International Journal of

Operations Research

International Journal of Operations Research Vol. 4, No. 3, 118-137 (2007)

Duality for Mixed-Integer Linear Programs

M. Guzelsoy * and T. K. Ralphs
Department of Industrial and Systems Engineering, Lehigh University, Bethlehem, PA 18015

Received May 2006; Revised November 2006; Accepted January 2007

bstract The theory of duality for linear programs is well-developed and has been successful in advancing both the
theory and practice of linear programming. In principle, much of this broad framework can be extended to mixed-integer
linear programs, but this has proven difficult, in part because duality theory does not integrate well with current
computational practice. This paper surveys what is known about duality for integer programs and offers some minor
extensions, with an eye towards developing a more practical framework.
Keywords Duality, Mixed-integer linear programming, Value function, Branch and cut

INTRODUCTION information in the case of branch and cut, which is the

most commonly employed solution algorithm for MILPs
Duality has long been a central tool in the development today.
of both optimization theory and methodology. The study The study of duality for MILPs can be considered to
of duality has led to efficient procedures for computing have two main goals: (1) to develop methods for deriving a
bounds, is central to our ability to perform post facto priori lower bounds on the optimal solution value of a
solution analysis, is the basis for procedures such as specific MILP instance and (2) to develop methods for
column generation and reduced cost fixing, and has yielded determining the effect of modifications to the input data
optimality conditions that can be used as the basis for on the optimal solution and/or optimal solution value post
“warm starting” techniques. Such procedures are useful facto. Methods for producing a priori lower bounds are
both in cases where the input data are subject to useful primarily as a means of solving the original problem,
fluctuation after the solution procedure has been initiated usually by embedding the bounding procedure into a
and in applications for which the solution of a series of branch-and-bound algorithm. Such bounding methods
closely-related instances is required. This is the case for a have received a great deal of attention in the literature and
variety of integer optimization algorithms, including are well-studied. Methods for producing dual information
decomposition algorithms, parametric and stochastic post facto, on the other hand, are useful for performing
programming algorithms, multi-objective optimization sensitivity analyses and for warm starting solution
algorithms, and algorithms for analyzing infeasible procedures. Such methods have received relatively little
mathematical models. attention in the literature. In both cases, the goal is to
The theory of duality for linear programs (LPs) is produce “dual information.” Methods of the second type,
well-developed and has been extremely successful in however, can take advantage of information produced as a
contributing to both theory and practice. By taking by-product of a primal solution algorithm.
advantage of our knowledge of LP duality, it has been The primary goal of this study is to survey previous
possible to develop not only direct solution algorithms for work on methods of the second type with an eye towards
solving LPs but also sophisticated dynamic methods developing a framework for MILP duality that can be
appropriate for large-scale instances. In theory, much of integrated with modern computational practice.
this broad framework can be extended to mixed-integer Computational methods have evolved significantly since
linear programs (MILPs), but this has proven largely most of the work on integer programming duality was
impractical because a duality theory well-integrated with done and a close reexamination of this early work is
practice has yet to be developed. Not surprisingly, it is needed. We have attempted to make the paper as general
difficult to develop a standard dual problem for MILP with and self-contained as possible by extending known results
properties similar to those observed in the LP case. Such from the pure integer to the mixed-integer case whenever
dual problems are generally either not strong or not possible. We have included proofs for as many results as
computationally tractable. Unlike the LP case, dual space would allow, concentrating specifically on results
information is not easy to extract from the most commonly whose proofs were not easily accessible or for which we
employed primal solution algorithms. In Section 4.7, we provide a generalization or alternative approach. The
discuss the challenges involved in extracting dual proofs for all results not included here can be found in the

*
Corresponding author’s email: [email protected]

1813-713X Copyright © 2007 ORSTW

Guzelsoy and Ralphs: Duality for Mixed-Integer Linear Programs 119
IJOR Vol. 4, No. 3, 118-137 (2007)

references cited. This survey draws heavily on the multiplication and

foundational work of a small cadré of authors who i. f (v1)  f(v2) = f(v1  v2) " v1, v2 Î V,
contributed greatly to the study of MILP duality, including ii. f ( v) = f (v) " v Î V , " Î ¡.
Gomory, Johnson, Wolsey, Blair, and Jeroslow, and is in l convex if V and epi (f ) = {(v, y): v Î V, y ³ f (v)} are
many respects an updating and expansion of the excellent convex sets, and
papers of Wolsey (1981) and Williams (1996). l polyhedral if epi (f ) is a polyhedron.

. Definitions and notation Definition 2. For a given k Î ¥ , let

k
l L k = {f | f : ¡ ® ¡ },
Before beginning, we briefly introduce some
terminology and notation. A linear program is the problem of l Lk = {f Î L k | f is linear},
minimizing a linear objective function over a polyhedral l Ck = {f Î Lk | f is convex},
feasible region
l Fk = {f Î Lk | f is subadditive}.
n
P = {x Î ¡ + | Ax = b} (1)
The notation éê ùú for a scalar is used to denote the
defined by rational constraint matrix A Î ¤m ´n and smallest integer greater than or equal to . Similarly, we let
right-hand side vector b Î ¡ m . A mixed-integer linear program ëê ûú = - êé - úù . For a function f Î Lk, êé f úù is the
is an LP with the additional constraint that a specified function defined by éê f úù ( d ) = êé f ( d )úù "d Î ¡ k .
subset of the variables must take on integer values. For the Finally, the l1 norm of a vector x = (x1, ..., xn) is denoted by
remainder of the paper, we address the canonical MILP
= å i =1 x i .
n
instance specified by the constraints in (1), with the integer x 1

variables (those required to take on integer values) indexed

by the set I = {1, ..., r} Í N = {1, ..., n} if r > 0 (otherwise, .2 Outline of the paper
I = Ø). The remaining variables, indexed by the set C = N\I,
constitute the continuous variables. The feasible region is then The outline of the paper is as follows. In Section 2, we
introduce several notions of MILP duality that have
given by S = P Ç ( ¢ r
¡ n r ) and the MILP instance can
appeared in the literature and discuss the relationships
be stated as that of determining between them. In Section 3, we discuss in more detail the
well-known subadditive dual, which yields a generalization of
z IP = min cx (2) many duality results from linear programming, but does
x S
not integrate easily with current computational practice. In
for c ¡ n . The rationality of the constraint matrix A is Section 4, we discuss methods for obtaining dual functions
required to ensure the consistency of (2) and guarantees that provide a lower bound on the objective function value
that any such program is either infeasible, unbounded or of MILP instances in the neighborhood of a given base
has a finite optimal value (Meyer (1974)). In the sequel, we instance and can be seen as solutions to certain dual
refer to this canonical instance of MILP as the primal problems we present in Sections 2 and 3. Finally, in Section
problem. 5, we discuss future research in this area and indicate how
some of the theory presented in this paper may be put into
For a given MILP, we call any x Î S a feasible solution
practice.
with solution value cx and any x* such that cx* = zIP an optimal
solution. Aside from determining the value zIP, the goal of 2. GENERAL FRAMEWORK
solving (2) is to find such an optimal solution. The LP
obtained from a given MILP by removing the integrality A common approach to obtaining an a priori lower
requirements on the variables, i.e., setting I = Ø, is referred bound for a single MILP instance is to construct an
to as the associated LP relaxation. The associated pure integer optimization problem of the form
linear program (PILP), on the other hand, is obtained by
requiring all variables to be integer, i.e., setting r = n. For z D = max f ( v ), (3)
v ÎV
any index set K Í N, AK is the submatrix consisting of the
corresponding columns of A and similarly, yK is the vector with objective function f : V ® ¡ and feasible region
consisting of just the corresponding components of a
vector y. V Í ¡ k for k Î ¥ such that zD £ zIP. Such a problem is
In what follows, we frequently refer to certain classes of called a dual problem and is a strong dual problem if zD = zIP.
functions, defined below. For any pair (f, V) that comprises a dual problem and any
v Î V, f(v) is a valid lower bound on zIP and the dual
Definition Let a function f be defined over domain V. problem is that of finding the best such bound. The
Then f is usefulness of such a dual problem may be rather limited,
l subadditive if f(v1)  f(v2) ³ f(v1  v2) " v1, v2, v1  v2 Î V. since we require only that it provides a valid bound for the
l linear if V is closed under addition and scalar single MILP instance being analyzed and since the pair f
Guzelsoy and Ralphs: Duality for Mixed-Integer Linear Programs 120
IJOR Vol. 4, No. 3, 118-137 (2007)

and V are not necessarily selected according to any In the remainder of the paper, we refer to the following
criterion for goodness. A number of methods for running example.
producing such dual problems directly from primal input
data are already known and include both the dual of the LP Example Consider the following MILP instance with
relaxation and the Lagrangian dual (see Section 4.5), which right-hand side b:
are generally easy to construct and solve. The bounds
produced by such methods are useful in helping to solve 1
the primal problem, but must necessarily be weak on their z IP = min x 1 + 2x 3 + x 4
2
own, since computing the exact optimal solution value zIP 3
of the MILP (2) is an NP-hard optimization problem in s.t. x 1 - x 2 + x 3 - x 4 = b and (6)
2
general.
x1 , x 2 Î ¢ + , x 3 , x 4 Î ¡ + .
Conceptually, one avenue for improvement in the
bound yielded by (3) is to let the dual problem itself vary
and try to choose the “best” among the possible In this case, the value function (pictured in Figure 1) can
alternatives. This leads to the generalized dual be represented explicitly in the form:

