

nug12 | 232 ; 144 | 84 (478) | 4068 | 50,000 | 9.8–7 | \({ 5.4}{-}{} { 6}\) | \({-}{} { 3.0}{-}{} { 5} | {-}{} { 1.1{-}4}\) | 1:32 | 6:34 |

nug14 | 313 ; 196 | 93 (610) | 4953 | 50,000 | 9.7–7 | \({ 6.9}{-}{} { 6}\) | \({-}{} { 2.6}{-}{} { 5} | {-}{} { 1.1{-}4}\) | 3:12 | 10:27 |

nug15 | 358 ; 225 | 102 (660) | 5627 | 50,000 | 7.0–7 | \({ 1.1}{-}{} { 5}\) | \({-}{} { 2.2}{-}{} { 5} | {-}{} { 1.7{-}4}\) | 4:12 | 12:42 |

nug16a | 406 ; 256 | 86 (530) | 4945 | 50,000 | 9.9–7 | \({ 7.2}{-}{} { 6}\) | \({-}{} { 2.3}{-}{} { 5} | {-}{} { 1.1{-}4}\) | 4:39 | 18:17 |

nug16b | 406 ; 256 | 97 (631) | 4777 | 50,000 | 9.9–7 |\({ 1.2}{-}{} { 5}\) | \({-}{} { 2.5}{-}{} { 5} | {-}{} { 2.0{-}4}\) | 5:19 | 19:06 |

nug17 | 457 ; 289 | 110 (772) | 5365 | 50,000 | 9.9–7 |\({ 1.3}{-}{} { 5}\) | \({-}{} { 2.4}{-}{} { 5} | {-}{} { 1.8{-}4}\) | 7:41 | 22:47 |

nug18 | 511 ; 324 | 85 (559) | 4367 | 50,000 | 9.9–7 |\({ 6.1}{-}{} { 6}\) | \({-}{} { 3.3}{-}{} { 5} | {-}{} { 9.9}{-}{} { 5}\) | 6:10 | 26:20 |

nug20 | 628 ; 400 | 114 (746) | 5220 | 50,000 | 9.9–7 |\({ 8.6}{-}{} { 6}\) | \({-}{} { 2.3}{-}{} { 5} | {-}{} { 1.3{-}4}\) | 19:25 | 55:36 |

nug21 | 691 ; 441 | 84 (569) | 4322 | 50,000 | 9.7–7 |\({ 6.8}{-}{} { 6}\) | \({-}{} { 4.0}{-}{} { 5} | {-}{} { 1.1{-}4}\) | 18:48 | 1:09:39 |

nug22 | 757 ; 484 | 121 (822) | 5822 | 50,000 | 9.6–7 |\({ 8.3{-}6}\) | \({-}{} { 4.1}{-}{} { 5} | {-}{} { 1.3{-}4}\) | 34:03 | 2:08:02 |

nug24 | 898 ; 576 | 89 (542) | 4345 | 50,000 | 9.9–7 |\({ 6.5{-}6}\) | \({-}{} { 3.2}{-}{} { 5} | {-}{} { 1.1{-}4}\) | 34:08 | 3:04:07 |

nug25 | 973 ; 625 | 129 (860) | 5801 | 50,000 | 9.9–7 |\({ 6.9{-}6}\) | \({-}{} { 2.2}{-}{} { 5} | {-}{} { 1.1{-}4}\) | 1:09:12 | 3:41:22 |

nug27 | 1132 ; 729 | 148 (951) | 8576 | 50,000 | 9.9–7 |\({ 8.3{-}6}\) | \({-}{} { 2.6}{-}{} { 5} | {-}{} { 1.3{-}4}\) | 2:09:22 | 4:59:06 |

nug28 | 1216 ; 784 | 119 (758) | 6389 | 50,000 | 9.7–7 |\({ 7.8{-}6}\) | \({-}{} { 2.9}{-}{} { 5} | {-}{} { 1.1{-}4}\) | 1:43:52 | 5:50:50 |

