Mixed Integer Linear Programming Formulation Techniques
Mixed Integer Linear Programming Formulation Techniques
Abstract. A wide range of problems can be modeled as Mixed Integer Linear Programming (MIP)
problems using standard formulation techniques. However, in some cases the resulting MIP
can be either too weak or too large to be effectively solved by state of the art solvers. In
this survey we review advanced MIP formulation techniques that result in stronger and/or
smaller formulations for a wide class of problems.
DOI. 10.1137/130915303
∗ Received by the editors April 2, 2013; accepted for publication (in revised form) July 21, 2014;
mit.edu).
3
4 JUAN PABLO VIELMA
Now, as illustrated in Figure 2.1 for our running example, the graph of a univariate
continuous piecewise linear function (with bounded domain) is the finite union of line
segments, which we can easily model with MIP. For instance, for the function defined
in (2.3) we can construct the textbook MIP formulation of gr(f ) given by
Formulation (2.5) illustrates how MIP formulations can use both continuous and inte-
ger variables to enforce requirements on the original variables. Formulations that use
auxiliary variables different from the original variables are usually denoted extended,
lifted, or higher-dimensional. In the case of formulation (2.5) these auxiliary variables
6 JUAN PABLO VIELMA
z = f (x)
3
0
0 1 2 3
Fig. 2.1 Graph of the continuous piecewise linear function defined in (2.3).
are necessary to construct the formulation. However, as we will see in section 4, aux-
iliary variables can provide an advantage even when they are not strictly necessary.
For this reason we use the following definition of an MIP formulation that always
considers the possible use of auxiliary variables.
Definition 2.1. For A ∈ Qm×n , B ∈ Qm×s , D ∈ Qm×t , and b ∈ Qm consider
the set of linear inequalities and continuous and integer variables of the form
(2.6a) Ax + Bλ + Dy ≤ b,
(2.6b) x ∈ Qn ,
(2.6c) λ ∈ Qs ,
(2.6d) y ∈ Zt .
We say (2.6) is an MIP formulation for a set S ⊆ Qn if the projection of (2.6) onto
the x variables is exactly S. That is, if we have x ∈ S if and only if there exist λ ∈ Qs
and y ∈ Zt such that (x, λ, y) satisfies (2.6).
Note that generic formulation (2.6) does not explicitly consider the inclusion of
integrality requirements on the original variables x. While example (2.5) does not
require such integrality requirements, it is common for at least some of the original
variables to be integers. Formulation (2.6) implicitly considers that possibility by
allowing the inclusion of constraints of the form xi = yj . Hence, we can replace (2.6b)
with x ∈ Qn1 × Zn2 without really changing the definition of an MIP formulation.
Using Definition 2.1 we can now write a generic MIP reformulation of (2.1) by
replacing every occurrence of (xi , z) ∈ gr(fi ) with an MIP formulation of gr(fi ) to
obtain
n
(2.7a) zMIP := min zi
i=1
s.t.
(2.7b) Ex ≤ h,
MIXED INTEGER LINEAR PROGRAMMING FORMULATION TECHNIQUES 7
(2.7c) 0 ≤ xi ≤ u ∀i ∈ {1, . . . , n} ,
xi
(2.7d) Ai + B i λi + Di y i ≤ bi ∀i ∈ {1, . . . , n} ,
zi
(2.7e) y i ∈ Zk ∀i ∈ {1, . . . , n} ,
where Ai , B i , Di , and bi are appropriately constructed matrices and vectors, and k
is an appropriate number of integer variables. There are many choices for (2.7d), but
as long as they are valid formulations of gr(fi ), we have zMP = zMIP and we can
extract a solution of (2.1) by looking at the x variables of an optimal solution of (2.7).
However, some versions of (2.7d) can perform significantly better when solved with
an MIP solver. We study this potential difference in the next subsection.
2.2. Strength, Size, and MIP Solvers. While all state-of-the-art MIP solvers are
based on the branch-and-bound algorithm [102], they also include a large number of
advanced techniques that make it hard to predict the specific impact of an alternative
formulation. However, there are two aspects of an MIP formulation that usually have
a strong impact on both simple branch-and-bound algorithms and state-of-the-art
solvers: the size and strength of the LP relaxation, and the effect of branching on the
formulation. Instead of giving a lengthy description of the branch-and-bound algo-
rithm and state-of-the-art solvers, we introduce the necessary concepts by considering
the first step of the solution of MIP reformulation (2.7) with a simple branch-and-
bound algorithm. For more details on basic branch-and-bound and state-of-the-art
solvers, we refer the reader to [22, 36, 131, 166] and [1, 2, 25, 23, 92, 112, 24], respec-
tively.
The first step in solving MIP formulation (2.7) with a branch-and-bound al-
gorithm is to solve the LP relaxation of (2.7) obtained by dropping all integrality
requirements. The resulting LP problem given by (2.7a)–(2.7d) is known as the root
LP relaxation and can be solved efficiently both in theory and in practice. Its optimal
n
value zLP := min { i=1 zi : (2.7b)–(2.7d)} provides a lower bound on zMIP known
as the LP relaxation bound. If the optimal solution to the LP relaxation satisfies the
dropped integrality constraints, then zMIP = zLP and this solution is also optimal
for (2.7). In contrast, if optimal solution x̄, z̄, λ̄, ȳ is such that for some i and j we
have ȳji ∈
/ Z, we can eliminate this infeasibility by branching on yji . To achieve this we
create two new LP problems by adding yji ≤ ȳji and yji ≥ ȳji to the LP relaxation,
respectively. These two new problems are usually denoted branches and are processed
in a similar way, which generates a binary tree known as the branch-and-bound tree.
The behavior of this first step is usually a good predictor of the performance of the
whole algorithm, so we now concentrate on the effect of an MIP formulation on this
step. We consider the effect of different formulations on subsequent steps in section 8.
To understand the effect of a formulation on the root LP relaxation we need
to understand what the LP relaxation of the formulation is modeling. Going back
to our running example, consider the LP relaxation of (2.5) given by (2.5a)–(2.5j).
Constraints (2.5a)–(2.5c) and (2.5i) show that any (x, z) that is part of a feasible
solution for this LP relaxation must be a convex combination of points (0, 2), (1, 3),
(2, 1), and (3, 1). Conversely, any (x, z) that is a convex combination of these points
is part of a feasible solution for the LP relaxation (let λ be the appropriate convex
combination multipliers, y1 = λ0 , y2 = 1−λ0 −λ3 , and y3 = λ3 ). Hence, the projection
onto the (x, z) variables of the LP Relaxation of (2.5) is equal to conv (gr (f )), the
convex hull of gr (f ). As illustrated in Figure 2.2(a), this convex hull corresponds to
the smallest convex set containing gr (f ). Now, if we used this formulation for fi in
8 JUAN PABLO VIELMA
(2.1) (for simplicity imagine that all these functions are equal to f ), the LP relaxation
of (2.7) would be equivalent to
n
(2.8a) zLP := min φi (xi )
i=1
s.t.
(2.8b) Ex ≤ h,
(2.8c) 0 ≤ xi ≤ u ∀i ∈ {1, . . . , n} ,
where φi : [0, u] → Q is the convex envelope or lower convex envelope [80] of fi . This
lower convex envelope is the tightest convex underestimator of f and is illustrated for
our running example in Figure 2.2(b).
z = f (x) z = f (x)
3 3
2 2
1 1
0 0
0 1 2 3 0 1 2 3
(a) gr (f ) in red and conv (gr (f )) in light blue. (b) conv (gr (f )) in light blue and lower convex
envelope of f in dark blue.
This reasoning suggests that, at least with respect to the LP relaxation bound,
formulation (2.5) of our example function is as strong as it can be. Indeed, the
projection onto the (x, z)-space of any formulation of f is a convex (and polyhedral)
set that must contain gr (f ) and the LP relaxation of (2.5) projects to the smallest
of such sets. By the same argument, the projection onto the x variables of the LP
relaxation of an MIP formulation of any set S ⊆ Qn must also contain conv(S)
and formulations whose LP relaxations project precisely to conv(S) yield the best
LP relaxation bounds. Jeroslow and Lowe [90, 113] denoted those formulations that
achieve this best possible LP relaxation bound as sharp. Other authors also denote
them convex hull formulations.
Definition 2.2. An MIP formulation of set S ⊆ Qn is sharp if and only if the
projection onto the x variables of its LP relaxation is exactly conv(S).
It is important to note that sharp formulations yield the best LP relaxation
bound among all MIP formulations for the sets we selected to model. However, the
LP relaxation bound can vary if we elect to model other sets. For instance, in our
MIXED INTEGER LINEAR PROGRAMMING FORMULATION TECHNIQUES 9
example, we elected to independently model gr(fi ) for all i ∈ {1, . . . , n}. However, we
could instead have elected to model
⎧ ⎫
⎪ n
⎪
⎪
⎪ ⎪
⎪
⎪
⎨ z= fi (xi ), ⎪
⎬
(2.9) S = (x, z) ∈ Q × Q :
n i=1
⎪
⎪ Ex ≤ h, ⎪
⎪
⎪
⎪ ⎪
⎪
⎩ ⎭
0 ≤ xi ≤ u ∀i ∈ {1, . . . , n}
to obtain a formulation that considers all possible nonconvexities at the same time.
A sharp formulation for this case would be significantly stronger as we can show that
for S defined in (2.9) we have min {z : (x, z) ∈ conv (S)} = zMP = zMIP . Because
the LP relaxation of a sharp formulation of S is equal to conv(S), we have that calcu-
lating the optimal value of (2.4) could be done by simply solving an LP problem. Of
course, unlike our original piecewise sharp formulation, constructing a sharp formula-
tion for this joint S would normally be harder than solving (2.4). Furthermore, this
approach can result in a significantly larger formulation. Selecting which portions of
a mathematical programming problem to model independently to balance size and
final strength is a crucial and nontrivial endeavor. However, the appropriate selection
is usually ad hoc to the specific structure of the problem considered. For instance,
Croxton, Gendron, and Magnanti [42] study the case where (2.1) corresponds to a
multicommodity network flow problem with piecewise linear costs. In this setting,
they show that a convenient middle ground between constructing independent formu-
lations of each gr (fi ) and a single formulation for the complete problem (2.9) is to
construct independent formulations for each gr i∈Ia fi (xi ) , where Ia corresponds
to the flow variables of all commodities in a given arc a of the network. From now
on we assume that a similar analysis has already been carried out and we focus on
constructing small and strong formulations for the individual portions selected. How-
ever, in section 10 we will succinctly consider the selection of such portions and the
combination of the associated formulations.
As we have mentioned, sharp formulations are strongest in one sense. However,
if we consider the integer variables of our MIP formulation we can construct even
stronger formulations. To illustrate this, let us go back to our example and consider
an optimal solution x̄, z̄, λ̄, ȳ of the LP relaxation of (2.7) given by (2.7a)–(2.7d).