z D = max max f ( v ), (4) ì M M

f ,V v ÎV ï
ï -d - , - < d £ - 3
3 5
ï 2 2 2
where each pair ( f , V ) considered in the above ï 3
maximization is required to comprise a dual problem of the ï 2d + 3, - < d £ -1
form (3). This dual may yield an improved bound, but it is ï 2
ï -d, -1 < d £ 0
not clear how to obtain an optimal solution over any ï
reasonable class of dual problems and even less clear how z (d ) = í 1 (7)
ï 2d , 0<d £
such a solution might help determine the effect of 2
ï
perturbations of the primal problem. ï -d + ,
3 1
<d £1
A second avenue for improvement is to focus not just ï 2 2
on producing a lower bound valid for a single instance, but ï
on constructing a dual function that can produce valid ï 2d - 3 , 1< d £
3
ï 2 2
bounds across a range of instances within a neighborhood ï
î M M
of a given base instance. Such dual functions may be
obtained as a by-product of primal solution algorithms and
are needed for effective post facto analysis. In what follows, By considering what optimal solutions to this simple
we generally refer to any function that takes as input an MILP instance look like as the right-hand side is varied, we
infinite family of MILP instances and returns as output a can get an intuitive feel for why the value function has the
valid lower bound on each instance as a “dual function.” shape that it does in this example.
Such a function is considered strong with respect to a Note that the slope near zero is exactly the objective
given primal problem if the bound returned for that function coefficients of the continuous variables, since
particular instance is zIP. these are the only variables that can have positive value for
Because the right-hand side can be thought of as values of d near zero. Furthermore, the gradient of the
describing the level of resources available within the system function alternates between these two slope values moving
being optimized, it is natural to consider the question of away from zero in both directions, as the continuous
how the optimal solution value of a MILP changes as a variables alternate in the role of ensuring that the fractional
function of the right-hand side. The value function of a MILP part of the left-hand side is consistent with that of d. The
is a function that returns the optimal solution value for any coefficients of the integer variables, on the other hand,
given right-hand side, i.e., it is a function z : ¡ m ® determine the breakpoints between the linear pieces. ■
¡ È {±¥} defined by
Although it is generally difficult to construct the value
function itself, it is much easier to obtain an approximation
z ( d ) = min cx , (5) that bounds the value function from below, i.e., a function
x ÎS ( d )
F: ¡ m ® ¡ that satisfies F(d) £ z(d) for all d Î ¡ m .
Given that we can do this, the question arises exactly how
where S(d) = {x Î ¢ r+ ´ ¡ n+- r | Ax = d} for d Î ¡ m .
to select such a function from among the possible
By convention, we define z(d) = ¥ if d Ï W, where alternatives. A sensible method is to choose one that
W = {d Î ¡ m | S(d) ¹ Ø}. As we discuss below, the value provides the best bound for the current right-hand side b.
function plays a central role in classical duality theory, but This results in the dual
computing it is generally difficult even though it has a
closed form. We consider properties of the value function z D = max{F ( b ) : F ( d ) £ z ( d ), d Î ¡ m , F Î ¡ m }, (8)
and its structure in more detail in Section 4.1.
Guzelsoy and Ralphs: Duality for Mixed-Integer Linear Programs 121
IJOR Vol. 4, No. 3, 118-137 (2007)

where ¡ m Í L m and the infinite family of constraints FLP ( d ) = max{vd : vA £ c , v Î ¡ m }. (9)

ensures that we only consider dual functions that
yield a valid bound for any right-hand side. Let F be defined by F(d) = FLP(d) "d Î W and F(d) = 0
Note that if the primal problem has a finite optimal elsewhere. Then F is feasible to (8) if F Î ¡ m and the
value and ¡ m º L m , (8) always has a solution F* that is a primal problem is bounded, since linear programming
strong dual function by setting F*(d) = z(d) when d Î W, duality tells us that FLP(d) £ z(d) for all d Î W. The
and F*(d) = 0 elsewhere. In this case, it also follows that a following example shows the result of computing this dual
dual function is optimal to (8) if and only if it bounds the function for the MILP instance from Example 1.
value function from below and agrees with the value
function at b. This means that not all optimal solutions to Example 2. Consider the value function of the LP
(8) provide the same bound for a given vector d. In fact, relaxation of problem (6),
there are optimal solutions to this dual that provide
arbitrarily poor estimates of the value function for FLP ( d ) = max vd
right-hand sides d ¹ b, even when d is in a local 1
neighborhood of b. It is thus an open question whether (8) s.t 0 £ v £ , and (10)
2
in fact provides the best criterion for selecting a dual v Î ¡,
function or whether it is possible to compute a dual
function guaranteed to produce “reasonable” bounds
within a specified neighborhood of b.
Consider the value function of the LP relaxation of the
MILP (2), given by

Figure 1. Value function of MILP from Example 1.

Figure 2. Value function of LP relaxation of problem (6).

Guzelsoy and Ralphs: Duality for Mixed-Integer Linear Programs 122
IJOR Vol. 4, No. 3, 118-137 (2007)

which can be written explicitly as (8) as the subadditive dual

ì0, d £ 0 z D = max F ( b )
ï
FLP ( d ) = í 1 . (11) F(a j ) £ c j " j Î I ,
ïî 2 d , d > 0 (14)
F ( a j ) £ c j " j Î C , and
This dual function is shown in Figure 2, along with the F Î Gm ,
value function of the original MILP. In this example, FLP
can be seen to be the best piecewise-linear, convex where a j is the jth column of A and the function F is
function bounding z from below. ■ defined by

By considering that F( d )
F ( d ) = lim sup "d Î ¡ m . (15)
d ®0 +
F ( d ) £ z ( d ), d Î ¡ Û F ( d ) £ cx , x Î S ( d ), d Î ¡
m m

(12)
Û F ( Ax ) £ cx , x Î ¢ r+ ´ ¡ n+-r , Here, F , first used by Gomory and Johnson (1972) in
the context of cutting plane algorithms, is the upper
we see that the dual problem (8) can be rewritten as d-directional derivative of F at zero. The next result reveals the
relation between F and F .
z D = max {F ( b ) : F ( Ax ) £ cx , x Î ¢ r+ ´ ¡ n+-r , F Î ¡ m }. (13)
Theorem (Johnson (1974), Jeroslow (1978), Nemhauser
In the next section, we will use this equivalence to derive and Wolsey (1988)) If F Î Gm, then "d Î ¡ m with
a simpler form of (8) in the case when ¡ m is restricted to F ( d ) < ¥ and ³ 0, F ( d ) £ F ( d ).
a particular class of subadditive functions.
Proof. Let > 0 and > 0. Setting q = - ëê ûú , we have
3. THE SUBADDITIVE DUAL

As stated, the dual (8) is rather general and perhaps only æ dö æ êë úû d q d ö

of theoretical interest. A natural question is whether it is F( d ) = F ç ÷=Fç + ÷
è ø è ø
possible to restrict the class of functions considered in (8)
in some reasonable way. Both linear and convex functions æ dö æq d ö
£ ëê úû F ç ÷+F ç ÷
are natural candidates for consideration. If we take è ø è ø
¡ m º Lm, then (8) reduces to zD = max{vb|vA £ c, v Î ¡ m }, æ dö æq d ö æ dö
= Fç ÷+F ç ÷ - qF ç ÷,
which is the dual of the continuous relaxation of the è ø è ø è ø
original MILP discussed earlier. Hence, this restriction
results in a dual that is no longer guaranteed to produce a
where the inequality follows from the fact that F Î Gm.
strong dual function (see Figure 2). Jeroslow (1979)
Now, letting = m1 , we get
showed that the optimum zD obtained by setting ¡ m º Cm
also results in the same optimal solution value obtained in
the linear case. F( d) æ F(q d ) F( d)ö
F( d ) £ +q ç - ÷. (16)
In a series of papers, Johnson (1973, 1974, 1979) and è q ø
later Jeroslow (1978) developed the idea of restricting ¡ m
to a certain subset of Fm (subadditive functions). The By taking the limit as ® 0+, we obtain
subadditive functions are a superset of the linear functions
that retain the intuitively pleasing property of “no F ( d ) £ F ( d ). (17)
increasing returns to scale” associated with linear functions.
A strong motivation for considering this class of functions Finally, note that
is that the value function itself is subadditive over the
domain W and can always be extended to a subadditive F ( ( d ))
function on all of ¡ m (see Theorem 5). This means that F ( d ) = lim sup
d ®0 +
this restriction does not reduce the strength of the dual (8). (18)
F( d)
To see why the value function is subadditive, let d1, d2 Î W = lim sup = F ( d ).
d ® 0 +
and suppose z(di) = cxi for some xi Î S(di), i = 1, 2. Then,
■
x1 + x2 Î S(d1 + d2) and hence z(d1) + z(d2) = c(x1 + x2) ³
z(d1 + d2). Example 3. Consider the d-directional derivative of the
If ¡m º Gm º {F Î Fm | F(0) = 0}, then we can rewrite value function for the MILP (6), shown in Figure 3:
Guzelsoy and Ralphs: Duality for Mixed-Integer Linear Programs 123
IJOR Vol. 4, No. 3, 118-137 (2007)

Figure 3. Directional derivative of the value function of problem (6).

ì -d , d £ 0 F (0) = 0 and F ( a j x j ) £ F ( a j )x j , x j Î ¡ + , j Î C
z (d ) = í . (19)
î 2d , d > 0 by Theorem 1. Therefore, the second inequality holds. For
the last inequality, F ( a j ) £ c j , j Î I and F ( a j ) £ cj, j Î C
Note that z is a piecewise linear convex function whose
by dual feasibility and xj is nonnegative for all j Î N by
directional derivatives near the origin coincide with that of primal feasibility. ■
z. As we pointed out in Example 1, these directional
derivatives are determined by the coefficients of the Example 4. For the MILP (6), the subadditive dual
continuous variables in (6). ■ problem is

