MINLP

ARGONNE NATIONAL LABORATORY
9700 South Cass Avenue

Argonne, Illinois 60439
Mixed-Integer Nonlinear Optimization
Pietro Belotti, Christian Kirches, Sven Leyffer, Jeff Linderoth, Jim Luedtke, and Ashutosh
Mahajan
Mathematics and Computer Science Division
Preprint ANL/MCS-P3060-1112
November 22, 2012
This work was supported by the Office of Advanced Scientific Computing Research, Office of Science, U.S. Department
of Energy, under Contract DE-AC02-06CH11357.
Contents
Table of Content ii
1 Mixed-Integer Nonlinear Programming Introduction 2

1.1 MINLP Notation and Basic Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Preview of Key Building Blocks of MINLP Algorithms . . . . . . . . . . . . . . . . . 4
1.3 Scope and Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2 Nonlinear Models with Integer Variables 10

2.1 Modeling Practices for MINLP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2 Design of Multiproduct Batch Plants . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.3 Design of Water Distribution Networks . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.4 A Dynamic Subway Operation Problem . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.5 Summary of MINLP Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3 Deterministic Methods for Convex MINLP 19

3.1 Nonlinear Branch-and-Bound . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.2 Multitree Methods for MINLP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.3 Single-Tree Methods for MINLP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.4 Presolve Techniques for MINLP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4 Cutting Planes for Convex MINLPs 41

4.1 Mixed-Integer Rounding Cuts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.2 Perspective Cuts for MINLP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.3 Disjunctive Cutting Planes for MINLP . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
4.4 Implementation of Disjunctive Cuts . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
4.5 Mixed-Integer Second-Order Cone Programs . . . . . . . . . . . . . . . . . . . . . . . 48
5 Nonconvex MINLP 54
5.1 Piecewise Linear Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
5.2 Generic Relaxation Strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
5.3 Spatial Branch-and-Bound . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
5.4 Relaxations of Structured Nonconvex Sets . . . . . . . . . . . . . . . . . . . . . . . . . 72
6 Heuristics for Solving MINLPs 78

6.1 Search Heuristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
6.2 Improvement Heuristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
7 Mixed-Integer Optimal Control Problems 84

7.1 Mixed-Integer Optimal Control Problem Class . . . . . . . . . . . . . . . . . . . . . . 85
7.2 Partial Outer Convexification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
ii
7.3 Relaxation of the Partially Convexified MIOCP . . . . . . . . . . . . . . . . . . . . . . 87
7.4 Constructing the Integer Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
7.5 Extensions to the Presented Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
8 Software for MINLP 91

8.1 Convex MINLP solvers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
8.2 Nonconvex MINLP solvers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
8.3 An MIOCP Solver . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
8.4 Modeling Languages and Online Resources . . . . . . . . . . . . . . . . . . . . . . . . 96
Acknowledgments 97
References 97
iii
Mixed-Integer Nonlinear Optimization
Pietro Belotti1 , Christian Kirches2 , Sven Leyffer3 , Jeff Linderoth4 , James Luedtke5 , and
Ashutosh Mahajan6
1
Department of Mathematical Sciences, Clemson University
2
Interdisciplinary Center for Scientific Computing, Heidelberg University, and
Mathematics and Computer Science Division, Argonne National Laboratory
3
Mathematics and Computer Science Division, Argonne National Laboratory
4
Department of Industrial and Systems Engineering, University of Wisconsin-Madison
5
Department of Industrial and Systems Engineering, University of Wisconsin-Madison
6
Industrial Engineering and Operations Research, Indian Institute of Technology Bombay
November 22, 2012
Abstract
Many optimal decision problems in scientific, engineering, and public sector applications
involve both discrete decisions and nonlinear system dynamics that affect the quality of the
final design or plan. These decision problems lead to mixed-integer nonlinear programming
(MINLP) problems that combine the combinatorial difficulty of optimizing over discrete vari-
able sets with the challenges of handling nonlinear functions. We review models and applica-
tions of MINLP, and survey the state of the art in methods for solving this challenging class of
problems.
Most solution methods for MINLP apply some form of tree-search. We distinguish two
broad classes of methods: single-tree and multitree methods. We discuss these two classes of
methods first in the case where the underlying problem functions are convex. Classical single-
tree methods include nonlinear branch-and-bound and branch-and-cut methods, while clas-
sical multitree methods include outer approximation and Benders decomposition. The most
efficient class of methods for convex MINLP are hybrid methods that combine the strengths of
both classes of classical techniques.
Nonconvex MINLPs pose additional challenges, because they contain nonconvex functions
in the objective or the constraints; hence even when the integer variables are relaxed to be con-
tinuous, the feasible region is generally nonconvex, resulting in many local minima. We discuss
a range of approaches to tackle this challenging class of problems, including piecewise linear
approximations, generic strategies for obtaining convex relaxations nonconvex functions, spa-
tial branch-and-bound methods, and a small sample of techniques that exploit particular types
of nonconvex structures to obtain improved convex relaxations.
Preprint ANL/MCS-3060-1112
1
2 Belotti, Kirches, Leyffer, Linderoth, Luedtke, & Mahajan
We finish our survey with a brief discussion of three important aspects of MINLP. First,
we review heuristic techniques that can obtain good feasible solution in situations where the
search-tree has grown too large or we require real-time solutions. Second, we describe an
emerging area of mixed-integer optimal control that adds systems of ordinary differential equa-
tions to MINLP. Third, we survey the state of the art in software for MINLP.
Keywords: Mixed-Integer Nonlinear Programming, Combinatorial Optimization, Noncon-
vex Programming, Global Optimization
AMS-MSC2010: 90C11, 90C26, 90C27, 90C30, 65K05
1 Mixed-Integer Nonlinear Programming Introduction

Many optimal decision problems in scientific, engineering, and public sector applications involve
both discrete decisions and nonlinear system dynamics that affect the quality of the final design
or plan. Mixed-integer nonlinear programming (MINLP) problems combine the combinatorial
difficulty of optimizing over discrete variable sets with the challenges of handling nonlinear func-
tions. MINLP is one of the most general modeling paradigms in optimization and includes both
nonlinear programming (NLP) and mixed-integer linear programming (MILP) as subproblems.
MINLPs are conveniently expressed as

minimize f (x),

x
subject to c(x) 0,

(1.1)

x X,

xi Z, i I,

where f : Rn R and c : Rn Rm are twice continuously differentiable functions, X Rn is a

bounded polyhedral set, and I {1, . . . , n} is the index set of integer variables. We note that we
can readily include maximization and more general constraints, such as equality constraints, or
lower and upper bounds l c(x) u. More general discrete constraints that are not integers can
be modeled by using so-called special-ordered sets of type I (Beale and Tomlin, 1970; Beale and
Forrest, 1976).
Problem (1.1) is an NP-hard combinatorial problem, because it includes MILP (Kannan and
Monma, 1978), and its solution typically requires searching enormous search trees; see Figure 1.
Worse, nonconvex integer optimization problems are in general undecidable (Jeroslow, 1973).
Jeroslow provides an example of a quadratically constrained integer program and shows that
no computing device exists that can compute the optimum for all problems in this class. In the
remainder of this paper, we concentrate on the case where (1.1) is decidable, which we can achieve
either by ensuring that X is compact or by assuming that the problem functions are convex.
1.1 MINLP Notation and Basic Definitions
Throughout this paper we use x(k) to indicate iterates of x and f (k) = f (x(k) ) to denote the evalua-
tion of the objective at x(k) . Similar conventions apply to constraints, gradients, or Hessian at x(k) ;
Mixed-Integer Nonlinear Optimization 3
Figure 1: Branch-and-bound tree without presolve after 360 s CPU time has more than 10,000
nodes.
for example, f (k) = f (x(k) ). We use subscripts to denote components; for example, xi is com-
ponent i of x. For a set J {1, . . . , n} we let xJ denote the components of x corresponding to J.
In particular, xI are the integer variables. We also define C = {1, . . . , n} I and let xC denote the
continuous variables. We denote by p the dimension of the integer space, p = |I|. We denote the
floor and ceiling operator by bxi c and dxi e, which denote the largest integer smaller than or equal
to xi and the smallest integer larger than or equal to xi , respectively. Given two n n matrices Q
and X, Q X = ni=1 nj=1 Qij Xij represents their inner product.
P P
In general, the presence of integer variables xi Z implies that the feasible set of (1.1) is not
convex. In a slight abuse of terminology, we distinguish convex from nonconvex MINLPs.
Definition 1.1. We say that (1.1) is a convex MINLP if the problem functions f (x) and c(x) are convex
functions. If either f (x) or any ci (x) is a nonconvex function, then we say that (1.1) is a nonconvex
MINLP.
Throughout this paper, we use the notion of a convex hull of a set S.
Definition 1.2. Given a set S, the convex hull of S is denoted by conv(S) and defined as
n o
conv(S) := x : x = x(1) + (1 )x(0) , 0 1, x(0) , x(1) S .
If X = {x Zp : l x u} and l Zp , u Zp , then conv(X) = [l, u] is simply the hyper-

cube. In general, however, even when X itself is polyhedral, it is not easy to find conv(X). The
convex hull plays an important role in mixed-integer linear programming: because an LP obtains
a solution at a vertex, we can solve an MILP by solving an LP over its convex hull. Unfortunately,
finding the convex hull of an MILP is just as hard as solving the MILP.
The same result does not hold for MINLP, as the following example illustrates:
n
X
minimize (xi 21 )2 , subject to xi {0, 1}.
x
i=1
The solution of the contiuous relaxation is x = ( 12 . . . 12 ), which is not an extreme point of the
feasible set and, in fact, lies in the strict interior of the MINLP; see Figure 2. Because the continuous
minimizer lies in the interior of the convex hull of the integer feasible set, it cannot be separated
from the feasible set. However, we can reformulate (1.1) by introducing an objective variable
and a constraint f (x) . We obtain the following equivalent MINLP:

minimize ,
,x

subject to f (x) ,

c(x) 0, (1.2)

x X,

xi Z, i I.
The optimal solution of (1.2) always lies on the boundary of the convex hull of the feasible set and
therefore allows us to use cutting-plane techniques.
x2
(
x1 , x
2 )
x1
Figure 2: Small MINLP to illustrate the need for a linear objective function.
1.2 Preview of Key Building Blocks of MINLP Algorithms
A wide variety of methods exists for solving MINLP. Here, we briefly introduce the two funda-
mental concepts underlying these algorithms: relaxation and constraint enforcement. A relaxation
is used to compute a lower bound on the optimal solution of (1.1). A relaxation is obtained by
enlarging the feasible set of the MINLP, for example, by ignoring some constraints of the problem.
Typically, we are interested in relaxations that are substantially easier to solve than the MINLP
itself. Together with upper bounds, which can be obtained from any feasible point, relaxations
allow us to terminate the search for a solution whenever the lower bound is larger than the cur-
rent upper bound. Constraint enforcement refers to procedures used to exclude solutions that are
feasible to the relaxation but not to the original MINLP. Constraint enforcement may be accom-
plished by refining or tightening the relaxation, often by adding valid inequalities, or by branching,
where the relaxation is divided into two or more separate problems.
In general, upper bounds are obtained from any feasible point. Often, we fix the integer vari-
ables at an integral value and solve the resulting NLP to obtain an upper bound (which we set to
infinity if the NLP is infeasible).
Relaxations. Formally, an optimization problem min{R (x) : x SR } is a relaxation of a problem

min{(x) : x S} if (i) SR S and (ii) R (x) (x) for each x S. The feasible set R of
a relaxation of a problem with feasible set F contains all feasible points of F. The main role of
the relaxation is to provide a problem that is easier to solve and for which we can obtain globally
optimal solutions that allow us to derive a lower bound. Relaxations that fall into this category are
convex NLPs, for which nonlinear optimization solvers will converge to the global minimum, and
MILPs, which can often be solved efficiently (for practical purposes) by using a branch-and-cut
approach.
Figure 3: Illustration of the two classes of relaxation. The left image shows the mixed-integer
feasible set, the top right image shows the nonlinear relaxation, and the bottom right shows the
polyhedral relaxation.
Several strategies are used to obtain relaxations of MINLPs.
1. Relaxing integrality. Integrality constraints xi Z can be relaxed to xi R for all i I. This

procedure yields a nonlinear relaxation of MINLP. This type of relaxation is used in branch-
and-bound algorithms (Section 3.1) and is given by

minimize f (x),

x
subject to c(x) 0, (1.3)

x X,

2. Relaxing convex constraints. Constraints c(x) 0 and f (x) z containing convex functions c
and f can be relaxed with a set of supporting hyperplanes obtained from first-order Taylor
series approximation,
T
z f (k) + f (k) (x x(k) ), (1.4)
(k)T
0 c(k) + c (x x(k) ), (1.5)
for a set of points x(k) , k = 1, . . . , K. When c and f are convex, any collection of such
hyperplanes forms a polyhedral relaxation of these constraints. This class of relaxations is
used in the outer approximation methods discussed in Section 3.2.1.
3. Relaxing nonconvex constraints. Constraints c(x) 0 and f (x) z containing nonconvex

functions require more work to be relaxed. One approach is to derive convex underestima-
tors, f(x) and c(x), which are convex functions that satisfy
f(x) f (x) and c(x) c(x), x conv(X). (1.6)
Then the constraints z f (x) and 0 c(x) are relaxed by replacing them with the con-
straints
z f(x) and 0 c(x).
In Section 5 we review classes of nonlinear functions for which convex underestimators are
known, and we describe a general procedure to derive underestimators for more complex
nonlinear functions.
All these relaxations enlarge the feasible set of (1.2), and they can be combined with each other.
For example, a convex underestimator of a nonconvex function can be further relaxed by using
supporting hyperplanes, yielding a polyhedral relaxation.
Figure 3 illustrates the relaxation of integrality constraints and convex nonlinear constraints.
The left image shows the mixed-integer feasible set (the union of the red lines), the top right image
shows the nonlinear relaxation obtained by relaxing the integrality constraints (the shaded area is
the NLP feasible set), and the bottom right figure shows a polyhedral relaxation (the union of the
red lines) as well as its LP relaxation (the shaded area). We note that an infinite number of possible
polyhedral relaxations exists, depending on the choice of the points x(k) conv(X), k = 1, . . . , K.
If the solution to a relaxation is feasible in (1.2), then it also solves the MINLP. In general,
however, the solution is not feasible in (1.2), and we must somehow exclude this solution from the
relaxation.
Constraint enforcement. Given a point x that is feasible to a relaxation but is not feasible to
the MINLP, the goal of constraint enforcement is to exclude this solution, so that the algorithm
can eventually converge to a solution that satisfies all the constraints. Two broad classes of con-
straint enforcement strategies exist: relaxation refinement and branching. Most modern MINLP
algorithms use both classes.
The goal of relaxation refinement is to tighten the relaxation in such a way that an infeasible
relaxation solution x is no longer feasible. Most commonly, this is achieved by adding a new
Figure 4: Separation of infeasible point (black dot) by adding a separating hyperplane. The green
line on the right shows a separating hyperplane.
valid inequality to the relaxation. A valid inequality is an inequality that is satisfied by all feasible
solutions to the MINLP. When a valid inequality successfully excludes a given infeasible solution,
it is often called a cut. Valid inequalities are usually linear but may be convex. For example, after
relaxing a convex constraint with a polyhedral relaxation, a valid inequality can be obtained by
linearizing the nonlinear functions about x . This valid inequality will successfully cut off x
, unless
x
satisfies the nonlinear constraints, c( x) 0; see Figure 4. This class of separation is used in
outer approximation, Benders decomposition, and the ECP algorithm discussed in Section 3.2.1.
Valid inequalities can also be useful after relaxing integrality. In this case, the goal is to obtain
an inequality that is valid because it does not cut off any integer feasible solution but will cut off
an integer infeasible solution x . This technique has been critical to the success of algorithms for
solving mixed-integer linear programs.
The second class of constraint enforcement strategy is branching: dividing the feasible region
into subsets such that every solution to MINLP is feasible in one of the subsets. When integrality
is relaxed, it can be enforced by branching on an integer variable that takes a fractional value x i for
some i I. Branching creates two new separate relaxations: the constraint xi b xi c is added to
the first relaxation, and the constraint xi d
xi e is added to the second relaxation; see Figure 5. All
solutions of the MINLP now lie in one of these two new relaxations. The resulting subproblems
are managed in a search tree that keeps track of all subproblems that remain to be solved. This
approach is the basis of the branch-and-bound algorithms described in detail in Section 3.1.
Constraint enforcement for relaxed nonconvex constraints involves a combination of branch-
ing and relaxation refinement. These techniques are discussed in detail in Section 5, but here we
outline the general idea using Figure 6. Following the solution of the relaxation (given by the
green objective on the left), we branch on a continuous variable and hence split its domain into two
subdomains. We then compute new underestimators for use in (1.6) that are valid on each of the
two subdomains (i.e., we refine the relaxation). In the example, these refined underestimators are
indicated by the two green objective functions on the right. This approach, which we refer to as
spatial branching, results in a branch-and-bound algorithm similar to the one for discrete variables.
We continue to divide the domain into smaller subdomains until the lower bound on a subdomain
Figure 5: Branching on the values of an integer variable creates two new nonlinear subproblems
that both exclude the infeasible point, denoted with the black dot.
is larger than the upper bound, at which point we can exclude this domain from our search. For
MINLPs having both integer variables and nonconvex constraints, branching may be required on
both integer and continuous decision variables.
1.3 Scope and Outline
The past 20 years or so have seen a dramatic increase in new mixed-integer nonlinear models
and applications, which has motivated the development of a broad range of new techniques to
tackle this challenging class of problems. This survey presents a broad overview of deterministic
methodologies for solving mixed-integer nonlinear programs. In Section 2 we motivate our interest in
MINLP methods by presenting some small examples, and we briefly discuss good modeling prac-
tices. In Section 3 we present deterministic methods for convex MINLPs, including branch and
bound, outer approximation, and hybrid techniques. We also discuss advanced implementation
considerations for these methods. Cutting planes have long played a fundamental role in mixed-
integer linear programming, and in Section 4 we discuss their extension to MINLP. We review a
range of cutting planes such as conic MIR cuts, disjunctive cuts, and perspective cuts. In Section 5
we outline methods for solving nonconvex MINLPs. A range of heuristics to obtain good incum-
bent solutions quickly is discussed in Section 6. We review two classes of deterministic heuristics:
search and improvement heuristics. In Section 7 we an emerging extension of MINLP, the mixed-
Figure 6: Constraint enforcement by using spatial branching for global optimization.
integer optimal control problem. In Section 8 we review the state of the art in software for MINLP
and categorize the different solvers within the context of the previous sections.
Given the wide applicability of MINLP and the ensuing explosion of numerical methods, it
would be prohibitive to discuss all methods. Instead, we focus this survey on deterministic meth-
ods that tackle both integer variables and nonlinear constraints. In particular, we do not survey
the two main subproblems of MINLP, mixed-integer linear programming and nonlinear program-
ming, for which there exist excellent textbooks (Nocedal and Wright, 1999; Fletcher, 1987; Wolsey,
1998; Nemhauser and Wolsey, 1988a) and recent survey articles (Gould and Leyffer, 2003). We
also do not cover three other related topics:
1. Algorithms that are polynomial when the number of variables is fixed, or which require a polyno-
mial number of calls to an oracle. Lenstras algorithm (H.W. Lenstra, 1983) solves integer lin-
ear programming problems in polynomial time when the number of variables in the prob-
lem is fixed. Khachiyan and Porkolab (2000) extended this algorithm to integer programs
with convex polynomial objective and constraints. Generalizations and improvements in
the complexity bound have been made in (Heinz, 2005; De Loera et al., 2006; Hildebrand

and Koppe, 2010; Dadush et al., 2011d). Using ideas from Graver bases, Hemmecke et al.
(2011) derive an algorithm that requires a polynomial number of augmentation steps to solve
specially structured convex minimization problems over integer sets. Baes et al. (2012) inves-
tigate the problem of minimizing a strongly convex Lipschitz continuous function over a set
of integer points in a polytope, obtaining an algorithm that provides a solution with a con-
stant approximation factor of the best solution by solving a polynomial number of specially
structured quadratic integer programs.
2. Properties of closures of convex integer sets. Recently, Dey and Vielma (2010) and Dadush et al.
(2011a,b,c) have been studying closures of certain classes of valid inequalities for convex
MINLPs. A closure is the set obtained after including all inequalities from a particular class.
In particular, for the class of Gomory-Chvatal (G-C) inequalities, inequalities which can be
obtained from a simple rounding argument, they have shown that the resulting closure is a
rational polyhedron if the original convex set is either strictly convex or compact. This is an
interesting result for two reasons: the original convex set may not be described by finitely
many linear inequalities, and also the number of G-C inequalities is not finite. In work in
the same spirit, Dey and Moran R. (2012) study conditions under which the convex hull of a
set of integer points in a convex set is closed, and polyhedral.
3. Mixed-integer derivative-free optimization problems. Such problems arise when the problem
functions are not given explicitly and can be evaluated only as the result of a (black-box)
simulation S(x). In this case, derivatives are typically not available, and the simulation is
often computationally expensive. One example of this class of problem is the design of land-
fill soil liners, where integer variables model the choice of material and the permeative and
decomposition processes are modeled through simulation (Bartelt-Hunt et al., 2006). Other
applications include the design of nanophotonic devices (Maria et al., 2009; Miller et al.,
2010) and the determination of loop unroll factors in empirical performance tuning (Bal-
aprakash et al., 2011). Algorithms for derivative-free mixed-integer nonlinear optimization
employ either global surrogates (Muller,
2012; Muller et al., 2012; Davis and Ierapetritou,
2009; Hemker, 2008; Hemker et al., 2008; Rashid et al., 2012), or mesh- and pattern-based
techniques (Liuzzi et al., 2011; Audet and Dennis, Jr., 2000; Fowler et al., 2008; Abramson
et al., 2009).
We mention a number of surveys and monographs on MINLP, including a review of MINLP and
disjunctive programming (Grossmann, 2002), a survey of MINLP algorithms (Grossmann and
Kravanja, 1997), a survey of algorithms and software for MINLP (Bonami et al., 2012), a compre-
hensive textbook on MINLP by Floudas (1995), and a collection of articles related to a recent IMA
workshop on MINLP (Lee and Leyffer, 2011).
2 Nonlinear Models with Integer Variables
In this section, we review a small number of models and modeling tricks to motivate the algo-
rithmic developments that are discussed in the subsequent sections. The models are chosen to
provide insight into the interactions of the two main modeling paradigms: integer variables and
nonlinear equations. We start by presenting a well-known application from chemical engineering
that models the design of a multiproduct batch plant (Kocis and Grossmann, 1988; Sargent, 1979),
which can be formulated as a convex MINLP. We then present a nonconvex MINLP that arises in
the design of water distribution networks (Eiger et al., 1994; Sherali and Smith, 1997; Sherali et al.,
2001; Bragalli et al., 2006, 2012). Finally, we present an time- and energy-optimal subway control
example (Bock and Longman, 1985) that adds time-dependent integer variables and constraints to
MINLP.
2.1 Modeling Practices for MINLP
Modeling plays a fundamental role in MILP (see the textbook by Williams (1999)) and is arguably
more critical in MINLP, because of the additional nonlinear relationships. The nonlinearities often
allow for equivalent formulations with more favorable properties. For example, in Section 2.2 we
present a model of a multiproduct batch plant and show that by using a nonlinear transformation,
we are able to reformulate the nonconvex problem as an equivalent convex problem, which is
typically easier to solve. Here, we briefly review a number of other modeling tricks that can be
useful in formulating easier-to-solve problems.
Convexification of binary quadratic programs. We can always make the quadratic form in a pure
binary quadratic program convex, because for any xi {0, 1} it follows that x2i = xi . For
x {0, 1}n we consider the quadratic form
q(x) = xT Qx + g T x,
and let be the smallest eigenvalue of Q. If 0, then q(x) is convex. Otherwise, we define
a new Hessian matrix W := Q I, where I is the identity matrix, and a new gradient
c := g + e, where e = (1, . . . , 1). It follows that
q(x) = xT Qx + g T x = xT W x + cT ,
where the second quadratic is now convex.
Exploiting low-rank Hessians. In many applications such as model structure determination and
parameter estimation problems (Skrifvars et al., 1998), the Hessian matrix is a large, dense,
and low-rank matrix. In this application, the Hessian matrix can be written as W = Z T R1 Z,
where R Rmm and Z Rmn , where m n and Z is sparse. Forming the Hessian matrix
and operating with it can be prohibitive. Some nonlinear solvers only require matrix-vector
products of the Hessian and can readily exploit this situation. An alternative is to introduce
additional variables z and constraints z = Zx, and then rewrite xT W x = z T R1 z, which
allows the solver to exploit the sparsity of the constraint matrix Z.
Linearization of constraints. A simple transformation is to rewrite x1 /x2 = a as x1 = ax2 , where

a is a constant. special-ordered sets provide a systematic way to formulate nonlinear expres-
sions as piecewise linear functions (see Section 5.1), and these can be effective for expressions
that involve a small number of variables.
Linearization of x1 x2 , for x2 {0, 1}. We can linearize the expression x1 x2 , when 0 x1 u and
x2 {0, 1} by observing that the product x1 x2 is either equal to zero (if x2 = 0) or equal to
x1 (if x2 = 1). By introducing the new variable x12 and adding the constraints,
0 x12 x2 u and u(1 x2 ) x1 x12 u(1 x2 ),
we can replace the nonconvex term x1 x2 by x12 . This trick readily generalizes to situations
where l x1 u.
Avoiding undefined nonlinear expressions. In many practical instances, the MINLP solvers fail,
because the nonlinear solver generates an iterate at which the nonlinear functions cannot be
evaluated. An example is the constraint,
c(x1 ) = ln(sin(x1 )) 0,
which cannot be evaluated whenever sin(x1 ) 0. Because NLP solvers typically remain
feasible with respect to simple bounds, we can reformulate this constraint equivalently as
c(x2 ) = ln(x2 ) 0, x2 = sin(x1 ), and x2 0.
Interior-point methods will never evaluate the constraint c(x2 ) at a point x2 0, and even
active-set methods can be expected to avoid this region, because the constraint violation be-
comes unbounded near x2 = 0. An additional benefit of this reformulation is that it reduced
the degree of nonlinearity of the resulting constraint set.
A common theme of these reformulation is that convex formulations are typically preferred over
nonconvex ones, and linear formulations over nonlinear formulations. More useful modeling
tricks for integer and binary variables are given in the classic textbook by Williams (1999).
2.2 Design of Multiproduct Batch Plants
Multiproduct batch plants produce a number of different products on parallel lines in a multistage
batch process. In the following, we use subscripts j to index the stages, and subscripts i to index
the products. A multiproduct batch plant is characterized by the following set of (fixed) parame-
ters:
M : the set of batch processing stages;
N : the set of different products;
H: the time horizon;
Qi : the required quantity of product i N ;
tij : Processing time product i stage j M ; and
Sij : Size Factor product i N stage j M .
Our goal is to find the minimum cost design by choosing optimal values for the design (free)
variables of the batch process, which are given by:
Bi : the batch size of product i N ;
Vj : the size of stage j M , which must satisfy Vj Sij Bi i, j;
Nj : the number of machines at stage j M ; and
Ci : the longest stage time for product i N , which satisfies Ci tij /Nj i, j.
Given a set of cost parameters, j , j > 0, we can formulate the minimum cost design of a
multiproduct batch plant as a nonconvex MINLP:

X
minimize j Nj Vj j ,

V,C,B,N

jM
subject to Vj Sij Bi 0 i N, j M,

Ci Nj tij i N, j M,

X Qi

Ci H,

Bi
iN

Vj [Vl , Vu ], Ci [Cl , Cu ], Bi [Bl , Bu ], i N, j M,

Nj {1, 2, . . . , Nu } j M.

Unfortunately, this model is a nonconvex MINLP, because the objective function, the horizon-time
constraint, and the constraint defining the longest stage time are nonconvex functions. Fortu-
nately, we can apply a variable transformation which convexifies the problem. We introduce new
log-transformed variables vj , nj , bi , and ci defined by
vj = ln(Vj ), nj = ln(Nj ), bi = ln(Bi ), ci = ln Ci .
This transformation provides an equivalent convex reformulation of the multiproduct batch plant
design problem:
X

minimize
v,c,b,n j enj +j vj ,
jM

subject to vj ln(Sij )bi 0, i N, j M,

ci + nj ln(tij ) i N, j M,

X
c b

Qi e i i
H,

iN

vj [vl , vu ], ci [cl , cu ], bi [bl , bu ], i N, j M,

nj {0, ln(2), . . . , ln(Nu )} j M,
where vl , vu , cl , cu , bl , and bu are suitably transformed bounds on the variables. The last constraint,
nj {0, ln(2), . . . , ln(Nu )} is difficult to enforce directly, but it can be modeled using a special
ordered set of type I (SOS-1) (Beale and Tomlin, 1970). By introducing binary variables ykj {0, 1}
we can replace the last constraint in the convex model by,
K
X K
X
ln(k)ykj = nj and ykj = 1, j M. (2.1)
k=1 k=1
The resulting model is available as batch on MacMINLP (Leyffer, 2003). Similar reformulations

have been proposed using x-transforms for models involving bilinear terms (Harjunkoski et al.,
1998).
2.3 Design of Water Distribution Networks
Many MINLPs model flows on networks. Examples include the optimal design of water net-
works (Burgschweiger et al., 2008; Bragalli et al., 2006, 2012) the design of gas-networks (Wolf
and Smeers, 2000) and to the optimal regulation of ventilation systems for mines (Wu et al., 1991;
Maonan and Wenjun, 1991). All applications share the same structural flow constraints. Here, we
concentrate on the design of water networks. The model is defined by a set on nodes, i cN , and
a set of arcs (i, j) A that define the (fixed) topology of the network. The goal is to determine
the pipe diameter for all arcs that minimizes the installation cost, meets the demand at all nodes,
and satisfies the hydraulic (flow) constraints along the arcs. The pipe diameters must be chosen
from a discrete set of available diameters, giving rise to integrality constraints, while the hydraulic
constraints involve nonlinear relationships, giving rise to a nonconvex MINLP.
We simplify the model by assuming that certain constants are uniform throughout the net-
work. The sets and (fixed) parameters that define the model are:
N: the set of nodes in the network;
S: the set of source nodes in the network, S N ;
A: the set of arcs in the network, A N N ;
Lij : the length of pipe (i, j) A;
Kij : a physical constant that models the roughness of pipe (i, j) A;
Di : the demand at node i N ;
Vmax : the maximum velocity of flow in all pipes;
Qmax : the maximum magnitude of flow in all pipes;
Pk : pipe diameters available to network, k = 1, . . . , r;
Hs : the hydraulic head at source node s S; and
Hl , Hu : are lower and upper bounds on the hydraulic head.
The variables of the model are:
qij : the flow in pipe (i, j) A (from node i to node j);
dij : the diameter of pipe (i, j) A, where dij {P1 , . . . , Pr };
hi : hydraulic head at node i N , where hs = Hs , s S, and Hl hi Hu , i N S;
zij : binary variables that model the direction of flow in pipe (i, j) A;
aij : the area of the cross section of pipe (i, j) A; and
yijk : a set of SOS-1 variables, see (2.1) that models the pipe-type on arc (i, j) A.
The conservation of flow at every node gives rise to a linear constraint
X X
qij qji = Di , i N S.
(i,j)A (j,i)A
Because our model contains cross section variables, aij = d2ij /4, we can model the bounds on the
flow along an arc depend as a linear set of constraints,
Vmax aij qij Vmax aij , (i, j) A.
The choice of pipe types, dij {P1 , . . . , Pr }, is modeled using the SOS-1 variables yijk , giving rise
to the set of constraints,
r
X r
X
yijk {0, 1}, k = 1, . . . , r, and yijk = 1, and Pk yijk = dij . (i, j) A,
k=1 k=1
The MINLP solvers can now ether branch on individual yijk or on SOS-1 sets (yij1 , . . . , yijr ), which
is generally more efficient. See Williams (1999) for a discussion on branching on special-ordered
sets. We can use the same SOS-1 set to linearize the nonlinear equation, aij = d2ij /4, by adding
the constraints,
Xr
(Pk /4)yijk = aij , (i, j) A.
k=1
The final set of constraints is an empirical model of the pressure loss along the arc (i, j) A,
c1
sgn(qij )|qij |c1 c2 Lij Kij
hi hj = , (i, j) A, (2.2)
dcij3
where c1 = 1.852, c2 = 10.7, and c3 = 4.87 are constants that depend on the medium of the
fluid. This last equation appears to be nonsmooth because it contains terms of the form, h(x) =
sgn(x)|x|c1 . However, it is easy to verify that h(x) is continuously differentiable, because c1 > 1.
To model h(x), we split the flow into its positive and negative part by adding binary variables
zij {0, 1} and the following set of constraints,
+ +
0 qij Qmax zij , 0 qij Qmax (1 zij ), qij = qij qij .
An alternative formulation based on complementarity constraints avoids the introduction of the

+
binary variables zij , and instead models the disjunction as 0 qij qij 0, where means that
+
either qij > 0 or qij > 0. With these new flow variables, we can now rewrite (2.2) equivalently as
h c1 c1 i
+ c1
qij qij c2 Lij Kij
hi hj = , (i, j) A.
dcij3
Finally, we can lower the degree of nonlinearity by substituting dij in this last constraint by
p
dij = 4aij /. This substitution generally provides better Taylor-series approximations than
the more nonlinear version involving dij , which ensures better practical convergence behavior
of the underlying NLP solvers.
2.4 A Dynamic Subway Operation Problem
As an example for a dynamic nonlinear mixed-integer problem, we present a control problem that
goes back to work by Bock and Longman (1985) optimizing the subway of the city of New York.
The problem has been treated by, for example, Sager (2005). We are interested in minimizing the
total energy consumption
Z T
minimize L(x(t), v(t)) dt (2.3)
x(),u(),v(),T 0
of a subway train, subject to a system of ordinary differential equations that models the dynamic
behavior of a subway trains state x(t) comprising position xs (t) and velocity xv (t),
x s (t) = xv (t) t [0, T ], (2.4)

x v (t) = fv (x(t), u(t), v(t)) t [0, T ], (2.5)
on a horizon [0, T ] R with free end time T . The continuous control 0 u(t) umax indicates
the braking deceleration. The integer control vector v(t) {1, 2, 3, 4} indicates the operation of
the subways two electric engines in one of four discrete modes that affect the subway trains
acceleration and energy consumption, namely
1. Serial, v(t) = 1. The energy consumption rate is given by

e p1 if xv (t) v1 ,

e p2 if v1 < xv (t) v2 ,

L(x(t), 1) = (2.6)
5
ci (1) (0.1xv (t))i if v2 < xv (t),
P
e

i=0
and the subway trains dynamics are described by

g e a1
if xv (t) v1 ,
Weff fv (t) = g e a2 if v1 < xv (t) v2 , (2.7)

g (e F (xv (t), 1) R(xv (t))) if v2 < xv (t).

2. Parallel, v(t) = 2. The energy consumption rate is given by

0 if xv (t) v2 ,

e p3 if v2 < xv (t) v3 ,

L(x(t), 2) = (2.8)
5
ci (2) (0.1xv (t) 1)i if v3 < xv (t),
P
e

i=0
and the subway trains dynamics are described by

0
if xv (t) v2 ,
Weff fv (t) = g e a3 if v2 < xv (t) v3 , (2.9)

g (e F (xv (t), 2) R(xv (t))) if v3 < xv (t).