Because the root LP relaxation is usually solved by a simplex algorithm, we can ex-
pect this to be an optimal basic feasible solution. However, analyzing basic feasible
solutions of (2.7b)–(2.7d) can be quite hard, so let us analyze basic feasible solutions
of the LP relaxations of the individual formulations of gr(fi ) (i.e., (2.5a)–(2.5j)) as a
reasonable proxy. We can check that one optimal basic feasible solution of minimizing
z over (2.5a)–(2.5j) is given by λ̄2 = λ̄3 = 1/2, λ̄0 = λ̄1 = 0, ȳ1 = ȳ3 = 1/2, ȳ2 = 0,
x̄ = 2.5, and z̄ = 1. Because (2.5) is a sharp MIP formulation, it is not surprising
that this gives the same optimal value as minimizing z over (2.5). Indeed we have
that 1 = f (2.5) = φ(2.5) and x = 2.5 is a minimizer of f . However, the basic feasible
solution obtained has some of its integer variables set at fractional values. Because
a general purpose branch-and-bound solver is not aware of the specific structure of
the problem, it would have no choice but to unnecessarily branch on one of these
variables (let’s say ȳ1 ) or to run a rounding heuristic to obtain an integer feasible
solution. Hence, while sharp formulations are the strongest possible with regard to
LP relaxation bounds, they can be somewhat weak with respect to finding optimal
or good quality integer feasible solutions. Fortunately, it is possible to construct MIP
formulations that are strong from both the LP relaxation bound and integer feasibility
10 JUAN PABLO VIELMA
perspectives. These formulations are those whose LP relaxations have basic feasible
solutions that automatically satisfy the integrality requirements on the y variables.
Such LP relaxations are usually denoted integral and formulations with this property
were denoted locally ideal by Padberg and Rijal [136, 137]. Ideal comes from the fact
that this is the strongest property we can expect from an MIP formulation from any
perspective. Locally serves to clarify that this property refers to a formulation for a
specifically selected portion and not to the whole mathematical programming problem
(following this convention, we should then refer to sharp formulations as locally sharp,
but we avoid it for simplicity and historical reasons). For simplicity, here we restrict
our attention to MIP formulations with LP relaxations that have at least one basic fea-
sible solution, which is the case for all practical formulations considered in this survey.
Definition 2.3. An MIP formulation (2.6) is locally ideal if and only if its LP
relaxation has at least one basic feasible solution and all such basic feasible solutions
have integral y variables.
As expected, the following simple proposition states that a locally ideal formu-
lation is at least as strong as a sharp formulation. We postpone the proof of this
proposition until section 3 where we introduce some useful results concerning the
feasible regions of LP problems.
Propositions 2.4. A locally ideal MIP formulation is sharp.
Finally, formulation (2.5) shows that being locally ideal can be strictly stronger
than being sharp. However, one case in which the converse of Proposition 2.4 holds is
that of traditional MIP formulations without auxiliary variables (i.e., (2.6) with s = 0
and such that for all i ∈ {1, . . . , t} constraints (2.6a) include an equality of the form
xj = yi for some j ∈ {1, . . . , n}). We again postpone the proof of this proposition
until section 3.
Propositions 2.5. For A ∈ Qm×n , b ∈ Qm , and n1 , n2 ∈ Z+ such that n =
n1 + n2 , let
(2.10a) Ax ≤ b,
(2.10b) x ∈ Qn1 × Zn2 ,
be an MIP formulation of S ⊆ Qn (i.e., S is precisely the feasible region of (2.10)).
If the LP relaxation of (2.10) has at least one basic feasible solution, then (2.10) is
locally ideal if and only if it is sharp.
k
3. Polyhedra. Most sets modeled with MIP are of the form S = i=1 P i , where
P i are rational polyhedra (e.g., the graph of a univariate piecewise linear function is
the union of line segments). Sets of this form usually appear as the feasible regions of
certain disjunctive programming problems [10, 11, 13]. For this reason we refer to such
sets as disjunctive constraints or disjunctive sets. While in the theory of disjunctive
programming these terms are also used to describe a slightly broader class of sets, in
this survey we concentrate on disjunctive sets that are the union of certain unbounded
rational polyhedra. To construct and evaluate MIP formulations for these and other
sets it will be convenient to use definitions and results from polyhedral theory that
we now review. We begin by considering some basic definitions and a fundamental
result that relates two natural definitions of polyhedra. We then consider the relation
between polyhedra and the feasible regions of MIP problems. After that we study
the linear transformations of polyhedra, which will be useful when analyzing the
strength of MIP formulations. Finally, we consider the smallest possible descriptions
of polyhedra to consider the real sizes of MIP formulations. We refer the reader to
[22, 131, 145, 166, 170] for omitted proofs and a more detailed treatment of polyhedra.
MIXED INTEGER LINEAR PROGRAMMING FORMULATION TECHNIQUES 11
k
(3.3a) D= Qi ,
i=1
⎧
⎪
⎪ m1 x + c1 , x ∈ Q1 ,
⎨
..
(3.3b) f (x) = .
⎪
⎪
⎩ mk x + c , x ∈ Qk .
k
k
(3.4a) gr(f ) = (x, z) ∈ Qn × Q : Ai x ≤ bi , mi x + ci = z ,
i=1
k
v
(3.4b) gr(f ) = conv .
f (v) v∈ext(Qi )
i=1
k
(3.5a) epi(f ) = (x, z) ∈ Qn × Q : Ai x ≤ bi , mi x + ci ≤ z ,
i=1
k
v 0n
(3.5b) epi(f ) = conv + cone ,
f (v) v∈ext(Qi ) 1
i=1
where in both cases the recession cone of all polyhedra considered is equal to
0n
cone = {(x, z) ∈ Qn × Q : x = 0, z ≥ 0} .
1
3.2. Fundamental Theorem of Integer Programming and Formulation
Strength. The finite V-representation of any polyhedron guaranteed by Theorem 3.6
yields a convenient way to prove Proposition 2.4 as follows.
Propositions 2.4. A locally ideal MIP formulation is sharp.
Proof. Let (2.6) be a locally ideal MIP formulation of a set S and let Q ⊆
Qn × Qs × Qt be the polyhedron described by (2.6a). We need to show that the
projection of Q onto the x variables is contained in conv(S).
By Lemma 3.5 and locally idealness of (2.6) we have that ext (Q) ⊆ Qn × Qs × Zt .
Then, by Theorem 3.6 and through an appropriate scaling of the extreme rays of Q,
p d
we have that there exist x̂j , ûj , ŷ j j=1 ⊆ Qn × Qs × Zt and x̃l , ũl , ỹ l l=1 ⊆
p d
Zn × Zs × Zt such that ext (Q) = x̂j , ûj , ŷ j j=1 , ray (Q) = x̃l , ũl , ỹ l l=1 , and
p d
Q = conv x̂j , ûj , ŷ j j=1
+ cone x̃l , ũl , ỹ l l=1
.
14 JUAN PABLO VIELMA
p
Then for any (x, u, y) ∈ Q there exist λ ∈ Δp := λ ∈ Qp+ : i=1 λi = 1 and μ ∈ Qd+
such that
p
d
(x, u, y) = λj x̂j , ûj , ŷ j + μl x̃l , ũl , ỹ l .
j=1 l=1
p
Let (x1 , u1 , y 1 ) := j=1 λj x̂j , ûj , ŷ j . Because points x̂j , ûj , ŷ j j=1 satisfy (2.6),
p
we have x̂j ∈ S for all j and hence x1 ∈ conv(S). Now, without loss of generality,
assume λ1 > 0 and let
p
d
(x2 , u2 , y 2 ) : = λj x̂j , ûj , ŷ j + α μl x̃l , ũl , ỹ l
j=1 l=1
p
= λ1 (x̄, ū, ȳ) + λj x̂j , ûj , ŷ j ,
j=2
d
where α≥ 1 is such that (α/λ1 ) l=1 μl ỹ l ∈ Zt and
d
(x̄, ū, ȳ) := x̂1 , û1 , ŷ 1 + (α/λ1 ) μl x̃l , ũl , ỹ l .
l=1
Then ȳ = ŷ 1 + (α/λ1 ) l=1 μl ỹ l ∈ Zt and (x̄, ū, ȳ) satisfies (2.6) by Theorem 3.6.
d
Theorem 3.9 shows that conv (P ∩ (Qn1 × Zn2 )) is the LP relaxation of a sharp
(by definition) formulation of P ∩ (Qn1 × Zn2 ). Furthermore, through Proposition 2.5
it shows that this formulation is additionally locally ideal.
Propositions 2.5. For A ∈ Qm×n , b ∈ Qm , and n1 , n2 ∈ Z+ such that n =
n1 + n2 , let
(2.10a) Ax ≤ b,
(2.10b) x ∈ Qn1 × Zn2 ,
x2 = (λ1 + λ2 ) x̂2 + j=3 λj x̂j are such that x1 , x2 ∈ Q and x1 = x2 . This again
p
with the extremality of y. Hence, y = L x̂j for some j ∈ {1, . . . , p}, which concludes
the proof.
Since the projection onto a set of variables is a linear transformation, Proposi-
tion 3.10 shows that the extreme points of the projection of a polyhedron are contained
in the projection of the extreme points of the same polyhedron. Hence, because pro-
jection preserves integrality, the projection of locally ideal formulations is also locally
ideal. However, in section 5 we will show that projecting a polyhedron (or formula-
tion) can result in a significant increment in the number of inequalities. To achieve
this we will need the following proposition that gives a more detailed description of
an H-representation of the projection of a polyhedron.
Proposition 3.11. Let A ∈ Qm×n , D ∈ Qm×p , b ∈ Qm ,
P = {x ∈ Qn : ∃w ∈ Qp s.t. Ax + Dw ≤ b} ,
and C = μ ∈ Qm : DT μ = 0, μ ≥ 0 . Then
P = x ∈ Qn : μT Ax ≤ μT b ∀μ ∈ ray(C) .
In particular we have that P∞ = {x ∈ Qn : ∃w ∈ Qp s.t. Ax + Dw ≤ 0}.
16 JUAN PABLO VIELMA
4. Size and Extended MIP Formulations. Strength and small size can some-
times be incompatible in MIP formulations. Fortunately, this can often be conciliated
by utilizing the power of auxiliary variables in extended formulations. We first il-
lustrate this by showing an example where the incompatibility between strength and
small size can be resolved by using the same binary variables that are required to
construct even the simplest formulation.
Example 2. Consider the set S := ni=1 P i where, for each i ∈ {1, . . . , n},
we have P i := {x ∈ Qn : |xi | ≤ 1, xj = 0 ∀j = i}. It is easy to check that an MIP
formulation of S is given by
Then, the formulation obtained by adding (4.2) to (4.1) is automatically sharp. How-
ever, formulation (4.1)–(4.2) has two problems. First, it is not locally ideal because,
for n = 3, we have that x1 = x2 = y2 = y3 = 1/2, x3 = y1 = 0 is an extreme point of
the LP relaxation of (4.1)–(4.2). Second, the formulation is extremely large, because
the number of linear inequalities described by (4.2) is 2n . Unfortunately, each one of
these inequalities defines a different facet of conv(S) and together they form a min-
imal representation of conv(S). Thus removing any of them destroys the sharpness
property. Fortunately, careful use of auxiliary variables y allows constructing a much
smaller and locally ideal formulation for S.