The use of F is required in (14) due to the presence of max F ( b )

the continuous variables and is not needed for pure integer 1
programs. Intuitively, the role of F is to ensure that F (1) £
2
solutions to (14) have gradients that do not exceed those of F ( - 32 ) £ 0
the value function near zero, since the subadditivity of F
alone is not enough to ensure this in the case of MILP. We F (1) £ 2 (20)
now show formally that (14) is in fact a valid dual problem. F ( -1) £ 1
F Î G1 .
Theorem 2. (Weak Duality by Jeroslow (1978, 1979)) Let
x be a feasible solution to the MILP (2) and let F be a
As described above, the last two constraints require that
feasible solution to the subadditive dual (14). Then, F(b) £
the slope of F going away from the origin (the d-directional
cx.
derivative) be less than or equal to that of the value
function, whereas the first two constraints require that F(1)
Proof. Let x be a given feasible solution to the MILP (2).
and F ( - 23 ) not exceed z(1) and z ( - 32 ) , respectively.
Note that
Note that in this example, the constraint F (-1) £ 1 is
F ( b ) = F ( Ax ) actually equivalent to the constraint F(-1) £ 1, but
£ F(å a j x j ) + F( å a j x j ) replacing F (1) £ 2 with F(1) £ 2 results in the admission
j ÎI j ÎC of invalid dual functions.
£ å F ( a )x j + å F ( a j )x j
j If we require integrality of all variables in (6), then the
j ÎI j ÎC value function becomes that shown in Figure 4, defined
£ cx . only at discrete values of the right-hand side d. In this case,
F is replaced by F in (20) and the third constraint becomes
The first inequality follows from the subadditivity of F. redundant. This can be seen by the fact that x3 cannot take
Next, F ( å j ÎI a j x j ) £ å j ÎI F ( a j )x j , since F is on a positive value in any optimal solution to the pure
integer restriction of (6). ■
subadditive, F(0) = 0 and x j Î ¢ + , j Î I. Similarly,
F ( å j ÎC a j x j ) £ å j ÎC F ( a j x j ) £ å j ÎC F ( a j )x j , since Although the value function itself yields an optimal
solution to the subadditive dual of any given MILP,
Guzelsoy and Ralphs: Duality for Mixed-Integer Linear Programs 124
IJOR Vol. 4, No. 3, 118-137 (2007)

irrespective of the value of the original right-hand side b, 2. The dual problem is infeasible if and only if z(0) < 0.
the set of all dual functions that are optimal to (14) can be
affected dramatically by the initial value of b considered. Proof. First, note that 0Î W and z(0) £ 0, since x = 0 is a
This is because F is required to agree with the value feasible solution to the MILP (2) with right-hand side 0.
function only at b and nowhere else. In the following 1. If b Î W and z(0) < 0, then there exist x Î S and
example, we consider optimal solutions to (20) for
different values of b. x̂ Î S(0) with cxˆ < 0. Then x + xˆ Î S for all
Î ¡ + and it follows that can be chosen to make
Example 5. Consider optimal solutions to (14) for the z(b) arbitrarily small. Conversely, if b Î W and z(0) = 0,
MILP (6) for different values of b. then we must also have that min{cx | Ax = 0, x Î ¡ n+ }
= 0. Otherwise, there must exist an xˆ Î ¤n+ for
1. F1(d) = d2 is an optimal dual function for b Î {0, 1,
2, ...} (see Figure 2), which Axˆ = 0 and cxˆ < 0, which can be scaled to
2. F2(d) = 0 is an optimal dual function for b Î { ..., -3, yield an integer solution to (2) with right-hand side 0,
contradicting the assumption that z(0) = 0. Since no
- 32 , 0} (see Figure 2).
such vector exists, the LP relaxation of (2), and hence
3. F3 ( d ) = max { éêd -
1
2
ù , 2d - 3 éd - êé éêd ùú - d úù ù is
ú
êé éêd ùú - d úù
2
ê
4 4
ú } 2.
the MILP itself, must be bounded.
If z(0) = 0, then min{cx|Ax = 0, x Î ¡ n+ } =
an optimal dual function for b Î {[0, 41 ] È [1, 54 ] È
max {v 0: vA £ c, v Î ¡ m } = 0 (see the proof of part 1
[2, 94 ] È ...} (see Figure 5).
above) and therefore, (14) is feasible by setting F(d) =
4. F4(d) = max { 32 éê 23d - 23 éê êé 23d úù - 23d ùú ùú - d , v*d "d Î ¡ m , where v* is the optimal dual solution.
This implies that if the dual is infeasible, then z(0) < 0.
- 43 êé 23d - 23 êé éê 23d ùú - 23d úù úù + d2 } is an optimal dual
If z(0) < 0, on the other hand, the dual cannot be
function for b Î { ¼È [ - 72 ,-3] È [-2, - 32 ] È [- 12 , feasible since any feasible solution F has to satisfy F(0)
0]} (see Figure 5). ■ = 0 and this would contradict weak duality. ■

As in LP duality, weak duality yields results concerning Corollary 4. For the MILP (2) and its subadditive dual
the relationship between primal and dual when no finite (14),
optimum exists. Before proving the main corollary, we
need the following important lemma. 1. If the primal problem (resp., the dual) is unbounded
then the dual problem (resp., the primal) is infeasible.
Lemma 3. For the MILP (2) and its subadditive dual (14), 2. If the primal problem (resp., the dual) is infeasible,
the following statements hold: then the dual problem (resp., the primal) is infeasible
or unbounded.
1. The primal problem is unbounded if and only if b Î W
and z(0) < 0.

Figure 4. Value function of problem (6) with all variables integer.

Guzelsoy and Ralphs: Duality for Mixed-Integer Linear Programs 125
IJOR Vol. 4, No. 3, 118-137 (2007)

Figure 5. Observe that F(d) = max{F3(d), F4(d)} is an optimal dual function for (20) for some values of b and only feasible
otherwise.

Proof. We will not give the proof here, but note that the idea
depends on iteratively adding columns orthogonal to the
1. This part follows directly from Theorem 2. span of the columns of A with objective function
2. Assume that the primal problem is infeasible. Then coefficients chosen so that ze(d) = z(d) whenever z(d) < ¥.
there are two cases. If z(0) < 0, the dual is infeasible by The following result then shows formally that the dual (14)
Lemma 3. Otherwise, by LP duality, $ v Î ¡ m with vA is strong.
£ c. Let F1 Î Gm be defined by F1(d) = vd, "d Î ¡ m .
Theorem 6. (Strong duality by Jeroslow (1978, 1979),
Note that F 1 = F1. Next, consider the value function
Wolsey (1981)) If the primal problem (2) (resp., the dual)
F2(d) = min{x0: Ax + x0d = d, x Î ¢ r+ ´ ¡ n+- r , has a finite optimum, then so does the dual problem (14)
x0 Î ¢ + }. F2 is defined and finite for all d Î ¡ m since (resp., the primal) and they are equal.
x = 0 with x0 = 1 is a feasible solution for any
right-hand side. Therefore, F2 Î Gm. Furthermore, for Poof. Note that if the primal or the dual has a finite
optimum, then Corollary 4 requires the other also to have a
any j Î I, F2 ( a j ) £ 0, since e j (the jth unit vector)
finite optimum. Now, we prove the claim by verifying that
together with x0 = 0 is a feasible solution to the the value function z (or an extension to z) is a feasible dual
corresponding problem. On the other hand, for any function whenever the primal has a finite optimum.
j Î C and > 0, F2( a j ) £ 0 due to the fact that
x = e j and x0 = 0 is feasible. Thus, F 2 ( a j ) £ 0, i. W º ¡ m : In this case, z Î G m , and with a similar
" j Î C. In addition, F2(b) = 1 since there cannot be argument in the second part of the proof of Corollary
an optimal solution with x0 = 0 as a consequence of 4, z is feasible to the dual problem.
S = Ø. Then, observe that for any scalar Î ¡ + , F1 + ii. W Ì ¡ m : By Theorem 5, $ ze Î G m with ze(d) = z(d)
F2 is dual feasible to (14), which means that the dual "d Î W and ze(d) <¥ "d Î ¡ m . By construction, ze
is unbounded as can be chosen arbitrarily large. must satisfy the constraints of the dual of the original
If the dual problem is infeasible, then, by Lemma 3, MILP (2), since the dual of the extended MILP from
z(0) < 0 and the primal problem is unbounded if Theorem 5 includes the constraints of (14) (Ie Ê I and
b Î W and infeasible otherwise. ■ Ne Ê N). Therefore, ze is feasible to the dual of the
original MILP and hence, this dual has an optimal
Before moving on to prove strong duality, we need the solution value of ze(b) = z(b). ■
following theorem that states that any given MILP can be One can further use the strong duality property of (14)
“extended” to one that is feasible for all right-hand sides to derive a generalization of Farkas’ Lemma. This result is
and whose value function agrees with that of the original stated more formally in the following corollary.
MILP for all right-hand sides d Î W.
Corollary 7. (Blair and Jeroslow (1982)) For the MILP (2),
Theorem 5. (Blair and Jeroslow (1977)) For the MILP (2), exactly one of the following holds:
there exists an extension ze(d) = min {cex: Aex = d, 1. S ¹ Ø.
x Î ¢ l+ ´ ¡ k+- l }, with ce and Ae obtained by adding new 2. There is an F Î G m with F( a j ) £ 0 "j Î I, F ( a j ) £
coefficients and columns to c and A, such that ze(d) = z(d) 0 "j Î C and F(b) > 0.
" d Î W and ze(d) < ¥ " d Î ¡ m .
Guzelsoy and Ralphs: Duality for Mixed-Integer Linear Programs 126
IJOR Vol. 4, No. 3, 118-137 (2007)

Proof. The proof follows directly from applying Corollary å F(a j

)x j + å F ( a j )x j ³ F ( d ) (23)
4 and Theorem 6 to the MILP (2) with c = 0. ■ j ÎI j ÎC

The subadditive dual (14) can also be used to extend is satisfied for all x Î S(d). The proof follows the same
familiar concepts such as reduced costs and the complementary
steps as that of weak duality, with x restricted to be in S(d).
slackness conditions to MILPs. For a given optimal solution
F* to (14), the reduced costs can be defined as cj - F*( a j ) Furthermore, the following result shows that any valid
* inequality is either equivalent to or dominated by an
for j Î I and cj - F ( a j ) for j Î C. These reduced costs inequality in the form of (23).
have an interpretation similar to that in the LP case, except
that we do not have the same concept of “sensitivity Theorem 9. (Johnson (1973), Jeroslow (1978)) For the
ranges” within which the computed bounds are exact. MILP (2) and p Î ¡ n , p0 Î ¡ , we have that px ³ p0
Complementary slackness conditions can be stated as "x Î S if and only if there is an Fp Î G m such that
follows.

