## On Lagrangian relaxation of quadratic matrix constraints (2000)

### Cached

### Download Links

Venue: | SIAM J. Matrix Anal. Appl |

Citations: | 46 - 17 self |

### BibTeX

@ARTICLE{Anstreicher00onlagrangian,

author = {Kurt Anstreicher and Henry Wolkowicz},

title = {On Lagrangian relaxation of quadratic matrix constraints},

journal = {SIAM J. Matrix Anal. Appl},

year = {2000},

pages = {41--55}

}

### Years of Citing Articles

### OpenURL

### Abstract

Abstract. Quadratically constrained quadratic programs (QQPs) play an important modeling role for many diverse problems. These problems are in general NP hard and numerically intractable. Lagrangian relaxations often provide good approximate solutions to these hard problems. Such relaxations are equivalent to semidefinite programming relaxations. For several special cases of QQP, e.g., convex programs and trust region subproblems, the Lagrangian relaxation provides the exact optimal value, i.e., there is a zero duality gap. However, this is not true for the general QQP, or even the QQP with two convex constraints, but a nonconvex objective. In this paper we consider a certain QQP where the quadratic constraints correspond to the matrix orthogonality condition XXT = I. For this problem we show that the Lagrangian dual based on relaxing the constraints XXT = I and the seemingly redundant constraints XT X = I has a zero duality gap. This result has natural applications to quadratic assignment and graph partitioning problems, as well as the problem of minimizing the weighted sum of the largest eigenvalues of a matrix. We also show that the technique of relaxing quadratic matrix constraints can be used to obtain a strengthened semidefinite relaxation for the max-cut problem. Key words. Lagrangian relaxations, quadratically constrained quadratic programs, semidefinite programming, quadratic assignment, graph partitioning, max-cut problems

### Citations

974 | Improved approximation algorithms for maximum cut and satisfiability problems using semidefinite programming
- Goemans, Williamson
- 1995
(Show Context)
Citation Context ... provide a tractable approach for finding good bounds for many hard combinatorial problems. The best example is the application of SDP to the max-cut problem, where a 87% performance guarantee exists =-=[11, 12]-=-. Other examples include matrix completion problems [23, 22], as well as graph partitioning problems and the quadratic assignment problem (references given below). In this paper we consider several qu... |

499 | Interior point methods in semidefinite programming with applications to combinatorial optimization
- Alizadeh
- 1995
(Show Context)
Citation Context ...values of L+U . We will now show that the Donath-Hoffman bound can be obtained by applying Lagrangian relaxation to an appropriate QQP relaxation of GP. (An SDP formulation for this bound is given in =-=[1]-=-.) Clearly if P is a partition matrix then x T i x i = 1, i = 1; : : : ; n, where x T i is the ith row of X . Moreover the columns of X are orthogonal with one another, and the norm of the jth column ... |

426 | The geometry of algorithms with orthogonality constraints
- Edelman, Arias, et al.
- 1998
(Show Context)
Citation Context ... [10, 20]. 3 Orthogonally Constrained Programs with Zero Duality Gaps Consider the orthonormal constraint X T X = I ; X 2 M m;n : 6 (The set of such X is sometimes known as the Stiefel manifold, e.g. =-=[7]-=-. Applications and algorithms for optimization on orthonormal sets of matrices are discussed in [7].) In this section we will show that for m = n, strong duality holds for a certain (nonconvex) quadra... |

379 |
Interior-point polynomial algorithms in convex programming
- Nesterov, Nemirovskii
- 1994
(Show Context)
Citation Context ...s. Moreover these relaxations can be shown to be equivalent to semidefinite programming (SDP) relaxations, and SDP problems can be solved efficiently, i.e. they are polynomial time problems, see e.g. =-=[32]-=-. SDP relaxations provide a tractable approach for finding good bounds for many hard combinatorial problems. The best example is the application of SDP to the max-cut problem, where a 87% performance ... |