Example 3. A polynomial-sized sharp formulation for set S in Example 2 is
given by
and 8). Formulation (4.3)’s only auxiliary variables are the binary variables that are
used to indicate which P i contains x. The following example shows that these binary
variables might not be enough to construct a polynomial-sized sharp formulation.
Example 4. For i ∈ {1, . . . , n}, let v i , wi ∈ Qn be defined by
) )
n, j = i, −1, j = i,
vji = wji =
−1, j = i, 0, j = i,
for all j ∈ {1, . . . , n}. In addition, let v 0 , w0 ∈ Qn be defined by vj0 = −wj0 = −1 for
all j ∈ {1, . . . , n}. n
Let S = (V × {0}) ∪ (W × {1}) ⊆ Qn+1 , where V = conv v i i=0 and W =
i n
conv w i=0 . S and conv(S) are depicted for n = 2 in Figure 4.2, where S is
shown in red and conv(S) is shown in blue. By noting that
⎧ ⎫
⎨ n ⎬
(4.4a) V × {0} = x ∈ Qn+1 : xn+1 = 0, xj ≤ 1, −xj ≤ 1 ∀j ∈ {1, . . . , n}
⎩ ⎭
j=1
and
⎧ ⎫
⎪
⎪
n
⎪
⎪
⎪
⎪ xn+1 = 1, − xj ≤ 1, ⎪
⎪
⎪
⎪
⎨ j=1
⎬
(4.4b) W × {1} = x ∈ Qn+1 :
⎪
⎪
n ⎪
⎪
⎪
⎪ (n + 1)xi − xj ≤ 1 ∀i ∈ {1, . . . , n}⎪
⎪
⎪
⎩ ⎪
⎭
j=1
(4.5c) xn+1 = 1 − y1 ,
n
(4.5d) (n + 1)xi − xj ≤ 1 + (n2 + n − 2)(1 − y2 ) ∀i ∈ {1, . . . , n},
j=1
n
(4.5e) − xj ≤ 1 + (n − 1)(1 − y2 ),
j=1
(4.5f) xn+1 = y2 ,
(4.5g) y1 + y2 = 1,
(4.5h) y ∈ {0, 1}2.
p(n)
By Lemma 3.5, the number of extreme rays of this cone is at most p(n)−3−k , which
is also a polynomial. Hence, no formulation of this form can have an LP relaxation
that projects to conv(S).
Fortunately, by allowing a growing number of auxiliary variables, the following
proposition by Balas, Jeroslow, and Lowe [13, 90, 113] yields polynomial-sized formu-
lations for a wide range of disjunctive constraints that include the set in Example 4.
We postpone the proof of this proposition to section 5.1, where we consider a slightly
more general version of this formulation.
k
Proposition 4.2. Let P i i=1 be a finite family of polyhedra with a common
recession cone (i.e., P∞i j
= P∞ for all i, j), such that P i = x ∈ Qn : Ai x ≤ bi for
k
all i. Then, a locally ideal MIP formulation of S = i=1 P i is given by
(4.6a) Ai xi ≤ bi yi ∀i ∈ {1, . . . , k},
k
(4.6b) xi = x,
i=1
MIXED INTEGER LINEAR PROGRAMMING FORMULATION TECHNIQUES 21
k
(4.6c) yi = 1,
i=1
(4.6d) yi ≥ 0 ∀i ∈ {1, . . . , k},
(4.6e) x ∈Q
i n
∀i ∈ {1, . . . , k},
(4.6f) y∈Z . k
k
(5.1) S= P i ∩ (Qn1 × Zn2 ) ,
i=1
i k
where P i=1 is a finite family of polyhedra with a common recession cone. Hooker
[76] showed that such sets can be modeled through a simple extension of formulation
(4.6). Hooker also showed that this extension is sharp if the formulations of these
mixed integer sets are sharp (i.e., if P i = conv P i ∩ (Qn1 × Zn2 ) ). However, achiev-
ing this could require a large number of inequalities in the descriptions of the P i s.
Fortunately, as noted in [37], we may significantly reduce the number of inequalities
by using auxiliary variables in the description of P i . This results in the following
generalization of formulation (4.6) that also yields a locally ideal or sharp formulation
when locally ideal or sharp extended formulations of P i ∩ (Qn1 × Zn2 ) are used.
k
Proposition 5.1. Let P i i=1 be a finite family of polyhedra with a common
recession
cone (i.e., P∞i j
= P∞ for all i, j ∈ {1, . . . , k}) and p ∈ Zk+ be such that
P = x ∈ Q : ∃w ∈ Q s.t. (x, w) ∈ Qi , where
i n pi
Qi = (x, w) ∈ Qn × Qpi : Ai x + Di w ≤ bi
for Ai ∈ Qmi ×n , Di ∈ Qmi ×pi , and bi ∈ Qmi for each i ∈ {1, . . . , k}. Then, for any
k
n1 , n2 ∈ Z+ such that n = n1 + n2 , an MIP formulation of S = i=1 P i ∩ (Qn1 × Zn2 )
is given by
Furthermore, if P i = conv P i ∩ (Qn1 × Zn2 ) for all i ∈ {1, . . . , k}, then (5.2) is
sharp, and if ext Qi ⊆ Qn1 × Zn2 × Qpi for all i ∈ {1, . . . , k}, then (5.2) is locally
ideal.
Proof. For validity of (5.2), without loss of generality, assume y1 = 1 and yi = 0
for all i ≥ 2. Then x1 ∈ P 1 and by Proposition 3.11 we have that xi ∈ P∞ i
for all
1
i ≥ 2. Then by the common recession cone assumption we have x ∈ P∞ for all i ≥ 2
i
and hence x ∈ P 1 .
MIXED INTEGER LINEAR PROGRAMMING FORMULATION TECHNIQUES 23
To prove sharpness let (x, w, y) be feasible for the LP relaxation of (5.2) and
I = {i ∈ {1, . . . , k} : yi > 0}. Then x = i∈I yi xi /yi , i∈I yi = 1, y ≥ 0, and
xi /yi , wi /yi ∈ Qi for all i ∈ I. By the assumption on P i we then have that
xi /yi ∈ conv P i ∩ (Qn1 × Zn2 ) and hence
k k
x ∈ conv conv P ∩ (Q × Z )
i n1 n2
= conv P ∩ (Q × Z ) = conv(S).
i n1 n2
i=1 i=1
To prove locally idealness first note that if (x, w, y) is an extreme point of the LP
relaxation of (5.2) and y ∈ {0, 1}k , we may again assume without loss of generality
that y1 = 1 and yi = 0 for all i ≥ 2. Then by extremality of (x, w, y) we have
that xi = 0 and wi = 0 for all i ≥ 2, x = x1 , and x1 , w1 ∈ ext Q1 . Then, by
the assumption on Q1 we have x = x1 ∈ Qn1 × Zn2 and hence (x, w, y) satisfies the
integrality constraints. To finish the proof, assume for a contradiction that (x, w, y)
is an extreme point of the LP relaxation of (5.2) such that y ∈ / {0, 1}k . Without loss
of generality we may assume that y1 , y2 ∈ (0, 1). Let ε = min{y1 , y2 , 1 − y1 , 1 − y2 } ∈
(0, 1),
⎧ ⎧
⎪
⎨yi + ε, i = 1, ⎪
⎨yi − ε, i = 1,
yi = yi − ε, i = 2, y i = yi + ε, i = 2,
⎪
⎩ ⎪
⎩
yi , otherwise, yi otherwise.
xi = (y i /yi )xi , w i = (y i /yi )wi , xi = (y i /yi )xi , and w i = (y i /yi )wi for i ∈ {1, 2},
k k
xi = xi = xi and w i = w i = wi for all i ∈/ {1, 2}, x = i=1 xi , and x = i=1 xi . Then
(x, w, y) = (x, w, y), (x, w, y) = (1/2)(x, w, y) + (1/2)(x, w, y), and (x, w, y), (x, w, y)
are feasible for the LP relaxation of (5.2), which contradicts (x, w, y) being an extreme
point.
Formulation 5.2 can be used to construct several known formulations for piecewise
linear functions and more general disjunctive constraints. The following proposition
illustrates this by constructing a variant of (5.2) that is convenient when P i are
described through their extreme points and rays. The resulting formulation is a
straightforward extension of a formulation for disjunctive constraints introduced by
Jeroslow and Lowe [90, 113].
k
Corollary 5.2. Let P i i=1 ⊆ Qn be a finite family of polyhedra with a com-
mon recession cone C. Then, for any n1 , n2 ∈ Z+ such that n = n1 + n2 , an MIP
k
formulation of S = i=1 P i ∩ (Qn1 × Zn2 ) is given by
k
(5.3a) vλiv + rμr = x,
i=1 v∈ext(P i ) r∈ray(C)
(5.3b) λiv = yi ∀i ∈ {1, . . . , k},
v∈ext(P i )
k
(5.3c) yi = 1,
i=1
(5.3d) λiv ≥ 0 ∀i ∈ {1, . . . , k}, v ∈ ext P i ,
(5.3e) μr ≥ 0 ∀r ∈ ray(C),
(5.3f) y ∈ {0, 1} ,
k
Furthermore, if P i = conv P i ∩ (Qn1 × Zn2 ) for all i ∈ {1, . . . , k}, then (5.3) is a
locally ideal formulation of S.
Proof. By Theorem 3.6 we have that P i is the projection onto the x variables
of
⎧ ⎫
⎪
⎪ vλv + rμr = x⎪ ⎪
⎪
⎪ ⎪
⎪
⎪
⎪ v∈ext(P i ) r∈ray(C) ⎪
⎪
⎪
⎪ ⎪
⎪
⎨ ⎬
ray(C) i
ext(P ) λ = 1
Q = (x, μ, λ) ∈ Q × Q
i n
×Q : v
⎪
⎪ v∈ext(P i ) ⎪
⎪
⎪
⎪ ⎪
⎪
⎪
⎪ ≥ ∀v ∈ i ⎪
⎪
⎪
⎪
λv 0 ext P ⎪
⎪
⎩ ⎭
μr ≥ 0 ∀r ∈ ray (C)
for all i ∈ {1, . . . , k}. Using these extended formulations of the Qi s we have that
formulation (5.2) for S is given by
(5.4a) vλiv + rμir = xi ∀i ∈ {1, . . . , k},
v∈ext(P i ) r∈ray(C)
(5.4b) λiv = yi ∀i ∈ {1, . . . , k},
v∈ext(P i )
k
(5.4c) xi = x,
i=1
k
(5.4d) yi = 1,
i=1
(5.4e) λiv ≥ 0 ∀i ∈ {1, . . . , k}, v ∈ ext P i ,
(5.4f) μir ≥ 0 ∀i ∈ {1, . . . , k}, r ∈ ray (C) ,
(5.4g) yi ≥ 0 ∀i ∈ {1, . . . , k},
(5.4h) x ∈Q
i n
∀i ∈ {1, . . . , k},
ext(P i )
(5.4i) λ ∈Q
i
∀i ∈ {1, . . . , k},
ray(C)
(5.4j) μ ∈Q
i
∀i ∈ {1, . . . , k},
(5.4k) y∈Z , k
We claim that the LP relaxation of (5.3) is equal to the image of the LP relaxation
of (5.4) through the linear transformation that projects out the xi and μi variables
k
and lets μ = i=1 μi . We first show that the LP relaxation of (5.3) is contained
in the image of the LP relaxation of (5.4). For this, simply note that any point in
the LP relaxation of (5.3) is the image of the point in the LP relaxation of (5.4)
obtained by letting xi = i
v∈ext(P i ) vλv + (1/k)
i
r∈ray(C) rμr and μr = (1/k)μr
for all i ∈ {1, . . . , k} and r ∈ ray (C). The reverse containment is straightforward.