Theorem (Jeroslow (1978), Johnson (1979), Bachem Fp ( a j ) £ p j "j Î I ,

and Schrader (1980), Wolsey (1981)) For a given right-hand F p ( a j ) £ p j "j Î C , and (24)
side b, let x* and F* be feasible solutions to the primal
Fp ( b ) ³ p 0 .
problem (2) and the subadditive dual problem (14). Then,
x* and F* are optimal if and only if
Proof. First assume that p Î ¡ n and p0 Î ¡ are given
x *j ( c j - F * ( a j )) = 0, "j Î I , such that px ³ p0 "x Î S. Consider the MILP
*
x *j ( c j - F ( a j )) = 0, "j Î C , and (21)
z p = min{p x | x Î S}. (25)
F * ( b ) = å F * ( a j )x *j + å F ( a j )x *j
*

j ÎI j ÎC
Clearly, zp ³ p0 because otherwise, there exists an
x Î S with p x < p0. Applying Theorem 6 to (25), we find
Proof. If x* and F* are optimal, then, from the properties
of F* and strong duality, that there must be a dual feasible function Fp satisfying
(24).
Conversely, assume that there exists an Fp Î G m
F * ( b ) = F * ( Ax * ) = å F * ( a j )x *j + å F ( a j )x *j = cx * . (22)
*

j ÎI j ÎC
satisfying (24) for a given p Î ¡ n and p0 Î ¡ . Then Fp is
also feasible to the subadditive dual of (25) and from weak
Then, we have duality, px ³ Fp(b) ³ p0 for all x Î S. ■

åx ( c j - F * ( a j )) + å x *j ( c j - F ( a j )) = 0. Example 6. The subadditive dual function F3(d) in

* *
j
j ÎI j ÎC Example 5 is feasible to (20). Since F3 (1) = 21 ,
F3 ( - 23 ) = - 12 , F 3 (1) = 2, F 3 ( -1) = 1, then
By primal and dual feasibility, x ³ 0 " j ÎN, *
j

cj - F * ( a j ) ³ 0 " j Î I and cj - F ( a j ) ³ 0 " j ÎC.

* x1 x 2
- + 2x 3 + x 4 ³ F3 ( b ) (26)
Therefore, (21) holds. 2 2
On the other hand, if the conditions (21) are satisfied,
is a valid inequality for (6). ■
then (22) holds, which in turn, yields F*(b) = cx*. ■
As an extension to this theorem, Bachem and Schrader
These conditions, if satisfied, yield a certificate of
optimality for a given primal-dual pair of feasible solutions. (1980) showed that the convex hull of S can be
They can further be used to develop an integer represented using only subadditive functions and that
programming analog of the well-known primal-dual rationality of A is enough to ensure the existence of such a
algorithm for linear programming. Llewellyn and Ryan representation, even if the convex hull is unbounded.
(1993) give the details of one such algorithm.
The subadditive duality framework also allows the use of Theorem (Jeroslow (1978), Blair (1978), Bachem and
subadditive functions to obtain inequalities valid for the Schrader (1980)) For any d Î W,
convex hull of S. In fact, subadditive functions alone can,
conv ( S ( d )) = {x : å F ( a j )x j
in theory, yield a complete description of conv(S ). It is j ÎI
(27)
easy to see that for any d Î W and F Î G m with + å F ( a j )x j ³ F ( d ), F Î G m , x ³ 0}.
*
F ( a j ) < ¥ "j Î C, the inequality j ÎC
Guzelsoy and Ralphs: Duality for Mixed-Integer Linear Programs 127
IJOR Vol. 4, No. 3, 118-137 (2007)

For a fixed right-hand side, it is clear that only finitely 4. CONSTRUCTING DUAL FUNCTIONS
many subadditive functions are needed to obtain a
complete description, since every rational polyhedron has It is reasonable to conclude that neither the general dual
finitely many facets. In fact, Wolsey (1979) showed that for problem (8) nor the subadditive dual problem (14) can be
PILPs, there exists a finite representation that is valid for formulated as manageable mathematical program solvable
all right-hand sides. directly using current technology.
However, there are a number of methods we can use to
Theorem (Wolsey (1979)) For a PILP in the form (2), obtain feasible (and in some cases optimal) dual functions
there exist finitely many subadditive functions Fi, i = 1, ..., indirectly. We focus here on dual functions that provide
k, such that bounds for instances of a given MILP after modifying the
right-hand side, since these are the ones about which we
n know the most. Such dual functions are intuitive because
conv ( S ( d )) = {x : å Fi ( a j )x j ³ Fi ( d ), they allow us to extend traditional notions of duality from
j =1 (28) the realm of linear programming to that of integer
i = 1, ..., k , x ³ 0} programming. However, we emphasize that they are not
the only dual functions of potential interest in practice.
for any d Î W. Dual functions that accommodate changes to the objective
function are also of interest in many applications,
Proof. Wolsey (1979) showed that when A Î ¢ m ´n , there particularly decomposition algorithms. Dual functions that
exists finitely many subadditive functions Fi, i = 1, ..., k, allow changes to the constraint matrix are closely related to
such that those for studying the right-hand side, but may also yield
further insight. Dual functions of these latter two types
n have not been well-studied. We discuss future research on
conv ( S ( d )) = {x : Ax = d , å F (a
j =1
i
j
)x j ³ Fi ( d ),
(29)
these dual functions in Section 5.
Dual functions of the right-hand side can be grouped
i = 1, ..., k , x ³ 0} "d Î ¢ . m
into three categories: (1) those constructed explicitly in
closed form using a finite procedure, (2) those obtained
However, the assumption that A Î ¢ m ´n is without loss from known families of relaxations, and (3) those obtained
of generality, since A can be scaled appropriately. After as a by-product of a primal solution algorithm, such as
branch and cut. In Sections 4.1 and 4.2 below, we discuss
scaling, we must have W Í ¢ m and the result follows. ■
two different methods of explicitly constructing the value
function of a PILP and give an idea of how those might be
Finally, it is possible to show not only that any facet can
extended to the MILP case. In Sections 4.3 through 4.5, we
be generated by a subadditive function, but that this is true
discuss methods for obtaining dual functions from
of any minimal inequality. Recall that p Î ¡ m and p0 Î ¡ relaxations. In Section 4.6, we discuss a method by which
define a minimal valid inequality if there is no other valid the subadditive dual of a bounded PILP can be formulated
inequality defined by p̂ Î ¡ m and p̂ 0 Î ¡ such that as a linear program. Finally, in Section 4.7, we discuss how
pˆ j £ pj for all j Î N and p̂ 0 ³ p0. Although the next to obtain a dual function as a by-product of the
theorem was originally stated for either rational constraint branch-and-cut algorithm, the method used most
matrices (Johnson (1974), Blair (1978)) or bounded feasible commonly in practice for solving MILPs.
regions (Jeroslow (1979)), Bachem and Schrader (1980)
showed that the same result holds without any restriction 4. The value function
on S. The value function itself is the most useful dual function
we can obtain for studying the effect of perturbations of
Theorem (Bachem and Schrader (1980)) If p Î ¡ m the right-hand side vector, since it provides an exact
and p0 Î ¡ define a minimal valid inequality for the MILP solution value for any right-hand side vector.
(2), then there is an F Î G m such that Unfortunately, it is unlikely that there exist effective
methods for producing the value function for general
F ( a j ) = p j = F (b ) - F (b - a j ) "j Î I , MILPs. For PILPs, Blair and Jeroslow (1982) showed that
a procedure similar to Gomory’s cutting plane procedure
F ( b )- F ( b - a )j
F ( a j ) = p j = lim  ® 0+  "j Î C and, can be used to construct the value function in a finite
F (b ) = p 0 . (30) number of steps. Unfortunately, the representation so
produced may have exponential size. From this procedure,
however, they were able to characterize the class of
The converse of Theorem 12 holds for any subadditive
functions to which value functions belong, namely, Gomory
function that is the value function for the MILP (2) with
functions, a subset of a more general class called Chvátal
objective function p, where p0 = min{px | x Î S} (as in functions.
(25)).
Guzelsoy and Ralphs: Duaity for Mixed-Integer Linear Programs 128
IJOR Vol. 4, No. 3, 118-137 (2007)

Definition 3. Chvátal functions are the smallest set of = 32 max{ êé 23d úù , éêd úù } - d " d Î W, which is a Gomory
functions C m such that function (see Figure 4). ■
1. If h Î Lm, then h Î C m.
2. If h1, h2 Î C m and , b Î ¤ + , then h1 + bh2 Î C m. For PILPs, it is also worth mentioning that there always
3. If h Î C m,then êéh úù Î C m. exists an optimal solution to the subadditive dual problem
(14) that is a Chvátal function.
Gomory functions are the smallest set of functions
G m Í C m with the additional property that Theorem (Blair and Jeroslow (1982)) For a PILP in the
4. If h1, h2 Î G m, then max{h1, h2} Î G m. form (2), if b Î W and z(b) > -¥, then there exists h Î C m
that is optimal to the subadditive dual (14).
The relationship between C m and G m is evident from the
following theorem. Proof. Note from the proof of Theorem 6 that either the
value function itself, or an extension of the value function
Theorem (Blair and Jeroslow (1982)) Every Gomory is a feasible solution to the subadditive dual. Denote this
function can be written as the maximum of finitely many function as ze. From Theorem 15, we know that there is a
Chvátal functions, that is, if g Î G m, then there exist g Î G m with g(d) = ze(d) for all d Î W and hence, feasible to
hi Î C m for i = 1, ..., k such that the subadditive dual (14). By Theorem 13, g is the
maximum of finitely many Chvátal functions, h1, ..., hk. For
g = max{h1 , ..., hk }. (31) right-hand side b, since ze(b) = max{h1(b), ..., hk(b)}, there
exists l Î {1, ..., k} with z(b) = ze(b) = hl(b). Then hl is an
This theorem also makes evident the relationship optimal solution to the subadditive dual (14) since it is
between G m and the property of subadditivity. Note that if subadditive, and hl( a j ) £ g( a j ) £ cj for all j Î I. ■
h1, h2 are subadditive and , b Î ¤ + , then it is easy to
show that the functions h1 + bh2 and éêh1 ùú are both Using a result similar to Corollary 7 above, Blair and
subadditive. Consequently, one can show that Chvátal Jeroslow (1982) introduced the more general concept of a
functions are subadditive by induction on the rank of consistency tester to detect the infeasibility of the problem for
functions (i.e., the number of operations of the type any right-hand side. They showed that for a given PILP,
specified in Definition 3 needed to derive a given Chvátal there is a g Î G m such that for every d Î ¡ m , g(d) £ 0 if
function from the base class Lm). Since max{h1, h2} is and only if d Î W. Using the consistency tester concept, we
can state a converse of Theorem 15. That is, for Gomory
subadditive, Gomory functions are also subadditive. As a
result of subadditivity, both Chvátal and Gomory functions functions g1, g2, there exist A , c such that g1(d) =
can be used for generating valid inequalities. The following min{ c x| Ax = d, x ³ 0 and integral} for all d with g2(d) £
lemma, needed for the proof of Theorem 15 shows that 0. In this sense, there is one-to-one correspondence
for PILPs, Chvátal functions can be used to obtain a between PILP instances and Gomory functions.
description of the convex hull of solutions. For MILPs, neither Theorem 15 nor its converse holds.
However, Blair and Jeroslow (1984) argue that the value
Lemma (Schrijver (1980)) The subadditive functions in function z can still be represented by a Gomory function if
Theorem 11 can be taken to be Chvátal functions. cj = 0 " j Î C or can be written as a minimum of finitely
many Gomory functions. A deeper result is contained in
The above lemma then allows us to characterize the the subsequent work of Blair (1995), who showed that the
value function of a PILP for which z(0) = 0. value function of a MILP can be written as a Jeroslow formula,
consisting of a Gomory function and a correction term.
Theorem (Blair and Jeroslow (1982)) For a PILP in the Here, rather than the formula itself (see Blair and Jeroslow
form (2), if z(0) = 0, then there is a g Î G m such that g(d) = (1984), Blair (1995) for details), we present a simplified
z(d) for all d Î W. version to illustrate its structure.
For a given d Î W, let the set E consist of the index sets
Proof. Consider the parameterized family of PILPs of dual feasible bases of the linear program
min{cx|x Î conv(S(d))} " d Î W, where conv(S(d)) is
represented by the finite set of Chvátal functions whose min{c C x C : AC x C = d , x ³ 0}. (32)
existence is guaranteed by Lemma 14. Applying LP duality,
we get g(d) = z(d) = max{ å i =1 v i Fi ( d ) |v Î V } where V is
k By the rationality of A, we can choose M Î ¢ + such that
the finite set of dual basic feasible solutions. Then the for any E Î E, MAE-1a j Î ¢ m for all j Î I. For E Î E, let
proof is complete by Theorem 13. ■ vE be the corresponding basic feasible solution to the dual
of
Example 7. The value function of problem (6) with all
variables assumed to be integer can be written as z(d)
Guzelsoy and Ralphs:  for Mixed-Integer Linear Programs 129
IJOR Vol. 4, No. 3, 118-137 (2007)