278 | Cones of matrices and setfunctions and 0-1 optimization
- Lovász, Schrijver
- 1991
(Show Context)
Citation Context ... strengthened SDP relaxations for the max-cut problem. Adding constraints to close the duality gap is akin to adding valid inequalities in cutting plane methods for discrete optimization problems. In =-=[2, 24]-=- this approach, inQUADRATIC MATRIX CONSTRAINTS 53 combination with a lifting procedure, is used to solve discrete optimization problems. In our case we add quadratic constraints. The idea of quadrati... |

234 |
Computing a trust region step
- Moré, Sorensen
- 1983
(Show Context)
Citation Context ...duality holds. Extensions of this result to a two-sided general, possibly nonconvex, constraint are discussed in [43, 29]. An algorithm based on Lagrangian duality appears in [40] and (implicitly) in =-=[30, 41]-=-. These algorithms are extremely efficient for the TRS problem, i.e. they solve this problem almost as quickly as an eigenvalue problem. 2.4 Two Trust Region Subproblem The two trust region subproblem... |

225 |
A lift-and-project cutting plane algorithm for mixed 0–1 programs
- Balas, Ceria, et al.
- 1993
(Show Context)
Citation Context ... strengthened SDP relaxations for the max-cut problem. Adding constraints to close the duality gap is akin to adding valid inequalities in cutting plane methods for discrete optimization problems. In =-=[2, 24]-=- this approach, inQUADRATIC MATRIX CONSTRAINTS 53 combination with a lifting procedure, is used to solve discrete optimization problems. In our case we add quadratic constraints. The idea of quadrati... |

219 | An interior-point method for semidefinite programming
- Helmberg, Rendl, et al.
- 1996
(Show Context)
Citation Context ...e Xs0; (2.3) which has proven to have very strong theoretical and practical properties, i.e. the bound has an 87% performance guarantee for the problem MC, and a 97% performance in practice, see e.g. =-=[12, 18, 15]-=-. Other theoretical results for general objectives and further relaxed constraints appear in [31, 46]. In [38], several unrelated bounds for MC are shown to be equivalent. These bounds include the box... |

147 | A spectral bundle method for semidefinite programming
- Helmberg, Rendl
(Show Context)
Citation Context ...theoretical properties. There have been attempts at further strengthening this relaxation. For example, a copositive relaxation is presented in [39]. Adding cuts to the SDP relaxation is discussed in =-=[15, 16, 17, 18]-=-. These improvements all involve heuristics such as which cuts to choose or solving a copositive problem, which is NP-hard in itself. The relaxation in (2.3) is obtained by lifting the vector x into m... |

146 |
Lower bounds for the partitioning of graphs
- Donath, Hoffman
(Show Context)
Citation Context ...he set of n \Theta k partition matrices (i.e. each column of X is the indicator function of the corresponding set; X ij = 1 if node i is in set j and 0 otherwise). The well known Donath-Hoffman bound =-=[6]-=- z DHsz for GP is z DH := max e T u=0 k X i=1 m isi (L + U); wheres1 (L+U)s2 (L+U)s: : :sn (L+U) are the eigenvalues of L+U . We will now show that the Donath-Hoffman bound can be obtained by applying... |

123 |
Quality of semidefinite relaxation for nonconvex quadratic optimization
- Nesterov
- 1997
(Show Context)
Citation Context ...s an 87% performance guarantee for the problem MC, and a 97% performance in practice, see e.g. [12, 18, 15]. Other theoretical results for general objectives and further relaxed constraints appear in =-=[31, 46]-=-. In [38], several unrelated bounds for MC are shown to be equivalent. These bounds include the box relaxation \Gammaesxse, the trust region relaxation P i x 2 i = n, and an eigenvalue relaxation. Fur... |

115 | The quadratic assignment problem
- Burkard, Çela, et al.
- 1998
(Show Context)
Citation Context ...atrices contains the permutation matrices, and the orthonormally constrained problem (3.1) is an important relaxation of QAP. The bounds obtained are usually called the eigenvalue bounds for QAP, see =-=[8, 13]-=-. Theorem 3.2 shows that the eigenvalue bounds are in fact obtained from a Lagrangian relaxation of (3.1), after adding the seemingly redundant constraint XX T = I . 4.2 Weighted Sums of Eigenvalues C... |