3. Coasting, v(t) = 3. The energy consumption rate is zero, L(x(t), 3) = 0, and the subway
trains dynamics are described by
Weff fv (t) = g R(xv (t)) C Weff (2.10)
4. Braking, v(t) = 4. The energy consumption rate is zero, L(x(t), 4) = 0, and the subway
trains dynamics are described by
fv (t) = u(t). (2.11)
The forces occurring in these dynamics are given by

1.3
R(xv (t)) = ca 2 xv (t)2 + bW xv (t) + W + 116, (2.12)
2000
5
bi (1) (0.1xv (t) 0.3)i ,
X
F (xv , 1) = (2.13)
i=0
5
bi (2) (0.1xv (t) 1)i .
X
F (xv , 2) = (2.14)
i=0
In addition, the system shall satisfy certain path and point constraints as follows. Initial and
terminal states for the system trajectory are constrained to
x(0) = (0, 0)T , x(T ) = (S, 0)T . (2.15)
A maximum on the allowable driving time to complete the distance S is imposed,
T Tmax . (2.16)
Different scenarios can be defined for this problem by prescribing values for the parameters S
and W . In addition, scenarios may include speed limits at certain points or on certain parts of the
track. A description of several scenarios, units and numerical values for the model parameters a,
a1 , a2 , a3 , b, bi (v), C, c, ci (v), e, g, , p1 , p2 , p3 , S, Tmax , umax , v1 , v2 , v3 , W , Weff , and analytical inves-
tigations along with numerical solutions can be found in, for example, Bock and Longman (1985)
or Sager (2005). This problem shows three challenging features that are typical for real-world dy-
namic mixed-integer problems: integer variables in time, system state dependent switches that
need to be modeled appropriately using, for example, additional integer variables, and higher-
order polynomial approximations to nonlinear and possibly nonconvex system characteristics.
2.5 Summary of MINLP Applications
In addition to the models discussed above, MINLPs arise in a broad range of engineering and sci-
entific applications, including chemical, civil, electrical, nuclear, and communication engineering,
as well as emerging scientific applications in the design of nanophotonic materials and devices.
Electrical engineering applications of MINLP include the efficient management of electricity
transmission (Bacher, 1997; Momoh et al., 1997), transmission expansion (Romero et al., 2002;
Garver, 1997), transmission switching (Bartholomew et al., 2008; Hedman et al., 2008), and con-
tingency analysis, and blackout prevention of electric power systems (Bienstock and Mattia, 2007;
Donde et al., 2005).
MINLP also arise in many chemical engineering applications, such as the design of water
(Bragalli et al., 2006; Karuppiah and Grossmann, 2006) and gas (Martin et al., 2006) distribution
networks, the minimization of the environmental impact of utility plants (Eliceche et al., 2007),
the integrated design and control of chemical processes (Flores-Tlacuahuac and Biegler, 2007),
and block layout design in the manufacturing and service sectors (Castillo et al., 2005), in the
optimization of polymerization plants (Prata et al., 2008). A related area is systems biology, where
topics such as circadian rythms (Shaik et al., 2008), and protein folding (Klepeis and Floudas, 2003)
are addressed using mixed-integer optimization.
Applications in communication engineering and computer science include the optimal re-
sponse to a cyber attack (Goldberg et al., 2012; Altunay et al., 2011), wireless bandwidth allo-
cation (Bhatia et al., 2006; Sheikh and Ghafoor, 2010; Costa-Montenegro et al., 2007), selective
filtering (Sinha et al., 2002; Soleimanipour et al., 2002), optical network performance optimiza-
tion (Elwalid et al., 2006), network design with queuing delay constraints (Boorstyn and Frank,
1977), and network design topology (Bertsekas and Gallager, 1987; Chi et al., 2008), multi-vehicle
swarm communication network optimization (Abichandani et al., 2008), the design of optimal
paths (i.e. minimum time) for robotic arms (Gentilini et al., 2012), the synthesis of periodic wave-
forms by tripolar pulse codes (Callegari et al., 2010), and the solution of MILP under uncertainty
of the parameters through robust optimization (Ben-Tal and Nemirovski, 1995).
Other engineering applications include the operational reloading of nuclear reactors (Quist
et al., 1998), the optimal response to catastrophic oil spills such as the recent Deepwater oil spill
in the Gulf of Mexico (You and Leyffer, 2010, 2011), the design of load-bearing thermal insu-
lation system for the Large Hadron Collider (Abramson, 2004; Abhishek et al., 2010), concrete
structure design (Guerra et al., 2011), and the design and operation of drinking water networks
(Burgschweiger et al., 2008), gas networks (Martin et al., 2006), electric distribution networks
(Lakhera et al., 2011), and mining networks (Pruitt et al., 2012). Applications in traffic model-

ing and optimization are found in (Fugenschuh et al., 2006), and (Soler et al., 2011) considers a
flight path optimization problem. An important financial application of MINLP arises in portfolio
optimization (Bienstock, 1996; Jobst et al., 2001).
MINLP models can also be stochastic service system design problems (Elhedhli, 2006), since
performance metrics are often nonlinear in the decision variables.
Another developing applications area of both game theory and optimization is resource allo-
cation for homeland security (see, e.g., Bier (2005); Sandler and Arce M (2003); Zhuang and Bier
(2007b)). Simple versions of these models (involving, for example, a single target and attacker)
have closed-form optimal (equilibrium) solutions (e.g., Powell (2007); Sandler and Siqueira (2006);
Zhuang and Bier (2007a)). In more realistic models, however, the best defensive strategy must be
computed. Such models often have important binary choices, including which targets to assign
a high priority in defending (Bier et al., 2005, 2007) and which strategies to employ in defending
a target (e.g., whether to attempt to deceive attackers about the level of resources invested to de-
fend a given target (Zhuang and Bier, 2007c; Zhuang, 2008)). Moreover, the utility functions of
the attackers and defenders in these models are highly nonlinear. Powerful MINLP solvers are
expected to provide decision support in this area.
An emerging area with challenging MINLP tasks is human decision analysis and support in
complex problem solving (Engelhart et al., 2012).
A special domain of MINLP are dynamic problems constrained by ordinary differential equa-
tions (ODEs) or differential-algebraic equations (DAE), often called mixed-integer optimal control
problems (MIOCPs). One of the earliest practical problems was the optimal switched operation of
the New York subway (Bock and Longman, 1985). Recent applications include gear shifts in auto-
motive control (Gerdts, 2005; Kirches et al., 2010), automated cruise controllers (Terwen et al., 2004;
et al., 2009; Kirches, 2011), superstructure detection in simulated moving bed processes
Hellstrom
(Sager et al., 2007), the optimization of batch distillation processes (Oldenburg et al., 2003).
3 Deterministic Methods for Convex MINLP

In general, we resolve the integrality constraints using some form of tree-search strategy. MINLPs
pose the additional challenge of having nonlinear functions. Consequently, two broad classes of
methods for solving (1.1) exist: single-tree methods and multitree methods. In this section we
concentrate on methods for convex objective functions and convex constraints. See Section 5 for a
discussion of methods for nonconvex MINLPs. Throughout this section, we make the following
assumptions.
Assumption 3.1. Consider Problem (1.1) and assume the following:
A1 The set X is a bounded polyhedral set.
A2 The functions f and c are twice continuously differentiable convex functions.
A3 Problem (1.1) satisfies a constraint qualification for every point in the convex hull of the feasible set of
(1.1).
The most restrictive assumption is the convexity assumption A2. Assumption A1 is rather
mild, and A3 is technical. We do not specify the particular constraint qualification; it is needed
merely to ensure the existence of multipliers and the convergence of the NLP solvers. We note that
if we assume the existence of a strictly interior feasible point for (1.1), then A3 follows from A2 as
a consequence of Slaters constraint qualification (Griva et al., 2009), which is one of the strongest
constraint qualifications.
We start by discussing nonlinear branch-and-bound, whose basic concepts underpin both
classes of methods and which is a first example of a single-tree method.
3.1 Nonlinear Branch-and-Bound
The nonlinear branch-and-bound method for MINLPs dates back to Dakin (1965); see also (Gupta
and Ravindran, 1985). The algorithm starts by solving the NLP relaxation of (1.1), the root node,
defined by relaxing the integrality conditions on the integer variables, xi , i I . If this relaxation
is infeasible, then MINLP is also infeasible. If the solution of the relaxation is integer, then it also
solves the MINLP. Otherwise, branch-and-bound searches a tree whose nodes correspond to NLP
subproblems, and whose edges correspond to branching decisions. We use both optimality and
feasibility of NLP subproblems to prune nodes in the tree. We define a node problem and then
define the branching and pruning operations before stating the algorithm, which is also illustrated
in Figure 7.
A node in the branch-and-bound tree is uniquely defined by a set of bounds, (l, u), on the
integer variables and corresponds to the NLP

minimize f (x),

x
subject to c(x) 0,

(NLP(l, u))

x X,

li xi ui , i I.

Figure 7: Illustration of a nonlinear branch-and-bound algorithm that traverses the tree by solving
NLPs at every node of the tree.
We note that the root node relaxation corresponds to NLP(, ). Next, we describe the branch-
ing and pruning rules for branch-and-bound.
Branching. If the solution x0 of (NLP(l, u)) is feasible but not integral, then we branch on any
nonintegral variable, say x0i . Branching introduces two new NLP nodes, also referred to as child
nodes of (NLP(l, u)). In particular, we initialize bounds for two new problems as (l , u ) := (l, u)
and (l+ , u+ ) := (l, u) and then modify the bound corresponding to the branching variable
u 0 + 0
i := bxi c, and li := dxi e. (3.1)
The two new NLP problems are then defined as NLP(l , u ) and NLP(l+ , u+ ). In practice, the
new problems are stored on a heap H, which is updated with these two new problems.
Pruning rules. The pruning rules for NLP branch-and-bound are based on optimality and feasi-
bility of NLP subproblems. We let U be an upper bound on the optimal value of (1.1) (initialized
as U = ).
Infeasible nodes. If any node, (NLP(l, u)) is infeasible, then any problem in the subtree
rooted at this node is also infeasible. Thus, we can prune infeasible nodes (indicated by a
red circle in Figure 7).
Integer feasible nodes. If the solution, x(l,u) of (NLP(l, u)) is integral, then we obtain a new
incumbent solution if f (x(l,u) ) < U , and we set x = x(l,u) and U = f (x(l,u) ). Otherwise, we
prune the node because its solution is dominated by the upper bound.
Upper bounds on NLP nodes. If the optimal value of (NLP(l, u)), f (x(l,u) ) (or in fact a lower
bound on the optimal value) is dominated by the upper bound, that is, if f (x(l,u) ) U , then
we can prune this node because there cannot be any better integer solution in the subtree
rooted at (NLP(l, u)).
The complete nonlinear branch-and-bound algorithm is described in Algorithm 1, and Proposi-

tion 3.1 establishes its convergence.
Branch-and-bound for MINLP

Choose a tolerance > 0, set U = , and initialize the heap of open problems H = .
Add (NLP(, )) to the heap: H = H {NLP(, )}.
while H 6= do
Remove a problem (NLP(l, u)) from the heap: H = H { NLP(l, u) }.
Solve (NLP(l, u)) and let its solution be x(l,u) .
if (NLP(l, u)) is infeasible then
Node can be pruned because it is infeasible.
else if f (x(l,u) ) > U then
Node can be pruned, because it is dominated by upper bound.
(l,u)
else if xI integral then
Update incumbent solution: U = f (x(l,u) ), x = x(l,u) .
else
(l,u)
BranchOnVariable(xi , l, u, H), see Algorithm 2
Algorithm 1: Branch-and-bound for MINLP.
Proposition 3.1. Consider solving (1.1) by nonlinear branch-and-bound. Assume that the problem func-
tions f and c are convex and twice continuously differentiable and that X is a bounded polyhedral set. Then
it follows that branch-and-bound terminates at an optimal solution after searching a finite number of nodes
or with an indication that (1.1) has no solution.
Proof. The assumptions in Proposition 3.1 ensure that every NLP node can be solved to global
optimality. In addition, the boundedness of X ensures that the search tree is finite. The proof
now follows similarly to the proof for MILP branch-and-bound; see, for example, Theorem 24.1 of
Schrijver (1986). 2
We note that the convergence analysis of nonlinear branch-and-bound requires us only to en-
sure that every node that is pruned is solved to global optimality. The convexity assumptions are
one convenient sufficient condition that ensures global optimality, but clearly not the only one!
(l,u) (l,u)
Subroutine: S BranchOnVariable (xi , l, u, H) // Branch on a fractional xi for i I
(l,u) + (l,u) +
Set ui = bxi c, l = l and li = dxi e, u = u.
Add NLP(l , u ) and NLP(l+ , u+ ) to the heap: H = H {NLP(l , u ), NLP(l+ , u+ )}.
Algorithm 2: Branch on a fractional variable.
Our description of branch-and-bound leaves open a number of important questions. In par-

ticular, two important strategic decisions concern the selection of branching variable and the next
problem to be solved. Details of these important algorithmic decisions can be found in (Achter-
berg et al., 2004) for the MILP case, and in (Bonami et al., 2011) for the MINLP case. It turns out
that most observations generalize from the MILP to the MINLP case. In particular, branching
decisions are best based on estimates of the effect of branching on a particular variables. On the
other hand, a depth-first search is typically preferred initially in order to obtain a good incumbent
solution, after which most solvers switch to a best-estimate node selection strategy. In the next
two sections we discuss the choice of branching variable and search strategy, before discussing
other implementation considerations and presenting a generic branch-and-cut approach for con-
vex MINLPs.
3.1.1 Selection of branching variable
Choosing a good branching variable is a crucial component of branch-and-bound. Ideally, we

would like to choose the sequence of branching variables that minimizes the size of the tree that
we need to search. Doing so however is impractical, because the sequence of branching variables
is not known a priori. A more achievable goal in selecting a branching variable is to choose a
variable that maximizes the increase in the lower bound at a node.
We denote by Ic I the set of all candidate branching variables at a particular node in the
branch-and-bound tree. For example, Ic could be the index set of all fractional integer variables
or a subset chosen according to some user-defined priorities (see, e.g., Williams (1999)). A simple
branching rule is to select the variable with the largest integer violation for branching, in other
words, choose
argmax {min (xi bxi c , dxi e xi )} ,
iIc
which is known as maximum fractional branching. In practice however, this branching rule is not
efficient: it performs about as well as randomly selecting a branching variable (Achterberg et al.,
2004).
The most successful branching rules estimate the change in the lower bound after branching.
Because we prune a node of the branch-and-bound tree whenever the lower bound for the node
is above the current upper bound, we want to increase the lower bound as much as possible. For
every integer variable xi , i I, we define degradation estimates Di+ and Di for the increase in the
lower bound value after branching up and down on xi , respectively. A reasonable choice would
be to select the variable for which both Di+ and Di are large. We combine the up and down
degradations Di+ and Di to compute a score for each candidate branching variable; the variable
of highest score is selected. A common formula for computing this score is
si := min(Di+ , Di ) + (1 ) max(Di+ , Di ),
where [0, 1] is a prescribed parameter typically close to 1. We then select the branching variable
that maximizes si :
argmax {si } .
iIc
We next describe two methods for estimating Di+ and Di and show how they can be combined.
Strong branching computes the degradations Di+ and Di by solving both child nodes for all
branching candidates, xi , i Ic , which requires the solution of 2 |Ic | NLPs. To this end, we let
the solution to the current node, (NLP(l, u)), be f (l,u) . For every branching candidate xi , i Ic we
create two temporary branching problems, NLPi (l , u ) and NLPi (l+ , u+ ) as in (3.1) and solve
them. If both NLPs are infeasible for an index i, then we can prune the parent node (NLP(l, u)); if
one of them is infeasible, then we can tighten that integer variable in the parent node and resolve
it; otherwise, we let the solution be fi+ and fi , respectively. Given the values fi+ and fi , we
compute Di+ and Di as
Di+ = fi+ f, and Di = fi f.
Strong branching can significantly reduce the number of nodes in a branch-and-bound tree, but
it is often slow overall because of the added computational cost of solving two NLP subproblems
for each fractional variable. In order to reduce the computational cost of strong-branching, it is
often efficient to solve the subproblems only approximately. If the relaxation is an LP, as in the
case of LP/NLP-BB (see Section 3.3.1), then one can limit the number of pivots. In the NLP case,
we can limit the number of iterations, but that does not reduce the solve time sufficiently. Instead,
we can use approximations of the NLP, which can be warm-started much faster than NLPs can.
One approach that has been shown to be efficient is to use the basis information from a quadratic
program solved at the parent node, and perform a limited number of pivots on this quadratic
approximation in order to obtain estimates for fi+ and fi , respectively (Bonami et al., 2011).
Pseudocosts branching keeps a history of the results of past branching decisions for every vari-
able and computes the degradations Di+ and Di by averaging the increase in the objective value

over the history. For every integer variable, we let n+ i , ni denote the number of times we have
solved the up/down node for variable i. We update the per unit change in the objective when

branching on xi by computing the pseudocost, p+ i , pi whenever we solve an up or down child
node:
fi+ f fi f
p+i = + p +
i , n +
i = n +
i + 1 or p i = + p
i , ni = ni + 1. (3.2)
dxi e xi xi bxi c
The pseudocosts are then used to estimate Di+ and Di whenever we need to make a branching
decision as
p+ p
Di+ = (dxi e xi ) i+ and Di = (xi bxi c) i .
ni ni
Pseudocosts are typically initialized by using strong branching. The update of the pseudocosts
is relatively cheap compared with that of strong branching, because the solutions of the parent
and child node are available. Statistical experience on MILP problems has shown that pseudo-
costs are reasonable estimates of the degradation in the objective after branching (Linderoth and
Savelsbergh, 1999). One difficulty that arises with pseudocosts, however, is how to update the
pseudocosts if the NLP is infeasible. Typically the update is skipped in this case, but a fruitful
alternative motivated by NLP might be to use the value of the `1 exact penalty functions (Fletcher,
1987, Chapter 12.3), which is readily available at the end of the NLP solve.
Reliability branching is an attractive hybrid between strong branching and pseudocost branch-

ing. It performs strong branching on a variable and updates the pseudocosts until n+ i or ni are
greater than a threshold (a typical value for is 5). This corresponds to using strong branching
early during the tree search, when branching decisions are more critical, and then switching to
pseudocost branching once reliable estimates are available.
Branching on general disjunctions is an alternative to branching on a variable. We can branch

in many ways. One popular way is to branch on special ordered sets (Beale and Tomlin, 1970). A
more general approach is to branch on split disjunctions of the form
aT xI b aT xI b + 1 ,

(3.3)
where a Zp and b Z. Split disjunctions have been shown to produce promising results for
MILP (Jobst et al., 2001) but have not been used in MINLP.
3.1.2 Node selection strategies
The second important strategic decision is which node should be solved next. The goal of this
strategy is to find a good feasible solution quickly in order to reduce the upper bound, and to prove
optimality of the current incumbent x by increasing the lower bound as quickly as possible. We
introduce two popular strategies, depth-first search and best-bound search, and discuss their strengths
and weaknesses. We also present two hybrid schemes that aim to overcome the weaknesses of
these two strategies.
Depth-first search selects the deepest node in the tree (or the last node that was added to H).
One advantage of this strategy is that it keeps the list of open nodes, H, as small as possible.
This property made it a popular strategy for the earliest implementations of branch-and-bound
(Dakin, 1965). Another advantage is that this strategy minimizes the change to subsequent NLP
relaxations, (NLP(l, u)), that are solved, because only a single bound is changed. This fact allows
us to exploit warm-start techniques that reuse the existing basis factors in MILP problems, or
make use of a good starting point in MINLP problems. Though some attempts have been made
to use warm-starts in MINLP, see (Bonami et al., 2011; Mahajan et al., 2012), they generally have
not been as successful in MINLP, mainly because the Jacobian and Hessian matrix change from
one node to another so that factorizations are always outdated. Unfortunately, depth-first search
can exhibit extremely poor performance if no upper bound is found, exploring many nodes with
a lower bound that is larger than the solution.
Best-bound search selects the node with the best lower bound. Its advantage is that for a fixed
sequence of branching decisions it minimizes the number of nodes that are explored, because all
nodes that are explored would have been explored independently of the upper bound. On the
other hand, the weaknesses of this strategy are that it may require significantly more memory
to store the open problems, the sequence of NLP subproblems does not readily allow for warm-
starts, and it usually does not find an integer feasible solution before the end of the search. This
last point is particularly relevant for very large problems or if the solution time is limited such
as in real-time applications, because best-bound search may fail to produce even a feasible point.
Like depth-first search, this strategy has also been used since the very first branch-and-bound
algorithms (Land and Doig, 1960; Lawler and Woods, 1966).
Variants of best-bound search. Two variants of best-bound search have been proposed. Both
try to estimate the effect of the branching on the bound by using pseudocosts. We let fp denote
the lower bound or estimate of problem p on the heap.
1. Best expected bound selects the node with the best expected bound after branching, which is
estimated as
+ p+ p
bp = fp + (dxi e xi ) + and bp = fp + (xi bxi c) i .
i
ni ni
The next node is selected as maxp min b+ .

p , bp
2. Best estimate chooses the node that is expected to contain the best expected integer solution
within its subtree based on pseudo costs. The best expected solution within a subtree can be
estimated as
p+ p
X
i i
ep = fp + min (dxi e xi ) + , (xi bxi c) ,
i:x fractional
ni ni
i
namely, by adding the pseudocost estimates for all non-integral integer variables to the lower
bound at that node. The next node to be solved is then chosen as maxp {ep }.
Both strategies have been reasonably successful in the context of MINLP (Goux and Leyffer, 2003).
Hybrid search strategies. Good search strategies try to combine depth-first and best-bound
search. Two such strategies are the two-phase method and the diving method (Eckstein, 1994;
Linderoth and Savelsbergh, 1999; Achterberg, 2005).
1. Two-phase methods start with depth-first search until one or a small number of integer solu-
tions have been found. It then switches to best-bound search in order to prove optimality.
If the tree becomes too large during this second phase, then the method switches back to
depth-first search in order to keep the number of open nodes manageable.
2. Diving methods are also two-phase methods. The method starts with depth-first search until
a leaf node (feasible or infeasible) is found. It then backtracks to the best bound on the tree
to start another depth-first dive. Diving methods continue to alternate between this diving
step and the backtracking step.
3.1.3 Other implementation considerations
Two other considerations are important when implementing an MINLP branch-and-bound solver:
(1) the degree to which inexact subproblem solves can be used during the tree search and (2) the
use of heuristics to find good incumbents quickly that will then reduce the size of the search tree.
Inexact NLP subproblem solves. There exists considerable freedom in how exactly NLP sub-
problems are solved in the tree search. In fact, at any non-leaf node of the tree, we need only
to provide a branching variable. We can exploit this freedom to consider inexact solves at NLP
subproblems in order to improve the performance of branch-and-bound. The first approach to
using inexact NLP solves is due to Borchers and Mitchell (1994). The authors consider a sequen-
tial quadratic programming (SQP) approach to solving every NLP subproblem and interrupt SQP
after every iteration to check whether SQP appears to converge to a nonintegral solution. If con-
vergence to nonintegral solution is detected, the authors stop the current NLP solve and proceed
to branching on any non-integral integer variable. Note, however, that the inexact solution of an
NLP subproblem does not provide a valid lower bound for the node. Hence, the authors propose
to occasionally solve an augmented Lagrangian dual to obtain a lower bound. This approach
can be improved by combining insights from outer approximation to obtain implicit bounds; see
(Leyffer, 2001). An alternative approach is presented by Mahajan et al. (2012). The authors search
the branch-and-bound tree using a single quadratic program generated at the root node. The ad-
vantage of this approach is that quadratic programs, unlike NLPs, can be warm-started efficiently
after branching by reusing factors of the basis of the parent node. The authors show how the
pruning rules described above can be adapted to ensure that the algorithm guarantees a global
solution to a convex MINLP. Numerical experience shows that this approach can reduce the CPU
time of branch-and-bound by several orders of magnitude for some problems.
Heuristics for finding good incumbents. The bounding in Algorithm 1 shows that it is impor-
tant to find good incumbents quickly in order to prune more parts of the tree. A range of heuristics
exists that can be used at the root node or at intermediate nodes of the tree; these methods are dis-
cussed in Section 6.
3.1.4 Cutting planes for nonlinear branch-and-bound
Branch-and-bound algorithms can be enhanced by adding cutting planes, as first described by

Stubbs and Mehrotra (1999) for convex mixed 0-1 nonlinear programs, based on Balas et al. (1996)
for MILP. The branch-and-cut algorithm extends the branch-and-bound Algorithm 1 by an ad-
ditional step during which one or more cutting planes may be generated and added to a node
(NLP(l, u)) in order to cut off a fractional optimal solution x(l,u) . A node is branched on only
if the relaxed optimal solution remains fractional even after a prescribed number of rounds of
adding cuts have been added or if no suitable cuts could be generated at all. The hope is that
adding cutting planes will lead to a significant reduction of the tree size or will ideally remove the
need for branching by producing a locally tight description of the convex hull. An outline of the
branch-and-cut algorithm for MINLP is given as Algorithm 3.
Branch-and-cut for MINLP

Choose a tolerance > 0, and set U = .
Initialize the heap of open problems H = .
Add (NLP(, )) to the heap: H = H {NLP(, )}.
while H 6= do
Remove a problem (NLP(l, u)) from the heap: H = H { NLP(l, u) }.
repeat
Solve (NLP(l, u)) and let its solution be x(l,u) .
if (NLP(l, u)) is infeasible then
Node can be pruned because it is infeasible.
else if f (x(l,u) ) > U then
Node can be pruned, because it is dominated by upper bound.
(l,u)
Update incumbent solution: U = f (x(l,u) ), x = x(l,u) .
else if more cuts shall be generated then
GenerateCuts(x(l,u) , j), see Algorithm 4.
until no new cuts generated
if (NLP(l, u)) not pruned and not incumbent then
(l,u)
BranchOnVariable(xj , l, u, H), see Algorithm 2
Algorithm 3: Branch-and-cut for MINLP.
The most significant difference from the branch-and-bound algorithm for MINLP is the gen-
eration of additional cutting planes. We rely on a generic subroutine GenerateCuts that produces
a new valid inequality for the current tree node. The high-level description of this step is to solve
(l,u)
a separation problem. Given a point x(l,u) with xj / {0, 1} and a feasible set F(l, u) associated
with the current node NLP(l, u), find a vector (0 , )T such that the inequality T x 0 holds
T
for all x F(l, u) but cuts off x(l,u) by satisfying T x(l,u) > 0 . Various approaches to this end are
discussed in Section 4.
Like branch-and-bound, a number of implementation issues are open. In their original work
Stubbs and Mehrotra (1999) generate cuts only at the root node. Such cuts will then be valid for the
entire branch-and-bound tree and are hoped to reduce it in size. In general, cuts generated in tree
(l,u)
Subroutine: GenerateCuts (x(l,u) , j) // Generate a valid inequality that cuts off xj
/ {0, 1}
(l,u)
Solve a separation problem in x(l,u)to obtain an inequality that cuts off xj
/ {0, 1} from
the feasible set of (NLP(l, u)). Add this inequality to (NLP(l, u)).
Algorithm 4: Generate a subgradient cut by solving a separation problem.
nodes will be valid only for that particular subtree. Lifting procedures can sometimes be obtained
in order to make locally generated cuts also globally valid for the entire branch-and-bound tree.
Implementations of branch-and-cut solvers may choose to maintain both a global and a local pool
of cuts.
Depending on the type of cut generated, the separation problems may themselves be difficult
to solve numerically. The performance of a branch-and-cut scheme may depend crucially on the
ability to reliably and precisely solve the arising separation problems. Solver failures may reduce
the number of valid cuts obtained. Care must be taken not to cut off the optimal solution because
of numerical difficulties.
3.2 Multitree Methods for MINLP
One drawback of nonlinear branch-and-bound is that it requires the solution of a large number
of NLPs that cannot be easily hot-started (unlike MILP where we can reuse the LP basis factors
after branching). This observation led to the development of another class of methods that we
term multitree methods because they decompose the MINLP into an alternating sequence of NLP
subproblems and MINLP relaxations. Here, we review three such methods: outer approxima-
tion (Duran and Grossmann, 1986; Fletcher and Leyffer, 1994; Bonami et al., 2008), generalized
Benders decomposition (Geoffrion, 1972; Van Roy, 1983), and the extended cutting-plane method
(Westerlund and Pettersson, 1995).
3.2.1 Outer approximation

(j)
We start by defining the NLP subproblem obtained by fixing the integer variables to xI in (1.1),

minimize f (x)

x
(j)
subject to c(x) 0 (NLP(xI ))

(j)
xX and xI = xI ,
(j)
and we let its solution be x(j) . If (NLP(xI )) is infeasible, then most NLP solvers will return a
solution to a feasibility problem of the form
X
minimize wi c+
i (x)
x

iJ

(j)
(F(xI ))

subject to ci (x) 0, i J
(j)
x X and xI = xI ,

where wi > 0 is a set of weights that can be chosen reduce to the `1 or ` norm minimization, J is
a set of constraints that can be satisfied, and its complement J is the set of infeasible constraints;
see, for example, (Fletcher and Leyffer, 1994, 2003; Gould et al., 2004). An optimal solution of
(j) (j)
(F(xI )) with a positive objective is a certificate that the corresponding NLP (NLP(xI )) is infea-
sible.
Next, we consider (1.2) and observe that the convexity of f and c implies that the linearization
(j) (j)
about the solution x(j) of (NLP(xI )) or (F(xI )), given by
T T
f (j) + f (j) (x x(j) ) and 0 c(j) + c(j) (x x(j) ), (3.4)
(j)
is an outer approximation of the feasible set of (1.2). We can show that if (NLP(xI )) is infeasible,
(j)
then the corresponding set of outer approximations ensures that xI = xI violates (3.4).
(j)
Lemma 3.1 (Lemma 1 in (Fletcher and Leyffer, 1994)). If (NLP(xI )) is infeasible and x(j) is an optimal
(j) (j)
solution of (F(xI )), then xI = xI violates (3.4) for all x X.
Next, we define an index set of all possible feasible integer assignments:

n o
(j) (j)
X := x(j) X : x(j) solves (NLP(xI )) or (F(xI )) . (3.5)
Because X is bounded, there exists only a finite (albeit exponentially large) number of different
(j)
integer points xI , and hence X is a finite set. Now we can construct an MILP that is equivalent
to (1.2):
minimize ,
,x

T
subject to f (j) + f (j) (x x(j) ), x(j) X

T
0 c(j) + c(j) (x x(j) ), x(j) X (3.6)

x X,

xi Z, i I.
We can show that (3.6) and (1.2) have the same optimal value and that any solution of (1.2) is
an optimal solution of (3.6). However, the converse is not true, as the following example from
Bonami et al. (2008) shows:
minimize x3 subject to (x1 12 )2 + x22 + x33 1, x1 Z [1, 2].

x
The MILP created from outer approximations contains no coefficient for x2 , because x2 = 0 is
optimal in all NLP subproblems. Hence, any value of x2 is optimal in the MILP.
Theorem 3.1. Assume that the assumptions in Proposition 3.1 hold, and let x solve (1.1). Then it follows
that x also solves (3.6). Conversely, if ( , x ) solves (3.6), then it follows that the optimal value of (1.1) is
and xi is an optimal solution in both problems.
Proof. The proof follows from the results by Bonami et al. (2008) and Fletcher and Leyffer (1994).
2
Of course, it is not practical to solve (3.6) because by the time it is set up, we already know the
solution of (1.1). Instead, Duran and Grossmann (1986) propose a relaxation algorithm that solves
(j)
an alternating sequence of MILP problems and NLP subproblems. Initially, we solve (NLP(xI ))
for a given initial point x(j) = x(0) and set up a relaxation of (3.6) in which we replace X by a
subset X k X , with X k = {0}. We also add an upper bound on corresponding to the best
solution found so far:
n o
(j)
< U k := min f (j) | (NLP(xI )) is feasible .
jk
We note, however, that this latter constraint is not enforceable in practice, and is typically replaced
by U k , where > 0 is a small tolerance. This upper bound ensures that once we have solved
(j)
(NLP(l, u)) and added its outer approximations (3.4) to (M (X k )), xi is not feasible in (M (X k ))
for k j, ensuring that outer approximation terminates finitely. Thus, the MILP master problem
solved at iteration k is given by

minimize ,

,x
subject to Uk

T

f (j) + f (j) (x x(j) ), x(j) X k

T
(M (X k ))

0 c(j) + c(j) (x x(j) ), x(j) X k

x X,

xi Z, i I.