nug30 | 1393 ; 900 | 105 (777) | 4912 | 50,000 | 9.9–7 |\({ 9.3{-}6}\) | \({-}{} { 2.6}{-}{} { 5} | {-}{} { 1.2{-}4}\) | 2:08:56 | 7:40:43 |

rou12 | 232 ; 144 | 78 (418) | 6600 | 50,000 | 9.9–7 |\({ 4.4{-}6}\) | \({-}{} { 2.2}{-}{} { 5} | {-}{} { 9.3}{-}{} { 5}\) | 2:13 | 6:36 |

rou15 | 358 ; 225 | 106 (639) | 5952 | 50,000 | 9.9–7 |\({ 5.4{-}6}\) | \({-}{} { 2.7}{-}{} { 5} | {-}{} { 1.0{-}4}\) | 4:12 | 13:02 |

rou20 | 628 ; 400 | 65 (359) | 4238 | 50,000 | 9.9–7 |\({ 3.6{-}6}\) | \({-}{} { 2.5}{-}{} { 5} | {-}{} { 6.1}{-}{} { 5}\) | 11:53 | 1:00:07 |

scr12 | 232 ; 144 | 56 (295) | 2205 | 50,000 | 9.9–7 |\({ 3.2{-}6}\) | \({-}{} { 7.5}{-}{} { 6} | {-}{} { 4.0}{-}{} { 5}\) | 43 | 6:57 |

scr15 | 358 ; 225 | 121 (769) | 5730 | 50,000 | 7.4–7 |\({ 1.0{-}5}\) | \({-}{} { 2.1}{-}{} { 5} | {-}{} { 1.9{-}4}\) | 4:39 | 12:45 |

scr20 | 628 ; 400 | 89 (590) | 5621 | 50,000 | 9.9–7 |\({ 8.3{-}6}\) | \({-}{} { 4.1}{-}{} { 5} | {-}{} { 1.6{-}4}\) | 18:49 | 1:01:19 |

tai12a | 232 ; 144 | 110 (807) | 6090 | 50,000 | 9.9–7 |\({ 8.9{-}6}\) | \({-}{} { 1.8}{-}{} { 5} | {-}{} { 1.3{-}4}\) | 2:40 | 6:15 |

tai12b | 232 ; 144 | 123 (856) | 6323 | 50,000 | 8.6–7 |\({ 7.3{-}6}\) | \({-}{} { 2.5}{-}{} { 5} | {-}{} { 1.1{-}4}\) | 2:43 | 6:19 |

tai15a | 358 ; 225 | 67 (405) | 4301 | 50,000 | 9.4–7 |\({ 3.0{-}6}\) | \({-}{} { 2.8}{-}{} { 5} | {-}{} { 5.9}{-}{} { 5}\) | 2:48 | 13:09 |

tai17a | 457 ; 289 | 95 (569) | 6142 | 50,000 | 9.9–7 |\({ 3.8{-}6}\) | \({-}{} { 2.0}{-}{} { 5} | {-}{} { 6.5}{-}{} { 5}\) | 6:33 | 19:23 |

tai20a | 628 ; 400 | 87 (498) | 4762 | 50,000 | 9.7–7 | \({ 3.0}{-}{} { 6}\) | \({-}{} { 2.2}{-}{} { 5} | {-}{} { 5.4}{-}{} { 5}\) | 15:00 | 1:00:55 |

tai25a | 973 ; 625 | 25 (138) | 3438 | 10,084 | 9.9–7 | \({ 9.9}{-}{} { 7}\) | \({ 9.4}{-}{} { 6} | {-}{} { 1.5}{-}{} { 5}\) | 17:46 | 47:55 |

tai25b | 973 ; 625 | 164 (1219) | 7803 | 50,000 | 9.8–7 | \({ 1.4}{-}{} { 5}\) | \({-}{} { 4.8}{-}{} { 5} | {-}{} { 2.6{-}4}\) | 1:33:02 | 3:33:13 |

