Fourth International Conference

Modelling and Development of Intelligent Systems

October 28 - November 1, 2015
“Lucian Blaga” University
Sibiu - Romania

A continuous reformulation of the quadratic unconstrained

binary optimization problem
Vasile Moraru, Sergiu Zaporojan

Abstract
In this paper we consider the Quadratic Unconstrained Binary Optimization (QUBO) Problem.
Using a suitable function and penalty parameter we can reformulate the original QUBO problem as a
continuous program. It is shown that the problem of large size can be reduced to two constraints. A
new convex formulation is then proposed.

1 Introduction
In this paper we consider the quadratic unconstrained binary optimization (QUBO) problem:
n n n ⎫
f ( x) = xT Qx + cT x = ∑∑
i =1 j =1
qij xi x j + ∑c x
i =1
i i → min ⎪
⎬ (1)
⎪
subject to xi ∈ {0, 1}, ∀i = 1,2, K , n, ⎭
where Q is a n × n real symmetric matrix, but not necessarily positive semidefinite, c is a constant
vector: c = (c1 , c2 , K , cn )T ∈ ℜ n and x is an n-vector of binary variables: x ∈ {0, 1}n . The superscript
“T” indicates transposition. ℜ n is the Euclidean space of n-dimensional column vectors with the
n
inner product xT y = ∑x y
i =1
i i and ℜ denote the set of real numbers.

The problem considered by quadratic programming (1) occurs in different applications [1], [2],
[3]. There are several problems of decisions whose modeling is reduced to the binary quadratic
optimization. Such problems are frequently found in Operations Research, in Graph Theory, in
Economic Science and in other areas [4], [5]. The QUBO problem is a problem of combinatorial
optimization and is well known as NP – hard [6] and so, is difficult to solve. For example, the
well studied max-cut problem is a problem of QUBO.
There were developed and proposed different methods of solving and various relaxation
techniques of the considered problem:
• Semidefinite Relaxation [7]
• Lagrangian Relaxations [7]
• Linearization Techniques [8], [9]
• Convex – programming Relaxation [10], [11], [12], [13]
• Heuristic Methods [4], [14].
The simplest relaxation technique is the continuous relaxation that consists in replacing the
discrete conditions xi ∈ {0, 1}, i = 1,2, K , n with continuous constraints 0 ≤ xi ≤ 1 for all i = 1,2, K , n .
In the present paper are considered some representations of the binary set {0, 1}n by means of which
continuous reformulations for the QUBO problem could be obtained. But there are many large
dimension problems at practical applications which are intractable.
This paper is organized as follows: in Section 2 we extend a well-known result on representation
the constraints xi ∈ {0, 1}, i = 1,2, K , n ; in the third Section we show how one can considerably

86
 Vasile Moraru, Sergiu Zaporojan

reduce a big number of simple constraints 0 ≤ xi ≤ 1, i = 1,2, K , n, to only two convex constraints. In
Section 4, by means of penalization parameters, the initial problem is reduced to a known problem
in literature - DC problem.

2 Continuous Reformulation of QUBO Problem

The QUBO problem can be transformed into a continuous optimization problem, represented the
binary set {0, 1}n through inequations system:
p( x) ≤ 0, ⎫
⎬ (2)
0 ≤ x ≤ e,⎭
where e = (1, 1, L ,1)T is the vector, all components of which are equal to one and p( x) : ℜ n → ℜ is a
continuous function that checks the inequality and the equality:
p( x) ≥ 0, ∀x ∈ [0,1]n , ⎫⎪
⎬ (3)
p( x) = 0 and if and only if x ∈ {0,1}n . ⎪⎭
Acting as p(x) with the above properties (3) has been proposed and used the concave functions [15],
[16]:
n
p( x) = ∑ x (1 − x ) = x
i =1
i i
T
(e − x )

or
n
p( x) = ∑ min( x ,1 − x ) .
i =1
i i

There is infinity of such functions. In what follows, we would propose three other continuous
functions with similar features.
Let the function
⎧2u (1 − 2u ), if u ≤ 0,
⎪ 2
p(u ) = (1 − u ) 1 − u + u u − (1 − 2u ) = ⎨2u (1 − u ), if 0 ≤ u ≤ 1, (4)
⎪2u (3 − 2u ) − 2, if u ≥ 1.
⎩
Function p(u ) is a concave function, continuously differentiable to derivatives:
⎧2(1 − 4u ), if u ≤ 0,
⎪
p′(u ) = ⎨2(1 − 2u ), if 0 ≤ u ≤ 1,
⎪2(3 − 3u ) − 2, if u ≥ 1.
⎩
It is found easily that the function