98 | Semidefinite programming in combinatorial optimization
- Goemans
- 1997
(Show Context)
Citation Context ... provide a tractable approach for finding good bounds for many hard combinatorial problems. The best example is the application of SDP to the max-cut problem, where a 87% performance guarantee exists =-=[11, 12]-=-. Other examples include matrix completion problems, [23, 22] as well as graph partitioning problems and the quadratic assignment problem (references given below). In this paper we consider several qu... |

96 | The quadratic assignment problem: A survey and recent developments
- Pardalos, Rendl, et al.
- 1994
(Show Context)
Citation Context ...ree applications of the above strong duality result. 4.1 Quadratic Assignment Problem Let A and B be n \Theta n symmetric matrices, and consider the homogeneous Quadratic Assignment Problem, see e.g. =-=[33]-=-, QAP min tr AXBX T s.t. X 2 \Pi; where \Pi is the set of n \Theta n permutation matrices. The set of orthonormal matrices contains the permutation matrices, and the orthonormally constrained problem ... |

74 | Approximating quadratic programming with bound and quadratic constraints
- Ye
- 1999
(Show Context)
Citation Context ...s an 87% performance guarantee for the problem MC, and a 97% performance in practice, see e.g. [12, 18, 15]. Other theoretical results for general objectives and further relaxed constraints appear in =-=[31, 46]-=-. In [38], several unrelated bounds for MC are shown to be equivalent. These bounds include the box relaxation \Gammaesxse, the trust region relaxation P i x 2 i = n, and an eigenvalue relaxation. Fur... |

72 | Semidefinite programming relaxations for the quadratic assignment problems
- Zhao, Karisch, et al.
- 1998
(Show Context)
Citation Context ...rect ordering of the eigenvalues. 2 The Lagrangian dual of QQPO is max S=S T min X tr AXBX T \Gamma tr S(XX T \Gamma I): (3.3) However, there can be a nonzero duality gap for the Lagrangian dual, see =-=[50]-=- for an example. The inner minimization in the dual problem (3.3) is an unconstrained quadratic minimization in the variables vec (X), with Hessian B\Omega A \Gamma I\Omega S: Clearly this minimizatio... |

65 |
A trust region strategy for nonlinear equality constrained optimization
- Celis, Dennis, et al.
- 1984
(Show Context)
Citation Context ...onstraint, i.e. two convex quadratic constraints. This problem arises in solving general nonlinear programs using a sequential quadratic programming approach, and is often called the CDT problem, see =-=[4]-=-. In contrast to the above single TRS, the TTRS can have a nonzero duality gap, see e.g. [2, 47, 48, 49]. This is closely related to quadratic theorems of the alternative, e.g. [5]. In addition, if th... |

63 | A semidefinite framework for trust region subproblems with applications to large scale minimization
- Rendl, Wolkowicz
- 1997
(Show Context)
Citation Context ...vex program where strong duality holds. Extensions of this result to a two-sided general, possibly nonconvex, constraint are discussed in [43, 29]. An algorithm based on Lagrangian duality appears in =-=[40]-=- and (implicitly) in [30, 41]. These algorithms are extremely efficient for the TRS problem, i.e. they solve this problem almost as quickly as an eigenvalue problem. 2.4 Two Trust Region Subproblem Th... |

56 | H.: Indefinite trust region subproblems and nonsymmetric eigenvalue perturbations
- Stern, Wolkowicz
- 1995
(Show Context)
Citation Context ... q 0 (x) s.t. x T x \Gamma ffi 2s0 (or = 0): Assuming that the constraint in TRS is written "," the Lagrangian dual is: DTRSs:= max 0 min x q 0 (x) + (x T x \Gamma ffi 2 ): This is equivalen=-=t to (see [43]-=-) the (concave) nonlinear semidefinite program DTRSs:= max g T 0 (Q + I) y g 0 \Gamma ffi 2 s.t. Q+ Is0s0: where \Delta y denotes Moore-Penrose inverse. It is shown in [43] that strong duality holds f... |