A description of the outer approximation algorithm is given in Algorithm 5, and its convergence
result is stated in Theorem 3.2.
Outer approximation
Given x(0) , choose a tolerance > 0, set U 1 = , set k = 0, and initialize X 1 = .
repeat
(j) (j)
Solve (NLP(xI )) or (F(xI )) and let the solution be x(j) .
(j)
if (NLP(xI )) is feasible & f (j) < U k1 then
Update current best point: x = x(j) and U k = f (j) .
else
Set U k = U k1 .
Linearize objective and constraint f and c about x(j) and set X k = X k1 {j}.
Solve (M (X k )), let the solution be x(k+1) and set k = k + 1.
until MILP (M (X k )) is infeasible
Algorithm 5: Outer approximation.
The algorithm also detects whether (1.2) is infeasible. If U k = on exit, then all integer
assignments visited by the algorithm are infeasible, and hence (1.2) in infeasible. The use of upper
(j)
bounds on and the definition of the set and X k ensure that no xI is replicated by the algorithm.
Thus, One can prove that the algorithm terminates after a finite number of steps, provided that
there is only a finite number of integer assignments.
Theorem 3.2. If Assumptions 3.1 hold and if the number of integer points in X is finite, then Algorithm 5
terminates in a finite number of steps at an optimal solution of (1.1) or with an indication that (1.1) is
infeasible.
A proof of Theorem 3.2 was given by Fletcher and Leyffer (1994). The main argument of the
(j)
proof is that the optimality of x(j) in (NLP(xI )) implies that f (j) for any feasible point in
(j)
(M (X k )). The upper bound f (j) therefore ensures that the choice xI = xI in (M (X k ))
in not feasible. Hence, the algorithm is finite. The optimality of the algorithm follows from the
convexity of f and c, which ensures that the linearizations are supporting hyperplanes.
Figure 8: 3D plot of worst-case example for outer approximation (3.7).
Worst-case complexity of outer approximation. In practice, outer approximation often works

efficiently. However, Hijazi et al. (2010) provide an example where outer approximation takes an
exponential number of iterations. In particular, they consider the following MINLP:
n
1 2 n1
X
minimize 0 subject to xi x {0, 1}n . (3.7)
x 2 4
i=1
Geometrically, the problem corresponds to a feasibility problem that seeks a point that lies in the
intersecting the n-dimensional `2 ball with radius n1 4 centered at x = ( 21 , . . . , 12 ) with the unit
hypercube {0, 1}n . Since the distance from x to any vertex of {0, 1}n is 2n > n1 4 , the problem is
infeasible. Figure 8 shows a 3D plot of this example. The black lines illustrate the unit hypercube,
and the green surface is the boundary of the nonlinear constraint in (3.7). The authors show that
any outer approximation cut cannot cut more than one vertex. Thus, outer approximation must
visit all 2n vertices of the hypercube. The example generalizes to Benders decomposition and
extended cutting plane methods, because these methods are relaxations of outer approximation.
3.2.2 Generalized Benders decomposition
Generalized Benders decomposition was developed before outer approximation; see (Geoffrion,
1972). Given outer approximation, however, it is straightforward to develop the Benders cuts
and present Benders decomposition. We start by considering the outer approximations (3.4) and
assume that the constraints X are inactive. Summing up the outer approximations (3.4) weighted
(j)
with (1, (j) ), where (j) 0 are the multipliers of (NLP(xI )), we obtain the following valid
inequality:
m
!T
T

(j) (j)
X
f (j) + (j) c(j) + f (j) + i ci (x x(j) ). (3.8)
i=1
We observe that (j)T

c(j) = 0 as a result of complementary slackness and that the continuous
variable component of the gradient
m
(j) (j)
X
C f (j) + i C ci =0
i=1
as a result of the optimality of x(j) . Thus, we can rewrite the cut (3.8) in the integer variables only
as !T
m
(j) (j) (j)
X
f (j) + I f (j) + i I ci (xI xI ), (3.9)
i=1
(j)
which is the Benders cut for feasible subproblems. We also observe that the optimality of (NLP(xI ))
(j) (j)
implies the existence of multipliers I of the bounds xI = xI and that their value is equal to the
gradient in the Benders cut. Thus, we can write the Benders cut compactly as
(j)T (j)
f (j) + I (xI xI ). (3.10)
(j)
With (F(xI )), a similar derivation shows that the Benders cut for infeasible subproblems is
(j)T (j)
X
0 wi c+
i (x) + I (xI xI ), (3.11)
iJ
(j) (j) (j)

where I are the multipliers of xI = xI in (F(xI )).
The advantage of the Benders cuts is that they involve only the integer variables and one
objective variable. A disadvantage is that the Benders cuts are almost always dense. Moreover,
the Benders cuts are weaker than the outer approximation cuts from which they are derived.
3.2.3 Extended cutting-plane method
The extended cutting-plane method (Westerlund and Lundqvist, 2001; Westerlund and Petters-
son, 1995) can be viewed as a variant of outer approximation that does not solve NLP subprob-
(j) (j)
lems such as (NLP(xI )) or (F(xI )). Instead, the extended cutting-plane method linearizes all
functions at the solution of the MILP master problem, x(k) . If x(k) satisfies all linearizations, then
we have solved the MINLP. Otherwise, we choose one (or a small number of) the most violated
linearizations and add it to the MILP master problem. The method alternates between the solution
of the master problem and the generation of linearizations that are underestimators if the MINLP
problem is convex.
Convergence of the extended cutting-plane method follows similar to that of outer approxi-
mation. The convexity of f and c ensures that the linearizations are a separating hyperplane, and
the convergence in the continuous space follows from the convergence of Kelleys cutting-plane
method (Kelley, 1960).
One weakness of the extended cutting-plane method is that it can produce the same integer
(j)
assignment multiple times, because the cuts are not generated from solutions of (NLP(xI )) or
(j)
(F(xI )). The rate of convergence of Kelleys cutting-plane method is in general linear, and hence
it may require a larger number of iterations. In practice, however, the extended cutting-plane
method is competitive with outer approximations, and the cutting planes it creates have been
used to accelerate outer-approximation-based schemes, such as the LP/NLP-based branch-and-
bound method discussed in Section 3.3.1; see for example (Abhishek et al., 2010).
3.3 Single-Tree Methods for MINLP
One single-tree method was already described above, namely, branch-and-bound. This section
shows how we can develop hybrid or integrated approaches that use outer approximation prop-
erties but require only a single MILP tree to be searched. The advantages of these approaches are
twofold. First, we avoid the need to re-solve related MILP master problems; second, we search a
tree whose nodes can be effectively warm-started by reusing basis information from parent nodes.
An alternative motivation for these hybrid techniques is to interpret them as branch-and-cut
algorithms for solving the large MILP, (3.6), with the full set of linearizations X as in (3.5). This
problem is clearly intractable, so instead we apply a delayed constraint generation technique of
the formulation constraints X k X . At integer solutions we can separate cuts by solving the
(j) (j)
NLP (NLP(xI )) or (F(xI )) for fixed integer variables.
3.3.1 LP/NLP-based branch-and-bound
Introduced by Quesada and Grossmann (1992), LP/NLP-based branch-and-bound (LP/NLP-BB)

methods have emerged as a powerful class of algorithms for solving convex MINLPs. LP/NLP-
BB improves outer approximation by avoiding the solution of multiple MILP master problems,
which take an increasing amount of computing time. Moreover, since these MILP relaxations are
strongly related to one another a considerable amount of information is regenerated each time a
relaxation is solved.
Instead, the LP/NLP-BB algorithm solves the continuous relaxation of (NLP(l, u)) and enforces
integrality of the xI variables by branching. Whenever a new integer solution is found, the tree
(j)
search is interrupted to solve (NLP(xI )), and the master MILP is updated with new outer ap-
proximations generated from the solution of the subproblem. Finally, the node corresponding to
the integer point is resolved, and the tree search continues.
The previous integer feasible node must be re-solved, because unlike ordinary branch-and-
bound a node cannot be pruned if it produces an integer feasible solution, since the previous
solution at this node is cut out by the linearizations added to the master program. Thus, only
infeasible nodes can be pruned.
We now formally define the LP/NLP-based branch-and-bound algorithm. It can be viewed as
a hybrid algorithm between nonlinear branch-and-bound (Algorithm 1) and outer approximation
(Algorithm 5); see also (Bonami et al., 2008). We denote by (LP(X k , li , ui )) the LP node (relaxation)
of the MILP master problem (M (X k )) with bounds li xI ui . In particular, (LP(X k , , )) is
the LP root node relaxation of (M (X k )). The algorithm uses an initial integer point x(0) to set up
the initial master problem, but it could also solve the NLP relaxation in order to derive the initial
outer approximations.
LP/NLP-Based branch-and-bound
Given x(0) , choose a tolerance > 0, set U 1 = , set k = 0, and initialize S 1 = .
Initialize MILP:
(j) (j)
Solve (NLP(xI )) or (F(xI )) and let the solution be x(j) .
Linearize objective and constraint f and c about x(j) and set S k = S k1 {j}.
(j)
if (NLP(xI )) is feasible then
Initialize MILP Search Tree:
Initialize the heap of open problems H = .
Add (LP(X k , , )) to the heap: H = H {LP(X k , , )}.
while H 6= do
Remove an LP problem from the heap: H = H {LP(X k , l, u)}.
Solve (LP(X k , l, u)) and let its solution be x(l,u) .
if (LP(X k , l, u)) is infeasible then
(j)
Node can be pruned because (LP(X k , l, u)) and hence (NLP(xI )) are infeasible.
(l,u)
(j) (l,u) (j) (j)
Set xI = xI and solve (NLP(xI )) or (F(xI )) and let the solution be x(j) .
Linearize objective and constraintf and c about x(j) and set S k = S k1 {j}.
(j)
if (NLP(xI )) is feasible & f (j) < U k then
else
Set U k = U k1 .
Add the LP Back to the heap: H = H {LP(X k , l, u)}.
Set k = k + 1.
else
(l,u)
Possibly add outer approximation cuts, or call BranchOnVariable(xi , l, u, H)
Algorithm 6: LP/NLP-based branch-and-bound.
As in the outer approximation algorithms the use of an upper bound implies that no integer
assignment is generated twice during the tree search. Since both the tree and the set of integer
variables are finite, the algorithm eventually encounters only infeasible problems, and the heap
is thus emptied so that the procedure stops. This provides a proof of the following corollary to
Theorem 3.2.
Theorem 3.3. If Assumptions 3.1 hold, and if |Y | < , then Algorithm 6 terminates in a finite number
of steps at a solution of (1.1) or with an indication that (1.1) is infeasible.
Figure 9 illustrates the progress of Algorithm 6. In (i), the LP relaxation of the initial MILP has
been solved, and two branches have been added to the tree. The LP that is solved next (indicated
by an *) does not give an integer feasible solution, and two new branches are introduced. The
next LP in (ii) produces an integer feasible solution indicated by a box. The corresponding NLP
subproblem is solved, and in (iii) all nodes on the heap are updated (indicated by the shaded
circles) by adding the linearizations from the NLP subproblem, including the upper bound U k
that cuts out the current assignment xI . Then, the branch-and-bound process continues on the
updated tree by solving the LP marked by an *.
(i) (ii)
*
*
Solution not integer feasible; Solution integer feasible;
branch solve the NLP-subproblem
(iii) (iv)
Update all problems Continue the branch and

on the stack bound process
Figure 9: Progress of LP/NLP-based branch-and-bound.
Algorithmic refinements of LP/NLP-BB. Abhishek et al. (2010) have shown that the branch-
and-cut LP/NLP-BB algorithm can be improved significantly by implementing it within a modern
MILP solver. Advanced MILP search and cut-management techniques improve the performance
of LP/NLP-BB dramatically. It is important to generate cuts at nodes that are not integer fea-
sible, in which case it is advantageous to generate outer approximations around the solution of
the LP relaxation, rather than solving an NLP (we term such cuts ECP cuts). We have shown
that LP/NLP-BB can be improved dramatically by exploiting MILP domain knowledge, such as
strong branching, adaptive node selection, and, most important, cut management. We have also
observed that weaker cuts, such as linearizations generated at LP solutions, improve the perfor-
mance of LP/NLP-BB.
3.3.2 Other single-tree approaches
It is straightforward to develop single-tree version of generalized Benders decomposition and the

extended cutting plane method. In the case of Benders decomposition, we need only to replace
the outer approximation cuts (3.4) by the Benders cut (3.10). In fact, the Benders cuts can already
be used to condense old outer approximation cuts in order to reduce the size of the LP relaxation.
The extended cutting plane method can be similarly generalized (see (Still and Westerlund, 2006))
by replacing the NLP solver with a simple function and gradient evaluation in order to generate
the outer approximations.
The convergence results are readily extended. In the case of Benders decomposition, conver-
gence follows from the convexity and the finiteness of the set of feasible integer variables, because
every integer assignment is visited at most once. In the case of the extended cutting plane method,
convergence follows from the finiteness of the integer set and the finite -convergence of the cut-
ting plane method.
3.4 Presolve Techniques for MINLP
A key component in successful MILP software is an efficient presolve. These techniques were pop-
ularized by Savelsbergh (1994), and some of these techniques can readily be extended to MINLP.
Here, we briefly review these techniques and demonstrate their importance with two examples:
coefficient tightening and constraint disaggregation. The goal of the presolve is to create an equiv-
alent but tighter LP (or NLP) relaxation that will likely result in a significantly smaller search tree.
Presolve techniques fall into two broad categories: basic functions for housekeeping and ad-
vanced functions for reformulation. Housekeeping include checking for duplicate rows (or con-
straints), tightening of bounds on the variables and constraints, fixing and removing variables,
and identifying redundant constraints. Reformulations include improvement of coefficients, dis-
aggregation of constraints, and derivation of implications or conflicts.
3.4.1 Coefficient tightening for MINLP
We observed very large search trees when we tried to solve certain chemical engineering synthe-
sis problem with a range of branch-and-bound solvers. For example, the search tree generated
by MINOTAUR (Mahajan et al., 2011) for problem Syn20M04M grows rapidly, as shown in Fig-
ure 10. The left figure shows the search tree after 75 and 200 seconds CPU time (the tree after 360
seconds is shown in Figure 1). The problem is not solved within 2 hours of CPU time, at which
point MINOTAUR has visited 264,000 nodes. Other solvers behave similarly: BONMIN-BB and
MINLPBB have searched about 150,000 nodes after 2 hours CPU time without finding the solu-
tion. On the other hand, this problem is solved in 9 seconds by using a hybrid outer approximation
branch-and-bound approach (Bonami et al., 2008). In this section we show that we can improve
the performance of branch-and-bound methods dramatically by extending coefficient tightening
to MINLP.
Figure 10: Left: branch-and-bound tree without presolve after 75 and 200 second CPU time. Right:
Complete tree after presolve and coefficient tightening were applied.
We start by describing the principle of coefficient tightening on a simple MILP, whose feasible
set is given by
x1 + 21x2 30, 0 x1 14, x2 {0, 1}. (3.12)
The feasible set is the union of the two red lines in Figure 11. If x2 = 1, then the constraint
x1 30 21x2 = 9 is tight; but if x2 = 0, then the x1 30 21x2 = 30 is loose. The shaded
area illustrates the feasible set of the LP relaxation, which is much larger than the convex hull of
the integer feasible set. We can tighten the formulation by changing the coefficient of the binary
variable x2 . It is easy to see that
x1 + 5x2 14, 0 x1 14, x2 {0, 1}
is equivalent to (3.12), and corresponds to the convex hull of the two red lines (see Figure 11,
right).
(9,1) (9,1)
(0,1) (0,1)
x1 + 21x2 30
(14,0) (14,0)
(0,0) (0,0) x1 + 5x2 14
Figure 11: Feasible set of MILP example (3.12) (left) and feasible set after coefficient tightening
(right).
We can extend this idea to MINLP problems, where we often encounter constraints of the form
c(x1 , . . . , xk ) b + M (1 y), y {0, 1}, and li xi ui , i = 1, . . . , k,
where M > 0 is a large constant. This form of constraints corresponds to on/off decisions that
arise, for example, in process synthesis problems (Turkay and Grossmann, 1996).
If the constraint c(x1 , . . . , xk ) b + M (1 y) is loose for y = 0, then we can reduce the value
of M by an amount determined by the following optimization problem:
(
maximize c(x1 , . . . , xk ),
x (3.13)
subject to li xi ui , i = 1, . . . , k.
Let us denote the optimal value by cu := c(x1 , . . . , xk ). If we have cu < b + M , then we can tighten
the coefficient M and arrive at the equivalent formulation
c(x1 , . . . , xk ) + (cu b)y cu , y {0, 1}, and li xi ui , i = 1, . . . , k.
Unfortunately, this approach requires solving an NLP for every set of constraints for which we
wish to tighten the formulation. Moreover, the optimization problem (3.13) is nonconvex, because
we maximize a convex function. Thus, we would have to apply global optimization techniques to
derive the bound cu , which would be prohibitive.
We can avoid the solution of this nonconvex NLP if the binary variable also appears as an
upper bound on the variables. This is indeed the case in many applications, such as the synthesis
design problem, where we find the following constraint structure:
c(x1 , . . . , xk ) b + M (1 y), y {0, 1}, and 0 xi ui y, i = 1, . . . , k, (3.14)
where y = 0 now switches the constraints and variables off. In this case, we can simply evaluate
the constraint cu := c(0, . . . , 0) and then derive the following tightened set of constraints (provided
that cu < b + M ):
c(x1 , . . . , xk ) + (cu b)y cu , y {0, 1}, and 0 xi ui y, i = 1, . . . , k. (3.15)

The effect of this reformulation is dramatic as can be seen from Figure 10. The right tree shows
the complete search tree from MINOTAUR with presolve, and the MINLP is now solved in 2.3
seconds. Similar improvements are obtained for other synthesis problems. The performance pro-
file (Dolan and More, 2002) in Figure 12 compares the performance of MINOTAURs presolve on
the Syn* and Rsyn* instances from the IBM/CMU library. Clearly, in almost all instances the
presolve helps to improve its performance. In addition, the presolve enables MINOTAUR to solve
20% more instances than the version without presolve.
0.8
Fraction of Instances
0.6
0.4
0.2
Minotaur with presolve
Minotaur without presolve
Bonmin
0
1 4 16 64 256 1024
Normalized Time
Figure 12: Performance profile comparing the effect of presolve on MINLP solver MINOTAUR for
Syn* and Rsyn* instances.
3.4.2 Constraint disaggregation for MINLP
The preceding section provides an example of how problem formulation can have a significant
impact on our ability to solve problems. Here, we present another idea, known as disaggregation
of constraints. A well-known example from MILP is the uncapacitated facility location problem
(see, e.g., Wolsey (1998)), which can be described as follows. Given a set of customers, i = 1, . . . , m,
and a set of facilities, j = 1, . . . , n, which facilities should we open (xj {0, 1}, j = 1, . . . , n) at
cost fj to serve the customers? The decision that facility j serves customer i is modeled with the
binary variable yij {0, 1}. The constraints that every customer be served by exactly one facility
and that only facilities that have customers assigned be open can be modeled as
n
X m
X
yij = 1, i = 1, . . . , m, and yij nxj , j = 1, . . . , n, (3.16)
j=1 i=1
respectively. A tighter formulation (in the sense that its LP relaxation is closer to the convex hull
of the feasible set) is given by the disaggregated form of the second constraints as
n
X
yij = 1, i = 1, . . . , m, and yij xj , i = 1, . . . , m, j = 1, . . . , n. (3.17)
j=1
As with (3.15), the difference in solution time can be dramatic. For a small random example with
n = m = 40, the CPU time is 53, 000 seconds for (3.16) versus 2 seconds for (3.17).
Similar disaggregation tricks have been applied to nonlinear functions. Tawarmalani and
Sahinidis (2005) consider constraint sets of the following form:
S := {x Rn : c(x) = h(g(x)) 0} , (3.18)
where g : Rn Rp is a smooth convex function and h : Rp R is a smooth, convex, and

nondecreasing function. These two conditions imply that c : Rn R is a smooth convex function.
We note that this functional form is related to the concept of group partial separability, which is
frequently used to compute gradients and Hessians efficiently in NLP (Griewank and Toint, 1984;
Gay, D.M., 1991; Bongartz et al., 1995).
We can derive a disaggregated version of the constraint set in (3.18) by introducing new vari-
ables y = g(x) Rp , which leads to the following convex set:
Sd := (x, y) Rn Rp : h(y) 0, y g(x) .

(3.19)
One can show that S is the projection of Sd onto x. This reformulation is of interest because any
outer approximation of Sd is stronger than the same outer approximation of S, and hence the
formulation (3.19) is preferred in any outer-approximation-based algorithm. In particular, given a
set of points X k := x(1) , . . . , x(k) we construct the outer approximations of S and Sd as

n T
o
S oa := x : c(l) + c(l) (x x(l) ) 0, x(l) X k (3.20)
and
n T T
o
Sdoa := (x, y) : h(l) + h(l) (y g(x(l) )) 0, y g (l) + g (l) (x x(l) ) x(l) X k , (3.21)
respectively. It can be shown that the projection of Sdoa onto x is contained in S oa . Tawarmalani
and Sahinidis (2005) give an example that shows that the outer approximation (3.21) is tighter
than (3.20). Moreover, for the two outer approximations to be equivalent, we may require an
exponentially larger number of linearization points x(l) in (3.20) compared with (3.21).
Hijazi et al. (2010) study a similar structure, namely, the partially separable constraint set:

q
X
x : c(x) := hj (aTj x + bj ) 0 , (3.22)

j=1
where hj : R R are smooth and convex functions, which ensures that this set is convex. We can
again derive a disaggregated form of this set by introducing new variables y Rq

q
X
(x, y) : yj 0, and yj hj (aTj x + bj ) . (3.23)

j=1
Again, one can show that the outer approximation of (3.23) is tighter than the outer approximation
of (3.22). We can apply this technique to the worst-case example, (3.7), choosing two linearization
points as x(1) {0, 1}n and its complement x(2) := e x(1) , where e = (1, . . . , 1) is the vector of all
ones. The combined outer approximation of (3.23) is then given by
n
X
3 1
yi , and xi 4 yi , and 4 xi yi ,
i=1
which together with xi {0, 1} implies that zi 0, which leads to zi n4 > n1

P
4 , which shows
(1) (2)
that any master problem that includes the linearizations from x and x is infeasible. Hijazi
et al. (2010) suggest two additional improvements for outer approximation for partially separable
constraints. The first improvement is to add additional linearization points for the univariate
functions hj (t) in order to obtain a better description of the nonlinear feasible set. The second
improvement is to employ an inner approximation of hj (t) in order to generate initial feasible
points in the case that aTj x + bj = xj .
4 Cutting Planes for Convex MINLPs

In this section we review different cutting planes for use in a branch-and-cut algorithm solving
convex MINLPs. We then review generalized disjunctive cuts and perspective cuts for MINLP. We
also cover details about the practical realization of disjunctive cuts in a branch-and-cut framework.
The final part of this section addresses the closely related problem class of mixed-integer second
order cone programs (MISOCPs).
4.1 Mixed-Integer Rounding Cuts
The LP/NLP-based branch-and-bound Algorithm 1 for MINLP solves MILP relaxations, which
we intend to strengthen by iteratively adding cuts to remove fractional solutions from these relax-
ations as outlined in Algorithm 3. We start by considering mixed-integer rounding cuts. These are
best introduced for the two-variable set
S := {(x1 , x2 ) R Z | x2 b + x1 , x1 0}, (4.1)
where R = {1}, I = {2}. Let f0 = b bbc, and observe that the inequality
x1
x2 bbc + (4.2)
1 f0
x1
x2
Figure 13: Mixed-integer rounding (MIR) cut. Feasible set of the LP relaxation (hatched), integer
feasible set (bold black lines), and MIR cut (grey) x2 2x1 derived from the inequality x2 12 +x1 .
is valid for X by verifying it for the two cases: x2 bbc and x2 bbc + 1. The situation is depicted
in Figure 13 for the set S = {(x1 , x2 ) R {0, 1} | x2 21 + x1 , 0 x1 2}.
For the general MILP case, it is sufficient to consider the set

X := {(x+ 2 p T + +
R , xR , xI ) R Z | aI xI + xR b + xR , xR 0, xR 0, xI 0}. (4.3)
This set describes a selected constraint row of a MILP, or a one-row relaxation of a subset of con-

straints aggregated in the vector a Rn and scalar b. Real variables are aggregated in x+R and xR
depending on the sign of their coefficient in aR . The extension from (4.1) is now straightforward
by observing the following inequality is valid for X:
X max{fi f0 , 0} x

R
bai c + xi bbc + , (4.4)
1 f0 1 f0
iI
where fi = ai bai c for i I and f0 = b bbc are the fractional parts of a and b.
Gomory cuts were originally derived by Gomory (1958, 1960) and Chvatal (1973) for integer
linear programs. In the mixed-integer case, a Gomory cut is given by the inequality
X X f0 (1 fi ) f0
fi xi + xi + x+
R+ x f0 (4.5)
fi 1 f0 R
iI1 iI2
where I1 = {i I | fi f0 } and I2 = I \ I1 . It can be seen to be an instance of a MIR cut by

considering the set

X = {(xR , x0 , xI ) R2 Z Zp | x0 + aTI xI + x+
R xR = b, xR 0, xI 0}, (4.6)
generating a MIR inequality from it, and eliminating the variable x0I .
To apply MIR cuts in MINLP, we can generate cuts from linearized inequality constraints and
the linearized objective constraint of the MINLP reformulation (1.2). Hence, it is sufficient to
consider cuts for MILPs. Akrotirianakis et al. (2001) report modest performance improvements
using this approach for MINLP.
4.2 Perspective Cuts for MINLP
Many MINLP problems employ binary variables to indicate the nonpositivity of continuous vari-
ables. If xi , i I is a binary variable and xj , j C is the associated continuous variable, the
relationship can be modeled with what is known as a variable upper bound constraint,
xj uj xi . (4.7)
Note that, since xi is a {0, 1} variable, if xj > 0, then xi = 1. If, in addition, the continuous variable
xj appears in a convex, nonlinear constraint, then a reformulation technique called the perspective
reformulation can lead significantly improved computational performance. Frangioni and Gentile
(2006) pioneered the use of this strengthened relaxation, using cutting planes known as perspective
cuts. The technique is perhaps best introduced with a simple example. Consider the following
mixed-integer set with three variables:
n o
S = (x1 , x2 , x3 ) R2 {0, 1} : x2 x21 , ux3 x 0 .
Figure 14 depicts the set S, which is the union of two convex sets S = S 0 S 1 , where
S 0 = (0, x2 , 0) R3 : x2 0 ,

S 1 = (x1 , x2 , 1) R3 : x2 x21 , u x1 0 .

y
y x2
z=1
Figure 14: The set S.
One can observe from the geometry that the convex hull of S requires the surface defines by
the family of line segments connecting the origin in the x3 = 0 plane to the graph of the parabola
x2 = x21 in the x3 = 1 plane. Using this geometric intuition, we can define the convex hull of S as
conv(S) = {(x1 , x2 , x3 ) R3 : x2 x3 x21 , ux3 x1 0, 1 x3 0, x2 0}. (4.8)
The expression x2 x3 x21 in (4.8) can be explained in terms of the perspective function of the left-
hand-side of the inequality x21 x2 0. For a convex function f : Rn R, the perspective function
P : Rn+1 R of f is (
0 if z = 0,
P(x, z) := (4.9)
zf (x/z) if z > 0.
The epigraph of P(x, z) is a cone pointed at the origin whose lower shape is f (x). If zi is an indi-
cator binary variable that forces some variables x to be 0, or else the convex nonlinear constraint
f (x) 0 must hold, then by replacing the constraint f (x) 0 with
zi f (x/zi ) 0, (4.10)
results in a convex inequality that describes a significantly tighter relaxation of the feasible region.
uk
Gunl and Linderoth (2012) (Lemma 3.2) slightly generalize this construction to the case where
the S0 side of the disjunction is an unbounded ray (like in Figure 14).
The general construction is as follows. We consider the problem
n o
min f (x) + cz | Ax bz ,
(x,z)Rn {0,1}
where (i) X = {x | Ax b} is bounded (also implying {x | Ax 0} = {0}), (ii) f (x) is a convex

X and
function that is finite on X, and (iii) f (0) = 0. Under these assumptions, for any x
subgradient s f (x), the inequality
x) + c + sT (x x
v f ( x) sT x
) + (c + f ( ))(z 1) (4.11)
is valid for the equivalent mixed-integer program

n o
min v | v f (x) + cz, Ax bz .
(x,z,v)Rn {0,1}R
Inequality (4.11), called the perspective cut, was introduced by Frangioni and Gentile (2006) and
used dynamically to build a tight formulation. (Gunl uk
and Linderoth, 2010) show that perspec-
tive cuts are indeed outer approximation cuts for the perspective reformulation for this MINLP.
Therefore, adding all (infinitely many) perspective cuts has the same strength as the perspective
reformulation. It may be computationally more efficient to use linear outer approximation in-
equalities of the form (1.5) instead of using the nonlinear form (4.10).
4.3 Disjunctive Cutting Planes for MINLP
Disjunctive cuts for use in a branch-and-cut procedure were first discussed by Stubbs and Mehro-
tra (1999) for convex mixed 0-1 nonlinear programs, based on Balas et al. (1996) for MILP. Simulta-
neously, Ceria and Soares (1999) derived the disjunctive programming arguments to be presented
in a more general setting. We consider the convex mixed 0-1 nonlinear program

minimize

x,

subject to f (x) ,

c(x) 0, (4.12)

x X,

xi {0, 1} i I.
For the continuous relaxation of this problem, optimal solutions are guaranteed to lie on the
boundary of the continuous relaxation C = {x X | f (x) , c(x) 0, 0 xI 1} of the
feasible set. We consider a node in a branch-and-bound tree, with optimal solution x0 , where x0j
is fractional for some j I. We denote by I0 I , I1 I the index sets of integer variables fixed
to zero or one, respectively, by the previous branching decisions that led to the tree node under
consideration. We denote by F the index set of free real and integer variables.
Instead of separating the fractional solution x0 by simply branching to x0j = 0 and x0j = 1,
we are interested in adding a valid inequality to the relaxation that cuts off x0 . To this end, we
consider the disjoint feasible sets obtained by fixing xj to either choice,
Cj0 = {x C | xj = 0, 0 xi 1 i I, i 6= j}, (4.13)

Cj1 = {x C | xj = 1, 0 xi 1 i I, i 6= j}. (4.14)
We are interested in a description of M j (C) = conv(C 0 C 1 ), the convex hull of the continuous
j j
relaxation C of the feasible set with either binary restriction on a single selected xj . For the set
M j (C), Stubbs and Mehrotra (1999) give the following description:

v0 + v1 = xF , v0j = 0, v1j = 1

Mj (C) = (xF , v0 , v1 , 0 , 1 ) 0 + 1 = 1,

0 , 1 0 , (4.15)

Pc (v0 , 0 ) 0, Pc (v1 , 1 ) 0, 1 i nc
i i
where Pf (v, ) is the perspective function (4.9) for the function f . This procedure can be gen-
eralized by repetition to describe the lifted convex hull of the disjoint feasible sets obtained for
multiple fractional xj , j I all at once. With that description of the convex hull in hand, we
can set up and solve a separation NLP to find a point x closest to the fractional one x0 and in the
convex hull:
minimize ||x x0 ||,

x,v0 ,v1 ,0 ,1

subject to (x, v0 , v1 , 0 , 1 ) M j (C)
(BC-SEP(x0 , j))

xi = 0, i I0

xi = 1, i I1 .

be an optimal solution of (BC-SEP(x0 , j)), and denote by F the Lagrange multipliers for the
Let x
equality constraint v0 + v1 = xF in (4.15). Then, an inequality that is valid for the current node
and cuts off x
from the feasible set is given by
FT xF FT x
F . (4.16)
As an example for deriving a disjunctive cut, consider the MINLP

minimize x1

x1 ,x2
subject to (x1 21 )2 + (x2 34 )2 1

(4.17)

2 x1 2

x2 {0, 1}

with the relaxed optimal solution x0 = (x01 , x02 ) = ( 21 , 34 ) shown in Figure 15 on the left. To
construct the separation problem for identifying a disjunctive cut that removed this solution from
the relaxation, we consider the individual convex hulls C 0 and C 1 for x2 = 0 and x2 = 1, where the
limits are found by solving the constraint (x1 12 )2 + (x2 43 )2 1 for x1 given a fixed x2 ,
0
n o
C = (x1 , 0) R {0, 1} 2 7 4x1 2 + 7 , (4.18)

n o
C1 = (x1 , 1) R {0, 1} 2 15 4x1 2 + 15 .

With the euclidean norm || ||2 , the minimum norm separation problem for x0 yields (approxi-
mately) the solution x
= ( 2 ) = (0.40, 0.78) with steepest descent direction = (0.1, 0.03)
x1 , x
for the norm objective. This identifies the separating hyperplane
! !
x 0.40
1
0.1 0.03 0.1 0.03 = x1 + 0.3x2 0.166, (4.19)
x2 0.78
C0 C1
shown in Figure 15 on the right.
x1
C0 C1
x
x2
x
= (
x1 , x
2 ) x
= (
x1 , x
2 )
Figure 15: Left: NLP relaxation C (grey), integer feasible convex hull (hatched), and disjoint convex
hulls C 0 and C 1 (bold black lines) for the MINLP example (4.17). Right: Solution to the minimum
norm problem, and resulting disjunctive cut for the MINLP example (4.17). The next NLP solution
including the disjunctive cut will produce the MINLP solution.
We are free to choose the norm in (BC-SEP(x0 , j)), but different choices lead to different refor-
mulations, for example for the 1-norm or the -norm. Stubbs and Mehrotra (1999) observed the
most favorable performance of the generated cuts when using the -norm.
The cut (4.16) is in general valid for the local subtree of the branch-and-bound tree only. Stubbs
and Mehrotra (1999) show that a lifting to a globally valid cut
T x T x
. (4.20)
may be obtained by assigning
i = min{eTi H0T 0 , eTi H1T 1 }, i

/F (4.21)
where ei denotes the ith unit vector; 0 = (0F , 0), 1 = (1F , 0), where 0F , 1F are the Lagrange
multipliers of the perspective inequalities from (4.15) active in x ; and H0 , H1 are matrices formed
T T
from subgradient rows v Pci (v0 , 0 ) , v Pci (v1 , 1 ) of those inequalities.
Disjunctive cuts are linear in the lifted variable space. In (Stubbs and Mehrotra, 2002), a step
toward nonlinear cutting planes is taken and convex quadratic inequalities are derived as cuts for
convex mixed 0-1 integer programs. Computational evidence so far suggests that such cuts do not
benefit branch-and-cut procedures in general, although nonlinear cuts for the mixed integer conic
case show otherwisesee the end of Section 4.5.
4.4 Implementation of Disjunctive Cuts
The nonlinear separation problem (BC-SEP(x0 , j)) has twice the number of unknowns as the orig-
inal problem and is not differentiable everywhere because of the perspective constraints. These
drawbacks have hindered the use of disjunctive cutting planes for convex 0-1 MINLP, and Stubbs
and Mehrotra (1999) reported computational results on only four instances with no more than
30 unknowns. Addressing this issue, we mention an LP-based iterative separation procedure by
Klnc et al. (2010) and Klnc (2011) that replaces (BC-SEP(x0 , j)), and a nonlinear separation ap-
proach by Bonami (2011) that circumvents the difficulties inherent in (BC-SEP(x0 , j)).
Klnc et al. (2010) and Klnc (2011) propose an iterative procedure to replace (BC-SEP(x0 , j))
by a sequence of cut generation LPs. They report both an increased number of solvable problems
and a significant reduction in runtime for a set of 207 instances. To describe this method, we let
B C be a relaxation of the original MINLP relaxation, and we consider the sets
Bj0 = {x B 0 | xj = 0}, Bj1 = {x B 0 | xj = 1}. (4.22)
Valid inequalities for conv(Bj0 Bj1 ) are also valid for conv(Cj0 Cj1 ). The separation program be-
comes a linear one if B is restricted to be a polyhedron and if an appropriate norm is chosen in
(BC-SEP(x0 , j)). Klnc et al. (2010) and Klnc (2011) iteratively tighten polyhedral outer approx-
imations of Cj0 and Cj1 using inequalities generated as disjunctive cuts. To this end, in iteration t
two sets K0t , K1t of linearization points are maintained, resulting in the polyhedral approximations
g(x0 )

Ft0 = x Rn | xi = 0, g(x0 ) + (x x0 ) 0 x Kt0 , (4.23)
x
g(x0 )

1 n 0 0 1
Ft = x R | xi = 1, g(x ) + (x x ) 0 x Kt . (4.24)
x
Since the sets Ft0 , Ft1 are polyhedral relaxations of C 0 and C 1 , valid disjunctive cuts can be gen-
erated. Initially empty, the sets Kt0 , Kt1 are augmented by the solutions x0t of the linear separa-
tion problems and with two so-called friendly points yt and zt , respectively, that satisfy x0t =
yt + (1 )zt for some [0, 1]. Klnc et al. (2010) prove this approach to be sufficient to ensure
that in the limit the obtained inequality is of the same strength as if the nonlinear separation prob-
lem (BC-SEP(x0 , j)) was solved. The process can be terminated early if it is observed that the cuts
are not effective in reducing the solution integrality gap. The procedure is applicable to general
convex MINLPs with xi Z as well.
g
A similar procedure based on the outer approximating set B = {x Rn | g(x0 )+ x (xx0 ) 0}
was proposed earlier by Zhu and Kuno (2006) but does not necessarily converge (Klnc et al., 2010;
Klnc, 2011).
Bonami (2011) addresses the difficulty of solving (BC-SEP(x0 , j)) in the special situation when
x0 with k < xj < k + 1 fractional for some j I is to be separated from a relaxation that was ob-
tained by a simple disjunction created from a split relaxation. In this case, the separation problem
allows an algebraic reduction to

maximize v1,j

v1
subject to c v1 f k v ,

i f0 0 1,j
0 (4.25)
x v

ci 1f0
1
f0 k v1,j ,

n
v1 R .
where f0 = x0j k > 0 is the fractional part of xj . Again, the approach is applicable to general
convex mixed-integer problems. In (4.25), the perspective is no longer required, the problem size
does not grow compared to the MINLP under consideration, and differentiability is maintained.
One can show that the optimal objective value is smaller than f0 k if and only if x0 / Cjk . The
separating inequality can be computed from an LP model as described for the approach by Klnc
et al. (2010) and the particular choice
0
x v1

v1
Kt0 = , Kt1 = , (4.26)
1 f0 f0
if v1 denotes the optimal solution of (4.25).
4.5 Mixed-Integer Second-Order Cone Programs
Both mixed-integer rounding cuts and disjunctive cuts can be generalized from the LP cone Rn+
to other cones such as second-order cones defined by the set {(x0 , x) R Rn | x0 ||x||2 }. We
consider the class of mixed-integer linear programs with a second-order cone constraint

minimize cT x

x
subject to x K

(MISOCP)

x X

xi Z i I.

