Lecture 11: 0-1 Quadratic Program and
Lower Bounds
(3 units)
Outline
I Problem formulations
I Reformulation: Linearization & continuous relaxation
I Branch & Bound Method framework
I Simple bounds, LP bound and semidefinite relaxation
Problem formulation
I Standard form with 0-1 variables: (0-1QP) min
x∈{0,1}nf(x) = 1 2x
TQx+cTx whereQ is an n ×n symmetric matrix and c ∈Rn.
I Homogenous form:
(0-1QPh) min x∈{0,1}nx
TQx.
I Binary variables:
(BQP) min x∈{−1,1}nx
TQx+cTx. Transformation: xi = 12(yi+ 1).
I Homogenous form with binary variables: (BQPh) min
x∈{−1,1}nx TQx,
I (BQP)⇒ (BQPh) with
Q :=
µ
0 1 2cT 1
2c Q
¶
, x:= (±1,xT)∈ {−1,1}n+1.
Max-Cut problem
I Consider a graph G = (E,V) with vertex setV ={1, . . . ,n}
and edge set E ={ij |1≤i <j ≤n}. For every edge ij ∈E, there is an associated weight wij.
I Cut: For a given set S ⊆V, a cutδ(S) is the set of all edges with one endpoint inS and the other inV \S, and the weight of cut δ(S) isPij∈δ(S)wij.
I Max-Cut: find a cut δ(S) with maximum weight.
I Binary quadratic problem: (Max-Cut) max1
2
X
1≤i<j≤n
wij(1−xixj)
s.t. x ∈ {−1,1}n.
I x ∈ {−1,1}n ⇔ S ={i ∈V |x
i = 1} andV \S =
Linearization Method
I 0-1 Quadratic problem:
(P) min
x∈{0,1}n Q(x) = n
X
i=1
cixi +
X
1≤i<j≤n
qijxixj.
I For xi,xj ∈ {0,1},yij =xixj iff
yij = max{xi +xj−1,0}, yij ∈ {0,1}, or
yij = min{xi,xj}, yij ∈ {0,1}.
I (P) is equivalent to the following 0-1 linear integer program: min
x,y n
X
i=1
cixi +
X
(i,j)∈I+
qijyij +
X
(i,j)∈I−
qijyij
s.t. yij ≤xi, yij ≤xj, (i,j)∈I− (qij <0)
yij ≥xi +xj −1, (i,j)∈I+ (qij ≥0),
xi ∈ {0,1}, i = 1, . . . ,n,
yij ∈ {0,1}, 1≤i <j ≤n.
I A polynomially solvable case: qij ≤0: min
x,y n
X
i=1
cixi +
X
1≤i<j≤n
qijyij
s.t. yij ≤xi, 1≤i <j ≤n
yij ≤xj, 1≤i <j ≤n
xi,xj,yij ∈ {0,1}, 1≤i <j ≤n.
The constraint matrix is totally unimodular!
Continuous relaxation
I Consider the continuous relaxation of (P): ( ¯P) min
x∈[0,1]n Q(x) = n
X
i=1
cixi +
X
1≤i<j≤n
qijxixj. Then, at least one of the optimal solutions of ( ¯P) is located at an extreme point of [0,1]n. Thereforev(P) =v( ¯P).
I Unfortunately, the objective function of ( ¯P) isnonconvex and
nonconcave.
I Define
Qp(x) =
n
X
i=1
cixi+xTQx−pxTx+peTx.
Qp(x) =Q(x) for x ∈ {0,1}n. For largep,Qp(x) is a
concave function. Thus, (P) is equivalent to the concave minimization problem:
(Pc) min
x∈[0,1]n Qp(x)
Branch and Bound Framework
I Computing lower bound; I Branching on xi = 0 or xi = 1;
Basic lower bounding methods
I Simple lower bounds
I Continuous relaxation
I LP relaxation
I Lagrangian relaxation & SDP relaxation
Simple bounds
I Lower bound 1. Let
Q(x) =
n
X
i=1
cixi +12xTQx =
n
X
i=1 ˜
cixi +12xTQx˜ .
I An obvious lower bound of Q(x) over {0,1}n is:
LBs1 = 1
2 n
X
i=1
X
j6=i
min(qij,0) + n
X
i=1
min(ci +12qii,0).
I Lower bound 2. An improved simple lower bound is derived
by noting that: sincex ≥0, if ˜Qx ≥a, then 12xTQx˜ ≥ 1 2aTx.
I Let ˜Qi denote theith row of ˜Q. Then
ai = min x∈{0,1}n
˜
Qix=X
j6=i
I So
min x∈{0,1}n
1 2x
TQx˜ + ˜cTx
≥ min
x∈{0,1}n( 1 2a+ ˜c)
Tx
= n
X
i=1
min{ci +1 2qii+
1 2
X
j6=i
min(qij,0),0}
= LBs2.
It is easy to show thatLB2
s is better thanLBs1, i.e.,
LBs2 ≥LBs1.