53 | A new lower bound via projection for the quadratic assignment problem
- Hadley, Rendl, et al.
- 1992
(Show Context)
Citation Context ... n symmetric matrices, and consider the orthonormally constrained homogeneous QQP QQPOsO := min tr AXBX T s:t: XX T = I : (3.1) This problem can be solved exactly using Lagrange multipliers, see e.g. =-=[14]-=-, or using the classical Hoffman-Wielandt inequality, e.g. [3]. We include a simple proof for completeness. Proposition 3.1 Suppose that the orthogonal diagonalizations of A; B are A = V \SigmaV T and... |

53 | A new matrix-free algorithm for the largescale trust-region subproblem
- Rojas, Santos, et al.
- 2000
(Show Context)
Citation Context ...duality holds. Extensions of this result to a two-sided general, possibly nonconvex, constraint are discussed in [43, 29]. An algorithm based on Lagrangian duality appears in [40] and (implicitly) in =-=[30, 41]-=-. These algorithms are extremely efficient for the TRS problem, i.e. they solve this problem almost as quickly as an eigenvalue problem. 2.4 Two Trust Region Subproblem The two trust region subproblem... |

52 |
Quadratic assignment problems
- Finke, Burkard, et al.
- 1987
(Show Context)
Citation Context ...atrices contains the permutation matrices, and the orthonormally constrained problem (3.1) is an important relaxation of QAP. The bounds obtained are usually called the eigenvalue bounds for QAP; see =-=[8, 13]-=-. Theorem 3.2 shows that the eigenvalue bounds are in fact obtained from a Lagrangian relaxation of (3.1) after adding the seemingly redundant constraint XX T = I. 4.2. Weighted sums of eigenvalues. C... |

50 | Cones of matrices and successive convex relaxations of nonconvex sets
- Kojima, Tunçel
(Show Context)
Citation Context ...riginal feasible set of QQP with the feasible set of the SDP relaxation. The connection is through valid quadratic inequalities, i.e. nonnegative (convex) combinations of the quadratic functions; see =-=[10, 20]-=-. 3 Orthogonally Constrained Programs with Zero Duality Gaps Consider the orthonormal constraint X T X = I ; X 2 M m;n : 6 (The set of such X is sometimes known as the Stiefel manifold, e.g. [7]. Appl... |

47 |
Quadratic optimization problems
- Shor
- 1987
(Show Context)
Citation Context ...ethods in [21]. Examples of approximations to QQPs also appear in [9]. The Lagrangian relaxation of a QQP is equivalent to the SDP relaxation, and is sometimes referred to as the Shor relaxation, see =-=[42]-=-. The Lagrangian relaxation can be written as an SDP if one takes into the account the hidden semidefinite constraint, i.e. a quadratic function is bounded below only if the Hessian is positive semide... |

44 | A recipe for semidefinite relaxation for 0-1 quadratic programming
- Poljak, Rendl, et al.
- 1995
(Show Context)
Citation Context ...nclude the box relaxation −e ≤ x ≤ e, the trust region relaxation ∑ i x2i = n, and an eigenvalue relaxation. Furthermore, these bounds are all shown to be equivalent to the Lagrangian relaxation; see =-=[37]-=-. Thus we see that the Lagrangian relaxation is equivalent to the best of these tractable bounds. 2.6. General QQP. The general, possibly nonconvex QQP has many applications in modeling and approximat... |

38 |
Perturbation bounds for matrix eigenvalues
- Bhatia
- 1987
(Show Context)
Citation Context ...ed homogeneous QQP QQPOsO := min tr AXBX T s:t: XX T = I : (3.1) This problem can be solved exactly using Lagrange multipliers, see e.g. [14], or using the classical Hoffman-Wielandt inequality, e.g. =-=[3]-=-. We include a simple proof for completeness. Proposition 3.1 Suppose that the orthogonal diagonalizations of A; B are A = V \SigmaV T and B = U U T ; respectively, where the eigenvalues in \Sigma are... |

38 |
A recipe for semidefinite relaxation for 0-1 quadratic programming
- Polijak, Rendl, et al.
- 1995
(Show Context)
Citation Context ...ude the box relaxation \Gammaesxse, the trust region relaxation P i x 2 i = n, and an eigenvalue relaxation. Furthermore, these bounds are all shown to be equivalent to the Lagrangian relaxation, see =-=[37]-=-. Thus we see that the Lagrangian relaxation is equivalent to the best known bound. 2.6 General QQP The general, possibly nonconvex, QQP has many applications in modeling and approximation theory, see... |