tai30a | 1393 ; 900 | 97 (546) | 4270 | 50,000 | 7.8–7 | \({ 2.8}{-}{} { 6}\) | \({-}{} { 1.4}{-}{} { 5} | {-}{} { 4.1}{-}{} { 5}\) | 1:22:01 | 7:21:02 |

tai30b | 1393 ; 900 | 162 (1229) | 6845 | 50,000 | 9.9–7 | \({ 1.2}{-}{} { 5}\) | \({-}{} { 3.3}{-}{} { 5} | {-}{} { 2.1{-}4}\) | 3:10:41 | 7:21:29 |

tai35a | 1888 ; 1225 | 95 (537) | 5781 | 50,000 | 9.9–7 | \({ 2.8}{-}{} { 6}\) | \({-}{} { 9.2}{-}{} { 6} | {-}{} { 3.5}{-}{} { 5}\) | 3:34:56 | 15:04:40 |

tai35b | 1888 ; 1225 | 152 (1195) | 5303 | 50,000 | 9.6–7 | \({ 1.2}{-}{} { 5}\) | \({-}{} { 3.8}{-}{} { 5} | {-}{} { 1.9{-}4}\) | 5:56:59 | 13:37:28 |

tai40a | 2458 ; 1600 | 79 (398) | 6381 | 50,000 | 9.1–7 | \({ 3.0}{-}{} { 6}\) | \({-}{} { 1.7}{-}{} { 5} | {-}{} { 3.4}{-}{} { 5}\) | 6:01:56 | 23:07:21 |

tho30 | 1393 ; 900 | 116 (761) | 5468 | 50,000 | 8.9–7 | \({ 8.0}{-}{} { 6}\) | \({-}{} { 3.1}{-}{} { 5} | {-}{} { 1.3{-}4}\) | 2:15:55 | 7:39:54 |

tho40 | 2458 ; 1600 | 122 (762) | 3834 | 50,000 | 9.9–7 | \({ 7.4}{-}{} { 6}\) | \({-}{} { 2.9}{-}{} { 5} | {-}{} { 1.0{-}4}\) | 6:28:52 | 26:40:25 |

### 6.4 Evaluation of Qsdpnal on instances generated from sensor network localization problems

*m*anchors and

*l*sensors:

*i*th unit vector in \(\mathfrak {R}^l\), and \(I_d\) is the \(d\times d\) identity matrix. Let \(g_{ik} = a_{ik}\) for \((i,k)\in \mathcal{M}\), \(g_{ij} = [e_{ij}; \mathbf{0}_m]\) for \((i,j)\in \mathcal{N}\), and

*l*points \(\{{\hat{x}}_{i}\in \mathfrak {R}^d \mid i=1,\ldots ,l\}\) in \([-0.5,0.5]^d\). Then, the edge set \(\mathcal{N}\) is generated by considering only pairs of points that have distances less than a given positive number

*R*, i.e.,

*m*anchors \(\{ a_{k}\in \mathfrak {R}^d \mid k=1,\ldots ,m\}\), the edge set \(\mathcal{M}\) is similarly given by

*i*and

*j*, \(\varepsilon _{ij}\) are assumed to be independent standard Normal random variables, \(\tau \) is the noise parameter. For the numerical experiments, we generate 10 instances where the number of sensors

*l*ranges from 250 to 1500 and the dimension

*d*is set to be 2 or 3. We set the noise factor \(\tau = 10\%\). The 4 anchors for the two dimensional case (\(d=2\)) are placed at

*E*is the \(3\times 3\) matrix of all ones.

| \(m_E\) | | iter.a | iter.b | \(\eta _{qsdp }\) | \(\eta _gap \) | Time |
---|---|---|---|---|---|---|

it (subs)|itSCB | a|b | a|b | a|b | |||

2 | 452 | 252 | 0.50 | 12 (74) | 652 | 24,049 | 7.1–7 | 9.9–7 | − 9.0–7 | 6.8–7 | 47 | 6:10 |