Continuous relaxation
I Since ˜Q(x) = 12xT(Q+diag(u))x+ (c−1
2u)Tx takes the same value on {0,1}n asQ(x), it is natural to compute a lower bound via solving the continuous relaxation:
( ¯P) β(u) = min x∈[0,1]n
1 2x
T(Q+diag(u))x+ (c−1 2u)
Tx.
I Some observations:
I Ifui’s are large enough, then ˜Q=Q+diag(u) will be
diagonally dominant (thus positive definite).
I u=λmine is an obvious choice to make ˜Q positive semidefinite
(but not necessarily the best one).
I The optimal solution to ( ¯P) tends to (1
2,12, . . . ,12)T asui’s are increased.
I Consider a small example of (P) where
Q =
µ
1 −3
−3 −1
¶ , c =
µ
1 1
¶ .
I For this example, we havex∗ = (1,1)T with Q(x∗) =−1.
The two simple bounds for this problem are: LBs1=−3 and
LB2
s =−1.
I The eigenvalues of Q is (−2,4).
0 0.2
0.4 0.6
0.8 1
0 0.2 0.4 0.6 0.8 1 −1 −0.5 0 0.5 1 1.5
0 0.2
0.4 0.6
0.8 1
0 0.2 0.4 0.6 0.8 1 −1.5 −1 −0.5 0 0.5 1 1.5
Figure: u=λmine,xu = (0.8604,1)T,β(u) =−1.0406.
0 0.2
0.4 0.6
0.8 1
0 0.2 0.4 0.6 0.8 1 −5 −4 −3 −2 −1 0 1 2
0 0.2
0.4 0.6
0.8 1
0 0.2 0.4 0.6 0.8 1 −25 −20 −15 −10 −5 0 5
Figure: u= 100e,xu= (0.5003,0.5101)T,β(u) =−24.7551.
I Another way of choosing u is to find au∗ such that
β(u∗) = max{β(u)|(Q−diag(u))º0, u ∈Rn}.
I The above problem is equivalent to a semidefinite quadratic
LP relaxation
I The continuous relaxation of the 0-1 linearized problem is a linear program (yij =xixj):
min x,y
X
1≤i<j≤n
qijyij +
n
X
i=1
(ci +12qii)xi
s.t. yij ≤xi, yij ≤xj, 1≤i <j ≤n, qij <0,
xi +xj −1≤yij, 1≤i <j ≤n, qij >0,
0≤xi ≤1, i = 1, . . . ,n,
yij ≥0, 1≤i <j ≤n.
SDP relaxation
I ConsiderQ(x) = 1
2xTQx+cT.
I The Lagrangian dual is:
v(D)
= max
λ∈Rnd(λ) = max
λ∈Rnxinf∈Rn{ 1 2x
T[Q+ 2diag(λ)]x+cTx−eTλ}
= max
(λ,τ)∈Rn+1 −τ
s.t. 1
2x
T[Q+ 2diag(λ)]x+cTx−eTλ≥ −τ
I 1
2xT[Q+ 2diag(λ)]x+cTx−eTλ≥ −τ, x ∈Rn ⇔ (xT,t)
µ
Q+ 2diag(λ) c
cT 2τ −2eTλ
¶ µ
x t
¶ ≥0 (t,x)∈Rn+1,
⇔ µ
Q+ 2diag(λ) c
cT 2τ−2eTλ
¶ º0,
Variable fixation
I Let x∗ denote the optimal solution of (P). Let
li =ci +
1 2qii+
X
j6=i
min(0,qij),
ui =ci +1
2qii +
X
j6=i
max(0,qij).
(i) Ifli >0 thenx∗
i = 0;
Fixing variable by outerbox
I Equivalent perturbed problem: (Pµ) min
x∈{0,1}nfµ(x) = 1 2x
T(Q+µI)x+ (c−1 2µe)
Tx,
whereµ≥0,I is then×n identity matrix and
e = (1, . . . ,1)T.
The shape of contour changes when µchanges. The
properties of (Pµ), e.g., the conditional number, change when µ changes.
−2 −1 0 1 2 3 −4
−3 −2 −1 0 1
x0(µ)
µ=0
µ=1
µ=3 (1,1)
(1,0) (0,1)
(0,0)
µ=5
µ=10 x
2
x
I Ellipsoid contour: Let x0 ∈ {0,1}n.
Eµ={x|fµ(x) =f(x0).
I Let the minimum box that contains ellipsoid Eµbe B = [a,b],
wherea=x0−s,b =x0+s. Then
I Let x∗ be an optimal solution to (P). Then
(i) Ifbi <1 for some i, then x∗
i = 0;
(ii) Ifai >0 for some i, then xi∗= 1.
−1.5 −1 −0.5 0 0.5 1 1.5 2 −1.5
−1 −0.5 0 0.5 1 1.5
(0,0)
Eµ(˜v)
x0(µ)
(0,1) (1,1)
Bµ(˜v)