Validity then follows directly from Proposition 5.1.
For locally idealness note that by Proposition 2.5 and the assumption on P i we
have that ext P i ⊆ Qn1 × Zn2 . By noting that (x, μ, λ) ∈ ext Qi if and only
if x ∈ ext P i , μ = 0, λx = 1, and λv = 0 for v ∈ ext P i \ {x}, we have that
i
ext Qi ⊆ Qn1 ×Zn2 ×Qray(C) ×Qext(P ) . Then by Proposition 5.1 we have that (5.4)
MIXED INTEGER LINEAR PROGRAMMING FORMULATION TECHNIQUES 25
is locally ideal and hence so is (5.3) by Proposition 3.10 and the linear transformation
described in the previous paragraph.
To distinguish extreme point/ray formulation (5.3) from inequality formulation
(5.2), we refer to the first one as the V-formulation and to the second as the H-
formulation. While V-formulation (5.3) can be derived from H-formulation (5.2),
their application to specific disjunctive constraints can yield formulations with very
different structures. The following two examples illustrate this for the case n2 = 0 in
both formulations and pi = 0 for all i ∈ {1,n. . . , k} in Proposition 5.1.
Example 6. Consider the set S = i=1 {x ∈ Qn : −1 ≤ xi ≤ 1, xj = 0 ∀j = i}
from Example 2. H-formulation (5.2) for S is given by
Similarly to the proof of Proposition 5.1, we can show that the projection of (5.5) onto
the x and y variables is precisely formulation (4.3) from Example 3, which is given
by
we obtain a standard MIP formulation for piecewise linear functions that is denoted
the multiple choice model in [155]. Similarly to the proof of Proposition 5.1, we can
show that the multiple choice model is the projection onto the x, xi , y, and z variables
of (5.8). If we instead use formulation (5.3) for characterization (3.4b) of gr(f ), we
obtain the formulation of gr(f ) given by
k
(5.10a) vλiv = x,
i=1 v∈ext(Qi )
k
(5.10b) f (v)λiv = z,
i=1 v∈ext(Qi )
(5.10c) λiv = yi ∀i ∈ {1, . . . , k},
v∈ext(Qi )
k
(5.10d) yi = 1,
i=1
(5.10e) λiv ≥ 0 ∀i ∈ {1, . . . , k}, v ∈ ext Qi ,
(5.10f) y ∈ {0, 1}k ,
which is a standard MIP formulation for piecewise linear functions that is denoted the
disaggregated convex combination model in [155].
For examples of formulation (5.2) with pi > 0 and n2 > 0 we refer the reader to
[37] and [76], respectively.
6. Projected Formulations. One disadvantage of basic formulations (5.2) and
(5.4) is that they require multiple copies of the original x variables (i.e., the xi vari-
ables) or of some auxiliary variables (i.e., the λi variables). In this section we study
formulations that do not use these copies of variables and are hence much smaller.
The cost of this reduction in variables is usually a loss of strength, but it some cases
MIXED INTEGER LINEAR PROGRAMMING FORMULATION TECHNIQUES 27
this loss can be avoided. We begin by considering the well-known Big-M approach
and an alternative formulation that combines the Big-M approach with a formulation
tailored to polyhedra with a common structure. We then consider the strength of
both classes of formulations in detail.
6.1. Traditional Big-M Formulations. One way to write MIP formulations with-
out using copies of the original variables is to use the Big-M technique. The following
proposition proven in [164] for the bounded case and in [76] for the unbounded de-
scribes the technique.
k
Proposition 6.1. Let P i i=1 be a finite family of polyhedra with a common
recession cone (i.e., P∞ i
= P∞ j
for all i, j), such that P i = x ∈ Qn : Ai x ≤ bi ,
where Ai ∈ Qmi ×n and bi ∈ Qmi for all i ∈ {1, . . . , k}. Furthermore, for each
i, j ∈ {1, . . . , k}, i = j, and l ∈ {1, . . . , mi }, let Mli,j ∈ Q be such that
(6.1) Mli,j ≥ max ai,l x : Aj x ≤ bj ,
x
where a ∈ Q is the lth row of A . Then, for any n1 , n2 ∈ Z+ such that n = n1 +n2 ,
i,l n i
k
an MIP formulation for S = i=1 P i ∩ (Qn1 × Zn2 ) is given by
(6.2a) Ai x ≤ bi + M i − bi (1 − yi ) ∀i ∈ {1, . . . , k},
k
(6.2b) yi = 1,
i=1
(6.2c) yi ≥ 0 ∀i ∈ {1, . . . , k},
(6.2d) y∈Z , k
Unfortunately, as the following example shows, even the strongest version of (6.2)
can fail to be sharp.
Example 9. The strongest version of formulation (6.2) for S from Example 4 is
precisely the nonsharp formulation (4.5).
For more discussion about Big-M formulations for general and specially structured
sets, we refer the reader to [18, 76, 79, 164].
6.2. Hybrid Big-M Formulations. A different class of projected formulations was
introduced by Balas, Blair, and Jeroslow [12, 26, 87] for families of polyhedra with
a common left-hand side matrix (i.e., where Ai = Aj for all i, j). Balas, Blair, and
Jeroslow showed that such formulations can have very favorable strength properties.
However, while the common left-hand side structure appears in many problems (see
Example 10), these formulations still have more limited applicability than the tra-
ditional Big-M formulation from Proposition 6.1. For this reason we here combine
the projected formulation of Balas, Blair, and Jeroslow with the traditional Big-M
formulation to introduce the following hybrid Big-M formulation that generalizes the
former and improves upon the latter. While this formulation does not require the
common left-hand side structure, it is equipped to exploit it when present.
p ks
Proposition 6.2. For k ∈ Zp+ , let s=1 P s,i i=1 be a finite family of polyhedra
s,i t,j
with a common recession cone (i.e., P∞ = P∞ for all s, t, i, j) such that
P s,i = x ∈ Qn : As x ≤ bs,i ,
where As ∈ Qms ×n and bs,i ∈ Qms for each s ∈ {1, . . . , p} and i ∈ {1, . . . , ks }.
Furthermore, for each s, t ∈ {1, . . . , p}, i ∈ {1, . . . , ks }, and l ∈ {1, . . . , ms }, let
Mls,t,i ∈ Q be such that
(6.3) Mls,t,i ≥ max as,l x : At x ≤ bt,i ,
x
where as,l ∈ Qn is the lth row of As . Then, for any n1 , n2 ∈ Z+ such that n = n1 +n2 ,
p ks
an MIP formulation for S = s=1 i=1 P s,i ∩ (Qn1 × Zn2 ) is given by
p
kt
(6.4a) As x ≤ M s,t,i yt,i ∀s ∈ {1, . . . , p},
t=1 i=1
p
ks
(6.4b) ys,i = 1,
s=1 i=1
(6.4c) ys,i ≥ 0 ∀s ∈ {1, . . . , p}, i ∈ {1, . . . , ks } ,
(6.4d) ys,i ∈ Z ∀s ∈ {1, . . . , p}, i ∈ {1, . . . , ks } ,
(6.4e) x ∈ Qn1 × Zn2 .
In particular, we may take M s,s,i = bs,i for all s ∈ {1, . . . , p}, i ∈ {1, . . . , ks }.
Validity of this formulation is again straightforward from finiteness of the Mls,t,i ,
which can be proven analogously to Proposition 6.1. Furthermore, the strongest
possible version of formulation (6.4) is again obtained when equality holds in (6.3).
In particular, M s,s,i = bs,i is the strongest possible coefficient unless some P s,i has
a redundant constraint. Of course, even in the case of a redundant constraint, it is
always valid and convenient to use Big-M constants such that M s,s,i ≤ bs,i . For this
reason, we assume this to be the case from now on.
MIXED INTEGER LINEAR PROGRAMMING FORMULATION TECHNIQUES 29
k
(6.5d) yi = 1,
i=1
(6.5e) λv ≥ 0 ∀v ∈ V,
(6.5f) μr ≥ 0 ∀r ∈ ray(C),
(6.5g) y ∈ {0, 1}k ,
(6.5h) x∈ Qn1 × Zn2 ,
and
(6.6a) vλv + rμr = x,
v∈V r∈ray(C)
(6.6b) λv = 1,
v∈V
(6.6c) λv ≥ yi ∀i ∈ {1, . . . , k},
v∈ext(P i )
k
(6.6d) yi = 1,
i=1
(6.6e) λv ≥ 0 ∀v ∈ V,
(6.6f) μr ≥ 0 ∀r ∈ ray(C),
(6.6g) y ∈ {0, 1}k ,
(6.6h) x∈ Qn1 × Zn2 .
Furthermore, if P i = conv P i ∩ (Qn1 × Zn2 ) for all i ∈ {1, . . . , k}, then both formu-
lations are sharp.
Proof. Let
⎧ ⎫
⎪
⎪ vλv + rμr = x⎪
⎪
⎪
⎪ ⎪
⎪
⎪
⎪ v∈V r∈ray(C) ⎪
⎪
⎪
⎪ ⎪
⎪
⎪
⎪ ⎪
⎪
⎪ λv = 1⎪
⎪
⎪
⎪
⎨ ⎪
⎬
v∈V
Q = (x, λ, μ) ∈ Qn × QV × Qray(C) :
i
.
⎪
⎪ λv ≤ 1 ∀v ∈ ext P i ⎪
⎪
⎪
⎪ ⎪
⎪ i ⎪
⎪
⎪ λv ≤ 0 ∀v ∈ V \ ext P ⎪⎪
⎪
⎪
⎪ ⎪
⎪
⎪
⎪ λv ≥ 0 ∀v ∈ V ⎪
⎪
⎪
⎪ ⎪
⎪
⎩ ⎭
μr ≥ 0 ∀r ∈ ray(C)
k
Validity of (6.5) follows by noting that i=1 P i is the projection onto the x variables
k k
of i=1 Qi and that (6.5) is the simple version of (6.4) for i=1 Qi . Validity of (6.6)
MIXED INTEGER LINEAR PROGRAMMING FORMULATION TECHNIQUES 31
For sharpness, note that the projection onto the x variables of the LP relaxation
of both formulations is contained in conv (V ) + C. The result then follows because,
under the assumption on P i and by Theorem 3.6, we have
k
conv(S) = conv P ∩ (Q × Z )
i n1 n2
i=1
k
= conv conv P ∩ (Q i n1
×Z ) n2
i=1
k
i
= conv P
i=1
k
i
= conv conv ext P +C
i=1
k
= conv conv ext P i +C
i=1
k
i
= conv ext P + C = conv(V ) + C.
i=1
While (6.5) and (6.6) are equivalent, their LP relaxations are not contained in one
another. Of course this can only happen because neither formulation is locally ideal.
