An Explicit Exact SDP Relaxation for Nonlinear 0-1 Programs Jean B. Lasserre LAAS-CNRS, 7 Avenue du Colonel Roche, 31077 Toulouse Cédex, France [email protected] http://www.laas.fr/˜lasserre Abstract. We consider the general nonlinear optimization problem in 01 variables and provide an explicit equivalent convex positive semidefinite program in 2n − 1 variables. The optimal values of both problems are identical. From every optimal solution of the former one easily find an optimal solution of the latter and conversely, from every solution of the latter one may construct an optimal solution of the former. 1 Introduction This paper is concerned with the general nonlinear problem in 0-1 variables IP → p∗ := min {g0 (x) | gk (x) ≥ 0, k = 1, . . . m} x∈{0,1}n (1) where all the gk (x) : IRn → IR are real-valued polynomials of degree 2vk −1 if odd or 2vk if even. This general formulation encompasses 0-1 linear and nonlinear programs, among them the quadratic assignment problem. For the MAX-CUT problem, the discrete set {0, 1}n is replaced by {−1, 1}n . In our recent work [4,5,6], and for general optimization problems involving polynomials, we have provided a sequence {Qi } of positive semidefinite (psd) programs (or SDP relaxations) with the property that inf Qi ↑ p∗ as i → ∞, under a certain assumption on the semi-algebraic feasible set {gk (x) ≥ 0, k = 1, . . . m}. The approach was based on recent results in algebraic geometry on the representation of polynomials, positive on a compact semi-algebraic set, a theory dual to the theory of moments. For general 0 − 1 programs, that is, with the additional constraint x ∈ {0, 1}n , we have shown in Lasserre [5] that this assumption on the feasible set is automatically satisfied and the SDP relaxations {Qi } simplify to a specific form with at most 2n − 1 variables, no matter how large is i. The approach followed in [4] and [5] is different in spirit from the lift and project iterative procedure of Lovász and Schrijver [7] for 0-1 programs (see also extensions in Kojima and Tunçel [3]), which requires that a weak separation oracle is available for the homogeneous cone associated with the constraints. In the lift and project procedure, the description of the convex hull of the feasible set is implicit (via successive projections) whereas we provide in this paper an K. Aardal, B. Gerards (Eds.): IPCO 2001, LNCS 2081, pp. 293–303, 2001. c Springer-Verlag Berlin Heidelberg 2001 294 J.B. Lasserre explicit description of the equivalent convex psd program, with a simple interpretation. Although different, our approach is closer in spirit to the successive LP relaxations in the RLT procedure of Sherali and Adams [10] for 0-1 linear programs in which each of the linear original constraints is multiplied by suitable polynomials of the form Πi∈J1 xi Πj∈J2 (1 − xj ) and then linearized in a higher dimension space via several changes of variables to obtain an LP program. The last relaxation in the hierarchy of RLT produces the convex hull of the feasible set. This also extends to a special class of 0-1 polynomial programs (see Sherali and Adams [10]). In the present paper, we show that in addition to the asymptotic convergence already proved in [4,5], the sequence of convex SDP relaxations {Qi } is in fact finite, that is, the optimal value p∗ is also min Qi for all i ≥ n + v with v := maxk vk . Moreover, every optimal solution y ∗ of Qi is the (finite) vector of moments of some probability measure supported on optimal solutions of IP. Therefore, every 0-1 program is in fact equivalent to