The conic constraint x K represents the product of k 1 smaller cones K := K1 . . . Kk ,

defined as
Kj := xj = (xj0 , xTj1 )T R Rnj 1 : ||xj1 ||2 xj0 , 1 j k,

(4.27)
and x = (xT1 , . . . , xTk )T . Convex MINLP solvers are usually not directly applicable to this problem
class because the conic constraint is not continuously differentiable. Drewes (2009) and Drewes
and Ulbrich (2012) propose a variant of the LP/NLP-based branch-and-bound, Algorithm 6, that
solves continuous relaxations of (MISOCP) instead of NLPs for a fixed integer assignment xkI :

minimize cT x

x
subject to x K,

(SOCP(xkI ))

x X,

xI = xkI .

The algorithm builds MIP outer approximations of (MISOCP) from subgradient information as
follows. Denote by (s, y) the dual variables associated with the conic and the linear constraints,
and define index sets Ja (x) and J00 (
x), J0+ (
x) of active conic constraints differentiable and subd-
ifferentiable, respectively, at x
by
x) := {j : gj (
Ja ( 6= 0},
x) = 0, x (4.28)
x, s) := {j : x
J0+ ( j = 0, sj0 > 0}, (4.29)
x, s) := {j : x
J00 ( j = 0, sj0 = 0}, (4.30)
where gi is the conic constraint function. Let S Rn denote the set of previous primal-dual
x, s) to (SOCP(xkI )). Then the linear outer approximation MIP for problem (MISOCP) is
solutions (

minimize
cT x

x
subject to x X

cT x cT x, I Zp
X, x
x

0 ||xj1 ||xj0 + xTj1 xj1 , j Ja (
x), X,
x (MIP(X))

1 T
0 xj0 sj0 sj1 xj1 , j J0+ ( X,
x, s), x

0 xj0 , j J00 (
x, s), x X,

xi Z, i I.
For infeasible SOC subproblems, solutions from feasibility problems can be incorporated into
(MIP(X)) in a similar way. Convergence under a Slater constraint qualification, or alternatively by
using an additional SOCP branching step if this CQ is violated, is shown by Drewes and Ulbrich
(2012).
Polyhedral approximations of the second order cone constraint are discussed for use in an
outer approximation algorithm by Vielma et al. (2008), who use the polynomial-size relaxation
introduced by Ben-Tal and Nemirovski (2001), while Krokhmal and Soberanis (2010) generalize
this to p-order cones, that is, sets of the form {x Rn+1 : xn+1 ||x||p }. The drawback here is that
one has to choose the size of the approximation a priori. Consequently, the LP becomes large when
the approximation has to be strengthened. An iterative scheme such as SOCP-based branch-and-
cut procedure for strengthening is hence preferable. The use of a polyhedral second-order conic
constraint has been generalized by Masihabadi et al. (2011).
Two efforts in this rapidly evolving area are the work by Dadush et al. (2011a), who prove that
conic quadratic inequalities are necessary to represent the split closure of an ellipsoid, and the
work by Belotti et al. (2012), who study the convex hull of the intersection of a disjunction A B
and a generic convex set E. In general, A = {x Rn : aT x }, B = {x Rn : bT x }; this is
therefore a nonparallel disjunction, and a more general disjunction than discussed in Section 3.1.1
and unlike the previous examples. The authors prove that the convex hull is given by intersecting
E with a cone K such that K A = E A and K B = E B, where S is the frontier of set S,
if one such cone exists. The authors then provide an algorithm to find this cone for MISOCP, and
they prove that it is a second-order cone. In general, this conic cut, which is shown in two and
three dimensions in Figure 16, is proved to be more effective than the conic MIR cut presented by
and Narayanan (2010) and is discussed below.
Atamturk
x2
A B x1
(a) (b) (c)
Figure 16: Disjunctive conic cuts as generated by Belotti et al. (2012). In (a), K is the disjunctive
cone generated when intersecting the ellipsoid E with the intersection A B. In (b), a disjunction
described by two halfspaces delimited by non-parallel hyperplanes is intersected with an ellipsoid
E, and the intersection with the arising disjunctive cone, also shown in (b), returns a tighter feasible
set depicted in (c).
4.5.1 Gomory cuts for MISOCP
Drewes (2009) describes Gomory cuts for (MISOCP) based on the work by C ezik and Iyengar
(2005) for pure integer conic programs. We assume here that the bounded polyhedral set X is
described by
X = {x Rn | Ax = b, l x u} , (4.31)
with A Rmn , b Rm . The additional nonnegativity requirement li 0 holds for all i I. Then,
the following theorem describes a Gomory cut for (MISOCP).
Theorem 4.1 (Theorem 2.2.6 in Drewes (2009)). Assume that the continuous relaxation (SOCP(xkI ))
and its dual have feasible interior points. Let x
with x / Zp be a solution of (SOCP(xkI )), and let (
I s, y) be
the corresponding dual solution. Then the following cut is a valid inequality for (MISOCP),
d(ATI ( sI eT sI d(
y y) y y)T be, (4.32)
where y solves
! !
AC cC
y = . (4.33)
AI 0
y y)T b
Furthermore, if ( / Z, then (4.32) cuts off x
from the integer feasible set.
Gomory cuts are of restricted applicability because the requirement that lI 0, which turns out
to be violated frequently by MISOCP instances of practical interest. Consequently, Gomory cuts
were found to be largely ineffective for those MISOCP instances evaluated in the computational
studies presented by Drewes (2009).
As an example due to Drewes (2009), we consider the MISOCP

minimize x2

x
subject to 3x2 + x3 0

2x2 + x3 3

(4.34)

0 x1 , x2 3

x1 ||(x2 , x3 )T ||2

x1 , x2 Z,

whose SOCP relaxation has the optimal solution (3, 12 9

5 , 5 ). The Gomory cut x2 2 that can be
deduced from this point is shown in Figure 17 and cuts off the relaxed solution.
2
x1
1
0 3
2 3
1
0 2
x3 -1
1
-2 x2
-3 0
Figure 17: MISOCP example (4.34). Feasible cone (rainbow), linear constraints (green planes),
integer feasible set (blue lines), relaxed optimal solution (red asterisk), and Gomory cut cutting off
the relaxed optimal solution (red plane).
4.5.2 Lift and project cuts for MISOCP
Next, we consider the restricted class of mixed 0-1 second-order cone programs,

minimize cT x

x

subject to

x K,
Ax = b, (4.35)

l x u,

xi {0, 1} i I,

and follow the notation of the lift and project procedure described for disjunctive cuts in Sec-
tion 4.3. We are again interested in a finite conic linear description of the convex hull of the feasible
set of (4.35), and we investigate the convex hull of the union of disjoint sets C j to this end. For
MISOCP, the convex hull of C j is described by the set M j (C):

v 0 + v 1 = x,

0 + 1 = 1, 0 , 1 0

Av 0 0 b = 0, Av 1 1 b = 0

j (C) :=
M (x, v 0 , v 1 , 0 , 1 )

. (4.36)

v 0 K, v1 K

vj0 = 0, vj1 = 1

0 vk0 0 , 0 vk1 1 , k J, k 6= j

B (C) is straightforward by repetition of

If we fix a set B J, the definition of the convex hull M
the lifting process. We are now prepared to state the subgradient cut theorem from Stubbs and
Mehrotra (1999) for MISOCP:
Theorem 4.2 (Proposition 2.1.6 in Drewes (2009)). Fix a set B I, let x

B (C), and let x be the
/M
solution of
min ||x x
||2 . (4.37)
B (C)
(x,v 0 ,v 1 ,0 ,1 )M
Then the subgradient cut

)T x x T (x x
(x x ) (SGC)
B (C) that cuts off x

is a valid linear inequality for all x M .
Drewes (2009) describes further lift-and-project cuts for MISOCP, based on work by C ezik
and Iyengar (2005) and Stubbs and Mehrotra (1999) for integer SOCPs and MILPs. Similar to the
procedure described for MINLP disjunctive cuts above, these cuts can be constructed by solving
linear, quadratic, and conic auxiliary programs, and have been found to be considerably more
efficient than Gomory cuts when used in a MISCOP branch-and-cut procedure.
4.5.3 Mixed-integer rounding cuts for MISOCP
and Narayanan (2010) describe mixed-integer rounding cuts for MISOCPs. To this end,
Atamturk
we consider the following formulation of problem (MISOCP):

minimize cT x

x
subject to ||Aj x bj ||2 dTj x hj0 , 1 j k

(4.38)

x0

xi Z i I,

where Aj Rnj n are matrices, bj Rnj , dj Rn are column vectors, and hj0 R are scalars.
and Narayanan (2010) introduce the following polyhedral second-order conic constraint for-
Atamturk
mulation that allows the exploitation of polyhedral information on conic constraints of this shape.
For simplicity of exposition, we consider the case of k = 1 conic constraint of dimension n1 only.
We introduce t = (t0 , t1 , . . . , tn1 ) R1+n1 , and we denote by aTl Rn the lth row of matrix
A Rn1 n :
minimize
cT x

x,t
subject to tl |aTl x bl |,

1 l n1

T
||t||2 t0 d x h0 (4.39)

x0

x Z i I.
i
We denote by Sl the feasible set considering a single component 1 l n1 of the polyhedral conic
constraint,
Sl := {x Rn , x 0, xi Z i I, t R : t |aTl x bl |}. (4.40)
A family of valid inequalities for Sl , called conic mixed-integer rounding (MIR) inequalities, is
given by the following theorem.
Theorem 4.3 (Theorem 1 in Atamturk and Narayanan (2010)). For any 6= 0 the conic mixed integer
rounding (MIR) inequality
X

f (ai /)xi f (b/) (t + x+
R + xR )/|| (4.41)
iI

is valid for the set Sl , where x+
R and xR aggregate the real variables xR with positive and negative coefficients
alR , and f : R R is the conic MIR function for 0 f := b/ bb/c 1,
(
(1 2f )p (a p) if p a < p + f
f (a) := , n Z. (4.42)
(1 2f )p + (a p) 2f if p + f a < p + 1
Moreover, the inequalities are shown to be facet defining under certain conditions. Further-
more, such inequalities can be used efficiently to cut fractional points off the relaxation of Sl :
Theorem 4.4 (Proposition 4 in Atamturk and Narayanan (2010)). Conic mixed integer equalities with
= alj , j I are sufficient to cut off all fractional extreme points of the LP relaxation of Sl .
5 Nonconvex MINLP
Nonconvex MINLPs are especially challenging because they contain nonconvex functions in the
objective or the constraints; hence even when the integer decision variables are relaxed to be con-
tinuous, the feasible region may be nonconvex. Therefore, more work needs to be done to obtain
an efficiently solvable (convex) relaxation for use in a branch-and-bound framework.
Nonconvex MINLP is closely related to global optimization, a topic that also seeks optimal
solutions to optimization problems having nonconvex functions, although the focus in global op-
timization has often been on problems with only continuous decision variables. Huge literature
on global optimization exists, including several textbooks (Hansen, 1992; Horst and Tuy, 1993;
Horst et al., 1995; Floudas, 2000). It is outside the scope of this paper to provide a comprehensive
review of global optimization. Our focus will be on describing techniques that either explicitly
consider the integer variables appearing in an MINLP or that are essential to many algorithms for
solving nonconvex MINLPs.
One approach to solving nonconvex MINLPs is to replace the nonconvex functions with piece-
wise linear approximations and solve the corresponding approximation by using mixed-integer
linear programming solvers. We discuss this approach in Section 5.1. In the remaining sections
we discuss components of methods for directly solving nonconvex MINLPs. In Section 5.2, we
discuss generic strategies for obtaining convex relaxations nonconvex functions. We then describe
in Section 5.3 how spatial branching can be used with these relaxations to obtain a convergent al-
gorithm. Section 5.4 provides a sample of some techniques to obtain improved convex relaxations
by exploiting particular types of nonconvex structures.
We refer the reader to the works of Tawarmalani and Sahinidis (2002) and Burer and Letchford
(2012a) for additional surveys focused on nonconvex MINLP.
5.1 Piecewise Linear Modeling
A common approach for approximately solving MINLPs with nonconvex functions is to replace
the nonlinear functions with piecewise linear approximations, leading to an approximation that
can be solved by mixed-integer linear programming solvers. In fact, the importance of being able
to model such functions using binary variables was recognized by some of the earliest work in
binary integer programming (Markowitz and Manne, 1957; Dantzig, 1960). Figure 18 shows an
example of a nonconvex function with a corresponding piecewise linear approximation.
We focus our attention on modeling piecewise linear approximations of a univariate function
f : [l, u] R, where l, u R. A multivariate separable function of the form
K
X
g(x) = fi (xi )
i=1
can be approximated by separately obtaining piecewise linear approximations of fi (xi ). In section

5.1.3 we briefly introduce extensions of the piecewise linear approximation approach to more
general multivariate functions.
Figure 18: Example of a function (solid line) and a piecewise linear approximation of it (dashed
line).
Using a piecewise linear modeling technique for solving MINLPs involves two steps: obtain-
ing piecewise linear approximations of the nonlinear functions and modeling the piecewise linear
functions in a way that mixed-integer linear programming solvers can handle. We discuss these
two steps in Sections 5.1.1 and 5.1.2, respectively. In Section 5.1.3 we provide a brief overview of
how these modeling approaches can be extended to multivariate function.
Our treatment of this approach is necessarily brief. For more details on this topic, we refer
to the reader to Geissler et al. (2012), who provide a recent survey of piecewise linear modeling
in MINLP, and to Vielma et al. (2010), who provide a detailed review of methods for modeling
piecewise linear functions using binary variables.
5.1.1 Obtaining a piecewise linear approximation
Given a function f : [l, u] R, we seek to obtain a piecewise linear function f : [l, u] such that
f(x) f (x) for all x [l, u]. If the piecewise linear function f consists of d linear segments, then
it may be specified by its break points l =: b0 < b1 < < bd := u and the corresponding function
values y k = f(bk ), for k = 0, 1, . . . , d. Then the function f is given by
k
y y k1

f (x) = y +k
(x bk1 ), x [bk1 , bk ], k = 1, . . . , d. (5.1)
bk bk1
Alternatively, if for each k = 1, . . . , d, we let mk = (y k y k1 )/(bk bk1 ) be the slope of the line
segment in interval k, then ak = y k mk bk1 is the y-intercept of the line defining the line segment
in interval k. Thus we can equivalently write f as:
f(x) = ak + mk x, x [bk1 , bk ], k = 1, . . . , d. (5.2)

Obtaining a piecewise linear approximation involves two competing objectives. The first is to ob-
tain an accurate approximation, where accuracy may be measured in multiple ways. A natural
measure of accuracy is the maximum absolute difference between the function and its approxima-
tion:
max |f (x) f(x)|.
x[l,u]
The second objective is to obtain an approximation that uses fewer linear segments d. This is im-
portant because the time to solve the resulting approximate problem increases with the number of
segments used in the approximation, possibly dramatically. Therefore, one always can obtain an
improved approximation by including more segments, this has to be weighed against the compu-
tational costs.
The simplest approach for obtaining a piecewise linear approximation is to simply choose a set
of break points, for example, uniformly in the interval [l, u], and then let y k = f (bk ) for each break
point bk . This approach is illustrated in Figure 18. For a given number of break points, however,
one can to obtain significantly better approximations by choosing the location of the break points
so that parts of the function that are more nonlinear have more break points. In addition, using a
value of y k 6= f (bk ) may also yield a better approximation.
The sandwich algorithm is another simple approach for obtaining a piecewise linear approxi-
mation. This approach begins with a single linear segment approximating the function by using
break points b0 = l and b1 = u and function values y 0 = f (b0 ) and y 1 = f (b1 ). At each iteration
k 1 we have break points b0 , . . . , bk , and we use y i = f (bi ) for i = 0, . . . , k. We then select the
index i {1, . . . , k} such that the error between the function f and the current approximation
over the interval [xi1 , xi ] is largest, according to whatever error measure we are interested in.
A new breakpoint in the interval (xi1 , xi ) is then selected and added to the set of break points.
Many possible rules for selecting the the new breakpoint exist, such as choosing the point where
the error is largest or choosing the midpoint of the interval. Rote (1992) analyzes these and two
other variants of this algorithm when applied to a function f that is either convex or concave, and
he shows that using any of these variants the error after k iterations is O(1/k 2 ).
Piecewise linear approximation methods appear in diverse fields, and so it is beyond the scope
of this paper to provide a thorough review of these techniques. We instead provide a sample of
references to other approaches. Geoffrion (1977) studies methods for finding optimal piecewise
linear approximations of convex or concave functions for a given number of breakpoints. Bellman
(1961) introduces a dynamic programming approach. Boyd and Vandenberghe (2004) consider
the problem of finding a piecewise linear convex function that best approximates a given finite set
of (x, f (x)) points, where they assume the break points are given and the problem is to find the
function values at the break points. Toriello and Vielma (2012) also consider the setting in which
a finite set of (x, f (x)) data points is given and provide optimization formulations for finding
piecewise linear approximations of these points. Notably, the approach of Toriello and Vielma
(2012) does not require that the piecewise linear approximations be convex (although this can be
enforced if desired) and that one simultaneously choose the break points and the function values
of the approximation.
5.1.2 Modeling piecewise linear functions
This section describes techniques for modeling a piecewise linear function f : [l, u] R, as defined
in (5.1). We introduce a variable y that will be used to model the value of this function. That is, we
provide formulations that enforce the condition
y = f(x).
We then use y in place of the function f (x) anywhere it appears in the optimization model. In our
development, we assume the piecewise linear function f is continuous, although many of these
formulations can be extended to lower semicontinuous piecewise linear functions (Vielma et al.,
2010). In describing the different approaches, we follow the names used by Vielma et al. (2010).
The first approach we present is the multiple choice model, which uses the representation of

f (x) given in (5.2). In this model, a set of binary variables zk , k = 1, . . . , d, is introduced, where
zk = 1 indicates that x is in intervral k, [bk1 , bk ]. In addition, for each interval k, a variable wk is
introduced, where wk will be equal to x if x is in interval k, and wk = 0 otherwise. The model is
as follows:
d
X d
X d
X
wk = x, (mk wk + ak zk ) = y, zk = 1, (5.3a)
k=1 k=1 k=1
k1 k
b z k wk b z k , zk {0, 1} k = 1, . . . , d. (5.3b)
This model was introduced by Jeroslow and Lowe (1984) and has been studied by Balakrishnan
and Graves (1989) and Croxton et al. (2003). In computational experiments conducted by Vielma
et al. (2010), this model frequently yielded the best computational performance when the number
of break points was small or moderate (e.g., d 16).
The second model is the disaggregated convex combination model, which uses the representation

of f (x) given in (5.1). In this model, a set of binary variables zk , k = 1, . . . , d, is introduced, where
zk = 1 indicates that x [bk1 , bk ]. For each interval, this model introduces a pair of continuous
variables k and k , which are zero if zk = 0, but otherwise describe x as a convex combination
of endpoints of the interval [bk1 , bk ]. Consequently, the function value y = f(x) can then be
described as the same convex combination of y k1 and y k . The formulation is as follows:
d
X d
X
(k bk1 + k bk ) = x, (k y k1 + k y k ) = y, (5.4a)
k=1 k=1
d
X
zk = 1, k + k = zk , k = 1, . . . , d, (5.4b)
k=1
k 0, k 0, zk {0, 1}, k = 1, . . . , d. (5.4c)
The constraints (5.4b) enforce that exactly one interval has zk = 1 that for this interval the variables
k and k sum to one, and that i = i = 0 for all other intervals i 6= k. Thus, the constraints (5.4a)
enforce that x and y be written as a convex combination of the end points of the selected interval,
and the function values at those end points, respectively. This approach has been presented in
(Meyer, 1976; Jeroslow and Lowe, 1984, 1985; Croxton et al., 2003; Sherali, 2001).
The next formulation is the convex combination model, also sometimes called the lambda method.
In this formulation, a single continuous variable is introduced for each break point. These vari-
ables are used to express x and y as a convex combination of the break points, and their function
values, respectively. As in the disaggregated convex combination model, binary variables are in-
troduced to determine which interval x lies in. These variables are then used to ensure that only
the convex combination variables associated with the end points of this interval are positive. The
formulation is as follows:
d
X d
X
k bk = x, k y k = y, (5.5a)
k=0 k=0
Xd d
X k1
X k
X
j zj , j zj , k = 1, . . . , d, (5.5b)
j=k j=k j=0 j=1
d
X d
X
k = 1 zk = 1, (5.5c)
k=0 k=1
k 0, k = 0, 1, . . . , d zk {0, 1}, k = 1, . . . , d. (5.5d)
The constraints (5.5c) enforce that when zk = 1, j = 0 for all j

/ {k 1, k}. In most presentations
of the convex combination model (Dantzig, 1960, 1963; Jeroslow and Lowe, 1985; Nemhauser and
Wolsey, 1988b; Wilson, 1998; Lee and Wilson, 2001; Vielma et al., 2010; Geissler et al., 2012) this
condition is instead formulated by using the following constraints:
0 z1 , d zd , k zk + zk+1 , k = 1, . . . , d 1. (5.6)
However, Padberg (2000) demonstrated that the formulation using (5.6) allows more continuous
solutions when the binary constraints on the zk variables are relaxed, which makes (5.5a) prefer-
able for computational purposes.
The next model we present is the incremental model, sometimes referred to as the delta method,
which was originally proposed by Markowitz and Manne (1957). In this model, continuous vari-
ables k , for each interval k = 1, . . . , d are introduced to determine what portion of interval k the
argument x has filled. Binary variables are introduced to enforce the condition that the intervals
are filled in order. This leads to the following formulation:
d
X d
X
b0 + k (bk bk1 ) = x, y0 + k (y k y k1 ) = y, (5.7a)
k=1 k=1
k+1 zk k , k = 1, . . . , d 1, (5.7b)
1 1, d 0, zk {0, 1}, k = 1, . . . , d 1 (5.7c)
If i < 1 for some i, then constraints (5.7b) combined with the binary restrictions on the z variables
enforce that zi = 0 and i+1 = 0 and then recursively that zj = j+1 = 0 for all j > i, which
is precisely the condition that the intervals should be filled in order. As pointed out by Padberg
(2000), the convex combination formulation (5.7) and the incremental formulation (5.7) are related
by a simple change of variables (i.e., k = dj=k j , k = 1, . . . , d and similarly for the respective zk
P
variables).
The fifth model we describe does not use binary variables at all. Instead, it uses the concept
of a special ordered set of variables of type II (SOS2) (Beale and Tomlin, 1970; Tomlin, 1981; Beale
and Forrest, 1976; Keha et al., 2006). An ordered set of variables = (0 , 1 , . . . , d ) is said to be
SOS2 if at most two of the variables in the set are nonzero and the nonzero variables are adjacent.
This is exactly the condition that the binary variables in the convex combination model (5.5) are
used to enforce. Therefore, by instead explicitly stating this condition, the following formulation
is obtained:
Xd Xd
k
k b = x, k y k = y, (5.8a)
k=0 k=0
Xd
k = 1, (5.8b)
k=0
k 0, k = 0, 1, . . . , d (0 , 1 , . . . , d ) is SOS2. (5.8c)
The condition that an ordered set of variables is SOS2 can be declared in most commercial mixed-
integer linear programming solvers (IBM Ilog CPLEX, 2009; Gurobi, 2012; FICO Xpress, 2009),
similar to declaring that an individual variable is integer or binary. This condition is relaxed to
obtain a linear programming relaxation and is progressively enforced through branching, just
as integer restrictions are. The main difference is how the branching is done. In an LP relaxation
then an index k {1, . . . , d}
solution, if the SOS2 condition is violated by the relaxation solution ,
is selected such that there exists an index j1 < k with j1 > 0 and also an index j2 > k with j2 > 0.
Then, two branches are created: one that enforces j = 0 for all j < k and the other that enforces
j = 0 for all j > k. Every solution that satisfies the SOS2 condition is feasible to one of the two
branching conditions, and hence this branching strategy does not exclude any feasible solutions.
In addition, the current relaxation solution, which violates the SOS2 condition, is eliminated from
both branches, enabling the relaxation bound to improve and ensuring that the SOS2 condition
will be satisfied after a finite number of branches. One also can derive valid inequalities based
on the SOS2 condition, analogous to the use of valid inequalities for mixed integer programming
(Keha et al., 2006).
Vielma and Nemhauser (2011) have recently developed a novel approach for modeling piece-
wise linear functions using binary variables. The interesting feature of this approach is that the
number of binary variables required is only logarithmic in the number of segments of the linear
function. In contrast, the binary formulations we presented here all require one binary variable for
each segment. Note, however, that the number of continuous variables required in the formula-
tions of Vielma and Nemhauser (2011) is still linear in the number of segments. The computational
results of Vielma and Nemhauser (2011) suggest that this approach is most beneficial when mod-
eling multivariate piecewise linear functions.
5.1.3 Multivariate functions
One approach to using piecewise linear models for problems with multivariate functions is to
attempt to reformulate the problem in such a way that only univariate functions appear, and then
apply the univariate modeling techniques. We already mentioned separable functions of the form
PK
g(x) = i=1 fi (xi ) as one example where this approach can work well. More generally, if an
algebraic description of a multivariate function g is available, then the techniques described in
Section 5.2.1 can be used to first obtain a reformulated problem that contains only univariate and
bivariate functions and then construct piecewise linear approximations of these functions. Using
further transformations, one may be able to eliminate even the bivariate functions. For example,
suppose we have a constraint
y = x1 x2
in our model, where y, x1 , and x2 are all decision variables, and because of other constraints in the
model we know that x1 > 0 and x2 > 0 in any feasible solution. Then, this constraint is equivalent
to
ln(y) = ln(x1 ) + ln(x2 ).
Each of the functions ln(y), ln(x1 ), and ln(x2 ) can then be approximated by using univariate piece-
wise linear models.
The approach of reducing multivariate functions to univariate functions has a few potential
disadvantages, which have been pointed out, for example, by Tomlin (1981); Lee and Wilson
(2001) and Vielma et al. (2010). First, it may not always be easy to find a systematic approach
that reduces a given model to one that contains only univariate functions. In our example above,
we required x1 > 0 and x2 > 0 in order for the log transformation to be valid. If this was not
the case, an alternative would be required. Second, and probably more important, this process of
reformulation may allow the errors introduced by the piecewise linear approximations to accu-
mulate and amplify, potentially leading to an approximate model that either is too inaccurate to
be useful or requires so many break points in the piecewise linear approximations that it becomes
intractable to solve. Finally, in some cases a function is not given analytically. Instead, we may
have only an oracle, or black box, that allows us to obtain function evaluations at given points,
for example, by running a complex simulation or even a physical experiment. In this case, we
may wish to obtain a piecewise linear function that approximates the data obtained at a set of trial
points.
One also can directly model nonseparable multivariate piecewise linear functions using bi-
nary variables. We refer the reader to the works of Vielma et al. (2010) and Geissler et al. (2012) for
details of these techniques. In particular, the convex combination method has been extended to
multivariate functions by Lee and Wilson (2001), and the incremental method has been extended
by Wilson (1998). Tomlin (1981) also extended the notion of special ordered sets for modeling mul-
tivariate piecewise linear functions. DAmbrosio et al. (2010) provide an interesting alternative for
approximating multivariate piecewise linear functions, in which the piecewise linear approxima-
tion is not fixed a priori, but is instead allowed to be chosen optimistically by the optimization
* ^
x1 log x2 3
x2
Figure 19: The expression tree of f (x1 , x2 ) = x1 log(x2 ) + x32 . Leaf nodes are for variables x1 and
x2 and for the constant 3.
algorithm. This modification enables a relatively compact formulation to be derived. We note,

however, that the number of pieces requires to obtain an acceptable piecewise linear approxima-
tion of a given nonlinear function may grow exponentially in the number of arguments to the
function. Hence, all these approaches are, in practice, limited to functions with at most a few
arguments.
5.2 Generic Relaxation Strategies
For general nonconvex MINLP problems, methods for finding a relaxation exploit the structure
of the problem. For a broad class of MINLP problems, the objective function and the constraints
are nonlinear but factorable, in other words, they can be expressed as the sum of products of unary
functions of a finite set Ounary = {sin, cos, exp, log, ||} whose arguments are variables, constants, or
other functions, which are in turn factorable. In other words, a factorable function can be written
by combining a finite number of elements from a set of operators O = {+, , /,, sin, cos, exp, log, |
Rx
|}. This excludes, among others, integral functions x0 h(x)dx whose antiderivative is unknown
and black-box functions, whose value can be computed by running, for instance, a simulation. The
approach described below is therefore suitable when the symbolic information about the objective
function and constraints, that is, their expressions, is known.
Factorable functions can be represented by expression trees. These are n-ary arborescences
whose leaves are constants or variables and whose non-leaf nodes are, in general, n-ary operators
whose children are the arguments of the operator, and which are in turn expression trees (Cohen,
2003). The expression tree of the function f (x1 , x2 ) = x1 log(x2 ) + x32 is depicted in Figure 19.
5.2.1 Relaxations of factorable functions
If the objective and the constraints of an MINLP of the form (1.1) are factorable, the problem
admits a representation where the expression trees of the objective function and all constraints
are combined. The root of each expression is one among c1 (x), c2 (x), . . . , cm (x), or f (x), and is
associated with a lower and an upper bound: [, 0] for ci (x), i = 1, 2, . . . , m, and [, ] for
+ + +
^ 4 sin * ^ 5
2 3
x1 x2
Figure 20: The DAG associated with the problem in (5.9). Each node without entering arcs is
associated with the root of the expression tree of either a constraint or the objective function. In
common with all constraints and the objective are the leaf nodes associated with variables x1 and
x2 .
f (x), where is the objective function value of a feasible solution for (1.1), if available, or . The
leaf nodes of all expression trees are replaced by a unique set of nodes representing the variables
x1 , x2 , . . . , xn of the problem. The result is a directed acyclic graph (DAG).
Example. Consider the following MINLP:
min x1 + x22
s.t. x1 + sin x2 4
x1 x2 + x32 5 (5.9)
x1 [4, 4] Z
x2 [0, 10] Z.
The DAG of this problem has three nodes without entering arcs (one for the objective function
and two for the constraints) and six leaf nodes: two for the variables x1 and x2 and four for the
constants 2, 3, 4, and 5. It is represented in Figure 20.
Factorable problems allow a reformulation of the problem (1.1) (McCormick, 1976; Smith and
Pantelides, 1997; Tawarmalani and Sahinidis, 2002). The reformulation is another MINLP as fol-
lows:

minimize xn+q

x

subject to x = (x)

k k k = n + 1, n + 2, . . . , n + q
li xi ui i = 1, 2, . . . , n + q (5.10)

x X,

xi Z, i I,

where k is an operator of the set O introduced above. The bound on all variables are written
explicitly here for the sake of clarity, but they are included in the definition of X; we will use this
notation throughout this section. The reformulation contains a set of q new variables known as
auxiliary variables (or more simply auxiliaries). By convention, the last auxiliary variable replaces
the objective function. Each of these variables is constrained to be equal to a function (x) such
that O. The lower and upper bounds on each auxiliary variable xk , as well as its integrality
constraint, depend on the operator k associated with it and on the bounds on the arguments of
k (and the integrality of these arguments).
Example. The MINLP shown in (5.9) admits the following reformulation:
min x9
s.t. x3 = sin x2 x7 = x5 + x6 5 0 x2 10 0 x6 1000
x4 = x1 + x3 4 x8 = x22 1 x3 1 45 x7 0
x5 = x1 x2 x9 = x1 + x8 9 x4 0 0 x8 100
x6 = x2 3 4 x1 4 40 x5 40 4 x9 104
x1 , x2 , x5 , x6 , x7 , x8 , x9 Z.
Note that x3 , the auxiliary associated with sin x2 , has bounds [1, 1] because x2 [0, 10], while
x7 := x5 + x6 5, with bounds [45, 1035], is further constrained by the right-hand side of the
constraint x1 x2 + x32 5. This variable is indeed associated with the root of the expression tree
of c2 (x) = x1 x2 + x32 5. The integrality of an auxiliary also depends on whether the function
associated with it can return only integer values given the integrality constraints of its arguments.
In this case, x3 and x4 are not constrained to be integer, while all other variables are because of the
integrality constraint on x1 and x2 .
Problems (5.10) and (1.1) are equivalent in that for any feasible (resp. optimal) solution x
R n+q of (5.10) one can obtain a feasible (resp. optimal) solution x n
R of (1.1) and vice versa.
Although still a nonconvex MINLP, (5.10) makes it easier to obtain a convex relaxation of (1.1).
Consider the nonconvex sets
k = {x Rn+q : xk = k (x), x X, l x u, xi Z, i I}, k = n + 1, n + 2, . . . , n + k.
Note that k are, in general, nonconvex because of the equation xk = k (x) and the integrality
constraints. Suppose that a convex set k k exists for each k = n + 1, n + 2, . . . , n + k. Then the
following is a convex relaxation of (5.10) intersects k for k = n + 1, n + 2, . . . , n + q, and hence it
is a relaxation of (1.1) (note that the integrality constraints are also relaxed):

minimize xn+q

x
k

subject to x k = n + 1, n + 2, . . . , n + q

li xi ui i = 1, 2, . . . , n + q

x X.

Convex sets k are generally polyhedral, that is, they are described by a system of mk linear
inequalities:
k = {x Rn+q : ak xk + B k x dk , x X, l x u},

xk xk
xk xk
li xi
ui
xj
xi xi
li ui li ui
xi
(a) (b) (c) (d)
Figure 21: Polyhedral relaxations k for several univariate and bivariate operators: xk = x2 ,
i
xk = x3i , xk = xi xj , and xk = x2i with xi integer. Note that these relaxations are exact at the bounds
on xi and xj .
where ak Rmk , B k Rmk (n+q) , and dk Rmk . Hence several practical MINLP solvers for (1.1)
use the following linear relaxation to obtain a lower bound:

minimize xn+q

x
subject to ak xk + B k x dk k = n + 1, n + 2, . . . , n + q

li xi ui i = 1, 2, . . . , n + q

x X.

