Professional Documents
Culture Documents
Textbook Disjunctive Programming Egon Balas Ebook All Chapter PDF
Textbook Disjunctive Programming Egon Balas Ebook All Chapter PDF
https://textbookfull.com/product/disjunctive-programming-balas-e/
https://textbookfull.com/product/bitcoin-and-blockchain-history-
and-current-applications-balas/
https://textbookfull.com/product/structures-of-computing-a-guide-
to-practice-oriented-theory-1st-edition-egon-borger/
https://textbookfull.com/product/disjunctive-prime-ministerial-
leadership-in-british-politics-from-baldwin-to-brexit-
christopher-byrne/
Handbook of Data Science Approaches for Biomedical
Engineering 1st Edition Valentina Emilia Balas
https://textbookfull.com/product/handbook-of-data-science-
approaches-for-biomedical-engineering-1st-edition-valentina-
emilia-balas/
https://textbookfull.com/product/data-management-analytics-and-
innovation-proceedings-of-icdmai-2018-volume-2-valentina-emilia-
balas/
https://textbookfull.com/product/emergence-of-pharmaceutical-
industry-growth-with-industrial-iot-approach-1st-edition-
valentina-emilia-balas/
https://textbookfull.com/product/a-handbook-of-internet-of-
things-in-biomedical-and-cyber-physical-system-valentina-e-balas/
https://textbookfull.com/product/soft-computing-applications-
proceedings-of-the-8th-international-workshop-soft-computing-
applications-sofa-2018-vol-i-valentina-emilia-balas/
Egon Balas
Disjunctive
Programming
Disjunctive Programming
Egon Balas
Disjunctive Programming
123
Egon Balas
Tepper School of Business
Carnegie Mellon University
Pittsburgh, PA, USA
This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Preface
This book is meant for students and practitioners of optimization, mainly integer and
nonconvex optimization, as an introduction to, and review of, the recently developed
discipline of disjunctive programming. It should be of interest to all those who are
trying to overcome the limits set by convexity to our problem-solving capability.
Disjunctive programming is optimization over disjunctive sets, the first large class
of nonconvex sets shown to be convexifiable in polynomial time.
There are no prerequisites for the understanding of this book, other than some
knowledge of the basics of linear and integer optimization. Clarity of exposition
was a main objective in writing it. Where background material is required, it is
indicated through references.
Habent sua fata libelli, goes the Latin saying: books have their own fate. The
basic document on Disjunctive Programming is the July 1974 technical report
“Disjunctive Programming: Properties of the convex hull of feasible points. MSRR
#348”, which however has not appeared in print until 24 years later, when it was
published as an invited paper with a remarkable foreword by Gerard Cornuejols
and Bill Pulleyblank [7] (for a history of this episode see the introduction to [11]).
The new theory, which was laid out without implementation of its cutting planes,
let alone computational experience, stirred little if any enthusiasm at the time of its
inception. But 15 years later, when Sebastian Ceria, Gerard Cornuejols and myself
recast essentially the same results in a new framework which we called lift-and-
project [19], the reaction was quite different. This time our work was focused on
algorithmic aspects, with the cutting planes generated in rounds and embedded into
an enumerative (branch and cut) framework, and was accompanied by an efficient
computer code, MIPO, that was able to solve many problem instances that had been
impervious to solution by branch-and-bound alone. This has led to a general effort
on the part of software builders to incorporate cutting planes into a new generation
of integer programming solvers, with the outcome known as the revolution in the
state of the art in mixed integer programming that took place roughly in the period
1990–2005.
From a broader perspective, disjunctive programming is one of the early bridges
between convex and nonconvex programming.
v
vi Preface
vii
viii Contents
References .. .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 233
Chapter 1
Disjunctive Programming and Its
Relation to Integer Programming
1.1 Introduction
min{cx : x ∈ PI }, (MIP)
1.2 Intersection Cuts 3
where PI represents the set of feasible integer points. The linear programming
relaxation of (MIP), of central importance in all major solution methods, is
min{cx : x ∈ P }, (LP)
where
P := {x ∈ RN : Ax ≥ b, x ≥ 0}
PI := P ∩ {x ∈ RN : xj ∈ Z, j ∈ N ⊆ N}.
optimal solution to (LP) with I and J the index sets of basic and nonbasic variables,
respectively, then the optimal simplex tableau is
xi = x̄i − āij xj , i ∈ I
j ∈J
xj ≥ 0, j ∈I ∪J
When convenient, we will use the notations xN and xJ for the subvectors of x
with components indexed by j ∈ N and j ∈ J , respectively.
