On Integrality in Semidefinite Programming for Discrete

Optimization
∗ †
Frank de Meijer Renata Sotirov

Abstract
arXiv:2306.09865v2 [math.OC] 8 Nov 2023

It is well-known that by adding integrality constraints to the semidefinite programming (SDP)

relaxation of the max-cut problem, the resulting integer semidefinite program is an exact formulation
of the problem. In this paper we show similar results for a wide variety of discrete optimization
problems for which SDP relaxations have been derived. Based on a comprehensive study on discrete
positive semidefinite matrices, we introduce a generic approach to derive mixed-integer semidefinite
programming (MISDP) formulations of binary quadratically constrained quadratic programs and bi-
nary quadratic matrix programs. Applying a problem-specific approach, we derive more compact
MISDP formulations of several problems, such as the quadratic assignment problem, the graph par-
tition problem and the integer matrix completion problem. We also show that several structured
problems allow for novel compact MISDP formulations through the notion of association schemes.
Complementary to the recent advances on algorithmic aspects related to MISDP, this work opens
new perspectives on solution approaches for the here considered problems.
Keywords: mixed-integer semidefinite programming, discrete positive semidefinite matrices, binary
quadratic programming, quadratic matrix programming, association schemes

1 Introduction
Semidefinite programming (SDP) deals with the optimization of a linear function over the cone of pos-
itive semidefinite matrices under the presence of affine constraints. Over the last decades, semidefinite
programs (SDPs) have proven themselves particularly useful in providing tight relaxations of discrete
optimization problems. Following the extension from linear programming to integer linear programming
initiated in the 1960s, a recent interest in incorporating integer variables in SDPs has arisen. Indeed,
many real-world decision problems are most naturally modeled by including integer variables in optimiza-
tion problems. When the variables in an SDP are required to be integer, we refer to the problem as an
integer semidefinite program (ISDP). When an SDP contains both integer and continuous variables, we
refer to the program as a mixed-integer semidefinite program (MISDP). As mixed-integer linear programs
(MILPs) form a subclass of MISDPs, mixed-integer semidefinite programming is in general N P-hard.
The combination of positive semidefiniteness and integrality induces a lot of structure in matrices.
Exploiting this fact, it has been shown that several structured discrete optimization problems allow for
a formulation as a (M)ISDP. To the best of our knowledge, the first ISDP formulation of a discrete
optimization problem is derived for the symmetric traveling salesman problem by Cvetković et al. [18].
Eisenblätter [24] derives an ISDP formulation of the minimum k-partition problem, which asks for a
partition of the vertex set of a graph into at most k sets such that the total weight of the edges within
the same set is minimized. Anjos and Wolkowicz [3] show that the standard SDP relaxation of the max-
cut problem becomes exact when adding integrality constraints. As an immediate consequence, also the
SDP relaxation of the max-2-sat problem, i.e., the maximum satisfiability problem where each clause has
at most two literals, see e.g., [31], can be modeled as an ISDP. An ISDP formulation of the chromatic
number of a graph is derived by Meurdesoif [54]. The quadratic traveling salesman problem (QTSP) is
formulated as an ISDP in [53]. Next to these classical textbook problems, integrality in SDPs has also
∗ Delft Institute of Applied Mathematics, Delft University of Technology, The Netherlands, [email protected]
† CentER, Department of Econometrics and OR, Tilburg University, The Netherlands, [email protected]

1
been at consideration in more applied problems. Yonekura and Kanno [65] formulate an optimization
problem in robust truss topology design as a MISDP, see also [15, 43, 51]. The problem of computing
restricted isometry constants also allows for a MISDP formulation, see [26]. A MISDP formulation of the
regularized cardinality constrained least squares problem is derived in [59]. Gil-González et al. [29] use
a MISDP to formulate an optimal location problem in power system analysis. Zheng et al. [66] model a
robust version of a power system unit commitment problem using a mixture of semidefinite constraints
and integer variables. Duarte [22] exploits MISDPs to find exact optimal designs of experiments in
the domain of surface response modeling in statistics. Finally, Wei et al. [64] show that mixed-integer
quadratic programs with indicator variables can be solved as a MISDP.
Despite the literature on these particular problems, a generic approach for deriving problem for-
mulations based on mixed-integer semidefinite programming has not been followed. Although there do
exist several approaches in the literature where SDP relaxations are used in a branching scheme, see
e.g., [44, 61], the branching strategies are based on the problem structure rather than on the matrix
variables being integer. Accordingly, exploiting integrality in the MISDP models itself has not been the
method of choice so far. This might be due to the fact that solving SDPs of large sizes is still practically
challenging, discouraging to look into the extension of adding hard integrality constraints to the model.
In this paper we refute these objections to consider MISDPs as a general solution technique, advocating
that they have a great potential to be also numerically advantageous. We particularly focus on binary
quadratic programs (BQPs), which aim to optimize a quadratic objective function g(x) over a feasible
set X defined by quadratic or linear constraints, where x is required to be binary, see Figure 1. A common
approach to solve these programs is by exploiting standard linearization techniques to model them as a
MILP. This is often done in a branch-and-bound setting, where the subproblems correspond to the linear
programming relaxations of the MILP. This research line is depicted in the top stream of Figure 1. An
alternative approach is to lift the vector variable x in a BQP to a matrix variable X = xx⊤ so as to
model the problem as an SDP with a nonconvex rank constraint. After relaxing this rank constraint,
we obtain an SDP relaxation of the problem, see e.g., [2]. This relaxation approach corresponds to the
bottom arrow in Figure 1. Apart from the particular problems mentioned earlier, it is disregarded up
to this point that this relaxation can also be obtained via relaxing integrality in a MISDP model that
is equivalent to the BQP. More precisely, there exists a bijection between the elements in X and the
integer points in the feasible set of the SDP relaxation. Realizing that fact, this provides a systematic
way of approaching BQPs via mixed-integer semidefinite programming. Comparing the three equivalent
formulations given in Figure 1, the MISDP formulation has the advantages to have both a linear objective
function (compared to the BQP formulation) and a convex relaxation that is often stronger than standard
linear programming relaxations. After reformulating the mixed-integer nonlinear program as a convex
mixed-integer nonlinear program possessing a tight relaxation, all solution techniques from convex mixed-
integer nonlinear programming can be applied to tackle the problem. With the advancing state of the
solution approaches in this field, the perspectives of this solution approach are hopeful.
There exist various related works on the representation of optimization problems as convex mixed-
integer nonlinear programs in the literature. Lubin et al. [50] focus on characterizations of general
mixed-integer convex representability. The special case of mixed-integer second-order cone programming
has received some more attention, see the survey by Benson and Sağlam [8]. To the best of our knowledge,
we are the first to address the case of mixed-integer semidefinite representability on a generic level.
The focus of this paper is primarily on the modeling aspect of discrete optimization problems as
(M)ISDPs, and less on the algorithmic aspects of solving these. With respect to the computational side,
several general-purpose solution approaches have been considered recently. Gally et al. [27] propose a
branch-and-bound framework for solving MISDPs, with the characteristic that strict duality is maintained
throughout the branching tree. Solver ingredients, such as dual fixing and branching rules, are also
considered in [27]. Kobayashi and Takano [41] propose a cutting-plane and a branch-and-cut algorithm
for solving generic MISDPs, where it is shown that the branch-and-cut algorithm performs best. This
branch-and-cut algorithm is strengthened in [53], where specialized cuts, such as Chvátal-Gomory cuts,
are incorporated in the approach. Presolving techniques for MISDPs have been studied by Matter
and Pfetsch [52]. Hojny and Pfetsch [38] consider reduction techniques for solving MISDPs based on
permutation symmetries. Another project that supports solving MISDPs while exploiting sparsity is
GravitySDP [36]. The computational ingredients of the above-mentioned approaches combined with

2
 Mixed-Integer Convex Mixed-
Nonlinear Integer Nonlinear
Programming Programming
Mixed-Integer
Linear
Programming Linear
Programming
min f (x, y) Relaxation
s.t. (x, y) ∈ XM ILP
min f (x, y)
s.t. (x, y) ∈ X M ILP
f (x, y) linear function
Binary XM ILP mixed-integer
polyhedral set
Quadratic
Programming

min g(x)
s.t. x ∈ X

g(x) quadratic function

Mixed-Integer
X integer quadratically
constrained set
Semidefinite
Programming Semidefinite
Programming
min h(x, X) Relaxation
s.t. (x, X) ∈ XM ISDP
min h(x, X)
h(x, X) linear function s.t. (x, X) ∈ X M ISDP
XM ISDP mixed-integer
spectrahedral
set

lift x to X and relax rank-constraint

Figure 1: Overview of various exact formulations of binary quadratic program and their relaxations. A
double arrow (⇐⇒) denotes equivalence between the formulations, while a solid arrow (→) denotes that
the formulation is relaxed from the former to the latter. The sets X , XMILP and XMISDP are defined
by nonconvex integer constraints, while X MILP and X MISDP are convex relaxations.

the theoretical framework of modeling problems as (M)ISDPs that we derive in this paper, provide a
complementary foundation of mixed-integer semidefinite programming in discrete optimization.

Main results and outline

This paper studies the theoretical role of mixed-integer semidefinite programming in discrete optimization.
We show that many problems can be modeled as a (M)ISDP, either by a generic approach for certain large
problem classes, or by a more problem-specific approach. Gradually, we cover and exploit results from
matrix theory, optimization, algebraic combinatorics and graph theory. Our approach is accompanied
with a large number of examples of various discrete optimization problems.
We start our approach with an extensive overview of results on the matrix theory of discrete positive
semidefinite (PSD) matrices. Without considering explicit optimization problems, we focus on the struc-
ture of PSD {0, 1}–, {±1}– and {0, ±1}–matrices. This overview reviews results from [9, 23, 48, 49], but
also introduces new results and formulations of these matrix sets, such as a combinatorial viewpoint of
PSD {0, 1}–matrices of rank at most r. We also extend results that are known for {0, 1}–matrices to the
other two matrix sets.
These matrix theoretical results are exploited when proving that many binary quadratic problems
allow for a formulation as a binary semidefinite program (BSDP). We establish this result for binary
quadratically constrained quadratic programs and, in particular, for binary quadratic matrix programs.
Problems that allow for a formulation as a binary quadratic matrix program, e.g., quadratic clustering
or packing problems, can be modeled as a compact BSDP with a PSD matrix variable of relatively low
order.
After that, we treat several specific problem classes for which we obtain MISDP formulations that
do not follow from the above-mentioned framework or for which it is possible to obtain a more efficient
formulation. Among these are the quadratic assignment problem, including its extensive number of
special cases, see e.g., [14], several graph partition problems and graph problems that can be modeled
based on association schemes. Also, as most formulations that we discuss include binary variables, we
present two problems that have a MISDP formulation where the variables are integer, but nonbinary,

3
namely the integer matrix completion problem [1] and the sparse integer least squares problem [58].
This paper is structured as follows. In Section 2 we present results on the matrix theory of discrete
PSD matrices. These results are exploited in Section 3, where MISDP formulations of generic quadrati-
cally constrained quadratic programs and quadratic matrix programs are derived. In Section 4 we treat
problem-specific MISDP formulations that do not follow from the previous section. Concluding remarks
are given in Section 5.

Notation
We denote by 0n ∈ Rn the vector of all zeros, and by 1n ∈ Rn the vector of all ones. The identity matrix
and the matrix of ones of order n are denoted by In and Jn , respectively. We omit the subscripts of these
matrices when there is no confusion about the order. The ith unit vector is denoted by ei and we define
Eij := ei e⊤
j . The set of n × n permutation matrices is denoted by Πn .
For n ∈ Z+ , we define the set [n] := {1, . . . , n}. For any S ⊆ [n], we let 1S be the binary indicator
vector of S. The support of x ∈ Rn is denoted by supp(x).
We denote the set of all n × n real symmetric matrices by S n . The cone of symmetric positive
n
semidefinite matrices is defined as S+ := {X ∈ S n : X P 0}, where X  0 denotes that X is PSD. The
n×m
trace of a square matrix X = (Xij ) is givenP by tr(X) = i Xii . For any X, Y ∈ R the trace inner
⊤ n Pm
product is defined as hX, Y i := tr(X Y ) = i=1 j=1 Xij Yij .
The operator diag : Rn×n → Rn maps a square matrix to a vector consisting of its diagonal ele-
ments. We denote by Diag : Rn → Rn×n its adjoint operator. The Hadamard product of two matri-
ces X = (Xij ), Y = (Yij ) ∈ Rn×m is denoted by X ◦ Y and is defined as (X ◦ Y )ij := Xij Yij . The direct
sum of matrices X and Y is defined as X ⊕ Y = ( X 0
0 Y ). The Kronecker product X ⊗ Y of matrices
p×q r×s
X ∈R and Y ∈ R is defined as the pr × qs matrix composed of pq blocks of size r × s, with block
ij given by xij Y , i ∈ [p], j ∈ [q].

2 Theory on discrete PSD matrices

Most discrete optimization problems that we consider in this paper are defined using binary variables, i.e.,
variables taking values in {0, 1} or {±1}, or ternary variables, i.e., variables whose values are in {0, ±1}. In
this section we derive several useful results on these matrix sets with respect to positive semidefiniteness.
We start by considering the PSD {0, 1}–matrices, after which we extend these results to PSD {±1}–
and {0, ±1}–matrices.

