Linear Programming (Optimization)

Download Report

Transcript Linear Programming (Optimization)

II.2
Strong Valid Inequalities and Facets for
Structured Integer Programs
1
1. Introduction
 Strength of the valid inequalities is important (want facet or high dimensional
face). Use the structure of the problem to obtain strong valid inequalities.
 Node packing problem: 𝑆 = π‘₯ ∈ 𝐡𝑛 : π‘₯𝑖 + π‘₯𝑗 ≀ 1, βˆ€ 𝑖, 𝑗 ∈ 𝐸
dim(conv(𝑆)) = 𝑛 (𝑛 unit vector + 0 vector)
If 𝐢 βŠ† 𝑁 is a clique,
π‘—βˆˆπΆ π‘₯𝑗
If 𝐢 is a maximal clique,
≀ 1 is a valid inequality.
π‘—βˆˆπΆ π‘₯𝑗
≀ 1 defines a facet of conv(𝑆).
Pf) enough to show that there exist 𝑛 affinely independent feasible points
satisfying the clique constraint at equality.
Let 𝐢 = {1, … , π‘˜} be a maximal clique. For each node 𝑗 βˆ‰ 𝐢, βˆƒ 𝑙(𝑗) ∈ 𝐢
which is not connected to 𝑗 since 𝐢 is a maximal clique.
Consider the characteristic vectors of the packings, {1}, … , {π‘˜}, {π‘˜ + 1, 𝑙(π‘˜ +
1)}, … , {𝑛, 𝑙(𝑛)}.
Integer Programming 2013
2
2
𝐺:
2
𝐻:
6
3
6
3
1
5
4
5
4
 π‘₯ 1 = 12(0, 1, 1, 1, 1, 1) is an extreme point of the initial polytope (clique
constraints and nonnegativities).
To cut off π‘₯ 1 , consider odd hole (chordless cycle, odd cycle with no crossing
edges). Let 𝐻 βŠ† 𝑉, H induces an odd hole, then
π‘—βˆˆπ» π‘₯𝑗
≀ ( 𝐻 βˆ’ 1)/2 is a valid inequality. (π‘₯2 + π‘₯3 + π‘₯4 + π‘₯5 + π‘₯6 ≀ 2)
Moreover, it gives a facet of the convex hull of node packings for the
subgraph with node set 𝐻 (five linearly independent vectors satisfying the
inequality at equality). But, it does not give a facet of conv(𝑆) for 𝐺.
Integer Programming 2013
3
 Consider Ξ±π‘₯1 + (π‘₯2 + π‘₯3 + π‘₯4 + π‘₯5 + π‘₯6 ) ≀ 2.
Can we make  > 0 and large while the inequality remains valid?
When π‘₯1 = 0, it is valid for any  > 0.
When π‘₯1 = 1, 𝛼 ≀ 2 - (π‘₯2 + π‘₯3 + π‘₯4 + π‘₯5 + π‘₯6 ). But π‘₯1 = 1 implies π‘₯2 =
β‹― = π‘₯6 = 0, so 𝛼 ≀ 2.
Therefore, 2π‘₯1 + (π‘₯2 + π‘₯3 + π‘₯4 + π‘₯5 + π‘₯6 ) ≀ 2 is a valid inequality.
Moreover, (1, 0, 0, 0, 0, 0) vector in addition to the five vectors we already
have are linearly independent and satisfy the constraint at equality ⟹ facet.
 Lifting procedure: Systematic method to obtain facet (high dimensional face)
from a facet for the low dimensional restricted problem
Integer Programming 2013
4
 Prop 1.1: Suppose 𝑆 βŠ† 𝐡𝑛 , 𝑆 𝛿 = 𝑆⋂{π‘₯ ∈ 𝐡𝑛 : π‘₯1 = 𝛿}, 𝛿 = 0,1 , and
𝑛
𝑗=2 πœ‹π‘— π‘₯𝑗
1
≀ πœ‹0 (1.7) is valid for 𝑆 0 . If 𝑆1 = βˆ…, then π‘₯1 ≀ 0 is valid for S. If
𝑆 β‰  βˆ…, then 𝛼1 π‘₯1 + 𝑛𝑗=2 πœ‹π‘— π‘₯𝑗 ≀ πœ‹0 (1.8) is valid for 𝑆 for any 𝛼1 ≀ πœ‹0 βˆ’ 𝜁,
where 𝜁 = max{ 𝑛𝑗=2 πœ‹π‘— π‘₯𝑗 : π‘₯ ∈ 𝑆1 }.
Moreover, if 𝛼1 = πœ‹0 βˆ’ 𝜁 and (1.7) gives a face of dimension π‘˜ of conv(𝑆0),
then (1.8) gives a face of dimension π‘˜ + 1 of conv(𝑆). [If (1.7) facet of
conv(𝑆0), (1.8) facet of conv(𝑆).]
Pf) If π‘₯ ∈ 𝑆 0 ⟹ 𝛼1 π‘₯1 +
If π‘₯ ∈ 𝑆1 ⟹ 𝛼1 π‘₯1 +
𝑛
𝑗=2 πœ‹π‘— π‘₯𝑗
𝑛
𝑗=2 πœ‹π‘— π‘₯𝑗
≀ πœ‹0 valid.
= 𝛼1 +
𝑛
𝑗=2 πœ‹π‘— π‘₯𝑗
≀ 𝛼1 + 𝜁 ≀ πœ‹0 valid.
(1.7) gives a π‘˜ βˆ’dimensional face of conv(𝑆0) ⟹ βˆƒ π‘₯ 𝑖 ∈ 𝑆 0 such that π‘₯1𝑖 =
0, 𝑖 = 1, … , π‘˜ + 1. Also π‘₯ 𝑖 satisfies 𝛼1 π‘₯1 + 𝑛𝑗=2 πœ‹π‘— π‘₯𝑗 ≀ πœ‹0 at equality.
Let 𝜁 =
𝑛
βˆ—
𝑗=2 πœ‹π‘— π‘₯𝑗 ,
π‘₯ βˆ— ∈ 𝑆1
𝛼1 = πœ‹0 βˆ’ 𝜁 ⟹ π‘₯ βˆ— satisfies 𝛼1 π‘₯1 +
𝑆1
𝑛
𝑗=2 πœ‹π‘— π‘₯𝑗
= πœ‹0 and π‘₯1βˆ— = 1 since π‘₯ βˆ— ∈
⟹ π‘₯ βˆ— cannot be written as affine combination of {π‘₯ 1 , … , π‘₯ π‘˜+1 }, so the π‘˜ + 2
vectors {π‘₯ βˆ— , π‘₯ 1 , … , π‘₯ π‘˜+1 } are affinely independent.