∑ [(1 − x )1 − x ]
n
p1 ( x) = i i + xi xi − (1 − 2 xi ) 2 (5)
i =1
has the required properties.
Another function would be
⎧ 1
⎪⎪2u, if u ≤ 2 ,
p(u ) = 1 − 2u − 1 = ⎨ (6)
⎪2(1 − u ), if u ≥ 1 ,
⎪⎩ 2
which generates the function

∑ [(1 − 2 x − 1 ] .
n
p2 ( x ) = i (7)
i =1
We bring another example of a function p(u ) that is twice continuously differentiable but is not
concave on the interval [0,1] :

87
 A continuous reformulation of the quadratic unconstrained binary optimization problem

⎧ 1
⎪ u (8u 2 − 9u ), if u ≤ ,
3⎪ 2
p(u ) = − 2u − 1 + 3u 2 − 3u + 1 = ⎨ (8)
⎪− 8u 3 + 15u 2 − 9u + 2, if u ≥ 1 .
⎪⎩ 2
which generates the function

∑ [− 2 x − 1 ]
n
3
p3 ( x) = i + 3 xi2 − 3 xi + 1 . (9)
i =1
Graphs functions (4), (6) and (8) are presented below (Fig. 1):
p p 0.4
p 1.0
0.4
0.3

0.2
0.2
0.5

-0.4 -0.2 0.2 0.4 0.6 0.8 1.0 1.2 1.4 0.1
u
-0.2
-0.2 -0.1 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 1.1 1.2
-0.4
-0.4 -0.2 0.2 0.4 0.6 0.8 1.0 1.2 1.4
u u
-0.1

-0.6 -0.2
-0.5

-0.8 -0.3

-1.0 -1.0 -0.4

2 3
p1 (u ) = (1 − u ) 1 − u + u u − (1 − 2u ) p 2 (u ) = 1 − 2u − 1 p 3 (u ) = − 2u − 1 + 3u 2 − 3u + 1
Fig. 1.The graphs of p1 (u ) , p1 (u ) and p3 (u ) .

Using the functions (5), (7) or (9), quadratic programming problem (1) can be representing in the
equivalent form:
f ( x) = x T Qx + c T x → min ⎫
⎪
subject to p1 ( x ) ≤ 0, ⎪
⎪
or p 2 ( x ) ≤ 0, ⎬ (10)
or p 3 ( x) ≤ 0, ⎪
⎪
and 0 ≤ x ≤ e. ⎪
⎭
We mentioned that p s ( x ) = 0 , s = 1, s = 2 or s = 3 for any 0 ≤ xi ≤ 1 , i = 1,2, K , n .

3 Reducing the Number of Constraints

The problem (10) includes 2n simple constraints:
xi ≥ 0, i = 1,2, K , n,⎫
⎬
xi ≤ 1, i = 1,2, K , n. ⎭
For sufficiently large n there are some major problems in solving the problem (10).
In what follows we will show how we can reduce these 2n simple restriction to only two convex
constraints.
The constraint u ≥ 0 is equivalent to inequation γ 1 (u ) ≤ 0 or with the inequation γ 2 (u ) ≤ 0 , and the
constraint u ≤ 1 is equivalent to γ 3 (u ) ≤ 0 or with γ 4 (u ) ≤ 0 where
⎧⎪2u 2 , if u ≤ 0,
γ 1 (u ) = u 2 − u u = ⎨
⎪⎩0, if u ≥ 0,
⎧- 2u , if u ≤ 0,
γ 2 (u ) = −u + u = ⎨
⎩0, if u ≥ 0,
⎧⎪0, if u ≤ 1,
γ 3 (u ) = (1 − u ) 2 − (1 − u ) 1 − u = ⎨
⎪⎩2(1 − u ) 2 , if u ≥ 1,
⎧0, if u ≤ 1,
γ 4 (u ) = −1 + u + 1 − u = ⎨
⎩2(−1 + u ), if u ≥ 1.

