Solutions and optimality criteria to box constrained nonconvex minimization problems

This paper presents a canonical duality theory for solving 
nonconvex polynomial programming problems subjected to box 
constraints. 
 It is proved that under certain conditions, 
 the constrained nonconvex problems can be converted to 
 the so-called canonical (perfect) dual problems, which can be solved 
 by deterministic methods. 
Both global and local extrema of the primal problems 
 can be identified by a triality theory proposed by the author. 
 Applications to 
 nonconvex integer programming and Boolean least squares problems 
 are discussed. 
 Examples are illustrated. A conjecture on NP-hard problems is proposed.

1. Primal problem and its dual form. The box constrained nonconvex minimization problem is proposed as a primal problem (P) given below: where X a = {x ∈ R n | ℓ l ≤ x ≤ ℓ u } is a feasible space, Q(x) = 1 2 x T Ax − c T x is a quadratic function, A = A T ∈ R n×n is a given symmetric matrix, ℓ l , ℓ u , and c are three given vectors in R n , W (x) is a nonconvex function. In this paper, we simply assume that W (x) is a so-called double-well fourth order polynomial function defined by where B ∈ R m×n is a given matrix and α > 0 is a given parameter. The notation |x| used in this paper denotes the Euclidean norm of x. Problems of the form (1) appear frequently in many applications, such as semilinear nonconvex partial differential equations [15], structural limit analysis, discretized optimal control problems with distributed parameters, information theory, and network communication. Particularly, if W (x) = 0, the problem (P) is directly related to certain successive quadratic programming methods ( [9,10,18]).
Moreover, if ℓ l = {0} and ℓ u = {1}, the problem leads to one of the fundamental problems in combinatorial optimization, namely, the integer programming problem [6]. Due to the nonconvexity of the cost function and the inequality constraints, traditional KKT theory and direct methods can only be used for solving the problem (P) to local optimality (cf. [23,25,26]). It was shown (see Murty and Kabadi (1987) [28] and Pardalos and Schnitger (1988) [29]) that if a pointx is degenerate (i.e., there is a componentx i ofx such that it lies at either the upper or lower bound and also has the gradient ∇ xi P (x) = 0), then even determining whether it is a local solution to a constrained quadratic programming problem (W (x) = 0) is an NP-complete problem. Therefore, necessary and sufficient conditions for global optimality are fundamentally important in nonconvex minimization problems. Much effort and progress have been made on solving box constrained nonconvex minimization problems during the last twenty years (see, for example, Al-Khayyal and Falk (1983), Sherali and Alameddine (1992), , Floudas (2000), and Floudas et al ( , 2005 and much more).
Canonical duality theory is a newly developed,potentially useful methodology [12,13]. This theory is composed mainly of a canonical dual transformation and a triality theory. The canonical dual transformation can be used to formulate perfect dual problems with zero duality gap, while the triality theory provides sufficient conditions for identifying global and local optimizers. The canonical duality theory has been used successfully for solving many difficult global optimization problems including polynomial minimization [14,19], nonconvex quadratic minimization with spherical, quadratical, box, and integer constraints (see [6,16,17,18]). Detailed discussion on canonical duality theory and its extensive applications can be found in the recent review articles [15,21].
In this paper, we shall demonstrate applications of the canonical duality theory by solving the general nonconvex problem given in (1) and its special cases. As indicated in [14], the key step of this canonical dual transformation is to introduce a geometrical operator such that both the nonconvex function W (x) and the constraints can be written in the so-called canonical form [12]. In order to do so, we assume without loss of generality that ℓ u = −ℓ l = ℓ necessary, a simple linear transformation can be used to convert the problem to this form). Thus, the geometrical operator Λ : R n → R 1+n and the associated canonical function U can be introduced as the following: where Clearly, the canonical function U (y) is convex and its effective domain is Thus, the primal problem (1) can be written in the following unconstrained canonical form (P) : min{P (x) = U (Λ(x)) + Q(x) : x ∈ R n }.
Let y * = ς σ ∈ R 1+n be a dual variable of y. The sup-Fenchel conjugate of U (y) can be defined by where From the theory of convex analysis, the following extended canonical duality relations (see [12]) hold: By the canonical dual transformation developed in [13,14], the canonical dual function of P (x) is defined by where Q Λ (y * ) is the so-called Λ-canonical dual transformation (see [12]) defined by where the notation sta { * : x ∈ R n } represents finding the stationary point of the statement { * } with respect to x ∈ R n , G(ς, σ) is a symmetrical matrix, defined by and Diag (σ) ∈ R n×n denotes a diagonal matrix with {σ i } (i = 1, 2, . . . , n) as its diagonal entries. Let S a ⊂ Y * denote the dual feasible space: The canonical dual problem of (P) can be proposed as the following Theorem 1 (Canonical Duality Theorem). Problem (P d ) is canonically dual to the primal problem (P) in the sense that ifȳ * = (ς,σ) T is a KKT point of (P d ), then the vector defined byx is a KKT point of (P), and Proof. Suppose thatȳ * = (ς,σ) T is a KKT point of (P d ), then the criticality condition δ ς P d (ς,σ) = 0 (where δ ς P d stands for Gâteaux derivative of P d with respect to ς) leads to In terms ofx = [G(ς,σ)] −1 c, the equation (14) is the canonical dual equation ς = 1 2 |Bx| 2 − α =ξ and the condition 1 2 |Bx| 2 ≥ 0 impliesς ≥ −α. On the other hand, the criticality condition for δ σ P d (ς,σ) ≤ 0 leads to the KKT condition This shows thatx(ς,σ) is also a KKT point of the primal problem (P). By the complementarity condition in (15), we haveσ i ℓ i =x 2 iσ i , i = 1, 2, . . . , n. Thus, in term ofx = [G(ς,σ)] −1 c and the Fenchel-Young equality 1 This proves the theorem.