2.1 Theory on PSD {0, 1}–matrices

In this section we consider the set of positive semidefinite {0, 1}–matrices. We derive and recall several
formulations of this matrix set, including a combinatorial, polyhedral and a set-completely positive de-
scription. We start this section with two known results, i.e., Theorem 1 and Proposition 1, after which
we present a series of new results.
Positive semidefinite {0, 1}–matrices are studied explicitly by Letchford and Sørensen [48]. They
derive the following decomposition result on PSD {0, 1}–matrices.
Pr
Theorem 1 ([48]). Let X ∈ {0, 1}n×n be a symmetric matrix. Then X  0 if and only if X = j=1 xj x⊤ j
for some xj ∈ {0, 1}n, j ∈ [r].
Given S ⊆ R+ , a matrix X is called S-completely positive if X = P P ⊤ for some P ∈ S n×k . In
case S = R+ , we call X completely positive. It follows from Theorem 1 that any PSD {0, 1}–matrix
is {0, 1}–completely positive, see also Berman and Xu [9].
The decomposition of PSD {0, 1}–matrices gives rise to a useful combinatorial interpretation on the
complete graph Kn . Viewing each vector xj ∈ {0, 1}n as an indicator vector on the vertices of Kn , the
matrix xj x⊤ j can be seen as the characteristic matrix of a clique in Kn . Given a decomposition X =
Pk ⊤
j=1 xj xj , the cliques indexed by j ∈ [k] are pairwise disjoint, since the diagonal of X is at most one.

4
Therefore, each PSD {0, 1}–matrix is the characteristic matrix of a set of pairwise disjoint cliques in Kn .
This combinatorial structure is in the literature also known as a clique packing.
As we will see in the next section, many SDP formulations arise from a lifting P P ⊤ , where P is an
appropriate n×r {0, 1}–matrix. Consequently, the resulting PSD {0, 1}–matrix has rank at most r. From
that perspective, it makes sense to consider the set of PSD {0, 1}–matrices that have an upper bound on
the rank. For positive integers r, n with r ≤ n, let us define the discrete set

Drn := X ∈ {0, 1}n×n : X  0, rank(X) ≤ r . (1)

Theorem 1 induces the following {0, 1}–completely positive description of Drn :


Drn = P P ⊤ : P ∈ {0, 1}n×r , P 1r ≤ 1n . (2)

Next, we will derive another formulation of Drn , where the constraint rank(X) ≤ r is established by an
appropriate linear matrix inequality. To that end, we exploit the following result that is implicitly proved
in many sources and explicitly by Dukanovic and Rendl [23].
Proposition 1 ([23]). Let X ∈ {0, 1}n×n be symmetric. Then, the following are equivalent:
(i) diag(X) = 1n , rank(X) = r, X  0.
(ii) There exists a permutation matrix Q such that QXQ⊤ = Jn1 ⊕ · · · ⊕ Jnr with n = n1 + · · · + nr .
(iii) diag(X) = 1n , rank(X) = r and X satisfies the triangle inequalities Xij + Xik − Xjk ≤ 1 for all
(i, j, k) ∈ [n] × [n] × [n].
(iv) diag(X) = 1n and (tX − J  0 ⇐⇒ t ≥ r).
Proposition 1 establishes the equivalence between useful characterizations of rank-r PSD {0, 1}–
matrices that have ones on the diagonal. In the following corollary we generalize this result by relaxing
the condition diag(X) = 1n . The proof is similar to the proof of Proposition 1.
Corollary 1. Let X ∈ {0, 1}n×n be symmetric. Then, the following statements are equivalent:
(i) rank(X) = r, X  0.
(ii) There exists a permutation matrix Q such that QXQ⊤ = Jn1 ⊕ · · · ⊕ Jnr ⊕ 0nz ×nz with n =
n1 + · · · + nr + nz .
(iii) rank(X) = r and X satisfies the triangle inequalities Xij ≤ Xii for all i 6= j and Xij + Xik − Xjk ≤
Xii for all j < k, i 6= j, k.
(iv) tX − diag(X)diag(X)⊤  0 if and only if t ≥ r.
Proof. Throughout the proof, let N1 := {i ∈ [n] : Xii = 1} and let N0 := [n]\N1 . Moreover, let Q′ denote
a permutation matrix corresponding to a permutation of [n] that maps the ordered set (1, 2, . . . , n) to an
ordered set where the elements in N1 occupy the first N1 positions.
(i) ⇐⇒ (ii) : Let X be positive semidefinite with rank(X) = r. Then, the rows and columns indexed
by N0 only contain zeros. As a consequence, Q′ X(Q′ )⊤ is of the form Y ⊕ 0|N0 |×|N0 | with diag(Y ) = 1|N1 |
and Y  0. By Proposition 1 there exists a permutation matrix Q̄ such that Q̄Y Q̄⊤ = Jn1 ⊕ · · · ⊕ Jnr
with |N1 | = n1 + · · · + nr . Let Q := (Q̄ ⊕ I|N0 | )Q′ , then QXQ⊤ = Jn1 ⊕ · · · ⊕ Jnr ⊕ 0|N0 |×|N0 | .
Conversely, suppose that QXQ⊤ = Jn1 ⊕ · · · ⊕ Jnr ⊕ 0nz ×nz with n = n1 + · · · + nr + nz for some
permutation matrix Q. Then, obviously, QXQ⊤ is positive semidefinite with rank(QXQ⊤ ) = r, from
which it follows that X  0 with rank(X) = r.
(i) ⇐⇒ (iii) : For n = 2, the inequalities Xij ≤ Xii , i 6= j, are trivially necessary and sufficient for
X  0. For n ≥ 3, the result follows from Letchford and Sørensen [48, Proposition 3].
(i) ⇐⇒ (iv) : Suppose X is PSD with rank(X) = r. Then, Q′ X(Q′ )⊤ = Y ⊕ 0|N0 |×|N0 | with diag(Y ) =
1|N1 | and Y  0. This leads to the following sequence of equivalences:

tX − diag(X)diag(X)⊤  0 ⇐⇒ t Q′ X(Q′ )⊤ − Q′ diag(X)diag(X)⊤ (Q′ )⊤  0

5
   ⊤

1|N1 | 1|N1 |
⇐⇒ t Y ⊕ 0|N0 |×|N0 | − 0
0|N0 | 0|N0 |

which holds if and only if t ≥ r, by statement (iv) of Proposition 1.

Conversely, suppose that tX − diag(X)diag(X)⊤  0 if and only if t ≥ r. If r = 0, then t = 0 induces
diag(X) = 0n , while t = 1 implies X − diag(X)diag(X)⊤ = X  0, since diag(X) = 0n . Hence, X must
be the zero matrix, which is positive semidefinite with rank zero. Now, assume that r ≥ 1. Then, rX −
diag(X)diag(X)⊤  0 can be written as X  r1 diag(X)diag(X)⊤  0. Let r∗ := rank(X). It follows
from the previously proven implication, (i) =⇒ (iv), that r∗ = min{t : tX − diag(X)diag(X)⊤  0}.
By assumption, this value equals r, so r∗ = r. We conclude that X is PSD with rank(X) = r.
Corollary 1 can be exploited to prove the following result.
 
r diag(X)⊤
Corollary 2. Let X ∈ {0, 1}n×n be symmetric. If Y = diag(X) X
 0, then X  0 with
rank(X) ≤ r.
Proof. The assertion X  0 is trivial, so it suffices to show that Y  0 implies rank(X) ≤ r. If r = 0,
then diag(X) = 0n . Since X  0, X must be the zero matrix and, thus, rank(X) = 0.
Now,
 let r ≥ 1. The Schur complement lemma implies that rX − diag(X)diag(X)⊤  0. Let r∗ :=
min t : tX − diag(X)diag(X)⊤  0 ≤ r. Since r∗ X − diag(X)diag(X)⊤  0 and X  0, it follows
that tX − diag(X)diag(X)⊤  0 for all t ≥ r∗ . Therefore, tX − diag(X)diag(X)⊤  0 if and only if
t ≥ r∗ . Corollary 1 then implies rank(X) = r∗ ≤ r.
Corollary 2 implies the following characterization of Drn , where the rank constraint is merged into a
lifted linear matrix inequality:
   
n n×n r diag(X)⊤
Dr = X ∈ {0, 1} : 0 . (3)
diag(X) X
For some optimization problems, the upper bound constraint on the rank of X is not sufficient, as we
require that X is exactly of rank r. The max k-cut problem, for instance, requires to partition the vertex
set of a graph into exactly k nonempty and pairwise disjoint subsets. The following two results show that
the description given in (3) can be extended to also include a lower bound on the rank of X.

 X ∈
Theorem 2. Let  {0, 1}
n×n
be symmetric. If there exists a matrix P ∈ {0, 1}n×r with P ⊤ 1 ≥ 1
⊤
such that Y = IPr PX  0, then X  0 with rank(X) ≥ r.

Proof. The assertion X  0 is trivial. It suffices to show that rank(X) ≥ r. As Y  0 and Y has binary
P u
u
⊤
entries, it follows from Theorem 1 that Y = kj=1 xjj xjj for some uj ∈ {0, 1}r and xj ∈ {0, 1}n, j ∈
Pk ⊤
[k]. Since j=1 uj uj = Ir , we must have k ≥ r. Moreover, the set {uj : j ∈ [k]} must contain e1 , . . . , er
and k − r copies of 0r . Without loss of generality, let us assume that the first r vectors in {uj : j ∈ [k]}
Pk Pr
correspond to the elementary vectors. Then, it follows that P = j=1 xj u⊤ j =
⊤
j=1 xj ej = [x1 . . . xr ].
⊤
Since P 1 ≥ 1, it follows that the vectors xj , j P
∈ [r], cannot be the zero vector. Since these are moreover
linearly independent, we have rank(X) ≥ rank( rj=1 xj x⊤ j ) = r.

Theorem 2 and Corollary 2 together impose the following integer semidefinite characterization of
PSD {0, 1}–matrices of rank r.
Corollary 3. Let X ∈ {0, 1}n×n
 be symmetric.
 If there exists a matrix P ∈ {0, 1}n×r with P ⊤ 1 ≥ 1,
⊤
P 1 = diag(X), such that Y = IPr PX  0, then X  0 with rank(X) = r.

Proof. It immediately follows from Theorem 2 that X  0 with rank(X) ≥ r. Moreover, since Y  0,
we also know that
 ⊤   


⊤ 1r Ir 1r 1⊤ ⊤
r P In r diag(X)⊤
1⊤r ⊕ In Y 1 ⊤
r ⊕ In = =  0.
In P 1r In XIn diag(X) X
It then follows from Corollary 2 that rank(X) ≤ r.

6
 The integer semidefinite characterization of Drn given in (3) shows that if a {0, 1}–matrix satisfies a
certain linear matrix inequality, then a rank condition is implied. For the case of rank-one matrices, we
can show that the converse implication does also hold, i.e., if a rank-one matrix satisfies a certain linear
matrix inequality, then its entries must be in {0, 1}.
⊤


Theorem 3. Let Y = x1 xX  0 with diag(X) = x. Then, rank(Y ) = 1 if and only if X ∈ {0, 1}n×n.

Proof. (=⇒) : If rank(Y ) = 1, then Y = x̄x̄⊤ with x̄ = [1 x⊤ ]⊤ ∈ Rn+1 and X = xx⊤ . From the positive
semidefiniteness of order two principal submatrices of Y we obtain 0n ≤ x ≤ 1n . Since diag(xx⊤ ) = x,
we have x2i = xi for all i ∈ [n], so x ∈ {0, 1}n. We conclude that X = xx⊤ ∈ {0, 1}n×n.
(⇐=) : Since X ∈ {0, 1}n×n and x = diag(X), it follows that Y ∈ {0, 1}(n+1)×(n+1). From Theorem 1
Pk
it follows that Y = j=1 xj x⊤ j for some xj ∈ {0, 1}
n+1
, j ∈ [k], i.e., Y can be decomposed in terms of
cliques. Since Y11 = 1 and diag(Y ) = (1, x⊤ )⊤ , all indices i ∈ [n + 1] for which Yii = 1 must be in the
same clique as the first index. Hence, the decomposition consists of only one clique and rank(Y ) = 1.
Theorem 3 plays a central role in deriving integer SDP formulations of binary quadratic problems
defined over vectors of variables in Section 3.1. However, Theorem 3 cannot be extended to matrices
with a rank larger than one. That is, if Y is a PSD matrix satisfying diag(Y ) = Y e1 and Y11 = 2, then
integrality of Y is not equivalent to be of rank 2. For example, the matrix Y = 12 (J3 + 3E11 ) satisfies
Y  0, diag(Y ) = Y e1 and rank(Y ) = 2, but Y is not integer.
The characterizations given in (2) and (3) rely on conditions involving discreteness. Let us now move
on to continuous descriptions. Of course, since Drn is itself a discrete set, a continuous description does
not aim at describing Drn , but rather its convex hull, i.e.,

Prn := conv(Drn ). (4)

Observe that although the matrices in Drn have an upper bound on the rank, the polytopes Prn are full-
dimensional, since n1 In ∈ Prn for all 1 ≤ r ≤ n. In order to gain more insight into the structure of Prn ,
we introduce the notion of a so-called packing family.
Definition 1. Let T be a finite set of elements. A collection F of nonempty subsets of T is called a
packing of T if the subsets in F are pairwise disjoint. The family of all packings of T is called the packing
family of T , denoted by F(T ).