88
 Vasile Moraru, Sergiu Zaporojan

The functions γ 1 (u ), γ 2 (u ), γ 3 (u ) and γ 4 (u ) are convex functions for ∀u ∈ ℜ , for example, γ 1 (u ) and
γ 3 (u ) look like that (Fig. 2):
0.5
0.5

0.375
0.375

0.25
0.25

0.125
0.125

0
0
-0.5 -0.375 -0.25 -0.125 0 0.125 0.25
-0.25 0 0.25 0.5 0.75 1 1.25 1.5
u
u

γ 1 (u ) = u 2 − u u γ 3 (u ) = (1 − u ) 2 − (1 − u ) 1 − u
Fig. 2.The graphs of γ 1 (u ) and γ 2 (u )

As xi2 − xi xi ≥ 0 and − xi + xi ≥ 0 for ∀xi ∈ ℜ the restrictions xi ≥ 0, i = 1,2, K , n are equivalent to the
constraint ϕ1 ( x) ≤ 0 or ϕ 2 ( x) ≤ 0 , where

∑ [x ]
n
2
ϕ1 ( x) = i − xi xi ,
i =1
(11)

∑ [− x + x ] .
n
ϕ 2 ( x) = i i
i =1
Similarly it is determined that the conditions xi ≤ 1, i = 1,2,K, n are equivalent to the convex constraint
g1 ( x) ≤ 0 or g 2 ( x) ≤ 0 , where

∑ [(1 − x ) ]
n
2
g1 ( x ) = i − (1 − xi ) 1 − xi ,
i =1
(12)

∑ [− 1 + x ]
n
g 2 ( x) = i + 1 − xi .
i =1
The functions ϕ1 ( x) , ϕ 2 ( x) , g1 ( x) and g 2 ( x) are convex and determines convex set. For example, for
n = 2 the functions ϕ1 ( x) and g1 ( x) are as follows (see Fig. 3):

4
4

3
3

2
2

1
1

0.5 0.5 0
-1 -1 0 0.75 0.75
-0.75 -0.75 1 1
-0.5 -0.5 1.25 1.25
-0.25 -0.25 x2 1.5 1.75 1.5
x2 0 0.25 0.25
0
22
1.75
0.5
0.5
x1 x1

Fig. 3. The graphs of ϕ1 ( x) and ϕ1 ( x) g 1 ( x)

So the problem (10) which has (2n+1) constrains can be reduced to a problem of optimization
only with three constrains:
f ( x) = xT Qx + cT x → min ⎫
⎪
subject to p1 ( x) ≤ 0, ⎪
⎬ (13)
ϕ1 ( x) ≤ 0, ⎪
g1 ( x) ≤ 0, ⎪
⎭

89
 A continuous reformulation of the quadratic unconstrained binary optimization problem

or
f ( x) = xT Qx + cT x → min ⎫
⎪
subject to p2 ( x) ≤ 0, ⎪
⎬, (14)
ϕ 2 ( x) ≤ 0, ⎪
g 2 ( x) ≤ 0, ⎪
⎭
where the convex functions ϕ1 ( x) , ϕ 2 ( x) , g1 ( x) , g 2 ( x) are determined by the formulas (11) and (12),
while the concave functions p1 ( x) and p2 ( x) by formulas (5) and respectively (7).

4 Reduction to a DC problem
In problems (13) and (14) the constraints p1 ( x) ≤ 0 or p2 ( x) ≤ 0 are difficult as they represent non-
convex sets. One way to ease the solution of problems (13) and (14) is to penalize these
constraints. This can be carried out in such a way: Let us consider the optimization problem
obtained from the problem (13):
~
f ( x) = x T Qx + c T x − τ p1 ( x) → min ⎫
⎪
subject to ⎪
⎬ (15)
ϕ1 ( x) ≤ 0, ⎪
g1 ( x) ≤ 0, ⎪
⎭
where τ > 0 is a sufficiently large positive number (parameter penalty). If the matrix Q is positive
~
semidefinite then the function f ( x) is represented as the difference of two convex functions and thus
the problem (15) becomes a DC (Difference of Convex Functions) Programming [17], [18]. It could be
found a τ 0 so that for ∀τ ≥ τ 0 the problems (13) and (15) have the same optimal solutions. To solve
such problems (15) can be used with success DC Algorithm [18].