Associated with every basic solution x̄ to (LP) is the LP-cone C(x̄) whose apex is
x̄ and whose facets are defined by the n hyperplanes that form the basis of x̄. In order
to avoid possible confusions in the case of a degenerate x̄ which is the intersection
of more than n hyperplanes, we will sometimes denote this cone by C(J ), meaning
that the n tight inequalities defining the cone are those that are nonbasic at x̄. Thus
j
C(J ) has n extreme rays with direction vectors r j , j ∈ J , with ri = −āij for i ∈ I
j j
and rj = 1, ri = 0 for i ∈ J \ {j }.
The convex hull of feasible integer points, conv(PI ), is also known as the integer
hull. The convex hull of integer points in C(J ) is known as the corner polyhedron
[82], denoted corner(J ). The relationship between the various polyhedral relax-
ations of PI introduced above can be summarized as C(J ) ⊃ P ⊃ conv(PI ) and
C(J ) ⊃ corner(J ) ⊃ conv(PI ).
The central object of research in integer programming is of course conv(PI ),
the integer hull, whose knowledge would transform a mixed integer program into a
linear one. This elusive object has defied the strenuous efforts spent on identifying
it, due to two circumstances: (a) its facets are hard to come by, and (b) their numbers
are typically exponential in the problem dimensions. Therefore attempts to identify
4 1 Disjunctive Programming and Its Relation to Integer Programming
r2
x r1
x̄
S
(b)
x
r2
1.2 Intersection Cuts 5
S := {x ∈ Rn : d i x ≤ d0i , i ∈ Q}
can also be viewed as a cut derived from the condition that every feasible point
satisfy the disjunction
(d i x ≥ d0i ), (1.1)
i∈Q
Clearly, (1.2) is stronger, i.e. more restrictive, than (1.1). But beyond this
consideration, the formulation (1.2) suggests a way to describe the family of valid
cuts, i.e. of inequalities satisfying condition (1.2), as those defining the convex hull
of a union of polyhedra. And this is the crux of the matter. But before we address
the question of finding the convex hull of a union of polyhedra, we have to address
the question of how to deal with disjunctive sets of inequalities that do not seem to
conform to this model? For instance, what about aset like
1the one1 defined
by the
g 1 x ≥ g01 h x ≤ h0
condition ((ax ≤ b) ∨ (cx ≥ d)) ⇒ ∨ .
g 2 x ≥ g02 h2 x ≤ h20
6 1 Disjunctive Programming and Its Relation to Integer Programming
A B A∨B A∧B
T T T T
T F T F
F T T F
F F F F
This table says that if both statement letters A and B are true, then both A ∨ B
and A ∧ B are true; if both A and B are false, then both A ∨ B and A ∧ B are false;
but if exactly one of A and B is true, then A ∨ B is true but A ∧ B is false.
Working with logical connectives requires knowledge of the basic laws of the
Algebra of Propositions, which are as follows:
Associative law: (A ∨ B) ∨ C = A ∨ (B ∨ C), (A ∧ B) ∧ C = A ∧ (B ∧ C)
Commutative law: A ∨ B = B ∨ A, A ∧ B = B ∧ A
Distributive law: A∨(B ∧C) = (A∨B)∧(A∨C), A∧(B ∨C) = (A∧B)∨(A∧C)
1.3 Inequality Systems with Logical Connectives 7
Ax ≥ b
xj ∈ {0, 1}, j = 1, . . . , n
Ax ≥ b
x ∈ {0, 1}n
8 1 Disjunctive Programming and Its Relation to Integer Programming
Note that the relationship between disjunction and conjunction (union and inter-
section) bears a certain similarity with the relationship between ordinary addition
and multiplication; namely, they both abide by the associative and commutative
laws. However, whereas multiplication is distributive with respect to addition, the
converse is false: addition is not distributive with respect to multiplication. By
contrast, disjunction and conjunction (union and intersection) are distributive with
respect to each other, which makes a significant difference.
It is precisely this distributive property of disjunction and conjunction with
respect to each other which makes it possible to put any logical statement in either
conjunctive or disjunctive normal form.
Example 1 Consider the expression
(AB ∨ CD) ∧ (E ∨ F ) ∧ G,
(A ∨ C) ∧ (A ∨ D) ∧ (B ∨ C) ∧ (B ∨ D) ∧ (E ∨ F ) ∧ G.