2 | 548 | 502 | 0.36 | 12 (62) | 1000 | 12,057 | 7.6–7 | 9.9–7 | \({-}{} { 9.2}{-}{} { 6} | {-}{} { 9.1}{-}{} { 6}\) | 1:49 | 17:25 |

2 | 633 | 802 | 0.28 | 17 (85) | 1000 | 27,361 | 3.0–7 | 9.9–7 | \({-}{} { 2.4}{-}{} { 6} | {-}{} { 9.9}{-}{} { 6}\) | 5:42 | 1:59:16 |

2 | 684 | 1002 | 0.25 | 17 (94) | 1000 | 50,000 | 4.1–7 | \({ 1.4}{{-}}{} { 5}\) | \({-}{} { 2.6}{-}{} { 6} | {-}{} { 3.0}{-}{} { 7}\) | 10:18 | 6:16:37 |

2 | 781 | 1502 | 0.21 | 21 (104) | 1000 | 50,000 | 3.6–7 | \({ 9.5{-}4}\) | \({-}{} { 6.3}{-}{} { 6} | { 5.1{-}3}\) | 23:05 | 13:47:39 |

2 | 774 | 2002 | 0.18 | 29 (156) | 1000 | 50,000 | 7.3–7 | \({ 2.1{-}3}\) | \({-}{} { 3.8}{-}{} { 6} | { 1.4{-}2}\) | 49:53 | 23:20:28 |

3 | 395 | 253 | 0.49 | 11 (31) | 408 | 1487 | 9.8–7 | 9.7–7 | \({-}{} { 1.4}{-}{} { 6} | { 1.3}{-}{} { 7}\) | 06 | 18 |

3 | 503 | 503 | 0.39 | 14 (61) | 877 | 7882 | 3.5–7 | 9.9–7 | \({-}{} { 1.1}{-}{} { 6} | { 2.5}{-}{} { 6}\) | 1:46 | 7:18 |

3 | 512 | 803 | 0.33 | 15 (85) | 1000 | 10,579 | 7.7–7 | 9.9–7 | \({-}{} { 1.3}{-}{} { 6} | { 4.2}{-}{} { 7}\) | 7:13 | 26:15 |

3 | 513 | 1003 | 0.31 | 16 (71) | 1000 | 14,025 | 2.5–7 | 9.9–7 | \({-}{} { 7.6}{-}{} { 7} | { 4.2}{-}{} { 6}\) | 8:46 | 1:02:07 |

3 | 509 | 1503 | 0.27 | 19 (83) | 1000 | 23,328 | 8.6–7 | 9.9–7 | \({-}{} { 4.5}{-}{} { 6} | { 7.2}{-}{} { 6}\) | 28:34 | 4:13:07 |

3 | 505 | 2003 | 0.24 | 19 (97) | 1000 | 50,000 | 9.5–7 | \({ 3.3{-}4}\) | \({-}{} { 1.4}{-}{} { 5} | {-}{} { 3.7{-}4}\) | 49:28 | 16:45:24 |

| \(m_E;m_I\) | | iter.a | iter.b | \(\eta _{qsdp }\) | \(\eta _gap \) | Time |
---|---|---|---|---|---|---|

it (subs)|itSCB | a|b | a|b | a|b | |||

2 | 452 ; 14,402 | 252 | 0.50 | 15 (119) | 603 | 50,000 | 6.4–7 | \(\textit{1.9-6}\) | − 9.1–7 | − 5.1–8 | 1:50 | 24:28 |

2 | 548 ; 55,849 | 502 | 0.36 | 16 (180) | 1357 | 29,565 | 4.4–7 | 9.9–7 | \({-}\)2.2–7 | \({-}{} { 2.3}{-}{} { 6}\) | 11:28 | 1:09:24 |

