a constructive characterization of the split closure of a mixed integer linear program
TRANSCRIPT
![Page 1: A Constructive Characterization of the Split Closure of a Mixed Integer Linear Program](https://reader031.vdocuments.net/reader031/viewer/2022021710/620b195c5237bc46f9609095/html5/thumbnails/1.jpg)
Introduction Characterization Lattices Polyhedrality
A Constructive Characterization of the SplitClosure of a Mixed Integer Linear Program
Juan Pablo Vielma
School of Industrial and Systems EngineeringGeorgia Institute of Technology
19th International Symposium on MathematicalProgramming, 2006
![Page 2: A Constructive Characterization of the Split Closure of a Mixed Integer Linear Program](https://reader031.vdocuments.net/reader031/viewer/2022021710/620b195c5237bc46f9609095/html5/thumbnails/2.jpg)
Introduction Characterization Lattices Polyhedrality
Outline
1 Introduction
2 Characterization
3 Lattices
4 Polyhedrality
![Page 3: A Constructive Characterization of the Split Closure of a Mixed Integer Linear Program](https://reader031.vdocuments.net/reader031/viewer/2022021710/620b195c5237bc46f9609095/html5/thumbnails/3.jpg)
Introduction Characterization Lattices Polyhedrality
History and Motivation
History:Split Cuts were introduced by [Cook, et. al. 1990]. Specialcase of Balas’s Disjunctive Cuts. “Equivalent” IntersectionCuts, Mixed Integer Gomory Cuts and MIR Cuts.The Split Closure is obtained by applying all split cuts.Split Closure is a polyhedron[Cook, et. al. 1990, Andersen, et. al. 2005].Non-constructive proofs.The Split Cloure has recently been studied by [Balas andSaxena, 2005] and by [Dash et. al. 2005].
Motivation of Constructive Characterization:Algorithm to generate Split Closure? (Naive).Helps understand Split Cuts better.
![Page 4: A Constructive Characterization of the Split Closure of a Mixed Integer Linear Program](https://reader031.vdocuments.net/reader031/viewer/2022021710/620b195c5237bc46f9609095/html5/thumbnails/4.jpg)
Introduction Characterization Lattices Polyhedrality
History and Motivation
History:Split Cuts were introduced by [Cook, et. al. 1990]. Specialcase of Balas’s Disjunctive Cuts. “Equivalent” IntersectionCuts, Mixed Integer Gomory Cuts and MIR Cuts.The Split Closure is obtained by applying all split cuts.Split Closure is a polyhedron[Cook, et. al. 1990, Andersen, et. al. 2005].Non-constructive proofs.The Split Cloure has recently been studied by [Balas andSaxena, 2005] and by [Dash et. al. 2005].
Motivation of Constructive Characterization:Algorithm to generate Split Closure? (Naive).Helps understand Split Cuts better.
![Page 5: A Constructive Characterization of the Split Closure of a Mixed Integer Linear Program](https://reader031.vdocuments.net/reader031/viewer/2022021710/620b195c5237bc46f9609095/html5/thumbnails/5.jpg)
Introduction Characterization Lattices Polyhedrality
Feasible Set of a (Mixed) Integer Linear Program andNatural Relaxations
Feasible set:
P := {x ∈ Rn : aTi x ≤ bi ∀ i ∈ M}
PI := {x ∈ P : xj ∈ Z ∀ j ∈ NI} forNI ⊆ {1, . . . , n}
P
![Page 6: A Constructive Characterization of the Split Closure of a Mixed Integer Linear Program](https://reader031.vdocuments.net/reader031/viewer/2022021710/620b195c5237bc46f9609095/html5/thumbnails/6.jpg)