In fact, Lee and Wilson [106, 165] show that constraints (6.6c) are valid inequalities for
the convex hull of integer feasible points of (6.5) and hence can be used to strengthen
it. These inequalities are sometimes facet defining and are part of a larger class
of inequalities that are enough to describe the convex hull of integer feasible points
of (6.5). Unfortunately, the number of such inequalities can be exponential in k.
However, in section 7 we show how an LP-based separation of these inequalities allows
us to construct a different formulation that ends up being equivalent to (5.3). If
we specialize formulation (6.5) to piecewise linear functions, we obtain the following
formulation from [106, 165].
Example 11. Let f : D ⊆ Qd → Q be a piecewise linear function defined by (3.3)
for a given finite family of polytopes {Qi }ki=1 . Formulation (6.5) for characterization
32 JUAN PABLO VIELMA
k
(6.7c) λv ≤ yi ∀v ∈ ext Qi ,
i:v∈ext(Qi ) i=1
k
(6.7d) yi = 1,
i=1
(6.7e) λv ≥ 0 ∀v ∈ ext Qi ,
(6.7f) y ∈ {0, 1}k ,
which is a standard MIP formulation for piecewise linear functions that is denoted the
convex combination model in [155].
6.3. Formulation Strength. The traditional Big-M formulation has been signif-
icantly more popular than the simple version of the hybrid Big-M formulation. One
reason for this is the somewhat limited applicability of the latter formulation. The
general version of the hybrid Big-M formulation resolves this issue as it can be used in
the same class of problems as the traditional Big-M formulation. In addition, the hy-
brid Big-M formulation provides an advantage over the traditional one with respect to
size, even if only partial common left-hand side structure is present in the polyhedra.
Indeed, both formulations have the same number of variables, but the hybrid formu-
p p
lation has 1 + s=1 ms constraints while the traditional one has 1 + s=1 ms × ks
constraints. Of course, such an advantage in size is meaningless if it is accompanied
by a significant reduction in strength. For this reason we now study the relative and
absolute strengths of these formulations. In particular, concerning the hybrid Big-
M formulation we show that it is always at least as strong as the traditional Big-M
formulation, that its simple version can be sharp and strictly stronger than the tra-
ditional formulation, but that even its strongest version can fail to be locally ideal
or sharp. We begin with the following proposition that shows that one case where
the hybrid and traditional formulations coincide is when S is the union of only two
polyhedra.
Proposition 6.4. If S is the union of two polyhedra whose description does not
include any redundant constraints, and equality is taken in (6.1) and (6.3), then the
LP relaxations of formulations (6.2) and (6.4) are equivalent.
Proof. For (6.2) the case of two polyhedra corresponds to k = 2, and in this case
(6.2a) is equal to
A1 x ≤ b1 + M 1,2 − b1 (1 − y1 ) ,
A2 x ≤ b2 + M 2,1 − b2 (1 − y2 ) .
A1 x ≤ M 1,1,1 y1 + M 1,1,2 y2
MIXED INTEGER LINEAR PROGRAMMING FORMULATION TECHNIQUES 33
and the equivalence follows from y1 + y2 = 1 by noting that for this case A1 =
A2 , M 1,2 = M 1,1,2 = b2 , and M 2,1 = M 1,1,1 = b1 , because of the nonredundancy
assumption. In the latter, (6.4) is equal to
A1 x ≤ M 1,1,1 y1 + M 1,2,1 y2 ,
A2 x ≤ M 2,1,1 y1 + M 2,2,1 y2 ,
and the equivalence follows from y1 + y2 = 1 by noting that for this case M 1,1,1 = b1 ,
M 2,2,1 = b2 , M 1,2,1 = M 1,2 , and M 2,1,1 = M 2,1 , because of the nonredundancy
assumption.
Unfortunately, as the following example shows, this equal strength can fail to
yield sharp formulations even for the case of equal left-hand side matrices and no
redundant constraints.
Example 12. For
⎡ ⎤ ⎛ ⎞ ⎛ ⎞
1 0 1 1 2
⎢ −1 0 1 ⎥ ⎜1⎟ ⎜2⎟
⎢ ⎥ ⎜ ⎟ ⎜ ⎟
A=⎢ ⎢ 0 1 1 ⎥ 1 ⎜ ⎟ 1
⎥ , b = ⎜2⎟ , b = ⎜1⎟ ,
⎜ ⎟
⎣ 0 −1 1 ⎦ ⎝ 2 ⎠ ⎝1⎠
0 0 −1 0 0
let P 1 = {x ∈ Q3 : Ax ≤ b1 } and P 2 = {x ∈ Q3 : Ax ≤ b2 }. The strongest versions
of formulations (6.2) and (6.4) for S = P 1 ∪ P 2 are equal to
⎛ ⎞
1
⎜ 1 ⎟
⎜ ⎟
(6.8a) Ax ≤ b + (1 − y1 ) ⎜
1 ⎟
⎜ −1 ⎟ ,
⎝ −1 ⎠
0
⎛ ⎞
−1
⎜ −1 ⎟
⎜ ⎟
(6.8b) Ax ≤ b2 + (1 − y2 ) ⎜
⎜ 1 ⎟,
⎟
⎝ 1 ⎠
0
(6.8c) y1 + y2 = 1,
(6.8d) y ∈ {0, 1}2.
Formulation (6.8) is not sharp because x1 = x2 = 0, x3 = 3/2, y1 = y2 = 1/2 is
feasible for its LP relaxation and x3 ≤ 1 for any x ∈ conv(S).
The lack of sharpness of formulation (6.8) in Example 12 can be corrected by
simply adding x3 ≤ 1, as conv(S) is exactly the projection of the LP relaxation of
(6.8) onto the x variables plus this inequality. However, this correction cannot always
be done with a polynomial number of inequalities (e.g., see Lemma 4.1). Furthermore,
as the following theorem by Blair [26] shows, recognizing sharpness of even the simple
version of hybrid Big-M formulation (6.4) is hard.
Theorem 6.5. Let p = 1 and M 1,1,i = b1,i for all i ∈ {1, . . . , k} in formulation
(6.4). Deciding whether the resulting formulation is sharp is NP-hard.
Fortunately, Balas, Blair, and Jeroslow gave some practical (but not exhaustive)
necessary and sufficient conditions for sharpness of the simple version of (6.4) [12,
26, 87]. An extremely useful example of such conditions is given by the following
proposition.
34 JUAN PABLO VIELMA
(6.9a) x1 ≤ y1 ,
(6.9b) xi ≤ yi−1 + yi ∀i ∈ {2, . . . , k},
(6.9c) xk+1 ≤ yk ,
k
(6.9d) yi = 1,
i=1
(6.9e) y ∈ {0, 1}k ,
(6.10a) x1 ≤ y1 ,
(6.10b) xj ≤ 1 − yi ∀j ∈ {1, . . . , k + 1} , i ∈ {1, . . . , k} \ {j − 1, j},
(6.10c) xk+1 ≤ yk ,
k
(6.10d) yi = 1,
i=1
Lemma 6.8. Hybrid Big-M formulation (6.9) is always sharp, while traditional
Big-M formulation (6.10) is not sharp for k ≥ 4.
Proof. The first formulation is sharp by Proposition 6.7.
For the lack of sharpness of the second formulation, note that x1 = xk+1 = 1/k,
xj = (k − 1)/k for j ∈ {2, . . . , k}, and yi = 1/k for i ∈ {1, . . . , k} is feasible for its LP
relaxation.
For k ≥ 4 this solution is not in conv(S) and hence this Big-M formulation is not
sharp (e.g., note that this x does not belong to the projection of the LP relaxation of
(6.9) onto the x variables).
Finally, the following proposition shows that the hybrid formulation is always at
least as strong as the traditional formulation if equivalent Big-M constants are used.
Proposition 6.9. If the Big-M constants in (6.1) and (6.3) are identical and
such that M s,s,i ≤ bs,i , then hybrid Big-M formulation (6.4) is at least as strong as
traditional Big-M formulation (6.2).
Proof. Using the notation of formulation (6.4) (i.e., accounting for the possible
common left-hand side matrices), (6.2a) from the traditional Big-M formulation is
equal to
s,i
As x ≤ bs,i + M − bs,i (1 − ys,i ) ∀s ∈ {1, . . . , p} , i ∈ {1, . . . , ks },
s,i
where M l = maxt∈{1,...,p}, j∈{1,...,kt }, (t,j)=(s,i) Mls,t,j . Using the fact that in this case
p ks
(6.2b) is equal to s=1 i=1 ys,i = 1, we can rewrite this equation as
s,i
As x ≤ bs,i ys,i + M yt,i ∀s ∈ {1, . . . , p} , i ∈ {1, . . . , ks }.
t∈{1,...,p},
j∈{1,...,kt }
(t,j)=(s,i)
The result then follows from assumption M s,s,i ≤ bs,i and because, by definition,
s,i
M ≥ M s,t,i .
36 JUAN PABLO VIELMA
Proposition 6.9 and the formulation sizes suggest that, at least from a theoretical
perspective, hybrid Big-M formulation (6.4) is always preferable to traditional Big-
M formulation (6.2). This theoretical advantage often results in a computational
advantage [156]. However, the complexities of state-of-the-art solvers could still allow
the traditional formulation to outperform the hybrid one, particularly when the size
and strength differences are small.
7. Large Formulations. The aim of the techniques reviewed in this survey is
to construct strong, but small MIP formulations. However, there are many cases in
which all known MIP formulations (strong or weak) are large. If these formulations
are such that the number of variables is small and the number of constraints is large,
but can be separated fast, it is sometimes possible to use them in a branch-and-
cut procedure [131, 39, 128]. Similarly, when only the number of variables is large,
the formulations can be used in column generation or branch-and-price procedures
[15]. These procedures can be very effective and so can extensions such as branch-
and-cut-and-price [48] and branch-and-price for extended formulations [142]. For this
reason, when both large and small (usually extended) formulations are available it is
not always clear which is more convenient. Sometimes it is advantageous to solve the
large formulation with one of these specialized procedures [40, 50, 32], and other times
it is more convenient to solve the small extended formulation directly [31]. Exploring
these alternatives is beyond the scope of this survey, so in this section we restrict
our attention to one class of large formulations from which alternative small extended
formulations can be easily constructed. Such a construction was introduced by Martin
[115] (see also [30]) and can be summarized in the following proposition.
