Page 1
SecondOrder Cone Relaxations for Binary Quadratic
Polynomial Programs
Bissan Ghaddar*, Juan C. Vera, Miguel F. Anjos†
Department of Management Sciences, University of Waterloo, 200 University Av. W., Waterloo, Ontario, N2L 3G1
bghaddar@uwaterloo.ca, jvera@uwaterloo.ca, anjos@stanfordalumni.org
Several types of relaxations for binary quadratic polynomial programs can be obtained using linear, second
order cone, or semidefinite techniques. In this paper, we propose a general framework to construct conic
relaxations for binary quadratic polynomial programs based on polynomial programming. Using our frame
work, we rederive previous relaxation schemes and provide new ones. In particular, we present three relax
ations for binary quadratic polynomial programs. The first two relaxations, based on secondorder cone and
semidefinite programming, represent a significant improvement over previous practical relaxations for several
classes of nonconvex binary quadratic polynomial problems. From a practical point of view, due to the
computational cost, semidefinitebased relaxations for binary quadratic polynomial problems can be used
only to solve small to midsize instances. To improve the computational efficiency for solving such problems,
we propose a third relaxation based purely on secondorder cone programming. Computational tests on dif
ferent classes of nonconvex binary quadratic polynomial problems, including quadratic knapsack problems,
show that the secondorder conebased relaxation outperforms the semidefinitebased relaxations that are
proposed in the literature in terms of computational efficiency and is comparable in terms of bounds.
Key words: binary quadratic polynomial program, polynomial programming, sumofsquares, secondorder
cone.
1. Introduction
Binary quadratic polynomial problems (BQPP) can be expressed as optimizing a quadratic poly
nomial objective subject to quadratic polynomial equalities and inequalities. Several types of
relaxations can be obtained using linear, secondorder cone [13; 15], or semidefinite techniques
[3; 7; 14; 25]. In this paper we study relaxations for general BQPPs based on polynomial program
ming.
∗Research supported by a Canada Graduate Scholarship from the Natural Sciences and Engineering Research Council
of Canada.
†Currently on leave at Universit¨ at zu K¨ oln, Institut f¨ ur Informatik, Pohligstrasse 1, 50969 K¨ oln, Germany. Research
supported by the Natural Sciences and Engineering Research Council of Canada, and by a Fellowship from the
Humboldt Foundation.
1
Page 2
2
Polynomial programming includes a broad class of problems and is known to be NPhard.
Polynomial programming problems can be relaxed to tractable problems by using sumofsquares
(SOS) decompositions which lead to semidefinite programming (SDP) relaxations. This technique
was first proposed by Shor [35] to obtain bounds on the optimal value of the unconstrained case.
This idea was then generalized by Parrilo [26; 28] and Lasserre [18] for the constrained case.
In this paper, we use a characterization of nonnegative linear polynomials over the ball to
propose secondorder cone (SOC) relaxations of binary quadratic polynomial problems. We use
the polynomial programming framework to rederive, compare and strengthen existing relaxation
schemes. We present a new secondorder and semidefinitebased construction where we are able to
theoretically show that the resulting relaxations provide bounds stronger than other computation
ally practical semidefinitebased relaxations proposed in the literature. Additionally, our proposed
framework enables us to isolate expensive components of existing relaxations, namely the semidefi
nite terms. By removing the semidefinite terms, we obtain relaxations based purely on secondorder
cones. We present computational tests exploring the performance of these relaxations, comparing
them to existing ones in terms of bounds and computational time on general quadratic constrained
problems, quadratic linear constrained problems, and quadratic knapsack problems. The compu
tational experiments confirm our theoretical results where we obtain that the SOCSDPbased
relaxations give the best bounds. Our experiments also show that the purely SOCbased relaxations
produce bounds that are competitive with the existing SDP bounds but computationally much
more efficient. Furthermore, our approach can be in principle extended to mixed binary polynomial
programs where some of the variables are continuous.
The paper is organized as follows. In Section 2, we present an overview of polynomial pro
gramming and its SOS and SOC relaxations. In Section 3, we describe our solution methodology
and present several relaxations for the binary quadratic polynomial problem including our three
new proposed relaxations. In Section 4, we apply our proposed relaxations to the three classes of
problems mentioned above, and theoretically compare them to other existing relaxations from the
literature. In Section 5, we report computational results for these problems. Finally, conclusions
and future research directions are discussed in Section 6.
2.
2.1.
Background
Preliminaries
Given an ntuple α = (α1,··· ,αn) where αi∈ Z+, the total degree of the monomial xα:=
xα1
n
degree at most d. A polynomial is a finite linear combination of monomials
1xα2
2···xαn
is the nonnegative integer d = α :=?n
iαi. There are N =?n+d
d
?
monomials of
Page 3
3
f(x)=
?
α
cαxα=
?
α
cαxα1
1···xαn
n,
where the vector of coefficients c∈RN. We denote the cone of real polynomials (of degree at most
d) that are SOS by Ψ ⊂ R[x] (resp. Ψd) where R[x] := R[x1,··· ,xn] (resp. Rd[x]) denotes the
set of polynomials in n variables with real coefficients (resp. of degree at most d). Notice that in
particular Ψd= {?N
denote Pd(S) := {p(x) ∈ Rd[x] : p(s) ≥ 0 for all s ∈ S} to be the cone of polynomials of degree at
most d that are nonnegative over S.
i=1pi(x)2: p(x) ∈ R?d
2?[x]} and Ψd= Ψd−1for every odd d. Given S ⊆ Rn, we
2.2.Polynomial Programming
Consider the multivariate polynomials f(x) and gj(x) for 1 ≤ j ≤ m with x ∈ Rn. A polynomial
programming problem has the form:
(PPP) sup f(x)
s.t. gj(x)≥0,
1≤j ≤m.
Equality constraints of the form hj(x)=0 can be included as they can be expressed as the inequal
ity constraints hj(x)≥0 and hj(x)≤0.
Solving polynomial programming problems is an area being actively studied. For the uncon
strained case, Shor introduced the idea of computing the minimum value λ such that λ−f(x) is a
SOS to obtain an upper bound for the supremum of f [35]. Such a minimum λ can be computed
in polynomialtime using semidefinite programming. This idea was further developed by Parrilo
[26] and Parrilo and Sturmfels [29] for the constrained case using SOS decompositions. Lasserre
[18] proposed a general solution approach for polynomial optimization problems via semidefinite
programming using methods based on moment theory. Refinements of such ideas have been used
in several instances. de Klerk and Pasechnik [8] approximated the copositive cone via a hierarchy
of linear or semidefinite programs of increasing size using decompositions into sumofsquares and
polynomials with nonnegative coefficients. Kojima, Kim, and Waki exploited the sparsity of the
polynomials to reduce the size of the semidefinite problem [16]. Pe˜ na, Vera, and Zuluaga [30] pre
sented solution schemes exploiting the equality constraints. In addition, the idea of approximating
a set of nonnegative polynomials is also present in the work of several authors such as Nesterov
[24], Parrilo [28; 27], Sturmfels, Demmel, and Nie [37], Laurent [20], and Zuluaga, Vera, and Pe˜ na
[40].
Page 4
4
Consider λ to be the optimal value for (PPP), then λ is the smallest value such that λ−f(x)≥0
for all x∈S :={x:gj(x)≥0; 1≤j ≤m}. As a result, we can express problem (PPP) as:
(PPD) inf λ
s.t. λ−f(x)≥0
∀x∈S.
(1)
To obtain computable relaxations (via SDP) of (1), one can use a SOS decomposition with
restricted degree of the (unknown) polynomials. This can be rephrased in terms of a linear system
of equations involving positive semidefinite matrices [40]. Thus, solving a polynomial problem can
be relaxed to solving an easier problem involving SOS which can be recast as a semidefinite
programming problem [35; 39].
The condition λ − f(x) ≥ 0 for all x ∈ S is NPhard in general. Relaxing this condition to
λ−f(x)∈K for a suitable K ⊆Pd(S) and defining
z∗
K= inf λ
s.t. f(x)−λ∈K,
we have z∗
K≥ z∗
PP. Finding a good approximation K of Pd(S) is a key factor in obtaining a good
bound of the original problem. At the same time, having a tractable approximation, i.e., one
that uses linear, secondorder, and semidefinite cones, is essential to be able to solve the resulting
relaxation efficiently using interiorpoint methods.