Introduction Characterization Lattices Polyhedrality
Feasible Set of a (Mixed) Integer Linear Program andNatural Relaxations
Feasible set:
P := {x ∈ Rn : aTi x ≤ bi ∀ i ∈ M}
PI := {x ∈ P : xj ∈ Z ∀ j ∈ NI} forNI ⊆ {1, . . . , n}
![Page 7: A Constructive Characterization of the Split Closure of a Mixed Integer Linear Program](https://reader031.vdocuments.net/reader031/viewer/2022021710/620b195c5237bc46f9609095/html5/thumbnails/7.jpg)
Introduction Characterization Lattices Polyhedrality
Feasible Set of a (Mixed) Integer Linear Program andNatural Relaxations
Feasible set:
P := {x ∈ Rn : aTi x ≤ bi ∀ i ∈ M}
PI := {x ∈ P : xj ∈ Z ∀ j ∈ NI} forNI ⊆ {1, . . . , n}
Relaxations:
P, LP Relaxation
P(B) := {x ∈ Rn : aTi x ≤ bi ∀ i ∈ B} for
B ∈ B := {B ⊆ M : |B| = n, {ai}i∈B l.i.}Basic or Conic Relaxation
x(B) unique solution to aTi x = bi ∀ i ∈ B
P
![Page 8: A Constructive Characterization of the Split Closure of a Mixed Integer Linear Program](https://reader031.vdocuments.net/reader031/viewer/2022021710/620b195c5237bc46f9609095/html5/thumbnails/8.jpg)
Introduction Characterization Lattices Polyhedrality
Feasible Set of a (Mixed) Integer Linear Program andNatural Relaxations
Feasible set:
P := {x ∈ Rn : aTi x ≤ bi ∀ i ∈ M}
PI := {x ∈ P : xj ∈ Z ∀ j ∈ NI} forNI ⊆ {1, . . . , n}
Relaxations:
P, LP Relaxation
P(B) := {x ∈ Rn : aTi x ≤ bi ∀ i ∈ B} for
B ∈ B := {B ⊆ M : |B| = n, {ai}i∈B l.i.}Basic or Conic Relaxation
x(B) unique solution to aTi x = bi ∀ i ∈ B
P(B)
![Page 9: A Constructive Characterization of the Split Closure of a Mixed Integer Linear Program](https://reader031.vdocuments.net/reader031/viewer/2022021710/620b195c5237bc46f9609095/html5/thumbnails/9.jpg)
Introduction Characterization Lattices Polyhedrality
Feasible Set of a (Mixed) Integer Linear Program andNatural Relaxations
Feasible set:
P := {x ∈ Rn : aTi x ≤ bi ∀ i ∈ M}
PI := {x ∈ P : xj ∈ Z ∀ j ∈ NI} forNI ⊆ {1, . . . , n}
Relaxations:
P, LP Relaxation
P(B) := {x ∈ Rn : aTi x ≤ bi ∀ i ∈ B} for
B ∈ B := {B ⊆ M : |B| = n, {ai}i∈B l.i.}Basic or Conic Relaxation
x(B) unique solution to aTi x = bi ∀ i ∈ B
P(B)
x(B)
![Page 10: A Constructive Characterization of the Split Closure of a Mixed Integer Linear Program](https://reader031.vdocuments.net/reader031/viewer/2022021710/620b195c5237bc46f9609095/html5/thumbnails/10.jpg)
Introduction Characterization Lattices Polyhedrality
Split Cuts are Constructed from Valid SplitDisjunctions
For (π, π0) ∈ Zn+1 let:
FlD(π,π0)
:= {x ∈ Rn : πT x ≤ π0}
FgD(π,π0)
:= {x ∈ Rn : πT x ≥ π0 + 1}
FD(π,π0) := FlD(π,π0)
∪ FgD(π,π0)
π0 π0 + 1
π
Fl Fg
![Page 11: A Constructive Characterization of the Split Closure of a Mixed Integer Linear Program](https://reader031.vdocuments.net/reader031/viewer/2022021710/620b195c5237bc46f9609095/html5/thumbnails/11.jpg)
Introduction Characterization Lattices Polyhedrality
Split Cuts are Constructed from Valid SplitDisjunctions
For (π, π0) ∈ Zn+1 let:
FlD(π,π0)
:= {x ∈ Rn : πT x ≤ π0}
FgD(π,π0)
:= {x ∈ Rn : πT x ≥ π0 + 1}
FD(π,π0) := FlD(π,π0)
∪ FgD(π,π0)
A split cut for D(π, π0) and P is an inequalityvalid for:
P ∩ FD(π,π0)
conv(P ∩ FD(π,π0))