The ability to put any logical statement, hence any disjunctive set, in any of these
two normal forms is crucial to the study of their properties: as it will soon become
obvious, each of the two crucial properties of disjunctive sets is derived from, and
applies to, one and only one of the two normal forms. This is remarkable in light
of the fact that the two normal forms are just two ways to describe the same entity,
namely the set of solutions to the system described in these two forms.
Of the two normal forms, the one that is of foremost interest to us, is the DNF,
which has a geometric interpretation as a union of polyhedra.
Once the inequalities of a disjunctive set are given, the disjunctive and conjunc-
tive normal forms are unique. However, it is important to notice that the inequalities
expressing the conditions of a given problem, can usually be chosen in more than
one way. For instance, the constraint set
3x1 + x2 − 2x3 + x4 ≤ 1,
x1 + x2 + x3 + x4 ≤ 1
xj = 0 or 1, j = 1, . . . , 4,
1.4 Valid Inequalities for Disjunctive Sets 9
3x1 + x2 − 2x3 + x4 ≤ 1
∨4i=1 (xi = 1, xj = 0, j = i) ∨ (xj = 0, ∀j ),
We start with a characterization of all valid inequalities for a disjunctive set. For that
purpose we need the disjunctive set in DNF.
Theorem 1.2 ([6, 8] Farkas’ Lemma for Disjunctive Sets) Let
F = Ph , Ph := {x ∈ Rn : Ah x ≥ bh }, h ∈ Q, (1.3)
h∈Q
αj = maxh∈Q∗ uh Ahj j ∈ N
(1.4)
α0 ≤ minh∈Q∗ uh bh
for some uh ≥ 0, h ∈ Q∗ .
10 1 Disjunctive Programming and Its Relation to Integer Programming
aij (−aij )
max , xj ≥ 1 (1.4 )
ai0 (1 − ai0 )
j ∈J
Note that the cut derived from xi ≤ 0, ∨xi ≥ 1 is by far not unique; it depends on
the set on nonbasic variables in terms of which xi as a basic variable is expressed.
An important feature of the general expression (1.4) is that putting a given cut
in the form (1.4) may suggest ways of improving it by an appropriate choice of the
multipliers.
We illustrate this by the following
Example 2 Consider the mixed integer program whose constraint set is
This problem is taken from the paper [87], which also lists six cutting planes
derived from the extreme valid inequalities for the associated group problem:
The first two of these inequalities are the mixed-integer Gomory cuts derived
from the row of x1 and x2 , respectively. To show how they can be improved, we first
derive them as they are. To do this, for a row of the form
xi = ai0 + aij (−xi ),
i∈J
Then every x which satisfies the above equation and the integrality constraints
on xj , j ∈ J1 ∪ {i}, also satisfies the condition
yi = ϕi0 + ϕij (−xj ), yi integer.
j ∈J
For the two equations of the example, the resulting conditions are
that is
The sum of coefficients on the left hand side has been reduced from 1.9875 to
−5.95.
Similarly, for the second cut, if instead of y2 ≤ 0 ∨ y2 ≥ 1 we use the disjunction
y2 ≤ 0
∨ {y2 ≥ 1},
x1 ≥ 0
Here the sum of left hand side coefficients has been reduced from 1.733 to −11.6.
In this section we state a duality theorem for disjunctive programs, which general-
izes to this class of problems the duality theorem of linear programming.
1.5 Duality for Disjunctive Programs 13
z0 = min cx,
Ah x ≥ b h (DP)
,
h∈Q x≥0
w0 = max w,
⎧ ⎫
⎪
⎪ w − uh b h ≤ 0 ⎪
⎪
⎨ ⎬ (DD)
u A ≤c .
h h
⎪ ⎪
h∈Q ⎪
⎩ ⎪
⎭
uh ≥ 0
Further, let
i.e., if (DP) is feasible and (DD) is infeasible, then there exists h ∈ Q such that
Xh = ∅, Uh = ∅.
Theorem 1.5 ([9, 10]) Assume that (DP) and (DD) satisfy the regularity condition.
Then exactly one of the following two situations holds.
(1) Both problems are feasible; each has an optimal solution and z0 = w0 .
(2) One of the problems is infeasible; the other one either is infeasible or has no
finite optimum.
14 1 Disjunctive Programming and Its Relation to Integer Programming
Proof
(1) Assume that both (DP) and (DD) are feasible. If (DP) has no finite minimum,
then there exists h ∈ Q such that X̄h = ∅ and x̄ ∈ X̄h such that cx̄ < 0. But
then Uh = ∅, i.e., (DD) is infeasible; a contradiction.
