IE 511: Integer Programming, Spring 2021 22 Apr, 2021

Lecture 25: Mixed Integer Cuts

Lecturer: Karthik Chandrasekaran Scribe: Karthik

Disclaimer: These notes have not been subjected to the usual scrutiny reserved for formal publi-
cations.

Today, we will learn how to obtain cuts for Mixed IPs.

Recall that mixed IPs (MIPs) have two kinds of variables: some variables are required to be integers
while the rest can take real values. The standard form of MIP is as follows:
MIP: max{cT x : Ax = b, x ≥ 0, x ∈ Zp × Rn−p }.
The feasible set of a MIP is known as a mixed-integer region (MIR).
Example 1. Consider the mixed-integer region S = {(x, y) : x, y ≥ 0, x + y ≥ 5/2, y ∈ Z}. This
set is plotted below. Note that the convex-hull of feasible points (i.e., convex-hull(S)) is different
from the convex-hull of integral points in S (i.e., convex-hull(S ∩ Z2 )).

Figure 25.1: Feasible region of a Mixed IP

The cutting plane approach naturally extends to MIPs with two changes: (1) the stopping criterion
now requires only certain variables to be integral and (2) we need an efficient procedure to generate
cuts—i.e., valid inequalities for the mixed-integer region that are violated by the current optimum.
For IPs, we saw that Gomory’s cut generation procedure is an efficient algorithm to generate
cuts/valid inequalities that are violated by the current optimum. How about MIPs?
Unfortunately, Gomory’s approach for generating cuts for IPs does not give valid cuts for MIPs.
We discuss this issue first. Recall that Gomory’s cuts for IPs are CG-cuts.

25-1
 Recap

CG-cut: wT x ≤ δ is valid for P and w is integral. If x is integral, then wT x ≤ bδc is valid

for PI .

Note that in a MIP, all variables of x are not integral. So we need an alternative approach to
obtain cuts/valid inequalities for the convex-hull of feasible solutions to a MIP. We will see such
an approach in this lecture. We start by understanding how to generate cuts for two-dimensional
feasible regions. We begin with the special case of regions defined by ≥ constraint.
Lemma 0.1. Let S ≥ := {(x, y) ∈ R × Z : x + y ≥ b, x ≥ 0} and f := b − bbc > 0. Then, the
inequality
x
+ y ≥ dbe
f
is valid for S ≥ .

Proof. Let (x̄, ȳ) ∈ S ≥ . We need to show that the point (x̄, ȳ) satisfies the inequality in the lemma.
If ȳ ≥ dbe, then x̄ ≥ 0 and hence, x̄ ≥ f (dbe − ȳ) (since the RHS is at most 0). Rewriting this last
inequality shows that (x̄, ȳ) satisfies the inequality in the lemma.
If ȳ < dbe, then x̄ ≥ b − ȳ = f + (bbc − ȳ) ≥ f + f (bbc − ȳ) = f (dbe − ȳ) (the last inequality holds
because bbc − ȳ ≥ 0 and f < 1). It implies that fx̄ + ȳ ≥ dbe.

Considering Example 1 above, we note that f = 1/2 and hence 2x + y ≥ 3 is valid for S.
Next, we address regions defined by ≤ constraint.
Corollary 0.1. Let S ≤ := {(x, y) ∈ R × Z : y ≤ b + x, x ≥ 0}. Suppose f := b − bbc > 0. Then,
the inequality
x
y ≤ bbc +
1−f
is valid for S ≤ .

Proof. We have y ≤ b + x iff x − y ≥ −b. Moreover, −b − b−bc = 1 − f . By Lemma 0.1,

x ≤
1−f − y ≥ d−be = −bbc is valid for S .

Note that when x = 0, we obtain a CG-cut/Gomory cut type inequality from the above lemma and
its corollary. Hence the above lemma and its corollary are generalizations of CG-cuts for mixed
integer sets.
Next, let us consider slightly more general mixed integer region. Let
S MIR := {(x, y) ∈ R × Z2 : x, y ≥ 0, a1 y1 + a2 y2 − x ≤ b}
with b ∈
/ Z.
Lemma 0.2 (Mixed Integer Rounding). Let f = b − bbc and fi = ai − bai c for i = 1, 2. If
f1 ≤ f ≤ f2 , then  
f2 − f x
ba1 cy1 + ba2 c + y2 ≤ bbc +
1−f 1−f
is valid for S MIR .

