Lift-and-project cuts

for
mixed 0-1 programs

Stefano Smriglio
Dipartimento di Informatica
Universitµ
a di L'Aquila
[email protected]
+ +

Outline

² Preliminaries

² Disjunctive Programming

² Application: sequential convexi¯cation

² Lift-and-project cutting planes

² Computational achievements

² Drawbacks

² Lift-and-project in Combinatorial Optimiza-

tion

² Research directions

+ 1
+ +

Basic notions

² a vector y 2 IRn is a linear combination of vectors

fx1; : : : ; xk g if real multipliers ¸1 ; : : : ; ¸k exist such
P
that y = ki=1 ¸i xi. A linear combination with ¸i ¸
0, i = 1; : : : ; k is said conic. A linear combination
Pk
with i=1 ¸i = 1 is said a±ne. A combination
both a±ne and conic is said convex

² Given a set S µ IRn, the convex hull conv(S) of S

is the set of all convex combinations of vectors in
S.

² A polyhedron P = fx 2 IRn : Ax ¸ bg is the set

of vectors satisfying a (¯nite) system of linear in-
equalities. A bounded polyedron is a polytope.

² The dimension dim(P ) of a polyhedron P is the

maximum number of a±nely independent points
in P minus one.

² A polyhedral cone is a polyhedron of the form

C = fx 2 IRn : Ax ¸ 0g. A ray r(x) of C induced
by a vector x 2 C is the conical hull of fxg. A ray
r(x) is extreme if x cannot be obtained as conic
combination of other vectors of C.

+ 2
+ +

Valid inequalities and facets

² An inequality ®x ¸ ¯ is valid for P if for every

¹ 2 P ®¹
x x ¸ ¯ holds;

² a valid inequality ®x ¸ ¯ de¯nes a face F = P \

fx 2 IRn : ®x = ¯g;

² if dim(F ) = dim(P ) ¡ 1 then F is a facet;

² we say that inequality ax ¸ ® dominates inequal-

ity cx ¸ ° if there exists ¸ > 0 such that a = ¸c
and ° ¸ ¸® (i.e., the set of solutions of the for-
mer inequality is contained in the one of the latter
inequality). If solution sets coincide the two in-
equalities equivalent;

² an inequality ax ¸ ® is valid for P if and only if it

is either equivalent or dominated by a conic com-
bination uAx · ub, u ¸ 0m of inequalities de¯ning
P;

² the representation Ax ¸ b of P is minimal if re-

moving any inequality yields a system A0 x ¸ b0 such
that P ½ fx 2 IRn : A0 x ¸ b0 g;

² when dim(P ) = n, the representation Ax ¸ b of P

is minimal if and only its inequalities biunivoquely
correspond to the facets of P .

+ 3
+ +

Mixed 0-1 Programs

min cx
s.t.
Ax ¸ b
xj 2 f0; 1g; j = 1; : : : ; p

De¯ne:

² generic formulation of the problem

P := fx 2 IRn : Ax ¸ b; xj ¸ 0; xj · 1; j =
~ ¸~
1; : : : ; pg := fx 2 IRn : Ax bg

² the feasible set

F := fx 2 P : xj 2 f0; 1g; j = 1; : : : ; pg

² the convex hull of feasible solutions (i.e.

\optimal formulation")

PF = conv(F )
+ 4
+ +

Formulation ranking

F = f(0; 1); (0; 1); (0; 1)g

x2
(1, 1)
(0, 1)

1/2 P

1/2 (1, 0) x1

x2
(1, 1)
(0, 1)

P¢ Í P

x1
1/2 (1, 0)

x2 (1, 1)
(0, 1)

conv (F )

x1

(1, 0)

+ 5
+ +

Cutting plane

x*

x¤ optimal solution to current LP relaxation

x*

P cut

cut-o®

+ 6
+ +

Projection of polyhedra

Let P = f(x; y) 2 IRp+q : Dx + By · dg. The

projection of P onto the x-space is:

P rojx(P ) = fx 2 IRp : 9y 2 IRq : (x; y) 2 P g

² method 1: Fourier-Motzkin

² method 2: Balas-Pulleyblank ('83): elimi-

nating several variables at a time.

De¯ne the projection cone W for P : W =

¹ be the set
fu 2 IRm : uB = 0; u ¸ 0g. Let W
of extreme rays of W .