Proposition 7.1. Let Q := {x ∈ Qn : Ax ≤ b, Dx ≤ d} and suppose there is a
separation LP problem for R := {x ∈ Qn : Dx ≤ d}. That is, there exist F ∈ Qr×n ,
H ∈ Qr×m , c ∈ Qm , and g ∈ Qr for which the LP problem given by
n
m
(7.1a) z(x) :=max xi πi + c j ρj ,
i=1 j=1
(7.1b) F π + Hρ ≤ g,
(7.1c) ρj ≥ 0 ∀j ∈ {1, . . . , m},
(7.2a) Ax ≤ b,
(7.2b) F T w = x,
(7.2c) H T w ≥ c,
(7.2d) g T w ≤ 0,
(7.2e) wj ≥ 0 ∀j ∈ {1, . . . , r}.
k
yi = 1,
i=1
λv ≥ 0 ∀v ∈ V,
μr ≥ 0 ∀r ∈ ray(C),
y ∈ {0, 1}k .
While (6.5) is sharp, it is not always locally ideal. However, all facet-defining in-
equalities of the convex hull of solutions to (6.5) are of the form λv ≥ 0, μr ≥ 0,
or
(7.3) yi ≤ λv
i∈I v∈U
k
(7.4a) max y i αi + λ v βv ,
i=1 v∈V
k
(7.5e) yi = 1,
i=1
(7.5f) λv ≥ 0 ∀v ∈ V,
(7.5g) μr ≥ 0 ∀r ∈ ray(C),
(7.5h) wvi ≥0 ∀i ∈ {1, . . . , k}, ∀v ∈ V,
(7.5i) y ∈ {0, 1}k .
It is interesting to note that by eliminating variables λv and renaming wvi to λiv we
obtain formulation (5.3). In other words, as expected, if we try to strengthen (6.5) or
(6.6) to recover what we lost by eliminating the copies of λ variables of (5.3) through
Corollary 6.3, we recover (5.3).
Similar formulations for the case in which the separation problem can be solved
using dynamic programming and other extensions can be found in [96, 116]. How-
ever, note that having a generic polynomial time algorithm for the separation is not
sufficient to obtain a polynomial-sized extended formulation. For instance, if S cor-
responds to the incidence vectors of perfect matchings of a complete graph as defined
in (4.8), then all inequalities of conv(S) can be separated in polynomial time [146].
However, Theorem 4.3 shows that S does not have a polynomial-sized sharp extended
formulation. Finally, we note that another way to obtain small formulations from
large ones in a systematic way is to define approximate versions that are slightly
weaker, but significantly smaller [161].
8. Incremental Formulations. All formulations for disjunctive constraints we
k
have considered so far include binary variables y such that i=1 yi = 1. The problem
with such a configuration is that it can be quite detrimental to branch-and-bound
algorithms. To see this, imagine that the optimal solution to the LP relaxation at a
node of the branch-and-bound tree (e.g., the root LP relaxation) is such that y i0 ∈ /Z
for some i0 ∈ {1, . . . , k}. As described in section 2.2, a branch-and-bound-based MIP
solver will branch on yi0 by creating two new LP problems by adding yi0 ≤ yi0 to the
LP relaxation in the first branch and yi0 ≥ y i0 in the second branch. Because the LP
relaxation includes constraints 0 ≤ yi0 ≤ 1, we have that this branching is equivalent
to fixing yi0 = 0 in the first branch and yi0 = 1 in the second one. These branches
are usually denoted down-branching and up-branching, respectively, and have very
different effects on the branch-and-bound tree. The difference is that up-branching
(i.e., fixing yi0 = 1) automatically forces all other yi variables to zero, while down-
branching (i.e., fixing yi0 = 0) does not imply any restriction on other yi variables.
This asymmetry results in what are usually denoted unbalanced branch-and-bound
trees, which can result in a significant slow-down of the algorithm [141].
One way to resolve this issue is to use specialized constraint branching schemes
[141]. In our particular case, an appropriate scheme is the SOS1 branching of Beale
k
and Tomlin [17] that can be described as follows. Because i=1 yi = 1 and y ≥ 0,
we have that for a fractional y there must be i0 < i1 ∈ {1, . . . , k} such that y i0 , y i1 ∈
(0, 1). We can then exclude this fractional solution by creating two new LP problems
by adding y1 = y2 = · · · = yi0 = 0 in the first branch and yi0 +1 = yi0 +2 = · · · = yk = 0
MIXED INTEGER LINEAR PROGRAMMING FORMULATION TECHNIQUES 39
in the second one. In contrast to the traditional up-down variable branching, this
approach usually fixes to zero around half the integer variables in each branch and
hence yields much more balanced branch-and-bound trees.
An alternative to constraint branching is to use a well-known MIP formulation
trick that uses a redefinition of variables y (e.g., [27, 111, 132, 149, 33, 129, 110]).
This transformation can be formalized
in the following straightforward
proposition.
k
Proposition 8.1. Let Δk := y ∈ Qk+ : i=1 y i = 1 and L : Q k
→ Qk be the
linear function defined as
k
L(y)i := yi .
j=i
and for S := {a1 , a2 , . . . , ak } consider the problem minx {|x| : x ∈ S}, which has an
optimal value of 1. This problem can be solved through the standard MIP formulation
given by
(8.1a) min z,
(8.1b) x ≤ z,
(8.1c) −x ≤ z,
k
(8.1d) x= ai y i ,
i=1
k
(8.1e) 1= yi ,
i=1
(8.1f) y ∈ {0, 1}k .
However, using Proposition 8.1 we can construct the alternative MIP formulation of
S given by
(8.2a) min z,
(8.2b) x ≤ z,
40 JUAN PABLO VIELMA
(8.2c) −x ≤ z,
k
(8.2d) x = a1 w1 + (ai − ai−1 )wi ,
i=2
(8.2e) wi ≥ wi+1 ∀i ∈ {1, . . . , k − 1},
(8.2f) w1 = 1,
(8.2g) w ∈ {0, 1}k .
k
v11 + v1i − vri−1
i
wi
i=2
k
ri
+ δji vji − v1i
i=1 j=2
(8.3a) + rμr = x,
r∈ray(C)
r1
(8.3b) δj1 ≤ 1,
j=2
MIXED INTEGER LINEAR PROGRAMMING FORMULATION TECHNIQUES 41
ri
(8.3c) δji ≤ wi ∀i ∈ {2, . . . , k},
j=2
by
k
(9.3a) yi = 1,
i=1
k
(9.3b) hi yi = w,
i=1
Recently, formulations that are essentially identical to this one were independently
proposed in [6, 74, 107, 159, 160]. However, the basic idea behind formulation (9.3)
has in fact been part of the mathematical programming folklore for a long time. For
k
instance, Glover [58] attributes (9.3) for a specific choice of hi i=1 to Sommer [148].
The following proposition from [74, 6] shows how to use (9.3) to construct formu-
lations for simple disjunctive constraints.
k
Proposition 9.2. Let hi i=1 ⊆ {0, 1}log2 (k) be such that hi = hj for any
i = j. Then a locally ideal formulation for S = {a1 , a2 , . . . , ak } is given by
k
(9.4a) yi = 1,
i=1
k
(9.4b) hi yi = w,
i=1
k
(9.4c) ai yi = x,
i=1
If log2 (k) ∈ Z and there exists (c0 , c) ∈ Q × Qlog2 (k) such that
log2 (k)
(9.5) ai = c0 + hij cj ∀i ∈ {1, . . . , k} ,
j=1
then formulation (9.4) can be reduced to the locally ideal formulation given by
log2 (k)
(9.6a) c0 + cj wj = x,
j=1
(9.7a)
k
vλiv + rμr = x,
i=1 v∈ext(P i ) r∈ray(C)
k
(9.7b) λiv = 1,
i=1 v∈ext(P i )
k
(9.7c) hi λiv = w,
i=1 v∈ext(P i )
The following proposition shows an entirely different technique that was intro-
duced in [81].
k
Proposition 9.4. Let P i i=1 be a finite family of polyhedra with a common
i
recession cone (i.e., P∞ = P∞ j
for all i, j), such that P i = x ∈ Qn : Ai x ≤ bi with
k
Ai ∈ Qmi ×n and bi ∈ Qmi for all i ∈ {1, . . . , k}. Also, let hi i=1 ⊆ {0, 1}log2 (k) be
k
such that hi = hj for any i = j. An MIP formulation for S = i=1 P i is given by
⎛ ⎞
log2 (k)
(9.8a) Ai x ≤ bi + M i ⎝ hij − wj + wj ⎠ ∀i ∈ {1, . . . , k},
j=1 j:hij =1 j:hij =0
.
r
kj
(10.1) S= P i,j ,
j=1 i=1
where P i,j are given polytopes. In this case, (10.1) can also be written using the
single disjunctive constraint,
.
r
(10.2) S= P sj ,j .
r
j=1 {1,...,kj }
s∈ j=1
Formulating this single combined constraint directly results in stronger, but larger
MIP formulations than when combining formulations for each constraint. However,
using the rules of propositional calculus we can construct other variations between
extreme cases (10.1) and (10.2) [11, 86]. Systematically constructing such variations
can lead to formulations that effectively balance size and strength (e.g., [139, 143,
144]). Furthermore, in some cases formulations based on (10.1) do not incur any loss
of strength [159, 160].
The reverse transformation from (10.2) to (10.1) is not always evident and some-
times requires the addition of auxiliary variables. Examples of this include the formu-
lations for joint probabilistic constraints introduced in [101, 114, 157]. In this case,
auxiliary binary variables that keep track of violated constraints are used to combine
formulations for single row or k-row probabilistic constraints into a formulation for
the full row joint probabilistic constraint.
For more information on propositional calculus, logic, and other techniques from
constraint programming with MIP formulations, we refer the reader to [78, 88, 164,
76, 79, 77].
11. MIP Representability. A systematic study of which sets can be modeled as
MIP problems began with the work of Meyer [118, 119, 120, 122, 121] and Ibaraki
[81], and the first precise characterization of these sets was given by Jeroslow and
Lowe [90, 113]. For general MIP formulations with unbounded integer variables this
characterization is quite technical, so Jeroslow and Lowe [90, 113] also introduced a
much simpler characterization for MIP formulations with binary or bounded integer
variables. This characterization was later extended by Hooker [76] to consider a
restricted use of unbounded integer variables that nevertheless allows for the modeling
of most sets that are used in practice, while preserving most of the simplicity of
Jeroslow and Lowe’s characterization. The first simple characterization to correspond
precisely to the sets modeled by basic formulations (4.6) was denoted bounded MIP
representability by Jeroslow and Lowe.
Definition 11.1. S ⊆ Qn is bounded MIP representable if and only if it has an
MIP formulation of the form (2.6), where y is additionally constrained to be a binary
vector.
MIXED INTEGER LINEAR PROGRAMMING FORMULATION TECHNIQUES 45
Jeroslow and Lowe showed that bounded MIP representable sets are precisely the
union of a finite number of polyhedra with a common recession cone.
Proposition 11.2. A set S ⊆ Qn is bounded MIP representable if and only if
k
there exists a finite family of polyhedra P i i=1 such that P∞
i j
= P∞ for all i, j and
k
(11.1) S= P i,
i=1
k
(11.2) S=C+ P i,
i=1
k
where P i i=1 is a finite family of polytopes (i.e., P∞ i
= {0} for all i) and C is a
polyhedral cone.