2.3. SumofSquares and SecondOrder Cone Relaxations
Consider a polynomial p(x) of degree d. A necessary condition for the polynomial p(x) to be non
negative for all x ∈ Rnis that the degree of p is even. A sufficient condition is the existence of a
sumofsquares decomposition, i.e., the existence of polynomials q1(x),··· ,qk(x) such that p(x) =
?k
polynomial for all values of x; however the inverse does not hold. A simple counterexample is the
Motzkin polynomial [23].
i=1qi(x)2, or equivalently, p∈Ψ. If p(x) is a sumofsquares polynomial then it is a nonnegative
SOS conditions can be written as SDP constraints by applying the following theorem:
Theorem 1. [35] A polynomial p(x) of degree d is SOS if and only if p(x) = σ(x)TQσT(x),
where σ is a vector of monomials in the xivariables, σ(x) = [xα] with α ≤d
?n+d/2
2and Q ∈ SN
+, N =
d/2
?=σ.
The size of the matrix Q in the corresponding SDP is?n+d/2
equality constraints. If d is fixed, then this problem is solvable in polynomialtime.
The following results will allow us to use secondorder cone relaxations when working with
d/2
?×?n+d/2
d/2
?. In addition, we have?n+d
d
?
nonnegative polynomials over the ball B :={x:?x?2=n}.
Page 5
5
Lemma 1. f(x)∈P1(B) if and only if f(x)=fT
order cone.
?√n
x
?
with f ∈Ln+1, where Ln+1is the second
Further, by the S−Lemma of Yakubovich (see [33]), the nonnegativity over the ball of a polynomial
f(x) of degree two can be represented using SOS:
Lemma 2. f(x)∈P2(B) if and only if f(x)=s(x)+t(n−?x?2), where s(x) is SOS and t∈R+.
The key feature of semidefinite and secondorder cones is their tractability. As a result, we can use
these techniques to compute global upper bounds for (PPP).
3.Binary Quadratic Polynomial Programming
Binary quadratic polynomial programming problem is a classical combinatorial problem. It is the
problem of minimizing or maximizing a quadratic function of several binary variables, subject to
quadratic and linear constraints. The problem can be formally expressed as:
(BQPP) max xTQx+pTx
s.t. aT
jx=bj
∀j ∈{1,··· ,t}
∀j ∈{1,··· ,u}
∀j ∈{1,··· ,v}
∀j ∈{1,··· ,w}
∀i∈{1,··· ,n}.
(2)
cT
jx≤dj
xTFjx+eT
(3)
jx=kj
(4)
xTGjx+hT
jx≤lj
(5)
xi∈{−1,1}
(6)
Note that constraint (6) can be modified to allow some continuous variables. In this paper we focus
on pure binary quadratic polynomial programs although our solution methodology can be applied
to mixedbinary quadratic polynomial programs with bounded continuous variables.
There are many wellknown problems that can be naturally written as binary quadratic poly
nomial problems. For instance, folding of proteins in threedimension by Phillips and Rosen [31],
machine scheduling and unconstrained task allocation by Alidaee, Kochenberger, and Ahmadian
[1], capital budgeting and financial analysis such as in Laughhunn [19], as well as other examples
arising in physics and engineering applications such as the spin glass problem and circuit board lay
out design by Gr¨ otschel, J¨ unger, and Reinelt [10]. Furthermore, Boros and Hammer [4] and Boros
and Prekopa [5] formulated many satisfiability problems as BQPPs. In addition, there are several
applications related to combinatorial problems such as the singlerow facility layout problem [2]
and the quadratic assignment problem [21].
Page 6
6
3.1.Polynomial ProgrammingBased Relaxations
Using (1), (BQPP) is equivalent to
min λ
s.t. λ−q(x)∈P2(H ∩S),
where q(x) =?
lj}, and H :={−1,1}n. Note that even checking if a polynomial is in P2(H) is NPhard, therefore
tractable approximations of P2(H ∩S) are needed. A hierarchy of approximations to P2(H ∩S) is
obtained using the cones
?
ij
?
⊆P2(H ∩S),
i,jQijxixj+?
ipixi, S = {x : aT
jx = bj,cT
jx ≥ dj,xTFjx+eT
jx = kj,xTGjx+hT
jx ≥
Kr:=Ψr+2+
?
(1−x2
i)Rr[x]+
?
jx)Rr[x]+
(bj−aT
jx)Rr+1[x]+
?
j
(dj−cT
jx)Ψr+1
?
+
j
(kj−xTFjx−eT
?
j
(lj−xTGjx−hT
jx)Ψr
∩R2[x]
for an integer r ≥0. The result is a hierarchy of relaxations:
(BQPPKr) min λ
s.t. λ−q(x)∈Kr
(7)
whose optimal value converges to the optimal value of (BQPP) due to the fact that at the limit the
cone Krcontains the interior of P2(H ∩S). The following theorem follows by applying Corollary 1
of [30] and Putinar’s theorem [34]
Theorem 2. The sequence of cones Krsatisfies
Kr⊆Kr+1⊆···⊆P2(H ∩S) and int(P2(H ∩S))⊆
∞
?
r=0
Kr⊆P2(H ∩S).
Hence λ∗
BQPPKr↑z∗
BQPP.
The size of the relaxations produced in the previous theorem grows exponentially in r. For this
reason, instead of looking at the hierarchy of relaxations, we will concentrate on the first and
simplest relaxation where r =0,
K0=Ψ2+
?
(kj−xTFjx−eT
i
(1−x2
i)R0+
?
jx)R0+
j
(bj−aT
jx)R1[x]+
?
?
j
(dj−cT
jx)R+
0
+
?
jj
(lj−xTGjx−hT
jx)R+
0.
We study how to improve the approximation of P2(H ∩S) using variations of the cone K0. The
fundamental tool that we use to construct such inner approximations of P2(H ∩S) is Lemma 1,
Page 7
7
a representation theorem for nonnegative linear polynomials over B which results in secondorder
cone conditions. These yield stronger approximations than K0with an insignificant impact on the
computational time.
3.2. New Conic Relaxations of BQPP
In this section, we present three relaxations for the BQPP problem. Two of these relaxations are
based on secondorder cone and semidefinite programming and the final relaxation is solely based
on secondorder cone programming.
3.2.1. SOCSDPbased Relaxations of BQPP
Recall the previous polynomial formulation of the binary quadratic polynomial problem. First,
notice that x∈H implies ?x?2
2=n. Therefore, S ∩H ⊆B and by defining¯K0as
?
jx)P1(B)+
j
¯K0=P2(B)+
?
(dj−cT
i
(1+xi)P1(B)+
i
(1−xi)P1(B)+
(kj−xTFjx−eT
?
i
(1−x2
i)R0+
?
?
j
(bj−aT
jx)R1[x]
+
?
j
?
jx)R0+
j
(lj−xTGjx−hT
jx)R+
0,
we have¯K0⊆P2(S ∩B∩H)=P2(S ∩H).
Using Lemmas 1 and 2, we can write the condition λ−q(x)∈¯K0as
?
+
ηj(x)(dj−cT
λ−q(x)=s(x)+
i
(1+xi)αi(x)+
?
?
i
(1−xi)βi(x)+
θj(kj−xTFjx−eT
?
jx)+
i
γi(1−x2
?
?
i)+
?
j
δj(x)(bj−aT
jx)
?
j
jx)+
jj
ξj(lj−xTGjx−hT
jx),
with s(x) =?1 xT?S
ηT
j
x
We then obtain the following relaxation of (BQPP):
?1
x
?
and S ∈ Sn+1
+
, αi(x) = αT
i
?√n
x
, βi(x) = βT
i
?√n
x
?
, and ηj(x) =
?√n
?
where αi,βi,ηj∈Ln+1, δj(x)∈R1[x], γi,θj∈R, and ξj∈R+.
(BQPPSS) min λ
s.t. λ−q(x)=?1 xT?S
?1
x
?
+
?
i)+
i
(1+xi)αT
i
?√n
x
?
jx)+
+
?
?
i
(1−xi)βT
i
?√n
x
?
+
?
?