Theorem 1

P rojx(P ) = fx 2 IRp : (uD)x · ub for all u 2 W

¹g

+ 7
+ +

Disjunctive Programming

² Optimization over unions of polyhedra,

non convex;

² solution sets of systems of linear inequali-

ties joined by logical operations (i.e., con-
junctions, disjunctions), where noncon-
vexity arise from disjunctions (£ _ ¤)

Logic statements: £ = x 2 S1, ¤ = x 2 S2

£ _ ¤ = x 2 S1 [ S2

Let P i = fx 2 IRn : Aix ¸ big, i 2 Q be convex

nonempty polyhedra, with Q ¯nite index set
and (Ai; bi) mi £ (n + 1) matrix.

Feasible region:

P~ := fx 2 IRn : _i2Q(Aix ¸ bi)g

+ 8
+ +

Disjunctive Principle

Recall that Aix ¸ bi ) (uiAi)x ¸ uibi, for any

ui ¸ 0

P 1] (u1A1)1x1 + : : : + (u1A1)nxn ¸ u1b1

P 2] (u2A2)1x1 + : : : + (u2A2)nxn ¸ u2b2
:::
P Q] (uQAQ)1x1 + : : : + (uQAQ)nxn ¸ uQbQ

such a collapsing of each P i to a single con-

straint followed by selection of worst cases
among i 2 Q yields the following inequality:

X
maxf(uiAi)j gxj ¸ minfuibig (1)
i2Q i2Q
j=1;:::;n

Theorem 2 (Jeroslow, '79) Every minimal

valid inequality for P~ can be computed via
the disjunctive principle (1) by an appropri-
ate choice of nonnegative u1; : : : ; uQ 2
[Proof by LP duality.]
+ 9
+ +

Geometry

P1 P2

Conv(P1ÈP2)

+ 10
 + +

Lift-and-project

Lift Compact representation of the convex hull

of a union of polyhedra in higher dimen-
sional space.

Project Projection of the above representation into

the original space

Given polyhedra P i := fx 2 IRn : Aix ¸ big 6

=
;, i 2 Q. Convex hull PQ of [i2QP i:

X
x¡ ¸i y i = 0
i2Q
Aiyi ¸ bi; i2Q
X
¸i = 1
i2Q
¸i ¸ 0; i2Q

non linear system in x; y i; ¸

+ 11
+ +

Representation of the convex hull

Theorem 3 (Balas '74) Given polyhedra P i :=

fx 2 IRn : Aix ¸ big 6 = ;, i 2 Q, the closed
convex hull PQ of [i2QP i is the set of those
x 2 IRn for which there exist vectors (yi; y0i ) 2
IRn+1, i 2 Q, satisfying

X
x¡ yi = 0
i2Q
Aiy i ¡ biy0i ¸ 0; i2Q
y0i ¸ 0; i2Q
X
y0i = 1
i2Q

² number of variables: n + (n + 1)jQj

² number of constraints: linear in jQj

+ 12
+ +

Application to 0-1 programs

Consider a 0-1 program

min cx
s.t.
~ ¸ ~
Ax b
xj 2 f0; 1g; j = 1; : : : ; p

Its feasible region can be represented by the

following disjunctive program: choose a vari-
able j and look at the two polyhedra

Pj0 := fx 2 IR+ ~ ¸~
n : Ax b; xj = 0g,

Pj1 := fx 2 IR+ ~ ¸~
n : Ax b; xj = 1g,

then, P~ = fPj0 _ Pj1g.

+ 13
+ +

0-1 disjunction

Disjunction of the form xj 2 f0; 1g, jQj = 2,

Pj0 := fx 2 IR+ ~ ¸~
n : Ax b; xj = 0g,

Pj1 := fx 2 IR+ ~ ¸~
n : Ax b; xj = 1g,

PQ = conv(Pj0 [Pj1) is the set of those x 2 IRn

for which there exist vectors (y; y0); (z; z0) 2
IRn+1, such that

x¡y¡z = 0
~ ¡~
Ay by0 ¸ 0
¡yj = 0
~ ¡~
Az bz0 ¸ 0
zj ¡ z0 = 0
y0 + z0 = 1

manageable size

+ 14
+ +

Projection

Theorem 4 (Balas, '74)

¤ g,
PQ = fx 2 IRn : ®x ¸ ¯ for all (®; ¯) 2 PQ
where PQ¤ := f(®; ¯) 2 IRn+1 :