Proof. The fact that a set of the form (11.1) is bounded MIP representable follows
from Proposition 4.2.
For the converse, let (2.6) be an MIP formulation of set S such that y is bounded.
Then the feasible region of this MIP formulation is a finite union of polyhedra (one
for each possible value of y) with the same recession cone. The result then follows by
projecting this union onto the x variables.
Of course, through unary or binary expansion, bounded MIP representability is
equivalent to asking for y to be bounded in (2.6). However, we can extend the class
of MIP representable sets by allowing unbounded integer variables. If we restrict
these unbounded integer variables to be part of the original variables x, we obtain
the second simple characterization introduced by Hooker. Because this representation
does not allow unbounded integer auxiliary variables, we denote it “projected MIP
representability.”
Definition 11.3. S ⊆ Qn is projected MIP representable if and only if it has
an MIP formulation of the form
(11.3a) Ax + Bλ + Dy ≤ b,
(11.3b) x ∈ Qn1 × Zn2 ,
(11.3c) λ ∈ Qs ,
(11.3d) y ∈ {0, 1}t.
Note that (11.3) is a special version of (2.6) in which the linear inequalities are
such that every integer auxiliary variable is either bounded or identical to one of the
original variables. Projected MIP representability certainly subsumes bounded MIP
representability, but as shown by Hooker [76] it can model a much wider class of sets.
Proposition 11.4. A set S ⊆ Rn is projected MIP representable if and only if
k
there exists a finite family of polyhedra P i i=1 such that P∞i
= P∞j
for all i, j and
k
(11.4) S= P i ∩ (Qn1 × Zn2 ) .
i=1
Proof. The fact that a set of the form (11.4) is projected MIP representable
follows from Proposition 5.1.
46 JUAN PABLO VIELMA
k
(11.5) S=M+ Pi
i=1
i k
for a finite family of polytopes P i=1 and a finitely generated monoid M .
Proof. For necessity assume S has a MIP formulation of the form (2.6) and let
Q ⊆ Qn ×Qs ×Qt be the polyhedron described by (2.6a). BypTheorem 3.6 and because
Q is a rational polyhedron, there exist points x̂j , ûj , ŷ j j=1 ⊆ Qn × Qs × Qt and
d
scaled rays x̃l , ũl , ỹ l l=1 ⊆ Zn × Zs × Zt such that for any (x, u, y) feasible for (2.6)
there exist λ ∈ Δp and μ ∈ Qd+ such that
p
d
(x, u, y) = λj x̂j , ûj , ŷ j + μl x̃l , ũl , ỹ l .
j=1 l=1
p
d
(x∗ , u∗ , y ∗ ) := λj x̂j , ûj , ŷ j + (μl − μl ) x̃l , ũl , ỹ l
j=1 l=1
and
d
(11.6) (x∞ , u∞ , y ∞ ) := μl x̃l , ũl , ỹ l .
l=1
For sufficiency, let S be of the form (11.5), let (4.6) be an MIP formulation for
k i d
i=1 P , and let M be generated by r i=1 ⊆ Z . Then an MIP formulation for S
i n
is given by
(11.7f) z ∈ Zd+ .
By comparing (11.2) and (11.5), we see that general MIP representability replaces
the continuous recession directions of polyhedral cone C with the discrete recession
directions of monoid M . The fact that this replacement gives further modeling power
stems from the following lemma, which shows that continuous recession directions
can be obtained from discrete recession directions and a polytope through Minkowski
addition of sets. d
i=1 μi r : μ ∈ Q+ be a polyhedral cone generated by
i d
Lemma 11.7. Let C =
i d d
r i=1 ⊆ Zn . Then C = M + P , where P := i=1 λi r : λ ∈ [0, 1]
i d
is a polytope
d
and M := μ
i=1 i r i
: μ ∈ Zd
+ is a finitely generated integral monoid.
Proof. The result follows from noting that
d
d
d
μi r i = μi r i + (μi − μi ) ri .
i=1 i=1 i=1
Using this lemma we have that bounded MIP characterization (11.2) is equivalent
to
k
(11.8) S = M + Q0 + Qi
i=1
k
for a finite family of polytopes Qi i=0 and a finitely generated integral monoid M .
k
We then obtain general MIP characterization (11.5) by noting that Q0 + i=1 Qi =
k i
i=1 Q0 + Q and that P i = Q0 + Qi is a polytope for every i. We could carry
out the reverse transformation if we could factor out a common polyhedron Q0 from
polytopes P i from bounded MIP characterization (11.2). However, as the following
example from [90, 113] shows, this factorization cannot always be done and general
characterization (11.5) includes more cases than bounded characterization (11.1)–
(11.2) and than projected MIP characterization (11.4).
Example 17. Let S = ({0} ∪ [2, ∞)) ∩ Z. S is a finitely generated integral
monoid and a general MIP formulation for S is given by
x − 2y1 − 3y2 = 0,
y ∈ Z2+ .
48 JUAN PABLO VIELMA
n−1
= {x ∈ Qn : xn = xi , xn ≥ xj ≥ 0 ∀j ∈ {1, . . . , n}} .
i=1
n−1
We have that conv(S) = x ∈ Qn : xn ≤ i=1 xi , xn ≥ xj ≥ 0 ∀j ∈ {1, . . . , n}
is a polyhedron, but S is not MIP representable.
We now illustrate some of these representability concepts by considering MIP
formulations of piecewise linear functions. However, we first introduce the following
example, which shows that rationality is crucial to the presented representability
results. For more information concerning MIP representability for nonpolyhedral or
nonrational polyhedral sets, we refer the reader to [49, 73].
Example √ Consider the MIP formulation
19. with irrational data given by S =
x ∈ R2 : x2 − 2x1 ≤ 0, x2 ≥ 0, x ∈ Z2 . We have that S is closed and the closure
of the convex hull of S is the nonrational polyhedron given by the LP relaxation of
this formulation. However, conv(S) is not closed. Furthermore, we have that S is an
integral monoid (it is composed by integer points and is closed under addition), but
by Lemma 3 in [83] we have that S is not finitely generated.
{ci }ki=1 ⊆ Q, and a finite family of polytopes {Qi }ki=1 such that
k
(11.9a) D= Qi ,
i=1
⎧
⎪ 1
⎪ m x + c1 , x ∈ Q1 ,
⎨
..
(11.9b) f (x) = .
⎪
⎪
⎩ mk x + c , x ∈ Qk .
k
Proof. For item 1 we first note that epi(f ) = C+ + gr(f ), where C+ := {(0, z) ∈
Qn ×Q : z ≥ 0}, and that for f with bounded domain gr(f ) is bounded. Now, if gr(f )
k
is bounded MIP representable, we have by Proposition 11.2 that gr(f ) = i=1 P i for
a finite family of polytopes {P i }ki=1 . Then, epi(f ) = C+ + ki=1 P i = ki=1 C+ + P i
k
and C+ + P i i=1 is a finite family of polyhedra with common recession cone C+ .
Hence, by Proposition 11.2 epi(f ) is bounded MIP representable.
For the “if” part of item 2 let P i := {(x, z) ∈ Qn × Q : x ∈ Qi , z = mi x + ci } for
each i ∈ {1, . . . , k}. Then {P i }ki=1 is a finite family of polytopes and gr(f ) = ki=1 P i ;
hence by Proposition 11.2 gr(f ) is bounded MIP representable. For the only if part,
k
if gr(f ) is bounded MIP representable, then, by Proposition 11.2, gr(f ) = i=1 P i
for a finite family of polytopes {P i }ki=1 . Let Qi be the projection onto the x variables
of polytope P i and let {v j }pj=1 be the extreme points of polytope Qi . We claim that
there exists a solution mi , ci ∈ Qn × Q to the system mi v j + ci = f (v j ) for all
j ∈ {1, . . . , p}. If such a solution exists we have that P i = {(x, z) ∈ Qn × Q : x ∈
Qi , mi x + ci = z}, which gives the desired result. To show the claim first assume
without loss of generality that v d+1 = 0. After that, assume again /without 0p loss of
generality that {v j }dj=1 are linearly independent where 1 < d = rank v j j=1 (if d =
0, then Qi = v d+1 = {0} and the claim is straightforward). Let mi ∈ Qn and ci ∈ Q
be such that ci = f v d+1 = f (0) and mi v j = f (v j ) − ci for all j ∈ {1, . . . , d}, which
exist by the linear independence assumption. To arrive at a contradiction assume
without loss of generality that mi v d+2 + ci > f v d+1 . Because of the assumption on
{v j }dj=1 , there exist λ ∈ Qd such that v d+2 = di=1 λi v i . Furthermore, without loss
of generality we may assume that λ1 ≤ λi for all i ∈ {1, . . . , d}. Let
d+1
1
d
ṽ(δ) := δ vi + (1 − δ)v d+2 = λ̃i (δ)v i ,
i=1
d+1 i=1
1
where λ̃i (δ) := δ d+1 + (1 − δ)λi for all i ∈ {1, . . . , d}. If λ1 ≥ 0, let δ1 := 0, and if
λ1 < 0, let
λ1
δ1 := ∈ (0, 1).
1
λ1 − d+1
d d
If i=1 λi ≤ 1, let δ2 := 0, and if i=1 λi > 1, let
d
λi − 1
δ2 := d
i=1
∈ (0, 1).
i=1 λi − d
d+1
50 JUAN PABLO VIELMA
Then, for δ ∗ = max {δ1 , δ2 } we have i=1 λ̃i (δ ∗ ) ≤ 1 and λ̃i (δ ∗ ) ≥ 0 for all i ∈
d
k
{1, . . . , d}. Hence, ṽ (δ ∗ ) ∈ conv {v j }d+1
j=1 . Now, because gr(f ) =
i
i=1 P , we
∗
have v , f (v ) ∈ P for all i ∈ {1, . . . , d + 2}. Then, by the definition of ṽ (δ ) and
i i i
1
convexity of P i , we have ṽ (δ ∗ ) , j=1 δ d+1
d+1
f (v i )+(1−δ)f (v d+2 ) ∈ P i . Furthermore,
by the definition of m , ci , the assumption on f (v d+2 ), and the fact that 1 − δ ∗ > 0,
i
This function is depicted in black Figure 11.1(a), where its epigraph is depicted in
red. The function is discontinuous at 1 so its graph is not closed and hence cannot
be MIP representable by Proposition 11.8. However, epi(f ) = P 1 ∪ P 2 where P 1 :=
{(x, z) : x + z ≥ 2, z − x ≥ 0, x ≥ 0} and P 2 := {(x, z) : z − x ≥ −1, x ≥ 1} are
polyhedra with a common recession cone. This is illustrated in Figure 11.1(b), where
P 1 is depicted in green and P 2 is depicted in blue. Proposition 11.2 then implies that
epi(f ) is bounded MIP representable.
Some continuous functions with unbounded domains also have bounded MIP
representable graphs, but the necessary conditions for this to hold are more restrictive.