2 | 633 ; 118,131 | 802 | 0.28 | 15 (226) | 2330 | 36,651 | 6.3–7 | 9.9–7 | \({-}{} { 2.4}{-}{} { 6} | {-}{} { 5.4}{-}{} { 6}\) | 1:06:04 | 3:43:51 |

2 | 684 ; 160,157 | 1002 | 0.25 | 20 (265) | 3384 | 50,000 | 4.5–7 | \({ 2.7}{-}{} { 6}\) | \({-}{} { 1.7}{-}{} { 6} | { 2.2}{-}{} { 6}\) | 2:36:41 | 8:30:28 |

2 | 724 ; 201,375 | 1202 | 0.23 | 21 (487) | 3115 | 50,000 | 9.8–7 | \({ 3.8}{-}{} { 6}\) | \({ 5.2}{-}{} { 6} | {-}{} { 2.8}{-}{} { 6}\) | 6:36:28 | 11:57:56 |

3 | 395 ; 16,412 | 253 | 0.49 | 12 (88) | 471 | 2897 | 3.1–7 | 9.8–7 | \({-}\)3.3–7 | \({-}\)5.7–7 | 46 | 1:18 |

3 | 503 ; 53,512 | 503 | 0.39 | 14 (136) | 949 | 11,003 | 4.4–7 | 9.9–7 | \({-}\)9.9–7 | \({-}\)3.1–7 | 8:23 | 20:26 |

3 | 512 ; 104,071 | 803 | 0.33 | 17 (145) | 1762 | 14,144 | 6.6–7 | 9.9–7 | \({-}{} { 2.6}{-}{} { 6}\) | 6.0–7 | 32:17 | 1:09:10 |

3 | 513 ; 139,719 | 1003 | 0.31 | 21 (198) | 2406 | 31,832 | 5.1–7 | 9.9–7 | 8.4–8 | 4.5–8 | 1:18:56 | 4:48:04 |

3 | 526 ; 180,236 | 1203 | 0.29 | 21 (250) | 2639 | 19,010 | 8.3–7 | 9.9–7 | \({-}\)3.8–8 | \({ 1.1}{-}{} { 5}\) | 2:15:54 | 4:16:09 |

*i*th sensor. To further test our algorithm Qsdpnal, we also generate the following valid inequalities and add them to problem (38)

## 7 Conclusions

We have designed a two-phase augmented Lagrangian based method, called Qsdpnal, for solving large scale convex quadratic semidefinite programming problems. The global and local convergence rate analysis of our algorithm is based on the classic results of proximal point algorithms [25, 26], together with the recent advances in second order variational analysis of convex composite quadratic semidefinite programming [7]. By devising novel numerical linear algebra techniques, we overcome various challenging numerical difficulties encountered in the efficient implementation of Qsdpnal. Numerical experiments on various large scale QSDPs have demonstrated the efficiency and robustness of our proposed two-phase framework in obtaining accurate solutions. Specifically, for well-posed problems, our Qsdpnal-Phase I is already powerful enough and it is not absolutely necessary to execute Qsdpnal-Phase II. On the other hand, for more difficult problems, the purely first-order Qsdpnal-Phase I algorithm may stagnate because of extremely slow local convergence. In contrast, with the activation of Qsdpnal-Phase II which has second order information wisely incorporated, our Qsdpnal algorithm can still obtain highly accurate solutions efficiently.