Observe that F = ∅ also belongs to F(T ). Next, we define the notion of an r-packing of T .
Definition 2. Let T be a finite set of elements. A packing F of T is called an r-packing of T if |F | ≤ r.
The family of all r-packings of T is called the r-packing family of T , denoted by Fr (T ).
The r-packing family of [n] can be exploited to describe Prn . Let X ∈ Drn . By Theorem 1 we know
that X is the sum of at most r rank-one PSD {0, 1}–matrices. From a combinatorial point of view, this
implies that X corresponds to an r-packing of [n]. In fact, there is aPbijection between the matrices in
Drn and the r-packings in Fr ([n]). For any r-packing F , let EF := S∈F 1S 1⊤ S . Then, we obtain the
following polyhedral description of Prn for all positive integers r ≤ n:
 
 X X 
Prn = X ∈ S n : X = λF EF , λF = 1, λF ≥ 0 for all F ∈ Fr ([n]) . (5)
 
F ∈Fr ([n]) F ∈Fr ([n])

We call the description above the packing description of Prn . Let us now consider the cardinality of the
vertices of Prn . S
In the vein of Definition 2, we call F ⊆ P([n]) an r-partition of [n] if it is an r-packing with S∈F = [n].
Here, P([n]) denotes the power set of [n]. The number of partitions of the set [n] into  k nonempty subsets
is in the literature known as the Stirling number of the second kind, denoted by nk . The total number
Pn
of partitions of [n] equals the Bell number Bn [7], for which we have Bn = k=0 nk . We can now show
the following result regarding the cardinality of Drn .

7
 Pr+1 n+1
Theorem 4. For n ≥ 1 and 0 ≤ r ≤ n, we have |Drn | = k=1 k . In particular, |D1n | = 2n
and |Dnn | = Bn+1 .
Proof. It follows from the discussion above that |Drn | equals the number of r-packings in Fr ([n]). In order
to count these, we count the number of packings that consist of exactly k subsets, while k ranges from 0
to r. Any packing of [n] into k subsets corresponds to a partition of [n + 1] into k + 1 subsets. To see this,
observe that to each packing F of [n] into k subsets one can add a (k+1)th set containing the element n+1
and the elements not covered by F . Conversely, given a partition of [n+1] into k +1 subsets, dropping the
set containing the element n + 1 yields a packing of [n] consisting
 of exactly k subsets. Hence, the number
Pr n+1 Pr+1 n+1
of packings of [n] consisting of exactly k subsets equals n+1k+1 and |D n
r | = k=0 k+1 = k=1 k .
n
n+1 n+1 2n+1 −2 n
For the special case r = 1, we obtain |D1 | = 1 + 2 = 1 + 2 = 2 . When r = n, we exploit
n+1 n
Pn+1 n+1 Pn+1 n+1
0 = 0 to conclude that |Dn | = k=1 k = k=0 k = Bn+1 .
The polytope Prn has several relationships with other well-known polytopes from the literature.
Letchford and Sørensen [48] study the polytope Pnn , albeit in a different embedding, and refer to it
as the binary PSD polytope of order n. They emphasize its relationship with the clique partition-
ing polytope that was introduced by Grötschel and Wakabayashi [33] and later studied in [5, 56].
Given the complete graph G = (V, E), a clique partition is a subset A ⊆ E such that there is a par-
V into nonempty disjoint sets V1 , . . . , Vk such that each Vj , j ∈ [k], induces a clique in G
tition of S
and A = j∈[k] {{i, ℓ} : i, ℓ ∈ Vj , i 6= ℓ}. The incidence vectors of clique partitions are only defined on
the edge set, and therefore the clique partition polytope can be seen as a projection of Pnn .
Among one of the first graph partition problems is the one considered by Chopra and Rao [17]. Given
an undirected graph G, the vertices need to be partitioned into at most k subsets so as to minimize
the total cost of edges cut by the partition. If G is the complete graph, the partition polytope P 1 (r)
considered in [17] coincides with Prn (apart from the embedding).
The polytope Pnn can also be related to the stable set polytope. Let GP = (VP , EP ) be the power
set graph, i.e., each vertex in VP corresponds to a nonempty subset of [n] and the edge set is defined
(2)
as EP := {{S, T } ∈ VP : S ∩ T 6= ∅}. A set of vertices is stable in GP if and only if its corresponding
collection of subsets is a packing of [n]. Hence, the packing family Fn ([n]) is the collection of all stable
sets in GP . It follows that there is a bijection between the elements in Pnn and the stable set polytope of
GP .
Finally, for r = 1, the r-packings of [n] are subsets of [n], so the polytope P1n simplifies to
 
 X X 
Rn1 := X ∈ S n : X = θS 1S 1⊤
S, θS = 1, θS ≥ 0 for all S ⊆ [n] . (6)
 
S⊆[n] S⊆[n]

The polytope Rn1 relates to the convex hull of the characteristic vectors of all cliques in Kn , i.e., the
clique polytope of Kn . This polytope is in the literature also known as the complete set packing polytope,
see [12]. Finally, apart from the embedding, the polytope Rn1 also coincides with the Boolean quadric
polytope [57].
Another continuous formulation of the convex hull of PSD {0, 1}–matrices is  given by a conic

de-
scription. The cone of completely positive matrices is defined as CP n := conv xx⊤ : x ∈ Rn+ . An
extension of the completely positive matrices are the so-called set-completely positive matrices, see e.g.,
[49, 10], where the membership condition x ∈ Rn+ is replaced by x ∈ K for a general convex cone K.
Lieder et al. [49] considered the following set-completely positive matrix cone:


SCP n := conv xx⊤ : x ∈ Rn+ , x1 ≥ xi for all i ∈ {2, . . . , n} . (7)

Since the membership condition given in (7) is more restricted than x ∈ Rn+ , we have SCP n ( CP n . Let
us now consider the following set-completely positive matrix set:
   
1 x⊤
C1n := X ∈ S n : ∈ SCP n+1 , diag(X) = x . (8)
x X

The following result follows from Lieder et al. [49].

8
Theorem 5 ([49]). We have P1n = C1n .
A natural question is whether the descriptions Rn1 and C1n for P1n given in (6) and (8), respectively,
can be extended to higher ranks. The extensions of these sets are as follows:
 
 X X X 
Rnr := X ∈ S n : X = θS 1S 1⊤
S, θS = r, θS ≤ 1 ∀i ∈ [n], θS ≥ 0 ∀S ⊆ [n] , (9)
 
S⊆[n] S⊆[n] S:i∈S
   
r diag(X)⊤
Crn := X ∈ S n : ∈ SCP n+1 , diag(X) ≤ 1n . (10)
diag(X) X

The extension from C1n to Crn follows from the intersection of the Minkowski sum of r copies of C1n with the
upper bound constraint X ≤ Jn . Since Xii ≥ Xij for all i, j ∈ [n] if X ∈ C1n , itPsuffices to add diag(X) ≤
1n . The extension from Rn1 to Rnr is derived as follows. If X ∈ Prn , then X = F ∈Fr ([n]) λF EF for some
nonnegative weights λF . By splitting each r-packing into its separate subsets, we obtain
X X X X X X
X= λF EF = λF 1S 1⊤
S = λF 1S 1⊤
S = θS 1S 1⊤
S,
F ∈Fr ([n]) F ∈Fr ([n]) S∈F S⊆[n] F ∈Fr ([n]): S⊆[n]
S∈F
P P P P
where θS := F ∈Fr ([n]):S∈F λF . Moreover, S⊆[n] θS = F ∈Fr ([n]) λF |F | ≤ r F ∈Fr ([n]) λF = r. By
P P
increasing θ∅ , we obtain S⊆[n] θS = r. Finally, since Xii ≤ 1 for i ∈ [n], we have S:i∈S θS ≤ 1. We
conclude that Prn ⊆ Rnr .
Unfortunately, for r ≥ 2, the sets Rnr and Crn do no longer exactly describe Prn . Namely, consider the
matrix X = 21 P P ⊤ where P = I3 + E23 + E32 + E31 + E41 . For this matrix one can verify that X ∈ R42
and X ∈ C24 , while X ∈/ P24 . For r ≥ 2, the following relationship between Prn , Crn , Rnr holds.
Theorem 6. We have Prn ⊆ Crn = Rnr , while for r = 1 the three sets are equal.
n n n n n
Proof. Since PP r
r = conv(Dr ), it suffices to consider membership of the elements in Dr in Cr . Let X ∈ Dr ,
⊤ n j ⊤
then X = j=1 xj xj for some xj ∈ {0, 1} , j ∈ [r]. Let Y := xj xj for all j ∈ [r]. We clearly have
 
1 diag(Y j )
diag(Y j ) Yj
∈ SCP n+1 for all j ∈ [r], from which it follows that

r 
X   
1 diag(Y j ) r diag(X)
= ∈ SCP n+1 .
diag(Y j ) Yj diag(X) X
j=1

Moreover, X ∈ {0, 1}n×n, so diag(X) ≤ 1n . We conclude that X ∈ Crn .

To prove Crn = Rnr , let X ∈ Crn . We define the matrix Y as
   
1 r diag(X)⊤ 1 diag( 1r X)⊤
Y := = . (11)
r diag(X) X diag( 1r X) 1
rX

From the fact that X ∈ Crn , it follows that Y ∈ SCP n+1 . Applying Theorem 5 to the matrix Y implies
that r1P
X is a convex combination of rankP ′
P θS ≥ 0 ′for all⊤ S ⊆ [n]
one binary PSD matrices, i.e., there exist
with S⊆[n] θS = 1, such that r X = S⊆[n] θS 1S 1S , or equivalently, X =
′ 1 ′ ⊤
S⊆[n] rθS 1S 1S . Now,
′
P
let θS := rθS for all S ⊆ [n], from which it follows that S⊆[n] θS = r. Since diag(X) ≤ 1n , it follows
P
that Xii = S:i∈S θS ≤ 1. We conclude that X ∈ Rnr .
Finally, observe that the argument above can also be followed in the converse direction. That is, given
X ∈ Rnr with corresponding weights θS for all S ⊆ [n], we define θS′ := r1 θS , S ⊆ [n], which implies that
1 n n+1
r X ∈ P1 . By Theorem 5, we know that Y , see (11), is contained in SCP , implying X ∈ Crn .

2.2 Theory on PSD {±1}–matrices

In this section we present several results for PSD matrices that have entries in {±1}. Let us first state
the following well-known result.

9
Proposition 2 ([3]). Let X be a symmetric matrix. Then, X  0, X ∈ {±1}n×n if and only if X = xx⊤
for some x ∈ {±1}n.
A simple necessary condition for X ∈ {±1}n×n to be PSD is that diag(X) = 1. The next result
establishes the equivalence between {0, 1}– and {±1}–PSD matrices by exploiting their rank.
Proposition 3. Let X ∈ {±1}n×n be a symmetric matrix and Y := 12 (X + J) ∈ {0, 1}n×n. Then, X  0
if and only if diag(Y ) = 1, Y  0 and rank(Y ) ≤ 2.
Proof. (=⇒): Let X  0. Since J  0, it follows that Y  0. Moreover, diag(X) = diag(J) = 1 implies
that diag(Y ) = 1. Finally, by Proposition 2 we know that X = xx⊤ for some x ∈ {±1}n. Therefore, Y
is the weighted sum of two rank one matrices, so rank(Y ) ≤ 2.
(⇐=): Let Y = 12 (X+J)  0, diag(Y ) = 1 and rank(Y ) ≤ 2 for some symmetric matrix X ∈ {±1}n×n.
Since Y is binary positive semidefinite with rank at most two, it follows from Theorem 1 that Y = x1 x⊤ ⊤
1 + x2 x2
n ⊤
for some x1 , x2 ∈ {0, 1} . Then, X = 2Y − J = (x1 − x2 )(x1 − x2 ) , which implies that X  0.
Note that the matrix Y from the previous theorem has rank one if and only if Y = X = J. Similar
to (1), we define the discrete set of all {±1}–matrices as

b n := X ∈ {±1}n×n : X  0 ,
D (12)

b n have rank one. Based on Proposition 2,

where the subscript r is not present anymore, as all matrices in D
b n
we can easily establish that |D | = 2 n−1
. Next, we summarize known results on sets related to {±1}–
matrices. The convex hull of all PSD {±1}–matrices is known as the cut polytope:

b n := conv(D
P b n ), (13)

see e.g., [46]. Also, we define the following set-completely positive matrix cone:


b n := conv
SC P xx⊤ : x ∈ Rn , x1 + xi ≥ 0, x1 − xi ≥ 0 for all i ∈ {2, . . . , n} . (14)

The cone SC Pbn is considered in [49], where the authors show that SC P b n and SCP n , see (7), are related
n
as follows: T (SCP ) = SC P b and T (SC P
n −1 b ) = SCP , where T is an appropriate linear mapping.
n n

Lieder et al. [49] consider the following set-completely positive matrix set:
   
1 x⊤
Cbn := X ∈ S n : ∈ SC Pb n+1 , diag(X) = 1n , (15)
x X

which is the analogue of the set C1n for {0, 1}–matrices, see (8).
b n = Cbn .
Theorem 7 ([49]). We have P
This theorem is the analogue of Theorem 5 that provides a result for {0, 1}–matrices. For the equiv-
alence transformation between {±1}– and {0, 1}–representations of SDP relaxations of binary quadratic
optimization problems, we refer the interested reader to Helmberg [34].