For more details concerning functions with unbounded domains, we refer the reader
to [84, 113, 120], and for discontinuous functions to [155] and the references therein.
3 3
2 2
1 1
0 0
0 1 2 3 0 1 2 3
Acknowledgment. The author would like to thank two anonymous referees and
one associate editor for their helpful feedback.
REFERENCES
[6] W. Adams and S. Henry, Base-2 expansions for linearizing products of functions of discrete
variables, Oper. Res., 60 (2012), pp. 1477–1490.
[7] W. Adams and H. Sherali, A tight linearization and an algorithm for zero-one quadratic
programming problems, Management Sci., 32 (1986), pp. 1274–1290.
[8] F. Al-Khayyal and J. Falk, Jointly constrained biconvex programming, Math. Oper. Res.,
8 (1983), pp. 273–286.
[9] G. Angulo, S. Ahmed, S. S. Dey, and V. Kaibel, Forbidden vertices, Math. Oper. Res., to
appear; doi:10.1287/moor.2014.0673.
[10] E. Balas, Disjunctive programming, Ann. Discrete Math., 5 (1979), pp. 3–51.
[11] E. Balas, Disjunctive programming and a hierarchy of relaxations for discrete optimization
problems, SIAM J. Algebraic Discrete Methods, 6 (1985), pp. 466–486.
[12] E. Balas, On the convex-hull of the union of certain polyhedra, Oper. Res. Lett., 7 (1988),
pp. 279–283.
[13] E. Balas, Disjunctive programming: Properties of the convex hull of feasible points, Discrete
Appl. Math., 89 (1998), pp. 3–44.
[14] K. M. Ball, An elementary introduction to modern convex geometry, in Flavors of Geometry,
S. Levy, ed., Math. Sci. Res. Inst. Publ. 31, Cambridge University Press, Cambridge, 1997,
pp. 1–58.
[15] C. Barnhart, E. Johnson, G. Nemhauser, M. Savelsbergh, and P. Vance, Branch-
and-price: Column generation for solving huge integer programs, Oper. Res., 46 (1996),
pp. 316–329.
[16] A. Barvinok and E. Veomett, The computational complexity of convex bodies, in Surveys
on Discrete and Computational Geometry: Twenty Years Later, J. Goodman, J. Pach,
and R. Pollack, eds., Contemp. Math. 453, AMS, Providence, RI, 2008, pp. 117–138.
[17] E. M. L. Beale and J. A. Tomlin, Special facilities in a general mathematical programming
system for non-convex problems using ordered sets of variables, in OR 69: Proceedings of
the Fifth International Conference on Operational Research, J. Lawrence, ed., Tavistock
Publications, 1970, pp. 447–454.
[18] N. Beaumont, An algorithm for disjunctive programs, European J. Oper. Res., 48 (1990),
pp. 362–371.
[19] A. Ben-Tal and A. Nemirovski, On polyhedral approximations of the second-order cone,
Math. Oper. Res., 26 (2001), pp. 193–205.
[20] D. Bertsimas, S. Gupta, and G. Lulli, Dynamic resource allocation: A flexible and tractable
modeling framework, European J. Oper. Res., 236 (2014), pp. 14–26.
[21] D. Bertsimas and J. Tsitsiklis, Introduction to Linear Optimization, Athena Scientific,
1997.
[22] D. Bertsimas and R. Weismantel, Optimization over Integers, Dynamic Ideas, 2005.
[23] R. Bixby, M. Fenelon, Z. Gu, E. Rothberg, and R. Wunderling, MIP: Theory and
practice—closing the gap, in System Modelling and Optimization, 1999, pp. 19–50.
[24] R. Bixby, M. Fenelon, Z. Gu, E. Rothberg, and R. Wunderling, Mixed-integer program-
ming: A progress report, in The Sharpest Cut: The Impact of Manfred Padberg and His
Work, SIAM, Philadelphia, PA, 2004, Chap. 18, pp. 309–325.
[25] R. Bixby and E. Rothberg, Progress in computational mixed integer programming: A look
back from the other side of the tipping point, Ann. Oper. Res., 149 (2007), pp. 37–41.
[26] C. Blair, Representation for multiple right-hand sides, Math. Programming, 49 (1990), pp. 1–
5.
[27] D. Bricker, Reformulation of special ordered sets for implicit enumeration algorithms, with
applications in nonconvex separable programming, AIIE Trans., 9 (1977), pp. 195–203.
[28] M. N. Broadie, A theorem about antiprisms, Linear Algebra Appl., 66 (1985), pp. 99–111.
[29] S. Burer and A. Letchford, Non-convex mixed-integer nonlinear programming: A survey,
Surv. Oper. Res. Manag. Sci., 17 (2012), pp. 97–106.
[30] R. D. Carr and G. Lancia, Compact vs. exponential-size LP relaxations, Oper. Res. Lett.,
30 (2002), pp. 57–65.
[31] R. D. Carr and G. Lancia, Compact optimization can outperform separation: A case study
in structural proteomics, 4OR, 2 (2004), pp. 221–233.
[32] R. Carvajal, M. Constantino, M. Goycoolea, J. P. Vielma, and A. Weintraub, Impos-
ing connectivity constraints in forest planning models, Oper. Res., 61 (2013), pp. 824–836.
[33] C. Cavalcante, C. C. de Souza, M. Savelsbergh, Y. Wang, and L. Wolsey, Scheduling
projects with labor constraints, Discrete Appl. Math., 112 (2001), pp. 27–52.
[34] J. Cochran, L. A. Cox, Jr., P. Keskinocak, J. Kharoufeh, and J. Smith, Wiley Ency-
clopedia of Operations Research and Management Science, John Wiley & Sons, 2011.
[35] M. Conforti, G. Cornuéjols, and G. Zambelli, Extended formulations in combinatorial
optimization, 4OR, 8 (2010), pp. 1–48.
MIXED INTEGER LINEAR PROGRAMMING FORMULATION TECHNIQUES 53
[123] R. Misener and C. A. Floudas, Global optimization of large-scale generalized pooling prob-
lems: Quadratically constrained MINLP models, Indust. Engrg. Chem. Res., 49 (2010),
pp. 5424–5438.
[124] R. Misener and C. A. Floudas, Piecewise-linear approximations of multidimensional func-
tions, J. Optim. Theory Appl., 145 (2010), pp. 120–147.
[125] R. Misener and C. A. Floudas, Global optimization of mixed-integer quadratically-
constrained quadratic programs (MIQCQP) through piecewise-linear and edge-concave
relaxations, Math. Programming, 136 (2012), pp. 155–182.
[126] R. Misener, C. Gounaris, and C. A. Floudas, Mathematical modeling and global optimiza-
tion of large-scale extended pooling problems with the (EPA) complex emissions con-
straints, Comput. Chem. Engrg., 34 (2010), pp. 1432–1456.
[127] R. Misener, J. Thompson, and C. A. Floudas, APOGEE: Global optimization of stan-
dard, generalized, and extended pooling problems via linear and logarithmic partitioning
schemes, Comput. Chem. Engrg., 35 (2011), pp. 876–892.
[128] J. E. Mitchell, Branch and cut, in Wiley Encyclopedia of Operations Research and Man-
agement Science, John Wiley & Sons, 2011.
[129] R. H. Möhring, A. S. Schulz, F. Stork, and M. Uetz, On project scheduling with irregular
starting time costs, Oper. Res. Lett., 28 (2001), pp. 149–154.
[130] F. M. Muldoon, W. P. Adams, and H. D. Sherali, Ideal representations of lexicographic
orderings and base-2 expansions of integer variables, Oper. Res. Lett., 41 (2013), pp. 32–
39.
[131] G. L. Nemhauser and L. A. Wolsey, Integer and Combinatorial Optimization, Wiley-
Interscience, 1988.
[132] W. Ogryczak, A note on modeling multiple choice requirements for simple mixed integer
programming solvers, Comput. Oper. Res., 23 (1996), pp. 199–205.
[133] M. Oral and O. Kettani, A linearization procedure for quadratic and cubic mixed-integer
problems, Oper. Res., 40 (1992), pp. 109–116.
[134] J. Owen and S. Mehrotra, On the value of binary expansions for general mixed-integer
linear programs, Oper. Res., 50 (2002), pp. 810–819.
[135] M. W. Padberg, The boolean quadric polytope: Some characteristics, facets and relatives,
Math. Programming, 45 (1989), pp. 139–172.
[136] M. W. Padberg, Approximating separable nonlinear functions via mixed zero-one programs,
Oper. Res. Lett., 27 (2000), pp. 1–5.
[137] M. W. Padberg and M. P. Rijal, Location, Scheduling, Design, and Integer Programming,
Springer, 1996.
[138] Y. Pochet and L. Wolsey, Production Planning by Mixed Integer Programming, Springer
Ser. Oper. Res. Financ. Eng., Springer, 2006.
[139] R. Raman and I. Grossmann, Modelling and computational techniques for logic based integer
programming, Comput. Chem. Engrg., 18 (1994), pp. 563–578.
[140] T. Rothvoß, The matching polytope has exponential extension complexity, in Symposium on
Theory of Computing (STOC 2014), D. B. Shmoys, ed., ACM, New York, 2014, pp. 263–
272.
[141] D. M. Ryan and B. A. Foster, An integer programming approach to scheduling, in Computer
Scheduling of Public Transport Urban Passenger Vehicle and Crew Scheduling, A. Wren,
ed., North-Holland, 1981, pp. 269–280.
[142] R. Sadykov and F. Vanderbeck, Column generation for extended formulations, Electron.
Notes Discrete Math., 37 (2011), pp. 357–362.
[143] N. Sawaya and I. Grossmann, A cutting plane method for solving linear generalized disjunc-
tive programming problems, Comput. Chem. Engrg., 29 (2005), pp. 1891–1913.
[144] N. Sawaya and I. Grossmann, A hierarchy of relaxations for linear generalized disjunctive
programming, European J. Oper. Res., 216 (2012), pp. 70–82.
[145] A. Schrijver, Theory of Linear and Integer Programming, Wiley Series in Discrete Mathe-
matics & Optimization, John Wiley & Sons, 1998.
[146] A. Schrijver, Combinatorial Optimization: Polyhedra and Efficiency, Algorithms and Com-
binatorics, Springer, 2003.
[147] H. Sherali and W. P. Adams, A Reformulation-Linearization Technique for Solving Discrete
and Continuous Nonconvex Problems, Kluwer Academic Publishers, 1999.
[148] D. Sommer, Computational Experience with the Ophelie Mixed Integer Code, talk presented
at the International TIMS Conference, Houston, 1972.
[149] C. C. Souza and L. A. Wolsey, Scheduling Projects with Labour Constraints, relatório
técnico ic-97-22, IC - UNICAMP, 1997.
[150] K. Spielberg, Minimal Preferred Variable Methods in 0-1 Programming, Tech. Report 320-
3013, IBM Philadelphia Scientific Center, 1972.
MIXED INTEGER LINEAR PROGRAMMING FORMULATION TECHNIQUES 57