Semidefinite Programming versus the Reformulation-Linearization Technique for Nonconvex Quadratically Constrained Quadratic Programming Kurt M. Anstreicher Department of Management Sciences University of Iowa Iowa City, IA 52242 USA May 2007 Abstract We consider relaxations for nonconvex quadratically constrained quadratic programming (QCQP) based on semidefinite programming (SDP) and the reformulation-linearization tech- nique (RLT). From a theoretical standpoint we show that the addition of a semidefiniteness condition removes a substantial portion of the feasible region corresponding to product terms in the RLT relaxation. On test problems we show that the use of SDP and RLT constraints together can produce bounds that are substantially better than either technique used alone. For highly symmetric problems we also consider the effect of symmetry-breaking based on tightened bounds on variables and/or order constraints. 1 1 Introduction We consider a quadratically constrained quadratic programming problem of the form: xT Q x aT x QCQP : max 0 + 0 T T s.t.xQix + ai x ≤ bi,i∈I T T x Qix + ai x = bi,i∈E l ≤ x ≤ u, n where x ∈ and I∪E = {1,...,m}. We assume that −∞ <li <ui < +∞ for each i,and the matrices Qi are all symmetric. If Q0 0, Qi 0fori ∈Iand Qi =0fori ∈E,then QCQP is a convex optimization problem. In general however QCQP is NP-hard. QCQP is a well- studied problem in the global optimization literature with many applications, frequently arising from Euclidean distance geometry. An example that has attracted considerable recent attention concerns localizing sensor networks given distance information [11]. Global optimization methods for QCQP are typically based on convex relaxations of the prob- lem. In this paper we compare two such relaxations, based on semidefinite programming (SDP) [14] and the reformulation-linearization technique (RLT) [8]. These two relaxations are described in the next section. In section 3 we analyze the effect that adding the semidefiniteness condition has on the feasible region for the three variables in the RLT relaxation corresponding to product terms induced by two original variables xi, xj. We show that for typical values of the original variables the semidefiniteness constraint removes a large fraction of this feasible region. In section 4 we consider computational results on two different classes of test problems. For nonconvex box- constrained QPs we show that the use of SDP and RLT constraints together produces bounds that are substantially better than when either technique is used alone. We also consider SDP and RLT relaxations applied to circle-packing problems in the plane. These problems are highly symmetric, and we examine the effect of partial symmetry-breaking based on tightened bounds for subsets of variables. In section 5 we consider the effect of further symmetry-breaking based on imposing ad- ditional order constraints. Computational results on these problems indicate unexpectedly regular solution values for the various relaxations, as well as an unexpected relationship between bounds from SDP relaxations and bounds from RLT relaxations with additional order constraints. Notation We use X 0 to denote that a symmetric matrix X is positive semidefinite. For n × n X Y X • Y X • Y n X Y e matrices and , denotes the matrix inner product = i,j=1 ij ij.Weuse to denote a vector with each component equal to one. 2 2 The SDP and RLT Relaxations Relaxations of QCQP based on SDP and RLT both utilize variables Xij that replace the product terms xixj of the original problem. The relaxations differ in the form of the constraints that are placed on these new variables. The SDP relaxation is based on the fact that since X = xxT in the actual solution of QCQP, one can obtain a relaxation of QCQP by imposing X xxT instead. The SDP relaxation of QCQP [14] may then be written Q • X aT x SDP : max 0 + 0 T s.t.Qi • X + ai x ≤ bi,i∈I T Qi • X + ai x = bi,i∈E l ≤ x ≤ u, X − xxT 0. Moreover it is very well known that the condition X − xxT 0isequivalentto 1 xT X˜ := 0, (1) xX and therefore SDP may be alternatively written in the form SDP : max Q˜0 • X˜ s.t. Q˜i • X˜ ≤ 0,i∈I Q˜i • X˜ =0,i∈E l ≤ x ≤ u, X˜ 0, where T −bi ai /2 Q˜i := . ai/2 Qi When the original QCQP is a convex problem (Q0 0, Qi 0fori ∈Iand Qi =0for i ∈E), it is straightforward to show that SDP is equivalent to QCQP. If QCQP is nonconvex, however, SDP may be unbounded even though all of the original variables have finite upper and lower bounds. This can easily be remedied by adding upper bounds to the diagonal components of 2 2 X. For example, it is obvious that Xii ≤ max{li ,ui }. Better upper bounds for Xii are obtained as part of the RLT relaxation that we describe below. An approximation result based on the SDP relaxation for a special case of QCQP (l = −e, u = e, I = ∅, ai =0andQi diagonal for i =1,...,m) is given in [18] The RLT relaxation of QCQP is based on using products of upper and lower bound constraints on the original variables to obtain valid linear inequality constraints on the new variables Xij [8]. 