25-2
Proof. The inequality ba1 cy1 + da2 ey2 ≤ b + x + (1 − f2 )y2 is valid for S MIR (because y1 ≥ 0 and
a2 = da2 e − (1 − f2 )). By Corollary 0.1,

x + (1 − f2 )y2
ba1 cy1 + da2 ey2 ≤ bbc +
1−f

is valid for S MIR . It means that

 
1 − f2 x
ba1 cy1 + da2 e + y2 ≤ bbc +
1−f 1−f

is valid for S MIR . Using da2 e = ba2 c + 1 gives the lemma.

Remark. Note that if f1 ≤ f ≤ f2 does not hold, then we can use one of the previous lemmas to
get a cut.
We now have the ingredients to obtain a valid inequality for a mixed integer region that is violated
by an extreme point optimum that does not satisfy the integral requirements.

25.1 Gomory’s Mixed Integer Cut

Now that we have seen how to obtain valid inequalities for a mixed-integer region, let us see how
to generate a valid inequality for the mixed-integer region that is violated by the current extreme
point.
Consider the standard form of MIP (where x variables are real and y variables are required to be
integral):
max{cT1 x + cT2 y : A1 x + A2 y = b, x, y ≥ 0, y ∈ Zp , x ∈ Rn−p }.
Let (x̄, ȳ) be a basic feasible solution to the LP-relaxation of this MIP. If ȳ is integral, then we are
done with the cutting plane approach. Suppose ȳ is not integral.
/ Z. Consider the ith row of the optimal tableau (obtained
Let yi be the basic variable with y¯i ∈
with knowledge of the basis):
X X
yi + āij yj + āij xj = b̄i (25.1)
j∈N1 j∈N2

where (yi , y, x) ∈ Z × Zn1 × Rn2 , where N1 is the subset of non-basic variables among y and N2 is
the subset of non-basic variables among x. Recall that the non-basic variables are set to 0 in the
solution (x̄, ȳ). Consider the mixed-integer region
 
 X X 
S i := (yi , y, x) ∈ Z × Zn1 × Rn2 : yi , y, x ≥ 0, yi + āij yj + āij yj = b̄i .
 
j∈N1 j∈N2

Lemma 0.3. Let fj := āij − bāij c for all j ∈ N1 ∪ N2 and f0 := b̄i − bb̄i c. Then
X X f0 X X  f0 
fj yj + (1−fj ) yj + āij xj + āij xj ≥ f0 (25.2)
1 − f0 1 − f0
j∈N1 :fj ≤f0 j∈N1 :fj >f0 j∈N2 :āij >0 j∈N2 :āij <0

is a valid inequality for S i that is violated by (x̄, ȳ) .

25-3
Inequality (25.2) is known as Gomory Mixed Integer Cut.

Proof. Violation follows as LHS evaluated at (x̄, ȳ) is 0 while RHS > 0. We now show validity. The
MIR inequality for S i (by Lemma 0.2) is
  X  āij 
X X fj − f0
yi + bāij cyj + bāij c + yj + xj ≤ bb̄i c.
1 − f0 1 − f0
j∈N1 :fj ≤f0 j∈N1 :fj >f0 j∈N2 :āij <0

Substituting yi from equation (25.1) gives the inequality (25.2).

Example. Consider the MIP

max 4y − x
7y − 2x ≤ 14 (1)
x≤3 (2)
2y − 2x ≤ 3 (3)
y, x ≥ 0
y∈Z

We introduce slack variables s1 , s2 , s3 for inequalities (1), (2), and (3) and solve the LP-relaxation.
It turns out that the optimal solution for the LP-relaxation is the one in which x, y and s3 are
basic variables and the corresponding solution is x̄ = 3, ȳ = 20/7, s̄3 = 23/7 (recall that non-basic
variables are set to 0). The optimal tableax (from the basis) is as follows:
59 4 1
z = max − s1 − s2
7 7 7
1 2 20
y + s1 − s2 =
7 7 7
x + s2 = 3
2 10 23
− s1 + s2 + s3 =
7 7 7
We note that ȳ is fractional. Therefore, the first row gives the MIR cut y ≤ 2. Substituting for y
in the MIR cut using the first equation gives 71 s1 − 27 s2 ≥ 67 . We add this cut and re-solve the new
LP to obtain an LP optimal solution y = 2, x = 1/2. Since y is integral, this is an optimal solution
for the MIP itself.
Similar to IPs, Gomory’s cutting plane algorithm for MIPs can also be shown to terminate in a
finite numbers of steps using a careful choice of variable for cut generation and a careful choice of
LP solving algorithm.

25-4