## Footnotes

## References

- 1.Alfakih, A.Y., Khandani, A., Wolkowicz, H.: Solving Euclidean distance matrix completion problems via semidefinite programming. Comput. Optim. Appl.
**12**, 13–30 (1999)MathSciNetCrossRefzbMATHGoogle Scholar - 2.Bauschke, H.H., Borwein, J.M., Li, W.: Strong conical hull intersection property, bounded linear regularity, Jameson’s property (G), and error bounds in convex optimization. Math. Program.
**86**, 135–160 (1999)MathSciNetCrossRefzbMATHGoogle Scholar - 3.Biswas, P., Liang, T.C., Toh, K.-C., Wang, T.C., Ye, Y.: Semidefinite programming approaches for sensor network localization with noisy distance measurements. IEEE Trans. Autom. Sci. Eng.
**3**, 360–371 (2006)CrossRefGoogle Scholar - 4.Bonnans, J.F., Shapiro, A.: Perturbation Analysis of Optimization Problems. Springer, New York (2000)CrossRefzbMATHGoogle Scholar
- 5.Burkard, R.E., Karisch, S.E., Rendl, F.: QAPLIB—a quadratic assignment problem library. J. Global Optim.
**10**, 391–403 (1997)MathSciNetCrossRefzbMATHGoogle Scholar - 6.Chen, L., Sun, D.F., Toh, K.-C.: An efficient inexact symmetric Gauss–Seidel based majorized ADMM for high-dimensional convex composite conic programming. Math. Program.
**161**, 237–270 (2017)MathSciNetCrossRefzbMATHGoogle Scholar - 7.Cui, Y., Sun, D.F., Toh, K.-C.: On the asymptotic superlinear convergence of the augmented Lagrangian method for semidefinite programming with multiple solutions, arXiv:1610.00875 (2016)
- 8.Clarke, F.: Optimization and Nonsmooth Analysis. Wiley, New York (1983)zbMATHGoogle Scholar
- 9.Facchinei, F., Pang, J.-S.: Finite-Dimensional Variational Inequalities and Complementarity Problems. Springer, New York (2003)zbMATHGoogle Scholar
- 10.Han, D., Sun, D., Zhang, L.: Linear rate convergence of the alternating direction method of multipliers for convex composite programming. Math. Oper. Res. (2017). https://doi.org/10.1287/moor.2017.0875
- 11.Higham, N.J.: Computing the nearest correlation matrix—a problem from finance. IMA J. Numer. Anal.
**22**, 329–343 (2002)MathSciNetCrossRefzbMATHGoogle Scholar - 12.Hiriart-Urruty, J.-B., Strodiot, J.-J., Nguyen, V.H.: Generalized Hessian matrix and second-order optimality conditions for problems with \({C}^{1,1}\) data. Appl. Math. Optim.
**11**, 43–56 (1984)MathSciNetCrossRefzbMATHGoogle Scholar - 13.Jiang, K., Sun, D.F., Toh, K.-C.: An inexact accelerated proximal gradient method for large scale linearly constrained convex SDP. SIAM J. Optim.
**22**, 1042–1064 (2012)MathSciNetCrossRefzbMATHGoogle Scholar - 14.Jiang, K., Sun, D.F., Toh, K.-C.: A partial proximal point algorithm for nuclear norm regularized matrix least squares problems. Math. Program. Comput.
**6**, 281–325 (2014)MathSciNetCrossRefzbMATHGoogle Scholar - 15.Krislock, N., Lang, J., Varah, J., Pai, D.K., Seidel, H.-P.: Local compliance estimation via positive semidefinite constrained least squares. IEEE Trans. Robot.
**20**, 1007–1011 (2004)CrossRefGoogle Scholar - 16.Li, L., Toh, K.-C.: An inexact interior point method for l1-regularized sparse covariance selection. Math. Program. Comput.
**2**, 291–315 (2010)MathSciNetCrossRefzbMATHGoogle Scholar - 17.Li, X.D.: A Two-Phase Augmented Lagrangian Method for Convex Composite Quadratic Programming, PhD thesis, Department of Mathematics, National University of Singapore (2015)Google Scholar
- 18.Li, X.D., Sun, D.F., Toh, K.-C.: A Schur complement based semi-proximal ADMM for convex quadratic conic programming and extensions. Math. Program.
