3.6 Conclusions and future work
4.1.1 Multiple-row cuts
Let us assume to be given a mixed integer set
S = {x ∈ Rn: Ax = b, x ≥ 0, xj ∈ Z ∀j ∈ J}, (4.1) with A ∈ Qm×n and b ∈ Qm, and denote the continuous relaxation of S as P = {x ∈ Rn: Ax = b, x ≥ 0}. We assume for the sake of simplicity the set S to be nonempty.
Given a basis B ⊂ {1, . . . , n} corresponding to the vertex x∗ ∈ P \ S of the polyhe- dron P , the set S can be rewritten as
xB = x∗ B+ P j∈Nrjxj, x ≥ 0, xj ∈ Z, j ∈ J, (4.2) where N denotes the set of nonbasic variables. A valid relaxation of S can be obtained by dropping the nonnegative restrictions on all the basic variables and considering a subset Q (with |Q| = q) of rows of (4.2) associated with basic integer-constrained variables (i.e., a subset of variables xi with i ∈ B ∩ J), thus getting
(SQ) xi = fi+ P j∈N rjxj, i ∈ Q xj ≥ 0, j ∈ N xj ∈ Z, j ∈ J (4.3)
with fi = x∗i− bx∗ic for any i ∈ Q and fi > 0 for some i ∈ Q. In the following we assume w.l.o.g. Q = {1, . . . , q}.
Set SQ can be further relaxed by dropping all the integrality requirements on the nonbasic variables, thus getting a system of the form
x = f +Pkj=1rjs j, x ∈ Zq, s ∈ Rk +, (4.4)
where all the continuous variables have been renamed as s and |N | = k. Recall that f, r1, . . . , rk∈ Qq and f 6∈ Zq, and denote as R
f(r1, . . . , rk) the convex hull of all vectors
s ∈ Rk for which there exists x ∈ Rq such that (x, s) satisfies (4.4). Since S in non empty and all the data are rational entries, Rf(r1, . . . , rk) is a rational full-dimensional polyhedron: i.e., it is a rational polyhedron and its recession cone is Rk
+.
Borozan and Cornu´ejols [39] considered relaxing the k-dimensional space of variables s = (s1, . . . , sk) to an infinite-dimensional space, where the variables sr are defined for any r ∈ Qq, thus getting the following semi-infinite relaxation, related to Gomory and Johnson’s infinite group problem [100]:
x = f +Pr∈Qqrsr, x ∈ Zq,
s ≥ 0 with finite support
(4.5) where s = (sr)r∈Qq is said to have a finite support if it has a finite numbers of nonzero components. Let now Rf be the convex hull of all s ∈ RQq
for which there exists x ∈ Rq such that (x, s) satisfies (4.5). In this general context, they showed that any valid inequality for Rf that cuts off the infeasible solution s = 0 can be written in the
form X
r∈Qq
where ψ : Qq−→ Q∪{+∞} is said to be a valid function if the corresponding inequality (4.6) is valid for Rf. Then, they provided a strong correspondence between minimal valid inequalities and maximal lattice-free convex sets2 as follows. Let define a minimal valid function as a valid function ψ such that there is no other valid function ψ0 with ψ0(r) ≤ ψ(r) for all r ∈ Qq and ψ0(r) < ψ(r) for some r ∈ Qq. For any valid function ψ, consider the set Bψ = {x ∈ Qq: ψ(x − f ) ≤ 1} associated with ψ and denote as cl(B
ψ) the topological closure of Bψ in Rq. The following theorem holds:
Theorem 4.1 (Borozan and Cornu´ejols [39]). Let f ∈ Qq \ Zq. A minimal valid
function ψ for Rf is nonnegative, piecewise linear, positively homogeneous and convex. Furthermore the set cl(Bψ) is a full-dimensional maximal lattice-free convex set contai- ning f . Conversely, for any full-dimensional maximal lattice-free convex set B ⊂ Rq
containing f , there exists a minimal valid function ψ for Rf such that cl(Bψ) = B, and when f is in the interior of B, this function is unique.
In practice, the above theorem shows that any minimal valid inequality for Rf arises from a maximal lattice-free convex set B containing f . Further, any maximal lattice- free convex set B containing f in the interior lead to a unique minimal valid inequality. However, the latter results does not hold if f is on the boundary of B; i.e., any maximal lattice-free convex set containing f on the boundary leads, in the general case, to several minimal valid inequalities arising from different minimal valid functions ψ. Let denote any maximal lattice-free convex set B containing f on the boundary as a degenerate set. With the same meaning, let denote any minimal valid function ψ associated with a degenerate set as a degenerate function. The following question naturally arises from Theorem 4.1: should one consider degenerate sets/functions in practice?
The answer has been provided by Zambelli [176] for the finite dimensional case (i.e., for the set Rf(r1, . . . , rk)):
Theorem 4.2 (Zambelli [176]). Given a minimal valid inequality Pkj=1αjsj ≥ 1 for
Rf(r1, . . . , rk), there exists a nondegenerate minimal valid function ψ such that ψ(rj) =
αj for all j = 1, . . . , k.
Thus, from the practical point of view of generating cuts for the finite dimensional set Rf(r1, . . . , rk), one does not need to be concerned with the complications arising from degenerate functions, since any minimal valid inequality can be obtained from a nondegenerate minimal valid function.
A first computational investigation on the possibility of generating cuts from a set of multiple rows has been provided by Espinoza [75], who considered relaxing the simplex tableau of a general MIP as a set of the form (4.4). This is obtained by selecting a subset Q of rows associated with basic integer-constrained variables and by dropping nonnegative requirements on all the basic variables and integrality requirements on all the nonbasic variables. The computational results reported in [75] on a large set of MIPLIB [36] instances showed that cutting planes from multiple-row sets are effective
in practice to improve on the lower bound at the root node yielded by the “classical” single-row cuts.