33 |
Semidefinite programming relaxation for nonconvex quadratic programs
- Fujie, Kojima
- 1997
(Show Context)
Citation Context ...riginal feasible set of QQP with the feasible set of the SDP relaxation. The connection is through valid quadratic inequalities, i.e. nonnegative (convex) combinations of the quadratic functions; see =-=[10, 20]-=-. 3 Orthogonally Constrained Programs with Zero Duality Gaps Consider the orthonormal constraint X T X = I ; X 2 M m;n : 6 (The set of such X is sometimes known as the Stiefel manifold, e.g. [7]. Appl... |

32 |
878-Approximation algorithms for MAX
- Goemans, Williamson
- 1994
(Show Context)
Citation Context ... provide a tractable approach for finding good bounds for many hard combinatorial problems. The best example is the application of SDP to the max-cut problem, where a 87% performance guarantee exists =-=[11, 12]-=-. Other examples include matrix completion problems, [23, 22] as well as graph partitioning problems and the quadratic assignment problem (references given below). In this paper we consider several qu... |

25 | Fixing variables in semidefinite relaxations
- Helmberg
(Show Context)
Citation Context ...theoretical properties. There have been attempts at further strengthening this relaxation. For example, a copositive relaxation is presented in [39]. Adding cuts to the SDP relaxation is discussed in =-=[15, 16, 17, 18]-=-. These improvements all involve heuristics such as which cuts to choose or solving a copositive problem, which is NP-hard in itself. The relaxation in (2.3) is obtained by lifting the vector x into m... |

25 | On a subproblem of trust region algorithms for constrained optimization
- Yuan
- 1990
(Show Context)
Citation Context ...linear programs using a sequential quadratic programming approach, and is often called the CDT problem, see [4]. In contrast to the above single TRS, the TTRS can have a nonzero duality gap, see e.g. =-=[2, 47, 48, 49]-=-. This is closely related to quadratic theorems of the alternative, e.g. [5]. In addition, if the constraints are not convex, then the primal may not be attained, see e.g. [27]. In [28], Martinez show... |

23 | Approximation algorithms for quadratic programming
- Fu, Luo, et al.
- 1998
(Show Context)
Citation Context ...QP The general, possibly nonconvex, QQP has many applications in modeling and approximation theory, see e.g. the applications to SQP methods in [21]. Examples of approximations to QQPs also appear in =-=[9]-=-. The Lagrangian relaxation of a QQP is equivalent to the SDP relaxation, and is sometimes referred to as the Shor relaxation, see [42]. The Lagrangian relaxation can be written as an SDP if one takes... |

23 |
A tour d’horizon on positive semidefinite and Euclidean distance matrix completion problems. Topics in semidefinite and interior-point methods
- Laurent
- 1996
(Show Context)
Citation Context ...y hard combinatorial problems. The best example is the application of SDP to the max-cut problem, where a 87% performance guarantee exists [11, 12]. Other examples include matrix completion problems, =-=[23, 22]-=- as well as graph partitioning problems and the quadratic assignment problem (references given below). In this paper we consider several quadratically constrained quadratic (nonconvex) programs arisin... |

21 |
An analytic center based column generation algorithm for convex quadratic feasibility problems
- Luo, Sun
- 1998
(Show Context)
Citation Context ...7]. However, the dual may not be attained, e.g. consider the convex program 0 = minfx : x 2s0g and its dual 0 = max 0 min x x + x 2 : Algorithmic approaches based on Lagrangian duality appear in e.g. =-=[19, 26, 32]-=-. 2.2 Rayleigh Quotient Suppose that A = A T 2 S n : It is well known that the smallest eigenvalues1 of A is obtained from the Rayleigh quotient, i.e.s1 = minfx T Ax : x T x = 1g: (2.1) Since A is not... |

20 |
Local minimizers of quadratic functions on Euclidean balls and spheres
- Martínez
- 1994
(Show Context)
Citation Context ....g. [2, 47, 48, 49]. This is closely related to quadratic theorems of the alternative, e.g. [5]. In addition, if the constraints are not convex, then the primal may not be attained, see e.g. [27]. In =-=[28]-=-, Martinez shows that the TRS can have at most one local and nonglobal optimum, and the Lagrangian at this point has one negative eigenvalue. Therefore, if we have such a case and add another ball con... |