ì1 1 ü is a feasible solution to (35) with right-hand side AI x I +

min í c C x C : AC x C = d , x ³ 0ý, (33)
îM M þ ëê AC x C ûú E and the last equality follows from the fact that

which is a scaled version of (32). Finally, for a right-hand v E ( d - ëêd ûú E ) = v E ( AC x C - ëê AC x C ûú E )

side d and E Î E, let ëêd ûú E = AE êë AE-1d úû .
= c E AE-1 ( AC x C - êë AC x C úû E ).

Theorem (Blair (1995)) For the MILP (2), there is a

g Î G m such that On the other hand, for any E Î E,

z ( d ) = min g ( ëêd ûú E ) + v E ( d - ëêd ûú E ) (34) g ( ëêd ûú E ) + v E ( d - ëêd ûú E ) = z JF ( ëêd ûú E ) + v E ( d - ëêd ûú E )

EÎE
³ z ( êëd úû E ) + z ( d - êëd úû E ) ³ z ( d ).
for any d Î W.
by the subadditivity of z. ■
Proof. Assume that cC and AC are scaled as in (33) and
consider the PILP instance Example . Consider the MILP (6). With M = 2, the set of
index sets of dual feasible bases of min{x3 + 12 x4| 12 x3
z JF ( ) = min cx + z (j ) y (35) - 12 x4 = d, x3, x4 ³ 0} is E = {{3}, {4}}. Furthermore, v{3}
s.t Ax + j y = = 2 and v{4} = -1. Since j = 12 - 12 = 0 and z(0) = 0, (35)
x Î ¢ , y Î ¢+ ,
n
+
reduces to zJF ( ) = { 12 x1 + x3 + 12 x4|x1 - 32 x2 + 12 x3 - 12 x4
= , xi Î ¢ + , i = 1, ..., 4}. The value function of this
where j = -åj ÎC a . Then we have the following:
j
problem is the same as z in Example 7. Thus, g(d)
= 32 max{ éê 23d ùú , êéd úù } - d solves this problem. Then the
1. For any E Î E and d Î ¡ m , (35) is feasible for value function (see Figure 1) of (6) is
= ëêd ûú E . To see this, observe that if êë AE d úû ³ 0, then
xE = êë AE-1d úû , xN\E = 0, y = 0 is a feasible solution. ïì 3 ïì é ê 2d ú ù é ê 2d ú ù ïü 3 é 2d ù
min í max í ê ë û ú , ê ë û ú ý + ê ú + 2d ,
Otherwise, there exists  Î ¢ + such that xE = ïî 2 ïî ê 3 ú ê 2 ú ïþ 2

( êë AE-1d úû  AE-1 åjÎE a j ) Î ¢ m+ , since AE-1 SjÎE ïì é é 2d ù ù é é 2d ù ù üï ïü

max í ê ê ú ú , ê ê ú ú ý - d ý .
3
a j = m. Therefore, together with xE, xI = 0, xj =  for 2 îï ê 3 ú ê 2 ú þï ïþ
j Î C\E and y =  is a feasible solution. ■
2. For any E Î E and d Î ¡ m , zJF ( êëd úû E ) ³ z( êëd úû E ).
4.2 Generating functions
To see this, assume that z(j ) = cx1 and zJF ( ëêd ûú E ) =
cx2 + z(j ) ŷ = c(x2 + x 1 ˆy ). But then, clearly, zJF ( ëêd ûú E ) Lasserre (2004a, b) recently introduced a different
method for constructing the value function of PILPs that
³ z( ëêd ûú E ) since (x2 + x 1 ˆy ) Î S( ëêd ûú E ).
utilizes generating functions. This methodology does not
fit well into a traditional duality framework, but
Now, we know from Theorem 15 that there is a g Î G m nevertheless gives some perspective about the role of basic
with g( ëêd ûú E ) = zJF ( ëêd ûú E ) for all d Î ¡ m , E Î E. Let feasible solutions of the LP relaxation in determining the
optimal solution of a PILP.
d Î W be given and x Î S(d). By LP duality, there is an
E Î E with c E AE-1 AC x C £ c C x C . Noting that êëd ûú E = Theorem (Lasserre (2003)) For a PILP in the form (2)
m ´n
ëê Ax ûú E = AI x I + ëê AC x C ûú E , we have with A Î ¢ , define

g ( ëêd ûú E ) + v E ( d - ëêd ûú E ) z ( d , c ) = min cx , (36)

x ÎS ( d )

= z JF ( AI x I + êë AC x C úû E ) + v E ( d - êëd úû E ) and let the corresponding summation function be

-1
£ c I x I + c E A êë AC x C úû E + v E ( d - ëêd ûú E )
E
-1
zˆ( d , c ) = å e cx "d Î ¢ m . (37)
= c I x I + c E A AC x C
E
x ÎS ( d )

£ cx,
Then the relationship between z and ẑ is
where the first inequality follows from the fact that xI
= x I , xj = 0 for j Î C\E, xE = AE-1 êë AC x C úû E , and y = 0
Guzelsoy and Ralphs:  for Mixed-Integer Linear Programs 130
IJOR Vol. 4, No. 3, 118-137 (2007)

1 4.3 Cutting plane method

e z ( d ,c ) = lim zˆ( d , qc ) q
or equivalently,
q ®-¥
(38) Cutting plane algorithms are a broad class of methods for
1
z ( d , c ) = lim ln zˆ( d , qc ). obtaining lower bounds on the optimal solution value of a
q ®-¥ q
given MILP by iteratively generating inequalities valid for
the convex hull of S (called cutting planes or cuts). The
In order to get a closed form representation of ẑ , we
procedure works by constructing progressively tighter
can solve the two sided ¢- transform Fˆ : £ m ® £ polyhedral approximations of conv(S ), over which a linear
defined by
program is then solved to obtain a bound. To be more
precise, in iteration k, the algorithm solves the following
Fˆ ( s , c ) = ås m
-d
zˆ( d , c ) (39) linear program:
d ΢

min cx
with sd = s 1d 1 ... s md m for d Î ¢ m . Substituting ẑ in this
s.t. Ax = b
formula, we get (42)
Õ x ³ Õ0
x ³ 0,
Fˆ ( s , c ) = å å s - d e cx
d ΢ x ÎS ( d )
m

where P Î ¡ k ´n and P0 Î ¡ k represents the cutting

= åe ås cx

d = Ax
-d

planes generated so far. At the time of generation, each of

x ΢n+
(40)
= åe cx - Ax
s the valid inequalities is constructed so as to eliminate a
portion of the feasible region of the current relaxation that
x ΢n+
n
1 contains the current solution to (42), but does not contain
=Õ j , any feasible solutions to the original MILP.
1 - e s -a
cj
j =1
As noted earlier, the LP dual of (42), i.e.,

where the last equality is obtained by applying Barvinok max vb + w Õ 0

(1993)’s short form equation for summation problems over
vA + w Õ £ c (43)
a domain of all non-negative integral points. The formula
j c v Ρ , w Ρ ,
m k
(40) is well-defined if s a > e j , j = 1, ..., n and the +

function ẑ is then obtained by solving the inverse problem is also a dual problem for the original MILP, but does not
yield a dual function directly because the cutting planes
1 generated may only be valid for the convex hull of
(2i p )m ò s =g
m
zˆ( d , c ) = Fˆ ( s , c )s d -1 ds solutions to the original MILP and not for instances with a
(41) modified right-hand side. However, one can extract such a
1
m ò s =g
...ò
m
= Fˆ ( s , c )s d -1 ds , dual function if it is possible to systematically modify each
(2i p ) 1 1 s m = g m
cut to ensure validity after replacing the original right-hand
side b with a new right-hand side d. Assuming that a
where is a vector satisfying aj
>e
cj
j = 1, ..., n and m = subadditive representation (23) of each cut is known, the ith
cut can be expressed parametrically as a function of the
(1, ..., 1) Î ¡ . m
right-hand side d Î ¡ m in the form
Although it is possible to solve (41) directly by Cauchy
residue techniques, the complex poles make it difficult.
One alternative is to apply Brion and Vergne’s (see Brion åF (j ÎI
i i ( a j ))x j + å F i ( i ( a j ))x j ³ Fi ( i ( d )),
j ÎC
(44)
and Vergne (1997), Lasserre (2003) for details) lattice
points counting formula in a polyhedron to get the reduced
where Fi is the subadditive function representing the cut,
form, which, for each d Î ¡ m , is composed of the optimal
solution value of the LP relaxation and a correction term. and the functions i, i : ¡ m ® ¡ m + i -1 are defined by
The correction term is the minimum of the sum of the • 1(d) = 1 (d) = d,
reduced costs of certain nonbasic variables over all basic • i(d) = [ d F1( 1(d)) ... Fi-1( i-1(d)) ] for i ³ 2, and
feasible solutions, obtained by the degree sum of certain • i (d) = [ d F 1 ( 1 (d)) ... F i -1 ( i -1 (d)) ] for i ³ 2.
real-valued univariate polynomials. Another approach
Furthermore, if (vk, wk) is a feasible solution to (43) in
using generating functions is to apply Barvinok (1994)’s
the kth iteration, then the function
algorithm for counting lattice points in a polyhedron of
fixed dimension to a specially constructed polyhedron that k
includes for any right-hand side the corresponding minimal FCP ( d ) = v k d + å w ik Fi ( i ( d )) (45)
test set (see Loera et al. (2004a, b) for details). i =1
Guzelsoy and Ralphs: ality for Mixed-Integer Linear Programs 131
IJOR Vol. 4, No. 3, 118-137 (2007)