2.3 Theory on PSD {0, ±1}–matrices

In the sequel we generalize several results from the previous sections to PSD {0, ±1}–matrices. The follow-
ing result shows that a PSD {0, ±1}–matrix is block-diagonalizable, which is the analogue of Proposition 1
for {0, 1}–matrices.
Proposition 4. Let X ∈ {0, ±1}n×n be symmetric. Then, the following statements are equivalent:
(i) diag(X) = 1n , rank(X) = r, X  0.
(ii) There exists a permutation matrix Q such that QXQ⊤ = Bn1 ⊕ · · · ⊕ Bnr , where Bni = bi b⊤
i ,
bi ∈ {±1}ni for i ∈ [r], n = n1 + · · · + nr .

10
Proof. Suppose that QXQ⊤ is in the block form given in (ii), then it trivially satisfies the conditions
given in (i). Conversely, let X ∈ {0, ±1}n×n satisfy (i). Let us consider the ith row in X. Suppose j
and k are two distinct indices not equal to i in the support of this row, i.e., Xij , Xik 6= 0. For the sake
of contradiction, suppose that Xjk = 0. Then, the submatrix of X induced by i, j and k is either one of
the following matrices:

i j k i j k i j k i j k
       
i 1 1 1 i 1 −1 −1 i 1 1 −1 i 1 −1 1
       
j 1 1 0 , j  −1 1 0 , j 1 1 0 , or j  −1 1 0 .
k 1 0 1 k −1 0 1 k −1 0 1 k 1 0 1

One easily checks that the determinants of these matrices are all negative, contradicting that X  0.
Hence, Xjk 6= 0. This argument can be repeated to conclude that the submatrix of X indexed by the
support of row i has entries in {±1}. Since the submatrix of X is also positive semidefinite, it follows
from Proposition 2 that the submatrix is of the form bb⊤ with b ∈ {±1}ni for some positive integer ni .
By the same argument, it follows that the other indices in the submatrix induced by row i have the
same support as row i. Indeed, if this would not be the case, one of the four matrices above should be a
submatrix of X. We conclude that X can be fully constructed from nonoverlapping submatrices of the
form bb⊤ with b ∈ {±1}ni for some positive integer ni . Since its rank equals r, there must be r of those
submatrices. From here the claim follows.
Proposition 4 extends easily to the following result.
Corollary 4. Let X ∈ {0, ±1}n×n be symmetric. Then, the following statements are equivalent:
(i) rank(X) = r, X  0.
(ii) There exists a permutation matrix Q such that QXQ⊤ = Bn1 ⊕ · · · ⊕Bnr ⊕0nz ×nz where Bni = bi b⊤
i ,
bi ∈ {±1}ni for i ∈ [r], n = n1 + · · · + nr + nz .
Proof. The proof is similar to the proof of Corollary 1.
Let X ∈ {0, ±1}n×n be given as in Corollary 4, then
r
X
QXQ⊤ = Bn1 ⊕ · · · ⊕ Bnr ⊕ 0nz ×nz = b1 b⊤ ⊤
1 ⊕ · · · ⊕ br br ⊕ 0nz ×nz = x̄i x̄⊤
i ,
i=1

where
Pr x̄⊤ ⊤ ⊤ ⊤ ⊤ ⊤ ⊤
1 = [b1 0n−n1 ], x̄2 = [0n1 b2 0n−n1 −n2 ], and so on. Let xi := Qx̄i for i ∈ [r], then X =
⊤ n
i=1 xi xi , where xi ∈ {0, ±1} . This construction yields the following decomposition of PSD {0, ±1}–
matrices.
Pr ⊤
Theorem 8. Let X ∈ {0, ±1}n×n be symmetric. Then, X  0 if and only if X = j=1 xj xj for
n
some xj ∈ {0, ±1} , j ∈ [r].
The previous result is an extension of Theorem 1 to {0, ±1}–matrices. We now consider an equivalence
between a PSD {0, ±1}–matrix of rank one and an extended linear matrix inequality, i.e., the analogue
of Theorem 3.
 
1 x⊤
Proposition 5. Let Y = ∈ S n+1 with supp(diag(X)) = supp(x). Then, Y ∈ {0, ±1}(n+1)×(n+1),
x X
Y  0 if and only if X = xx⊤ .
Proof. Let Yij ∈ {0, ±1} for all i, j ∈ [n + 1] and Y  0. Then x ∈ {0, ±1}n. The Schur complement
lemma implies X − xx⊤  0. If Xii = 0 then xi = 0, and if Xii = 1 then xi = 1 or xi = −1. Thus
diag(X − xx⊤ ) = 0, from which it follows that X = xx⊤ . The converse direction is trivial.
Clearly, the condition supp(diag(X)) = supp(x) can be replaced by diag(X)ii = |xi | for all i ∈ [n],
where | · | denotes the absolute value.

11
3 Binary quadratic optimization problems
In this section we exploit the theoretical results on discrete PSD matrices from the previous section to
derive exact reformulations of binary quadratic programs (BQPs) as binary semidefinite programs. In
Section 3.1 we consider the general class of binary quadratically constrained quadratic programs. In
Section 3.2 we consider a subclass of these programs that allow for a formulation as a binary quadratic
matrix program.

3.1 Binary quadratically constrained quadratic programs

A quadratically constrained quadratic program (QCQP) is an optimization problem with a quadratic
objective function under the presence of quadratic constraints. Many discrete optimization problems can
be formulated as QCQPs.
Let Q0 , Qi ∈ S n , c0 , ci ∈ Rn for all i ∈ [m], and ai ∈ Rn , bi ∈ R for all i ∈ [p], where m, p ∈ N. We
consider binary programs of the following form:

min x⊤ Q0 x + c⊤
0x

s.t. x⊤ Qi x + c⊤
i x ≤ di ∀i ∈ [m]
(QCQP )
⊤
ai x = b i ∀i ∈ [p]
x ∈ {0, 1}n.

The quadratic terms in (QCQP ) can be written as hQi , Xi+ c⊤ ⊤

i x for all i, where we substitute X for xx .
This yields the following exact reformulation of (QCQP ):

min hQ0 , Xi + c⊤
0x

s.t. hQi , Xi + c⊤
i x ≤ di ∀i ∈ [m]
⊤
ai x = bi ∀i ∈ [p]
 
1 x⊤
Y =  0, diag(X) = x, rank(Y ) = 1.
x X

Here we used the conventional notion of exactness, i.e., the nonconvex constraint rank(Y ) = 1. We also
exploit here Theorem 3 in order to not explicitly require x to be binary. However, one can utilize an
alternative notion of exactness in terms of integrality, namely by exploiting Theorem 3. This leads to the
following binary semidefinite program (BSDP):

min hQ0 , Xi + c⊤
0x

s.t. hQi , Xi + c⊤
i x ≤ di ∀i ∈ [m]
⊤
ai x = bi ∀i ∈ [p] (BSDPQCQP )
 
1 x⊤
 0, diag(X) = x, x ∈ {0, 1}n.
x X

Observe that it is sufficient to impose integrality on the diagonal of X. Namely, it follows from the
determinants of the 3 × 3 principal submatrices of the matrix Y that Xij ∈ {0, 1} whenever Xii , Xjj ∈
{0, 1} for all i and j, see e.g., [35, Section 3.2]. Note that a binary matrix X that satisfies the linear
matrix inequality from (BSDPQCQP ) with x = diag(X), is an element of D1n , see (3). The next result
follows directly from the previous discussion and Theorem 3.
Theorem 9. (BSDPQCQP ) is equivalent to (QCQP ).
To provide a more compact BSDP formulation of (QCQP ), we prove the following result.
P
−b
⊤

Lemma 1. Let S = pi=1 −b
⊤
ai
i
ai
i
and Y = x1 xX  0, where diag(X) = x and X ∈ {0, 1}n×n.
Then, a⊤
i x = bi for all i ∈ [p] if and only if hS, Y i = 0.

12
 ⊤
Proof. It follows from Theorem 3 that Y = ( x1 )( x1 ) . If a⊤ i x = bi for all i ∈ [p], it is not difficult
Pp D −bi
−bi
⊤ 1 1 ⊤ E
to verify that hS, Y i = 0. Conversely, let hS, Y i = 0. Then, 0 = i=1 ai ai , ( x )( x ) =
Pp ⊤ 2 ⊤
i=1 (bi − ai x) , from which it follows that ai x = bi for all i ∈ [p].

Lemma 1 induces the following compact BSDP that is equivalent to (QCQP ):

min hQ0 , Xi + c⊤ x
s.t. hQi , Xi + c⊤
i x ≤ di ∀i ∈ [m]
p
*     +
X ⊤
−bi −bi 1 x⊤
, =0
ai ai x X
i=1
 
1 x⊤
 0, diag(X) = x, x ∈ {0, 1}n.
x X
There are various equivalent formulations of the binary quadratic program (QCQP ) in the literature.
We finalize this subsection by mentioning below only those that are closely related to our approach.
Assume that Qi = 0, ci = 0, and di = 0 for all i ∈ [m] in (QCQP ). Burer [13] proved that the
resulting optimization problem with a quadratic objective and linear constraints is equivalent to the
following completely positive program:
min hQ0 , Xi + c⊤ x
s.t. ai ⊤ x = bi ∀i ∈ [p], hai ai ⊤ , Xi = b2i ∀i ∈ [p]
 
1 x⊤
∈ CP n+1 , diag(X) = x,
x X
provided that the inequalities 0 ≤ xi ≤ 1 for i ∈ [n] are implied by the constraints of the original problem.
Here CP n+1 is the cone of completely positive matrices.
On the other hand, Lieder et al. [49] proved the following equivalent formulation of the BQP with
quadratic objective and linear constraints:
min hQ0 , Xi + c⊤ x
s.t. ai ⊤ x = bi ∀i ∈ [p]
 
1 x⊤
∈ SCP n+1 , diag(X) = x,
x X

where the cone SCP n+1 is defined in (7). The authors of [49] also proved that, under mild assumptions,
the binary quadratic problem (QCQP ) with also quadratic constraints can be equivalently reformulated
as an optimization problem over the set-completely positive matrix cone SCP n+1 .
We end this section by providing an example of a problem that can be modeled as (BSDPQCQP ).
Example 1 (The stable set problem). Let G = (V, E) be a simple graph on n vertices. A stable set in
G is a subset S ⊆ V such that no two vertices in S are adjacent in G. The stable set problem (SSP) asks
for the largest size of a stable set in G. To model this problem, let x ∈ {0, 1}n be such that xi = 1 if
i ∈ S and xi = 0 otherwise. Then, x is the characteristic vector of a stable set in G if x⊤ (Eij + E⊤
ij )x = 0
for all {i, j} ∈ E. The cardinality of the stable set equals x⊤ x, hence the SSP is of the form (QCQP ).
Applying Theorem 9, the following BSDP models the SSP:
α(G) := max hIn , Xi
s.t. Xij = 0 ∀{i, j} ∈ E
  (16)
1 x⊤
 0, diag(X) = x, x ∈ {0, 1}n .
x X
The doubly nonnegative relaxation of (16) obtained after replacing X ∈ {0, 1}n×n by 0 ≤ X ≤ J, is
well-studied in the literature, see e.g., [32]. It is equivalent to a strengthened version of the Lovász theta
number, known as the Schrijver ϑ′ -number [62].

13
3.2 Binary quadratic matrix programs
A quadratic matrix program (QMP) [6] is a programming formulation where the objective and constraints
are given by

tr(P ⊤ Qi P ) + 2tr(Bi⊤ P ) + di (17)

for some Qi ∈ S n , Bi ∈ Rn×k and ci ∈ R, where P is an n × k matrix variable. QMPs are a special case
of QCQPs and are particularly useful to model optimization problems where the matrix P has entries
in {0, 1} and represents a classification of n objects over k classes, i.e., Pij = 1 if and only if object i is
assigned to class j. For example, if each object needs to be assigned in exactly (resp. at most) one class,
we call P a partition (resp. packing) matrix.
In this section we consider two different binary QMPs of increasing generality and show how these
can be reformulated as BSDPs. For both QMPs, we consider some well-known problems that fit into the
framework.
Our first QMP incorporates a specific objective and constraint structure, while optimizing over the
packing or partition matrices. Let Q0 , Qi ∈ S n , di ∈ R for all i ∈ [m], ai ∈ Rn and bi ∈ R+ for all i ∈ [p].
We consider the binary quadratic matrix program

min tr(P ⊤ Q0 P )
s.t. tr(P ⊤ Qi P ) + di ≤ 0 ∀i ∈ [m], P ⊤ ai ≤ bi 1k ∀i ∈ [p] (QM P1 )
n×k
P 1k ≤ 1n , P ∈ {0, 1} .