![Page 12: A Constructive Characterization of the Split Closure of a Mixed Integer Linear Program](https://reader031.vdocuments.net/reader031/viewer/2022021710/620b195c5237bc46f9609095/html5/thumbnails/12.jpg)
Introduction Characterization Lattices Polyhedrality
Split Cuts are Constructed from Valid SplitDisjunctions
For (π, π0) ∈ Zn+1 let:
FlD(π,π0)
:= {x ∈ Rn : πT x ≤ π0}
FgD(π,π0)
:= {x ∈ Rn : πT x ≥ π0 + 1}
FD(π,π0) := FlD(π,π0)
∪ FgD(π,π0)
A split cut for D(π, π0) and P is an inequalityvalid for:
P ∩ FD(π,π0)
conv(P ∩ FD(π,π0))
![Page 13: A Constructive Characterization of the Split Closure of a Mixed Integer Linear Program](https://reader031.vdocuments.net/reader031/viewer/2022021710/620b195c5237bc46f9609095/html5/thumbnails/13.jpg)
Introduction Characterization Lattices Polyhedrality
Split Cuts are Constructed from Valid SplitDisjunctions
For (π, π0) ∈ Zn+1 let:
FlD(π,π0)
:= {x ∈ Rn : πT x ≤ π0}
FgD(π,π0)
:= {x ∈ Rn : πT x ≥ π0 + 1}
FD(π,π0) := FlD(π,π0)
∪ FgD(π,π0)
A split cut for D(π, π0) and P is an inequalityvalid for:
P ∩ FD(π,π0)
conv(P ∩ FD(π,π0))
![Page 14: A Constructive Characterization of the Split Closure of a Mixed Integer Linear Program](https://reader031.vdocuments.net/reader031/viewer/2022021710/620b195c5237bc46f9609095/html5/thumbnails/14.jpg)
Introduction Characterization Lattices Polyhedrality
Valid Split Disjunctions don’t Cut Integer FeasiblePoints
For fixed NI we are interested in D(π, π0)such that, for any P:
PI ⊆ FD(π,π0) ( Rn
x1
x2
NI = {1}
![Page 15: A Constructive Characterization of the Split Closure of a Mixed Integer Linear Program](https://reader031.vdocuments.net/reader031/viewer/2022021710/620b195c5237bc46f9609095/html5/thumbnails/15.jpg)
Introduction Characterization Lattices Polyhedrality
Valid Split Disjunctions don’t Cut Integer FeasiblePoints
For fixed NI we are interested in D(π, π0)such that, for any P:
PI ⊆ FD(π,π0) ( Rn
so we study
Π(NI) := {(π, π0) ∈ (Zn \ {0}) × Z :πj = 0, j /∈ NI}
x1
x2
NI = {1}
![Page 16: A Constructive Characterization of the Split Closure of a Mixed Integer Linear Program](https://reader031.vdocuments.net/reader031/viewer/2022021710/620b195c5237bc46f9609095/html5/thumbnails/16.jpg)
Introduction Characterization Lattices Polyhedrality
The Split Closure is the Polyhedron Formed by AllSplit Cuts
The split closure [Cook, et. al. 1990] of PI is
SC :=⋂
(π,π0)∈Π(NI )
conv(P ∩ FD(π,π0)).
Theorem
[Cook, et. al. 1990] SC is a polyhedron
![Page 17: A Constructive Characterization of the Split Closure of a Mixed Integer Linear Program](https://reader031.vdocuments.net/reader031/viewer/2022021710/620b195c5237bc46f9609095/html5/thumbnails/17.jpg)
Introduction Characterization Lattices Polyhedrality
Sufficient to Study Split Cuts for Basic Relaxations
For B ∈ B let
SC(B) :=⋂
(π,π0)∈Π(NI )
conv(P(B) ∩ FD(π,π0)).
![Page 18: A Constructive Characterization of the Split Closure of a Mixed Integer Linear Program](https://reader031.vdocuments.net/reader031/viewer/2022021710/620b195c5237bc46f9609095/html5/thumbnails/18.jpg)
Introduction Characterization Lattices Polyhedrality
Sufficient to Study Split Cuts for Basic Relaxations
For B ∈ B let
SC(B) :=⋂
(π,π0)∈Π(NI )
conv(P(B) ∩ FD(π,π0)).