is a feasible solution to the subadditive dual problem (14). for all x Î {0, 1}n satisfying px £ p 0 , where p 0 is the
As noted earlier, Wolsey (1981) showed how to modified right-hand side.
construct a dual function optimal to the subadditive dual
for a given PILP using the Gomory fractional cutting plane In the same paper, it is further discussed that a similar
algorithm under the assumption that cuts are generated construction can also be obtained for lifted cover
using a method guaranteed to yield a sequence of LPs with inequalities where some of the coefficients of the left side
lexicographically increasing solution vectors (this method is of the cover inequality are increased to strengthen the
needed to guarantee termination of the algorithm in a finite inequality.
number of steps with either an optimal solution or a proof
that original problem is infeasible). In Gomory’s procedure, 4.4 Corrected linear dual functions
the subadditive function Fi, generated for iteration i, has
the following form A natural way in which to account for the fact that linear
functions are not sufficient to yield strong dual functions in
i -1
ém ù the case of MILPs is to consider dual functions that consist
Fi ( d ) = ê å ki -1d k + å i -1
m +k Fk ( d )ú of a linear term (as in the LP case) and a correction term
ê k =1 k =1 ú (46)
accounting for the duality gap. One way to construct such
where i -1 = ( 1i -1 , ..., i -1
m + i -1 ) ³ 0. a function is to consider the well-known group relaxation.
Let B be the index set of the columns of a dual feasible
Assuming that b Î W, z(b) > -¥, and that the algorithm basis for the LP relaxation of a PILP and denote by N\B
terminates after k iterations, the function FG defined by the index set of the remaining columns. Consider the
function FB defined as
k
FG ( d ) = v k d + å w ik Fi ( d ) (47) FB ( d ) = min c B x B + c N \ B x N \ B
i =1
s.t. AB x B + AN \ B x N \ B = d (49)
is optimal to the subadditive dual problem (14). Note that x B Î ¢ , x N \B Î ¢
m n -m
.
+
FG is a Chvátal function and hence, this can be seen as an
alternative proof for Theorem 16.
In practice, it is generally not computationally feasible to Substituting xB = AB-1 d - AB-1 AN \ B x N \ B in the objective
determine a subadditive representation for each cut added function, we obtain the group relaxation (Gomory (1969))
to the LP relaxation. However, since our goal is simply to
ensure the validity of each cut after modification of the FB ( d ) = c B AB-1d - max c N \ B x N \ B
right-hand side, an alternative approach that is feasible for
AB x B + AN \ B x N \ B = d , (50)
some classes of valid inequalities is simply to track the
n-m
dependency of each cut on the original right-hand side in x B Î ¢ , x N \B Î ¢
m
+
some other way. If this information can be functionally
encoded, as it is with the subadditive representation, the where c N / B = ( c B AB-1 AN / B - c N / B ). Here, dual feasibility
right-hand side of each cut can be modified to make it
valid for new instances and these functions can be used to of the basis AB is required to ensure that c N / B £ 0.
obtain a dual function similar in form to (45). As an FB is feasible to the subadditive dual (14). To see this,
example of this, Schrage andWolsey (1985) showed how to note that FB is subadditive since it is the sum of a linear
construct a function tracking dependency on the function and the value function of a PILP. Also, we have
right-hand side for cover inequalities by expressing the FB( a j ) £ c B AB-1a j - ( c B AB-1a j - cj) = cj, j Î N\B and
right-hand side of a cut of this type as an explicit function FB( a j ) = cB AB-1a j = cj, j Î B. Therefore, for the PILP (2),
of the right-hand side of the original knapsack constraint.
FB(b) £ z(b). Gomory (1969) further discusses sufficient
To illustrate, suppose that p Î ¡ n and p0 Î ¡ is such that
conditions for FB to be strong. Observe that FB(b) = z(b)
p ³ 0 and p0 ³ 0. We define U Í {1, ..., n} to be a cover if
when there exists an optimal solution to (50) with xB ³ 0.
åjÎUpj > p0. It is then well-known that åjÎUxj £ |U| - 1
Another way to construct an optimal solution to the
for all x Î {0, 1}n satisfying px £ p0. The following
subadditive dual using a linear function with a correction
proposition shows how to modify the given inequality so
term is given by Klabjan (2002).
that it remains valid if p0 is changed to p Î ¡.
Theorem (Klabjan (2002)) For a PILP in the form (2),
Theorem . (Schrage and Wolsey (1985)) Let pv =
and a given vector v Î ¡ m , define the function Fv as
max{pj| j Î U} for a given knapsack constraint with
nonnegative parameters and a cover U. Then,
{
Fv ( d ) = vd - max ( vADv - c Dv )x | ADv x £ d , x Î ¢ +Dv , }
ê å j ÎU p j - p 0 úú
åx
j ÎU
j £ êU -
êë pv úû
(48) where Dv = {i Î I : va i > ci}. Then, Fv is a feasible
Guzelsoy and Ralphs: ality for Mixed-Integer Linear Programs 132
IJOR Vol. 4, No. 3, 118-137 (2007)

solution to the subadditive dual problem (14) and Lagrangian relaxation. Generally, this relaxation is
furthermore, if b Î W and z(b) > -¥, there exists a v constructed so that it is much easier to solve than the
Î ¡ m such that Fv(b) = z(b). original MILP, in which case a dual problem can be
constructed as follows. Suppose for a given d Î ¡ m that the
Proof. For a given v, Fv is subadditive using an argument inequalities defined by matrix A and right-hand side d are
similar to that made above for group relaxation problems. partitioned into two subsets defined by matrices A1 and A2
Now, consider the problem max{ ( vADv - c Dv )x | ADv x £ and right-hand sides d 1 and d 2 . Furthermore, let
a i , x Î ¢ +Dv } for a given i. If i Î I\Dv, x = 0 is feasible. SLD( d 2 ) = {x Î ¢ r+ ´ ¡ n+- r : A2x = d 2 }. Then, for a given
Otherwise the ith unit vector is a feasible solution. Thus, for penalty multiplier v Î ¡ m - l , the corresponding Lagrangian
any i Î I, Fv( a i ) £ ci. Therefore, Fv is a feasible solution to relaxation can be formulated as
the subadditive dual (14) and Fv(b) £ z(b).
Next, suppose that the original PILP is solved with L ( d , v ) = min 2 cx + v ( d 1 - A1 x ) (52)
x ÎS LD ( d )
Gomory’s procedure (42) after k iterations. Let the set of
generated Chvátal inequalities be represented by ( p j , p 0j ) Assuming z(0) = 0 and that x*(d) is an optimal solution
for j Î J = {1, ..., k}. Let vk and wk be the corresponding to the original MILP with right-hand side d, we have L(d, v)
components of the optimal dual solution with respect to £ cx* (d) + v(d1 - A1x*(d)) = cx* (d) = z(d) " v Î ¡ m - l . Thus,
the set of original constraints and the set of valid the Lagrangian function defined by
Dk
inequalities. With x Î {x Î ¢ + v | AD k x = b},
v
L D ( d ) = max{L ( d , v ) : v Î V }, (53)

( v k AD k - c D k )x £
v v
å åp
i ÎD k j Î J
i
j
w kj x i
with V º ¡ m - l , is a feasible dual function in the sense that
v
LD(d) £ z(d) " d Î W.
= -å w kj åp j
xi
jÎ J i ÎD k
i
Note that for a given d Î W, L(d, v) is a concave,
v
piecewise-polyhedral function. Therefore, the set Vd of
£ -å w kj p 0j
jÎJ
extreme points of epi(L(d, v)) is finite. Setting VW = ÈdÎWVd,
we can rewrite LD(d) = max{L(d, v): v Î VW}. It follows
= v b - z ( b ),
k
that if VW is finite, then LD reduces to the maximization of
finitely many subadditive functions and therefore, is
where the first inequality follows from the dual feasibility subadditive and feasible to the subadditive dual problem
of v k and w k , i.e., v k a i + å j Î J p i j w kj £ ci, i Î Dv k , (14). Furthermore, in the PILP case, LD corresponds to a
and the last inequality follows from the fact that p j x ³ Gomory function, since for a fixed v, (52) can be
Dk
represented by a Gomory function and the maximum of
p 0j , j Î J, are valid inequalities for { AD k x = b, x Î ¢ + v } finitely many Gomory functions is also a Gomory function.
v
LD above is a weak dual function in general, but Blair
and wk ³ 0. Rearranging, we have
and Jeroslow (1979) showed that it can be made strong for
PILP problems by introducing a quadratic term. To show
z ( b ) £ v k b - ( v k AD k - c D k )x £ Fv k ( b ). (51) this, we first need the following proximity relation.
v v

Combining this result with weak duality, we get z(b) Theorem (Blair and Jeroslow (1977)) For a given PILP
= Fv k ( b ). ■ with z(0) = 0, there is a constant > 0 such that

Klabjan (2002) also introduced an algorithm that finds z (d 1 ) - z (d 2 ) £ d1 - d 2 1 . (54)

the optimal dual function utilizing a subadditive approach
from (Burdet and Johnson (1977)) together with a row for all d1, d2 Î W.
generation approach that requires the enumeration of
feasible solutions. Unfortunately, even for the set Let the quadratic Lagrangian relaxation be defined as
partitioning problems that the author reports on, this
algorithm seems not to be practical. ì m
ü
L ( d , v , ) = minn í( c - vA )x +
x ΢ +
î
å( A x - d
i =1
i i )2 + vd ý , (55)
þ
4.5 Lagrangian relaxation