αi,βi,ηj∈Ln+1,
i
γi(1−x2
θj(kj−xTFjx−eT
?
j
δj(x)(bj−aT
j
(dj−cT
jx)ηT
j
?√n
x
?
+
j
jx)+
?
j
ξj(lj−xTGjx−hT
ξj∈R+.
jx),
S ∈Sn+1
+
,γi,θj∈R,
To strengthen this relaxation we can add valid inequalities to the original problem (BQPP) which
is equivalent to adding more variables to the relaxation due to the next lemma.
Page 8
8
Lemma 3. For any S, d, and f ∈Rd[x]
Pd(S ∩{x:f(x)≥0})⊇Pd(S)+f(x)Pd−deg(f)(S).
Notice that products of linear constraints, such as (dk− cT
cT
kx)(1 + xi), (dk− cT
kx)(1 − xi), (dk−
kx)(dl−cT
inequalities and can be added to (BQPPSS) to further strengthen the relaxation. Hence we obtain
lx), (1−xj)(1−xi), (1+xj)(1+xi), and (1−xj)(1+xi) are also considered as valid
(BQPPSS+) min λ
s.t. λ−q(x)=?1 xT?S
?1
x
?
+
?
i)+
i
(1+xi)αT
i
?√n
x
?
jx)+
+
?
?
i
(1−xi)βT
i
?√n
x
?
+
?
?
?
?
?
αi,βi,ηj∈Ln+1,
i
γi(1−x2
θj(kj−xTFjx−eT
σik(dk−cT
νkl(dk−cT
ωij(1+xi)(1+xj)+
?
j
δj(x)(bj−aT
j
(dj−cT
jx)ηT
j
?√n
x
?
+
j
jx)+
?
?
j
ξj(lj−xTGjx−hT
µik(dk−cT
?
φij(1−xi)(1+xj)
γi,θj∈R,
jx)
+
i,k
kx)(1+xi)+
i,k
kx)(1−xi)
+
k≤l
kx)(dl−cT
lx)+
?
i≤j
τij(1−xi)(1−xj)
+
i≤j
i,j
S ∈Sn+1
+
,ξj,σik,µik,νkl,τij,ωij,φij∈R+.
3.2.2.Pure SOCbased Relaxations of BQPP
The relaxation (BQPPSS) can further be relaxed by removing the positive semidefinite variable
leading to the following relaxation:
min λ
s.t. λ−q(x)=
?
+
i
(1+xi)αT
i
?√n
x
?
jx)+
+
?
?
jx),
i
(1−xi)βT
i
?√n
x
?
+
?
?
i
+
γi(1−x2
?
i)
?
?
j
δj(x)(bj−aT
ξj(lj−xTGjx−hT
γi,θj∈R,
j
(dj−cT
jx)ηT
j
?√n
x
j
θj(kj−xTFjx−eT
jx)
+
j
αi,βi,ηj∈Ln+1,ξj∈R+.
One type of valid inequalities that we consider for BQPP is:
−1≤xixj≤1.
(8)
These inequalities are not violated in the presence of the SDP term. However, once the SDP term
is removed these constraints are no longer satisfied and adding them will strengthen the SOC
relaxation.
Page 9
9
Hence, we obtain our proposed SOCbased relaxation:
(BQPPSOC) min λ
s.t. λ−q(x)=
?
+
i
(1+xi)αT
i
?√n
x
?
jx)+
+
?
?
jx)+
i
(1−xi)βT
i
?√n
x
?
+
?
?
i
+
γi(1−x2
?
?
i)
?
?
j
δj(x)(bj−aT
ξj(lj−xTGjx−hT
γi,θj∈R,
j
(dj−cT
?
ξj,h+
jx)ηT
j
?√n
x
j
θj(kj−xTFjx−eT
jx)
+
ji<j
h+
ij(1+xixj)+
i<j
h−
ij(1−xixj)
αi,βi,ηj∈Ln+1,
ij,h−
ij∈R+.
By construction we have the following theorem relating the three presented relaxations:
Theorem 3. Let λ∗
BQPPSOC, λ∗
BQPPSS, and λ∗
BQPPSS+be the optimal solution value of (BQPPSOC),
(BQPPSS), and (BQPPSS+) respectively, then
λ∗
BQPPSOC≥λ∗
BQPPSS≥λ∗
BQPPSS+≥z∗
BQPP.
4.
In this section, we apply our proposed framework to the following classes of constrained BQPPs:
Applications
• General quadratic polynomial problems;
• Quadratic linear constrained problems;
• Quadratic knapsack problems.
First, we start with the most general class of binary quadratic polynomial problems (BQPP)
where we have quadratic and linear constraints. Then we consider the special case with only linear
constraints and finally we consider problems with a single linear constraint. We rederive existing
relaxations that have been proposed in the literature for each of these problems and theoretically
compare our proposed two SOCSDPbased relaxations to them. We show theoretically that we
obtain stronger relaxations based on applying the methodology of Section 3. In addition, in Section
5 we compare the relaxations computationally for each of these three classes of binary quadratic
problems. Our computational results show that more efficient relaxations are obtained if the SDP
term is omitted.
4.1.General Quadratic Polynomial Problems
In this section, we consider the general binary quadratic problem (BQPP). Lasserre [17] introduced
SDP relaxations for binary polynomial programs by approximating P2(H ∩S) using the cone
?
ii
?
for even r ≥0.
Γr:=Ψr+2+
?
(1−x2
i)Ψr+
?
?
(x2
i−1)Ψr+
?
ix−ki)Ψr+
i
(bi−aT
ix)Ψr+
?
i
(aT
ix−bi)Ψr+
?
?
i
(di−cT
ix)Ψr
+
i
(ki−xTFix−eT
ix)Ψr+
i
(xTFix+eT
?
i
(li−xTGix−hT
ix)Ψr
∩R2[x],
Page 10
10
Lemma 4. Γr⊆Kr.
Notice that from the definition of Ψr, Γris equal to Γr−1for all odd r. Taking r =0, we obtain the
following relaxation for the BQPP problem:
(BQPPLas) min λ
s.t. λ−q(x)∈Γ0.
Theorem 4. Let λ∗
BQPPLas, λ∗
BQPPSS, and λ∗
BQPPSS+be the optimal solution value of (BQPPLas),
(BQPPSS), and (BQPPSS+) respectively, then
λ∗
BQPPLas≥λ∗
BQPPSS≥λ∗
BQPPSS+≥z∗
BQPP.
Proof: Define
H1=Ψ2+
?
(ki−xTFix−eT
i
(1−x2
i)R0+
?
ix)R0+
?
i
(bi−aT
ix)R1[x]+
?
(1+xi)P1(B)+
?
?
?
i
(di−cT
ix)P1(B)
+
?
ii
(li−xTGix−hT
ix)R+
0.
H2=H1+(di−cT
H3=H2+
+
ix)P1(B)+
Ψ0(1+xi)(dk−cT
Ψ0(1+xi)(1+xj)+
i
?
i
(1−xi)P1(B)=¯K0.
?
i,k
kx)+
i,k
Ψ0(1−xi)(dk−cT
Ψ0(1−xi)(1−xj)+
kx)+
?
?
k≤l
Ψ0(dk−cT
kx)(dl−cT
lx)
?
i≤ji≤j
i,j
Ψ0(1+xi)(1−xj).
We have
Ψ1=R+
0⊆P1(B)⇒K0⊆H1.
In addition, from Lemma 4, by setting r to zero we have Γ0⊆K0and therefore,
Γ0⊆K0⊆H1⊆H2⊆H3.
?
From Theorem 4, we obtain that (BQPPSS+) provides the best bound for the BQPP problem
while (BQPPSS) has a better bound than Lasserre’s relaxation. Further, as presented in Table 1
the computational complexity of the three problems is similar. Table 1 summarizes the number of
variables (and for SDPs, the dimension) for each of the resulting optimization problems. Recall
that the (BQPP) problem has t linear equalities, u linear inequalities, v quadratic equalities, w
quadratic inequalities, and n binary variables.
Page 11
11
Table 1
Variables
SDP
SOC
Linear Nonnegative 2n+2t+2v +u+w
Linear Free
Problem dimension for various BQPP relaxations.