Thus (DP) has an optimal solution, say x̄. Then the inequality cx ≥ z0 is
a consequence of the constraint set of (DP); i.e., x ∈ Xh implies cx ≥ z0 ,
∀h ∈ Q. But then for all h ∈ Q∗ , there exists uh ∈ Uh such that uh bh ≥ z0 .
Further, since (DD) is feasible, for each h ∈ Q \ Q∗ there exists ûh ∈ Uh ; and
since Xh = ∅ (for h ∈ Q \ Q∗ ), there also exists ūh ∈ Ūh such that ūh bh > 0,
∀h ∈ Q \ Q∗ . But then, defining
cx̄ = min{cx|x ∈ Xh }.
= max{w|w − uh bh ≤ 0, uh ∈ Uh }
"
"#
"
≥ max w " (w − u b ≤ 0, u ∈ Uh )
h h h
"h∈Q
uh (λ) ∈ Uh , h ∈ Q, for all λ > 0, and since ūh bh > 0, ∀h ∈ Q, w can be made
arbitrarily large by increasing λ; i.e., (DD) has no finite optimum.
Conversely, if (DD) is infeasible, then either (DP) is infeasible and we are
done, or else, from the regularity condition, Q∗ \Q∗∗ = ∅; and for h ∈ Q∗ \Q∗∗
there exists x̂ ∈ Xh and x̄ ∈ X̄h such that cx̄ < 0. But then
x(μ) = x̂ + μx̄
1.5 Duality for Disjunctive Programs 15
is a feasible solution to (DP) for any μ > 0, and since cx̄ < 0, z can be made
arbitrarily small by increasing μ; i.e., (DP) has no finite optimum.
The above theorem asserts that either situation 1 or situation 2 holds for (DP)
and (DD) if the regularity condition is satisfied. The following Corollary shows that
the condition is not only sufficient but also necessary.
Corollary 1.6 If the regularity condition does not hold, then if (DP) is feasible
and (DD) is infeasible, (DP) has a finite minimum (i.e., there is a “duality gap”).
Proof Let (DP) be feasible, (DD) infeasible, and Q∗ \ Q∗∗ = ∅, i.e., for every
h ∈ Q∗ , let Uh = ∅. Then for each h ∈ Q∗ , min{cx|x ∈ Xh } is finite, hence (DP)
has a finite minimum.
Remark 1.7 The theorem remains true if some of the variables of (DP) [of (DD)] are
unconstrained, and the corresponding constraints of (DD) [of (DP)] are equalities.
The regularity condition can be expected to hold in all but some rather peculiar
situations. In linear programming duality, the case when both the primal and the
dual problem is infeasible only occurs for problems whose coefficient matrix A has
the very special property that there exists x = 0, u = 0, satisfying the homogeneous
system
Ax ≥ 0, x ≥ 0;
uA ≤ 0, u ≥ 0.
In this context, our regularity condition requires that, if the primal problem is
feasible and the dual is infeasible, then at least one of the matrices Ah whose
associated set Uh are infeasible, should not have the above mentioned special
property.
Though most problems satisfy this requirement, nevertheless there are situations
when the regularity condition breaks down, as illustrated by the following example.
Consider the disjunctive program
max w,
w + 2u12 ≤ 0,
− u11 − u12 ≤ −1,
u11 − u12 ≤ −2,
(DD)
w − u22 ≤ 0,
− u1 + u22
2 ≤ −1,
u21 − u22 ≤ −2,
uki ≥ 0, i = 1, 2; k = 1, 2.
The primal problem (DP) has an optimal solution x̄ = (0, 2), with cx̄ = −4;
whereas the dual problem (DD) is infeasible. This is due to the fact that Q∗ = {1},
Q \ Q∗∗ = {2} and X2 = ∅, U2 = ∅, i.e., Q∗ \ Q∗∗ = ∅, hence the regularity
condition is violated. Here
" " 2
" −x1 + x2 ≥ 0 " −u + u2 ≤ −1
2 " 2 " 1 2
X2 = x ∈ R+ " , U2 = u ∈ R+ " .
" x1 − x2 ≥ 1 " u2 − u2 ≤ −2
1 2
Chapter 2
The Convex Hull of a Disjunctive Set
There are two ways to describe the convex hull of a disjunctive set, and each one
provides its insights. Here we discuss the first one.