Theorem
[Andersen, et. al. 2005] SC =⋂
B∈BSC(B)
Theorem
[Andersen, et. al. 2005] SC(B) is a polyhedron for all B ∈ B.Hence SC is a polyhedron.
Let P = P(B) = {x ∈ Rn : Bx ≤ b}, for B ∈ Qn×n,rank(B) = n
![Page 19: A Constructive Characterization of the Split Closure of a Mixed Integer Linear Program](https://reader031.vdocuments.net/reader031/viewer/2022021710/620b195c5237bc46f9609095/html5/thumbnails/19.jpg)
Introduction Characterization Lattices Polyhedrality
Proposition
[Andersen, et. al. 2005, Balas and Perregaard, 2003,Caprara and Letchford, 2003] All non-dominated validinequalities for conv(P ∩ FD(π,π0)) are of the form δT x ≤ δ0 where
δ = BTµl + µl0π = BTµg − µg
0π
δ0 = bTµl + µl0π0 = bTµg − µg
0(π0 + 1)
for µl0, µ
g0 ∈ R+ and µl, µg ∈ Rm
+ solutions to
BTµg − BTµl = π
bTµg − bTµl = π0 + µg0
µl0 + µg
0 = 1, µg0 ∈ (0, 1), µl
i · µgi = 0
![Page 20: A Constructive Characterization of the Split Closure of a Mixed Integer Linear Program](https://reader031.vdocuments.net/reader031/viewer/2022021710/620b195c5237bc46f9609095/html5/thumbnails/20.jpg)
Introduction Characterization Lattices Polyhedrality
Proposition
conv(P ∩ FD(π,π0)) = {x ∈ P : δT x ≤ δ0}
where δ(µ)T x ≤ δ0(µ) is defined equivalent to
(µ−)T(Bx − b) + (1 − f (µT b))(µT Bx − ⌊µTb⌋) ≤ 0
for µ unique solution (if it exists) to
BTµ = π µ ∈ Rr
µTb /∈ Z π0 = ⌊µTb⌋
(y− = max{−y, 0} , f (y) = y − ⌊y⌋ and operations over vectorsare componentwise)
![Page 21: A Constructive Characterization of the Split Closure of a Mixed Integer Linear Program](https://reader031.vdocuments.net/reader031/viewer/2022021710/620b195c5237bc46f9609095/html5/thumbnails/21.jpg)
Introduction Characterization Lattices Polyhedrality
What Multipliers Induce Valid Split Disjunctions?
We haveΠ(NI) := {(π, π0) ∈ (Zn \ {0}) × Z : πj = 0, j /∈ NI} and
BTµ = π µ ∈ Rr
µTb /∈ Z π0 = ⌊µTb⌋
Let B = [BIBC] for BI ∈ Rn×|NI | and BC ∈ Rn×(n−|NI |)
corresponding to the integer and continuous variables ofPI. Multipliers that induce valid split disjunctions are
L(B) := {µ ∈ Rn : BITµ ∈ Z|NI |, BC
Tµ = 0}
![Page 22: A Constructive Characterization of the Split Closure of a Mixed Integer Linear Program](https://reader031.vdocuments.net/reader031/viewer/2022021710/620b195c5237bc46f9609095/html5/thumbnails/22.jpg)
Introduction Characterization Lattices Polyhedrality
What Multipliers Induce Valid Split Disjunctions?
We haveΠ(NI) := {(π, π0) ∈ (Zn \ {0}) × Z : πj = 0, j /∈ NI} and
BTµ = π µ ∈ Rr
µTb /∈ Z π0 = ⌊µTb⌋
Let B = [BIBC] for BI ∈ Rn×|NI | and BC ∈ Rn×(n−|NI |)
corresponding to the integer and continuous variables ofPI. Multipliers that induce valid split disjunctions are
L(B) := {µ ∈ Rn : BITµ ∈ Z|NI |, BC
Tµ = 0}
![Page 23: A Constructive Characterization of the Split Closure of a Mixed Integer Linear Program](https://reader031.vdocuments.net/reader031/viewer/2022021710/620b195c5237bc46f9609095/html5/thumbnails/23.jpg)
Introduction Characterization Lattices Polyhedrality
Valid Split Disjunctions are Related to Integer Lattices
For {vi}ri=1 ⊆ Rn l.i. a lattice is
L := {µ ∈ Rn : µ =
r∑
i=1
kivi ki ∈ Z}
L(B) is a lattice,
⌈µ−⌉T(Bx−b)+(1−f (µT b))(µT Bx−⌊µTb⌋) ≤ 0
is valid for PI and cuts x(B).[Köppe and Weismantel, 2004].