Although finding a polyhedral superset of k is nontrivial, for each operator of a finite set O we
can define a method to find one. We provide two examples. LP relaxations for monomials of
odd degree such as xk = x2p+1 i , with p Z+ , were proposed by Liberti and Pantelides (2003).
For products of two variables xk = xi xj , the following four inequality proposed by McCormick
(1976) provide the convex hull of the set k = {(xi , xj , xk ) : xk = xi xj , (li , lj , lk ) (xi , xj , xk )
(ui , uj , uk )}, as proved by Al-Khayyal and Falk (1983):
xk lj xi + li xj li lj xk lj xi + ui xj ui lj
(5.11)
xk uj xi + ui xj ui uj xk uj xi + li xj li uj .
Figure 21 shows polyhedral relaxations for xk = x2i , xk = x3i , and xk = xi xj . If the argument
xi of a function k is integer, the linear relaxation can be strengthened. Suppose xj = (xi )2 and
xi Z [2, 1], as in Figure 21(d). Then we can add linear inequalities that can be violated by
points ( 2 ) if x
x, x / Z.
Convex relaxations k of the univariate or multivariate operator k are required to be exact at
the frontier of the bound interval of the arguments of k . In particular, let xi1 , xi2 , . . . , xik be the
arguments of k , each of them bounded by the interval [lij , uij ]. Define for each j = 1, 2, . . . , k a
value bij {lij , uij }. Then
k {x Rn : xi = bi } = k {x Rn : xi = bi }.
j j j j
A direct consequence is the convergence result shown in Section 5.3.
5.2.2 Disjunctive cuts
In the class of convex MINLPs, the only nonconvex constraints are represented by integer vari-
ables, and these nonconvexities are resolved by integer branching, which represents a specific class
of disjunctions. There are classes of nonconvex MINLP whose nonlinear objective and constraint
introduce a different type of disjunction that can be used to create a valid cut. Disjunctive cuts
for several classes of convex MINLP problems have been discussed in Section 4.3. For nonconvex,
factorable MINLP, disjunctions arise from branching rules xi b xi b on continuous variables.
Other disjunctions that have important application arise from complementarity constraints, that is,
constraints of the form xi xj = 0 that are equivalent to the disjunction xi = 0 xj = 0. These

constraints are the basis of disjunctive cuts developed by Judice et al. (2006).
Disjunctive cuts can be developed by using the branching disjunction xi b xi b in
MINLPs with factorable functions. Suppose a branching rule is enforced because an auxiliary
variable xk and its related nonconvex constraint xk = k (x) are such that x k 6= k (x) for a given
LP solution x . A branching rule allows us to refine the LP relaxation of a subproblem as shown
in Figure 22. Figure 22(a) depicts the set j for xj = x2 and shows that x
j 6= x2i . The disjunction
i
xi b xi b can be used to create two new subproblems, NLP(l , u ) and NLP(l+ , u+ ), and
consequently two tighter LP relaxations, LP(l , u ) and LP(l+ , u+ ), shown in Figure 22(b); note
that both relaxations exclude ( j ). For an example with a different function xj = exi , see Figure
xi , x
22(c). Belotti (2012b) creates disjunctive cuts by means of a cut generating LP (CGLP) (Balas et al.,
1993), a linear optimization problem used to devise a disjunctive cut that maximizes the violation
. The CGLP obtains a disjunctive cut that is valid for both LP(l , u ) and LP(l+ , u+ ) and
w.r.t. x
that hence exploits the disjunction to eliminate infeasible solutions without actually creating two
subproblems.
This procedure retains the pros and cons of its MILP version: although it allows one to avoid
creating two new subproblems and can be effective, each of these cuts is generated by solving very
large LPs and can be computationally expensive for large MINLPs.
5.3 Spatial Branch-and-Bound
The best-known method for solving nonconvex MINLP problems is branch-and-bound (BB). Most
of the global optimization community usually refers to these methods as spatial BB (sBB). As out-
lined in Section 3, a BB method is an implicit enumeration technique that recursively partitions
the feasible set. This partitioning yields two or more subproblems whose solution sets are, ideally,
disjoint from one another, in order to avoid evaluating a feasible solution in more than one sub-
problem. Most BB implementations for MINLP use the reformulation scheme outlined in Section
5.2.1 to obtain a lower bound (Sahinidis, 1996; Ryoo and Sahinidis, 1996; Tawarmalani and Sahini-
dis, 2002; Ryoo and Sahinidis, 1995; Tawarmalani and Sahinidis, 2004; Smith and Pantelides, 1997;
Belotti et al., 2009).
Other approaches employ a relaxation technique called -convexification (Androulakis et al.,

1995). For a nonconvex quadratic function f (x) = xT Qx + cT x with x [l, u], a valid lower bound
is provided by
n
f(x) = xT Qx + cT x +
X
(xi li )(xi ui ).
i=1
Note that f(x) f (x) for x [l, u] and that f(x) can be rewritten as a quadratic function
xT P x + dT x where P = Q + I, and f(x) is convex if P 0. Therefore it suffices to set
= MIN (Q), namely, the opposite of the minimum eigenvalue of Q, to obtain a convex re-
laxation. Some implementation of the -convexification adopt a quadratic approximation of the
objective and the constraints and hence guarantee optimality only for quadratic problems (Nowak
et al., 2003). This method can be extended to nonquadratic functions while preserving the validity
of the lower bound. A generalization of this method is at the base of the MINLP solver GloMIQO
(Misener and Floudas, 2012b).
A BB algorithm requires (i) a procedure to compute a lower bound on the optimal objective
function value of a subproblem and (ii) a procedure for partitioning the feasible set of a subprob-
lem. In general, the former consists of obtaining a convex relaxation of a subproblem NLP(l, u)
and solving it to optimality, while the latter generates two new subproblems, NLP(l , u ) and
NLP(l+ , u+ ), by appropriately setting new variable bounds. We discuss these techniques in detail
below.
The structure of a branch-and-bound for nonconvex MINLP follows the scheme described in
Section 3, and we will not repeat it here. The generic subproblem NLP(l, u) of the BB can be
defined as a restriction of the original MINLP (1.1) as follows:

minimize f (x),

x

subject to c(x) 0,

xX (5.12)

li xi ui i = 1, 2, . . . , n

x Z, i I.

i
At subproblem NLP(l, u), the BB algorithm seeks a lower bound of the optimal value of f (x)
by solving a convex relaxation such as the LP relaxation LP(l, u):

minimize xn+q

x
subject to ak xk + B k x dk k = n + 1, n + 2, . . . , n + q

(5.13)

li xi ui i = 1, 2, . . . , n + q

x X.

Suppose an optimal solution x

of LP(l, u) is found. If x
is feasible for (5.12) and hence for (1.1), sub-
problem NLP(l, u) can be eliminated. If x is infeasible for (5.12), then at least one of the following
two holds:
is not integer feasible, i.e., i I : x

1. x i
/ Z.
2. At least one of the continuous nonconvex constraints of the reformulation is violated, that
is,
k {n + 1, n + 2, . . . , n + q} : x
k 6= k (
x).
In the first case, one can generate two new subproblems, NLP(l , u ) and NLP(l+ , u+ ), whose
feasible sets F(l , u ) and F(l+ , u+ ) are amended new bounds on xi through the branching rule
xi bxi c xi dxi e. In the second case, branching may be necessary on a continuous variable.
In that case, suppose that xi is among the arguments of function k . Then the branching rule
xi x
i xi xi creates two new subproblems whose feasible sets have a nonempty intersection,
where xi = x i . This constitutes a strong point of departure with the subclass of pure integer
nonconvex MINLPs, where all variables are integer, and with convex MINLP discussed in Section
3. For these subclasses, branching is necessary only on integer variables. Finite bounds on integer
variables ensures finite termination of the branch-and-bound algorithm.
Consider the feasible set of a subproblem NLP(l, u) of (1.1):
F(l, u) = {x [l, u] : ci (x) 0 i = 1, 2, . . . , m, x X, xi Z, i I},
whose only difference from the feasible set of (1.1) is new bounds on the variables, as dictated
by the branching rules. Consider a branching rule on a continuous variable xi , subdividing the
feasible set of subproblem NLP(l, u) into two subproblems, NLP(l , u ) and NLP(l+ , u+ ), with
feasible sets F(l , u ) and F(l+ , u+ ). A bounding operation yields two subproblems, NLP(l , u )
and NLP(l+ , u+ ), by applying a branching rule, and lower bounds F (l ,u ) , F (l+ ,u+ ) and upper
bounds F (l ,u ) , F (l+ ,u+ ) for the new subproblems. Such a bounding operation is said consistent
if, at every step, subsets F(l , u ) and F(l+ , u+ ) either are fathomed or can be further refined in
such a way that, for any infinite sequence {Fh } resulting from applying bounding operations, one
can guarantee that (Horst and Tuy, 1993)
lim Fh Fh = 0.
h
In addition, a bounding operation is finitely consistent if any sequence {Fh } of successively refined
partitions of F is finite. Branching on continuous variables does not imply directly that a finite
number of branching rules will be used, yet both in theory and in practice BB algorithms do have
finite termination properties, as shown by McCormick (1976) and Horst and Tuy (1993).
Theorem 5.1 (McCormick (1976); Horst and Tuy (1993)). If the bounding operation in the BB algorithm
is finitely consistent, the algorithm terminates in a finite number of steps.
The value of this result can be made clearer if one considers an MINLP with even just one
continuous variable x1 : by branching only on integer variables (in a finite number of BB nodes if
all integer variables are bounded), one eventually obtains a possibly nonconvex continuous opti-
mization problem. Therefore, branching will become necessary on the continuous variable x1 as
well, although termination is no longer guaranteed by integrality. The result by McCormick (1976)
states that convergence is still ensured as long as the bounding operation is finitely consistent.
5.3.1 Spatial branching
Partitioning the feasible set of a subproblem NLP(l, u) yields h 2 new subproblems NLP(l0 , u0 ),
NLP(l00 , u00 ), . . . , NLP(l(h) , u(h) ), whose lower bounds NLP(l0 ,u0 ) , NLP(l00 ,u00 ) , . . . , NLP(l(h) ,u(h) ) are
not smaller than that of NLP(l, u). We will assume w.l.o.g. that two new problems NLP(l , u ) and
NLP(l+ , u+ ) are created. Most practical implementation adopt a variable branching xi b xi b.
The performance of the BB algorithm depends strongly on the choice of i and b (Belotti et al., 2009;
Tawarmalani and Sahinidis, 2002). An integer variable is obviously a candidate for selection as
a branching variable if its value is fractional in the LP solution. In the remainder of this section,
we assume that all integrally constrained variables are integer and hence no branching is possible
(integer branching has been discussed in Section 3) and that branching is done because of an
auxiliary xk such that x k 6= k ( x).
An ideal choice of i should balance more than one objective: it should (i) increase both lower
bounds NLP(l ,u ) and NLP(l+ ,u+ ) ; (ii) shrink both feasible sets F(l , u ) and F(l+ , u+ ); and (iii)
allow for a balanced BB tree, among other criteria.
Suppose an optimal solution x of the LP relaxation LP(l, u) is found. A continuous variable xi
is a candidate for branching if it is not fixed (i.e., its lower and upper bound do not coincide), it
is an argument of a function k (x) associated with an auxiliary variable xk , and x k 6= k (x). For
k 6= x
example, if xk = k (x) = xi xj , x i x
j , and li < ui , then xi is a candidate for branching.
Upon branching, the two generated subproblems each will obtain a lower bound by solving
two tighter relaxations than that of their ancestor. A geometrical intuition is provided in Figure 22.
Suppose the auxiliary xj is defined as xj = j (xi ) = (xi )2 and xi [li , ui ] for this subproblem.
Because the LP solution x is such that x xi )2 (see Figure 22(a)), one can generate two new
j 6= (
subproblems using the branching rule xi bxi b. The new linear relaxations are the polytopes
in Figure 22(b), which are disjoint except for the point (b, b2 ) and which exclude the point (xi , x
j ).
Figure 22(c) provides a similar example for xj = j (xi ) = e . x i
Tawarmalani and Sahinidis (2002) introduce violation transfer (VT) as a variable selection tech-
nique. VT identifies the variable xi that has the largest impact on the violation of the nonconvex
constraints xk = k (x) for all k = 1, 2, . . . , n+q such that xi is an argument of k . Strong branching,
pseudocost branching, and reliability branching, discussed in Section 3, can be applied with lit-
tle modification to nonconvex MINLP and have been implemented in nonconvex MINLP solvers
(Belotti et al., 2009).
The choice of branching point is also crucial and differs from integer branching in that one has
the freedom of choosing a branching point for variable xi that can differ from x i . A branching rule
should ensure that x + +
is infeasible for both LP(l , u ) and LP(l , u ); hence the sole branching rule
xi x i will not suffice. However, the refined linear relaxations LP(l , u ) and LP(l+ , u+ )
i xi x
will be obtained by adding linear inequalities that are violated by x . While the linear inequalities
depend on the new bounds on xi , setting the branch point to a suitable b 6= xi does not prevent
one from excluding x .
xj
xj xj
(x^ i ,x^ j) (x^ i ,x^ j)
xi xi
xi
li ui li b ui li b ui
(a) (b) (c)
Figure 22: Polyhedral relaxations upon branching: In (a), the set k is shown with the components
(
xi , x
j ) of the LP solution. Branching on xi excludes the LP solutionsee (b). In (c), the LP relax-
ation before and after branching is shown for xj = exi in lighter and darker shade, respectively.
5.3.2 Bounds tightening
Bounds tightening (also referred to as bounds reduction or domain reduction) is a class of algo-
rithms aiming at reducing the bound intervals on the variables of (1.1). Although these algorithms
are optional in a BB solver, they are crucial to obtaining an optimal solution in reasonable time and
therefore are implemented in the vast majority of MINLP solvers.
Their importance is directly connected to the LP relaxation (5.13): the tighter the variable
bounds, the tighter the linear polyhedra k for each auxiliary variable xk and hence the better
the lower bound on the objective. Some MINLP solvers use bound reduction as the sole means
of obtaining a lower bound on the optimal objective function value of the subproblem (Messine,
2004).
The usefulness of bound reduction is not limited to a tighter bound hyperrectangle: as a result
of bound reduction, the feasible set might become empty, or the lower bound ln+q on xn+q is above
the cutoff value of the problem, namely, the objective function value of a feasible solution of (1.1).
In these two cases, the procedure proves that the current node is infeasible and can be fathomed
without the need of computing a lower bound by solving the convex relaxation.
Consider again the feasible set of a MINLP problem:
F = {x [l, u] : ci (x) 0 i = 1, 2, . . . , m, x X, xi Z, i I},
and suppose that a feasible solution of (1.1) is known with value z. For each variable xi , i =
1, 2, . . . , n, valid (and possibly tighter) lower and upper bounds are given by
li0 = min{xi : x S, f (x) z}; u0i = max{xi : x S, f (x) z}. (5.14)
Solving the 2n optimization problems above would yield tighter bounds, but these problems can
be as hard as problem (1.1) itself. The two most important bound reduction techniques are feasi-
bility based (FBBT) and optimality-based bound tightening (OBBT). Other commonly used techniques
are probing and reduced cost tightening; we present these techniques below.
Feasibility-based bound tightening. FBBT has been used in the artificial intelligence literature
(Davis, 1987) and is a strong component of constraint programming solvers. It is part of nonlinear
optimization solvers (Messine, 2004) as well as of MILP solvers (Andersen and Andersen, 1995;
Savelsbergh, 1994).
FBBT works by inferring tighter bounds on a variable xi as a result of a changed bound on
one or more other variables xj that depend, directly or indirectly, on xi . For example, if xj = x3i
and xi [li , ui ], then the bound interval of xj can be tightened to [lj , uj ] [li3 , u3i ]. Vice versa, a
tightened bound lj0 on xj implies a possibly tighter bound xi , namely, li0 = 3 lj . Another example
p
is given by xk = xi xj , with (1, 1, 0) (xi , xj , xk ) (5, 5, 2). Lower bounds li = lj = 1 imply a

tighter lower bound lk = li lj = 1 > 0, while the upper bound uk = 2 implies that xi uljk and
xj ulik , and hence u0i = u0j = 2 < 5.
Perhaps the best-known example applies to affine functions. Suppose xk is an auxiliary vari-
able defined as xk = a0 + nj=1 aj xj , with k > n. Suppose also that J + = {j = 1, 2, . . . , n : aj > 0}
P
and J = {j = 1, 2, . . . , n : aj < 0}. Then valid bounds on xk are

X X X X
a0 + aj uj + aj lj xk a0 + aj lj + aj uj .
jJ jJ + jJ jJ +
Moreover, explicit bounds [lk , uk ] on xk imply new (possibly tighter) bounds lj0 , u0j on xj , j =
1, 2, . . . , n : aj 6= 0:

j : aj > 0, lj0 = a1j lk a0 + iJ + \{j} ai ui + iJ ai li
P P

u0j = a1j uk a0 + iJ + \{j} ai li + iJ ai ui
P P
(5.15)
j : aj < 0, lj0 = a1j uk a0 + iJ + ai li + iJ \{j} ai ui
P P

u0j = a1j lk a0 + iJ + ai ui + iJ \{j} ai li .
P P
These implied bounds are commonly used as a preprocessing technique (Andersen and Andersen,
1995) prior to solving MILP problems. For both MILP and MINLP problems, bound reduction can
be obtained by using pairs of inequality (Belotti, 2012a), specifically through the convex combina-
tion of two inequalities aT x and bT x using a parameter [0, 1]. The resulting inequality
(a + (1 )b)T x + (1 ) yields bounds on the variables similar to (5.15), but these are a
function of and can be shown to be tighter than those obtained by single inequalities.
For the general nonlinear case, variable bounds are propagated by using the DAG of the prob-
lem. For instance, consider problem (5.9) and bounds [4, 4] and [0, 10] on x1 and x2 , respectively.
We rewrite the DAG of this problem to reflect these bounds and to show each auxiliary variable
next to the root of the expression tree associated to it; see Figure 23.
If a solution x
is found with f (
x) = 10, then an upper bound on the objective function x9 :=
x1 + x8 and the lower bound on x1 4 imply that x8 14 < 100. This in turn is propagated
x7
+ x9 + x4 +
^ x8 4 sin x 3 * x5 ^ x6 5
2 3
x 1 [0,4] x 2 [10,10]
Figure 23: Association between auxiliary variables and the nodes of the DAG related with the
problem in (5.9).

to the expression x8 = x22 , which implies that 14 x2 14, thus tightening x2 . No other
variables are tightened because of to the new cutoff. In the general case, this procedure propagates
throughout the DAG of the problem and repeats while there are tightened bounds, terminating
when no more bound is reduced.
FBBT algorithms allow for fast implementation and are commonly used in problems even of
very large size. However, they may exhibit convergence issues even at very small scale: consider
the trivial problem min{x1 : x1 = x2 , x2 = x1 , x1 [1, 1]}, with R \ {0, 1}. Although
by inspection one can see that the only feasible solution x = (0, 0) is also optimal, FBBT will not
terminate in a finite number of steps. In fact, a first pass will tighten x2 to [ 1 , 1 ]; this will trigger
a reduction of the bound interval of x1 to [ 12 , 12 ], which in turn will propagate to yield new
bounds [ 13 , 13 ] on x2 . This procedure does not terminate unless tolerances or iteration limits are
imposed, and it does not achieve its fixed point in finite time. A linear optimization problem has
been proposed for MINLP that achieves the fixed point of a FBBT algorithm applied to the linear
relaxation of (1.1) (Belotti et al., 2010).
Optimality-based bounds tightening. Solving problems (5.14) is impractical because of the non-
convexity of their feasible set, which is the same feasible set of (1.1). A more practical approach
considers the feasible set of a convex relaxation of (1.1), such as the one in (5.13):

a k x + B k x dk k = n + 1, n + 2, . . . , n + q

k
F(l, u) = x Rn+q : li xi ui i = 1, 2, . . . , n + q .

xX

Then the following are valid bounds on variable xi :
li0 = min{xi : x F(l, u), f (x) z}; u0i = max{xi : x F(l, u), f (x) z}.
Empirical evidence shows that this technique is effective in obtaining tight bounds (Belotti et al.,
2009), but it requires solving 2n linear programming problems. Thus, its use is limited to the root
node or to the nodes of small depth.
Probing and reduced-cost bounds tightening. Consider the bounds [li , ui ] on a variable xi as
defined in (5.13), and set the upper bound to a fictitious value u0i < ui , regardless of whether u0i is
valid. Then apply a bound-tightening procedure such as FBBT. If the procedure indicates that the
tightened bound hyperrectangle renders the problem infeasible or drives the lower bound ln+q on
xn+q above the cutoff value, then we have a proof that no optimal solution exists with xi [li , u0i ],
and the bounds on xi become [u0i , ui ].
The same procedure can be repeated by imposing a fictitious lower bound li0 on xi and ver-
ifying through bound tightening whether tightening xi to [li0 , ui ] yields a problem that can be
fathomed. Applying this procedure to all variables, possibly in a repeated fashion, can lead to
massive reduction in bounds, but it is computationally expensive given that it requires multiple
calls to other bounds tightening procedures. Probing is used in MILP (Savelsbergh, 1994) and
MINLP (Belotti et al., 2009; Tawarmalani and Sahinidis, 2002) on binary, integer, and continuous
variables.
Reduced cost bound tightening (Ryoo and Sahinidis, 1995), akin to the reduced cost fixing tech-
nique utilized in MILP (Nemhauser and Wolsey, 1988a), uses the solution of an LP relaxation of
(1.1) to infer new, and possibly tighter, bounds on the variables. Suppose that an optimal solution
x
of (5.13) has a variable xi at its lower bound li . Suppose also that the optimal solution has an
objective function value zLP = x n+q and that a cutoff is known for (1.1) with value z. If the re-
duced cost i of xi is positive, then increasing xi by yields an increase in the objective function
of i . Then a valid upper bound on xi is u0i = li + zz 0
i , which constitutes a tightening if ui < ui .
LP
Similarly, if for the optimal solution x? one has xi = ui and a negative reduced cost i , then a valid
lower bound on xi is li0 = ui + zz
i .
LP
5.4 Relaxations of Structured Nonconvex Sets
The methods described in Sections 5.2 and 5.3 are broadly applicable. This approach can be used
to relax any constraint containing a nonlinear function that can be factored into simpler primitive
functions for which we have known relaxations, and then to refine this relaxation after spatial
branching. When combined with relaxation and branching on integer variables, this leads to algo-
rithms that can (theoretically) solve almost any MINLP with explicitly given nonlinear constraints.
The drawback of this general approach is that the relaxation obtained may be weak compared with
the tightest possible relaxation, the convex hull of feasible solutions, leading to an impractically
large branch-and-bound search tree.
As a simple example, consider the nonconvex constraint in two variables x1 and x2 :
x21 + x22 1 (5.16)
and suppose x1 , x2 [0, 2]. The set defined by these constraints is the dark shaded area in Figure
24. One can easily see that the convex hull of this set is given by the bounds on the variables, plus
the inequality x1 + x2 1 (the dashed red line). Now consider the relaxation approach of Section
5.2. We first introduce two new decision variables, x3 and x4 , with x3 x21 and x4 x22 , and
x2
x1
Figure 24: The dark shaded area is the feasible region. The convex hull is the entire shaded area,
and is defined by the red dashed line, x1 + x2 1. The blue dashed line corresonds to the weaker
inequality x1 + x2 1/2.
replace the constraint (5.16) with
x3 + x4 1. (5.17)
The nonconvex constraints x3 x21 and x4 x22 are then relaxed (in the best possible way given
the bounds on x1 and x2 ) with x3 2x1 and x4 2x2 . To compare this with the convex hull, we
can then eliminate the variables x3 and x4 by substituting these inequalities into (5.17), obtaining
2x1 + 2x2 x3 + x4 1, or x1 + x2 1/2 (the blue dashed line in the figure). This relaxation is
therefore significantly weaker than the convex hull.
Examples like those in the preceding paragraph motivate the study of relaxations that consider
more of the problem simultaneously, rather than just separately relaxing all components. Of course,
considering the entire MINLP feasible region is in general an intractable task. One common strat-
egy, however, is to identify specific structures that may appear in many MINLP problems and
study improved relaxations for these structures. A huge variety of such structures exists, so we
cannot provide a complete survey in this paper; but in the following subsections we highlight a
couple of examples.
5.4.1 Nonconvex quadratic functions
One structure that appears in many MINLP problems is the presence of (nonconvex) quadratic or
bilinear functions in either the constraints or the objective. These quadratically constrained quadratic
programs (QCQPs) may also include integer variables and linear constraints. A generic QCQP is a
special case of MINLP of the following form:

minimize xT Q0 x + cT0 x,

x
subject to xT Qk x + cTk x bk , k = 1, . . . , q

(5.18)

Ax b,

0 x u, xi Z, i I,

where for each k = 0, 1, . . . , q, Qk is an n n symmetric matrix, and A is an m n matrix. The

matrices Qk are not assumed to be positive semidefinite, so this problem is nonconvex even when
the integrality constraints are relaxed. It is also possible to have nonzero lower bounds x `, but
we assume here simply x 0 to simplify exposition.
Many relaxation strategies for QCQPs are based on introducing additional variables Xij for all
i, j pairs, and reformulating (5.18) as follows:

minimize

Q0 X + cT0 x,
x
T

subject to Qk X + ck x bk , k = 1, . . . , q

Ax b, (5.19)

0 x u, xi Z, i I,

X = xxT ,

where X is the n n matrix containing all the Xij variables, so that the constraint X = xxT
records the nonconvex constraints Xij = xi xj for all i, j = 1, . . . , n. Observe that these constraints
are the only nonlinear constraints in this reformulation. Obtaining a relaxation of (5.19) can then
be accomplished by relaxing the constraint X = xxT . We discuss two general approaches for re-
laxing this constraint: the reformulation-linearization technique (RLT) (Adams and Sherali, 1986;
Sherali and Alameddine, 1992; Sherali and Adams, 1998) and semidefinite programming. Both
approaches have been widely studied, and an exhaustive literature review is beyond the scope of
this work. Instead, we introduce the basic idea of each as an example of how structured nonconvex
constraints can be relaxed.
In its most basic form, RLT relaxes the constraint Xij = xi xj , for any fixed i, j by first deriving
the following nonlinear nonconvex constraints, based on multiplying pairs of the nonnegative
quantities xi , xj , ui xi , and uj xj :
xi xj 0, (ui xi )(uj xj ) 0, xi (uj xj ) 0, (ui xi )xj 0.
These inequalities are then linearized by replacing the products xi xj with the variable Xij , yielding
Xij 0, Xij ui xj + uj xi ui uj , Xij uj xi , Xij ui xj .
Observe that these inequalities are exactly the special case of the inequalities (5.11) where the lower
bounds on the variables being multiplied are 0. Using these inequalities in place of X = xxT yields
a polyhedral relaxation of (5.19). Two other techniques are commonly used to further strengthen
the RLT relaxation. First, if a decision variable xi , for i I is binary (i.e., ui = 1), then it holds
that x2i = xi , and hence the linear constraint Xii = xi is added to the relaxation. A generalization
to this technique to general integer variables, based on Lagrange interpolating polynomials, has
been proposed as well (Adams, 2011). The second major technique for further improving the RLT
relaxation is to multiply linear constraints together to obtain additional quadratic constraints that
can then be linearized. For example, multiplying a nonnegative decision variable xi with a linear
constraint bt nj=1 atj xj 0 yields the inequality
P
n
X
bt xi atj xi xj 0
j=1
which can then be linearized as

n
X
bt xi atj Xij 0.
j=1
Similar inequalities can be derived by multiplying linear constraints with the nonnegative terms
(ui xi ), and also by multiplying linear constraints with each other, although deriving inequalities
from all possible pairs of linear inequalities may yield a very large linear program.
The other general technique for relaxing the constraint X = xxT in (5.19) is via semidefinite
programming. The key observation is that this constraint X xxT can be relaxed to the constraint
X xxT 0, which is equivalent to
" #
1 xT
0
x X
and thus this yields a semidefinite programming relaxation. Just as in RLT, the corresponding
relaxation can be improved by including the constraint Xii = xi for binary variables xi . When
the QCQP contains linear equations, such as at x = bt , additional linear constraints for the SDP re-
laxation can be obtained by squaring the constraints and then linearizing the variables (Anstre-
icher, 2009):
aTt Xat = b2t .
SDP relxations can be used within a branch-and-bound algorithm to solve QCQPs to optimality
(Burer and Vandenbussche, 2009; Buchheim and Wiegele, 2012).
Anstreicher (2009) compares the relaxations obtained using the RLT and SDP approaches, find-
ing that neither strictly dominates the other, and that combining the two approaches can yield a
relaxation significantly better than obtained by using either approach individually. Additional
linear inequalities related to those obtained for the Boolean Quadric Polytope (Padberg, 1989) can
also be added to further strengthen the relaxation (Yajima and Fujie, 1998; Burer and Letchford,
2009). Anstreicher (2012) demonstrated that the resulting relaxations can be exceptionally tight,
very often yielding a bound equal to the optimal objective value. Unfortunately, the resulting re-
laxations, although convex, may be computationally demanding. Consequently, linear cuts have
been introduced by Sherali and Fraticelli (2002) and further refined by Qualizza et al. (2012) to
provide a polyhedral approximation of the SDP constraint.
For both the RLT and SDP relaxation approach, improved relaxations can also be obtained
by further multiplying linear constraints with each other, obtaining higher-order polynomial con-
straints. Additional variables can then be introduced, as in (5.19), to linearize these constraints,
and then the constraints defining these variables (such as Xi,j,k = xi xj xk ) can be relaxed using an
approach similar to that used for (5.19). This approach leads to a heirarchy of relaxations of im-
proving quality (Adams and Sherali, 2005; Lasserre, 2000, 2001). The drawback of this approach is
that the size of these formulations grows dramatically, limiting their current practical use. Indeed,
even the formulation (5.19) may be significantly larger than the original formulation (5.18), lead-
ing Saxena et al. (2011) to study of relaxations of QCQPs that do not use the additional variables
Xij .
For MINLPs with only a quadratic objective (i.e., problem (5.18) with q = 0), Burer (2009)
derived another important link with conic optimization. In particular, he showed that an exact
reformulation of such a problem can be obtained using a reformulation similar to (5.19) with cer-
tain enhancements (such as using including the constraints Xii = xi for binary variables) and
replacing the constraint X = xxT with the conic constraint
" #
1 xT
C,
x X
where C is the completely positve cone, the set of matrices Y such that Y = BB T for some matrix B
that is componentwise nonnegative.
We close this section by mentioning a few other approaches for solving or relaxing QCQPs.
Saxena et al. (2010) study techniques for deriving disjunctive cuts based on the nonconvex con-
straint xxT X 0, which is also implied by the constraint X = xxT . Vandenbussche and
Nemhauser (2005b,a) studied polyhedral relaxations of box-constrained QPs, in which a general
quadratic function is to be minimized subject to bound constraints on the continuous decision
variables. Linderoth (2005) studied a simplicial branch-and-bound algorithm for QCQP. Burer
and Letchford (2012b) study relaxations of QCQPs with unbounded integer decision variables.
The general -BB relaxation approach Androulakis et al. (1995) has also been used for solving
QCQPs, although Anstreicher (2012) showed that the bounds obtained from SDP relaxation are
always at least as good as those obtained from the -BB approach. Misener and Floudas (2012a)
study the additional use of piecewise linear and edge-concave relaxations of QCQPs. Bao et al.
(2009) and Luedtke et al. (2012) consider related approaches for relaxing multilinear functions in
which the decision variables are bounded, which can also be applied to QCQPs.
5.4.2 Bilinear covering sets
Tawarmalani et al. (2010) study a framework for generating valid inequalities for MINLPs that
have a certain orthogonal disjunction structure. We do not review this general work but instead
highlight the results they obtain by applying this framework to sets they refer to as bilinear covering
sets.
First consider pure integer covering set defined below:

n n
X o
I n n
B := (x, y) Z+ Z+ | xi yi r
i=1
where r is a positive number. Note that, for any i, the convex hull of the two variable integer set
BiI := {(xi , yi ) Z+ Z+ | xi yi r}
is a polyhedron defined by d dre + 1 linear inequalities. Let us denote the inequalities defining
the convex hull of BiIC by
ak xi + bk yi 1, k = 1, . . . , d (5.20)
where we can assume (by scaling) that each inequality has a right-hand side 1. In particular,
these inequalities include the constraints xi 1 and yi 1. The remaining inequalities can be
computed, for example, by finding all inequalities of the form axi + byi 1 that do not cut off any
of the points (xti , yit ) = (t, dr/te) for t = 1, . . . , dre and that are exactly satisfied by two of these
points.
Now, let be the collection of all possible mappings of the form : {1, . . . , n} {1, . . . , d}.
That is, if , then for each i 1, . . . , n, (i) selects an inequality in the description of conv(BiI ).
Then, conv(B I ) is characterized as follows.
Theorem 5.2 (Proposition 6 in Tawarmalani et al. (2010)). The convex hull of B I is given by the set of
x Rn+ , y Rn+ that satisfy the inequalities
n
X
(a(i) xi + b(i) yi ) 1, . (5.21)
i=1
While there are an exponential number of inequalities in (5.21), given a point ( x, y) Rn+
Rn+ , separation can be accomplished efficiently by independently considering each x i , yi pair and
setting (i) to the index of the most violated constraint in (5.20). Tawarmalani et al. (2010) provide
a a very similar result for the case of a bilinear covering set similar to B I , but where one of the sets
of variables is continuous.
We turn next to the case of a continuous bilinear covering set of the form
n n
X o
C n n
B := (x, y) R+ R+ | (ai xi yi + bi xi + ci yi ) r ,
i=1
where r > 0 and ai , bi , ci > 0 for i = 1, . . . , n.
Theorem 5.3 (Proposition 9 in Tawarmalani et al. (2010)). The convex hull of B C is given by the set of
x Rn+ , y Rn+ that satisfy the inequality:
n
1 X p
bi xi + ci yi + (bi xi + ci yi )2 + 4ai rxi yi r.
2
i=1
Unfortunately, when the bounds on the variables are also considered (e.g., if x is a set of binary
variables), the corresponding covering sets become much more difficult to analyze. In particular,
Chung et al. (2011) show that optimizing a linear function over such a set is N P-hard. Chung
et al. (2011) did, however, study valid inequalities for this set that may be useful in strengthening
the continuous relaxation.
6 Heuristics for Solving MINLPs