(BQPPLas)
1, (n+1)×(n+1) 1, (n+1)×(n+1)
(BQPPSS)(BQPPSS+)
1, (n+1)×(n+1)
(2n+u) , (n+1)
?+2?n+1
 (2n+u) , (n+1)
w w+2tn+?t+1
22
?+n2

n+(n+1)t+vn+(n+1)t+v
4.2.Quadratic Linear Constrained Problems
Without loss of generality, we formulate the binary quadratic linear constrained problem as:
(QLCP) max xTQx+pTx
s.t. aT
jx≤bj
x∈{−1,1}n.
∀j ∈{1,··· ,m}
Specializing the results of Section 3.2 to (QLCP), we obtain the following polynomial programming
relaxations:
(QLCPSS) min λ
s.t. λ−q(x)=?1 x?S
?1
x
?
+
m
?
?√n
j=1
(bj−aT
?
S ∈Sn+1
jx)dT
j
?√n
x
?
+
n
?
i=1
(1+xi)fi
T
?√n
x
?
+
n
?
fi,gi,dj∈Ln+1,
i=1
(1−xi)giT
x
+
n
?
i=1
ci(1−x2
;
i)
ci∈R,
+
(QLCPSS+) min λ
s.t. λ−q(x)=?1 x?S
?1
x
?
+
m
?
?√n
j=1
(bj−aT
?
jx)dT
j
?√n
αik(1+xi)(bk−aT
n
?
?
lx)+
i=1
+,fi,gi,dj∈Ln+1,
x
?
+
n
?
i=1
(1+xi)fi
T
?√n
x
?
+
n
?
n
?
n
?
m
?
αik,βik,γij,δij,ζij,ηkl∈Rn
i=1
(1−xi)giT
m
?
n
?
m
?
x
+
n
?
kx)+
i=1
m
?
k=1
kx)
+
i=1
k=1
βik(1−xi)(bk−aT
i=1
?
?
n
?
ζij(1+xi)(1−xj)
n
ci(1−x2
j=i
γij(1+xi)(1+xj)
+
i=1j=i
δij(1−xi)(1−xj)+
n
i=1
n
j=1
+
k=1l=k
ηkl(bk−aT
kx)(bl−aT
i)
ci∈R,S ∈Sn+1
+
;
Page 12
12
(QLCPSOC) min λ
s.t. λ−q(x)=
m
?
+
j=1
?
h+
(bj−aT
jx)dT
j
?√n
?
fi,gi,dj∈Ln+1.
x
?
+
n
?
i=1
(1+xi)fi
T
?√n
x
?
+
n
?
i=1
(1−xi)giT
?√n
x
?
n
i=1
ci(1−x2
i)+
i<j
h+
ij(1+xixj)+
n
?
i<j
h−
ij(1−xixj)
ci∈R,
ij,h−
ij∈R+
4.2.1.The Relaxation of Burer and Lov´ aszSchrijver
Burer [6] presented an SDPbased relaxation for the QLCP where the variables are 01. We
introduce the following relaxation that is at least as strong as the relaxation presented by Burer
[38]:
(QLCPBurer’) min λ
s.t. λ−q(x)=?1 x s t?(M +N)
?
ci∈R,
1
x
s
t
+
n
?
i=1
cixi(1−xi)+
n
?
i=1
(1−xi−si)li(x)
+
m
j=1
li,ki∈R1[x,s,t],
(bj−aT
jx−tj)kj(x)
M ∈S2n+m+1
+
,N ∈R2n+m+1
+
,
where m is the number of linear constraints. Further (QLCPBurer’) is equivalent to:
min λ
s.t. λ−q(x)=?1 x 1−x b−aTx?(M +N)
ci∈R,
1
x
1−x
b−aTx
,
+
n
?
i=1
cixi(1−xi)
M ∈S2n+m+1
+
,N ∈R2n+m+1
+
Page 13
13
which can be written as
min λ
s.t. λ−q(x)=?1 x?M?
?1
αikxi(bk−aT
x
?
+
n
?
i=1
cixi(1−xi)
+
n
?
n
?
n
?
αik,βik,γij,δij,ζij,ηkl∈R+,
i=1
m
?
n
?
n
?
k=1
kx)+
n
?
δij(1−xi)(1−xj)
m
?
i=1
m
?
k=1
βik(1−xi)(bk−aT
kx)
+
i=1j=i
γijxixj+
n
?
i=1
n
?
j=i
+
i=1
j=1
ζijxi(1−xj)+
k=1
m
?
l=k
ηkl(bk−aT
M?∈Sn+1
kx)(bl−aT
lx)
ci∈R,
+
.
Notice that (QLCPBurer’) reduces to the N+relaxation of Lov´ asz and Schrijver [22] by setting the
variables γij,δij,ζij, and ηklto zero. That is, N+is equivalent to the following relaxation:
(QLCPN+) min λ
s.t. λ−q(x)=?1 x?S
?1
βik(1−xi)(bk−aT
αik,βik∈R+,
x
?
+
n
?
i=1
cixi(1−xi)+
n
?
i=1
m
?
k=1
αikxi(bk−aT
kx)
+
n
?
i=1
m
?
k=1
kx)
ci∈R,S ∈Sn+1
+
.
4.2.2.Comparing Relaxations for QLCP
We can prove the following result:
Theorem 5. Let λ∗
QLCPN+, λ∗
QLCPBurer’, and λ∗
QLCPSS+
be the optimal solution value of
(QLCPN+),(QLCPBurer’), and (QLCPSS+) respectively, then
λ∗
QLCPN+≥λ∗
QLCPBurer’≥λ∗
QLCPSS+≥z∗
QLCP.
Proof: Define
H4=Ψ2+
?
?
i,k
Ψ0(1+xi)(bk−aT
Ψ0(1+xi)(1+xj)+
?
?
kx)+
?
Ψ0(1−xi)(1−xj)+
i,k
Ψ0(1−xi)(bk−aT
kx)+
?
?
Ψ0(1+xi)(1−xj)
i
(1−x2
i)R0
H5=H4+
i≤j
?
i≤j
i,j
+Ψ0
k≤l
(bj−aT
(bk−aT
kx)(bl−aT
lx)
?
H6=H5+
j
jx)P1(B)+
i
(1+xi)P1(B)+
?
i
(1−xi)P1(B).
Hence,
H4⊆H5⊆H6.
Page 14
14
After a simple change of variables from {−1,1} to {0,1}, H4 and H5 correspond to the repre
sentations (QLCPN+) and (QLCPBurer’) respectively, while H6corresponds to the representation
(QLCPSS+).
Table 2 lists the number of variables required to formulate the various relaxations for the QLCP
?
problem of Theorem 5 where we have m linear constraints and n binary variables. While both
relaxations have the same computational complexity, (QLCPSS+) provides the best bounds as shown
in Theorem 5 and confirmed by the computational results of Section 5.
Table 2 Problem dimension for various QLCP relaxations.
Variables
SDP
SOC
Linear Nonnegative
Linear Free
(QLCPN+)
1, (n+1)×(n+1)
(QLCPBurer’)
1, (n+1)×(n+1)
(QLCPSS+)
1, (n+1)×(n+1)
(2n+m),(n+1)
2nm+n2+2?n+1

?
2nm 2nm+n2+2?n+1
2
?+?m+1
2
2
?+?m+1
2
?
nnn
Remark 1. We are unable to compare theoretically (QLCPSS) with (QLCPN+) and
(QLCPBurer’). However in our computational experiments in Section 5.2, (QLCPSS) always provides
a strictly better bound than (QLCPN+) while (QLCPBurer’) provides a strictly better bound than
(QLCPSS).
4.3. Quadratic Knapsack Problem
In this section, we consider the quadratic knapsack problem (QKP) which is the particular case
of QLCP where m=1. The QKP was introduced by Gallo, Hammer, and Simeone [9] and is NP
hard. The QKP can be interpreted as follows: we are given n items with a nonnegative weight
wiassigned to item i, and a (n+1)×(n+1) symmetric matrix Q with real entries. The QKP is
the problem of selecting a subset of items so as to maximize the overall profit such that the total
weight of the selected items does not exceed a given capacity c. Introducing the binary variable xi
such that
?
−1
the problem may be formulated as:
xi=
1 if item i is selected
otherwise,
(QKPP) max q(x)=?1 x?Q
s.t. wTx≤c
?1
x
?
x∈{−1,1}n.