Theorem 1 shows that there is no duality gap between the primal problem (P) and its canonical dual (P d ). In the dual feasible space S a , although it is required that det G(ς, σ) = 0, which is essentially not a constraint since in the case that the matrix G(ς, σ) is singular, the vector x = [G(ς, σ)] −1 c = 0 ∈ R n is a trivial KKT point of the primal problem.
It is known that the KKT conditions are only necessary for local minimizers of the nonconvex quadratic programming problem (P). The next section will show that the global and local extrema of the primal problem (P) depend on the canonical dual solutions.
2. Global and local optimality criteria. In order to identify global and local extrema among the KKT points of the nonconvex problem (P), we need to introduce some useful feasible spaces: Theorem 2 (Triality Theorem). Suppose that the vectorȳ * = (ς,σ) T is a KKT point of the canonical dual function P d (y * ) andx = [G(ς,σ)] −1 c. Ifȳ * = (ς,σ) T ∈ S + a , thenȳ * is a global maximizer of P d on S + a , the vectorx is a global minimizer of P on X a , and holds, or Proof. In the canonical form of the primal problem (5), replacing U (Λ(y)) by the Fenchel-Young equality (Λ(x)) T y * − U ♯ (y * ), the Gao-Strang type total complementary function (see [22]) associated with (P) can be obtained as By Theorem 1 we know that the vectorȳ * ∈ S a is a KKT point of the problem (P d ) if and only ifx = [G(ς,σ)] −1 c is a KKT point of the problem (P) and Since the total complementary function Ξ is a saddle function on R n × S + a , i.e., it is convex in x ∈ R n and concave in y * ∈ S + a . Thus, we have due to the fact that U (Λ(x)) = max otherwise.
From Theorem 1 we have (18). On the other hand, ifȳ * ∈ S − a , the matrix G(ς,σ) is negative definite. In this case, the total complementary function Ξ(x, y * ) defined by (21) is a so-called super-Lagrangian (see [12]), i.e., it is locally concave in both x ∈ X o ⊂ X a and y * ∈ S o ⊂ S − a . Thus, by the triality theory developed in [12], we have either This proves the statements (19) and (20).
In the case that W (x) = 0, the problem (P) is reduced to a box constrained quadratic minimization problem:

DAVID Y. GAO
The canonical dual problem for this box constrained quadratic minimization problem is a special case of the problem (P d ) (see [18]), i.e., where G q (σ) = A + 2Diag (σ) and Furthermore, we let is positive definite}. Then combining Theorem 1 and 2, we have the following result: Ifσ ∈ S + q , thenσ is a global maximizer of P d q (σ) on S + q , the vectorx is a global minimizer of the problem (P q ), and This Corollary shows that the box constrained nonconvex quadratic minimization problem (P q ) can be converted to a concave maximization canonical dual problem Clearly, if for a given matrix A ∈ R n×n and σ ∈ S + q such that the vector c is in the collum space of G q (σ) = A + 2Diag (σ), the canonical dual problem (27) has a unique solutionσ which leads to a global minimizerx = G q (σ) −1 c. Based on this Corollary, certain algorithms can be developed for solving the box constrained nonconvex minimization problem (P q ).
3. Integer programming and boolean least squares problem. As it was indicated in [10] and in Section 1 of this paper that the integer programming is a special case of box constrained optimization. Particularly, if we let X ip be a subset of X a , defined by where the notation {−1, 1} n denotes integer vectors of R n with components either −1 or 1, the primal problem (P) is a nonconvex integer programming problem (denoted by (P ip )): Canonical duality theory for solving quadratic 0 -1 integer programming problems was first studied in the joint work with Fang et al [6]. In the current nonconvex problem (P ip ), we have ℓ i = 1 (i = 1, 2, . . . , n) so that the inequality constraints in X ip can be written in the canonical form ǫ = {x 2 i − 1} ≤ 0 ∈ R n . It is easy to prove that the canonical duality relations Clearly, the complementarity condition leads to the integer constraint Thus, replacing the dual feasible spaces S a and S + a by respectively, the canonical dual problem of (P ip ) can be proposed as (31) Similar to Theorem 1 and 2, we have is canonically dual to the primal problem (P ip ) in the sense that ifȳ * = (ς,σ) T ∈ S ip is a KKT point of (P d ip ), then the vector x = [G(ς,σ)] −1 c is a KKT point of (P ip ), and (32) Ifȳ * = (ς,σ) T ∈ S + ip , thenȳ * is a global maximizer of P d ip on S + ip , the vectorx is a global minimizer of P on X ip , and Since the canonical dual function P d ip (y * ) is concave on S + ip , this theorem shows that the discrete nonconvex integer minimization problem (P ip ) can be converted into a continuous concave maximization dual problem (P d ip ), which can be solved by any well-developed nonlinear optimization method. Detailed study on the canonical duality theory for solving integer programming was given in [6].