Lifting is a topic introduced into the discrete optimization literature as the lifting
of inequalities, in the following sense: given a (MIP) in Rn as defined in Sect. 1.1,
suppose we know q < n coefficients of a valid inequality; how do we find values
for the remaining n − q coefficients that leave the inequality valid? There is a rich
literature on this and related topics. Here we refer to lifting in a broader sense,
as an operation which in a way is the reverse of projection. Typically, discrete
or combinatorial optimization problems can be formulated in more than one way.
Given a formulation in, say, Rn , by lifting we mean finding a formulation in a higher
dimensional space that offers some advantage. In the current context, our objective
is to find a description of the convex hull of a disjunctive set, i.e. of a union of
polyhedra, and we attain our objective by formulating the convex hull in a higher
dimensional space and projecting back the resulting formulation onto the original
space.
One possible way to derive this higher-dimensional formulation is the following.
Consider a disjunctive program
min cx
$ (DP)
h∈Q (A
hx ≥ bh )
Its dual, defined (for the case where x ≥ 0) in Sect. 1.5, can be stated as the linear
program
max w
w − uh b h ≤ 0
(DD)
uh A h = c
uh ≥ 0 h∈Q
(Here equation replaces inequality in (DD) because x is not sign restricted in (DP).)
Taking the linear programming dual of (DD) we get (denoting |Q| = q)
min cy 1 + · · · + cy q
A1 y 1 − b1y01 ≥0
.. ..
. .
q
Aq y q − b q y0 ≥ 0
q
y01 + ··· + y0 = 1
q
y01 , . . . , y0 ≥ 0
If we now represent the sum of the y h by the new variable x ∈ Rn , the result is
min cx
x− yh =0
h∈Q
Ah y h − bh y0h ≥ 0 h ∈ Q
(DP )
y0h = 1
h∈Q
y0h ≥ 0, h ∈ Q.
The fact that (DP ) is the linear programming dual of (DD) suggests that this
problem is equivalent to the disjunctive program (DP). This is indeed the case, and
furthermore the constraint set of (DP ) describes, in its higher-dimensional space,
the convex hull of the feasible set of (DP).%Indeed, consider the polyhedra Ph :=
{x ∈ Rn : Ah x ≥ bh }, h ∈ Q, and let F := h∈Q Ph , Q∗ := {h ∈ Q : Ph = ∅}.
2.1 The Convex Hull Via Lifting and Projection 19
Theorem 2.1 ([6]) The closed convex hull of F , cl conv F , is the set of those x for
which there exist vectors (y h , y0h ) ∈ Rn+1 , h ∈ Q∗ , satisfying
x− yh =0
h∈Q∗
Ah y h − bh y0h ≥ 0
(2.1)
y0h ≥ 0 h ∈ Q∗
y0h = 1.
h∈Q∗
with θ hi ≥ 0, i ∈ Uh , σ hk ≥ 0, k ∈ Vh , and i∈Uh θ hi = ȳ0h .
For h ∈ Q∗2 , ȳ h is either
0 or a nontrivial solution to the homogeneous system
A y h ≥ 0, hence ȳ h = k∈Vh σ hk v hk , σ hk ≥ 0, k ∈ Vh for some v hk ∈ dir Ph ,
h
Another random document with
no related content on Scribd:
Puhemies Näppisen kysyttyä, oliko vielä muita asioita, huusi Kaisa
Kompura kahvilan puolelta, että entäs se vallankumouksen
toimeenpaneminen, jokos siitä on päätetty?
Pekka Suova sanoi, että jos tässä sortaminen olisi kyseessä, niin
olisi hän ensin mennyt tuonne kahvilan puolelle pyytämään
maistellakseen sitä Reeta Sinkkosen limonaatia, jolla on niin muikea
haju, että tuntuu pihalle asti.
8 §.
Kokouksen jatkuessa sääntöperäisessä järjestyksessä ilmoitti
puhemies Näppinen kysymystä olleen kommuunin keskuudessa
yksityisesti, olisiko kommunistien ryhdyttävä keskustelemaan
noskelaisten kanssa yhteistoiminnan aikaansaamisesta
eduskuntavaaleissa köyhälistön rintaman lujittamiseksi porvarillista
taantumusta vastaan.
Paavo Pellikka veti esiin näytöksiä sos. dem. lehdistä, että nämä
tuomitsevat ankarasti porvarillisen oikeuspuntarin, jos se tahtoo
takseerata tuomiota vallankumouksen valmistamisesta, mihin
sihteeri Mikko Tarjus liitti välihuudahduksen, että ne ovat vain
salataantumuksen grogotiilin mustia murhekyyneliä
vallankumouksellisen käsityksen hämäämiseksi.
11 §.
12 §.