• No results found

Canonical duality for binary quadratic problems

lems

5.3.1

Canonical dual problem

The integer constraints in the problem (Pbqp) is treated as equality ones, x2i = 1, i =

in Chapter 3, we can easily get the canonical dual problem. Let

G( ) = Q + diag( ), and

Sa ={ 2 Rn| det(G( )) 6= 0}.

The canonical dual function is formulated as

⇧d( ) = 1

2f

TG( ) 1f 1

2e

T , (5.22)

and the canonical dual problem is defined as

(Pbqpd ) ext{⇧d( )| 2 Sa}. (5.23)

Notice that if f = 0, the function ⇧d( ) is equal to the objective function in

the problem (5.12), which is thus same to the problem of maximizing ⇧d over the

positive semidefinite region defined by

Sc+ ={ 2 Sa| G( ) ⌫ 0}.

5.3.2

Global optimality conditions

The following theorem characterizes the primal-dual relation.

Theorem 30 (Complementary-dual principle) If ¯ 2 Sa is a critical point of

⇧d( ), the vector ¯ x = G( ¯ ) 1f (5.24) is a feasible solution of (Pbqp). Let ¯x2 X and ¯ = ¯x f x (Q ¯¯ x). (5.25)

If ¯ 2 Sa, then ¯ is a critical point of ⇧d( ).

For both statements, we have

⇧( ¯x) = ⇧d( ¯ ). (5.26) Proof: From the assumption of ¯ being a critical point and the definition of ¯x, we have r ⇧d( ¯ ) = ⇢ 1 2f TG( ¯ ) 1I iG( ¯ ) 1f n i=1 1 2e = 1 2x¯ x¯ 1 2e = 0,

where Ii denotes an all-zero matrix except for the entry (i, i), which is equal to one.

It is verified that ¯x is a feasible solution of the primal problem (Pbqp). The definition

of ¯x in equation (5.24) also implies that ¯x and ¯ satisfy the equilibrium equation G( ¯ ) ¯x f = 0, (5.27)

from which we have

¯

x ¯ = f Q ¯x. (5.28) The equality (5.28) is further equivalent to

¯

x x¯ ¯ = ¯x f x (Q ¯¯ x). Thus, we have

¯ = ¯x f x (Q ¯¯ x) = diag( ¯x)f diag( ¯x)Q ¯x. (5.29)

Now we are ready to prove the equation (5.26):

⇧d( ¯ ) = 1 2f TG( ¯ ) 1f 1 2e T¯ = 1 2f Tx¯ 1 2e T (diag( ¯x)f diag( ¯x)Q ¯x) = 1 2x¯ TQ ¯x fTx = ⇧( ¯¯ x).

Let ¯ be defined by the equation (5.25). From the equivalence between (5.27) and (5.29), it can be easily proved that ¯ is a critical point if ¯ 2 Sa.

The theorem is proved. 2

Besides the positive semidefinite regionS+

c , we also introduce the negative semidef-

inite region:

Sc ={ 2 Sa| G( ) 0}.

First,S+

c and Sc are convex sets. Second, from the expression of Hessian of the dual

function ⇧d( ), we notice that ⇧d( ¯ ) is concave on S+

c and convex on Sc . Hence,

any critical point in S+

c is a maximizer of ⇧d over Sc+, and any critical point in Sc

is a minimizer of ⇧d over S

c . We have the following result.

Theorem 31 For any given matrix Q 2 Sn and vector f 2 Rn, suppose ¯ is a

critical point of the dual function ⇧d( ) and ¯x = G( ¯ ) 1f . 1. If ¯ 2 S+

c , then ¯x is a global minimizer of ⇧(x) on X ; we have

⇧( ¯x) = min

x2X⇧(x) = max2S+ c

2. If ¯ 2 Sc , then ¯x is a global maximizer of ⇧(x) on X ; we have ⇧( ¯x) = max

x2X ⇧(x) = min2Sc

⇧d( ) = ⇧d( ¯ ). (5.31)

Proof: The last equalities in (5.30) and (5.30) are obvious, since ⇧d( ) is concave

onS+

c and convex on Sc . For the minimizer, the weak duality in (3.30) shows that

we always have

⇧d( ¯ ) = max

2Sc+

⇧d( ) min

x2X⇧(x).

Thus, by Theorem 30, we have

⇧( ¯x) = ⇧d( ¯ ) = min x2X⇧(x),

and it is proved that ¯x is a minimizer of ⇧(x) on X .