a continuous convex psd program in 2n − 1 variables for which an explicit form as well as a simple interpretation are available. The projection of the feasible set defined by the LMI constraints of this psd program onto the subspace spanned by the first n variables, provides the convex hull of the original feasible set. Note that the result holds for arbitrary 0-1 constrained programs, that is, with arbitrary polynomial criteria and constraints (no weak separation oracle is needed as in the lift and project procedure [7], and no special form of IP is assumed as in [10]). This is because the theory of representation of polynomials positive on a compact semi-algebraic set and its dual theory of moments, make no assumption on the semi-algebraic set, except compactness (it can be nonconvex, disconnected). For illustration purposes, we provide the equivalent psd programs for quadratic 0-1 programs and MAX-CUT problems in IR3 . For practical computational purposes, the preceding result is of little value for the number of variables grows exponentially with the size of the problem. Fortunately, in many cases, the optimal value is also the optimal value of some Qi for i n. For instance, on a sample of 50 randomly generated MAX-CUT problems in IR10 , the optimal value p∗ was always obtained at the Q2 relaxation (in which case Q2 is a psd program with “only” 385 variables to compare with 210 − 1 = 1023). 2 Notation and Definitions We adopt the notation in Lasserre [4] that for sake of clarity, we reproduce here. Given any two real-valued symmetric matrices A, B let hA, Bi denote the usual scalar product trace(AB) and let A B (resp. A B) stand for A − B positive semidefinite (resp. A − B positive definite). Let 1, x1 , x2 , . . . xn , x21 , x1 x2 , . . . , x1 xn , x2 x3 , . . . , x2n , . . . , xr1 , . . . , xrn , (2) An Explicit Exact SDP Relaxation for Nonlinear 0-1 Programs 295 be a basis for the r-degree real-valued polynomials and let s(r) be its dimension. Therefore, a r-degree polynomial p(x) : IRn → IR is written p(x) = X pα xα , x ∈ IRn , α s(r) αn 1 α2 the coefficients of the where xα = xα 1 x2 . . . xn . Denote by p = {pα } ∈ IR polynomial p(x) in the basis (2). Hence, the respective vectors of coefficients of the polynomials gi (x), i = 0, 1, . . . m, in (1), are denoted {(gi )α } = gi ∈ IRs(wi ) , i = 0, 1, . . . m if wi is the degree of gi . We next define the important notions of moment matrix and localizing matrix. 2.1 Moment Matrix Given a s(2r)-sequence (1, y1 , . . . , ), let Mr (y) be the moment matrix of dimension s(r) (denoted M (r) in Curto and Fialkow [1]), with rows and columns labelled by (2). For instance, for illustration purposes, consider the 2-dimensional case. The moment matrix Mr (y) is the block matrix {Mi,j (y)}0≤i,j≤2r defined by yi+j,0 yi+j−1,1 . . . yi,j yi+j−1,1 yi+j−2,2 . . . yi−1,j+1 . (3) Mi,j (y) = ... ... ... ... yj,i yi+j−1,1 . . . y0,i+j To fix ideas, when n = 2 and r = 2, one obtains 1 y10 y01 y20 y11 y10 y20 y11 y30 y21 y01 y11 y02 y21 y12 M2 (y) = y20 y30 y21 y40 y31 y11 y21 y12 y31 y22 y02 y12 y03 y22 y13 y02 y12 y03 . y22 y13 y04 Another more intuitive way of constructing Mr (y) is as follows. If Mr (1, i) = yα and Mr (y)(j, 1) = yβ , then Mr (y)(i, j) = yα+β , with α + β = (α1 + β1 , · · · , αn + βn ). (4) Mr (y) defines a bilinear form h., .iy on the space Ar of polynomials of degree at most r, by hq(x), v(x)iy := hq, Mr (y)vi, q(x), v(x) ∈ Ar , and if y is a sequence of moments of some measure µy , then Mr (y) 0 because Z q(x)2 µy (dx) ≥ 0, ∀q ∈ IRs(r) . (5) hq, Mr (y)qi = 296 2.2 J.B. Lasserre Localizing Matrix If the entry (i, j) of the matrix Mr (y) is yβ , let β(i, j) denote the subscript β of yβ . Next, given a polynomial θ(x) : IRn → IR with coefficient vector θ, we define the matrix Mr (θy) by X θα y{β(i,j)+α} . (6) Mr (θy)(i, j) = α For instance, with x → 7 θ(x) = a − x21 − x22 , we obtain a − y20 − y02 , ay10 − y30 − y12 , ay01 − y21 − y03 M1 (θy) = ay10 − y30 − y12 , ay20 − y40 − y22 , ay11 − y31 − y13 . ay01 − y21 − y03 , ay11 − y31 − y13 , ay02 − y22 − y04 If y is a sequence of moments of some measure µy , then Z θ(x)q(x)2 µy (dx), ∀q ∈ IRs(r) , hq, Mr (θy)qi = (7) so that Mr (θy) 0 whenever µy has its support contained in the set {θ(x) ≥ 0}. In Curto and Fialkow [1], Mr (θy) is called a localizing matrix. The theory of moments identifies those sequences {yα } with Mr (y) 0, such that the yα ’s are moments of a representing measure µy . The IK-moment problem identifies those yα ’s whose representing measure µy has its support contained in the semi-algebraic set IK. In duality with the theory of moments is the theory of representation of positive polynomials (and polynomials positive on a semi-algebraic set IK), which dates back to Hilbert’s 17th problem. For details and recent results, the interested reader is referred to Curto and Fialkow [1], Schmüdgen [9] and the many references therein. Both sides (moments and positive polynomials) of this same theory is reflected in the primal and dual SDP relaxations proposed in Lasserre [4], and used below in the present context of nonlinear 0-1 programs. 3 Main Result Consider the 0-1 optimization problem IP in (1). Let the semi-algebraic set IK := {x ∈ {0, 1}n | gi (x) ≥ 0, i = 1, . . . m} be the feasible set. For the sake of simplicy in the proofs derived later on, we treat the 0-1 integrality constraints x2i = xi as two opposite inequalities gm+i (x) = x2i − xi ≥ 0 and gm+n+i (x) = xi − x2i ≥ 0, and we redefine the set IK to be IK = {gi (x) ≥ 0, i = 1, . . . m + 2n}. (8) However, in view of the special form of the constraints gm+k (x) ≥ 0, k = 1, . . . 2n, we will provide a simpler form of the LMI relaxations {Qi } below. An Explicit Exact SDP Relaxation for Nonlinear 0-1 Programs 297 Depending on its parity, let wk := 2vk or wk := 2vk − 1 be the degree of the polynomial gk (x), k = 1, . . . m + 2n. When needed below, for i ≥ maxk wk , the vectors gk ∈ IRs(wk ) are extended to vectors of IRs(i) by completing with zeros. As we minimize g0 we may and will assume that its constant term is zero, that is, g0 (0) = 0. For i ≥ maxk vk , consider the following family {Qi } of convex psd programs X (g0 )α yα inf y α (9) Qi Mi (y) 0 M (g y) 0, k = 1, . . . m + 2n, i−vk k with respective dual problems m+2n X −X(1, 1) − gk (0)Zk (1, 1) sup X,Z 0 k ∗ k=1 Qi m+2n X hZk , Cαk i = (g0 )α , ∀α 6= 0, hX, Bα i + (10) k=1 where we have written X X Bα yα ; Mi−vk (gk y) = Cαk yα , k = 1, . . . m + 2n, Mi (y) = α α for appropriate real-valued symmetric matrices Bα , Cαk , k = 1, . . . m + 2n. Interpretation of Qi . The LMI constraints of Qi state necessary conditions for y to be the vector of moments up to order 2i, of some probability measure µy with support contained in IK. This clearly implies that inf Qi ≤ p∗ because the vector of moments of the Dirac measure at a feasible point of IP, is feasible for Qi . Interpretation of Q∗i . Let X, Zk 0 be a feasible solution of Q∗i with value ρ. Write X X tj t0j and Zk = tkj t0kj , k = 1, . . . m + 2n. X = j j Then, from the feasibility of (X, Zk ) in Q∗i , it was shown in Lasserre [4,5] that m+2n X