Another widely used framework for generating dual where v Î ¡ m , Î ¡ + and Ai is the ith row of A.
problems is that of Lagrangian duality (Fisher (1981)). A
mathematical program obtained by relaxing and Theorem 22. (Blair and Jeroslow (1979)) For a PILP in
subsequently penalizing the violation of a subset of the the form (2), denote the quadratic Lagrangian dual
original constraints, called the complicating constraints, is a function as
Guzelsoy and Ralphs: ality for Mixed-Integer Linear Programs 133
IJOR Vol. 4, No. 3, 118-137 (2007)

L D ( d , v ) = max L ( d , v , ). (56) 4.6 Linear representation of the subadditive dual

r Ρ +

For bounded PILPs with A Î ¤m+ ´n , the subadditive

Then for a given v Î ¡ , LD(d, v) £ z(d) " d Î W and
m
dual can be reformulated as an equivalent LP
furthermore, if b Î W and z(b) > -¥, then for any v Î ¡ m ,
LD(b, v) = z(b). max (b )
s.t. ( ) + ( ) ³ ( + ),
Proof. The first part follows from the fact that for any 0£ £ b, 0 £ £ b, 0 £ + £ b, (60)
d Î W and Î ¡ + ,
( a ) £ c j , j = 1, ..., n
j

(0) = 0,
ì m
ü
L ( d , v , ) £ min í( c - vA )x +
x ÎS ( d )
î
å
i =1
( Ai x - d i )2 + vd ý
þ (57)
after scaling A and b to be integer. This follows from the
= min cx = z ( d ). fact that the subadditive dual function in this case can be
x ÎS ( d )
represented by the values it takes over the finite domain
For the second part, we show that for right-hand side { Î ¢ m+ | £ b} (Gomory (1969), Johnson (1979)). The
b Î W with z(b) > -¥ and a given v Î ¡ m , there exists variables in the above LP represent the values of the
(v) Î ¡ + such that, L(b, v, (v)) = z(b). Let (v) = 1 + + subadditive function to be constructed at each point in this
domain and the constraints ensure that the function : { |
v 1 , with defined as in (54), assume that x Î ¢ n+ is an
£ b}® R is actually subadditive.
optimal solution to yield L(b, v, (v)) and let b = A x . Lasserre (2004c, b) further decreases the row dimension
Then, of this LP using a discrete version of Farkas’ lemma. Let
¡ [s1, ..., sm] be the ring of real-valued polynomials in the
m
variables si, i = 1, ...,m. Then, a polynomial Q Î ¡ [s1, ..., sm]
( c - vA )x + ( v )å ( Ai x - bi )2 + vb
i =1
can be written as
m
= cx + v ( b - Ax ) + ( v )å ( b i - bi )2 Q( s ) = å  
s = å  1
s 1 ...s m m ,
i =1  Îz  Îz

³ z ( b ) + v (b - b ) + ( v ) b - b (58)
1

where Ì ¢ m+ and Ρ" Î .
³ z (b ) - b -b - v 1
b - b + (v ) b - b
1 1 1

= z (b ) + b - b Theorem 23. (Lasserre (2003)) The following two

1
properties are equivalent:
³ z (b )
1. Ax = b has a solution x Î ¢ n+ .
2. The real valued polynomial sb - 1 can be written as
by Theorem 21 and the fact that b - b £ å i =1 ( bi - b i )2 .
m

Therefore, L(b, v, (v)) ³ z(b) and due first part, LD(b, v) =

n
s b - 1 = å Q j ( s )( s a - 1)
j
(61)
L(b, v, (v)) = z(b). ■ j =1

Note that one can verify that (56) attains its maximum at for some real-valued polynomials Qj Ì ¡ [s1, ..., sm], j =
a point x* that is also optimal to the PILP. This is because 1, ..., n, all with nonnegative coefficients.
in order to get equality in (58), the conditions b = b and cx*
= z(b) have to be satisfied at the same time. Otherwise, Proof. (1) ® (2). Let x Î S. Writing
LD(b, v) > z(b). In addition, it is clear that ( v ) can be
replaced by any such that ³ ( v ) for a given v in 1 1 2
n -1
å j =1 a j x j n
s b - 1 = s a x1 - 1 + s a x1 ( s a x2
- 1) + ... + s ( s a x n - 1)
(58). In fact, if we let v be the optimal solution to the
dual of the LP relaxation of PILP, then choosing > z(b)
with
- vb is adequate, since

- 1 = ( s a - 1) é1 + s a + ... + s ù , j = 1,..., n ,
a jx j j j a j ( x j -1)
s
ë û
m
( c - vA )x + å (b
i =1
i - bi ) + vb ³
2
+ vb > z ( b ). (59)
we obtain
Due to dual feasibility, L( b , v , ) is forced to have its
j -1 k

Q j ( s ) = s å k=1 é1 + s a j + ... + s a ù , j = 1,..., n .

j
a xk ( x j - 1)
infimum at an x* that is also optimal to the PILP, since (62)
equality in (59) is attained only in that case. ë û
Guzelsoy and Ralphs:  ality for Mixed-Integer Linear Programs 134
IJOR Vol. 4, No. 3, 118-137 (2007)

(2) ® (1). Let q Î ¡ k+ be the vector of nonnegative subadditive formulation that is also dual to PILP.
coefficients of all polynomials Qj , j = 1, ..., n, and M Compared to (60), the number of variables is the same,
Î ¡ p ´k be such that the set of constraints defining the however, this one has O(np) constraints, whereas (60) has
polyhedron Q = {q|Mq = , q ³ 0} ¹ Ø equalizes the O(p2) constraints.
respective coefficients of the polynomials sb - 1 and
å
n j
Q j ( s )( s a - 1). It is easy to show that each Qj, j = 4.7 Branch and cut
j =1

1, ..., n, may be restricted to contain only monomials {s!: The most common technique for solving MILPs in
£ b -a j, Î ¢ m+ } and therefore practice today is the branch-and-cut algorithm. Developing
a procedure for obtaining a dual function as a by-product
m of this procedure is of great importance if duality is to be
p = Õ ( bi + 1) made computationally useful. Here we discuss “vanilla”
i =1
branch and cut, in which branching is done only by
n m
k = å k j with k j = Õ ( bi - a i j + 1), j = 1,..., n . restricting variable bounds and no standard computational
j =1 i =1 enhancements, such as preprocessing, are used. Such an
algorithm works by attempting to solve the subproblem of
In other words, p is the number of monomials y # with each branch-and-cut tree node utilizing a cutting plane
method, as described in Section 4.3. If the subadditive
£ b and kj is the number of monomials y # with - characterization or a functional encoding of the right-hand
a £ b. With this construction, it is not hard to see that M
j
side dependency is available for each cut, then we can
is totally unimodular and each extreme point of Q, if it obtain a dual function for the corresponding subproblem.
exists, is integral, since is also integral. Below, we show how this dual information can be gathered
k k
Next, recall that j = (1, ..., 1) Î ¡ j , j = 1, ..., n, and together to yield a feasible dual function for the original
n ´k problem.
let K Î ¢ + be the n-block diagonal matrix, whose each
Assume that the MILP (2) has a finite optimum and has
k
diagonal block is a row vector 1 j , that is, been solved to optimality with a branch-and-cut algorithm.
Let T be the set of leaf nodes of the tree and let (t) be the
é1k1 0 ¼ 0 ù number of cuts generated so far on the path from the root
ê ú node to node t Î T (including the ones generated at t). To
0 1k2 0 ¼ ú
K =ê . obtain a bound for this node, we solve the LP relaxation of
ê¼ ¼ ¼ ¼ ú
ê the following problem
kn ú
ëê 0 ¼ 0 1 ûú
z t ( b ) = min cx
(63)
Now, let Qj, j = 1, ..., n, be the set of polynomials s.t. x Î St ( b ),
satisfying (61). Then, Q ¹ Ø and there exists an integral
q Î Q. If we denote by Q j , j = 1, ..., n, the where the feasible region St(b) = {x Î ¢ r ´ ¡ n - r |Ax = b, x
corresponding monomials q represents and take the ³ l t , -x ³ -u t , P t x ³ P t0 } and u t , l t Î ¢ n+
derivative of both sides with respect to s i , i = 1, ..., m, at are the branching bounds applied to the integer variables,
(1, ..., 1), we get P t Î ¡n ( t )´n and P t0 Î ¡n ( t ) .
n n
For each cut k, k = 1, ..., (t), suppose that the
bi = å Q j (1,...,1)a i = å a i ( Kq ) j , i = 1,..., m .
j j
subadditive representation Fkt is known and let the
j =1 j =1
function kt be defined for (63) as in Section 4.3,
Observe that setting x = K q completes the proof. ■ considering also the branching bounds. For each feasibly
pruned node t Î T, let ( v t , v t , v t , w t ) be the
The converse of the last part of the proof is also valid, corresponding dual feasible solution used to obtain the
i.e., for any x Î S, x = Kq for some q Î Q. As a bound that allowed the pruning of node t. Note that such a
solution is always available if the LP relaxations are solved
consequence, we have the following corollary.
using a dual simplex algorithm. For each infeasibly pruned
Corollary 24. (Lasserre (2004c)) For a PILP in the form (2) node t Î T, let ( v t , v t , v t , w t ) be a corresponding dual
with A Î ¢ m+ ´n , let K, M, be defined as before. Then, z(b) feasible solution that can be obtained from the parent of
node t.
= min{cKq|Mq = , q ³ 0}. Moreover, if q* is an optimal
solution, then x* = Kq* is an optimal solution to the PILP.
Theorem 25. If b Î W and z(b) > -¥, then the function
Lasserre further shows that the LP dual of the problem
in the first part of Corollary 24 can be reduced to a
Guzelsoy and Ralphs: $%ality for Mixed-Integer Linear Programs 135
IJOR Vol. 4, No. 3, 118-137 (2007)