Every µ ∈ L(B) s.t. µTb /∈ Z induces a validsplit disjunction.[Bertsimas and Weismantel, 2005].
v1
v2
v1 + 2v2
![Page 24: A Constructive Characterization of the Split Closure of a Mixed Integer Linear Program](https://reader031.vdocuments.net/reader031/viewer/2022021710/620b195c5237bc46f9609095/html5/thumbnails/24.jpg)
Introduction Characterization Lattices Polyhedrality
Valid Split Disjunctions are Related to Integer Lattices
For {vi}ri=1 ⊆ Rn l.i. a lattice is
L := {µ ∈ Rn : µ =
r∑
i=1
kivi ki ∈ Z}
L(B) is a lattice,
⌈µ−⌉T(Bx−b)+(1−f (µT b))(µT Bx−⌊µTb⌋) ≤ 0
is valid for PI and cuts x(B).[Köppe and Weismantel, 2004].
Every µ ∈ L(B) s.t. µTb /∈ Z induces a validsplit disjunction.[Bertsimas and Weismantel, 2005].
v1
v2
v1 + 2v2
![Page 25: A Constructive Characterization of the Split Closure of a Mixed Integer Linear Program](https://reader031.vdocuments.net/reader031/viewer/2022021710/620b195c5237bc46f9609095/html5/thumbnails/25.jpg)
Introduction Characterization Lattices Polyhedrality
Valid Split Disjunctions are Related to Integer Lattices
For {vi}ri=1 ⊆ Rn l.i. a lattice is
L := {µ ∈ Rn : µ =
r∑
i=1
kivi ki ∈ Z}
L(B) is a lattice,
⌈µ−⌉T(Bx−b)+(1−f (µT b))(µT Bx−⌊µTb⌋) ≤ 0
is valid for PI and cuts x(B).[Köppe and Weismantel, 2004].
Every µ ∈ L(B) s.t. µTb /∈ Z induces a validsplit disjunction.[Bertsimas and Weismantel, 2005].
v1
v2
v1 + 2v2
![Page 26: A Constructive Characterization of the Split Closure of a Mixed Integer Linear Program](https://reader031.vdocuments.net/reader031/viewer/2022021710/620b195c5237bc46f9609095/html5/thumbnails/26.jpg)
Introduction Characterization Lattices Polyhedrality
Proposition
SC(B) =⋂
µ∈L(B)µT b/∈Z{x ∈ P(B) : δ(µ)T x ≤ δ0(µ)}.
![Page 27: A Constructive Characterization of the Split Closure of a Mixed Integer Linear Program](https://reader031.vdocuments.net/reader031/viewer/2022021710/620b195c5237bc46f9609095/html5/thumbnails/27.jpg)
Introduction Characterization Lattices Polyhedrality
Proposition
SC(B) =⋂
µ∈L(B)µT b/∈Z{x ∈ P(B) : δ(µ)T x ≤ δ0(µ)}.