X X (11) tj (x)2 + gk (x) tkj (x)2 , g0 (x) − ρ = j k=1 j where the polynomials {tj (x)} and {tkj (x)} have respective coefficient vectors {tj } and {tkj }, in the basis (2). As ρ ≤ p∗ , g0 (x) − ρ is nonnegative on IK (strictly positive if ρ < p∗ ). One recognizes in (11) a decomposition into a weighted sum of squares of the polynomial g0 (x) − p∗ , strictly positive on IK, as in the theory of representation 298 J.B. Lasserre of polynomials, strictly positive on a compact semi-algebraic set IK (see e.g., Schmüdgen [9], Putinar [8], Jacobi and Prestel [2]). Indeed, when the set IK has a certain property (satisfied here), the “linear” representation (11) holds (see Lasserre [4,5]). Hence, both programs Qi and Q∗i illustrate the duality between the theory of moments and the theory of positive polynomials. Among other results, it has been shown in Lasserre [5, Th. 3.3] that inf Qi ↑ p∗ as i → ∞. (12) From the construction of the localizing matrices in (6), and the form of the polynomials gi for i > m, the constraints Mi−1 (gm+k y) 0 and Mi−1 (gm+n+k y) 0 for k = 1, . . . n (equivalently Mi−1 (gm+k y) = 0), simply state that the variable yα with α = (α1 , . . . αn ), can be replaced with the variable yβ with βi := 1 whenever αi ≥ 1. Therefore, a simpler form of Qi is obtained as follows. Ignore the constraints Mi−vk (gk y) 0 for k = m + 1, . . . m + 2n, and make the above substitution yα → yβ in the matrices Mi (y) and Mi−vk (gk (y)), k = 1, . . . m. For instance, in IR2 (n = 2), the matrix M2 (y) now reads 1 y10 y01 y10 y11 y01 y10 y10 y11 y10 y11 y11 y01 y11 y01 y11 y11 y01 , M2 (y) = (13) y10 y10 y11 y10 y11 y11 y11 y11 y11 y11 y11 y11 y01 y11 y01 y11 y11 y01 and only the variables y10 , y01 , y11 appear in all the relaxations Qi . Interpreted in terms of polynomials, the integrality constraints x2i = xi , i = 1, . . . n, imply αn 1 α2 that a monomial xα 1 x2 · · · xn can be replaced by 0 if αi = 0 (14) xβ1 1 xβ2 2 · · · xβnn with βi = 1 if αi ≥ 1 Therefore, there are are no more than 2n − 1 variables yβ , the number of monomials as in (14), of degree at most n. Remark 1. At any relaxation Qi , the matrix Mi (y) can be reduced in size. When looking at the kth column Mi (y)(., k), if Mi (y)(1, k) = Mi (y)(p) for some p < k, then the whole column Mi (y)(., k) as well as the corresponding line Mi (y)(k, .) can be deleted. For instance, with M2 (y) as in (13), the constraint M2 (y) 0 can be replaced by 1 y10 y01 y11 c2 (y) := y10 y10 y11 y11 0, M y01 y11 y01 y11 y11 y11 y11 y11 for the 4th and 6th columns of M2 (y) are identical to the 2nd and 3rd columns, ci (y) of Mi (y), one retains only the respectively. Thus, in the simplified form M An Explicit Exact SDP Relaxation for Nonlinear 0-1 Programs 299 columns (and the rows) corresponding to the monomials in the basis (2) that are distinct after the simplification x2i = xi . The same simplification occurs for the matrices of the LMI constraints Mi−vk (gk y) 0, k = 1, . . . m. Next, we begin with the following crucial result: Proposition 1. (a) All the relaxations Qi involve at most 2n − 1 variables yα . (b) For all the relaxations Qi with i > n, one has rank Mi (y) = rank Mn (y). (15) Proof. (a) is just a consequence of the comment preceding Proposition 1. To get (b) observe that with i > n, one may write Mn (y) | B − Mi (y) = − 0 B |C for appropriate matrices B, C, and we next prove that each column of B is identical to some column of Mn (y). Indeed, remember from (4) how an element Mi (y)(k, p) can be obtained. Let yγ = Mi (y)(k, 1) and yα = Mi (y)(1, p). Then Mi (y)(k, p) = yη with ηi = γi + αi , i = 1, . . . n. (16) Now, consider a column Bj of B, that is, some column Mi (y)(., p) of Mi (y), with first element yα = Bj (1) = Mi (y)(1, p). Therefore, the element B(k) (or, equivalently, the element Mi (y)(k, p)) is the variable yη in (16). Note that α correspond to a monomial in the basis (2) of degree larger than n, say α1 · · · αn . Associate to this column Bj the column v := Mn (y)(., q) of Mn (y) whose element v(1) = Mn (y)(1, q) = yβ (for some q) with βi = 0 if αi = 0 and 1 otherwise, for all i = 1, . . . n. Then, the element v(k) = Mn (y)(k, q) is obtained as v(k) = yδ with δi = γi + βi , i = 1, . . . n. But then, as for each entry yα of Mj (y), we can make the substitution αi ↔ 1 whenever αi ≥ 1, it follows that the element v(k) is identical to the element B(k). In other words, each column of B is identical to some columnof M n (y). Mn (y) B If we now write Mj (y) = [A|D] with A := , and D := , then, B0 C with exactly same arguments, every column of D is also identical to some column of A, and consequently, (15) follows. t u For instance, when n = 2 the reader can check that M3 (y) has same rank as M2 (y) in (13). We now can state our main result: Theorem 1. Let IP be the problem defined in (1) and let v := maxk=1,...m vk . Then for every i ≥ n + v: (a) Qi is solvable with p∗ = min Qi , and to every optimal solution x∗ of IP corresponds the optimal solution y ∗ := (x∗1 , · · · , x∗n , · · · , (x∗1 )2i , · · · , (x∗n )2i ) of Qi . (17) 300 J.B. Lasserre (b) Every optimal solution y ∗ of Qi is the (finite) vector of moments of a probability measure finitely supported on s optimal solutions of IP, with s = rank Mi (y) = rank Mn (y). Proof. Let y be an admissible solution of Qn+v . From Proposition 1, we have that rank Mi (y) = rank Mn (y) for all i > n (in particular for i = n + v). From a deep result of Curto and Fialkow [1, Th 1.6, p. 6], it follows that y is the vector of moments of some rank Mn (y) atomic measure µy . Mn+1 (y) is called a flat positive extension of Mn (y) and it follows that Mn+1 (y) admits unique flat extension moment matrices Mn+2 (y), Mn+3 (y), etc. (see Curto and Fialkow [1, p. 3]). Moreover, from the constraints Mn+v−vk (gk y) 0, for all k = 1, . . . m + 2n, it also follows that µy is supported on IK. Theorem 1.6 in Curto and Fialkow [1] is stated in dimension 2 for the complex plane, but is valid for n real or complex variables (see comments page 2 in [1]). In the notation of Theorem 1.6 in Curto and Fialkow [1], we have M (n) 0 and M (n + 1) has a flat positive extension M (n + 1) (hence unique flat positive extensions M (n + k) for all k ≥ 1), with Mgk (n + vk ) 0, k = 1, . . . m + 2n. But then, as µy is supported on IK, it also follows that Z X (g0 )α yα = g0 (x) µy (dx) ≥ p∗ . IK α From this and inf Qi ≤ p∗ , the first statement (a) in Theorem 1 follows for i = n + v. For i > n + v, the result follows from p∗ ≥ inf Qi+1 ≥ inf Qi , for all i. Finally, y ∗ in (17) is obviously admissible for Qi with value p∗ , and therefore, is an optimal solution of Qi . (b) follows from same arguments as in (a). First observe that from (a), Qi is solvable for all i ≥ n+v. Now, let y ∗ be an optimal solution of Qi . From (a), y ∗ is the vector of moments of an atomic measure µy∗ supported on s (= rank Mn (y)) points x1 , · · · xs ∈ IK, that is, with δ{.