The QKP is a generalization of the linear knapsack problem (where the objective function is
linear). As in the case of the linear knapsack problem, the QKP often appears as a subproblem to
Page 15
15
other complex problems such as the graph partitioning problem described in Johnson, Mehrotra,
and Nemhauser [12]. Since the QKP is a constrained version of the binary quadratic problem, all
valid inequalities for the unconstrained BQPP problem are also valid for the QKP and hence they
can be used to tighten bounds for this problem. Using the same approach as in Section 3.2, we
obtain the following relaxations of (QKPP):
(QKPSS) min λ
s.t. λ−q(x)=?1 x?S
?1
x
?
+(c−wTx)dT
?√n
?√n
ci(1−x2
;
x
?
+
n
?
i)
i=1
(1+xi)fi
T
?√n
x
?
+
n
?
fi,gi,d∈Ln+1,
i=1
(1−xi)giT
x
?
S ∈Sn+1
+
n
?
i=1
ci∈R,
+
(QKPSS+) min λ
s.t. λ−q(x)=?1 x?S
?1
x
?
+(c−wTx)dT
?√n
?√n
αi(1+xi)(c−wTx)+
?
x
?
+
n
?
i=1
(1+xi)fi
T
?√n
n
?
?
x
?
+
n
?
?
n
?
αi,βi,γij,δij,ζij∈R+,
i=1
(1−xi)giT
γij(1+xj)(1+xi)+
x
?
+
n
?
i=1i=1
βi(1−xi)(c−wTx)
ζij(1−xj)(1+xi)
+
i≤ji≤j
δij(1−xj)(1−xi)+
i,j
+
i=1
ci(1−x2
i)
ci∈R,fi,gi,d∈Ln+1,S ∈Sn+1
+
;
(QKPSOC) min λ
s.t. λ−q(x)=(c−wTx)dT
?√n
x
?
+
n
?
h+
i=1
(1+xi)fi
T
?√n
?
x
?
+
n
?
i=1
(1−xi)giT
?√n
x
?
+
n
?
h+
i=1
ci(1−x2
i)+
?
fi,gi,d∈Ln+1.
i<j
ij(1+xixj)+
i<j
h−
ij(1−xixj)
ci∈R,
ij,h−
ij∈R+
4.3.1.HelmbergRendlWeismantel QKP Relaxation
Helmberg et al. [11] presented four SDPbased relaxations for the QKP where the discrete set
is {0,1}n. These relaxations are obtained by considering the semidefinite matrix X = xxT. In
particular they studied the relaxation
Page 16
16
(QKPHRW4) max ?P,X?+cst
s.t.
?
X −diag(X)diag(X)T?0,
j
wjXij−¯ cXii≤01≤i≤n
where ¯ c =1
4Qii−4?
from {−1,1} to {0,1}. Helmberg et al. [11] showed that the optimal objective value of (QKPHRW4),
λ∗
2(?
iwi− c), P is an n × n matrix with entries Pij= 4Qij (for i ?= j) and Pii=
jQij+4Q0i, and cst=Q00−2?
iQ0i+?
i,jQijare obtained by mapping the variables
QKPHRW4, provides the best bound among the SDP relaxations they provided. Actually, (QKPHRW4)
provides the tightest known SDP relaxation for the QKP in the literature. We will be using this
relaxation for comparison purposes in our computational results. In addition, Helmberg et al.
[11] strengthen these proposed relaxations by using cutting planes that are valid for BQPP. To
illustrate the quality of these SDP relaxations and of the cutting planes, Helmberg et al. [11] present
computational results on instances with up to 61 items.
4.3.2. Comparing Relaxations for QKP
In this section, we compare (QKPHRW4) and our proposed relaxation. First we rederive (QKPHRW4)
in a different way by considering the problem
(QKPD) min λ
s.t. λ−p(x)∈P2({0,1}n∩{x:(¯ c−wTx)≥0}),
where p(x)=?
i,jPijxixj+cst. This problem can be relaxed using
P2({0,1}n∩{x:(¯ c−wTx)≥0})⊇Ψ2+
?
i
Ψ0xi(¯ c−wTx)+
?
i
xi(1−xi)R0,
obtaining
min λ
s.t. λ−p(x)=?1 x?S
, di∈ R+, and ci∈ R. By equating the coefficients of the monomials of the above
problem, we rewrite it as
?1
x
?
+
?
i
dixi(¯ c−wTx)+
?
i
cixi(1−xi),
where S ∈ Sn+1
+
(QKPHRW4D) min λ
s.t. λ−cst−S00=0
ci+¯ cdi+Si0+S0i=0
diwj+djwi
2
S ?0,
−Sij+ciδi=j=Pij
di≥0.
1≤i≤j ≤n
where δi=jequals 1 if i=j and 0 otherwise. Taking the dual of (QKPHRW4D), we obtain
Page 17
17
max
s.t.¯ X00=1
¯ Xii−¯ Xi0=0
n
?
¯ X ?0,
?¯P,¯ X?
(9)
1≤i≤n
(10)
j=1
wj¯ Xij−¯ c¯ Xii≤01≤i≤n
(11)
(12)
?
where¯P =
?cst 0
0 P
?
. Since X −diag(X)diag(X)T?0 is equivalent to¯ X =
0, the above problem is a reformulation of (QKPHRW4). Taking X = I, X is strictly feasible for
?
1 diag(X)T
X
diag(X)
?
(QKPHRW4), therefore Slater’s constraint qualification is satisfied for (QKPHRW4). In addition, X −
diag(X)diag(X)T? 0 implies −1
?
Theorem 6. Let λ∗
8≤ Xij≤ 1 [11]. As a result, the objective ?P,X? is bounded by
i,jPij and we have strong duality.
QKPHRW4Dand λ∗
QKPSS+be the optimal solution values of (QKPHRW4D) and
(QKPSS+) respectively, then
λ∗
QKPHRW4D=λ∗
QKPHRW4≥λ∗
QKPSS+≥z∗
QKP.
Proof: Define
H7=Ψ2+
H8=H7+
?
?
?
i
Ψ0(1+xi)(c−wTx)+
Ψ0(1−xi)(c−wTx)+
Ψ0(1+xi)(1−xj)+(c−wTx)P1(B)+
?
?
i
(1−x2
Ψ0(1+xi)(1+xj)+
i)R0
i
i≤j
?
i≤j
Ψ0(1−xi)(1−xj)
(1+xi)P1(B)+
+
i,j
?
i
?
i
(1−xi)P1(B).
Hence,
H7⊆H8.
After mapping the variables from {−1,1} to {0,1}, H7 corresponds to the approximation of
P2({0,1}n∩ {x : (¯ c − wTx) ≥ 0}) that is equivalent to (QKPHRW4D) and H8 corresponds to the
representation (QKPSS+).
?
Table 3 presents the number of variables for the relaxations (QKPHRW4D) and (QKPSS+). Notice
that both relaxations have the same computational complexity. However, the (QKPSS+) relaxation
provides the best bounds as shown in Theorem 6.
Remark 2. In some instances, even when using the weaker relaxation (QKPSS), we obtain a
strictly better bound than (QKPHRW4) as shown in Section 5.3. For those instances (QKPSS+) is
also strictly better than (QKPHRW4).
Page 18
18
Table 3
Variables
SDP
SOC
Linear Nonnegative
Linear Free
Problem dimension for various QKP relaxations.
(QKPHRW4D)
1, (n+1)×(n+1) 1, (n+1)×(n+1)
(QKPSS+)
 (2n+1),(n+1)
2n+n2+2?n+1
n
n
2
?
n
5.Computational Results
In this section, we present computational results obtained by implementing the proposed relax
ations of Section 3.2 to the three classes of BQPP problems considered in Section 4. We conduct
comparisons based on computational time and on the quality of the bounds. The focus is on veri
fying the efficiency of the proposed SOC relaxations compared to the SOS/SDPbased relaxations.
All relaxations were implemented with MATLAB 7.9.0 for constructing the problems and SeDuMi
solver version 1.3 [36] was used to solve the conic problems. The experiments were done on a 1200
MHz Sun Sparc machine and the reported computational time is in cpu seconds.