Integer Programming 2013
5
 Lifting is also possible from 𝑆1 to S.
 Prop 1.2: Suppose (1.7) is valid for 𝑆1 . If 𝑆 0 = βˆ…, then π‘₯1 β‰₯1 is valid for 𝑆.
If 𝑆 0 β‰  βˆ…, then 𝛾1 π‘₯1 + 𝑛𝑗=2 πœ‹π‘— π‘₯𝑗 ≀ πœ‹0 + 𝛾1 (1.9) is valid for 𝑆 for any 𝛾1 β‰₯
𝜁 βˆ’ πœ‹0 , where 𝜁 = max{ 𝑛𝑗=2 πœ‹π‘— π‘₯𝑗 : π‘₯ ∈ 𝑆 0 }.
Moreover, if 𝛾1 = 𝜁 βˆ’ πœ‹0 and (1.7) gives a face of dimension π‘˜ of conv(𝑆1),
then (1.9) gives a face of dimension π‘˜ + 1 of conv(𝑆). [If (1.7) facet of
conv(𝑆0), (1.8) facet of conv(𝑆).]
 When 𝛼1 = πœ‹0 βˆ’ 𝜁 in Prop 1.1 or when 𝛾1 = 𝜁 βˆ’ πœ‹0 in Prop 1.2, we say that
the lifting is maximum.
 Lifting is done sequentially one variable at a time and the strength of the lifted
inequality depends on the sequence of the variables lifted.

Integer Programming 2013
6
 Prop 1.3: Let 𝑁\𝑁1 = {1, 2, … , 𝑑} and suppose when Prop 1.1 is applied
sequentially using maximum lifting in the order (𝑖1 , 𝑖2 , … , π‘–π‘˜βˆ’1 , π‘–π‘˜ , … , 𝑖𝑑 ), the
inequality π‘—βˆˆπ‘\𝑁1 𝛼𝑗 π‘₯𝑗 + π‘—βˆˆπ‘1 πœ‹π‘— π‘₯𝑗 ≀ πœ‹0 is obtained.
β€²
Then for any order (𝑖1β€² , 𝑖2β€² , … , π‘–π‘˜βˆ’1
, π‘–π‘˜β€² , … , 𝑖𝑑′ ) with 𝑖𝑗′ = 𝑖𝑗 for 𝑗 = 1, … , π‘˜ βˆ’ 1, the
resulting inequality π‘—βˆˆπ‘\𝑁1 𝛼𝑗′ π‘₯𝑗 + π‘—βˆˆπ‘1 πœ‹π‘— π‘₯𝑗 ≀ πœ‹0 obtained by maximum
lifting has π›Όπ‘–β€²π‘˜ ≀ π›Όπ‘–π‘˜ .
Pf) π‘–π‘˜ = 𝑖𝑠′ for some 𝑠 > π‘˜. Then
π›Όπ‘–π‘˜ = πœ‹0 βˆ’ max{
0 for 𝑗 > π‘˜}
= πœ‹0 βˆ’ max{
π‘˜βˆ’1
𝑗=1 𝛼𝑖𝑗 π‘₯𝑖𝑗
π‘˜βˆ’1
𝑗=1 𝛼𝑖𝑗 π‘₯𝑖𝑗
+
𝑛 : π‘₯ = 1 and π‘₯ =
πœ‹
π‘₯
∢
π‘₯
∈
𝑆
∩
{π‘₯
∈
𝐡
𝑗
𝑗
π‘–π‘˜
𝑖𝑗
π‘—βˆˆπ‘1
+
π‘—βˆˆπ‘1 πœ‹π‘— π‘₯𝑗
: π‘₯ ∈ 𝑆 ∩ {π‘₯ ∈ 𝐡𝑛 : π‘₯𝑖𝑠′ = 1 π‘Žπ‘›π‘‘ π‘₯𝑖 β€² =
+
π‘—βˆˆπ‘1 πœ‹π‘— π‘₯𝑗
+
𝑗
0 for π‘˜ ≀ 𝑗 ≀ 𝑑, 𝑗 β‰  𝑠}
β‰₯ πœ‹0 βˆ’ max{
π‘˜βˆ’1
𝑗=1 𝛼𝑖𝑗 π‘₯𝑖𝑗
π‘ βˆ’1 β€²
𝑗=π‘˜ 𝛼𝑖𝑗′ π‘₯𝑖𝑗′
: π‘₯ ∈ 𝑆 ∩ {π‘₯ ∈ 𝐡𝑛 : π‘₯𝑖𝑠′ =
1 π‘Žπ‘›π‘‘ π‘₯𝑖 β€² = 0 for 𝑗 > 𝑠}
𝑗
= π›Όπ‘–β€²π‘˜

Integer Programming 2013
7