} the Dirac measure at 00 .00 , µ y∗ = s X γj δxj with γj ≥ 0, j=1 s X γj = 1. j=1 Therefore, from g0 (xj ) ≥ p∗ for all j = 1, . . . s, and p∗ = min Qi = Z = IK X α (g0 )α yα∗ g0 (x) µy∗ (dx) = s X γj g0 (xj ), j=1 it follows that each point xj must be an optimal solution of IP. t u Hence, Theorem 1 shows that every 0-1 program is equivalent to a convex psd program with 2n − 1 variables. The projection of the feasible set of Qn+v onto An Explicit Exact SDP Relaxation for Nonlinear 0-1 Programs 301 the space spanned by the n variables y10...0 , . . . , y0...01 is the convex hull of IK in (8). Interpretation . The interpretation of the relaxation Qn+v is as follows. The unknown variables {yα } should be interpreted as the moments of some probability measure µ, up to order n + v. The LMI constraints Mn+v−vk (gk y) 0 state that Z for all polynomials q of degree at most n, gk (x)q(x)2 dµ ≥ 0, (as opposed to only gk (x) ≥ 0 in the original description). As the integral constraints x2i = xi are included in the constraints, and due to the result of Curto and Fialkow, the above constraint (for k = 1, . . . m + 2n) will ensure that the support of µ is concentrated on {0, 1}n ∩ [∩m k=1 {gk (x) ≥ 0}]. For illustration purposes, we provide below the explicit description of the equivalent convex psd programs for quadratic 0-1 programs and MAX-CUT programs in IR3 , respectively. 3.1 Quadratic 0-1 Programs Consider the quadratic program min{x0 Ax | x ∈ {0, 1}n } for some real-valued symmetric matrix A ∈ IRn×n . As the only constraints are the integral constraints x ∈ {0, 1}n , with n = 3, it is equivalent to the convex psd program inf A11 y100 + A22 y010 + A33 y001 + 2A12 y110 + 2A13 y101 + 2A23 y011 1 y100 y100 y100 y010 y110 c3 (y) = y001 y101 M y110 y110 y101 y101 y011 y111 y111 y111 y010 y110 y010 y011 y110 y111 y011 y111 y001 y101 y011 y001 y111 y101 y011 y111 y110 y110 y110 y111 y110 y111 y111 y111 y101 y101 y111 y101 y111 y101 y111 y111 y011 y111 y011 y011 y111 y111 y011 y111 y111 y111 y111 y111 y111 y111 y111 y111 0, c3 (y) is the simplified form of M3 (y) (see Remark 1). where M 3.2 MAX-CUT Problem In this case, g0 (x) := xAx for some real-valued symmetric matrix A = {Aij } ∈ IRn×n with null diagonal and the constraint set is {−1, 1}n . Now, we look for solutions in {−1, 1}n , which yields the following obvious modifications in the relaxations {Qi }. In the matrix Mi (y), replace any entry yα by yβ with βi = 1 whenever αi is odd, and βi = 0 otherwise. As for 0-1 programs, the matrix Mi (y) 302 J.B. Lasserre ci (y) (by eliminating identical rows can be reduced in size to a simplified form M and columns). Hence, MAX-CUT in IR3 is equivalent to the psd program inf A12 y110 + A13 y101 + A23 y011 1 y100 y100 1 y010 y110 y001 y101 c M3 (y) = y110 y010 y101 y001 y011 y111 y111 y011 y010 y110 1 y011 y100 y111 y001 y101 y001 y101 y011 1 y111 y100 y010 y110 y110 y010 y100 y111 1 y011 y101 y001 y101 y001 y111 y100 y011 1 y110 y010 y011 y111 y001 y010 y101 y110 1 y100 y111 y011 y101 y110 0. y001 y010 y100 1 Despite its theoretical interest, this result is of little value for computational purposes for the number of variables is exponential in the problem size. However, in many cases, low order SDP relaxations Qi , that is, with i n, will provide the optimal value p∗ . To illustrate the power of the SDP relaxations Qi , we have run a sample of 50 MAX-CUT problems in IR10 , with non diagonal entries randomly generated (uniformly) between 0 and 1 (in some examples, zeros and negative entries were allowed). In all cases, min Q2 provided the optimal value. The SDP ci (y) 0) of dimension 56 × 56 and 385 relaxation Q2 has one LMI constraint (M variables yβ . Remark 2. Consider the MAX-CUT problem with equal weights, that is, X min n xi xj . p∗ = x∈{−1,1} From i<j " n " n #2 # X X X 1 1 n xi xj + xi + (1 − xi )2 , = 2 2 2 i<j i=1 i=1 we can prove that min Q1 = −n/2. Indeed, with en ∈ IRn being a vector of ones, let X ∈ IR(n+1)×(n+1) , Zk ∈ IR, be defined as 1 0 1 × [0, en ], (18) Zk = , k = 1, . . . n; X := 2 2 en Pn so that X 0 and −X(1, 1) − k=1 Zk (1, 1) = −n/2. Thus, (X, {Zk }) is admissible for Q∗1 with value −n/2. Let n = 2p, and let x ∈ IRn be defined as xi = 1, i = 1, . . . p; xi = −1, i = p + 1, . . . n. Let y be the vector of first and second moments of the Dirac at x (which, from M1 (y) 0, implies that y is admissible for Q1 ). Its value, i.e., the sum of all yα ’s corresponding to xi xj , ∀i < j, is −n/2. Moreover, " n #2 X 1 0 0 xi = 0, ,X i = hM1 (y), Xi = h en 2 en i=1 An Explicit Exact SDP Relaxation for Nonlinear 0-1 Programs 303 so that y and (X, {Zk }) are optimal solutions of Q1 and Q∗1 respectively. It also follows that −n/2 = p∗ , as x is feasible for IP. Next, consider the case where n is odd. Let y be defined with the correspondances xi → yα := 0; i = 1, 2, . . . n; xi xj → yα := −1/(n − 1), ∀i < j. With X ∈ IR(n+1)×(n+1) , Zk ∈ IR as in (18), one may check that M1 (y) 0 and hM1 (y), Xi = 0. The sum of all yα corresponding to xi xj , i < j is −n(n − 1)/2 × (n − 1)−1 = −n/2. Hence, again, y and (X, {Z}k ) are optimal solutions of Q1 and Q∗1 respectively. But y is not a moment vector and −n/2 is only a (strict) lower bound on p∗ . 4 Conclusion We have provided an explicit equivalent continuous convex psd program for arbitrary constrained nonlinear 0-1 programs, the last in a finite hierarchy of SDP relaxations. For practical computation, in many cases, an SDP relaxation of low order is exact, i.e., provides the optimal value p∗ , as confirmed on the MAX-CUT problem as well as on a (small) sample of nonconvex continuous problems in Lasserre [4]. However, for large or even moderate size problems, the resulting SDP relaxations Qi might still be too large for the present status of SDP software packages. Therefore, we hope that this work will help stimulate further effort in designing efficient solving procedures for large SDP programs. References 1. Curto, R.E. , Fialkow, L.A.: The truncated complex K-moment problem. Trans. Amer. Math. Soc. 352 (2000) 2825–2855. 2. Jacobi, T., Prestel, A.: On special representations of strictly positive polynomials. Technical-report, Konstanz University, Germany (2000). 3. Kojima, M., Tunçel, L.: Cones of matrices and successive convex relaxations of non convex sets. SIAM J. Optim. 10 (2000) 750–778. 4. Lasserre, J.B.: Global optimization with polynomials and the problem of moments. SIAM J. Optim. 11 (2001) 796–817. 5. Lasserre, J.B.: Optimality conditions and LMI relaxations for 0-1 programs. Technical report #00099, LAAS-CNRS Toulouse, France (2000). 6. Lasserre, J.B.: New LMI relaxations for the general nonconvex quadratic problem. In: Advances in Convex Analysis and Global Optimization, Eds: N. Hadjisavvas and P.M. Pardalos, Kluwer Academic Publishers, 2001. 7. Lovász, L., Schrijver, A.: Cones of matrices and set-functions and 0-1 optimization. SIAM J. Optim. 1 (1991) 166–190. 8. Putinar, M.: Positive polynomials on compact semi-algebraic sets. Ind. Univ. Math. J. 42 (1993) 969–984. 9. Schmüdgen, K.: The K-moment problem for compact semi-algebraic sets. Math. Ann. 289 (1991) 203–206. 10. Sherali, H.D., Adams, W.P.: A hierarchy of relaxations between the continuous and convex hull representations for zero-one programmimg problems. SIAM J. Disc. Math. 3 (1990) 411–430.
© Copyright 2025 Paperzz