5.1. General BQPPs Computational Results
In this section, we compare our proposed relaxations with the approach adopted by Lasserre [17]
to solve general binary quadratic problems. We compare the following four relaxations:
(BQPPLas): the relaxation presented in Section 4.1;
(BQPPSS+): the relaxation presented in Section 3.2;
(BQPPSS): the SOCSDPbased relaxation presented in Section 3.2;
(BQPPSOC): the SOC relaxation presented in Section 3.2.
We consider 100 randomly generated instances that vary in size, n, from 10 items up to 70 and
density from 20% to 100%. In addition the number of linear constraints, m, varies from 1 to
n
2
and the number of quadratic constraints, m, also varies from 1 to
n
2. We implemented Lasserre’s
relaxation using our code. In Table 4, we report the average gap and the average computational
time of all four relaxations (the average is computed over 5 instances for each value of n and m).
The gap (in %) is calculated as follows:
gap=100×ubrelaxation−ubbest
ubbest
%,
where the best upper bound is the one obtained by the (BQPPSS+) relaxation.
From Table 4 and Figure 1, we notice that (BQPPSOC) is the most computationally efficient
relaxation in most cases. When the number of linear constraints has a value of
n
2, then (BQPPLas)
is slightly more efficient but for those cases the bounds provided by (BQPPLas) are weaker than
Page 19
19
Table 4 Computational results for the BQPP instances. The avg. gaps are with respect to (BQPPSS+).
nm (BQPPSS+)(BQPPSS)
TimeGapTime
101 2.03 0.851.98
5 1.992.15 1.71
201 5.42 0.24 5.36
5 10.602.64 6.37
10 14.964.266.66
30
1 22.33 1.0916.39
5 35.95 2.8419.17 23.88
15 73.29 10.7422.94 32.92
401 78.181.6656.30
5 122.372.33 67.54
20306.31 5.71 88.80
50
1 268.93 0.68179.74
5 397.34 3.44193.86 17.71
25 1245.49 12.27258.77
601 970.003.15 626.87 19.61
51169.37 3.69 663.09
305637.189.42 850.83
701 2793.310.93 2515.31
53848.182.50 2532.18
3515420.53 14.85 2429.09
Avg. 4.27
(BQPPLas)
Gap
12.50
32.40
7.96
20.37
72.30
2.36
(BQPPSOC)
Gap
2.40
2.63
1.78
Time
1.49
1.27
4.18
5.56 16.59
5.36
10.32
14.30
17.02
34.59 29.97
44.66 28.18
48.27 38.60
112.49
122.32
142.29 43.08
375.24
397.93 39.75
473.50 52.10
Time
1.47
1.07
1.81
2.12
2.42
7.69
9.11
11.19
33.07
37.47
44.11
48.72
117.75
190.33
94.16
183.34
650.46
165.63
549.22
8.08
28.11
26.21
53.21
34.89
36.38
50.60
5.12
15.16
39.05
94.54
65.83
40.75
58.95
29.44 1214.23
53.64 1245.09 26.98
47.51 1446.99 46.99 1818.69
34.69
31.51
30.06
Figure 1Computational time for BQPP (logarithmic scale).
those provided by (BQPPSOC).
The bound of (BQPPSS+) is the strongest among the four relaxations. Therefore, we compare the
average gaps of (BQPPSOC), (BQPPLas), and (BQPPSS) relative to (BQPPSS+). Further, (BQPPSS)
provides better gaps than (BQPPSOC) and (BQPPLas) for all instances and for the latter two
relaxations we indicate the gap with lower value in bold. Notice that from Table 4, (BQPPSOC)
Page 20
20
frequently has better gaps than (BQPPPLas).
5.2.QLCP Computational Results
In this section, we compare our proposed relaxations of QLCP with the approach proposed by
Burer [6] to solve binary quadratic polynomial problems with linear constraints. Table 5 reports
the average gap (in %) between each relaxation’s upper bound and the optimal objective value
(known a priori), as well as the average computational time. We compare five relaxations:
(QLCPBurer’): the relaxation presented in Section 4.2;
(QLCPN+): the Lov´ aszSchrijver relaxation presented in Section 4.2;
(QLCPSS+): the strengthened SDP relaxation presented in Section 4.2;
(QLCPSS): the SOCSDP relaxation presented in Section 4.2;
(QLCPSOC): the SOC relaxation presented in Section 4.2.
We consider 732 instances that vary in size from 10 up to 50 items, and with density varying
from 1% to 100%. The number of the linear constraints varies from 1 to 25. The data for the
instances and their optimal objective values, as well as the upper bounds and computational time
of Burer’s specialized implementation, labeled as Time1 in Table 5, were all provided by Burer [6].
We also implemented Burer’s relaxation using our code (as described in Section 4.2) and we report
the average computational time we obtained for it as Time2 in Table 5.
Table 5
n
Computational results for the QLCP instances.
m
(QLCPSS+)
GapTime
1 7.761.54
5 11.282.82 11.30
1 3.72 6.90
5 8.4414.23
10 10.62 20.10 10.64
1 1.7420.52
5 5.75 47.14
15 10.0976.75 10.14
1 1.2672.38
5 2.77150.43
209.94297.52 10.01
1 1.07 222.98
5 2.64 495.88
259.57 1163.82
 6.19
(QLCPBurer’)
Gap Time1 Time2
7.77 0.66
0.72
3.751.24
8.48 2.01
2.22
1.80 2.17
5.793.35
4.76
1.313.76
2.81 5.30 128.28
10.21 245.19 16.03
1.114.96 200.60
2.71 8.00 447.68
9.7718.13 865.09 16.78 199.60 30.64 365.35 34.76 158.80
6.249.16
(QLCPSS)
Gap
9.00
(QLCPN+)
Gap
(QLCPSOC)
Gap
9.63
Time
1.44 11.27
1.77 21.59
5.19
6.58 17.58
7.11 21.86
13.66
17.96 13.49
20.52 22.66
41.09
54.23
66.26 26.89 120.24 33.11
1.31 134.422.07 104.82
4.00 161.796.69 204.87 17.50 109.17
Time
1.11
1.63 18.73
4.40
7.95 17.86
11.63 23.19
12.00
22.45 19.43
34.63 29.21
34.29
68.33 14.19
Time
0.94
1.15
1.80
2.31
2.45
8.50
10.09
11.84
29.63
37.28
42.54
32.99
101.17
2.18 16.60
5.95
12.56 12.09
17.53 15.70
17.64
40.39
64.28 15.41
63.50
20 4.94 7.89 7.84
30 2.32
8.40
3.40 5.94
401.84
3.77
2.91
5.91
8.86
504.79
Avg. 13.92  17.50
From Table 5, we see that Burer’s relaxation is the most efficient in terms of computational time
but this is due to the fact that Burer’s algorithm is specialized for solving problems of this form.
Page 21
21
Figure 2Computational time for QLCP (logarithmic scale).
However, in theory, it is an SDPbased relaxation and thus the computational time has a higher
order of complexity than the SOCbased relaxation, (QLCPSOC). This can be seen when comparing
Time2 with the computational time of the (QLCPSOC) relaxation where the latter is on average 4
times more efficient for large n (see Figure 2). Among the four SDPbased relaxations, (QLCPSS)
is the most computationally efficient as seen from Figure 2.
As shown in Theorem 5 and Table 2, (QLCPSS+) provides the strongest bounds for the QLCP
relaxation and has the same computational complexity as (QLCPBurer’). On the other hand, both
(QLCPN+) and (QLCPSS) are semidefinitebased relaxations but with less computational com
plexity than (QLCPSS+) and (QLCPBurer’). We notice that (QLCPSS) provides better bounds than
(QLCPN+) for all instances and is more computationally efficient. The average percentage gap for
(QLCPSS) is 9.16% while that of (QLCPN+) is 13.92%. In addition,(QLCPSOC) provides compara
ble bounds with (QLCPN+) with an average percentage gap of 17.50% but is computationally the
most efficient.
5.3. QKP Computational Results
In this section, we compare the performance of our proposed relaxations for the QKP with the
relaxation of Helmberg et al. [11] presented in Section 4.3.1. We generated test instances using the
approach proposed in [32]. The Pijand wjvalues are discrete taken from a uniform random distri
bution in [1,100] and [1,50] respectively. The capacity ¯ c is uniformly distributed in [50,?n
The density d of the P matrix varies from 10 to 90 %.