Observe that P 1k ≤ 1n implies that P is a packing matrix. This constraint is replaced by P 1k = 1n in

case we deal with partition matrices. The constraints tr(P ⊤ Qi P ) + di ≤ 0 and P ⊤ ai ≤ bi 1k might follow
from the structure of the problem under consideration. Observe that these constraints differ from the
general form (17) in the sense that the linear part tr(Bi⊤ P ) is only included in a very specific form.
A possible way to deal with the quadratic terms in (QM P1 ) is by lifting the variables in a higher-
dimensional space. By vectorizing the matrix P , the problem (QM P1 ) can be written in the form (QCQP ),
after which we can follow the approach of Section 3.1. This results in a BSDP where the matrix variable
is of order nk + 1. Since the resulting program is obtained from a lifting of the vectorization of P , we
say that we applied a vector-lifting approach. To obtain a more compact problem formulation where the
matrix variable is of lower order, we here consider a matrix-lifting approach. In particular, the objective
function can be written as tr(P ⊤ Q0 P ) = tr(Q0 P P ⊤ ) = tr(Q0 X), where X = P P ⊤ . By doing so, we
obtain the following BSDP:
min hQ0 , Xi
s.t. hQi , Xi + di ≤ 0 ∀i ∈ [m], Xai ≤ bi x ∀i ∈ [p]
  (BSDPQMP 1 )
k x⊤
 0, diag(X) = x, X ∈ {0, 1}n×n.
x X

If a QMP is defined over the partition matrices, then P 1k ≤ 1n is replaced by P 1k = 1n in (QM P1 ),

and consequently diag(X) = x is replaced by diag(X) = 1n in (BSDPQMP 1 ). By exploiting theory from
Section 2.1, we show the following equivalence.
Theorem 10. (BSDPQMP 1 ) is equivalent to (QM P1 ).
Proof. Let P be feasible for (QM P1 ) and define X = P P ⊤ and x = P 1k . Since P represents a
packing matrix, we have X ∈ {0, 1}n×n, where x is a {0, 1}–vector indicating whether object i is
packed in one of the classes or not. Then, hQi , Xi + di = hQi , P P ⊤ i + di = tr(P ⊤ Qi P ) + di ≤ 0 for
all i ∈ [m]. Moreover, we have Xai = P P ⊤ ai ≤ bi P 1k = bi x. To show that diag(X) = x, ob-
P P

serve that Xii = kj=1 Pij2 = kj=1 Pij = e⊤ i P 1k = xi . Finally, we can decompose the matrix
k x⊤
x X
⊤

 ⊤  1⊤ ⊤
into xk xX = 1Pk k
P
, showing that it is PSD. We conclude that X and x are feasible for (BSDPQMP 1 ).
To show the converse inclusion, let X and x = diag(X) be feasible for (BSDPQMP 1 ). It follows from
Corollary 2 that X can be decomposed as the sum of at most k rank-one symmetric {0, 1}–matrices. By

14
adding copies of the zero matrix in case rank(X) < k, we may assume that there exist x1 , . . . , xk ∈ {0, 1}n
P P
such that X = kj=1 xj x⊤ j . Now, let P = [x1 . . . xk ]. Then, P ∈ {0, 1}
n×k
with P 1k = kj=1 xj =
diag(X) ≤ 1n . To prove that P ⊤ ai ≤ bi 1k , consider column j ∗ of P . If all entries in P ej ∗ (= xj ∗ ) are
zero, implying that e⊤ ⊤ ∗
j ∗ P ai = 0 ≤ bi , since bi ∈ R+ . Otherwise, there exists a row i such that Pi∗ j ∗ = 1.
P k
For the i∗ th row of X, we know e⊤ i∗ X =
⊤ ⊤ ∗
j=1 (xj )i∗ xj = xj ∗ . The i th row of the system Xai ≤ bi x
then reads xj ∗ ai ≤ bi xi∗ = bi . Hence, P ai ≤ bi 1k . Finally, the constraint tr(P ⊤ Qi P ) + di ≤ 0 follows
⊤ ⊤

immediately from hQi , Xi + di ≤ 0 for all i ∈ [m]. Thus, P is feasible for (QM P1 ).
As the objective functions of (QM P1 ) and (BSDPQMP 1 ) clearly coincide with respect to the given
mapping between P and X, we conclude that the two programs are equivalent.
The matrix P no longer appears explicitly in (BSDPQMP 1 ), and therefore we will not be able to
write all quadratic problems over the packing or partition matrices in this form. The typical problems
that can be modeled as (BSDPQMP 1 ), are the ones that are symmetric over the classes [k], i.e., we do
not add constraints for one specific class. Below we discuss two examples from the literature that fit in
the framework of (QM P1 ).
Example 2 (The maximum k-colorable subgraph problem). Let G = (V, E) be a simple graph with n :=
|V | and m := |E|. Given a positive integer k, a graph is called k-colorable if it is possible to assign to
each vertex in V a color in [k] such that any two adjacent vertices get assigned a different color. The
maximum k-colorable subgraph (MkCS) problem, see e.g., [45, 55], asks to find an induced subgraph
G′ = (V ′ , E ′ ) of G that is k-colorable such that |V ′ | is maximized.
The MkCS problem can be modeled as (QM P1 ) where P ∈ {0, 1}n×k is such that Pij = 1 if and only
if vertex i ∈ [V ] is in color class j ∈ [k]. In order to model that P induces a coloring in G, we include the
constraints tr(P ⊤ (Eij + Eji )P ) = 0 for all {i, j} ∈ E. Additional constraints of the form P ⊤ ai ≤ bi 1k
do not appear in the formulation.
Now, it follows from Theorem 10 that the MkCS problem can be modeled as the following BSDP:

max hIn , Xi
 
k x⊤ (18)
s.t. Xij = 0 ∀{i, j} ∈ E,  0, diag(X) = x, X ∈ {0, 1}n×n.
x X

The doubly nonnegative relaxation of (18) obtained after replacing X ∈ {0, 1}n×n by 0 ≤ X ≤ J, equals
the formulation θk3 (G) derived in [45].
The next example shows that the parameter k in (BSDPQMP 1 ) can also be used as a variable in
order to quantify the number of classes in the solution.
Example 3 (The quadratic bin packing problem). Let a set of n items be given, each with a positive
weight wi ∈ R+ , i ∈ [n]. Assume an unbounded number of bins is available, each with total capacity W ∈
R+ and cost c ∈ R+ . Moreover, let D = (dij ) ∈ S n denote a dissimilarity matrix, where dij equals the
cost of packing item i and j in the same bin. The goal of the quadratic bin packing problem (QBPP),
see e.g., [16], is to assign each item to exactly one bin, under the condition that the total sum of weights
for each bin does not exceed W , such that the sum of the total dissimilarity and the cost of the used bins
is minimized.
Let us first consider the related problem where the number of available bins equals k. This problem
can be modeled in the form (QM P1 ), where P ∈ {0, 1}n×k is a matrix with Pij = 1 if and only if item
i is contained in bin j. Since all items need to be packed, we require P to be a partition matrix, i.e.,
P 1k = 1n . Moreover, the capacity constraints can be modeled as P ⊤ w ≤ W 1k . Theorem 10 shows that
this problem can be modeled as a BSDP where the number of bins k appears as a parameter. If we
replace k by a variable z, we obtain the following formulation of the QBPP:
  
z 1⊤ n
min , c⊕D
1n X
  (19)
z 1⊤ n n×n
s.t. Xw ≤ W 1n , diag(X) = 1n ,  0, X ∈ {0, 1} , z ∈ R.
1n X

15
The variable z is not explicitly restricted to be integer, since at an optimal solution it will always be
equal to rank(X).
The quadratic matrix program (QM P1 ) only includes specific types of constraints of the form (17).
We now consider a generalization of (QM P1 ). Let Q0 , Qi ∈ S n , B0 , Bi ∈ Rn×k and d0 , di ∈ R for
all i ∈ [m] and consider the quadratic matrix program

min tr(P ⊤ Q0 P ) + 2tr(B0⊤ P ) + d0

s.t. tr(P ⊤ Qi P ) + 2tr(Bi⊤ P ) + di ≤ 0 ∀i ∈ [m] (QM P2 )
P 1k ≤ 1n , P ∈ {0, 1}n×k .

Again, the constraint P 1k ≤ 1n can be replaced by P 1k = 1n when optimizing over partition matrices.
Now, let us consider the BSDP
 d   
k Ik
0
B0⊤ Ik P ⊤
min ,
B0⊤ Q0 P X
 d   
k Ik
i
Bi⊤ Ik P ⊤
s.t. , ≤ 0 ∀i ∈ [m] (BSDPQMP 2 )
Bi⊤ Qi P X
 
Ik P ⊤
 0, diag(X) = P 1k , X ∈ {0, 1}n×n, P ∈ {0, 1}n×k ,
P X

which is equivalent to (QM P2 ), as shown below.

Theorem 11. (BSDPQMP 2 ) is equivalent to (QM P2 ).


⊤  
⊤
Proof. Let P be feasible for (QM P2 ) and define Y ∈ {0, 1}(n+k)×(n+k) as Y = IPk IPk = IPk PX ,
P P
where X := P P ⊤ . Clearly, we have Y  0 and Xii = kj=1 Pij2 = kj=1 Pij = e⊤ i P 1k for all i ∈ [n],
showing that diag(X) = P 1k . Moreover, we have
 d   
⊤ ⊤ ⊤
i
Ik Bi⊤ Ik P ⊤
tr(P Qi P ) + 2tr(Bi P ) + di = tr(Qi X) + 2tr(Bi P ) + di = k ,
Bi⊤ Qi P X

for all i ∈ [m] and i = 0. Hence, X and P are feasible for (BSDPQMP 2 ) and the objective functions
coincide.
Conversely, let P ∈ {0, 1}n×k and X ∈ {0, 1}n×n be feasible for (BSDPQMP 2 ). Following the proof
of Theorem 2, it follows that there exist x1 , . . . , xk′ ∈ {0, 1}n with k ′ ≥ k such that P = [x1 . . . xk ]
Pk ′ ⊤ ⊤
and X = j=1 xj xj . Since diag(X) = P 1k , it follows that for all i ∈ [n] we have Xii = ei P 1k
Pk ′ P k
implying that j=1 (xj )2i = j=1 (xj )i . Since (xj )i ∈ {0, 1}, the equality above only holds if (xj )i = 0
for all j = k + 1, . . . , k . As this is true for all i ∈ [n], we have xj = 0n for all j = k + 1, . . . , k ′ , implying
′
P
that X = kj=1 xj x⊤ ⊤
j = P P . We can now follow the derivation of the first part of the proof in the
converse order to conclude that P is feasible for (QM P2 ).
Typical problems that fit in the framework of (QM P2 ) and (BSDPQMP 2 ) are quadratic matrix
programs over the packing or partition matrices that require constraints for specific classes, see e.g.,
Example 4. Another important feature of (BSDPQMP 2 ) is that it is possible to impose a condition
on the rank of X. Corollary 3 implies that if we add the constraint P ⊤ 1n ≥ 1k to (BSDPQMP 2 ), the
resulting matrix X has rank exactly k. This makes this formulation suitable for quadratic classification
problems that require an exact number of classes, e.g., the (capacitated) max-k-cut problem [28].
Example 4 (The quadratic multiple knapsack problem). Let a set of n items be given, each with a
weight wi ∈ R+ and a profit pi ∈ R+ , i ∈ [n]. We are also given a set of k knapsacks, each with a
capacity cj ∈ R+ , j ∈ [k]. Finally, let R = (riℓ ) denote a revenue matrix, where riℓ denotes the revenue of
including items i and ℓ in the same knapsack. The quadratic multiple knapsack problem (QMKP) aims
at allocating each item to at most one knapsack such that we maximize the total profit of the included
items and their interaction revenues, see [37].

16
 Let P ∈ {0, 1}n×k be a packing matrix where Pij = 1 if and only if item i is allocated to knapsack j.
The capacity constraint can be modeled as P ⊤ w ≤ c, where w ∈ Rn+ and c ∈ Rk+ denote the vector of item
weights and knapsack capacities, respectively. The total profit can be computed as hR, P P ⊤ i + p⊤ P 1k ,
where p ∈ Rn denotes the vector of item profits. It follows from Theorem 11 that we can model the
QMKP as the following binary SDP:
 1 ⊤
  
0 2 1k p Ik P ⊤
max 1 ⊤ ,
2 p1k R P X
s.t. P ⊤ w ≤ c, diag(X) = P 1k (20)
 
Ik P ⊤
 0, X ∈ {0, 1}n×n, P ∈ {0, 1}n×k .
P X

4 Problem-specific formulations
In this section we consider MISDP formulations of problems that do not belong to the binary quadratic
problems or for which the reformulation technique differs from the ones in Section 3.