Some real-world applications cannot be solved to global optimality by using the methods de-
scribed in Sections 3 to 5, because the problems are too large, generate a huge search tree, or must
be solved in real time. In these situations it is more desirable to obtain a good solution quickly than
to wait for an optimal solution. In such situations, we may resort to heuristic search techniques
that provide a feasible point without any optimality guarantees. Heuristics can also accelerate
deterministic techniques by quickly identifying an incumbent with a low value of the objective
function. This upper bound can then be used to prune a larger number of the nodes in the branch-
and-bound algorithm of Section 3.1 and in the branch-and-cut algorithm of Section 3.1.4. An
incumbent solution may additionally be used for guided dives (Danna et al., 2005), that is, se-
lecting a child node during a dive. Bounds tightening described in Section 5.3.2 can also be applied
to the objective function more effectively if a tight upper bound is known.
We distinguish two classes of heuristic search techniques: probabilistic search and determinis-
tic search. Probabilistic search refers to techniques that require at each iteration a random choice
of a candidate solution or parameters that determine a solution. Simulated annealing (Kirkpatrick
et al., 1983), ant colony optimization (Dorigo et al., 1996), particle-swarm optimization (Kennedy
and Eberhart, 1995), cross-entropy (Rubinstein and Kroese, 2004), tabu search (Glover, 1989, 1990)
and genetic algorithms (Goldberg, 1989) are some methods that fall under this category. Although
simple to design and applicable to many combinatorial optimization problems, these methods re-
quire implementation and modifications specific to the structure of the problem being solved. We
therefore focus only on more general approaches.
We use the term deterministic rather loosely, since the methods in this category may also
sometimes need randomization in certain iterations. To begin with, all deterministic techniques
discussed in Sections 3 to 5 can be run as heuristics. For example, we can run branch-and-bound
for a fixed time or fixed number of nodes or until it finds its first incumbent. In this section, we
discuss more efficient alternatives to such simple heuristics. These heuristics can be classified into
two types: search heuristics, which search for a solution without the help of any known solutions,
and improvement heuristics, which improve upon a given solution or a set of solutions.
Notation. Throughout this section we use a unified notation to refer to incumbents and solutions
of the various steps of the heuristics: x refers to the current incumbent, which is feasible in (1.1); x0
refers to a (local) solution of the continuous relaxation (1.3); x denotes the solution to a polyhedral
(j)
relaxation of (1.1); and x(j) is a (local) solution of an NLP with fixed integer variables, (NLP(xI )).
Given this notation, we can now describe the deterministic search techniques within a unified
terminology.
6.1 Search Heuristics
Several heuristics to search for a feasible solution of a MINLP have been proposed recently. They
all make clever use of LP, MILP, and NLP solvers to solve problems easier than the MINLP to
obtain a feasible point. Some of these heuristics may completely ignore the objective function and
focus on finding only a feasible solution. They may use the solution of the relaxation at any node
in the branch-and-bound as a starting point and hence try to make up for the lack of focus on the
objective function of the MINLP.
6.1.1 MILP-based rounding
Finding a locally optimal solution to the continuous relaxation (1.3) of the MINLP (1.1) is usu-
ally easier and computationally faster than solving the MINLP itself. Given a solution x0 of the
continuous relaxation, one can try rounding fractional values of integer-constrained variables.
Unfortunately, such a simple rounding usually will not produce a feasible solution. Nannicini
and Belotti (2012) propose to overcome this difficulty by solving an MILP. The constraints of the
MILP are linear relaxations of the original MINLP obtained by methods described in Section 5.2.
The objective function is the `1 norm kx x0 k1 . The solution x of the MILP satisfies the inte-
grality constraints but not necessarily the nonlinear constraints. Another NLP is now solved, this
time with all integer variables fixed to the values in xI . The process is repeated until we obtain a
feasible solution or reach termination criteria (limits on time or iterations).
Nannicini and Belotti (2012) suggest several practical measures for implementing the above
general scheme. First, it is not necessary to fully solve the MILP. We can stop as soon as a feasible
point of the MILP is found. Second, different initial points x0 can be used to initialize the heuristic.
In particular, if we are solving the NLP with an interior point method, we can stop the NLP if
it finds a feasible point even when the log-barrier coefficient is not close to zero. Third, no-good
cuts are added to ensure that the MILP does not include integer solutions (x ) found in previous
iterations of the heuristic. A no-good cut is used to model the constraint
X
|xi xi | 1. (6.1)
iI
When all integer variables are binary, the constraint (6.1) is simplified to
X X
xi + (1 xi ) 1. (6.2)
iI:xi =0 iI:xi =1
Auxiliary binary variables may need to be introduced when some variables are general integers
instead of binary.
6.1.2 Feasibility pump
The feasibility pump heuristic was introduced by Fischetti et al. (2005) in the context of MILP and
improved by Achterberg and Berthold (2007) and Fischetti and Salvagnin (2009). Bonami et al.
(2009) extend this idea to MINLPs. The main idea, like that in MILP based rounding described
above, is that an NLP solver can be used to find a solution that satisfies nonlinear constraints. In-
tegrality is enforced by solving an MILP. An alternating sequence of NLP and MILP is solved that
may lead to a solution feasible to the MINLP. The main difference from the rounding approach of
Section 6.1.1 is the way MILP is set up. Suppose x0 is a locally optimal solution of the NLP relax-
ation of the MINLP (1.1). Bonami et al. (2009) obtain the MILP using the linearization (3.4) used
for outer approximation. Thus, if we have a convex MINLP, the MILP is a relaxation. Otherwise,
it is only an approximation. The objective function is again the `1 norm kx x0 k1 .
If the solution x to the MILP satisfies the nonlinear MINLP constraints, c(x ) 0, then we
have found a new incumbent. Otherwise, we solve an NLP with all the integer variables fixed
to the values of x . The objective of this NLP is kx x k2 . Now, x0 is updated to the solution of
this NLP. If x0 does not satisfy integrality constraints, new linearizations are added to the previous
MILP.
For a convex MINLP, the more restricted MILP does not contain x , and hence the no-good
cuts are not required. In addition, Bonami et al. (2009) add the valid inequality
(x0 x )(x x0 ) 0
to the MILP. They also prove that in the case of a convex MINLP, the heuristic does not cycle and
always terminates in finite number of iterations. They call their version of feasibility pump for
convex MINLPs enhanced feasibility pump. A simpler version of the feasibility pump (Bonami
and Goncalves, 2012) does not involve solving an MILP. Instead one just rounds x0 to the nearest
point satisfying integrality constraints. This version requires random changes in x0 whenever
cycling occurs, similar to the work of (Fischetti et al., 2005).
DAmbrosio and Liberti (2012) note that feasibility-pump-based heuristics can be viewed as
applications of the sequential projection method (SPM) or the alternating projection methods.
SPM has been used extensively for solving convex feasibility problems. We refer to the survey
by Bauschke and Borwein (1996) for theory and algorithms. The feasibility of MINLPs is not a
convex problem, and so the heuristics along the lines of these algorithms cannot be expected to
have a similar performance or to even converge. DAmbrosio and Liberti (2012) consider two sets
related to the feasible region of the MINLP (1.1):
A = {x | c(x) 0, x X} , and (6.3)

B = {x | cc (x) 0, x X, xi Z i I} , (6.4)
where cC refers to the convex constraints in the MINLP. Set A is in general nonconvex, while B is
a set of feasible points of a convex MINLP. One can now solve optimization problems over A and
B repeatedly in order to obtain two sequences of solutions x

i and x
i as follows:
i = min kx x
x i1 k,
xA
i = min kx x
x i1 k.
xB
Both the above problems are NP-hard, and DAmbrosio and Liberti (2012) suggest solving them
using well-known heuristics. For instance, a locally optimal solution of the first problem can be
obtained by most NLP solvers. The latter problem can be solved as a convex MINLP by methods
described in Section 3. By selecting different methods to solve these problems, one can obtain
several variants of the feasibility pump.
6.1.3 Undercover
The Undercover heuristic (Berthold and Gleixner, 2012) is specially designed for nonconvex MINLPs.
The basic idea is to fix certain variables in the problem to specific values so that the resulting re-
striction becomes easier to solve. The restriction that Berthold and Gleixner (2012) obtain is an
MILP. This MILP can then be solved either exactly or heuristically. Since the MILP is a restriction
of the MINLP, any feasible solution of MILP will also satisfy the MINLP.
In order to be successful, the heuristic should fix a minimal number of variables lest the re-
duction be too restrictive and good solutions be cut off. The sparsity pattern of the Hessian of
the Lagrangian tells us which variables appear in nonlinear functions. The authors create a graph
G(V, E) where each vertex vi V, i = 1, . . . , n denotes a variable of the MINLP. An edge eij is
added to the graph if the Hessian of the Lagrangian of the NLP relaxation has a nonzero entry
(i, j). G may also contain loops if a diagonal entry in the Hessian is nonzero. We illustrate a
graph for a simple example below. To make the problem linear, we can fix the variables x1 , x3 ,
minimize x31 + x2 + x3 x4 + x5 1 3 5
x
subject to x4 x5 1
x1 , x2 , x3 , x4 , x5 [0, 10] 2 4
Figure 25: Graph (right) denoting the nonzero structure of the Hessian of the example on the left.
and x5 . However, this choice is not minimal because we can also fix x1 and x4 to obtain a linear
problem. Berthold and Gleixner (2012) observe that a minimal vertex-cover of G is also a minimal
set of variables that can be fixed to make the problem linear. They solve the minimal vertex-cover
problem using an MILP solver that is usually available in a MINLP framework.
The fixed values of variables in a cover are obtained from a solution of the NLP relaxation or
an LP relaxation. The variables are fixed sequentially. Each time a variable is fixed, domain prop-
agation is invoked to tighten bounds or to fix other variables. This heuristic works particularly
well for problems with quadratic constraints and objective.
6.1.4 RENS: Relaxation Enforced Neighborhood Search
The RENS (Berthold, 2012) heuristic searches for a feasible solution of an MINLP around a point
that does not satisfy integrality constraints. Consider an NLP solution x0 and suppose F = {i
I | x0i
/ Z} is the set of all variables that violate integrality constraints at x0 . Keeping the values of
the variables in the set I \ F fixed to those of x0 , one has 2|F | ways of rounding up or down x0 so
that it satisfies integrality constraints. The RENS heuristic seeks to systematically search through
these 2|F | combinations to find a feasible solution of the MINLP.
Berthold (2012) creates a possibly much smaller MINLP in order to search for the feasible
solution. First, he fixes all variables in the set I that have an integer value in x0 . Next, the bounds
of each variable xi , i F are changed to [bx0i c, dx0i e]. The resulting MINLP is then solved by using
a MINLP solver. If the restricted MINLP is considerably smaller than the original, then the solver
can solve it quickly. Additional limits need to be imposed on the solver so that the heuristic does
not take excessive time. By means of this heuristic, the authors show that more than half the
test-instances have solutions that could be obtained by rounding.
6.1.5 Diving
The fundamental idea of all diving heuristics is to conduct a depth-first exploration of a possi-
ble path from the root node to a leaf node of the branch-and-bound tree, before searching other
branches. The hope is that this will lead to a feasible solution and hence an upper bound early in
the MINLP solution process.
Bonami and Goncalves (2012) propose to start the diving process by solving the NLP relaxation
of the MINLP to obtain a relaxed solution x0 Rn . They then fix x0i 6= Z, i I to bx0i c or dx0i e and
resolve the modified NLP. This process is iterated until all integer variables have been fixed. This
heuristic is successful if the obtained leaf NLP is feasible for the MINLP. It reports a failure if the
obtained leaf NLP is infeasible, the objective function exceeds the incumbent bound, or an NLP
solver termination criterion such as a limit on the number of iterations is met.
Selection of the variable to be fixed and the side to which it is fixed leaves room for some
tailoring of diving heuristics. Bonami and Goncalves (2012) describe fractional diving, vector length
diving, and a modification of these heuristics we refer to as nonlinear diving:
In fractional diving, the variable to be rounded is selected from the set of smallest values |x0j
[xj ]|, where the bracket [] indicates rounding to the nearest integer. The selected variable is
fixed to the nearest integer.
In vector length diving, the variable is selected from the set of smallest ratios
(dx0j e x0j )gj + (bx0j c x0j )gj +
if gj 0, and otherwise, j I.
Aj + 1 Aj + 1
0
Here, gj = fx(xj ) , the constant Aj indicates the number of problem functions for which xj ,
j I has a nonzero coefficient in the linearization, and is chosen to be a small positive con-
stant, for example = 106 . The selected variable is rounded up if the gradient with respect
to xj is nonnegative, and it is rounded down otherwise. The selection favors rounding of a

variable that incurs a small objective function change but affects a large number of problem
constraints.
The nonlinear diving heuristic may be applied to either of the above criteria. Here, xi , i I is
selected from the subset of nonlinear fractional variables only, with the aim of obtaining a MILP
that can then be solved by a (black-box) MILP solver. In a leaf node where all nonlinear integer
variables have been fixed and at least one linear integer variable is still fractional, this MILP is
obtained by fixing all continuous nonlinear variables to their current value.
Diving heuristics need to take into account the computational effort required for repeated re-
solves of the modified NLP. To mitigate this cost, Bonami and Goncalves (2012) propose to fix
K > 1 variables at the same time before resolving the modified NLP. Mahajan et al. (2012) pro-
pose to solve QPs instead of NLPs, which speeds up the diving process by exploiting the warm
starting capabilities of active set QP solvers.
6.2 Improvement Heuristics
Improvement heuristics start with a given feasible point x of the MINLP and try to find a better
point. Two well-known heuristics for searching a better solution in the neighborhood of a known
solution have been adapted from MILP to MINLP. We describe them next.
6.2.1 Local branching
Local branching is a heuristic for MINLPs, where all integer variables are binary, that is, xi
{0, 1}, i I. It was first introduced in the context of MILP by Fischetti and Lodi (2003) and
generalizes readily to convex MINLPs. We start by describing local branching for convex MINLPs
and then describe an extension to nonconvex MINLPs.
The main idea behind local branching is to use a generic MILP solver at a tactical level that is
controlled at a strategic level by a simple external branching framework. Assume that we are given
a feasible incumbent x of (1.1), and consider the following disjunction (generalized branching)
for a fixed constant k Z:
kxI xI k1 k (left branch) or kxI xI k1 k + 1 (right branch). (6.5)
This disjunction corresponds to the Hamming distance of xI from xI , and the left branch can also
be interpreted as an `1 trust region around the incumbent. In the case of binary variables, we can
rewrite (6.5) as two linear constraints:
X X X X
xi + (1 xi ) k (left) or xi + (1 xi ) k + 1 (right). (6.6)
iI:xi =0 iI:xi =1 iI:xi =0 iI:xi =1
The left branch is constructed in such a way that it is much easier to solve than (1.1), typically by
choosing k [10, 20]. We start by solving the left branch using any of the methods introduced
in Section 3 and obtain a new incumbent. We can then either solve the right branch or again
divide the right branch using a new disjunction (6.6). This creates an outer branch-and-bound
tree where each node corresponds to an MINLP. If this local branching tree has been searched to
completion, we have solved the MINLP. In general, however, we do not run local branching to
completion, because it would be inefficient to regenerate pseudocosts for every MINLP solve, for
example. Fischetti and Lodi (2003) propose two enhancements: first they impose a time limit or
node limit on each MINLP solve, and second they introduce a diversification mechanism in case
the left branch does not improve the solution. If the left branch is not solved completely, one can
still obtain a complete search algorithm by modifying the local branching strategy.
Local branching has been extended to nonconvex MINLPs (Nannicini et al., 2008). The ex-
tension is based on solving an alternating sequence of (local) NLP relaxations and fixings and
(global) MILP outer approximations and is closely related to the feasibility pump described in
Section 6.1.2. The local branching is used only as a no-good cut (6.1) and not viewed as a strategic
(outer) branching technique. A related heuristic is RECIPE (Liberti et al., 2011).
6.2.2 RINS: Relaxation-Induced Neighborhood Search
In the RINS (Danna et al., 2005) heuristic, one searches for better solutions in the neighborhood of
an already known solution, much like local branching. However, instead of imposing a distance
constraint (6.5) to determine a neighborhood, variables are fixed to certain values. The variables
to be fixed are selected on the basis of the solution of the relaxation x0 , and the already known
incumbent x . For all i I the variables are fixed to xi , x0i = xi . If the fixing reduces the problem
size considerably, then it can be solved by calling the solver again.
Bonami and Goncalves (2012) extend this idea from MILP to the NLP-based branch-and-bound
algorithm of Section 3.1 for convex MINLP. Once they fix the integer variables as above, they solve
the smaller MINLP using the LP/NLP-BB algorithm mentioned in Section 3.3.1. They show that
the LP/NLP-BB algorithm is much faster on the smaller problems than is the NLP-based branch-
and-bound algorithm.
7 Mixed-Integer Optimal Control Problems
In this section, we are interested in mixed-integer nonlinear optimization problems constrained

by a system of ordinary differential equations, also called mixed-integer optimal control problems
(MIOCPs). In principle, these problems can be discretized in time and solved by using any of
the techniques of Sections 3 to 5. Unfortunately, such a discretization destroys the structure of
the problem and does not provide successful algorithms. We present an alternative partial outer
convexification approach that exploits problem structures and provides an efficient approach to
this class of problems.
7.1 Mixed-Integer Optimal Control Problem Class
We start by defining a class of MIOCPs:

minimize J(x(T ))

x(),u(),v()

subject to x(t)
= fODE (x(t), u(t), v(t)) t [0, T ],

x(0) = x0 , (MIOCP)

u(t) U t [0, T ],
0 c(x(t), u(t)) t [0, T ],

v(t) t [0, T ],

where we strive to determine differential states x : [0, T ] Rnx , continuous controls u : [0, T ]
U Rnu , and integer controls v : [0, T ] Rnv on the time horizon [0, T ] such that a Mayer
term J : Rnx R evaluated at the end of this horizon is minimized subject to a nonlinear system
fODE : Rnx Rnu Rnx of ordinary differential equations (ODEs) with fixed initial values
x0 Rnx . In addition, we consider nonlinear path constraints c : Rnx U Rnc independent
of the integer control. The set U describes a feasible domain for the continuous control u(), and
the set := {v 1 , v 2 , . . . , v n } contains n < discrete choices v i Rnv , i I := {1, . . . , n }
admissible for the integer control.
Mixed-integer optimal control problems can be discretized in time by using, for example, di-
rect collocation or direct shooting discretizations and can then be solved by any of the MINLP
algorithms in Sections 3 to 5. The discretized MINLPs, however, are typically challenging. Strong
nonlinear and transient phenomena inherent in the dynamic model fODE () may require a fine-
grained time discretization, which leads to a large number of integer variables in the discretized
MINLP. While this issue typically is successfully addressed by error estimation and adaptivity, in-
corporating adaptive discretizations into the MINLP model is difficult. Moreover, simply relaxing
the integrality in (MIOCP) to v(t) conv yields a lower bound on the optimal solution that can
be obtained by solving a continuous optimal control problem. Unfortunately, this relaxation in
general does not give any guarantees on integer feasibility or on the distance to an integer optimal
solution. In addition, the ODE constraint poses a particular challenge since the dynamics need
not remain well-defined on the continuous relaxation.
To address these issues, we review an NLP-MILP decomposition approach for mixed-integer
optimal control, as developed by Sager (2005); Sager et al. (2009) and Sager et al. (2012). The
fundamental insight here is that it is indeed sufficient to compute a relaxed optimal solution for
a special convexified reformulation of the mixed-integer optimal control problem. With a direct
method for continuous optimal control, this essentially amounts to solving an NLP. From this, we
obtain a lower bound on the MIOCP solution, which, unlike in MINLP, can be shown to be tight.
An integer feasible solution can be obtained constructively by applying a certain rounding scheme
or by solving an MILP.
7.2 Partial Outer Convexification
To tighten the lower bound obtained from a relaxation and to overcome the problem of ill-defined
dynamics, Sager (2005) proposes to convexify the problem with respect to the integer control v().
In doing so, binary multiplier functions i () {0, 1} are introduced for each admissible choice of
v(t). We represent the integer control as a convex combination,
t [0, T ] : v(t) = v i i (t) = 1, j (t) = 0 j I, j 6= i. (7.1)
This gives rise to a partially convexified MIOCP

minimize J(x(T ))

x(),u(),()

= iI i (t)fODE (x(t), u(t), v i )
P
subject to x(t)
t [0, T ],

x(0) = x0 ,

u(t) U t [0, T ], (MIOCP-BC)

0 c(x(t), u(t)) t [0, T ],

P
1 = iI i (t) t [0, T ],

(t) {0, 1}n t [0, T ].

This problem is convex only if fODE and c are convex in x and u. Clearly, the sets of optimal
solutions of (MIOCP) and (MIOCP-BC) coincide. Given optimal binary multiplier functions (),
the corresponding original integer control is easily determined. If v() is binary and enters linearly,
then (MIOCP) already is in the convexified form. Otherwise the special ordered set of type-1,
SOS1, may be used. Introduced by Beale and Tomlin (1970) in the context of MILP, an SOS1 set
can be used to model the condition (7.1) by means of the equations
X
v(t) = v i i (t) t [0, T ],
iI
X
i (t) = 1 t [0, T ]. (7.2)
iI
We are interested in the relaxation of (MIOCP-BC) obtained by replacing i () by relaxed convex

multiplier functions i () [0, 1]. This relaxation is a purely continuous optimal control problem,
denoted by (MIOCP-RC). Direct and all-at-once methods have emerged as the methods of choice
to solve it efficiently.
As an illustrating example, we consider a energy optimal rocket-car control problem in its
integer nonlinear formulation on the left and its partially convexified formulation on the right for
t [0, 32]. The multiplier function 1 (t) lends itself to elimination by using the SOS1 constraint.

minimize x 2 (32)
minimize x2 (32)

x(),v()

x(),()

subject to x 0 (t) = x1 (t) t

subject to x 0 (t) = x1 (t) t

x 1 (t) = v(t) t

x 1 (t) = 21 (t) + 2 (t) t
2

t

x 2 (t) = v (t) x 2 (t) = 41 (t) + 2 (t) t (7.3)

x(0) = (0, 0, 0)

x(0) = (0, 0, 0)

x 0 (32) = 300

x0 (32) = 300
v(t) {2, 1} t (t) {0, 1}2 t

1 (t) + 2 (t) = 1 t

For our example, we use the multiple shooting code MUSCOD-II for direct mixed-integer op-
timal control (Bock and Plitt, 1984; Leineweber et al., 2003; Sager, 2005; Kirches and Leyffer, 2011)
available on NEOS (Czyzyk et al., 1998; Dolan et al., 2002).
Figure 26: The nonlinear relaxed solution to the energy optimal rocket-car example yields a weak
lower bound x2 (32) 37 to the MIOCPs optimal objective. The figure shows v() [2, 1] relaxed
(top row, left), and partial outer convexification v() = 21 (t) + 12 (t) (bottom row, left), as well
as the resulting differential state trajectories x0 (), x1 (), x2 () (from left to right) for an equidistant
piecewise constant control discretization of N = 80 intervals.
7.3 Relaxation of the Partially Convexified MIOCP
Optimal solutions to control problems permit a classification into arcs, or sections, of [0, T ] on
which the solution has a particular structure; see, for example, Bryson and Ho (1975) for an in-
troduction. Of interest for (MIOCP-RC) are arcs on which the optimal solution to the relaxed
multipliers () comes to lie on the boundary of the feasible set and is hence feasible also for
(MIOCP-BC). One example of such behavior is bang-bang arcs. For particular problem classes,
such as time-optimal linear control problems, a bang-bang principle proves existence of an opti-
mal solution that consists of bang-bang arcs only; see the work of (Kirches et al., 2010) for a MIOCP
example. Solving (MIOCP-RC) then already yields an optimal solution of (MIOCP). In example
(7.3), (MIOCP-RC) has a bang-bang arc on [0, 3.5] approximately; see Figure 26.
Arcs on which () is not binary feasible require application of a rounding scheme to construct
a binary feasible control (). We postpone discussion of such rounding schemes until the next sec-
tion and consider first the work of (Sager et al., 2009, 2012) who investigate the difference between
the state trajectories generated by relaxed multipliers () and binary ones (). We consider the
ODE right-hand side f(x, u) composed from columns fODE (x, u, v i ), i I, and we fix a measurable
control u ().
Theorem 7.1 (Theorem 2 and Corollary 6 in (Sager et al., 2012)). Let x () and x () be the solution
of the initial value problems
(
x (t) = f(x (t), u (t)) (t), x (0) = x0
(7.4)
x (t) = f(x (t), u (t)) (t), x (0) = x0
with t [0, T ], , : [0, T ] [0, 1]n measurable functions and f : Rnx +1 Rnx Rn essentially
bounded on [0, T ] by M R+ and t-differentiable almost everywhere. If there exist C, L R+ such that

df
dt (x(t), u (t)) C, f(x (t), u (t)) f(x (t), u (t)) L ||x (t) x (t)|| (7.5)

for t [0, T ] almost everywhere, and there exists R+ such that

Z T

(t) (t) dt , (7.6)
0
then for all t [0, T ] it holds that
||x (t) x (t)|| (M + Ct) exp(Lt). (7.7)
This theorem differs remarkably from the MINLP setting. Given a relaxed optimal solution
(x (), u (), ()), if one can construct a binary solution () that satisfies assumption (7.6) on the
control deviation bound , then the deviation of the resulting state trajectory x () from the relaxed
optimal one x () is bounded linearly in .
7.4 Constructing the Integer Control
A sum-up rounding scheme for construction of () from a relaxed optimal solution () is shown
to satisfy assumption (7.6) by Sager et al. (2012). We first present the case for an affine-linear binary
control function v(), when the SOS1 constraint is absent.
Theorem 7.2 (Theorem 3 in (Sager et al., 2012)). Let : [0, T ] [0, 1]n and a time grid 0 = t0 <
t1 < . . . < tN = T be given, and define : [0, T ] {0, 1}n by (t) := pi for t [ti , ti+1 ), 0 i < N ,
where
tiR+1 i1
P
pi,j := j (s) ds pk,j (tk+1 tk ),
0 k=0
( (7.8)
1 if pi,j 21 (ti+1 ti )
pi,j := , j I.
0 else
Then it holds that

T
Z
1

(t) (t) dt max {ti+1
2 0i<N ti } =: . (7.9)
0
After elimination of 1 (), example (7.3) is covered by this theorem. Figure 27 shows the sum-
up rounding solution constructed from the convexified relaxed solution of Figure 26 (bottom).
Figure 27: Integer feasible sum-up rounding solutions to the energy optimal rocket-car example.
The objective is x2 (32) = 64, and the lower bound provided by the convexified relaxed solution is
exact in this case (cf. Figure 26, bottom row). The figure shows the integer control v() (left) and
the resulting state trajectories x0 (), x1 (), x2 () (from left to right).
When the SOS1 constraint is present in (MIOCP-RC), () constructed by the above rounding
may violate it. Hence, a more elaborate rounding scheme is required:
Theorem 7.3 (Theorem 5 in (Sager et al., 2012)). Let : [0, T ] [0, 1]n and a time grid 0 = t0 <
t1 < . . . < tN = T be given, and define : [0, T ] {0, 1}n by (t) := pi for t [ti , ti+1 ), 0 i < N ,
where (
1 if k 6= j : pi,j pi,k and k 6= j, pi,j = pi,k : j < k
pi,j := , j I, (7.10)
0 else
and the pi,j are defined as above. Then it holds that
T
Z
(t) (t) dt (n ) max {ti+1 ti } =: .

(7.11)

0 0i<N
For the constant (n ), the best-known result currently is (n ) := n 1 by Sager et al. (2012),
who conjecture (n ) O(log n ). The practical implication of these theorems is that the bound
of Theorem 7.1 can be made arbitrarily small by using a sufficiently fine time discretization, that is,
N large enough to allow for max0i<N {ti+1 ti } to be sufficiently small. As a consequence of the-
orems 7.1 and 7.2 or 7.3, the increase in the objective function value and the constraint infeasibility
are bounded as well.
Theorem 7.4 (Corollary 8 in (Sager et al., 2012)). Assume that J and c are continuous functions, and
that the assumptions of Theorem 7.1 hold. Then for all > 0 there exists t := max0i<N {ti+1 ti }
sufficiently small such that
|J(x (T )) J(x (T ))| , (7.12)

|ci (x (t), u (t)) ci (x (t), u (t))| , 1 i nc . (7.13)
The binary control () obtained from the constructive sum-up rounding process may switch
frequently, and the number of switches may increase along with N and need not be bounded,
as an example due to Fuller (1963) in Figure 28 shows. Sager et al. (2011) hence consider this
rounding scheme as a special-case solution to the following switch cost MILP, in which k,max is a
fixed upper bound on the number of switches allowed in the integer control,

Pi
minimize maximize |(jk pjk )(tj+1 tj )|

p 0i<N,kI j=1
NP1 (7.14)

subject to k,max |pik pi+1,k | k I,

i=0
pi {0, 1}n

0 i < N.

Here, () is determined by the solution of this MILP, rather than by Theorem 7.2 or 7.3. If the
switch cost constraint is inactive, these theorems give the linear-time solution to this MILP.
Figure 28: Fullers problem (Fuller, 1963) has an optimal solution that, in the limit N ,
switches infinitely often on a finite time horizon. Relaxed optimal control (left), sum-up rounding
control with optimized switching times (center), and optimal state trajectory (right) for N = 80.
Partial outer convexification, relaxation, and sum-up rounding can hence be seen as an NLP-
MILP decomposition approach to solving a time-discretized MIOCP that enjoys -optimality and
-feasibility certificates. The computational advantage of this approach over MINLP branch-and-
bound algorithms for MIOCP can hence be arbitrarily large, as demonstrated by Kirches et al.
(2010) for a benchmark problem due to Gerdts (2005).
7.5 Extensions to the Presented Theory
Problem class (MIOCP) is a special case of a larger class of mixed-integer optimal control problems
also covered by the presented theory. An integral Lagrange term in the objective, time-dependent
ODE systems, and free continuous model parameters constant in time are easily covered by in-
troducing additional artificial ODE states. Free initial and final times subject to optimization can
be realized by introducing two free parameters and an affine-linear transformation of the free
horizon. Problem class (MIOCP) is also easily extended to allow for free initial values, for point
constraints on the state x(), including end-point constraints, and for constraints coupled in time,
such as periodicity constraints.
Mixed-integer optimal control of systems described by differential-algebraic equations is dis-
cussed by Gerdts and Sager (2012). Because of time discretization, the switching times are deter-
Table 1: MINLP Solvers (Convex). A means that we do not know the language the solver is
written in.
Name Algorithm(s) Interfaces Language Open-Source
-ECP Extended cutting-plane Customized, GAMS Fortran-90 No
BONMIN NLP-BB, LP/NLP-BB, AMPL, C++, GAMS, C++ Yes
OA, Hybrid MATLAB
DICOPT OA GAMS No
FilMINT LP/NLP-BB AMPL C No
MILANO NLP-BB, OA MATLAB MATLAB Yes
MINLPBB NLP-BB AMPL, Fortran Fortran-77 No
MINOPT OA Customized C No
MINOTAUR NLP-BB, QP-Diving AMPL, C++ C++ Yes
SBB NLP-BB GAMS No
mined up to the grid granularity only. Sager (2005) uses a switching time optimization approach
to refine the integer solution once the optimal switching structure has been determined and fixed.
When (MIOCP) contains integer parameters subject to optimization, a combination of MIOCP
techniques and MINLP techniques has to be used.
Constraints that depend on the integer control v(t) are not immediately covered by (MIOCP)
and the approximation theory based on Theorem 7.1, because violations of constraints after round-
ing cannot be ruled out. Kirches (2011) discusses this case as well as extensions to mixed-integer
nonlinear feedback control. An online benchmark library of mixed-integer optimal control prob-
lems is presented by Sager (2012).
8 Software for MINLP
The availability as well as the maturity of software for modeling and solving MINLP has increased
significantly in the past fifteen years, and now includes a number of open-source and commercial
solvers. We briefly survey the solvers currently available and describe their salient features. Re-
cent surveys of Bussieck and Vigerske (2010) and DAmbrosio and Lodi (2011) also provide an
excellent description of available MINLP solvers. We divide the solvers into those for convex and
nonconvex MINLPs; little intersection exists between the two categories.
Key characteristics and features of the solvers are summarized in Tables 1 and 2, where we use
the following abbreviations to indicate the type of MINLP method that the solver implements:
NLP-BB for nonlinear branch-and-bound (Section 3.1); LP/NLP-BB for LP/NLP-based branch-
and-bound (Section 3.3.1); OA for outer approximation (Section 3.2.1); Hybrid is a hybrid between
OA and LP/NLP-BB; QP-Diving (Section 3.1.3); -BB for -branch-and-bound (Section 5.3); and
LP-BB for LP-based branch-and-bound (Section 5.3).
We are aware of two software packages, MUSCOD-II and MINOPT, that can solve MIOCPs.
Table 2: MINLP Solvers (Nonconvex). A means that we do not know the language the solver
is written in.
Name Algorithm(s) Interfaces Language Open-Source
-BB -BB Customized No
BARON LP-BB AIMMS, GAMS No
COCONUT LP-BB AMPL, GAMS C Yes
COUENNE LP-BB AMPL, C++, GAMS C++ Yes
GloMIQO LP-BB C++, GAMS C++ No
LGO Sampling, AIMMS, AMPL, GAMS, C No
Heursitics MATHEMATICA
LindoGlobal LP-BB C++, GAMS, Lindo No
SCIP LP-BB AMPL, C, GAMS, MATLAB, C Yes
OSiL, ZIMPL
While MINOPT relies on MINLP techniques, MUSCOD-II relies on the partial outer convexifica-
tion approach reviewed in Section 7. We briefly describe it in Section 8.3. Expressing and model-
ing MINLPs is significantly different from LPs or MILPs because general nonlinear functions are
much more difficult to represent with data structures. Hence, good modeling tools are indispens-
able for MINLPs. In Section 8.4 we describe some tools available for modeling and reformulating
MINLPs.
8.1 Convex MINLP solvers
-ECP is a solver written by Westerlund and Lundqvist (2005) to solve convex MINLPs by using
the extended-cutting plane method (Section 3.2.3). It also implements methods for MINLPs with
pseudoconvex functions (Westerlund and Porn, 2002). The solver can read MINLPs in an extended
LP format and can be called through the GAMS (Brooke et al., 1992) modeling system. It requires
user to specify an MILP solver for solving the MILP in each iteration. It also provides a graphical
interface for the MS-Windows operating system.
BONMIN stands for Basic Open-source Nonlinear Mixed Integer optimizer. It is an open-source
solver available at the COIN-OR website (Bonami et al., 2008). It implements nonlinear branch-
and-bound (Section 3.1), LP/NLP-based branch-and-bound (Section 3.3.1), and outer approxima-
tion (Section 3.2.1) algorithms. It also implements a hybrid of outer approximation and LP/NLP-
based branch-and-bound. It features several primal heuristics, including the feasibility pump,
diving, and RINS. It uses the CBC solver (https://projects.coin-or.org/Cbc) for per-
forming all the MILP operations, such as management of cuts and tree-search. It can solve NLPs
using IPOPT (Wachter and Biegler, 2006) or Filter-SQP (Fletcher and Leyffer, 1998). Source code
and documentation are available at the website (https://projects.coin-or.org/Bonmin).
DICOPT stands for Discrete and Continuous Optimizer. It implements a variant of outer ap-
proximation (Section 3.2.1), which has been generalized to tackle nonconvex MINLPs through a
penalty function heuristic; see (Viswanathan and Grossmann, 1990). The problem is input through
the GAMS modeling system. The user can specify options to select both the MILP solver and the
NLP solver at each iteration of the algorithm.
FilMINT (Abhishek et al., 2010) implements the LP/NLP-BB algorithm (Section 3.3.1) with sev-
eral practical improvements. It exploits the presolve, cutting planes, searching rules, and other
MILP tools of the MINTO solver (Nemhauser et al., 1994). Filter-SQP (Fletcher and Leyffer, 1998)
is used to solve NLPs. It also implements some of the disjunctive cuts described in Section 4.3 and
the feasibility pump heuristic (Section 6.1.2).
MILANO is a MATLAB-based solver for convex MINLPs. It implements the nonlinear branch-
and-bound (Section 3.1) and outer approximation (Section 3.2.1). The source code of MILANO
is available on the project website (http://www.pages.drexel.edu/hvb22/milano). The
main focus of this code is to develop efficient warm-starting methods for interior-point methods
(Benson, 2011, 2012) so as to make them more effective in solving MINLPs.
MINLPBB (Leyffer, 1998) is a Fortran-based nonlinear branch-and-bound solver (Section 3.1)

for convex MINLPs. Filter-SQP (Fletcher and Leyffer, 1998) is used to solve the NLP relaxations.
It also has the ability to restart the NLP iterations from different remote points in order to ensure
better solutions for nonconvex MINLPs, and it provides options for choosing different branching-
rules and tree-search strategies, see Sections 3.1.1 and 3.1.2.
MINOPT is a framework for both modeling and solving MINLPs and MIOCPs Developed in
1998, MINOPT (Schweiger, 1999). It implements generalized Benders decomposition (Section 3.2.2)
and outer approximation (Section 3.2.1). MINOPT requires linking with an NLP solver and an
MILP solver for which it has built-in routines for different solvers. License for MINOPT can be
obtained by contacting the authors. More information, a reference manual and examples are avail-
able on the project website (http://titan.princeton.edu/MINOPT).
MINOTAUR stands for Mixed-Integer Nonlinear Optimization Toolkit: Algorithms, Under-

estimators and Relaxations. It is a new open-source toolkit for MINLPs. Currently, it only
implements nonlinear branch-and-bound (Section 3.1) and QP-Diving (Section 3.1.3) for convex
MINLPs. It has interfaces to NLP, QP and LP solvers. MINOTAUR has the ability to create
and modify computational graphs of nonlinear functions. It can be used to reformulate non-
linear constraints and objective functions. The source code and documentation is available online
(http://wiki.mcs.anl.gov/minotaur/).
MISQP is a solver designed for practical problems where the nonlinear functions cannot be
evaluated when the variables xi , i I are not integers. This solver evaluates the functions and
derviates at the integer points only. The algorithm does not guarantee an optimal solution even for
convex MINLP. It generalizes the sequential-quadratic programming method with a trust region
(Exler and Schittkowski, 2007) to MINLPs. Exler et al. (2012) provide documentation of the solver
along with examples. More information is available online (http://www.ai7.uni-bayreuth.
de/misqp.htm).
SBB stands for Simple Branch-and-Bound. Implemented in GAMS, it allows the user to choose
an NLP solver for the NLP-BB algorithm (Section 3.1). It also can handle SOS1 and SOS2 (see (7.2)
and (5.8)) constraints. A user manual is available online (http://www.gams.com/dd/docs/
solvers/sbb.pdf).
8.2 Nonconvex MINLP solvers
Most modern MINLP solvers designed for nonconvex problems utilize a combination of the tech-
niques outlined in the previous sections; in particular, they are branch-and-bound algorithms with
at least one rudimentary bound-tightening technique and a lower-bounding procedure. The tech-
nique for factorable functions described in Section 5.2.1 is most commonly used.
-BB is a branch-and-bound solver that uses the -convexification (Adjiman et al., 1998) and its
variants to obtain quadratic underestimators of nonconvex functions in the constraints and ob-
jective. A number of special underestimators are available for specific commonly used functions.
More information is available on the project website (http://titan.princeton.edu/tools).
BARON is the acronym for Branch And Reduce Optimization Navigator (Sahinidis, 1996).
It implements a branch-and-bound algorithm that computes a lower bound at each subproblem
by means of a linear relaxation of (1.1), as discussed in Section 5.2. It includes various bound
tightening techniques (Section 5.3.2) such as probing and the violation transfer outlined in Section
5.3.1. It is available through the GAMS and AIMMS modeling systems. More information, ex-
amples, and documentation are available at the project website (http://archimedes.cheme.
cmu.edu/?q=baron).
COCONUT is an open-source environment for global optimization problems (Coconut, 2004;