The second part of the theorem can be similarly proved by applying the fact that the maximization of ⇧(x) is equivalent to the minimization of ⇧(x) since there are finite feasible solutions inX . 2 The equation (5.30) shows that the critical point ¯ is the maximizer of the dual function ⇧d( ¯ ) over S+

c . On the other hand, if the maximizer is a critical point of

⇧d( ¯ ), we can claim that the corresponding ¯x defined by the equation (5.24) is a

global optimal solution of the primal problem. Follow the discussion in Section 3.3.4, the maximizer can be found by solving the following SDP problem:

max ,⌧ ⌧ (5.32) s.t. ✓ 2G( ) f fT 1 2e T ◆ ⌫ 0

Let ( ¯ , ¯⌧ ) be a maximizer. Then, if G( ¯ ) 0, ¯ must be a critical point of ⇧d and ¯

x must be a global solution. While if det(G( ¯ )) = 0, ¯ may not be a critical point and ⇧d( ¯ ) is only a lower bound for the primal problem. It shows that the integer

problem can be converted into a convex optimization problem, which can be solved efficiently by well-developed convex optimization methods.

Corollary 32 Suppose that ¯ is a maximizer of the problem (5.32) and ¯x = G( ¯ ) 1f .

If G( ¯ ) 0, then ¯x is a global optimal solution of the problem (Pbqp).

5.3.3

Existence and uniqueness

The following result gives a criterion of existence and uniqueness of a critical point inS+

c .

Theorem 33 If, for any 0 with det(G( 0)) = 0 and G( 0)⌫ 0 and any 2 Sc+,

we have

lim

t!0+⇧

d(

0+ t ) = 1, (5.33)

then the canonical dual problem (Pd

Proof: The assumption in the equation (5.33), plus the fact that ⇧d( ) approaches

to minus infinity as any entry of increases infinitely, implies that the function ⇧d( ) is coercive on the convex set S+

c . Since ⇧d( ) is concave over Sc+, it has at

least one maximizer, which must be a critical point. We use ¯ to denote one of the maximizers. The uniqueness results from the fact that the Hessian of ⇧d( ) is

negative definite at the critical point ¯ . 2

5.3.4

Examples

Example 1 Let Q =  2 3 3 1 , and f = ✓ 1 2 ◆ .

In this case, the dual function has four critical points,

¯1 = (4, 6), ¯2 = (6, 2), ¯3 = (0, 0), and ¯4 = ( 2, 4),

with function values

⇧d( ¯1) = 5.5 < ⇧d( ¯2) = 3.5 < ⇧d( ¯3) = 1.5 < ⇧d( ¯4) = 4.5.

The corresponding solutions of the primal problem are

¯

x1 = ( 1, 1), ¯x2 = (1, 1), ¯x3 = (1, 1), and ¯x4 = ( 1, 1).

By checking the eigenvalues of Q+diag( ), we find Q+diag( 1)⌫ 0, Q+diag( 4)

0, and Q+diag( 2) and Q+diag( 3) are indefinite. Thus, Theorem 30 and Theorem

31 are demonstrated. Example 2 Let Q = 2 4 922 1409 16 1 6 80 3 5 , and f = 0 @ 26 1 1 A . The dual function has eight critical points:

¯1 = (12, 128, 73), ¯2 = (10, 119, 72), ¯3 = (11, 134, 7),

¯4 = (11, 125, 8), ¯5 = (19, 21, 78), ¯6 = (3, 12, 79),

¯7 = (20, 15, 2). ¯8 = (2, 6, 1)

The corresponding primal solutions are ¯ x1 = (0, 1, 1), x¯2 = (1, 1, 1), x¯3 = (0, 1, 0), ¯ x4 = (1, 1, 0), x¯5 = (1, 0, 1), x¯6 = (0, 0, 1), ¯ x7 = (1, 0, 0). x¯8 = (0, 0, 0)

with function values

⇧( ¯x1) = 97, ⇧( ¯x2) = 96, ⇧( ¯x3) = 64, ⇧( ¯x4) = 64,

⇧( ¯x5) = 47, ⇧( ¯x6) = 39, ⇧( ¯x7) = 9, ⇧( ¯x8) = 0.

It can be verified that Q + diag( ¯1) is positive definite, Q + diag( ¯8) is negative

definite and all Q + diag( ¯i) for i = 2, . . . , 7 are indefinite. The function values show

that ¯x1 is the minimizer and ¯x8 is the maximizer. Thus, Theorem 30 and Theorem

31 are explained.

Related documents