& (t ) & (t )
FBC ( d ) = min{v t d + v t l t - v t u t + å w kt Fkt ( t
k ( d ))} (64) FBCS ( d ) = max{v t d + å w kt Fkt ( t
k ( d ))}, d Î ¡ m + 2 n (69)
t ÎT t ÎT
k =1 k =1

is an optimal solution to the dual (8). is feasible to the subadditive dual problem (14) of the
MILP (67).
Proof. The proof follows the outline of Wolsey (1981)’s
proof for validating an optimal dual function for the Proof. For any t Î T, LP duality yields
branch-and-bound algorithm. Because of the way branch
& (t )
and cut partitions S, we are guaranteed that for any d Î W
c j ³ v t a j + å w kt P tj j = 1,..., n .
and x̂ Î S(d), there must exist a leaf node t Î T such that k =1

x̂ Î St (d). Then, from LP duality,

Therefore, it is clear that cj ³ FBCS( a j ) if j Î I and
c j xˆ j ³ v t a j xˆ j + v tj xˆ j - v jt xˆ j + w t P tj xˆ j j = 1,..., n , (65) likewise, cj ³ F BCS ( a j ) when j Î C. In addition, since
Fkt Î G m + 2 n + k -1 , k = 1, ..., (t), FBCS Î G m + 2n . ■
where P j is the jth column of P . Adding the above
t t

Note that in this case, the dual function may not be

inequalities over all columns, we get strong. As in Theorem 19, it is not strictly necessary to
have a subadditive representation of each cut in order to
cxˆ ³ v t Axˆ + v t xˆ - v t xˆ + w t P t xˆ apply the results of this section. They remain valid as long
&(t ) as a functional dependency of each cut on the
³ v t d + v t l t - v t u t + å w kt Fkt ( t
k ( d )) (66) right-hand-side is known (see Section 4.3).
k =1

³ FBC ( d )
5. CONCLUSION
Now assume that x* is an optimal solution to MILP with
right-hand side b. In this case, we know that for some node In this paper, we presented a survey of existing theory
t* t* and methodology for constructing dual functions. Many of
t*, z(b) = cx* = z ( b ) and we also have that zt(b) ³ z ( b )
the ideas presented here were developed more than three
for all t Î T. Therefore, FBC(b) = z(b). ■ decades ago and it would seem that little progress has been
made towards a practical framework. From the standpoint
Unfortunately, (64) is not subadditive due to the the of computational practice, the importance of these
constant term resulting from the bounds imposed by methodologies is that they may allow us to extend to the
branching and hence is not feasible for the subadditive dual realm of MILP some of the useful techniques already
(14). One can, however, obtain a subadditive dual function well-developed for linear programming, such as the ability
in the case where the original MILP has explicit upper and to perform post facto sensitivity analyses and the ability to
lower bounds on all variables by including these bounds as warm start solution processes. The development of such
part of the right-hand side. Suppose that techniques is the underlying motivation for our work.
Constructing a strong dual function for a given MILP is at
z ( b ) = min{cx | Ax ³ b , x Î ¢ r ´ ¡ n - r } (67) least as difficult as solving the primal problem, so there is
little hope of or use for constructing such functions
with A = [ A I - I ]¢ and b = [ b l - u ] where l and u are independent of primal solution algorithms. This leaves two
possible options—constructing dual functions that are not
the lower and upper bounds pre-defined on the variables.
necessarily strong, but may still give us some ability to
With this construction, at each node t Î T, we solve the LP analyze the effect of changes in the input data, or
relaxation of the following subproblem constructing dual functions as a by-product of existing
t
primal solution algorithms, namely branch and cut.
z t ( b ) = min cx Currently, the techniques discussed in Section 4.7
t represent the clearest path to bringing a practical notion of
s.t. Ax ³ b
duality to fruition, since these techniques work in tandem
P t x ³ P t0 with algorithms that are already effective in solving the
x Î ¢r ´ ¡n -r (68) primal problem. Execution of the branch-and-cut
algorithm produces a tremendous amount of dual
t information, most of which is normally discarded. By
with b = [ b l t - u t ]. retaining this information and using it effectively, one may
be able to develop procedures for both sensitivity analysis
Theorem 26. If S( b ) ¹ Ø and z ( b ) > -¥, then the and warm starting. Ralphs and Guzelsoy (2005) describe an
function implementation of these techniques, using the
Guzelsoy and Ralphs: '(ality for Mixed-Integer Linear Programs 136
IJOR Vol. 4, No. 3, 118-137 (2007)

SYMPHONY MILP solver framework, that supports both continuous functions related to corner polyhedra, II.
warm starting and basic sensitivity analysis for branch and Mathematical Programming, 3: 359-389.
bound. They also describe recent results using warm 15. Jeroslow, R.G. (1978). Cutting plane theory: Algebraic
starting to accelerate the solution process for algorithms methods. Discrete Mathematics, 23: 121-150.
that involve solution of a sequence of related MILPs 16. Jeroslow, R.G. (1979). Minimal inequalities.
(Ralphs and Guzelsoy (2006)). These results are still Mathematical Programming, 17: 1-15.
preliminary, but demonstrate the potential for further 17. Johnson, E.L. (1973). Cyclic groups, cutting planes and
development. shortest paths. In: T.C. Hu and S.M. Robinson (Eds.),
Mathematical Programming, Academic Press, New York,
pp. 185-211.
ACKNOWLEDGEMENT 18. Johnson, E.L. (1974). On the group problem for
This research was partially supported by National mixed integer programming. Mathematical Programming
Science Foundation Grants DMI- 0534862 and Study, 2: 137-179.
DMI-0522796. 19. Johnson, E.L. (1979). On the group problem and a
subadditive approach to integer programming. Annals
REFERENCES of Discreet Mathematics, 5: 97-112.
20. Klabjan, D. (2002). A new subadditive approach to
1. Bachem, A. and Schrader, R. (1980). Minimal equalities integer programming: Theory and algorithms.
and subadditive duality. Siam Journal on Control and Proceedings of the Ninth Conference on Integer Programming
Optimization, 18(4): 437-443. and Combinatorial Optimization, Cambridge, MA., pp.
2. Barvinok, A.I. (1993). Computing the volume, 384-400.
counting integral points, and exponential sums. Discrete 21. Lasserre, J.B. (2003). Duality and farkas lemma for
& Computational Geometry, 10: 1-13. integer programs. In: E. Hunt and C.E.M. Pearce
3. Barvinok, A.I. (1994). Polynomial time algorithm for (Eds.), Optimization Structure and Applications, Applied
counting integral points in polyhedra when the Optimization Series, Kluwer Academic Publishers.
dimension is fixed. Math of Operations Research, 19: 22. Lasserre, J.B. (2004a). Generating functions and
769-779. duality for integer programs. Discrete Optimization, 1:
4. Blair, C.E. (1978). Minimal inequalities for mixed 167-187.
integer programs. Discrete Mathematics, 24: 147-151. 23. Lasserre, J.B. (2004b). Integer programming, duality
5. Blair, C.E. (1995). A closed-form representation of and superadditive functions. Contemporary Mathematics,
mixed-integer program value functions. Mathematical 374: 139-150
Programming, 71: 127-136. 24. Lasserre, J.B. (2004c). The integer hull of a convex
6. Blair, C.E. and Jeroslow, R.G. (1977). The value rational polytope. Discrete & Computational Geometry,
function of a mixed integer program: I. Discrete 32(1): 129-139.
Mathematics, 19: 121-138. 25. Llewellyn, D.C. and Ryan, J. (1993). A primal-dual
7. Blair, C.E. and Jeroslow, R.G. (1979). The value integer programming algorithm. Discrete Applied
function of a mixed integer program: II. Discrete Mathematics, 45: 261-276.
Mathematics, 25: 7-19. 26. Loera, J.D., Haws, D., Hemmecke, R., Huggins, P.,
8. Blair, C.E. and Jeroslow, R.G. (1982). The value Sturmfels, B., and Yoshida, R. (2004a). Short rational
function of an integer program. Mathematical functions for toric algebra and applications. Journal of
Programming, 23: 237-273. Symbolic Computation, 38: 959-973.
9. Blair, C.E. and Jeroslow, R.G. (1984). Constructive 27. Loera, J.D., Haws, D., Hemmecke, R., Huggins, P.,
characterization of the value function of a and Yoshida, R. (2004b). Three kinds of integer
mixed-integer program: I. Discrete Applied Mathematics, 9: programming algorithms based on barvinok’s rational
217-233. functions. Computer Science, 3064: 244-255.
10. Brion, M. and Vergne, M. (1997). Residue formulae, 28. Meyer, R.R. (1974). On the existence of optimal
vector partition functions and lattice points in rational solutions to integer and mixed integer programming
polytopes. Journal of the American Mathematical Society, 10: problems. Mathematical Programming, 7: 223-235.
797-833. 29. Nemhauser, G.L. and Wolsey, L.A. (1988). Integer and
11. Burdet, C. and Johnson, E.L. (1977). A subadditive Combinatorial Optimization, Wiley, New York.
approach to solve integer programs. Annals of Discreet 30. Ralphs, T.K. and Guzelsoy, M. (2006). Duality and
Mathematics, 1: 117-144. warm starting in integer programming. Proceedings of the
12. Fisher, M.L. (1981). The lagrangian relaxation method 2006 NSF Design, Service, and Manufacturing Grantees and
for solving integer programming problems. Management Research Conference.
Science, 27: 1-17. 31. Ralphs, T.K. and Guzelsoy, M. (2005). The
13. Gomory, R.E. (1969). Some polyhedra related to SYMPHONY callable library for mixed-integer linear
combinatorial problems. Linear Algebra and Its programming. Proceedings of the Ninth INFORMS
Applications, 2: 451-558. Computing Society Conference, Annapolis, MD, pp. 61-76.
14. Gomory, R.E. and Johnson, E.L. (1972). Some 32. Schrage, L. and Wolsey, L.A. (1985). Sensitivity
Guzelsoy and Ralphs: ),ality for Mixed-Integer Linear Programs 137
IJOR Vol. 4, No. 3, 118-137 (2007)

analysis for branch and bound integer programming.

Operations Research, 33(5): 1008-1023.
33. Schrijver, A. (1980). On cutting planes. Annals of
Discrete Mathematics, 9: 291-296.
34. Williams, H.P. (1996). Duality in mathematics and
linear and integer programming. Journal of Optimization
Theory and Applications, 90(2): 257-278.
35. Wolsey, L.A. (1979). The b-hull of an Integer Program,
London School of Economics and CORE.
36. Wolsey, L.A. (1981). Integer programming duality:
Price functions and sensitivity analaysis. Mathematical
Programming, 20: 173-195.