Schichl, 2004). Although it does not solve problems with integer variables, we include it in this
section because it uses various techniques common to MINLP solvers: bounds tightening (Sec-
tion 5.3.2), reformulation (Section 5.2), and heuristics. The source code and documentation are
available at the project homepage (http://www.mat.univie.ac.at/coconut/coconut-
environment).
COUENNE or Convex Over- and Under-ENvelopes for Nonlinear Estimation (Belotti, 2009)
is an open-source branch-and-bound algorithm that, similarly to BARON, obtains a lower bound
through an LP relaxation using the reformulation technique outlined in Section 5.2.1. It also im-
plements several bound tightening procedures (Section 5.3.2) as well as a recently introduced
feasibility-pump heuristic (Section 6.1.2), a separator of disjunctive cuts (Section 5.2.2), and dif-
ferent branching schemes including strong, pseudocost, and reliability branching (Section 5.3.1).
Recently, it also introduced the linear cuts described by Qualizza et al. (2012) and briefly discussed
in Section 5.4.1. Source code and documentation are available online (https://projects.
coin-or.org/Couenne).
GloMIQO is an evolution of -BB with additional algorithms based on the work of Misener and
Floudas (2012b). It dramatically improves the lower bounding procedure used originally in the -
BB method. It can solve only quadratically constrained quadratic problems. The solver is available
through the GAMS modeling system. Related publications and more information are available on
the project website (http://helios.princeton.edu/GloMIQO/publications.html).
LaGO is a branch-and-bound algorithm that is guaranteed to return the global optimum for
mixed-integer quadratic problems (Nowak et al., 2003). It uses -convexification (see Section 5.3)
to obtain a lower bound on each subproblem. Though this technique is guaranteed only to solve
quadratic problems to global optimality, LaGO can be used as a heuristic in other cases. Source-
code and documentation are available online (https://projects.coin-or.org/LaGO).
LGO or the Lipschitz (Continuous) Global Optimizer implements a set of heuristics and exact
methods for global optimization. A constrained local optimization approach is used to obtain up-
per bounds on the objective value. Lower bounds are estimated through sampling. LGO assumes
that the functions in the objective and the constraints of the problem are Lipschitz-continuous.
Thus, it can find a global solution when the Lipschitz-constants for all functions in the prob-
lem are known. One advantage of LGO is that it does not require gradients of the functions
and hence can be applied to problems where the functions are not explicitly known: they could
come from a black box or a simulation. License for LGO can be purchased from the website
(http://www.pinterconsulting.com).
LindoGlobal is the MINLP solver of the LINGO modeling suite (Lin and Schrage, 2009). It is
in the same class as BARON and COUENNE in that it employs a lower bounding technique for
factorable functions, as described in Section 5.2.1. This solver can be purchased through LINDO
Systems (http://www.lindo.com).
SCIP started as a MILP solver (Achterberg, 2005) but evolved first into a solver for MINLPs
with quadratic objective function and constraints (Berthold et al., 2010) and, more recently, into
a solver for nonconvex MINLP (Berthold et al., 2012). Following the basic approach of BARON
and COUENNE, it implements a branch-and-bound algorithm (Section 5.2.1) with linear relax-
ation, various heuristics, and bound-tightening procedures. Source code and documentation are
available at the project website (http://scip.zib.de).
8.3 An MIOCP Solver
MUSCOD-II started as a reference implementation of direct multiple shooting, a direct and all-
at-once method for ODE-constrained optimal control problems (Bock and Plitt, 1984), and was
later extended to DAE-constrained problems (Leineweber et al., 2003) and mixed-integer opti-
mal control problems (Sager, 2005). Kirches and Leyffer (2011) propose an extension for model-
ing MIOCPs in the symbolic modeling language AMPL and present an interface to MUSCOD-
II. More information about this solver is available from its website on NEOS (http://www.
neos-server.org/neos/solvers/miocp:MUSCOD-II/AMPL.html).
8.4 Modeling Languages and Online Resources
Diverse modeling languages and online resources make it easy to specify and solve MINLP prob-
lems without needing to install software, code nonlinear functions, or derivatives. The wide avail-
ability of these tools means that MINLP has become accessible to the broader scientific and engi-
neering community. In this section, we briefly summarize these tools.
Modeling languages enable scientists and engineers to express optimization problems in a more
natural algebraic form that is close to a mathematical representation of the problem. Most mod-
eling languages include automatic differentiation tools (Griewank, 2000) that provide (exact) first
and second derivatives of the problem functions and relieve the user of the error-prone tasks of
coding derivatives. The most popular modeling languages are AIMMS (Bisschop and Entriken,
1993), AMPL (Fourer et al., 1993), GAMS (Brooke et al., 1992), MOSEL (Colombani and Heipcke,
2002), and TomLab (Holmstrom et al., 2010). TomLab is built on top of MATLAB, while the other
systems are domain-specific languages that define a syntax for specifying optimization problems
that can be parsed by the respective system to provide function and derivative information to the
solver through a back-end. Recently, an open-source modeling system, Pyomo (Hart et al., 2011),
that enables users to express MINLPs using the PYTHON scripting language has been developed.
Opti Toolbox (Currie and Wilson, 2012) is another open-source modeling tool. It enables users
to express and solve MINLPs from within the MATLAB environment. Currently, users can call
BONMIN and SCIP solvers from this toolbox.
Online resources for optimization have grown dramatically over the past 15 years. There exist
many libraries of test or benchmark problems in AMPL (http://wiki.mcs.anl.gov/leyffer/
index.php/MacMINLP and http://minlp.org/) and GAMS (http://www.gamsworld.org/
minlp/ and http://minlp.org/). Arguably the most important factor in making optimization
solvers widely available has been the NEOS server (Czyzyk et al., 1998). Many of the solvers de-
scribed above are now available on NEOS (http://www.neos-server.org/neos/. NEOS
provides a collection of state-of-the-art optimization software. Optimization problems are sub-
mitted through a web interface (or from within a modeling language session) and solved re-
motely. The MINLP solvers available on NEOS are AlphaECP, BARON, Bonmin, Couenne, DI-
COPT, FilMINT, LINDOGlobal, MINLPBB, SBB, and SCIP.
Acknowledgments
This work was supported by the Office of Advanced Scientific Computing Research, Office of
Science, U.S. Department of Energy, under Contract DE-AC02-06CH11357. This work was also
supported by the U.S. Department of Energy through grant DE-FG02-05ER25694. This work was
also supported through NSF grant CCF-0830035.
References
Abhishek, K., Leyffer, S., and Linderoth, J. T. (2010). FilMINT: An outer-approximation-based
solver for nonlinear mixed integer programs. INFORMS Journal on Computing, 22:555567.
DOI:10.1287/ijoc.1090.0373.
Abichandani, P., Benson, H., and Kam, M. (2008). Multi-vehicle path coordination under commu-
nication constraints. In American Control Conference, pages 650656.
Abramson, M., Audet, C., Chrissis, J., and Walston, J. (2009). Mesh adaptive direct search algo-
rithms for mixed variable optimization. Optimization Letters, 3:3547.
Abramson, M. A. (2004). Mixed variable optimization of a load-bearing thermal insulation system

using a filter pattern search algorithm. Optimization and Engineering, 5:157177.
Achterberg, T. (2005). SCIP a framework to integrate constraint and mixed integer pro-
Informationstechnik
gramming. Technical Report ZIB-Report 04-19, Konrad-Zuse-Zentrum fur
Berlin, Takustr. 7, Berlin.
Achterberg, T. and Berthold, T. (2007). Improving the feasibility pump. Discrete Optimization,
4(1):7786.
Achterberg, T., Koch, T., and Martin, A. (2004). Branching rules revisited. Operations Research
Letters, 33:4254.
Adams, W. (2011). Use of Lagrange interpolating polynomials in the RLT. Wiley Encyclopedia of
Operations Research and Management Science.
Adams, W. and Sherali, H. (1986). A tight linearization and an algorithm for zero-one quadratic
programming problems. Management Science, 32(10):12741290.
Adams, W. and Sherali, H. (2005). A hierarchy of relaxations leading to the convex hull represen-
tation for general discrete optimization problems. Annals of Operations Research, 140(1):2147.
Adjiman, C. S., Androulakis, I., and Floudas, C. (1998). A global optimization method, BB, for
general twice-differentiable constrained NLPs - II. implementation and computational results.
Computers & Chemical Engineering, 22:11591179.
Akrotirianakis, I., Maros, I., and Rustem, B. (2001). An outer approximation based branch-and-cut
algorithm for convex 0-1 MINLP problems. Optimization Methods and Software, 16:2147.
Al-Khayyal, F. A. and Falk, J. E. (1983). Jointly constrained biconvex programming. Mathematics

of Operations Research, 8:273286.
Altunay, M., Leyffer, S., Linderoth, J. T., and Xie, Z. (2011). Optimal security response to attacks
on open science grids. Computer Networks, 55:6173.
Andersen, E. D. and Andersen, K. D. (1995). Presolving in linear programming. Mathematical

Programming, 71:221245.
Androulakis, I. P., Maranas, C. D., and Floudas, C. A. (1995). BB : A global optimization method
for general constrained nonconvex problems. Journal of Global Optimization, 7:337363.
Anstreicher, K. (2012). On convex relaxations for quadratically constrained quadratic program-

ming. Mathematical Programming, pages 119. 10.1007/s10107-012-0602-3.
Anstreicher, K. M. (2009). Semidefinite programming versus the reformulation-linearization tech-

nique for nonconvex quadratically constrained quadratic programming. Journal of Global Opti-
mization, 43:471484.

Atamturk, A. and Narayanan, V. (2010). Conic mixed-integer rounding cuts. Mathematical Pro-
gramming A, 122(1):120.
Audet, C. and Dennis, Jr., J. E. (2000). Pattern search algorithms for mixed variable programming.
SIAM Journal on Optimization, 11(3):573594.
Bacher, R. (10-12 December 1997). The Optimal Power Flow (OPF) and its solution by the interior
point approach. EES-UETP Madrid, Short Course.
Baes, M., Del Pia, A., Nesterov, Y., Onn, S., and Weismantel, R. (2012). Minimizing lipschitz-
continuous strongly convex functions over integer points in polytopes. Mathematical Program-
ming, 134:305322. 10.1007/s10107-012-0545-8.
Balakrishnan, A. and Graves, S. (1989). A composite algorithm for a concave-cost network flow
problem. Networks, 19(2):175202.
Balaprakash, P., Wild, S. M., and Hovland, P. D. (2011). Can search algorithms save large-scale
automatic performance tuning? Procedia Computer Science (ICCS 2011), 4:21362145.
Balas, E., Ceria, S., and Cornuejols, G. (1993). A lift-and-project cutting plane algorithm for mixed
01 programs. Mathematical Programming, 58:295324.
Balas, E., Ceria, S., and Cornuejols, G. (1996). Mixed 0-1 programming by lift-and-project in a
branch-and-cut framework. Management Science, 42:12291246.
Bao, X., Sahinidis, N., and Tawarmalani, M. (2009). Multiterm polyhedral relaxations for noncon-
vex quadratically constrained quadratic programs. Optimization Methods and Software, 24:485
504.
Bartelt-Hunt, S., Culver, T., Smith, J., Matott, L. S., and Rabideau, A. (2006). Optimal design of
a compacted soil liner containing sorptive amendments. Journal of Environmental Engineering,
132(7):769776.
Bartholomew, E. F., ONeill, R. P., and Ferris, M. C. (2008). Optimal transmission switching. IEEE
Transactions on Power Systems, 23:13461355.
Bauschke, H. H. and Borwein, J. M. (1996). On projection algorithms for solving convex feasibility
problems. SIAM Review, 38(3):367426.
Beale, E. and Tomlin, J. (1970). Special facilities in a general mathematical programming system
for nonconvex problems using ordered sets of variables. In Lawrence, J., editor, Proceedings of
the 5th International Conference on Operations Research, pages 447454, Venice, Italy.
Beale, E. M. L. and Forrest, J. J. H. (1976). Global optimization using special ordered sets. Mathe-
matical Programming, 10:5269.
Bellman, R. (1961). On the approximation of curves by line segments using dynamic program-
ming. Commun. ACM, 4(6):284.
Belotti, P. (2009). COUENNE: a users manual. Technical report, Lehigh University.
Belotti, P. (2012a). Bound reduction using pairs of linear inequalities. Journal of Global Optimization.
DOI :10.1007/s10898-012-9848-9.
Belotti, P. (2012b). Disjunctive cuts for non-convex MINLP. In Mixed Integer Nonlinear Pro-
gramming, volume 154 of IMA Volume Series in Mathematics and its Applications, pages 117144.
Springer.
Belotti, P., Cafieri, S., Lee, J., and Liberti, L. (2010). Feasibility-based bounds tightening via fixed
points. In Wu, W. and Daescu, O., editors, Combinatorial Optimization and Applications, volume
6508 of Lecture Notes in Computer Science, pages 6576. Springer Berlin / Heidelberg.

Belotti, P., Goez,
J., Polik, I., Ralphs, T., and Terlaky, T. (2012). A conic representation of the
convex hull of disjunctive sets and conic cuts for integer second order cone optimization. Tech-
nical report, n. 12T-009, Lehigh University, Department of Industrial and Systems Engineering.
http://www.optimization-online.org/DB_FILE/2012/06/3494.pdf.
Belotti, P., Lee, J., Liberti, L., Margot, F., and Wachter, A. (2009). Branching and bounds tightening
techniques for non-convex MINLP. Optimization Methods and Software, 24(4-5):597634.
Ben-Tal, A. and Nemirovski, A. (1995). Optimal design of engineering structures. Optima, 47:48.
Ben-Tal, A. and Nemirovski, A. (2001). On polyhedral approximations of the second-order cone.

Mathematics of Operations Research, 26(2):193205.
Benson, H. Y. (2011). Mixed integer nonlinear programming using interior point methods. Opti-
mization Methods and Software, 26(6):911931.
Benson, H. Y. (2012). Using interior-point methods within an outer approximation framework

for mixed integer nonlinear programming. In Mixed Integer Nonlinear Programming, The IMA
Volumes in Mathematics and its Applications, pages 225243.
Berthold, T. (2012). RENS - the optimal rounding. ZIB-Report 12-17, Zuse Institut Berlin.
Berthold, T., Gamrath, G., Gleixner, A., Heinz, S., Koch, T., and Shinano, Y. (2012). Solving mixed
integer linear and nonlinear problems using the SCIP optimization suite. ZIB-Report 12-27,
Zuse Institut Berlin.
Berthold, T., Gleixner, A., Heinz, S., Koch, T., and Vigerske, S. (2010). Extending SCIP for solving
MIQCPs. In Proceedings of the European Workshop on Mixed Integer Nonlinear Programming, pages
181196.
Berthold, T. and Gleixner, A. M. (2012). Undercover: a primal MINLP heuristic exploring a largest
sub-MIP. ZIB-Report 12-07, Zuse Institut Berlin.
Bertsekas, D. and Gallager, R. (1987). Data Networks. Prentice-Hall, Endlewood Cliffs, NJ.
Bhatia, R., Segall, A., and Zussman, G. (2006). Analysis of bandwidth allocation algorithms for
wireless personal area networks. Wireless Networks, 12:589603.
Bienstock, D. (1996). Computational study of a family of mixed-integer quadratic programming

problems. Mathematical Programming, 74:121140.
Bienstock, D. and Mattia, S. (2007). Using mixed-integer programming to solve power grid black-
out problems. Discrete Optimization, 4:115141.
Bier, V. M. (2005). Game-theoretic and reliability methods in counterterrorism and security. In

Wilson, A., Limnios, N., Keller-McNulty, S., and Armijo, Y., editors, Mathematical and Statistical
Methods in Reliability, Series on Quality, Reliability and Engineering Statistics, pages 1728. World
Scientific, Singapore.
Bier, V. M., Nagaraj, A., and Abhichandani, V. (2005). Protection of simple series and parallel
systems with components of different values. Reliability Engineering System Safety, 87(3):315
323.
Bier, V. M., Oliveros, S., and Samuelson, L. (2007). Choosing what to protect. J. Public Economic
Theory, 9(4):563587.
Bisschop, J. and Entriken, R. (1993). AIMMS The Modeling System. Paragon Decision Technology.
Bock, H. and Longman, R. (1985). Computation of optimal controls on disjoint control sets for
minimum energy subway operation. Advances in the Astronautical Sciences, 50:949972. Proceed-
ings of the American Astronomical Society Symposium on Engineering Science and Mechanics,
Taiwan, 1982.
Bock, H. and Plitt, K. (1984). A multiple shooting algorithm for direct solution of optimal control
problems. In Proceedings of the 9th IFAC World Congress, pages 242247, Budapest. Pergamon
Press.
uk,
Bonami, P. (2011). Lift-and-project cuts for mixed integer convex programs. In Gunl O. and
Woeginger, G., editors, Integer Programming and Combinatoral Optimization, volume 6655 of Lec-
ture Notes in Computer Science, pages 5264. Springer, Berlin.
Bonami, P., Biegler, L., Conn, A., Cornuejols, G., Grossmann, I., Laird, C., Lee, J., Lodi, A., Margot,
F., Sawaya, N., and Wachter, A. (2008). An algorithmic framework for convex mixed integer
nonlinear programs. Discrete Optimization, 5(2):186204.
Bonami, P., Cornuejols, G., Lodi, A., and Margot, F. (2009). A feasibility pump for mixed integer
nonlinear programs. Mathematical Programming, 119:331352.
Bonami, P. and Goncalves, J. P. M. (2012). Heuristics for convex mixed integer nonlinear programs.
Computational Optimization and Applications, 51:729747.
Bonami, P., Klnc, M., and Linderoth, J. (2012). Algorithms and software for convex mixed integer
nonlinear programs. IMA Volumes in Mathematics and its Applications, 154:6192.
Bonami, P., Lee, J., Leyffer, S., and Wachter, A. (2011). More branch-and-bound experiments in
convex nonlinear integer programming. Preprint ANL/MCS-P1949-0911, Argonne National
Laboratory, Mathematics and Computer Science Division.
Bongartz, I., Conn, A. R., Gould, N. I. M., and Toint, P. L. (1995). CUTE: Constrained and Uncon-
strained Testing Enviroment. ACM Transactions on Mathematical Software, 21:123160.
Boorstyn, R. and Frank, H. (1977). Large-scale network topological optimization. IEEE Transactions
on Communications, 25:2947.
Borchers, B. and Mitchell, J. E. (1994). An improved branch and bound algorithm for mixed integer
nonlinear programs. Computers & Operations Research, 21:359368.
Boyd, S. and Vandenberghe, L. (2004). Convex Optimization. Cambridge University Press, Cam-
bridge.
Bragalli, C., DAmbrosio, C., Lee, J., Lodi, A., and Toth, P. (2006). An MINLP solution method for
a water network problem. In Algorithms - ESA 2006 (14th Annual European Symposium. Zurich,
Switzerland, September 2006, Proceedings), pages 696707. Springer.
Bragalli, C., DAmbrosio, C., Lee, J., Lodi, A., and Toth, P. (2012). On the optimal design of water
distribution networks: a practical minlp approach. Optimization and Engineering, 13:219246.
Brooke, A., Kendrick, D., Meeraus, A., and Raman, R. (1992). GAMS, A Users Guide. GAMS
Development Corporation.
Bryson, A. and Ho, Y.-C. (1975). Applied Optimal Control. Wiley, New York.
Buchheim, C. and Wiegele, A. (2012). Semidefinite relaxations for non-convex quadratic mixed-
integer programming. Mathematical Programming, pages 118. 10.1007/s10107-012-0534-y.
Burer, S. (2009). On the copositive representation of binary and continuous nonconvex quadratic
programs. Mathematical Programming, 120:479495.
Burer, S. and Letchford, A. (2009). On nonconvex quadratic programming with box constraints.
SIAM Journal on Optimization, 20(2):1073 89.
Burer, S. and Letchford, A. (2012a). Non-convex mixed-integer nonlinear programming: A survey.

Surveys in Operations Research and Management Science, 17:97106.
Burer, S. and Letchford, A. (2012b). Unbounded convex sets for non-convex mixed-integer
quadratic programming. Mathematical Programming, pages 126. 10.1007/s10107-012-0609-9.
Burer, S. and Vandenbussche, D. (2009). Globally solving box-constrained nonconvex quadratic

programs with semidefinite-based finite branch-and-bound. Computational Optimization and Ap-
plications, 43(2):181 195.
Burgschweiger, J., Gnadig, B., and Steinbach, M. (2008). Optimization models for operative plan-
ning in drinking water networks. Optimization and Engineering, 10(1):4373.
Bussieck, M. R. and Vigerske, S. (2010). MINLP solver software. In Cochran, J. J., Cox, L. A., Ke-
skinocak, P., Kharoufeh, J. P., Jeffrey, P., and Smith, J. C., editors, Wiley Encyclopedia of Operations
Research and Management Science. Wiley.
Callegari, S., Bizzarri, F., Rovatti, R., and Setti, G. (2010). On the approximate solution of a class
of large discrete quadratic programming problems by modulation: The case of circulant
quadratic forms. IEEE Transactions on Signal Processing, 58(12):61266139.
Castillo, I., Westerlund, J., Emet, S., and Westerlund, T. (2005). Optimization of block layout deisgn
problems with unequal areas: A comparison of MILP and MINLP optimization methods. Com-
puters & Chemical Engineering, 30:5469.
ezik, M. and Iyengar, G. (2005). Cuts for mixed 0-1 conic programming. Mathematical Program-
C
ming A, 104:179202.
Ceria, S. and Soares, J. (1999). Convex programming for disjunctive optimization. Mathematical
Programming, 86:595614.
Chi, K., Jiang, X., Horiguchi, S., and Guo, M. (2008). Topology design of network-coding-based
multicast networks. IEEE Transactions on Mobile Computing, 7(4):114.
Chung, K., Richard, J.-P., and Tawarmalani, M. (2011). Lifted inequalities for 0-1 mixed-integer
bilinear covering sets. Available at http://www.optimization-online.org/DB_FILE/
2011/03/2949.pdf.
Chvatal, V. (1973). Edmonds polytopes and a hierarchy of combinatorial problems. Discrete Math-
ematics, 4:305337.
Coconut (2004). The COCONUT benchmark: A benchmark for global optimization and constraint
satisfaction. http://www.mat.univie.ac.at/neum/glopt/coconut/benchmark.
html.
Cohen, J. S. (2003). Computer algebra and symbolic computation: elementary algorithms. Universities
Press.
Colombani, Y. and Heipcke, S. (2002). Mosel: An extensible environment for modeling and pro-
gramming solutions. In Jussien, N. and Laburthe, F., editors, Proceedings of the Fourth Interna-
tional Workshop on Integration of AI and OR Techniques in Constraint Programming for Combinatorial
Optimisation Problems (CP-AI-OR02), pages 277290.
Costa-Montenegro, E., Gonzalez-Castano, F. J., Rodriguez-Hernandez, P. S., and Burguillo-Rial,

J. C. (2007). Nonlinear optimization of IEEE 802.11 mesh networks. In ICCS 2007, Part IV, pages
466473, Springer Verlag, Berlin.
Croxton, K., Gendron, B., and Magnanti, T. (Sept. 2003). A comparison of mixed-integer program-
ming models for nonconvex piecewise linear cost minimization problems. Management Science,
49:126873.
Currie, J. and Wilson, D. I. (2012). OPTI: Lowering the Barrier Between Open Source Optimizers
and the Industrial MATLAB User. In Sahinidis, N. and Pinto, J., editors, Foundations of Computer-
Aided Process Operations, Savannah, Georgia, USA.
Czyzyk, J., Mesnier, M., and More, J. (1998). The NEOS server. IEEE Journal on Computational
Science and Engineering, 5:6875.
Dadush, D., Dey, S., and Vielma, J. P. (2011a). The split closure of a strictly convex body. Operations
Research Letters, 39(2):121 126.
Dadush, D., Dey, S. S., and Vielma, J. P. (2011b). The chvatal-gomory closure of a strictly convex
body. Mathematics of Operations Research, 36(2):227 239.
Dadush, D., Dey, S. S., and Vielma, J. P. (2011c). On the chvatal-gomory closure of a compact
convex set. In Lecture Notes in Computer Science, volume 6655 LNCS, pages 130 142, New York,
NY. Springer.
Dadush, D., Peikert, C., and Vempala, S. (2011d). Enumerative lattice algorithms in any norm
via m-ellipsoid coverings. In Foundations of Computer Science (FOCS), 2011 IEEE 52nd Annual
Symposium on, pages 580 589.
Dakin, R. J. (1965). A tree search algorithm for mixed programming problems. Computer Journal,
8:250255.
DAmbrosio, C. and Liberti, L. (2012). A storm of feasibility pumps for nonconvex MINLP. Math-
ematical Programming Series B. To appear, available online.
DAmbrosio, C. and Lodi, A. (2011). Mixed integer nonlinear programming tools: a practical
overview. 4OR, 9(4):329349.
DAmbrosio, C., Lodi, A., and Martello, S. (2010). Piecewise linear approximation of functions of
two variables in MILP models. Operations Research Letters, 38(1):3946.
Danna, E., Rothberg, E., and LePape, C. (2005). Exploring relaxation induced neighborhoods to
improve MIP solutions. Mathematical Programming, 102:7190.
Dantzig, G. B. (1960). On the significance of solving linear programming problems with some
integer variables. Econometrica, 28(1):3044.
Dantzig, G. B. (1963). Linear Programming and Extensions. Princeton University Press, Princeton,
NJ.
Davis, E. (1987). Constraint propagation with interval labels. Artificial Intelligence, 32(3):281331.
Davis, E. and Ierapetritou, M. (2009). A kriging based method for the solution of mixed-integer
nonlinear programs containing black-box functions. Journal of Global Optimization, 43(2-3):191
205.
De Loera, J. A., Hemmecke, R., Koppe, M., and Weismantel, R. (2006). Integer polynomial opti-
mization in fixed dimension. Mathematics of Operations Research, 31(1):147 153.
Dey, S. S. and Moran R., D. A. (2012). Some properties of convex hulls of integer points contained
in general convex sets. Mathematical Programming, pages 1 20.
Dey, S. S. and Vielma, J. P. (2010). The chvatal-gomory closure of an ellipsoid is a polyhedron. In

Lecture Notes in Computer Science, volume 6080 LNCS, pages 327 340, Lausanne. Springer.
Dolan, E., Fourer, R., More, J., and Munson, T. (2002). Optimization on the NEOS server. SIAM
News, 35:89.
Dolan, E. and More, J. (2002). Benchmarking optimization software with performance profiles.
Mathematical Programming, 91:201213.
Donde, V., Lopez, V., Lesieutre, B., Pinar, A., Yang, C., and Meza, J. (2005). Identification of severe
multiple contingencies in electric power networks. In Proceedings 37th North American Power
Symposium. LBNL-57994.
Dorigo, M., Maniezzo, V., and Colorni, A. (1996). The ant system: optimization by a colony of
cooperating agents. IEEE Transactions on Systems, Man and Cybernetics - Part B, 26(1):113.
Drewes, S. (2009). Mixed Integer Second Order Cone Programming. PhD thesis, Technische Univer-
sitat Darmstadt.
Drewes, S. and Ulbrich, S. (2012). Subgradient based outer approximation for mixed integer sec-
ond order cone programming. In Mixed Integer Nonlinear Programming, volume 154 of The IMA
Volumes in Mathematics and its Applications, pages 4159. Springer, New York. ISBN 978-1-4614-
1926-6.
Duran, M. A. and Grossmann, I. (1986). An outer-approximation algorithm for a class of mixed-

integer nonlinear programs. Mathematical Programming, 36:307339.
Eckstein, J. (1994). Parallel branch-and-bound algorithms for general mixed integer programming
on the CM-5. SIAM Journal on Optimization, 4:794814.
Eiger, G., Shamir, U., and Ben-Tal, A. (1994). Optimal design of water distribution networks. Water
Resources Research, 30(9):26372646.
Elhedhli, S. (2006). Service System Design with Immobile Servers, Stochastic Demand, and Con-
gestion. Manufacturing & Service Operations Management, 8(1):9297.
Eliceche, A. M., Corvalan, S. M., and Martnez, P. (2007). Environmental life cycle impact as a tool
for process optimisation of a utility plant. Computers & Chemical Engineering, 31:648656.
Elwalid, A., Mitra, D., and Wang, Q. (2006). Distributed nonlinear integer optimization for data-
optical internetworking. IEEE Journal on Selected Areas in Communications, 24(8):15021513.
Engelhart, M., Funke, J., and Sager, S. (2012). A decomposition approach for a new test-scenario
in complex problem solving. Journal of Computational Science. (to appear).
Exler, O., Lehmann, T., and Schittkowski, K. (2012). MISQP: A fortran subroutine of a trust re-
gion SQP algorithm for mixed-integer nonlinear programming - users guide. Technical report,
Department of Computer Science, University of Bayreuth.
Exler, O. and Schittkowski, K. (2007). A trust region SQP algorithm for mixed-integer nonlinear
programming. Optimization Letters, 1:269280.
FICO Xpress (2009). FICO Xpress Optimization Suite: Xpress-BCL Reference manual. Fair Isaac Cor-
poration.
Fischetti, M., Glover, F., and Lodi, A. (2005). The feasibility pump. Mathematical Programming,
104:91104.
Fischetti, M. and Lodi, A. (2003). Local branching. Mathematical Programming, 98:2347.
Fischetti, M. and Salvagnin, D. (2009). Feasibility pump 2.0. Mathematical Programming Computa-
tions, 1:201222.
Fletcher, R. (1987). Practical Methods of Optimization. John Wiley & Sons, Chichester.
Fletcher, R. and Leyffer, S. (1994). Solving mixed integer nonlinear programs by outer approxima-
tion. Mathematical Programming, 66:327349.
Fletcher, R. and Leyffer, S. (1998). User manual for filterSQP. University of Dundee Numerical
Analysis Report NA-181.
Fletcher, R. and Leyffer, S. (2003). Filter-type algorithms for solving systems of algebraic equations
and inequalities. In di Pillo, G. and Murli, A., editors, High Performance Algorithms and Software
for Nonlinear Optimization, pages 259278. Kluwer, Dordrecht.
Flores-Tlacuahuac, A. and Biegler, L. T. (2007). Simultaneous mixed-integer dynamic optimization

for integrated design and control. Computers & Chemical Engineering, 31:648656.
Floudas, C. (1995). Nonlinear and Mixed-Integer Optimization. Topics in Chemical Engineering.

Oxford University Press, New York.
Floudas, C. A. (2000). Deterministic Global Optimization: Theory, Algorithms and Applications. Kluwer
Academic Publishers.
Fourer, R., Gay, D. M., and Kernighan, B. W. (1993). AMPL: A Modeling Language for Mathematical
Programming. The Scientific Press.
Fowler, K. R., Reese, J. P., Kees, C. E., Dennis, J. E., Kelley, C. T., Miller, C. T., Audet, C., Booker,
A. J., Couture, G., Darwin, R. W., Farthing, M. W., Finkel, D. E., Gablonsky, J. M., Gray, G. A.,
and Kolda, T. G. (2008). A comparison of derivative-free optimization methods for water supply
and hydraulic capture community problems. Advances in Water Resources, 31(5):743757.
Frangioni, A. and Gentile, C. (2006). Perspective cuts for a class of convex 0-1 mixed integer
programs. Mathematical Programming, 106:225236.

Fugenschuh, A., Herty, M., Klar, A., and Martin, A. (2006). Combinatorial and continuous models
for the optimization of traffic flows on networks. SIAM Journal on Optimization, 16(4):11551176.
Fuller, A. (1963). Study of an optimum nonlinear control system. Journal of Electronics and Control,
15:6371.
Garver, L. L. (1997). Transmission network estimation using linear programming. IEEE Transac-
tions on Power Apparatus Systems, 89:16881697.
Gay, D.M. (1991). Automatic differentiation of nonlinear AMPL models. In Griewank, A. and
Corliss, G.F., editors, Automatic Differentiation of Algorithms: Theory, Implementation, and Applica-
tion, SIAM Proceedings, pages 6173, Philadelphia. SIAM.
Geissler, B., Martin, A., Morsi, A., and Schewe, L. (2012). Using piecewise linear functions for
solving MINLPs. In Lee, J. and Leyffer, S., editors, Mixed Integer Nonlinear Programming, volume
154 of The IMA Volumes in Mathematics and its Applications, pages 287314. Springer New York.
Gentilini, I., Margot, F., and Shimada, K. (2012). The travelling salesman problem with
neighbourhoods: MINLP solution. Optimization Methods and Software, pages 115. DOI:
10.1080/10556788.2011.648932.
Geoffrion, A. M. (1972). Generalized Benders decomposition. Journal of Optimization Theory and

Applications, 10(4):237260.
Geoffrion, A. M. (1977). Objective function approximations in mathematical programming. Math-

ematical Programming, 13:2337.
Gerdts, M. (2005). Solving mixed-integer optimal control problems by Branch&Bound: A case

study from automobile test-driving with gear shift. Optimal Control Applications and Methods,
26:118.
Gerdts, M. and Sager, S. (2012). Mixed-integer DAE optimal control problems: Necessary condi-
tions and bounds. In Biegler, L., Campbell, S., and Mehrmann, V., editors, Control and Optimiza-
tion with Differential-Algebraic Constraints, pages 189212. SIAM.
Glover, F. (1989). Tabu search: part I. ORSA Journal on Computing, 1(3):190206.
Glover, F. (1990). Tabu search: part II. ORSA Journal on Computing, 2(1):432.
Goldberg, D. E. (1989). Genetic algorithms in search, optimization, and machine learning. Addison-
Wesley, Boston.
Goldberg, N., Leyffer, S., and Safro, I. (2012). Optimal response to epidemics and cyber attacks
in networks. Preprint ANL/MCS-1992-0112, Argonne National Laboratory, Mathematics and
Computer Science Division.
Gomory, R. E. (1958). Outline of an algorithm for integer solutions to linear programs. Bulletin of
the American Mathematical Monthly, 64:275278.
Gomory, R. E. (1960). An algorithm for the mixed integer problem. Technical Report RM-2597,
The RAND Corporation.
Gould, N. I. M. and Leyffer, S. (2003). An introduction to algorithms for nonlinear optimization.