® = uiAi; ¯ = uibifor some ui ¸ 0; i 2 Q g 2

¤ is the set
in the case of 0 ¡ 1 disjunction PQ
of those (®; ¯) 2 IRn+1 for which there exists
vectors u; v 2 IRm+n+p and u0; v0 2 IR satisfy-
ing:

~ ¡ u0ej
® ¡ uA =0
® ~ ¡ v0ej = 0
¡ vA
u~
b =¯
v~
b + v0 = ¯
u; v ¸ 0

+ 15
+ +

Mixed 0-1 Programs

min cx
s.t.
Ax ¸ b
xj 2 f0; 1g; j = 1; : : : ; p

De¯ne:

² generic formulation of the problem

P := fx 2 IRn : Ax ¸ b; x ¸ 0; xj · 1; j =
~ ¸~
1; : : : ; pg := fx 2 IRn : Ax bg

² the feasible set

F := fx 2 F : xj 2 f0; 1g; j = 1; : : : ; pg

² the convex hull of feasible solutions (i.e.

\optimal formulation")

PF = conv(F )
+ 16
+ +

Sequential convexi¯cation

1. select an index j 2 f1; : : : ; pg

~ ¸ ~
2. Multiply Ax b with 1 ¡ xj and xj to
obtain the non-linear system

~ ¡~
(1 ¡ xj )(Ax b) ¸ 0 (2)
~ ¡~
xj (Ax b) ¸ 0 (3)

3. Linearize (2)(3) by substituting yi for xixj ,

i = 1; : : : ; n, i 6
= j and xj for x2
j . Denote
by Mj (P ) the resulting polyhedron.

4. Project Mj (P ) onto the x-space. Call the

resulting polyhedron P roj j (P ).

+ 17
+ +

Tightening

Theorem 5 (Balas, Ceria, Cornuejols, '93)

P roj j (P ) = conv(P \ fx 2 IRn : xj 2 f0; 1gg)

2

Corollary 1 conv(F ) µ P roj j (P ) µ P 2

x2
(1, 1)
(0, 1)

(0, 0) (1, 0) x1

(1, 1)
(0, 1)
P Ç { x Î Â 2 : x1 Î {0,1}}

(0, 0) (1, 0) x1

P roj j (P )
+ 18
+ +

Insights

~ ¡~
(1 ¡ xj )(Ax b) ¸ 0 (2)
~ ¡~
xj (Ax b) ¸ 0 (3)

² any x 2 F satis¯es (2)-(3), while any

x 2 f0; 1gn n F does not satisfy (2)-(3)
) (2)-(3) is a non-linear formulation of
the problem;

² any x 2 P satis¯es both (2) and (3) ) no

tightening;

+ 19
+ +

Tightening

~ ¡~
(1 ¡ xj )(Ax b) ¸ 0 (2)
~ ¡~
xj (Ax b) ¸ 0 (3)

² replacing xixj with yi does not tighten (2)

and (3);

² yet, Corollary 1 holds (striclty unless the

0-1 constraint is redundant for xj ): re-
placing x2
j with xj yields tightening, since
it may cut o® fractional points.

+ 20
+ +

Generating conv(F )

Theorem 6
P roj j (P ) = conv(P \ fx 2 IRn : xj 2 f0; 1gg)
2

De¯ne, for t ¸ 2,

P roj j1;:::;jt = P roj it (P roj it¡1 : : : (P roj i1 (P )) : : :)

Theorem 7
P roj j1;:::;jt (P ) = conv(P \ fx 2 IRn : xj 2
f0; 1gg; j = j1; : : : ; jt)

Corollary 2
P roj 1;:::;p(P ) = conv(F )

+ 21
+ +

Example

F = fx 2 f0; 1g2 : 2x1 + 2x2 ¸ 1g =

= f(1; 0); (0; 1); (1; 1)g

P = fx 2 IRn:
2x1 + 2x2 ¸ 1
x1 ¸ 0
x2 ¸ 0
1 ¡ x1 ¸ 0
1 ¡ x2 ¸ 0g

x2
(1, 1)
(0, 1)