**155**, 333–373 (2016)MathSciNetCrossRefzbMATHGoogle Scholar - 19.Nie, J.W., Yuan, Y.X.: A predictor-corrector algorithm for QSDP combining Dikin-type and Newton centering steps. Ann. Oper. Res.
**103**, 115–133 (2001)MathSciNetCrossRefzbMATHGoogle Scholar - 20.Pang, J.-S., Sun, D.F., Sun, J.: Semismooth homeomorphisms and strong stability of semidefinite and Lorentz complementarity problems. Math. Oper. Res.
**28**, 39–63 (2003)MathSciNetCrossRefzbMATHGoogle Scholar - 21.Povh, J., Rendl, F.: Copositive and semidefinite relaxations of the quadratic assignment problem. Discrete Optim.
**6**, 231–241 (2009)MathSciNetCrossRefzbMATHGoogle Scholar - 22.Qi, H.D.: Local duality of nonlinear semidefinite programming. Math. Oper. Res.
**34**, 124–141 (2009)MathSciNetCrossRefzbMATHGoogle Scholar - 23.Qi, H.D., Sun, D.F.: A quadratically convergent Newton method for computing the nearest correlation matrix. SIAM J. Matrix Anal. Appl.
**28**, 360–385 (2006)MathSciNetCrossRefzbMATHGoogle Scholar - 24.Rockafellar, R.T.: Conjugate Duality and Optimization, CBMS-NSF Regional Conf. Ser. Appl. Math. vol. 16. SIAM, Philadelphia (1974)Google Scholar
- 25.Rockafellar, R.T.: Monotone operators and the proximal point algorithm. SIAM J. Control Optim.
**14**, 877–898 (1976)MathSciNetCrossRefzbMATHGoogle Scholar - 26.Rockafellar, R.T.: Augmented Lagrangians and applications of the proximal point algorithm in convex programming. Math. Oper. Res.
**1**, 97–116 (1976)MathSciNetCrossRefzbMATHGoogle Scholar - 27.Sun, D.F., Sun, J.: Semismooth matrix-valued functions. Math. Oper. Res.
**27**, 150–169 (2002)MathSciNetCrossRefzbMATHGoogle Scholar - 28.Sun, D.F., Toh, K.-C., Yang, L.: A convergent 3-block semiproximal alternating direction method of multipliers for conic programming with 4-type constraints. SIAM J. Optim.
**25**, 882–915 (2015)MathSciNetCrossRefzbMATHGoogle Scholar - 29.Sun, D.F., Toh, K.-C., Yang, L.: An efficient inexact ABCD method for least squares semidefinite programming. SIAM J. Optim.
**26**, 1072–1100 (2016)MathSciNetCrossRefzbMATHGoogle Scholar - 30.Sun, J., Zhang, S.: A modified alternating direction method for convex quadratically constrained quadratic semidefinite programs. Eur. J. Oper. Res.
**207**, 1210–1220 (2010)MathSciNetCrossRefzbMATHGoogle Scholar - 31.Toh, K.-C.: An inexact primal-dual path following algorithm for convex quadratic SDP. Math. Program.
**112**, 221–254 (2008)MathSciNetCrossRefzbMATHGoogle Scholar - 32.Toh, K.-C., Tütüncü, R., Todd, M.: Inexact primal-dual path-following algorithms for a special class of convex quadratic SDP and related problems. Pac. J. Optim.
**3**, 135–164 (2007)MathSciNetzbMATHGoogle Scholar - 33.Yang, L., Sun, D.F., Toh, K.-C.: SDPNAL+: a majorized semismooth Newton-CG augmented Lagrangian method for semidefinite programming with nonnegative constraints. Math. Program. Comput.
**7**, 331–366 (2015)MathSciNetCrossRefzbMATHGoogle Scholar - 34.Zhao, X.Y.: A Semismooth Newton-CG Augmented Lagrangian Method for Large Scale Linear and Convex Quadratic SDPs, PhD thesis, Department of Mathematics, National University of Singapore (2009)Google Scholar
- 35.Zhao, X.Y., Sun, D.F., Toh, K.-C.: A Newton-CG augmented Lagrangian method for semidefinite programming. SIAM J. Optim.
**20**, 1737–1765 (2010)MathSciNetCrossRefzbMATHGoogle Scholar