The presented computational results are based on the following four types of relaxations for the
j=1wj].
quadratic knapsack problem:
(QKPHRW4): the Helmberg et al. SDP relaxation presented in Section 4.3.1;
Page 22
22
(QKPSS+): the relaxation presented in Section 4.3;
(QKPSS): the SOCSDP relaxation presented in Section 4.3;
(QKPSOC): the SOC relaxation presented in Section 4.3.
Table 6 reports results for 45 instances. These instances vary in size and density. The size varies
from 20 to 100 items and the density varies from 10 to 90% with a step size of 20%. For each
instance, we report the upper bound and the solution time in seconds.
In terms of computational time, (QKPSOC) is the most computationally efficient for all instances.
For example, for the largest instances (n = 100) the (QKPSOC) relaxation is on average 23 times
faster than the (QKPSS+), 19 times faster than the (QKPSS) relaxation, and 10 times faster than
the (QKPHRW4) relaxation (see Figure 3).
Further for all the tested instances, the (QKPSS+) and (QKPSS) bounds are strictly tighter than
the ones provided by (QKPHRW4), even though the bounds for the (QKPHRW4) relaxation are known
to be strong [11; 32]. In addition, we report the gap between the bounds of (QKPSS+), (QKPHRW4),
and (QKPSOC) and the bound of (QKPSS+). Over all instances, the percentage gap of the (QKPSOC)
relaxation with respect to the (QKPHRW4) relaxation ranges from 8% to around 31% with an
average of 4.39%, where a negative sign implies that the (QKPSOC) relaxation is better. Notice
that (QKPSOC) performs particularly well for instances with high density. In particular, (QKPSOC)
obtains better bounds than (QKPHRW4) for all the instances with d=90%.
Figure 3 Computational time for QKP (logarithmic scale).
Page 23
23
Table 6 Computational results for the QKP instances. The gaps are with respect to (QKPSS+).
nd (QKPSS+)
UB
(QKPSS)
UB Gap
(QKPHRW4)
UB Gap
(QKPSOC)
UBTime Time TimeGapTime
20
20
20
20
20
30
30
30
30
30
40
40
40
40
40
50
50
50
50
50
60
60
60
60
60
70
70
70
70
70
80
80
80
80
80
90
90
90
90
90
100
100
100
100
100
Avg.
10
30
50
70
90
10
30
50
70
90
10
30
50
70
90
10
30
50
70
90
10
30
50
70
90
10
30
50
70
90
10
30
50
70
90
10
30
50
70
90
10
30
50
70
90
809.00
2617.50
1120.90
2340.94
6082.09
1011.34
3451.65
8116.24
8042.65
5114.00
3845.33
11807.67
4298.30
17415.63
25599.30
2316.83
11414.34
23823.61
32567.32
17658.96
7173.33
26403.91
13853.47
56556.58
62009.00
3961.79
20191.73
45493.48
1621.19
32850.56
13062.74
1480.00
23126.43
58613.63
112167.40
6189.28
30656.56
81336.10
8004.38
55262.87
23941.78
40216.48
11707.00
122205.33
63378.00
8.66
4.01
7.52
5.53
5.61
20.83
24.15
17.14
15.01
15.96
51.43
40.09
93.95
76.24
64.70
274.29
186.91
270.40
133.12
167.46
705.30
552.28
726.82
663.95
357.10
2969.84
2698.05
2760.52
2900.58
1777.27
4407.65
3327.65
6694.43
5422.86
4178.58
15610.86
16455.66
10319.41
12082.24
11603.07
23975.63
31499.01
27958.75
20428.73
12182.11
811.22
2619.34
1137.25
2356.25
6083.70
1022.20
3470.97
8125.16
8047.03
5127.57
3853.49
11809.44
4309.76
17424.10
25612.48
2353.89
11433.16
23846.12
32571.10
17671.03
7188.68
26496.51
13871.42
56561.20
62009.00
4036.66
20208.57
45507.07
1631.57
32857.31
13074.13
1480.00
23141.40
58621.35
112184.20
6311.21
30710.62
81344.17
8014.95
55285.71
23951.45
40257.87
11737.62
122215.14
63378.00
0.27
0.07
1.46
0.65
0.03
1.07
0.56
0.11
0.05
0.27
0.21
0.02
0.27
0.05
0.05
1.60
0.16
0.09
0.01
0.07
0.21
0.35
0.13
0.01
0.00
1.89
0.08
0.03
0.64
0.02
0.09
0.00
0.06
0.01
0.01
1.97
0.18
0.01
0.13
0.04
0.04
0.10
0.26
0.01
0.00
0.30
6.19
5.68
6.09
5.51
5.72
18.24
16.37
19.48
18.20
15.78
55.43
59.27
69.20
60.41
59.17
158.24
188.84
181.09
213.29
168.55
673.37
644.20
682.12
797.11
478.40
1689.45
2262.87
2407.57
2308.23
2574.93
5008.22
4388.67
4650.24
5419.69
5052.10
7057.15
9398.88
12623.02
11942.53
8883.04
18831.11
17673.77
18867.58
24684.33
14881.31
814.84
2623.98
1175.07
2397.20
6086.12
1044.39
3511.30
8142.11
8073.14
5150.78
3864.51
11828.42
4365.56
17446.14
25630.04
2412.48
11485.59
23863.04
32626.49
17682.63
7215.96
26530.82
13895.51
56583.48
62015.61
4109.61
20275.13
45573.21
1882.75
32913.98
13118.78
1537.29
23220.33
58649.30
112202.99
6447.72
30829.68
81393.43
8312.97
55305.54
23977.05
40370.14
11879.03
122305.61
63411.61
0.72
0.25
4.83
2.40
0.07
3.27
1.73
0.32
0.38
0.72
0.50
0.18
1.56
0.18
0.12
4.13
0.62
0.17
0.18
0.13
0.59
0.48
0.30
0.05
0.01
3.73
0.41
0.18
16.13
0.19
0.43
3.87
0.41
0.06
0.03
4.18
0.56
0.07
3.86
0.08
0.15
0.38
1.47
0.08
0.05
1.34
4.10
3.10
4.14
4.14
3.85
9.31
9.77
12.25
10.78
9.35
38.72
32.40
34.06
35.92
39.15
96.62
114.12
116.33
113.25
91.98
394.70
312.36
355.53
343.42
391.59
954.02
1237.78
1224.95
1081.38
1157.09
2584.26
2143.41
2494.70
2979.25
2958.44
4818.87
5587.52
6233.23
4292.94
5640.32
9883.59
9832.83
8553.04
9482.50
10280.16
811.74
2619.48
1262.98
2540.40
6083.80
1129.01
3939.00
8127.76
8108.38
5136.34
3875.12
11811.54
5161.31
17447.01
25615.00
2846.05
12050.94
23850.99
32575.12
17672.78
7410.08
26502.66
14396.64
56561.20
62009.00
5104.22
21826.79
45752.77
1737.92
32876.13
13506.53
1532.02
25240.44
59322.02
112184.53
8500.89
36535.46
81385.48
8297.26
55291.14
24021.78
45597.97
13937.02
122476.61
63378.00
0.34
0.08
12.68
8.52
0.03
11.63
14.12
0.14
0.82
0.44
0.77
0.03
20.08
0.18
0.06
22.84
5.58
0.11
0.02
0.08
3.30
0.37
3.92
0.01
0.00
28.84
8.10
0.57
7.20
0.08
3.40
3.51
9.14
1.21
0.02
37.35
19.18
0.06
3.66
0.05
0.33
13.38
19.05
0.22
0.00
5.81
4.54
1.97
1.42
1.69
1.75
6.91
6.01
9.83
6.94
8.81
32.86
34.71
26.29
31.18
36.12
44.32
64.22
27.62
26.29
23.05
138.25
79.46
78.18
58.54
38.21
230.87
296.70
154.61
143.12
102.06
564.75
264.94
365.01
270.39
185.92
517.90
740.54
426.81
458.36
295.34
1867.44
973.49
1308.24
431.02
484.25

6. Conclusion and Future Work
In this research we used polynomial programming approaches to produce tractable relaxations for
general binary quadratic polynomial optimization problems. These approximations utilize linear,
secondorder and semidefinite cones over which it is known how to optimize efficiently. We proposed
Page 24
24
a secondorder cone relaxation for the general BQPP and applied it to several binary quadratic
polynomial problems. When compared to SDPbased relaxations, these SOCbased relaxations are
significantly more computationally efficient with only a small degradation of bounds.