1/2 P

1/2 (1, 0) x1

+ 22
+ +

Example

Multiplying for 1¡x1 and x1, we have (System

(2)-(3)):
2x1 + 2x2 ¡ 1 ¡ 2x2
1 ¡ 2x1x2 + x1 ¸0
2x1 + 2x1x2 ¡ x1 ¸0
x2
1 ¸0
x1 ¡ x2
1 ¸0
x2 ¡ x1x2 ¸0
x1 ¡ x1x2 ¸0
x1x2 ¸0
1 ¡ 2x1 + x21 ¸0
1 ¡ x1 ¡ x2 + x1x2 ¸0
linearizing, we get M1(P ):
2x1 + 2x2 ¡ 1 ¡ 2x1 ¡ 2y + x1 ¸0
2x1 + 2y ¡ x1 ¸0
x1 ¸0
x2 ¡ y ¸0
x1 ¡ y ¸0
y ¸0
1 ¡ 2x1 + x1 ¸0
1 ¡ x1 ¡ x2 + y ¸0
+ 23
+ +

Cut o®

M1(P ) = f(x; y) 2 IR3 :

2x1 + 2x2 ¡ 1 ¡ 2x1 ¡ 2y + x1 ¸ 0
2x1 + 2y ¡ x1 ¸ 0; 1 ¡ 2x1 + x1 ¸ 0
x2 ¡ y ¸ 0; x1 ¡ y ¸ 0
1 ¡ x1 ¡ x2 + y ¸ 0; x1 ¸ 0; y ¸ 0g

x2
(1, 1)
(0, 1)

1/2 P

(1, 0)
1/2 (2/3, 0) x1

substituting (2=3; 0) in M1(P ):

y · ¡1=6, y ¸ 0:

) (2=3; 0) 6
2 M1(P )
+ 24
+ +

Tightening

x1 + 2x2 ¡ 2y ¸ 1
x1 + 2y ¡ x1 ¸ 0; 1 ¡ 2x1 + x1 ¸ 0
x2 ¡ y ¸ 0; x1 ¡ y ¸ 0
1 ¡ x1 ¡ x2 + y ¸ 0; x1 ¸ 0; y ¸ 0g

combining ¯rst and last inequality we obtain:

x1 + 2x2 ¡ 2y ¸ 1

x2
(1, 1)
(0, 1)

Proj1(P)
1/2

(1, 0)
1/2 x1

+ 25
+ +

Convexi¯cation by Disjunctive
Programming

Look at the disjunction:

à ! à !
~ ¸~
Ax b ~ ¸~
Ax b
Pj0 = _ = Pj1
¡xj ¸ 0 xj ¸ 1
the lifted representation of PQ = conv(Pj0 [
Pj1) coincides with the polyhedron Mj (P ) (Balas,
Ceria, Cornuejols, '93)

and

P roj j (P ) = PQ = fx 2 IRn : ®x ¸ ¯ for all (®; ¯) 2

PQ¤ g, where P ¤ is the set of those (®; ¯) 2
Q
IRn+1 for which there exists vectors u; v 2
IRm+n+p and u0; v0 2 IR satisfying:

~ ¡ u0ej
® ¡ uA =0
® ~ ¡ v0ej = 0
¡ vA
u~
b =¯
v~
b + v0 = ¯
u; v ¸ 0
+ 26
+ +

Cutting planes

Using inequalities ®x ¸ ¯ valid for P roj j (P )

as cutting planes amounts to ensuring that
¤ , i.e., solving a linear
(®; ¯) is a ray of cone PQ
program:

Cut Generating LP

max a® + b¯
~ ¡ u0ej
® ¡ uA =0
® ~ ¡ v0ej = 0
¡ vA
u~
b =¯
v~
b + v0 = ¯
X
(ui + vi) + u0 + v0 = 1
u; v ¸ 0

² the normalization contraint avoids the prob-

lem being unbounded;

+ 27
+ +

A look at the matrix

u u0 v v0 a b
AT I -I ej -I 0
bT -1 T -1 0
AT I -I ej -I
=
0
bT -1 T 1 -1 0

m n n n

² number of variables: 2n + 5m + 3;

² number of constraints: 2(n + 1).

+ 28
+ +

Deepest cut

Let x
¹ be the current fractional point. The
objective function of CGLP is chosen so as
to maximize the amount ¯ ¡ ®x by which x
¹ is
cut o® (violation):

max ¡ x¹® + ¯
~ ¡ u0ej
® ¡ uA =0
® ~ ¡ v0ej = 0
¡ vA
u~
b =¯
v~
b + v0 = ¯
X
(ui + vi) + u0 + v0 = 1
u; v ¸ 0

+ 29
+ +

Size reduction

1. Lifting

² Working on a subspace of the frac-

tional variables (integer variables which
are not at their bounds) and continu-
ous positive (at x ¹) variables. Compu-
tationally cheaper (signi¯cantly !) and
essential for branch-and-cut;

² Ceria ('92) proved that we can solve

the problem in the subspace, and gen-
erate a solution to the full space by
using a simple formula.