Proposition
For µ ∈ L(B) s.t µTb /∈ Z split cut
(µ−)T(Bx − b) + (1 − f (µT b))(µT Bx − ⌊µTb⌋) ≤ 0
dominates
⌈µ−⌉T(Bx − b) + (1 − f (µTb))(µT Bx − ⌊µT b⌋) ≤ 0
![Page 28: A Constructive Characterization of the Split Closure of a Mixed Integer Linear Program](https://reader031.vdocuments.net/reader031/viewer/2022021710/620b195c5237bc46f9609095/html5/thumbnails/28.jpg)
Introduction Characterization Lattices Polyhedrality
Studying L(B) in Each Orthant Decomposes SC(B) tothe Intersection of a Finite Number of Sets
For σ ∈ {0, 1}n let
L(B, σ) := {µ ∈ L(B) : (−1)σiµi ≥ 0, ∀ i ∈ {1, . . . , n}}
so thatSC(B) =
⋂
σ∈{0,1}n
SC(B, σ)
where
SC(B, σ) =⋂
µ∈L(B,σ)µT b/∈Z {x ∈ P(B) : δ(µ)T x ≤ δ0(µ)}
![Page 29: A Constructive Characterization of the Split Closure of a Mixed Integer Linear Program](https://reader031.vdocuments.net/reader031/viewer/2022021710/620b195c5237bc46f9609095/html5/thumbnails/29.jpg)
Introduction Characterization Lattices Polyhedrality
Studying L(B, σ) Allows Detecting Dominated Cuts
Lemma
Let σ ∈ {0, 1}n and let µ ∈ L(B, σ) with µ = α + β forα, β ∈ L(B, σ) such that βT b ∈ Z. Then δ(µ)T x ≤ δ0(µ) isdominated by δ(α)T x ≤ δ0(α) in P(B).
Proof.
Uses the fact that for α, β in the same orthant|α + β| = |α| + |β|.
![Page 30: A Constructive Characterization of the Split Closure of a Mixed Integer Linear Program](https://reader031.vdocuments.net/reader031/viewer/2022021710/620b195c5237bc46f9609095/html5/thumbnails/30.jpg)
Introduction Characterization Lattices Polyhedrality
A Finite Integral Generating Set (FIGS) of L(B, σ)Induces a Finite Subset of L(B, σ)
Let {vi}i∈V(σ) ⊆ L(B, σ) be a (FIGS), i.e. a finite set suchthat
L(B, σ) = {µ ∈ Rr : µ =∑
i∈V(σ)
kivi ki ∈ Z+}
We want µTb /∈ Z, so for i ∈ V(σ) let
mi = min{m ∈ Z+ \ {0} : m bTvi ∈ Z}and define the following finite subset of L(B, σ).
L0(B, σ) := {µ ∈ L(B, σ) : µ =∑
i∈V(σ)
rivi, ri ∈ {0, . . . , mi−1}}
![Page 31: A Constructive Characterization of the Split Closure of a Mixed Integer Linear Program](https://reader031.vdocuments.net/reader031/viewer/2022021710/620b195c5237bc46f9609095/html5/thumbnails/31.jpg)
Introduction Characterization Lattices Polyhedrality
A Finite Integral Generating Set (FIGS) of L(B, σ)Induces a Finite Subset of L(B, σ)
Let {vi}i∈V(σ) ⊆ L(B, σ) be a (FIGS), i.e. a finite set suchthat
L(B, σ) = {µ ∈ Rr : µ =∑
i∈V(σ)
kivi ki ∈ Z+}
We want µTb /∈ Z, so for i ∈ V(σ) let
mi = min{m ∈ Z+ \ {0} : m bTvi ∈ Z}and define the following finite subset of L(B, σ).
L0(B, σ) := {µ ∈ L(B, σ) : µ =∑
i∈V(σ)
rivi, ri ∈ {0, . . . , mi−1}}
![Page 32: A Constructive Characterization of the Split Closure of a Mixed Integer Linear Program](https://reader031.vdocuments.net/reader031/viewer/2022021710/620b195c5237bc46f9609095/html5/thumbnails/32.jpg)
Introduction Characterization Lattices Polyhedrality
Proving the Polyhedrality of SC(B, σ) Yields thePolyhedrality of SC
Theorem
SC(B, σ) the polyhedron given by
SC(B, σ) =⋂
µ∈L0(B,σ)µT b/∈Z {x ∈ P(B) : δ(µ)T x ≤ δ0(µ)}
Corollary
SC(B) is a polyhedron for all B ∈ B. SC is a polyhedron.
![Page 33: A Constructive Characterization of the Split Closure of a Mixed Integer Linear Program](https://reader031.vdocuments.net/reader031/viewer/2022021710/620b195c5237bc46f9609095/html5/thumbnails/33.jpg)
Introduction Characterization Lattices Polyhedrality
Proof Idea.