In Blowey, J., Craig, A., and Shardlow, T., editors, Frontiers in Numerical Analysis, pages 109197.
Springer Verlag, Berlin.
Gould, N. I. M., Leyffer, S., and Toint, P. L. (2004). A multidimensional filter algorithm for nonlin-
ear equations and nonlinear least squares. SIAM Journal on Optimization, 15(1):1738.
Goux, J.-P. and Leyffer, S. (2003). Solving large MINLPs on computational grids. Optimization and
Engineering, 3:327354.
Griewank, A. (2000). Evaluating Derivatives: Principles and Techniques of Algorithmic Differentiation.

Number 19 in Frontiers in Applied Mathematics. Society for Industrial and Applied Mathemat-
ics, Philadelphia.
Griewank, A. and Toint, P. L. (1984). On the exsistence of convex decompositions of partially

separable functions. Mathematical Programming, 28:2549.
Griva, I., Nash, S. G., and Sofer, A. (2009). Linear and Nonlinear Optimization. SIAM, 2nd edition.
Grossmann, I. E. (2002). Review of nonlinear mixedinteger and disjunctive programming tech-

niques. Optimization and Engineering, 3:227252.
Grossmann, I. E. and Kravanja, Z. (1997). Mixed-integer nonlinear programming: A survey of

algorithms and applications. In L.T. Biegler, T.F. Coleman, A. C. and Santosa, F., editors, Large-
Scale Optimization with Applications, Part II: Optimal Design and Control, Springer, New York.
Guerra, A., Newman, A. M., and Leyffer, S. (2011). Concrete structure design using mixed-
integer nonlinear programming with complementarity constraints. SIAM Journal on Optimiza-
tion, 21(3):833863.
uk,
Gunl O. and Linderoth, J. (2010). Perspective relaxation of mixed integer nonlinear programs
with indicator variables. Mathematical Programming Series B, 104:186203.
O. and Linderoth, J. T. (2012). Perspective reformulation and applications. In IMA Vol-

uk,
Gunl
umes, volume 154, pages 6192.
Gupta, O. K. and Ravindran, A. (1985). Branch and bound experiments in convex nonlinear integer
programming. Management Science, 31:15331546.
Gurobi (2012). Gurobi Optimizer Reference Manual, Version 5.0. Gurobi Optimization, Inc.
Hansen, E. (1992). Global Optimization Using Interval Analysis. Marcel Dekker, Inc., New York.

Harjunkoski, I., Westerlund, T., Porn, R., and Skrifvars, H. (1998). Different transformations for
solving non-convex trim loss problems by MINLP. European Journal of Operational Research,
105:594603.
Hart, W. E., Watson, J.-P., and Woodruff, D. L. (2011). Pyomo: modeling and solving mathematical
programs in Python. Mathematical Programming Computations, 3:219260.
Hedman, K. W., ONeill, R. P., Fisher, E. B., and Oren, S. S. (2008). Optimal transmission switching
- sensitivity analysis and extensions. IEEE Transactions on Power Systems, 23:14691479.
Heinz, S. (2005). Complexity of integer quasiconvex polynomial optimization. Journal of Complex-

ity, 21(4):543 56.
E., Ivarsson, M., Aslund, J., and Nielsen, L. (2009). Look-ahead control for heavy trucks
Hellstrom,
to minimize trip time and fuel consumption. Control Engineering Practice, 17:245254.
Hemker, T. (2008). Derivative Free Surrogate Optimization for Mixed-Integer Nonlinear Black Box Prob-
lems in Engineering. PhD thesis, Technischen Universitat Darmstadt, Darmstadt, Germany.
Hemker, T., Fowler, K., Farthing, M., and von Stryk, O. (2008). A mixed-integer simulation-based
optimization approach with surrogate functions in water resources management. Optimization
and Engineering, 9:341360.
Hemmecke, R., Onn, S., and Weismantel, R. (2011). A polynomial oracle-time algorithm for convex
integer minimization. Mathematical Programming, 126:97117. 10.1007/s10107-009-0276-7.
Hijazi, H., Bonami, P., and Ouorou, A. (2010). An outer-inner approximation for separable
MINLPs. Technical report, LIF, Faculte des Sciences de Luminy, Universite de Marseille.

Hildebrand, R. and Koppe, M. (2010). A new Lenstra-type algorithm for quasiconvex polynomial
integer minimization with complexity 2O(n log n) . arxiv.org/abs/1006.4661.

Holmstrom,
K., Goran, A. O., and Edvall, M. M. (2010). Users guide for TOMLAB 7. Tomlab
Optimization Inc.
Horst, H., Pardalos, P. M., and Thoai, V. (1995). Introduction to Global Optimization. Kluwer, Dor-
drecht.
Horst, R. and Tuy, H. (1993). Global Optimization. Springer-Verlag, New York.
H.W. Lenstra, J. (1983). Integer programming with a fixed number of variables. Mathematics of
Operations Research, 8:538548.
IBM Ilog CPLEX (2009). IBM Ilog CPLEX V12.1: Users Manual for CPLEX. IBM Corp.
Jeroslow, R. and Lowe, J. (1984). Modelling with integer variables. Mathematical Programming
Studies, 22:16784.
Jeroslow, R. and Lowe, J. (1985). Experimental results on the new techniques for integer program-
ming formulations. Journal of the Operational Research Society, 36(5):393403.
Jeroslow, R. G. (1973). There cannot be any algorithm for integer programming with quadratic
constraints. Operations Research, 21(1):221224.
Jobst, N. J., Horniman, M. D., Lucas, C. A., and Mitra, G. (2001). Computational aspects of alterna-
tive portfolio selection models in the presence of discrete asset choice constraints. Quantitative
Finance, 1:489501.

Judice, J. J., Sherali, H. D., Ribeiro, I. M., and Faustino, A. M. (2006). A complementarity-based
partitioning and disjunctive cut algorithm for mathematical programming problems with equi-
librium constraints. Journal of Global Optimization, 36:89114.
Kannan, R. and Monma, C. (1978). On the computational complexity of integer programming

problems. In Henn, R., Korte, B., and Oettli, W., editors, Optimization and Operations Research,,
volume 157 of Lecture Notes in Economics and Mathematical Systems, pages 161172. Springer.
Karuppiah, R. and Grossmann, I. E. (2006). Global optimization for the synthesis of integrated
water systems in chemical processes. Computers & Chemical Engineering, 30:650673.
Keha, A. B., De Farias Jr., I. R., and Nemhauser, G. L. (2006). A branch-and-cut algorithm without
binary variables for nonconvex piecewise linear optimization. Operations Research, 54(5):847
858.
Kelley, J. E. (1960). The cutting plane method for solving convex programs. Journal of the SIAM,
8:703712.
Kennedy, J. and Eberhart, R. (1995). Particle swarm optimization. In IEEE International Conference
on Neural Networks, volume 4, pages 19421948.
Khachiyan, L. and Porkolab, L. (2000). Integer optimization on convex semialgebraic sets. Discrete
& Computational Geometry, 23:207224. 10.1007/PL00009496.
Klnc, M. (2011). Disjunctive Cutting Planes and Algorithms for Convex Mixed Integer Nonlinear
Programming. PhD thesis, Department of Industrial and Systems Engineering, University of
Wisconsin-Madison.
Klnc, M., Linderoth, J., and Luedtke, J. (2010). Effective separation of disjunctive cuts for convex
mixed integer nonlinear programs. Technical Report 1681, Computer Sciences Department,
University of Wisconsin-Madison.
Kirches, C. (2011). Fast numerical methods for mixed-integer nonlinear model-predictive con-
trol. In Bock, H., Hackbusch, W., Luskin, M., and Rannacher, R., editors, Advances in Numerical
Mathematics. Springer Vieweg, Wiesbaden. ISBN 978-3-8348-1572-9. PhD thesis, Ruprecht-Karls-
Universitat Heidelberg.
Kirches, C. and Leyffer, S. (2011). TACO a toolkit for AMPL control optimization. Preprint
ANL/MCS-P1948-0911, Mathematics and Computer Science Division, Argonne National Labo-
ratory, 9700 South Cass Avenue, Argonne, IL 60439, U.S.A.

Kirches, C., Sager, S., Bock, H., and Schloder, J. (2010). Time-optimal control of automobile test
drives with gear shifts. Optimal Control Applications and Methods, 31(2):137153.
Kirkpatrick, S., Gelatt, C. D., and Vecchi, M. P. (1983). Optimization by simulated annealing.
Science, 220(4598):671680.
Klepeis, J. L. and Floudas, C. A. (2003). ASTRO-FOLD: a combinatorial and global optimization

framework for ab initio prediction of three-dimensional structures of proteins from the amino
acid sequence. Biophysical Journal, 85:21192146.
Kocis, G. R. and Grossmann, I. E. (1988). Global optimization of nonconvex mixed-integer non-

linear programming (MINLP) problems in process synthesis. Industrial Engineering Chemistry
Research, 27:14071421.
Krokhmal, P. A. and Soberanis, P. (2010). Risk optimization with p-order conic constraints: A
linear programming approach. European Journal of Operational Research, 201(3):653671.
Lakhera, S., Shanbhag, U. V., and McInerney, M. (2011). Approximating electrical distribution
networks via mixed-integer nonlinear programming. International Journal of Electric Power and
Energy Systems, 33(2):245257.
Land, A. H. and Doig, A. G. (1960). An automatic method for solving discrete programming
problems. Econometrica, 28:497520.
Lasserre, J. (2000). Convergent LMI relaxations for nonconvex quadratic programs. In Proceedings
of the 39th IEEE Conference on Decision and Control (Cat. No.00CH37187), volume vol.5, pages 5041
6, Piscataway, NJ, USA.
Lasserre, J. (2001). An explicit exact SDP relaxation for nonlinear 0-1 programs. In Aardal, K. and
Gerards, A., editors, Integer Programming and Combinatorial Optimization 2001, Lecture Notes in
Computer Science, Vol. 2081, pages 293 303, Berlin, Germany.
Lawler, E. L. and Woods, D. E. (1966). Branch-and-bound methods: A survey. Operations Research,

14(4):699719.
Lee, J. and Leyffer, S., editors (2011). Mixed Integer Nonlinear Programming, IMA Volume in Math-
ematics and its Applications. Springer, New York.
Lee, J. and Wilson, D. (2001). Polyhedral methods for piecewise-linear functions I: The lambda
method. Discrete Applied Mathematics, 108(3):269 285.

Leineweber, D., Bauer, I., Schafer, A., Bock, H., and Schloder, J. (2003). An efficient multiple
shooting based reduced SQP strategy for large-scale dynamic process optimization (Parts I and
II). Computers & Chemical Engineering, 27:157174.
Leyffer, S. (1998). User manual for MINLP-BB. University of Dundee.
Leyffer, S. (2001). Integrating SQP and branch-and-bound for mixed integer nonlinear program-
ming. Computational Optimization & Applications, 18:295309.
Leyffer, S. (2003). MacMINLP: Test problems for mixed integer nonlinear programming. http:
//www.mcs.anl.gov/leyffer/macminlp.
Liberti, L., Mladenovic, N., and Nannicini, G. (2011). A recipe for finding good solutions to
MINLPs. Mathematical Programming Computations, 3:349390.
Liberti, L. and Pantelides, C. C. (2003). Convex envelopes of monomials of odd degree. Journal of
Global Optimization, 25(2):157168.
Lin, Y. and Schrage, L. (2009). The global solver in the LINDO API. Optimization methods and
software, 24(4):657668.
Linderoth, J. T. (2005). A simplicial branch-and-bound algorithm for solving quadratically con-

strained quadratic programs. Mathematical Programming, Series B, 103:251282.
Linderoth, J. T. and Savelsbergh, M. W. P. (1999). A computational study of search strategies in

mixed integer programming. INFORMS Journal on Computing, 11:173187.
Liuzzi, G., Lucidi, S., and Rinaldi, F. (2011). Derivative-free methods for bound constrained mixed-
integer optimization. Computational Optimization and Applications, pages 122.
Luedtke, J., Namazifar, M., and Linderoth, J. (2012). Some results on the strength of relaxations of
multilinear functions. Mathematical Programming, pages 127. 10.1007/s10107-012-0606-z.
Mahajan, A., Leyffer, S., and Kirches, C. (2012). Solving mixed-integer nonlinear programs by
QP-diving. Preprint ANL/MCS-2071-0312, Argonne National Laboratory, Mathematics and
Computer Science Division.
Mahajan, A., Leyffer, S., Linderoth, J., Luedtke, J., and Munson, T. (2011). MINOTAUR: a toolkit for
solving mixed-integer nonlinear optimization. wiki-page. http://wiki.mcs.anl.gov/minotaur.
Maonan, L. and Wenjun, H. (1991). The study of choosing opiimal plan of air quantities regulation
of mine ventilation network. In Proceedings of the 5th US Mine Ventilation Symposium, pages
427421.
Maria, J., Truong, T. T., Yao, J., Lee, T.-W., Nuzzo, R. G., Leyffer, S., Gray, S. K., and Rogers, J. A.
(2009). Optimization of 3D plasmonic crystal structures for refractive index sensing. Journal of
Physical Chemistry C, 113(24):1049310499.
Markowitz, H. M. and Manne, A. S. (1957). On the solution of discrete programming problems.

Econometrica, 25(1):84110.

Martin, A., Moller, M., and Moritz, S. (2006). Mixed integer models for the stationary case of gas
network optimization. Mathematical Programming, 105:563582.
Masihabadi, S., Sanjeevi, S., and Kianfar, K. (2011). n-step conic mixed integer rounding inequal-
ities. Optimization Online. http://www.optimization-online.org/DB_HTML/2011/
11/3251.html.
McCormick, G. P. (1976). Computability of global solutions to factorable nonconvex programs:

Part I Convex underestimating problems. Mathematical Programming, 10:147175.
Messine, F. (2004). Deterministic global optimization using interval constraint propagation tech-
niques. RAIRO-RO, 38(4):277294.
Meyer, R. (1976). Mixed integer minimization models for piecewise-linear functions of a single
variable. Discrete Mathematics, 16(2):16371.
Miller, R., Xie, Z., Leyffer, S., Davis, M., and Gray, S. (2010). Surrogate-based modeling of the
optical response of metallic nanostructures. Journal of Physical Chemistry C, 114(48):2074120748.
DOI:10.1021/jp1067632.
Misener, R. and Floudas, C. (2012a). Global optimization of mixed-integer quadratically-

constrained quadratic programs (MIQCQP) through piecewise-linear and edge-concave relax-
ations. Mathematical Programming, pages 128. 10.1007/s10107-012-0555-6.
Misener, R. and Floudas, C. (2012b). GloMIQO: Global mixed-integer quadratic optimizer. Journal
of Global Optimization, pages 148.
Momoh, J., Koessler, R., Bond, M., Stott, B., Sun, D., Papalexopoulos, A., and Ristanovic, P. (1997).
Challenges to optimal power flow. IEEE Transaction on Power Systems, 12:444455.

Muller, J. (2012). Surrogate Model Algorithms for Computationally Expensive Black-Box Global Opti-
mization Problems. PhD thesis, Tampere University of Technology, Tampere, Finland.

Muller, J., Shoemaker, C. A., and Piche, R. (2012). SO-MI: A surrogate model algorithm for com-
putationally expensive nonlinear mixed-integer black-box global optimization problems. Com-
puters & Operations Research. To appear.
Nannicini, G. and Belotti, P. (2012). Rounding-based heuristics for nonconvex MINLPs. Mathe-
matical Programming Computation, 4:131.
Nannicini, G., Belotti, P., and Liberti, L. (2008). A local branching heuristic for MINLPs.
arXiv:0812.2188v1 [math.CO]. http://arxiv.org/abs/0812.2188.
Nemhauser, G. and Wolsey, L. A. (1988a). Integer and Combinatorial Optimization. John Wiley and
Sons, New York.
Nemhauser, G. L., Savelsbergh, M. W. P., and Sigismondi, G. C. (1994). MINTO, a Mixed INTeger
Optimizer. Operations Research Letters, 15:4758.
Nemhauser, G. L. and Wolsey, L. A. (1988b). Integer and Combinatorial Optimization. Wiley, New
York.
Nocedal, J. and Wright, S. (1999). Numerical Optimization. Springer, New York.
Nowak, I., Alperin, H., and Vigerske, S. (2003). LaGO an object oriented library for solving
MINLPs. In Bliek, C., Jermann, C., and Neumaier, A., editors, Proceedings of the 1st Global Op-
timization and Constraint Satisfaction Workshop (COCOS 2002), number 2861 in Lecture Notes in
Computer Science, pages 3242, Berlin/Heidelberg. Springer.
Oldenburg, J., Marquardt, W., Heinz, D., and Leineweber, D. (2003). Mixed logic dynamic opti-
mization applied to batch distillation process design. AIChE Journal, 49(11):29002917.
Padberg, M. (1989). The boolean quadric polytope: some characteristics, facets and relatives.
Mathematical Programming, Series B, 45(1):139 72.
Padberg, M. (2000). Approximating separable nonlinear functions via mixed zero-one programs.
Operations Research Letters, 27(1):15.
Powell, R. (2007). Defending against terrorist attacks with limited resources. American Political
Science Review, 101(3):527541.
Prata, A., Oldenburg, J., Kroll, A., and Marquardt, W. (2008). Integrated scheduling and dynamic
optimization of grade transitions for a continuous polymerization reactor. Computers & Chemical
Engineering, 32:463476.
Pruitt, K. A., Leyffer, S., Newman, A. M., and Braun, R. (2012). Optimal design and dispatch of
distributed generation systems. Preprint ANL/MCS-2004-0112, Argonne National Laboratory,
Mathematics and Computer Science Division.
Qualizza, A., Belotti, P., and Margot, F. (2012). Linear programming relaxations of quadratically
constrained quadratic programs. In Mixed Integer Nonlinear Programming, volume 154 of IMA
Volume Series in Mathematics and its Applications, pages 407426. Springer.
Quesada, I. and Grossmann, I. E. (1992). An LP/NLP based branch-and-bound algorithm for

convex MINLP optimization problems. Computers & Chemical Engineering, 16:937947.

Quist, A. J., van Gemeert, R., Hoogenboom, J. E., Illes, T., Roos, C., and Terlaky, T. (1998). Ap-
plication of nonlinear optimization to reactor core fuel reloading. Annals of Nuclear Energy,
26:423448.
Rashid, K., Ambani, S., and Cetinkaya, E. (2012). An adaptive multiquadric radial basis function
method for expensive black-box mixed-integer nonlinear constrained optimization. Engineering
Optimization, to appear:122.
Romero, R., Monticelli, A., Garcia, A., and Haffner, S. (2002). Test systems and mathematical mod-
els for transmission network expansion planning. IEEE Proceedings Generation, Transmission
and Distrbution., 149(1):2736.
Rote, G. (1992). The convergence rate of the sandwich algorithm for approximating convex func-
tions. Computing, 48:33761.
Rubinstein, R. Y. and Kroese, D. P. (2004). The cross-entropy method: A Unified Approach to Combina-
torial Optimization, Monte-Carlo Simulation and Machine Learning. Springer, New York.
Ryoo, H. S. and Sahinidis, N. V. (1995). Global optimization of nonconvex NLPs and MINLPs with
applications in process design. Computers & Chemical Engineering, 19:552566.
Ryoo, H. S. and Sahinidis, N. V. (1996). A branch-and-reduce approach to global optimization.

Journal of Global Optimization, 8:107139.
Sager, S. (2005). Numerical methods for mixed-integer optimal control problems. Der andere Verlag,

Tonning,
Lubeck, Marburg. ISBN 3-89959-416-9.
Sager, S. (2012). A benchmark library of mixed-integer optimal control problems. In Lee, J. and
Leyffer, S., editors, Mixed Integer Nonlinear Programming, pages 631670. Springer.
Sager, S., Bock, H., and Diehl, M. (2012). The integer approximation error in mixed-integer optimal
control. Mathematical Programming A, 133(12):123.

Sager, S., Diehl, M., Singh, G., Kupper, A., and Engell, S. (2007). Determining SMB superstructures
by mixed-integer control. In Proceedings of OR2006, pages 3744, Karlsruhe. Springer.
Sager, S., Jung, M., and Kirches, C. (2011). Combinatorial integral approximation. Mathematical
Methods of Operations Research, 73(3):363380.
Sager, S., Reinelt, G., and Bock, H. (2009). Direct methods with maximal lower bound for mixed-
integer optimal control problems. Mathematical Programming, 118(1):109149.
Sahinidis, N. V. (1996). BARON: A general purpose global optimization software package. Journal
of Global Optimization, 8:201205.
Sandler, T. and Arce M, D. G. (2003). Terrorism and game theory. Simulation Gaming, 34:319337.
Sandler, T. and Siqueira, K. (2006). Global terrorism: Deterrence versus preemption. Canadian
Journal of Economics, 39(4):13701387.
Sargent, I. E. G. R. W. H. (1979). Optimal design of multipurpose batch plants. Industrial & Engi-
neering Chemistry Process Design and Development, 18:343348.
Savelsbergh, M. W. P. (1994). Preprocessing and probing techniques for mixed integer program-
ming problems. ORSA Journal on Computing, 6:445454.
Saxena, A., Bonami, P., and Lee, J. (2010). Convex relaxations of non-convex mixed inte-
ger quadratically constrained programs: extended formulations. Mathematical Programming,
124:383411.
Saxena, A., Bonami, P., and Lee, J. (2011). Convex relaxations of non-convex mixed inte-
ger quadratically constrained programs: projected formulations. Mathematical Programming,
130:359413.
Schichl, H. (2004). Global optimization in the COCONUT project. Numerical Software with Result
Verification, pages 277293.
Schrijver, A. (1986). Theory of Linear and Integer Programming. Wiley, New York.
Schweiger, C. A. (1999). Process Synthesis, Design, and Control: Optimization with Dynamic Models
and Discrete Decisions. PhD thesis, Princeton University, Princeton, NJ.
Shaik, O., Sager, S., Slaby, O., and Lebiedz, D. (2008). Phase tracking and restoration of circadian
rhythms by model-based optimal control. IET Systems Biology, 2:1623.
Sheikh, W. and Ghafoor, A. (2010). An optimal bandwidth allocation and data droppage scheme
for differentiated services in a wireless network. Wireless Communications and Mobile Computing,
10(6):733747.
Sherali, H. and Adams, W. (1998). A Reformulation-Linearization Technique for Solving Discrete and
Continuous Nonconvex Problems. Kluwer, Dordrecht.
Sherali, H. and Alameddine, A. (1992). A new reformulation-linearization technique for bilinear

programming problems. Journal of Global Optimization, 2(4):379410.
Sherali, H. and Smith, E. (1997). A global optimization approach to a water distribution network
design problem. Journal of Global Optimization, 11:107132.
Sherali, H. D. (2001). On mixed-integer zero-one representations for separable lower-

semicontinuous piecewise-linear functions. Operations Research Letters, 28:155 160.
Sherali, H. D. and Fraticelli, B. M. P. (2002). Enhancing RLT relaxations via a new class of semidef-
inite cuts. Journal of Global Optimization, 22:233261.
Sherali, H. D., Subramanian, S., and Loganathan, G. V. (2001). Effective relaxations and parti-
tioning schemes for solving water distribution network design problems to global optimality.
Journal of Global Optimization, 19:126.
Sinha, R., Yener, A., and Yates, R. D. (2002). Noncoherent multiuser communications: Multistage
detection and selective filtering. EURASIP Journal on Applied Signal Processing, 12:14151426.
Skrifvars, H., Leyffer, S., and Westerlund, T. (1998). Comparison of certain minlp algorithms when
applied to a model structure determination and parameter estimation problem. Computers &
Chemical Engineering, 22(12):1829 1835.
Smith, E. M. B. and Pantelides, C. C. (1997). Global optimization of nonconvex MINLPs. Computers

& Chemical Engineering, 21:S791S796.
Soleimanipour, M., Zhuang, W., and Freeman, G. H. (2002). Optimal resource management
in wireless multimedia wideband CDMA systems. IEEE Transactions on Mobile Computing,
1(2):143160.
Soler, M., Olivares, A., Staffetti, E., and Bonami, P. (2011). En-route optimal flight planning con-
strained to pass through waypoints using MINLP. In Proceedings of 9th USA/Europe Air Traffic
Management Research and Development Seminar, Berlin.
Still, C. and Westerlund, T. (2006). Solving convex MINLP optimization problems using a sequen-
tial cutting plane algorithm. Comput. Optim. Appl., 34(1):6383.
Stubbs, R. and Mehrotra, S. (1999). A branch-and-cut method for 0-1 mixed convex programming.
Mathematical Programming, 86:515532.
Stubbs, R. and Mehrotra, S. (2002). Generating convex polynomial inequalities for mixed 0-1
programs. Journal of Global Optimization, 24:311332.
Tawarmalani, M., Richard, J.-P., and Chung, K. (2010). Strong valid inequalities for orthogonal
disjunctions and bilinear covering sets. Mathematical Programming, 124:481512. 10.1007/s10107-
010-0374-6.
Tawarmalani, M. and Sahinidis, N. V. (2002). Convexification and Global Optimization in Continuous

and Mixed-Integer Nonlinear Programming: Theory, Algorithms, Software, and Applications. Kluwer
Academic Publishers, Boston MA.
Tawarmalani, M. and Sahinidis, N. V. (2004). Global optimization of mixed integer nonlinear

programs: A theoretical and computational study. Mathematical Programming, 99:563591.
Tawarmalani, M. and Sahinidis, N. V. (2005). A polyhedral branch-and-cut approach to global

optimization. Mathematical Programming, 103(2):225249.
Terwen, S., Back, M., and Krebs, V. (2004). Predictive powertrain control for heavy duty trucks. In
Proceedings of IFAC Symposium in Advances in Automotive Control, pages 451457, Salerno, Italy.
Tomlin, J. (1981). A suggested extension of special ordered sets to non-separable non-convex

programming problems. Annals of Discrete Mathematics, 11:359370.
Toriello, A. and Vielma, J. P. (2012). Fitting piecewise linear continuous functions. European Journal
of Operational Research, 219:8695.

Turkay, M. and Grossmann, I. E. (1996). Logic-based MINLP algorithms for the optimal synthesis
of process networks. Computers & Chemical Engineering, 20(8):959978.
Van Roy, T. J. (1983). Cross decomposition for mixed integer programming. Mathematical Program-
ming, 25:145163.
Vandenbussche, D. and Nemhauser, G. L. (2005a). A branch-and-cut algorithm for nonconvex

quadratic programs with box constraints. Mathematical Programming, 102:559575.
Vandenbussche, D. and Nemhauser, G. L. (2005b). A polyhedral study of nonconvex quadratic

programs with box constraints. Mathematical Programming, 102:531557.
Vielma, J. P., Ahmed, S., and Nemhauser, G. (2010). Mixed-integer models for nonsepara-
ble piecewise-linear optimization: Unifying framework and extensions. Operations Research,
58(2):303315.
Vielma, J. P., Ahmed, S., and Nemhauser, G. L. (2008). A lifted linear programming branch-and-
bound algorithm for mixed integer conic quadratic programs. INFORMS Journal on Computing,
20(3):438450.
Vielma, J. P. and Nemhauser, G. (2011). Modeling disjunctive constraints with a logarithmic num-
ber of binary variables and constraints. Mathematical Programming, 128(1):4972.
Viswanathan, J. and Grossmann, I. E. (1990). A combined penalty function and outer-

approximation method for MINLP optimization. Computers & Chemical Engineering, 14(7):769
782.
Wachter, A. and Biegler, L. T. (2006). On the implementation of a primal-dual interior point fil-
ter line search algorithm for large-scale nonlinear programming. Mathematical Programming,
106(1):2557.
Westerlund, T. and Lundqvist, K. (2001). Alpha-ECP, Version 5.01: An interactive MINLP-solver

based on the Extended Cutting Plane Method. Technical Report Report 01-178-A, Process De-

sign Laboratory at Abo University.
Westerlund, T. and Lundqvist, K. (2005). Alpha-ECP, version 5.101: An interactive MINLP-solver

based on the Extended Cutting Plane Method. Technical Report Report 01-178-A, Process De-

sign Laboratory at Abo University.
Westerlund, T. and Pettersson, F. (1995). A cutting plane method for solving convex MINLP prob-
lems. Computers & Chemical Engineering, 19:s131s136.

Westerlund, T. and Porn, R. (2002). Solving pseudo-convex mixed integer optimization problems
by cutting plane techniques. Optimization and Engineering, 3(3):253280.
Williams, H. P. (1999). Model Building in Mathematical Programming. John Wiley & Sons.
Wilson, D. L. (1998). Polyhedral Methods for Piecewise-Linear Functions. PhD thesis, University of
Kentucky.
Wolf, D. D. and Smeers, Y. (2000). The gas transmission problem solved by an extension of the
simplex algorithm. Management Science, 46(11):14541465.
Wolsey, L. A. (1998). Integer Programming. John Wiley and Sons, New York.
Wu, X., Topuz, E., and Karfakis, M. (1991). Optimization of ventilation control device locations
and sizes in underground mine ventilation systems. In Proceedings of the 5th US Mine Ventilation
Symposium, pages 391399.
Yajima, Y. and Fujie, T. (1998). Polyhedral approach for nonconvex quadratic programming prob-
lems with box constraints. Journal of Global Optimization, 13(2):151 170.
You, F. and Leyffer, S. (2010). Oil spill response planning with MINLP. SIAG/OPT Views-and-News,
21(2):18.
You, F. and Leyffer, S. (2011). Mixed-integer dynamic optimization for oil-spill response planning
with integration of a dynamic oil weathering model. AIChe Journal. Published online: DOI:
10.1002/aic.12536.
Zhu, Y. and Kuno, T. (2006). A disjunctive cutting-plane-based branch-and-cut algorithm for 0-1
mixed-integer convex nonlinear programs. Industrial and Engineering Chemistry Research, 45:187
196.
Zhuang, J. (2008). Modeling Secrecy and Deception in Homeland Security Resource Allocation. Ph.D.
thesis, University of Wisconsin-Madison.
Zhuang, J. and Bier, V. M. (2007a). Balancing terrorism and natural disasters defensive strategy
with endogenous attacker effort. Operations Research, 55:976991.
Zhuang, J. and Bier, V. M. (2007b). Investment in security. Industrial Engineer, 39:5354.
Zhuang, J. and Bier, V. M. (2007c). Modeling secrecy and deception in homeland security resource
allocation. Submitted.
The submitted manuscript has been created by the UChicago Argonne, LLC, Operator of Argonne National Laboratory (Argonne)
under Contract No. DE-AC02-06CH11357 with the U.S. Department of Energy. The U.S. Government retains for itself, and others
acting on its behalf, a paid-up, nonexclusive, irrevocable worldwide license in said article to reproduce, prepare derivative works,
distribute copies to the public, and perform publicly and display publicly, by or on behalf of the Government.

MINLP

Uploaded by

Copyright:

Available Formats

MINLP

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

MINLP

Uploaded by

Copyright:

Available Formats

ARGONNE NATIONAL LABORATORY

9700 South Cass Avenue

Mixed-Integer Nonlinear Optimization

Mathematics and Computer Science Division

November 22, 2012

1 Mixed-Integer Nonlinear Programming Introduction 2

2 Nonlinear Models with Integer Variables 10

3 Deterministic Methods for Convex MINLP 19

4 Cutting Planes for Convex MINLPs 41

6 Heuristics for Solving MINLPs 78

7 Mixed-Integer Optimal Control Problems 84

8 Software for MINLP 91

November 22, 2012

1 Mixed-Integer Nonlinear Programming Introduction

where f : Rn R and c : Rn Rm are twice continuously differentiable functions, X Rn is a

1.1 MINLP Notation and Basic Definitions

Throughout this paper, we use the notion of a convex hull of a set S.

If X = {x Zp : l x u} and l Zp , u Zp , then conv(X) = [l, u] is simply the hyper-

1.2 Preview of Key Building Blocks of MINLP Algorithms

Relaxations. Formally, an optimization problem min{R (x) : x SR } is a relaxation of a problem

Several strategies are used to obtain relaxations of MINLPs.

1. Relaxing integrality. Integrality constraints xi Z can be relaxed to xi R for all i I. This

3. Relaxing nonconvex constraints. Constraints c(x) 0 and f (x) z containing nonconvex

f(x) f (x) and c(x) c(x), x conv(X). (1.6)

1.3 Scope and Outline

Figure 6: Constraint enforcement by using spatial branching for global optimization.

2 Nonlinear Models with Integer Variables

2.1 Modeling Practices for MINLP

where the second quadratic is now convex.

Linearization of constraints. A simple transformation is to rewrite x1 /x2 = a as x1 = ax2 , where

0 x12 x2 u and u(1 x2 ) x1 x12 u(1 x2 ),

c(x2 ) = ln(x2 ) 0, x2 = sin(x1 ), and x2 0.

2.2 Design of Multiproduct Batch Plants

multiproduct batch plant as a nonconvex MINLP:

vj = ln(Vj ), nj = ln(Nj ), bi = ln(Bi ), ci = ln Ci .

2.3 Design of Water Distribution Networks

Vmax aij qij Vmax aij , (i, j) A.

An alternative formulation based on complementarity constraints avoids the introduction of the

2.4 A Dynamic Subway Operation Problem

x s (t) = xv (t) t [0, T ], (2.4)

1. Serial, v(t) = 1. The energy consumption rate is given by

and the subway trains dynamics are described by

2. Parallel, v(t) = 2. The energy consumption rate is given by

and the subway trains dynamics are described by

Weff fv (t) = g R(xv (t)) C Weff (2.10)

The forces occurring in these dynamics are given by

x(0) = (0, 0)T , x(T ) = (S, 0)T . (2.15)

A maximum on the allowable driving time to complete the distance S is imposed,

2.5 Summary of MINLP Applications

3 Deterministic Methods for Convex MINLP

A2 The functions f and c are twice continuously differentiable convex functions.

3.1 Nonlinear Branch-and-Bound

The complete nonlinear branch-and-bound algorithm is described in Algorithm 1, and Proposi-

Branch-and-bound for MINLP

Algorithm 1: Branch-and-bound for MINLP.

Our description of branch-and-bound leaves open a number of important questions. In par-

3.1.1 Selection of branching variable

Choosing a good branching variable is a crucial component of branch-and-bound. Ideally, we

Branching on general disjunctions is an alternative to branching on a variable. We can branch