The canonical duality theory can be applied to solve many constrained nonconvex integer programming problems. Particularly, if the function W (x) in (P ip ) is an indicator of the feasible space X b := {x ∈ X ip | Bx = b}, i.e., where b ∈ R m is a given vector, and the quadratic function Q(x) can be written in the least square form: where A could be any given p × n matrix and c ∈ R p is a given vector, then the primal problem (P ip ) is the so-called Boolean least squares problem: This problem arises from a large number of applications in communication systems such as the channel decoding, MIMO detection, multiuser detection, equalization, resource allocation in wireless systems, etc (see [4,5]). Since the quadratic function Q(x) is concave, its minimizers are located on the boundary of the feasible set X ip .
Traditional methods for solving this discrete concave minimization problem (P b ) is very difficult. We assume that m < n and rank B = m so that the problem (P b ) is not overconstrained. Let N B ∈ R n×r (r = n − m) be the null space of B, i.e., BN B x o = 0 ∈ R m ∀x o ∈ R r , thus, the general solutions of the constraint equation Bx = b is where x b ∈ R n is a particular solution of Bx = b, i.e., Bx b = b, and x o ∈ R r is a free vector. We let . Then, on the dual feasible space the canonical dual of the Boolean least squares problem can be formulated as is a critical point of (P) and Moreover, if G a (σ) is positive definite, thenσ is a global maximizer of (P d b ) on In the case that there is no equilibrium constraint, the primal problem (P b ) is a so-called lattice-decoding-type problem: In this case, the canonical dual problem has a simple form: Corollary 2. Ifσ ∈ S b is a KKT point the canonical dual problem (P d bo ), then x = −[G a (σ)] −1 A T c is a KKT point of the Boolean least squares problem (P bo ). If σ ∈ S + b , thenx is a global minimizer of P bo (x) on X ip and P bo (x) = min A detailed study for the canonical duality approach for solving quadratic 0 -1 integer programming can be found in [6].

4.2.
Two-dimensional box constrained problem. We now consider the following box constrained concave minimization problem where A = {a ij } is an arbitrarily given 2×2 matrix. If we choose a 11 = −1.0, a 12 = 0, a 21 = −1, a 22 = −2, and c = (3, 2) T , the dual function has four critical points: The corresponding primal solutions It is easy to check that we have only one critical point σ 1 ∈ S + b . By Corollary 2, we know that x 1 is a global minimizer (see Fig. 2). It is easy to verify that P (x k ) = P d (σ k ) k = 1, 2, 3, 4 and P bo (x 1 ) = −20.5 < P bo (x 2 ) = −8.5 < P bo (x 3 ) = −6.5 < P bo (x 4 ) = −2.5 We note that if the inequality constraints in (48) are replaced by x 2 i = 1, i = 1, 2, then the problem (47) is a Boolean lease square problem. Actually, since all the four dual solutions σ k = 0 (k = 1, 2, 3, 4), it turns out that the primal solutions x k are all integer vectors. Applications of the canonical duality theory have been demonstrated by solving the box constrained nonconvex optimization problems (P), (P ip ), and their special cases. Theorems proved in this paper show that by the canonical dual transformation, nonconvex primal problems can be converted into canonical dual problems. By the fact that the canonical dual function P d (y * ) is concave on the dual feasible space S + a , if S + a is non-empty, i.e., the dual problem (P d ) has at least one KKT point in S + a , the canonical dual problem max can be solved by well-developed deterministic optimization methods. Since the primal problem (P) could be NP-hard, a conjecture can be proposed as the following.
Conjecture. The box constrained nonconvex minimization problem (P) is NP-Hard if its canonical dual (P d ) has no KKT point in S + a .
Generally speaking, the function W (x) in the primal problem (P) could be any given nonconvex/nonsmooth function as long as it can be written in the canonical form, i.e., there exists a nonlinear operator y = Λ(x) and a canonical function U (y) such that W (x) = U (Λ(x)). Thus, the theorems proposed in this paper can be generalized for solving more complicated problems in global optimization.
In nonlinear analysis and mechanics, the canonical duality theory has been successfully applied for solving many nonconvex/nonsmooth variational/boundary value problems [11,20]. In finite deformation elasticity, the canonical duality theory is the so-called Gao principle [27]. Detailed discussion on canonical duality theory and its extensive applications can be found in the recent review articles [15,21].