4.1 The QAP as a MISDP

We present a MISDP formulation of the quadratic assignment problem (QAP) that is derived by a
matrix-lifting approach. To the best of our knowledge, our QAP formulation provides the most compact
convex mixed-integer formulation of the problem in the literature. The formulation is motivated by the
matrix-lifting SDP relaxations of the QAP derived in [21].
The quadratic assignment problem is an optimization problem of the following form:

min tr(AXBX ⊤ ) + tr(CX ⊤ ), (21)

X∈Πn

where A, B ∈ S n , C ∈ Rn×n and Πn is the set of n × n permutation matrices. The QAP is among
the most difficult N P-hard combinatorial optimization problems to solve in practice. The QAP was
introduced in 1957 by Koopmans and Beckmann [42] as a model for location problems. Nowadays, the
QAP is known as a generic model for various (real-life) problems.
By exploiting properties of the Kronecker product and Theorem 3, one can lift the QAP into the space
of (n2 + 1) × (n2 + 1) {0, 1}–matrix variables and obtain a BSDP formulation of the QAP, see Section 3.1.
Since this vector-lifting approach results in a problem formulation with a large matrix variable, we
consider here a matrix-lifting approach for the QAP. Ding and Wolkowicz [21] introduce several matrix-
lifting SDP relaxations of the QAP with matrix variables of order 3n. By imposing integrality on the
matrix variable X in one of these SDP relaxations, i.e., the relaxation MSDR 0 in [21], we obtain the
following MISDP:

min hA, Y i + hC, Xi

 
In X ⊤ R⊤
s.t.  X In Y   0, R = XB (22)
R Y Z
X ∈ Πn , R ∈ Rn×n , Y, Z ∈ S n .

Note that if B is an integer matrix, then R is also an integer matrix. However, we do not have to impose
integrality on R explicitly.
The Schur complement lemma implies that the linear matrix inequality in (22) is equivalent to
   
In Y XX ⊤ XR⊤
−  0. (23)
Y Z RX ⊤ RR⊤

Now, we are ready to prove the following result.

17
Proposition 6. The MISDP (22) is equivalent to (21).
 
In −XX ⊤ Y −XR⊤
Proof. Let (X, Y, Z, R) be feasible for (22). Then XX ⊤ = In and Y −RX ⊤ Z−RR⊤
 0 imply
⊤ ⊤ ⊤ ⊤
that Y = XR . Thus, Y = XB X = XBX , meaning that the two objectives coincide.
Conversely, let X be feasible for (21). Define R := XB, Y := XR⊤ and Z := RR⊤ . It trivially follows
that the constraints in (22) are satisfied and that the two objective functions coincide.
Many combinatorial optimization problems can be formulated as the QAP, see e.g., [14]. We provide
an example below.
Example 5 (The traveling salesman problem). Given is a complete undirected graph Kn = (V, E)
with n := |V | vertices and a matrix D = (dij ) ∈ S n , where dij is the cost of edge {i, j} ∈ E. The
goal of the traveling salesman problem (TSP) is to find a Hamiltonian cycle of minimum cost in Kn .
Let B be the adjacency matrix of the tour on n vertices, i.e., B is a symmetric Toeplitz matrix whose
first row is [0 1 0⊤n−3 1]. It is well-known, see e.g., [40], that (21) with this matrix B and A = D is a
formulation of the TSP. Therefore, a MISDP formulation of the TSP is the optimization problem (22)
where the objective is replaced by 21 hD, Y i.
Another MISDP formulation of the TSP is given in Section 4.3, see also (35). The latter formulation
is, to the best of our knowledge, the most compact formulation of the TSP.

4.2 MISDP formulations of the graph partition problem

We present here various MISDP formulations of the graph partition problem (GPP). Several of the here
derived formulations cannot be obtained by using results from Section 3.1 and Section 3.2.
The GPP is the problem of partitioning the vertex set of a graph into a fixed number of sets, say k,
of given sizes such that the sum of weights of edges joining different sets is optimized. If all sets are of
equal size, then the corresponding problem is known as the k-equipartition problem (k-EP). The case of
the GPP with k = 2 is known as the graph bisection problem (GBP). To formalize, let G = (V, E) be
an undirected graph on n := |V | vertices and let W := (wij ) ∈ S n denote a weight matrix with wij = 0
if {i, j} ∈
/ E. The graph partition problem aims to partition the vertices of G into k (2 ≤ k ≤ n − 1)
Pk
disjoint sets S1 , . . . , Sk of specified sizes m1 ≥ · · · ≥ mk ≥ 1, j=1 mj = n such that the total weight of
edges joining different sets Sj is minimized.
For a given partition of V into k subsets, let P = (Pij ) ∈ {0, 1}n×k be the partition matrix,
where Pij = 1 if and only if i ∈ Sj for i ∈ [n] and j ∈ [k]. The total weight of the partition equals:

1
1
tr W (Jn − P P ⊤ ) = tr(LP P ⊤ ), (24)
2 2
where L := Diag(W 1n ) − W is the weighted Laplacian matrix of G. The GPP can be formulated as the
following quadratic matrix program:
1
min hL, P P ⊤ i s.t. P 1k = 1n , P ⊤ 1n = m, P ∈ {0, 1}n×k , (25)
2
where m = [m1 . . . mk ]⊤ . The formulation (25) is a special case of the quadratic matrix program (QM P2 ).
Therefore, applying Theorem 11, the GPP can be modeled as follows:
1
min hL, Xi
2
s.t. P 1k = 1n , P ⊤ 1n = m, diag(X) = 1n (26)
 
Ik P ⊤
 0, X ∈ {0, 1}n×n, P ∈ {0, 1}n×k .
P X

18
The doubly nonnegative relaxation of (26) is similar to the relaxation for the k-partition problem from
[25]. For the k-EP and the GBP, we can derive simpler formulations by removing P from the model.
In the case of the k-EP, the QMP (25) is a special case of (QM P1 ), and therefore k-EP can be modeled
as follows:
1
min hL, Xi
2
n (27)
s.t. diag(X) = 1n , X1n = 1n
k
kX − Jn  0, X ∈ S n , X ∈ {0, 1}n×n.
This result follows from Theorem 10. An alternative proof is provided below.
n
Proposition 7. Let m = k 1k . Then, the QMP (25) for the k-EP is equivalent to the BSDP (27).
Proof. Let P be feasible for (25) where m = nk 1k . We define X := P P ⊤ . The first and second constraint
in (27), as well as X ∈ {0, 1}n×n follow by direct verification. Let pi be the ith column of P for i ∈ [k],
then
k k
! k !⊤
X X X X
⊤ ⊤ ⊤
kX − Jn = kP P − 1n 1n = k pi pi − pi pi = (pi − pj )(pi − pj )⊤  0.
i=1 i=1 i=1 i<j

Conversely, let X be feasible for (27). Then, it follows

Pr from ⊤Theorem P 1 and Proposition 1 that there
r
exist xi ∈ {0, 1}n, i ∈ [r], k ≥ r such that X = i=1 xi xi where i=1 xi = 1n . Since the con-
straint X1n = nk 1n is invariant under permutation of rows and columns of X, we have that the sum of
the elements in each row and column of the block matrix Jn1 ⊕ · · · ⊕ Jnr equals n/k. From this it follows
that r = k and 1⊤ n xi = n/k for i ∈ [k]. It is easy to verify that P := [x1 . . . xk ] ∈ {0, 1}
n×k
is feasible
for (25). Since the two objectives coincide, the result follows.
Next result shows that the MISDP (26) also simplifies for the GBP. It has to be noted, however, that
the GBP is not a special case of (QM P1 ).
Proposition 8. Let m = [m1 n − m1 ]⊤ , 1 ≤ m1 ≤ n/2. Then, the QMP (25) for the GBP is equivalent
to the following BSDP:
1
min hL, Xi
2
s.t. diag(X) = 1n , hJn , Xi = m21 + (n − m1 )2 (28)

2X − Jn  0, X ∈ S n , X ∈ {0, 1}n×n.
Proof. Let P be feasible for (25). We define X := P P ⊤ . The first and second constraint in (28) follow
by direct verification. Let pi be the ith column of P for i ∈ [2], then
2 2
! 2 !⊤
X X X
2X − Jn = 2P P ⊤ − 1n 1⊤ n =2 pi p⊤
i − pi pi = (p1 − p2 )(p1 − p2 )⊤  0.
i=1 i=1 i=1

Conversely, let X be feasible for (28). Then, it follows from Theorem 1 and Corollary 2 that there
exist x1 , x2 ∈ {0, 1}n such that X = x1 x⊤1 + x2 x2
⊤
where x1 + x2 = 1n . Note that X cannot have
rank one or zero for 1 ≤ m1 < n. From hJn , Xi = m21 + (n − m1 )2 , it follows that 1⊤ n x1 = m1 or
1⊤ ⊤
n x1 = n − m1 . Without loss of generality, we assume that 1 x1 = m1 . Clearly, P := [x1 x2 ] is feasible
for (25). Moreover, the two objective functions coincide.
In the remainder of this section, we derive yet another alternative MISDP formulation of the GPP,
different from (26). For that purpose we notice that the GPP can also be formulated as a QMP of the
following form:
min tr(P ⊤ Q0 P ) + tr(P C0 P ⊤ ) + 2tr(B0⊤ P ) + d0
s.t. tr(P ⊤ Qi P ) + tr(P Ci P ⊤ ) + 2tr(Bi⊤ P ) + di ≤ 0 ∀i ∈ [m] (QM P3 )
n×k
P ∈R ,

19
where Qi ∈ S n , Ci ∈ S k , Bi ∈ Rn×k , di ∈ R for i = 0, 1, . . . , m. Note that (QM P2 ) is a special case
of (QM P3 ). Examples of problems that are of this form are quadratic problems with orthogonality
constraints, see e.g., [4]. The GPP can be formulated as follows, see e.g., [20]:
1
min hL, P P ⊤ i
2
s.t. P ⊤ 1n = m, P ⊤ P = Diag(m) (29)

diag(P P ⊤ ) = 1n , P ≥ 0, P ∈ Rn×k .

To reformulate (29) as a MISDP we introduce matrices X1 ∈ S n and X2 ∈ S k such that X1 = P P ⊤

and X2 = P ⊤ P and relax these matrix equalities to the linear matrix inequalities (LMIs) X1 − P P ⊤  0
and X2 − P ⊤ P  0, respectively. These can be rewritten as
   
Ik P ⊤ In P
 0 and  0.
P X1 P ⊤ X2

After introducing the constraints diag(X1 ) = 1n and X2 = Diag(m), we obtain the following MISDP:
1
min hL, X1 i
2
s.t. P 1k = 1n , diag(X1 ) = 1n , X2 = Diag(m) (30)
   
Ik P ⊤ In P
 0,  0, X1 ∈ S n , X2 ∈ S k , P ∈ {0, 1}n×k .
P X1 P ⊤ X2

We prove below that (30) is an exact formulation of the GPP.

Proposition 9. The MISDP (30) is an exact formulation of the GPP.
Proof. We prove the result by showing the equivalence between (29) and (30).
P
Let P ∈ Rn×k be feasible for (29). Then, it follows from diag(P P ⊤ ) = 1n that (P P ⊤ )ii = kj=1 Pij2 =
⊤
1 for i ∈ [n]. From
P this and P ≥ 0, we obtain 0 ≤ Pij ≤ 1 for all i ∈ [n], j ∈ [k]. From
PP 12n = m
⊤ ⊤
it follows that i,j Pij = n and from P P = Diag(m) that tr(P P ) = n, and thus i,j Pij = n.
⊤
Therefore, Pij ∈ {0, 1} for all i ∈ [n], j ∈ [k]. The equality diag(P P ) = 1n then implies that P 1k = 1n .
It follows from the discussion prior to the proposition that X1 := P P ⊤ and X2 := P ⊤ P are feasible
for (30).
Conversely, let X1 , X2 and P be feasible for (30). From P ∈ {0, 1}n×k and P 1k = 1n it follows that
diag(P P ⊤ ) = 1n . From X2 − P ⊤ P  0 and 1⊤ k (X2 −PP ⊤ P )1k = P
0 it follows that (X2 − P ⊤ P )1k = 0
⊤ ⊤ n 2 n
and thus P 1n = m. Moreover, we have (P P )ii = j=1 Pji = j=1 Pji = mi for i ∈ [k], implying
that diag(P ⊤ P ) = m. Finally, we have X2 − P ⊤ P = Diag(m) − P ⊤ P  0, where it follows from above
that the latter matrix has a diagonal of zeros. Thus, we must have P ⊤ P = Diag(m), which concludes
the proof.
The MISDP (30) has two LMIs and requires integrality constraints only on a matrix of size n × k,
while (26) has only one LMI and asks for integrality on matrices of size n × n and n × k.

4.3 MISDP formulations via association schemes

Association schemes provide a unifying framework for the treatment of problems in several branches of
mathematics, including algebraic graph theory and coding theory. Delsarte [19] was the first to combine
association schemes and linear programming for codes. Schrijver [63] refined the Delsarte bound by using
semidefinite programming. De Klerk et al. [39] introduce a framework for deriving SDP relaxations of
optimization problems on graphs by using association schemes. In Section 4.3.1 (resp. Section 4.3.2) we
derive MISDP formulations of the TSP (resp. k-EP) by exploiting association schemes.
An association scheme of rank r is a set {A0 , A1 , . . . , Ar } ⊆ Rn×n of {0, 1}–matrices that satisfy the
following properties:

20
 Pr
(i) A0 = In and i=0 Ai = Jn ;
(ii) A⊤
i ∈ {A0 , A1 , . . . , Ar } for i ∈ {0, 1, . . . , r};

(iii) Ai Aj = Aj Ai for i, j ∈ {0, 1, . . . , r};

Pr
(iv) There exist scalars phij such that Ai Aj = h=0 phij Ah for i, j ∈ {0, 1, . . . , r}.