18 |
Cones of matrices and setfunctions and 0-1 optimization
- 'ASZ, SCHRIJVER
- 1991
(Show Context)
Citation Context ... strengthened SDP relaxations for the max-cut problem. Adding constraints to close the duality gap is akin to adding valid inequalities in cutting plane methods for discrete optimization problems. In =-=[24, 25]-=- this approach, in combination with a lifting procedure, is used to solve discrete optimization problems. In our case we add quadratic constraints. 13 The idea of quadratic valid inequalities has been... |

18 | Convex relaxations of (0, 1)-quadratic programming
- Poljak, Wolkowicz
- 1995
(Show Context)
Citation Context ...ormance guarantee for the problem MC, and a 97% performance in practice, see e.g. [12, 18, 15]. Other theoretical results for general objectives and further relaxed constraints appear in [31, 46]. In =-=[38]-=-, several unrelated bounds for MC are shown to be equivalent. These bounds include the box relaxation \Gammaesxse, the trust region relaxation P i x 2 i = n, and an eigenvalue relaxation. Furthermore,... |

17 | Copositive relaxations for general quadratic programming
- Quist, Klerk, et al.
(Show Context)
Citation Context ...n for MC performs very well in practice, and has strong theoretical properties. There have been attempts at further strengthening this relaxation. For example, a copositive relaxation is presented in =-=[39]-=-. Adding cuts to the SDP relaxation is discussed in [15, 16, 17, 18]. These improvements all involve heuristics such as which cuts to choose or solving a copositive problem, which is NP-hard in itself... |

17 |
dual algorithm for minimizing a quadratic function with two quadratic constraints
- Yuan, “A
- 1991
(Show Context)
Citation Context ...linear programs using a sequential quadratic programming approach, and is often called the CDT problem, see [4]. In contrast to the above single TRS, the TTRS can have a nonzero duality gap, see e.g. =-=[2, 47, 48, 49]-=-. This is closely related to quadratic theorems of the alternative, e.g. [5]. In addition, if the constraints are not convex, then the primal may not be attained, see e.g. [27]. In [28], Martinez show... |

16 | Bounds for the Quadratic Assignment Problem Using Continuous Optimization Techniques
- Hadley, Rendl, et al.
- 1990
(Show Context)
Citation Context ...atrices contains the permutation matrices, and the orthonormally constrained problem (3.1) is an important relaxation of QAP. The bounds obtained are usually called the eigenvalue bounds for QAP, see =-=[8, 13]-=-. Theorem 3.2 shows that the eigenvalue bounds are in fact obtained from a Lagrangian relaxation of (3.1), after adding the seemingly redundant constraint XX T = I . 4.2 Weighted Sums of Eigenvalues C... |

16 | Generalization of the trust region problem
- Moré
- 1993
(Show Context)
Citation Context ...the eigenvalue case, we see that this is an example of a nonconvex program where strong duality holds. Extensions of this result to a two-sided general, possibly nonconvex constraint are discussed in =-=[43, 28]-=-. An algorithm based on Lagrangian duality appears in [40] and (implicitly) in [29, 41]. These algorithms are extremely efficient for the TRS problem, i.e., they solve this problem almost as quickly a... |

14 |
On the convergence of the method of analytic centers when applied to convex quadratic programs
- Jarre
- 1991
(Show Context)
Citation Context ...7]. However, the dual may not be attained, e.g. consider the convex program 0 = minfx : x 2s0g and its dual 0 = max 0 min x x + x 2 : Algorithmic approaches based on Lagrangian duality appear in e.g. =-=[19, 26, 32]-=-. 2.2 Rayleigh Quotient Suppose that A = A T 2 S n : It is well known that the smallest eigenvalues1 of A is obtained from the Rayleigh quotient, i.e.s1 = minfx T Ax : x T x = 1g: (2.1) Since A is not... |