Goal: For µ ∈ L(B, σ), δ(µ)T x ≤ δ0(µ) is dominated byδ(α)T x ≤ δ0(α) for some α ∈ L0(B, σ).How:
For µ ∈ L(B, σ) show that µ = α + β for α, β such that:α ∈ L
0(B, σ), β ∈ L(B, σ)βT b ∈ Z
Use Lemma.
![Page 34: A Constructive Characterization of the Split Closure of a Mixed Integer Linear Program](https://reader031.vdocuments.net/reader031/viewer/2022021710/620b195c5237bc46f9609095/html5/thumbnails/34.jpg)
Introduction Characterization Lattices Polyhedrality
Proof Idea.
Goal: For µ ∈ L(B, σ), δ(µ)T x ≤ δ0(µ) is dominated byδ(α)T x ≤ δ0(α) for some α ∈ L0(B, σ).How:
For µ ∈ L(B, σ) show that µ = α + β for α, β such that:α ∈ L
0(B, σ), β ∈ L(B, σ)βT b ∈ Z
Use Lemma.
![Page 35: A Constructive Characterization of the Split Closure of a Mixed Integer Linear Program](https://reader031.vdocuments.net/reader031/viewer/2022021710/620b195c5237bc46f9609095/html5/thumbnails/35.jpg)
Introduction Characterization Lattices Polyhedrality
Proof of Theorem.
Let {vi}i∈V(σ) be a FIGS for L(B, σ) and let {ki}i∈V(σ) ⊆ Z+ besuch that
µ =∑
i∈V(σ)
kivi.
![Page 36: A Constructive Characterization of the Split Closure of a Mixed Integer Linear Program](https://reader031.vdocuments.net/reader031/viewer/2022021710/620b195c5237bc46f9609095/html5/thumbnails/36.jpg)
Introduction Characterization Lattices Polyhedrality
Proof of Theorem.
Let {vi}i∈V(σ) be a FIGS for L(B, σ) and let {ki}i∈V(σ) ⊆ Z+ besuch that
µ =∑
i∈V(σ)
kivi.
For each i ∈ V(σ) we have
ki = nimi + ri
for some ni, ri ∈ Z+, 0 ≤ ri < mi.Let
α =∑
i∈V(σ)
rivi and β =
∑
i∈V(σ)
nimivi
We have α ∈ L0(B, σ) and, as mi is such that mi bTvi ∈ Z wehave bTβ ∈ Z.
![Page 37: A Constructive Characterization of the Split Closure of a Mixed Integer Linear Program](https://reader031.vdocuments.net/reader031/viewer/2022021710/620b195c5237bc46f9609095/html5/thumbnails/37.jpg)
Introduction Characterization Lattices Polyhedrality
Final Remarks
The proof of the Theorem gives a way of enumerating theinequalities of SC(B, σ), SC(B) and SC:
Not practical for anything buy toy problems.There is redundancy in the enumeration for SC and SC(B).There is also redundancy in the enumeration of SC(B, σ). Infact we can reduce L0(B, σ) to
L0(B, σ) := {µ ∈ L(B, σ) : µ =∑
i∈V(σ)
rivi, ri ∈ {0, . . . , mi−1}
and {ri}i∈V(σ) are relatively prime}
[Dash et. al. 2005] also give a constructivecharacterization with similar properties.
![Page 38: A Constructive Characterization of the Split Closure of a Mixed Integer Linear Program](https://reader031.vdocuments.net/reader031/viewer/2022021710/620b195c5237bc46f9609095/html5/thumbnails/38.jpg)
Introduction Characterization Lattices Polyhedrality
Final Remarks
The proof of the Theorem gives a way of enumerating theinequalities of SC(B, σ), SC(B) and SC:
Not practical for anything buy toy problems.There is redundancy in the enumeration for SC and SC(B).There is also redundancy in the enumeration of SC(B, σ). Infact we can reduce L0(B, σ) to
L0(B, σ) := {µ ∈ L(B, σ) : µ =∑
i∈V(σ)
rivi, ri ∈ {0, . . . , mi−1}
and {ri}i∈V(σ) are relatively prime}
[Dash et. al. 2005] also give a constructivecharacterization with similar properties.