5 Conclusions
In this work we are interested in the resolution of quadratic optimization problem in variable
binary 0-1.The main idea is to transform QUBO problem into a new problem of optimization with
~
only two convex constraints. The objective function f (x) is reformulated a new function f ( x) ,
convex and equal to f (x) for any admissible solutions QUBO problem. The results obtained with
this approach are promising. On all convex functions that we considered p1 ( x) , ϕ1 ( x) and g1 ( x) to
have been very effective.

References
[1] L. A. Wolsey. Integer Programming, Wiley- Interscience, 1998.
[2] C.A. Floudas and V. Visweswaran. Quadratic Optimization. In “Handbook of Optimization”, R. Horst,
P. M. Pardalos (eds), Kluwer Academic Publishers, Dordrecht/ Boston/ Condon, 217-270,1995.
[3] A. Billionnet.Quadratic 0-1 Bibliography. http://cedric.cnam.fr/fichiers/RC611.pdf
[4] P.L. Hammer (Ivănescu), S. Rudeanu. Boolean Methods in Operation Research and Related Areas. Springer
Berlin Heidelberg. 2014. 331 p. (reprint of the 1st edition 1968).
[5] A. Boros, P. L. Hammer. Pseudo-Boolean Optimization. Discrete Applied Mathematics. Vol. 123, 155-225,
2002.
[6] M. Garey, D.Johnson. Computers and Intractibility: A Guide to the Theory of NP-Completeness. W.H.
Freeman & Company, 1979.
[7] C. Lemarechal and F. Qustry. Semidefinite Relaxation and Lagrangian Duality whith Application to
Combinatorial Optimization. Research Report N. 3710, INRIA Rhone-Alpes, France, 1999.
[8] H. D. Sherali and W. P. Adams. A Reformulation – Linearization Technique for Solving Discrete and
Continuous Nonconvex Problems. Kluwer Academic Publishers,, Norwell, MA, 1999.

90
 Vasile Moraru, Sergiu Zaporojan

[9] W. P. Adams, R. Forrester and F. Glover. Comparisons and Enhancement Strategies for Linearizing
Mixed 0-1 Quadratic Programms. Discrete Optimization. 1(2): 99-120, 2004.
[10] S. Poljak and H. Wolkowics. Convex Relaxations of (0,1) Quadratic Programming. Mathematics of
Operations Research, V. 20, 550-561.
[11] P. L. Hammer and A. A. Rubin. Some Remark on Quadratic Programming with 0-1 Variables.
RAIRO, 3:, 67-79, 1970.
[12] A. Billonnet and S. Elloumi. Using a Mixed Integer Quadratic Programming Solver for the
Unconstrained Quadratic 0-1 Problem. Mathematical Programming, Published online, 2006.
[13] M. W. Carter. The Indefinite Zero-One Quadratic Problem. Discrete Applied Mathematics, 7:,23-44,
1984.
[14] J. E. Beasley. Heuristics Algorithms for the Unconstrained Binary Quadratic Programming Problem. 1998.
http://mscmga.ms.ic.ac.uk/jeb/jeb.html.
[15] Le Thi Hoai An, Tao Pham Dinh, T. Huynh Van Ngai. Exact Penalty Techniques in DC Programming.
SIAM Conference on Optimization, 2005.
[16] Le Thi Hoai An, Tao Pham Dinh. A Continuous Approach for Globally Solving Mixed Integer
Programming. Sixth SIAM Conference on Optimization. Georgia, USA, Atlante, May 10-12, 1999.
[17] R. Host, N. V. Thoai. DC Programming: Overview. Journal of Optimization and Applications, 105 (1):
1-43, 1999. http://link.springer.com/article/10.1023%2FA%3A1021765131316#page-2
[18] Le Thi Hoai An, Tao Pham Dinh. The DC (Difference of Convex Functions) Programming and DCA
Revisited with DC Models of Real World Nonconvex Optimization Problems. Annals of Operations
Research, 133 (1), 23-46, 2005. http://link.springer.com/article/10.1007%2Fs10479-004-5022-1.

Vasile Moraru Sergiu Zaporojan

Technical University of Moldova Technical University of Moldova
Applied Informatics Department Computer Science Department
168, Stefan cel Mare str., Chisinau, 2004 168, Stefan cel Mare str., Chisinau, 2004
MOLDOVA Republic of MOLDOVA Republic of
E-mail: [email protected] E-mail: [email protected]

91