14 |
SQ2P, sequential quadratic constrained quadratic programming
- Kruk, Wolkowicz
- 1998
(Show Context)
Citation Context ...ion is equivalent to the best known bound. 2.6 General QQP The general, possibly nonconvex, QQP has many applications in modeling and approximation theory, see e.g. the applications to SQP methods in =-=[21]-=-. Examples of approximations to QQPs also appear in [9]. The Lagrangian relaxation of a QQP is equivalent to the SDP relaxation, and is sometimes referred to as the Shor relaxation, see [42]. The Lagr... |

14 |
Goemans, Semidefinite programming in combinatorial optimization
- X
- 1997
(Show Context)
Citation Context ... provide a tractable approach for finding good bounds for many hard combinatorial problems. The best example is the application of SDP to the max-cut problem, where a 87% performance guarantee exists =-=[11, 12]-=-. Other examples include matrix completion problems [23, 22], as well as graph partitioning problems and the quadratic assignment problem (references given below). In this paper we consider several qu... |

13 |
On lp programming
- Terlaky
- 1985
(Show Context)
Citation Context ...alification is satisfied, i.e. there is no duality gap and the dual is attained. However, surprisingly, if the primal value of CQP is bounded then it is attained and there is no duality gap, see e.g. =-=[44, 35, 36, 34]-=- and, more recently, [27]. However, the dual may not be attained, e.g. consider the convex program 0 = minfx : x 2s0g and its dual 0 = max 0 min x x + x 2 : Algorithmic approaches based on Lagrangian ... |

12 |
ASZ. On the Shannon capacity of a graph
- LOV
- 1979
(Show Context)
Citation Context ... strengthened SDP relaxations for the max-cut problem. Adding constraints to close the duality gap is akin to adding valid inequalities in cutting plane methods for discrete optimization problems. In =-=[24, 25]-=- this approach, in combination with a lifting procedure, is used to solve discrete optimization problems. In our case we add quadratic constraints. 13 The idea of quadratic valid inequalities has been... |

12 |
Geometric Programming: Duality in Quadratic Programming and lp Approximation III
- Peterson, Ecker
- 1970
(Show Context)
Citation Context ...alification is satisfied, i.e. there is no duality gap and the dual is attained. However, surprisingly, if the primal value of CQP is bounded then it is attained and there is no duality gap, see e.g. =-=[44, 35, 36, 34]-=- and, more recently, [27]. However, the dual may not be attained, e.g. consider the convex program 0 = minfx : x 2s0g and its dual 0 = max 0 min x x + x 2 : Algorithmic approaches based on Lagrangian ... |

12 |
Perturbation bounds for matrix eigenvalues. Pitman research notes in mathematics series 162
- Bhatia
- 1987
(Show Context)
Citation Context ...eneous QQP (3.1) QQPO µ O := min tr AXBX T s.t. XXT = I. This problem can be solved exactly using Lagrange multipliers (see, e.g., [14]) or using the classical Hoffman–Wielandt inequality (see, e.g., =-=[3]-=-). We include a simple proof for completeness. Proposition 3.1. Suppose that the orthogonal diagonalizations of A, B are A = V ΣV T and B = UΛU T , respectively, where the eigenvalues in Σ are ordered... |

9 | On extensions of the Frank-Wolfe theorems
- Lou, Zhang
- 1999
(Show Context)
Citation Context ...is no duality gap and the dual is attained. However, surprisingly, if the primal value of CQP is bounded then it is attained and there is no duality gap, see e.g. [44, 35, 36, 34] and, more recently, =-=[27]-=-. However, the dual may not be attained, e.g. consider the convex program 0 = minfx : x 2s0g and its dual 0 = max 0 min x x + x 2 : Algorithmic approaches based on Lagrangian duality appear in e.g. [1... |

5 |
Generalizations of the trust region problem
- MOR
- 1993
(Show Context)
Citation Context ...he eigenvalue case, we see that this is an example of a nonconvex program where strong duality holds. Extensions of this result to a two-sided general, possibly nonconvex, constraint are discussed in =-=[43, 29]-=-. An algorithm based on Lagrangian duality appears in [40] and (implicitly) in [30, 41]. These algorithms are extremely efficient for the TRS problem, i.e. they solve this problem almost as quickly as... |