2. Constraint selection

² dropping columns of CGLP associated

to (functional) constraints of P not
tight at x
¹ reduces signi¯cantly com-
puting times, even if it may yield weaker
cuts (Ceria, '92).

+ 30
+ +

Cut generating procedure

² RESTRICT the problem to a subspace

de¯ned from the current LP optimum,
and choose a disjunction;

² LIFT the disjunctive set to describe its

convex hull in a higher dimensional space;

² PROJECT the convex hull onto the orig-

inal (restricted) space, generating cuts;

² LIFT the cuts into the original full space;

² STRENGTHEN the lifted cuts.

Termination of the overall cutting planes for

general MIPs proved by Balas, Ceria and Cor-
nuejols ('93).

Computational advance: multiple cuts from

the current fractional solution.
+ 31
+ +

Computational achievements

² Lift-and-project based branch-and-cut vs.

Cplex

{ Balas, Ceria, Cornuejols ('96) proved

that the approach is computationally
practicable;

{ comparison with Cplex5.0 (cut-and-branch)

on 18 hard MIPLIB instances (Ceria,
Pataki, '98):
¤ time saving for 14 instances (for six
the gain was more than 4-fold);

¤ two instances solved in few minutes

by LP could not be solved by Cplex;

{ set1ch was solved in 28s. while Cplex

fails;

² In late nineties Cplex outperformed the

above results by improving other branch-
and-cut features;

+ 32
+ +

Computational achievements

² A computational breakthrough is due to

Zongao Gu (ISMP'2000), who deviced an
algorithm for generating cuts from 0 ¡
1 disjunctions without explicitly solving
the CGLP (theory in Balas and Perregard
2003). This started the integration of
disjunctive cutting planes into Cplex 8.0.

² The e±cacy of disjunctive cuts still looks

problem-dependent. In general, Cplex \ag-
gressive" separation mode is outperformed
by moderate (default) settings.

² \Seymour" is a very hard Set Covering

problem with 4944 rows and 1372 columns.
It was ¯rst solved to optimality with mas-
sive use of disjunctive cuts (Ceria, Ferris,
Linderoth, Pataki, Schmieta, 2001). Up
to now no other algorithms could solve
Seymour to optimality.

+ 33
+ +

Drawbacks

² cut density: both LP and CGLP sizes in-

crease;

² cut selection: deepest cut is often not the

right choice;

+ 34
+ +

Disjunctive cuts in combinatorial

optimization

Separation problem: given a class F of lin-

ear inequalities and a point x¤ 2 IRn, ¯nd an
inequality in F violated by x¤ or prove that
none exist.

In the classical polyhedral approach (Padberg,

'80s) facet inducing inequalities are used as
cutting planes.

² Good news: many facets of di®erent com-

binatorial problems can be obtained by
simple disjunctions. This yields new poly-
nomial separation algorithms (Letchford,
2000).

² Bad news: producing e®ective practical

implementations of separation algorithms
is challenging.

+ 35
+ +

Example

Stable set problem: given an undirected

graph G = (V; E), ¯nd a maximum size set
S µ V of pairwise non-adjacent vertices.

Let Stab(G) be the convex hull of incidence

vectors of stable sets of G. Consider an chord-
less cycle C induced by vertex set C, with jCj
odd. Then,
X jCj
xj · b c
j2C 2

is a facet of Stab(C) (odd hole inequality) 2

Odd hole inequalities can be obtained by 0-1

disjunctions.