The numbers phij are called the intersection numbers of the association scheme. For more background on
association schemes, we refer to [11, 30].
We restrict here to symmetric association schemes, i.e., we assume that all matrices Ai are symmetric.
Matrices Ai are linearly independent and they generate a commutative (r + 1)-dimensional algebra of
symmetric matrices. This algebra is called the Bose-Mesner algebra of the association scheme. Since the
matrices Ai for all i = 0, 1, . . . , r commute, they can be simultaneously diagonalized. Moreover, there
exists a unique basis for the Bose-Mesner algebra consisting of minimal Pr idempotents Ei , i = 0, 1, . . . , r.
These matrices satisfy Ei Ej = δij Ei for all i, j ∈ {0, 1, . . . , r} and i=0 Ei = In , where δij is the Kro-
necker delta function. We assume without loss of generality that E0 = n1 Jn . Bases {A0 , A1 , . . . , Ar }
and {E0 , E1 , . . . , Er } are related by:
r
X r
1X
Aj = Pij Ei and Ej = Qij Ai j = 0, 1, . . . , r. (31)
i=0
n i=0

Parameters Pij and Qij for i, j = 0, 1, . . . , r are known as the eigenvalues of the association scheme
and P = (Pij )ri,j=0 and Q = (Qij )ri,j=0 are the eigenmatrices of the association scheme.
Distance-regular graphs are closely related to association schemes. Classes of distance-regular graphs
include complete graphs, complete bipartite graphs and cycle graphs. Let us define the distance-i graph
Γi as the graph with vertex set V , in which two vertices x and y are adjacent if and only if the length
of the shortest path between x and y in G is i. The adjacency matrix Ai of Γi is called the distance-i
matrix of G for i = 0, 1, . . . , d. It is known that the set of so generated matrices {A0 , A1 , . . . , Ar } form
an association scheme. It follows from the construction that this association scheme forms a (regular)
partition of the edge set of a complete graph.
If {A0 , A1 , . . . , Ar } is an association scheme and P ∈ Πn , then {P ⊤ A0 P, P ⊤ A1 P, . . . , P ⊤ Ar P } is also
an association scheme. Let us define the following polytope:

H(A0 , A1 , . . . , Ar ) := conv (P ⊤ A0 P, P ⊤ A1 P, . . . , P ⊤ Ar P ) : P ∈ Πn . (32)

The vertices of H(A0 , A1 , . . . , Ar ) are of the form P ⊤ A0 P, P ⊤ A1 P, . . . , P ⊤ Ar P with P ∈ Πn . De
Klerk et al. [39] show that many combinatorial optimization problems may be formulated as optimization
problems over the polytope (32). The key is to formulate a combinatorial optimization problem as the
problem of finding an optimal weight distance-regular subgraph of a weighted complete graph. Since
the distance matrices of a distance-regular graph form an association scheme, the problem can then be
reformulated as an optimization problem over the polytope (32) for an appropriate association scheme.
For a given association scheme {A0 , A1 , . . . , Ar }, De Klerk et al. [39] approximate H(A0 , A1 , . . . , Ar )
by the following larger set:
 r
X r
X 
A := (X0 , . . . , Xr ) : X0 = In , Xi , = Jn , Qij Xi  0, j ∈ [r], Xi ≥ 0, Xi = Xi⊤ , i ∈ [r] , (33)
i=0 i=0

where Qij are the dual eigenvalues of the association scheme {A0 , A1 , . . . , Ar }. They exploit the set A to
derive SDP relaxations of various combinatorial optimization problems. The LMIs in A follow from the
properties of the basis of minimal idempotents, see (31) and [39] for details. In the sequel, we show that
by imposing integrality to only one of the matrices in (33), one obtains an exact problem formulation.

21
4.3.1 The traveling salesman problem
Let us reconsider the traveling salesman problem (TSP), see Example 5. In this section we assume that
n is odd. The results for n even can be derived similarly.
The adjacency matrix of a Hamiltonian cycle is a symmetric circulant matrix that belongs to the
association scheme of symmetric circulant matrices, known as the Lee scheme [47]. The dual eigenvalues
of the Lee scheme are Q0j = 2 for j ∈ [r] and Qi0 = 1 for i = 0, 1, . . . , r, where r = ⌊n/2⌋. After
incorporating those dual eigenvalues in (33), the authors of [39, 40] derive an SDP relaxation for the
TSP. By imposing integrality on the variable X1 in the SDP relaxation from [39, 40] we obtain the
following MISDP:
1
min hD, X1 i
2
Xr Xr  
2ijπ (34)
s.t. In + Xi = Jn , In + cos Xi  0 ∀j ∈ [r]
i=1 i=1
n
X1 = X1⊤ ∈ {0, 1}n×n, Xi ≥ 0, Xi ∈ S n ∀i = 2, . . . , r,

where r = ⌊n/2⌋. Next, we show that the MISDP (34) is an exact model of the TSP.
Theorem 12. The MISDP (34) is an exact formulation of the TSP.
Proof. To prove the result we use the following ISDP formulation of the TSP by Cvetković et al. [18]:
1
min hD, Xi
2   
2π (35)
s.t. 2In − X + 2 1 − cos (Jn − In )  0
n
X1n = 21n , diag(X) = 0, X ∈ S n , X ∈ {0, 1}n×n.

The ISDP (35) is derived by exploiting the algebraic connectivity of a cycle on n vertices.
Let X be feasible for (35). Then, X is the adjacency matrix of a Hamiltonian cycle, hence X is
a symmetric circulant matrix. Define X0 := In , X1 := X and X2 := X12 − 2In , and the recurrence
relation X1 Xi = Xi+1 + Xi−1 for i = 2, . . . , r − 1. Then, the set {X0 , X1 , . . . , Xr } where r = ⌊n/2⌋
forms an association scheme of (permuted) symmetric circulant matrices. Note that Xi i = 0, 1, . . . , r
are the distance matrices of the Hamiltonian cycle whose adjacency matrix is X1 . The LMIs in (34)
Pr from the fact that the corresponding basis of minimal idempotents {E0 , . . . , Er } satisfies Ej =
follow
1
n i=0 Qij Xi  0, for j = 0, 1, . . . , r. See also [40] for a direct proof of the validity of these LMIs that
exploits the fact that the Xi ’s may be simultaneously diagonalized.
Conversely, let Xi for i ∈ [r] be feasible for (34). Define X := X1 . Now, Pr similarly as in [40, Theorem
4.1] one can show that 2In + (1 − 2 cos (2π/n)) X1 + (2 − 2 cos (2π/n)) Pr i=2 Xi  0 may be obtained as
a nonnegative linear combination of r LMIs from (34) and In + i=1 Xi  0. From [39, Theorem 7.3],
it follows that X1n = 21n . Since the two objectives clearly coincide, the result follows.

4.3.2 The k-equipartition problem

In Section 4.2 we derived various MISDP formulations of the GPP, including the special case of the k-EP.
We now derive an alternative formulation of the k-EP by exploiting an appropriate association scheme.
Let n, k, m ∈ Z+ , and D be a nonnegative symmetric matrix of order n, where n = mk. The k-EP
can be equivalently formulated as finding a complete regular k-partite subgraph on n vertices in Kmk of
minimum weight. The complete regular k-partite graph is strongly regular. The dual eigenvalues of the
corresponding association scheme are given in the following matrix:
 
1 (m − 1)k k − 1
Q = 1 0 −1  . (36)
1 −k k−1

22
De Klerk et al. [39] combine (33) and (36) to derive an association scheme-based SDP relaxation of
the k-EP. By imposing integrality on the matrix variable X2 in the SDP relaxation from [39] we obtain:
min hD, X1 i
s.t. (m − 1)In − X2  0, (k − 1)In − X1 + (k − 1)X2  0 (37)
n n×n
In + X1 + X2 = Jn , X1 , X2 ∈ S , X1 ≥ 0, X2 ∈ {0, 1} .
In the sequel we show that the MISDP (37) is an exact formulation of the k-EP. Note that in (37) it is
sufficient to require that X2 ∈ {0, 1}n×n, since the integrality of X1 follows from In + X1 + X2 = Jn .
Proposition 10. The BSDP (27) is equivalent to the MISDP (37).
Proof. Let X1 , X2 be feasible for (37). We define matrix X := In + X2 , and show that it is feasible
for (27). It follows trivially that diag(X) = 1n . Further, X1n = (In + X2 )1n = (In + A2 )1n = m1n
follows from the fact that X2 1n = A2 1n , see [39, Theorem 7.3], where A2 := Diag(1k ) ⊗ (Jm − Im ). To
verify the PSD constraint we proceed as follows:
(k − 1)In − X1 + (k − 1)X2 = k(In + X2 ) − (In + X1 + X2 ) = kX − Jn  0.
Conversely, let X be feasible for (27). Define X1 := Jn − X and X2 := X − In , from where it follows
that X1 , X2 ∈ {0, 1}n×n. Then, the first constraint in (37) is trivially satisfied and
(k − 1)In − X1 + (k − 1)X2 = (k − 1)In − Jn + X + (k − 1)(X − In ) = kX − Jn  0.
For the second LMI, we have (m − 1)In − X2 = mIn − X  0, where we use the fact that the spectral
radius ρ(X) ≤ kXk∞ = m. The two objective functions clearly coincide.
Note that the MISDP (37) has two LMIs and the BSDP (27) only one. Moreover, observe that one
may replace in (37) the constraint (m − 1)In − X2  0 by X2 1n = (m − 1)1n and obtain a MISDP for
the k-EP with only one PSD constraint. This follows directly from the proof of Proposition 10.

4.4 MISDP formulations beyond binarity

Almost all problem formulations that have been discussed before involve matrix variables with entries
in {0, 1}. In this section we consider several problems that allow for semidefinite formulations where
(some of) the variables are integers, but not necessarily restricted to {0, 1}.
Example 6 (The integer matrix completion problem). A well-known problem in data analysis is the
problem of low-rank matrix completion. Suppose a partially observed data matrix is given, i.e., let Ω ⊆
[n] × [m] denote the set of observed entries and let D ∈ Rn×m denote a given data matrix that has its
support on Ω. The goal of the low-rank matrix completion problem is to find a minimum rank matrix
X ∈ Rn×m such that X coincides with D on the set Ω, e.g., see [1].
Since minimizing rank(X) leads to a nonconvex and therefore hard problem,
Pn a related but tractable
alternative is given by minimizing the nuclear norm of X, i.e., ||X||∗ := i=1 σi (X), where σi denotes
the ith singular value of X. Hence, we consider the following program:
min ||X||∗ s.t. Xij = Dij for all (i, j) ∈ Ω.
X∈Rn×m

As shown by Recht et al. [60], the optimization problem above is equivalent to the following SDP:
min hIn , Z1 i + hIm , Z2 i
 
Z1 X
s.t.  0, Xij = Dij for all (i, j) ∈ Ω.
X ⊤ Z2
Since X can be thought of as a multiplicative model underlying the data observed in D, see [1], a possible
generalization would be to require the entries in X to be integer. This yields the following integer matrix
completion problem:
min hIn , Z1 i + hIm , Z2 i

23
 s.t. Xij = Dij for all (i, j) ∈ Ω, Xij ∈ S for all (i, j) ∈
/Ω
 
Z1 X
 0.
X ⊤ Z2

where S ⊆ Z is a discrete set.

Example 7 (The sparse integer least squares problem). In the integer least squares problem we are given
a matrix M ∈ Rn×k and a column b ∈ Rn and we seek the closest point to b in the lattice spanned by the
columns of M . Del Pia and Zhou [58] consider the related sparse integer least squares (SILS) problem,
which can be formulated as
1
min ||M x − b||22 s.t. x ∈ {0, ±1}k , ||x||0 ≤ K. (38)
n
The SILS problem has applications in, among other, multiuser detection and sensor networks, see [58]
and the references therein. Now, consider the following ternary SDP:
   ⊤ 
1 1 x⊤ b b −b⊤ M
min ,
n x X −M ⊤ b M ⊤ M
s.t. tr(X) ≤ K, diag(X) = y1 + y2 , x = y1 − y2 (39)
   
1 x⊤ 1 x⊤
 0, ∈ {0, ±1}(k+1)×(k+1) , y1 , y2 ∈ Rn+ .
x X x X

It is easy to verify that if x is a solution to (38), then x, X = xx⊤ , y1 = max(x, 0) and y2 = max(−x, 0)
is feasible for (39) with the same objective value. Conversely, if (x, X, y1 , y2 ) is feasible for (39), it follows
from Proposition 5 that x = y1 − y2 is a solution to (38) with the same objective value.