For the general BQPP, we proposed two SOCSDPbased relaxations and compared them theo
retically and experimentally with Lasserre’s relaxation. By exploiting the linear constraints using
secondorder cones we are able to obtain a stronger relaxation than Lasserre’s. We also conducted
computational results on binary quadratic linear constrained problems and showed that the qual
ity of the bounds provided by our SOCSDPbased relaxation is competitive with those from the
very recent specialized relaxation of Burer for this problem [6]. Finally, for the quadratic knapsack
problem we showed that the two proposed SOCSDPbased relaxations are a strict improvement
on the best relaxation in the literature. Theoretical results as well as computational experiments
show that our SOCSDPbased relaxation outperforms the relaxation of Helmberg et al. [11] in
terms of bound while both relaxations are comparable in terms of computational time. We also
relaxed our proposed relaxation to obtain a weaker SOConly relaxation that is computationally
more efficient while still providing comparable bounds to [11] and for problems with high density
it provides better bounds.
The main objective of our research is to develop an exact algorithm for solving general binary
quadratic polynomial problems. Our SOC relaxations show strong potential, both in terms of
bounds and of computational time, to be used in an exact algorithm scheme to find optimal
solutions for large instances of such problems in a reasonable time. Future research will investigate
the use of SOCbased relaxations with additional valid inequalities. In particular, we are developing
nonlinear cuts based on polynomial programming to further strengthen the proposed relaxations.
Acknowledgments
The authors thank Sam Burer for providing us with the QLCP instances and their optimal values.
References
[1]B. Alidaee, G. Kochenberger, and A. Ahmadian. 01 quadratic programming approach for the optimal
solution of two scheduling problems. International Journal of Systems Science, 25:401–408, 1994.
[2]M. F. Anjos, A. Kennings, and A. Vannelli. A semidefinite optimization approach for the singlerow
layout problem with unequal dimensions. Discrete Optimization, 2(2):113–122, 2005.
[3]F. Barahona, M. J¨ unger, and G. Reinelt. Experiments in quadratic 01 programming. Mathematical
Programming, 44:127–137, 1989.
[4]E. Boros and P. Hammer. Pseudoboolean optimization. Discrete Applied Mathematics, 123(13):155–225,
2002.
Page 25
25
[5]E. Boros and A. Prekopa. Probabilistic bounds and algorithms for the maximum satisfiability problem.
Annals of Operations Research, 21:109–126, 1989.
[6]S. Burer. Private communication.
[7]S. Burer, R. Monteiro, and Y. Zhang. Ranktwo relaxation heuristics for maxcut and other binary
quadratic programs. SIAM Journal of Optimization, 12(2):503–521, 2002.
[8]E. de Klerk and D. Pasechnik. Approximation of the stability number of a graph via copositive program
ming. SIAM Journal on Optimization, 12(4):875–892, 2002.
[9]G. Gallo, P. Hammer, and B. Simeone. Quadratic knapsack problems. European Journal of Operational
Research, 12:132–149, 1980.
[10]M. Gr¨ otschel, M. J¨ unger, and G. Reinelt. An application of combinatorial optimization to statistical
physics and circuit layout design. Operations Research, 36(3):493–513, 1988.
[11]C. Helmberg, F. Rendl, and R. Weismantel. A semidefinite programming approach to the quadratic
knapsack problem. Combinatorial Oprimization, 4(2):197–215, 2000.
[12]E. L. Johnson, A. Mehrotra, and G. L. Nemhauser. Mincut clustering. Mathematical Programming,
62:133–151, 1993.
[13]S. Kim and M. Kojima. Secondorder cone programming relaxation of nonconvex quadratic optimization
problems. Optimization Methods and Software, 15(34):201–224, 2001.
[14]S. Kim and M. Kojima. Exact solutions of some nonconvex quadratic optimization problems via SDP
and SOCP relaxations. Computational Optimization and Applications, 26(2):143–154, 2003.
[15]S. Kim, M. Kojima, and M. Yamashita. Second order cone programming relaxation of a positive semidef
inite constraint. Optimization Methods and Software, 18(5):535–541, 2003.
[16]M. Kojima, S. Kim, and H. Waki. A general framework for convex relaxation of polynomial optimization
problems over cones. Journal of Operations Research Society of Japan, 46:125–144, 2003.
[17]J. Lasserre. An explicit equivalent positive semidefinite program for nonlinear 01 programs. SIAM
Journal on Optimization, 12(3):756–769, 2001.
[18]J. Lasserre. Global optimization problems with polynomials and the problem of moments. SIAM Journal
on Optimization, 11:796–817, 2001.
[19]D. Laughhunn. Quadratic binary programming with application to capitalbudgeting problems. Opera
tions Research, 18:454–461, 1970.
[20]M. Laurent. Semidefinite representations for finite varieties. Mathematical Programming, 109(Ser. A):1–
26, 2007.
[21]E. Lawler. The quadratic assignment problem. Management Science, 9:586–599, 1963.
[22]L. Lov´ asz and A. Schrijver. Cones of matrices and setfunctions and 01 optimization. SIAM Journal on
Optimization, 1:166–190, 1991.
Page 26
26
[23]T. S. Motzkin. The arithmeticgeometric inequality. In Inequalities (Proceedings of Symposium Wright
Patterson Air Force Base, Ohio, 1965), pages 205–224. Academic Press, New York, 1967.
[24]Y. Nesterov. Semidefinite relaxation and nonconvex quadratic optimization. Optimization Methods and
Software, 12:1–20, 1997.
[25]P. Pardalos and G. Rodgers. Computational aspects of a branch and bound algorithm for quadratic
zeroone programming. Computing, 45:131–144, 1990.
[26]P. Parrilo. Structured semidefinite programs and semialgebraic geometry methods in robustness and opti
mization. PhD thesis, Department of Control and Dynamical Systems, California Institute of Technology,
Pasadena, CA., 2000.
[27]P. Parrilo. An explicit construction of distinguished representations of polynomials nonnegative over
finite sets. Technical report, IfA Technical Report AUT0202, Zurich  Switzerland, 2002.
[28]P. Parrilo. Semidefinite programming relaxations for semialgebraic problems. Mathematical Programming,
96(2):293–320, 2003.
[29]P. Parrilo and B. Sturmfels. Minimizing polynomial functions, algorithmic and quantitative real algebraic
geometry. DIMACS Series in Discrete Mathematics and Theoretical Computer Science, 60:83–89, 2003.
[30]J. F. Pe˜ na, J. C. Vera, and L. F. Zuluaga. Exploiting equalities in polynomial programming. Operations
Research Letters, 36(2), 2008.
[31]A. Phillips and J. Rosen. A quadratic assignment formulation of the molecular conformation problem.
Journal of Global Optimization, 4:229–241, 1994.
[32]D. Pisinger. The quadratic knapsack problema survey. Discrete Applied Mathematics, 155(5):623–648,
2007.
[33]I. P´ olik and T. Terlaky. A survey of the Slemma. SIAM Review, 49:371–418, 2007.
[34]M. Putinar.Positive polynomials on compact semialgebraic sets.
Journal, 42:969–984, 1993.
Indiana University Mathematics
[35]N. Shor. A class of global minimum bounds of polynomial functions. Cybernetics, 23(6):731–734, 1987.
[36]J. Sturm. Using SeDuMi 1.02, a MATLAB toolbox for optimization over symmetric cones. Optimization
Methods and Software, 1112, 1999.
[37]B. Sturmfels, J. Demmel, and J. Nie. Minimizing polynomials via sum of squares over the gradient ideal.
Mathematical Programming, 106:587–606, 2006.
[38]J. C. Vera, L. Zuluaga, and J. Pe˜ na. Positive polynomials on equality constrained domains. Working
Paper.
[39]H. Wolkowicz, R. Saigal, and L. Vandenberghe. Handbook of Semidefinite programming Theory, Algo
rithms, and Applications. Kluwer, 2000.
[40]L. Zuluaga, J. C. Vera, and J. Pe˜ na. LMI approximations for cones of positive semidefinite forms. SIAM
Journal on Optimization, 16(4), 2006.
Download fulltext