+ 36
 5-hole
1
+ 1/2 +

5 1/2 1/2 2

1/2 1/2
4 3

look at the disjunction (x1 = 0) _ (x1 = 1):

¡x1 ¡ x2 ¸ ¡1 ¡x1 ¡ x2 ¸ ¡1
¤ ¡x2 ¡ x3 ¸ ¡1 ¡x2 ¡ x3 ¸ ¡1 ¤
¡x3 ¡ x4 ¸ ¡1 ¡x3 ¡ x4 ¸ ¡1 ¤
¤ ¡x4 ¡ x5 ¸ ¡1 ¡x4 ¡ x5 ¸ ¡1 ¤
¡x1 ¡ x5 ¸ ¡1 ¡x1 ¡ x5 ¸ ¡1
x1 ¸ 0 x1 ¸ 0
x2 ¸ 0 x2 ¸ 0
x3 ¸ 0 x3 ¸ 0 ¤
x4 ¸ 0 x4 ¸ 0 ¤
x5 ¸ 0 x5 ¸ 0
¡x1 ¸ ¡1 ¡x1 ¸ ¡1
¡x2 ¸ ¡1 ¡x2 ¸ ¡1
¡x3 ¸ ¡1 ¡x3 ¸ ¡1
¡x4 ¸ ¡1 ¡x4 ¸ ¡1
¡x5 ¸ ¡1 ¡x5 ¸ ¡1
¤ ¡x1 ¸ 0 x1 ¸ 1 ¤

summing up inequalities (¤) we obtain (from

both sides): ¡x1 ¡ x2 ¡ x3 ¡ x4 ¡ x5 ¸ ¡2

+ 37
+ +

Pitfall of Lift-and-project

1/2
1

1/2 5 2 1/2
6

1/2 4 3 1/2

¡x1 ¡ x2 ¡ x6 ¸ ¡1 ¡x1 ¡ x2 ¡ x6 ¸ ¡1 ¤
¤ } ¡x2 ¡ x3 ¡ x6 ¸ ¡1 ¡x2 ¡ x3 ¡ x6 ¸ ¡1 }
¡x3 ¡ x4 ¡ x6 ¸ ¡1 ¡x3 ¡ x4 ¡ x6 ¸ ¡1 ¤ }
¤ } ¡x4 ¡ x5 ¡ x6 ¸ ¡1 ¡x4 ¡ x5 ¡ x6 ¸ ¡1 }
¡x1 ¡ x5 ¡ x6 ¸ ¡1 ¡x1 ¡ x5 ¡ x6 ¸ ¡1 ¤
} x6 ¸ 0 x6 ¸ 0 ¤ }
¤ } ¡x1 ¸ 0 x1 ¸ 1 ¤ }

² summing up inequalities (¤) (from both sides):

¡x1 ¡ x2 ¡ x3 ¡ x4 ¡ x5 ¡ 2x6 ¸ ¡2 (wheel inequality)

² combining inequalities (}): ¡x1 ¡x2 ¡x3 ¡x4 ¡x5 ¸ ¡2

(odd hole inequality)

² the wheel inequality dominates the odd hole inequal-

ity, but they are violated by the same amount (!)

+ 38
+ +

Research directions

² Integrating LP and local cuts;

² integrating LP and branching (Avella, Ce-

ria, Rossi, '98);

² cut generation speed-up with general dis-

junctions;

² using LP to improve other families of cuts

for MIPs.

² deriving separation algorithms by disjunc-

tive interpretation (theory)

² connection to other projective methods

(Sherali and Adams, '90) (Lovasz and Schri-
jver, '91)

+ 39
+ +

References

1. E. Balas, Disjunctive Programming: properties of

the convex hull of feasible points, Discrete Applied
Mathematics, 89, 1998, 3-44.

2. E. Balas, Disjunctive Programming, Annals of Dis-

crete Mathematics, 5, 1979, 3 -51.

3. Balas, E., Ceria, S. and G. Cornuejols, A lift-and-

project cutting plane algorithm for mixed-integer
programs, Math. Prog., 58, 1993, 295 - 324.

4. Balas, E., Ceria, S. and G. Cornuejols, Mixed 0-1

programming by lift-and-project in a branch-and-
cut framework, Man. Sci., 42, 1996, 1229 - 1246.

5. Balas, E. and M. Perregard, Lift-and-Project for

Mixed 0-1 programming: recent progress, Dis-
crete Applied Mathematics, 123, 2002, 129 - 154.

6. Balas, E. and M. Perregard, A precise correspon-

dence between lift-and-project cuts, simple dis-
junctive cuts, and mixed integer gomory cuts for
0-1 programming, Math. Prog., 94, 2003, 221 -
245.

7. S. Ceria, Lift-and-Project Methods for Mixed 0-1

Programs, PhD dissertation in Industrial Admin-
istration, Graduate School of Industrial Adminis-
tration, Carnegie Mellon University, 1993.

+ 40