![Page 39: A Constructive Characterization of the Split Closure of a Mixed Integer Linear Program](https://reader031.vdocuments.net/reader031/viewer/2022021710/620b195c5237bc46f9609095/html5/thumbnails/39.jpg)
Introduction Characterization Lattices Polyhedrality
Final Remarks
The proof of the Theorem gives a way of enumerating theinequalities of SC(B, σ), SC(B) and SC:
Not practical for anything buy toy problems.There is redundancy in the enumeration for SC and SC(B).There is also redundancy in the enumeration of SC(B, σ). Infact we can reduce L0(B, σ) to
L0(B, σ) := {µ ∈ L(B, σ) : µ =∑
i∈V(σ)
rivi, ri ∈ {0, . . . , mi−1}
and {ri}i∈V(σ) are relatively prime}
[Dash et. al. 2005] also give a constructivecharacterization with similar properties.
![Page 40: A Constructive Characterization of the Split Closure of a Mixed Integer Linear Program](https://reader031.vdocuments.net/reader031/viewer/2022021710/620b195c5237bc46f9609095/html5/thumbnails/40.jpg)
Introduction Characterization Lattices Polyhedrality
Final Remarks
The proof of the Theorem gives a way of enumerating theinequalities of SC(B, σ), SC(B) and SC:
Not practical for anything buy toy problems.There is redundancy in the enumeration for SC and SC(B).There is also redundancy in the enumeration of SC(B, σ). Infact we can reduce L0(B, σ) to
L0(B, σ) := {µ ∈ L(B, σ) : µ =∑
i∈V(σ)
rivi, ri ∈ {0, . . . , mi−1}
and {ri}i∈V(σ) are relatively prime}
[Dash et. al. 2005] also give a constructivecharacterization with similar properties.
![Page 41: A Constructive Characterization of the Split Closure of a Mixed Integer Linear Program](https://reader031.vdocuments.net/reader031/viewer/2022021710/620b195c5237bc46f9609095/html5/thumbnails/41.jpg)
Introduction Characterization Lattices Polyhedrality
Final Remarks
The proof of the Theorem gives a way of enumerating theinequalities of SC(B, σ), SC(B) and SC:
Not practical for anything buy toy problems.There is redundancy in the enumeration for SC and SC(B).There is also redundancy in the enumeration of SC(B, σ). Infact we can reduce L0(B, σ) to
L0(B, σ) := {µ ∈ L(B, σ) : µ =∑
i∈V(σ)
rivi, ri ∈ {0, . . . , mi−1}
and {ri}i∈V(σ) are relatively prime}
[Dash et. al. 2005] also give a constructivecharacterization with similar properties.
![Page 42: A Constructive Characterization of the Split Closure of a Mixed Integer Linear Program](https://reader031.vdocuments.net/reader031/viewer/2022021710/620b195c5237bc46f9609095/html5/thumbnails/42.jpg)
Introduction Characterization Lattices Polyhedrality
D. Bertsimas, R. Weismantel.Optimization over Integers.Dynamic Ideas, Belmont, 2005.
K. Andersen, G. Cornuejols, Y. LiSplit Closure and Intersection Cuts.Mathematical Programming, 102:457–493. 2005.
E. Balas, M. PerregaardA precise correspondence between lift-and-project cuts,simple disjunctive cuts and mixed integer Gomory cuts for 01 programming.Mathematical Programming 94:221–245. 2003.
A. Caprara, A.N. LetchfordOn the separation of split cuts and related inequalities.Mathematical Programming 94:279–294. 2003.
![Page 43: A Constructive Characterization of the Split Closure of a Mixed Integer Linear Program](https://reader031.vdocuments.net/reader031/viewer/2022021710/620b195c5237bc46f9609095/html5/thumbnails/43.jpg)
Introduction Characterization Lattices Polyhedrality
W. Cook, R. Kannan, A. Schrijver.Chvátal closures for mixed integer programming problems.Mathematical Programming, 47:155–174. 1990.
S. Dash, O. Günlük, A. LodiOn the MIR closure of polyhedra.Working Paper.
M. Köppe, R. WeismantelCutting planes from a mixed integer Farkas lemma.Operations Research Letters 32:207–211. 2004