5 Conclusions
In this paper we showed that the class of mixed-integer semidefinite programs embodies a rich structure,
allowing for compact formulations of many well-known discrete optimization problems. Due to the recent
progress in computational methods for solving MISDPs [27, 36, 38, 41, 52, 53], these formulations can be
exploited to obtain alternative methods for solving the problems to optimality.
As most problems are naturally encoded using binary or ternary variables, we started our research
with a study on the general theory related to PSD {0, 1}–, {±1}– and {0, ±1}–matrices. Section 2
provides a comprehensive overview on this matter, including known and new results. In particular, we
presented a combinatorial, polyhedral, set-completely positive and integer hull description of the set of
PSD {0, 1}–matrices bounded by a certain rank, see Section 2.1. Several of these results are extended to
matrices having entries in {±1} and {0, ±1}.
Based on these matrix results, in particular Theorem 1–3 and Corollary 3, we derived a generic
approach to model binary quadratic problems as BSDPs. We derived a BSDP for the class of binary
quadratically constrained quadratic programs, see (BSDPQCQP ), and for two types of binary quadratic
matrix programs, see (BSDPQMP 1 ) and (BSDPQMP 2 ). These results are widely applicable to a large
number of discrete optimization problems, see also the examples in Section 3.
We moreover we considered problem-specific MISDP formulations that are derived in a different way
than through this generic approach. We provided compact MISDP formulations of the QAP, see (22),
and various variants of the GPP, see (26), (27) and (28). We derived several MISDP formulations of
discrete optimization problems that can be modeled using association schemes, see Section 4.3. We also
considered problems that have discrete but non-binary variables, e.g., the integer matrix completion
problem and the sparse integer least squares problem, see Example 6 and 7, respectively.
Given the wide range of discrete optimization problems for which we derived new formulations based
on mixed-integer semidefinite programming, we expect more problems to allow for such representations.
It is also interesting to study the behaviour of MISDP solvers on the presented formulations to see whether
this leads to competitive solution approaches for the considered problems.

24
References
[1] acm sigkdd and netflix. Proceedings of KDD Cup and Workshop. Available at
http://www.cs.uic.edu/~liub/KDD-cup-2007/proceedings.html, 2007.
[2] M.F. Anjos and J.B. Lasserre. Handbook on semidefinite, conic and polynomial optimization, volume
166. Springer Science & Business Media, 2011.
[3] M.F. Anjos and H. Wolkowicz. Strengthened semidefinite relaxations via a second lifting for the
max-cut problem. Discrete Appl. Math., 119(1–2):79–106, 2002.
[4] K. Anstreicher and H. Wolkowicz. On Lagrangian relaxation of quadratic matrix constraints. SIAM
J. Matrix Anal. Appl., 22(1):41–55, 2000.
[5] H.J. Bandelt, M. Oosten, J.H.G.C. Rutten, and F.C.R. Spieksma. Lifting theorems and facet char-
acterization for a class of clique partitioning inequalities. Oper. Res. Lett., 24:235–243, 1999.
[6] A. Beck. Quadratic matrix programming. SIAM J. Optim., 17:1224–1238, 2007.
[7] E.T. Bell. Exponential numbers. Amer. Math. Monthly, 41:411–419, 1934.
[8] H.Y. Benson and Ü. Sağlam. Mixed-integer second-order cone programming: A survey. In Theory
Driven by Influential Applications, pages 13–36. INFORMS, 2013.
[9] A. Berman and C. Xu. {0,1} completely positive matrices. Linear Algebra Appl., 399:35–51, 2005.
[10] I.M. Bomze and M. Gable. Optimization under uncertainty and risk: Quadratic and copositive
approaches. Eur. J. Oper. Res., 310:449–476, 2023.
[11] A.E. Brouwer and W.H. Haemers. Handbook of Combinatorics, chapter Association schemes, pages
747–771. Elsevier Science, Amsterdam, 1995.
[12] T. Bulhões, A. Pessoa, F. Protti, and E. Uchoa. On the complete set packing and set partitioning
polytopes: Properties and rank 1 facet. Oper. Res. Lett., 46:389–392, 2018.
[13] S. Burer. On the copositive representation of binary and continuous nonconvex quadratic programs.
Math. Program., 120:479–495, 2009.
[14] R. Burkard, M. Dell’Amico, and S. Martello. Assignment Problems. Society for Industrial and
Applied Mathematics, Philadelphia, PA, USA, 2009.
[15] A. Cerveira, A. Agra, F. Bastos, and J. Gromicho. A new branch and bound method for a discrete
truss topology design problem. Comput. Optim. Appl., 54(1):163–187, 2013.
[16] V.G. Chagas. Exact Algorithms for the Quadratic Bin Packing Problem. PhD thesis, Universidade
Estadual de Campinas, 2021.
[17] S. Chopra and M.R. Rao. The partition problem. Math. Program., 59:87–115, 1993.
[18] D. Cvetković, M. Čangalović, and V. Kovačević-Vujčić. Semidefinite programming methods for the
symmetric traveling salesman problem. In G. Cornuj́ols, R.E. Burkard, and G.J. Woeginger, editors,
Integer programming and Combinatorial Optimization (IPCO 1999), volume 1610 of Lecture Notes
in Computer Science. Springer, Berlin, Heidelberg, 1999.
[19] P. Delsarte. An algebraic approach to the association schemes of coding theory. Philips Research
Reports Suppl, 10, 1973.
[20] Y. Ding, D. Ge, and H. Wolkowicz. On equivalence of semidefinite relaxations for quadratic matrix
programming. Math. Oper. Res., 36(1):88–104, 2011.
[21] Y. Ding and H. Wolkowicz. A low-dimensional semidefinite relaxation for the quadratic assignment
problem. Math. Oper. Res., 34(4):1008–1022, 2009.
[22] B.P.M. Duarte. Exact optimal designs of experiments for factorial models via mixed-integer semidef-
inite programming. Math., 11(4):854, 2023.
[23] I. Dukanovic and F. Rendl. Semidefinite programming relaxations for graph coloring and maximal
clique problems. Math. Program., 109:345–365, 2007.
[24] A. Eisenblätter. Frequency assignment in GSM networks: Models, heuristics, and lower bounds.
PhD thesis, 2001.
[25] Jamie Fairbrother and Adam N. Letchford. Projection results for the k-partition problem. Discrete
Optim., 26:97–111, 2017.

25
[26] T. Gally and M.E. Pfetsch. Computing restricted isometry con-
stants via mixed-integer semidefinite programming. Optimization Online:
http://www.optimization-online.org/DB_FILE/2016/04/5395.pdf, 2016.
[27] T. Gally, M.E. Pfetsch, and S. Ulbrich. A framework for solving mixed-integer semidefinite programs.
Optim. Methods Softw., 33(3):594–632, 2018.
[28] D.R. Gaur, R. Krishnamurti, and R. Kohli. The capacitated max k-cut problem. Math. Program.,
115:65–72, 2008.
[29] W. Gil-González, A. Molina-Cabrera, O.D. Montoya, and L.F. Grisales-Noreña. An MI-SDP model
for optimal location and sizing of distributed generators in dc grids that guarantees the global
optimum. Appl. Sci., 10(21):7681, 2020.
[30] C.D. Godsil. Algebraic Combinatorics. Chapman & Hall, New York, 1993.
[31] M. X. Goemans and D. P. Williamson. Improved approximation algorithms for maximum cut and
satisfiability problems using semidefinite programming. J. ACM, 42(6):1115–1145, 1995.
[32] M. Grötschel, L. Lovász, and A. Schrijver. Geometric algorithms and combinatorial optimization,
volume 2. Springer, Berlin, 1988.
[33] M. Grötschel and Y. Wakabayashi. Facets of the clique partitioning polytope. Math. Program.,
47:367–387, 1990.
[34] C. Helmberg. Fixing variables in semidefinite relaxations. SIAM J. Matrix Anal. Appl., 21(2):952–
–969, 2000.
[35] C. Helmberg. Semidefinite programming for combinatorial optimization. habilitation, Germany,
2000.
[36] H. Hijazi, G. Wang, and C. Coffrin. GravitySDP: A solver for sparse mixed-integer semidefinite
programming, 2018. Available at https://github.com/coin-or/Gravity/tree/GravitySDP.
[37] A. Hiley and B.A. Julstrom. The quadratic multiple knapsack problem and three heuristic approaches
to it. In P. Heggernes, editor, Genetic and Evolutionary Computation Conference (GECCO), pages
547–552, 2006.
[38] C. Hojny and M.E. Pfetsch. Handling symmetries in mixed-integer semidefinite programs. In Interna-
tional Conference on Integration of Constraint Programming, Artificial Intelligence, and Operations
Research, pages 69–78. Springer, 2023.
[39] E. de Klerk, F.M. Oliveira Filho, and D.V. Pasechnik. Relaxations of combinatorial problems via
association schemes. In Miguel F. Anjos and Jean B. Lasserre, editors, Handbook on Semidefinite,
Conic and Polynomial Optimization, International Series in Operations Research & Management
Science, chapter 0, pages 171–199. Springer, June 2012.
[40] E. de Klerk, D.V. Pasechnik, and R. Sotirov. On semidefinite programming relaxations of the
traveling salesman problem. SIAM J. Optim., 19(4):1559–1573, 2009.
[41] K. Kobayashi and Y. Takano. A branch-and-cut algorithm for solving mixed-integer semidefinite
optimization problems. Comput. Optim. Appl., 75:493–513, 2020.
[42] T. C. Koopmans and M. Beckmann. Assignment problems and the location of economic activities.
Econometrica, 25(1):53–76, 1957.
[43] M. Kočvara. Truss topology design with integer variables made easy. Optimization Online:
https://optimization-online.org/2010/05/2614/, 2010.
[44] N. Krislock, J. Malick, and F. Roupin. A semidefinite branch-and-bound method for solving binary
quadratic problems. ACM Trans. Math. Softw., 43(4), 2017.
[45] O. Kuryatnikova, R. Sotirov, and J.C. Vera. The maximum k-colorable subgraph problem and
related problems. INFORMS J. Comput., 34(1):656–669, 2021.
[46] M. Laurent, S. Poljak, and F. Rendl. Connections between semidefinite relaxations of the max-cut
and stable set problems. Math. Program., 77:225–246, 1997.
[47] S.L. Lee, Y.L. Luo, B.E. Sagan, and Y.N. Yeh. Eigenvector and eigenvalues of some special graphs.
IV. multilevel circulants. Int. J. Quantum Chem., 41(1):105–116, 1992.
[48] A.N. Letchford and M.M. Sørensen. Binary positive semidefinite matrices and associated integer
polytopes. Math. Program. Series A, 131:253–271, 2012.

26
[49] F. Lieder, F.B.A. Rad, and F. Jarre. Unifying semidefinite and set-copositive relaxations of binary
problems and randomization techniques. Comput. Optim. Appl., 61:669–688, 2015.
[50] M. Lubin, J.P. Vielma, and I. Zadik. Mixed-integer convex representability. arXiv:1706.05135v3,
2020.
[51] S. Mars. Mixed-integer semidefinite programming with an application to truss topology design. PhD
thesis, FAU Erlangen-Nürnberg, 2013.
[52] F. Matter and M.E. Pfetsch. Presolving for mixed-integer semidefinite optimization. INFORMS J.
Optim., 5(2):131–154, 2022.
[53] F. de Meijer and R. Sotirov. The Chvátal-Gomory procedure for integer SDPs with applications in
combinatorial optimization. arXiv:2201.10224v2, 2023.
[54] P. Meurdesoif. Strengthening the Lovász bound for graph coloring. Math. Program., 102:577 – 588,
2005.
[55] G. Narasimhan. The maximum k-colorable subgraph problem. PhD thesis, University of Wisconsin-
Madison, 1989.
[56] M. Oosten, J.H.G.C. Rutten, and F.C.R. Spieksma. The clique partitioning problem: Facets and
patching facets. Networks, 38:209–226, 2001.
[57] M. Padberg. The Boolean quadric polytope: some characteristics, facets and relatives. Math.
Program., 45:139–172, 1989.
[58] A. Del Pia and D. Zhou. An SDP relaxation for the sparse integer least square problem.
arXiv:2203.02607, 2023.
[59] M. Pilanci, M.J. Wainwright, and L. El Ghaoui. Sparse learning via Boolean relaxations. Math.
Prog., 151:63–87, 2015.
[60] B. Recht, M. Fazel, and P.A. Parrilo. Guaranteed minimum-rank solutions of linear matrix equations
via nuclear norm minimization. SIAM Rev., 52(3):471–501, 2010.
[61] F. Rendl, G. Rinaldi, and A. Wiegele. Solving max-cut to optimality by intersecting semidefinite
and polyhedral relaxations. Math. Program., 121(2):307, 2010.
[62] A. Schrijver. A comparison of the Delsarte and Lovász bounds. IEEE Trans. Inform. Theory,
25:425–429, 1979.
[63] A. Schrijver. New code upper bounds from the Terwilliger algebra and semidefinite programming.
IEEE Transactions on Information Theory, 51:2859–2866, 2005.
[64] L. Wei, A. Atamtürk, A. Gómez, and S. Küçükyavuz. On the convex hull of convex quadratic
optimization problems with indicators. arXiv:2201.00387v2, 2022.
[65] K. Yonekura and Y. Kanno. Global optimization of robust truss topology via mixed integer semidef-
inite programming. Optim. Eng., 11(3):355–379, 2010.
[66] X. Zheng, H. Chen, Y. Xu, Z. Li, Z. Lin, and Z. Liang. A mixed-integer SDP solution to distribu-
tionally robust unit commitment with second order moment constraints. CSEE J. Power Energy
Syst., 6(2):374–383, 2020.

27