3 For two variables xi and xj we have constraints xi − li ≥ 0, ui − xi ≥ 0, xj − lj ≥ 0, uj − xj ≥ 0. Multiplying each of the constraints involving xi by a constraint involving xj, and replacing the product term xixj with the new variable Xij, we obtain the constraints Xij − lixj − ljxi ≥−lilj, Xij − uixj − ujxi ≥−uiuj, Xij − lixj − ujxi ≤−liuj, Xij − ljxi − uixj ≤−ljui, i, j =1,...,n. Note that these constraints also hold when i = j,inwhichcasethelasttwo constraints are identical. Moreover the last two constraints are identical for all i, j once the condition Xij = Xji is imposed. The resulting relaxation of QCQP can then be written Q • X aT x RLT : max 0 + 0 T s.t.Qi • X + ai x ≤ bi,i∈I T Qi • X + ai x = bi,i∈E X − lxT − xlT ≥−llT X − uxT − xuT ≥−uuT X − lxT − xuT ≤−luT l ≤ x ≤ u, X = XT . Using the fact that Xij = Xji, the result is an ordinary linear programming (LP) problem with n(n +3)/2 variables and a total of m + n(2n + 3) constraints. (In fact it is known that the original bound constraints l ≤ x ≤ u are redundant and could be removed [10, Proposition 1].) If QCQP contains linear constraints other than the simple bounds (Qi =0forsome0<i≤ m), then additional constraints can be imposed on the variables X in the SDP and/or RLT relaxations. For RLT the standard methodology is to form all possible products of pairs of linear inequality constraints, including the bound constraints on the variables [10]. If |{i ∈I: Qi =0}| = k this would add an additional 2kn + k(k − 1)/2 linear inequality constraints to the RLT relaxation. If there are linear equality constraints then it suffices to consider the constraints on X obtained by multiplying individual equality constraints by each variable xj [8, Remark 8.1], so if |{i ∈E: Qi = 0}| = p, a total of pn additional linear equality constraints would be added to the RLT relaxation. For linear equality constraints the standard approach in forming SDP relaxations (see for example [17, Remark 13.4.1]) is to add only the “squared” constraints T T 2 ai x = bi ⇒ ai Xai = bi . 4 The treatment of linear inequality constraints in SDP relaxations is less obvious, because the T constraint obtained by “squaring” a linear inequality ai x−bi ≥ 0 is already implied by X˜ 0. In [2] some theoretical justification is given for generating additional constraints from linear inequalities by first adding slack variables to obtain equalities and then forming the squared equality constraints. 3 Adding SDP to RLT In this section we examine the effect of adding the semidefiniteness condition X xxT to the RLT relaxation of QCQP. We will focus on the effect that adding semidefiniteness has on the feasible values for the product variables Xij. It is well known that the RLT relaxation is invariant with respect to an invertible affine transformation of the original variables [8, Proposition 8.4], and it is easy to show that such an invariance also holds for the SDP relaxation. As a result we may assume without loss of generality that l =0,u = e. We will consider two variables xi, xj,andfor convenience assume that i =1,j = 2. By interchanging and/or complementing the variables we may further assume that 0 ≤ x1 ≤ x2 ≤ .5. Under these assumptions the RLT constraints on X11, X22 and X12 become 0 ≤ X11 ≤ x1, (2a) 0 ≤ X22 ≤ x2, (2b) 0 ≤ X12 ≤ x1. (2c) Next we consider imposing the semidefiniteness condition X − xxT 0.Asdescribedinthe previous section this is equivalent to X˜ 0, where X˜ is defined as in (1). Restricting attention to the principal submatrix of X˜ corresponding to x1 and x2, we certainly have ⎛ ⎞ 1 x1 x2 ⎜ ⎟ ⎜ ⎟ ⎝ x1 X11 X12 ⎠ 0, (3) x2 X12 X22 and it is straightforward to show that for X12 ≥ 0, (3) is equivalent to the constraints X ≥ x2, 11 1 (4a) 2 X22 ≥ x , (4b) 2 2 2 X12 ≤ x1x2 + (X11 − x )(X22 − x ), (4c) 1 2 X ≥ x x − X − x2 X − x2 . 12 1 2 ( 11 1)( 22 2) (4d) 5 Our goal is to compare, for fixed values of x1 and x2, the three-dimensional feasible regions for (X11,X22,X12) corresponding to (2) before and after the addition of (4).