duality for mixed integer linear...
TRANSCRIPT
lehigh-logo
Duality for Mixed Integer Linear Programming
TED RALPHSMENAL GUZELSOY
ISE DepartmentCOR@L Lab
Lehigh [email protected]
University of Newcastle, Newcastle, Australia13 May 2009
Thanks:Work supported in part by the National Science Foundation
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 1 / 56
lehigh-logo
Outline
1 Duality TheoryIntroductionIP Duality
2 The Subadditive DualFormulationProperties
3 Constructing Dual FunctionsThe Value FunctionGomory’s ProcedureBranch-and-Bound MethodBranch-and-Cut
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 2 / 56
lehigh-logo
What is Duality?
It is difficult to give a general definition of mathematical duality, thoughmathematics is replete with various notions of it.
Set Theory and Logic (De Morgan Laws)Geomety (Pascal’s Theorem & Brianchon’s Theorem)Combinatorics (Graph Coloring)
The duality we are interested in is a sort offunctional duality.
We define a generic optimization problem to be a mappingf : X → R, whereXis the set of possibleinputsandf (x) is theresult.
Duality may then be defined as a method of transforming a givenprimal problemto an associateddual problemsuch that
the dual problem yields a bound on the primal problem, andapplying a related transformation to the dual produces the primal again.
In many case, we would also like to require that the dual boundbe “close” to theprimal result for a specific input of interest.
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 3 / 56
lehigh-logo
Duality in Mathematical Programming
In mathematical programming, the input is the problem data (e.g., the constraintmatrix, right-hand side, and cost vector for a linear program).
We view the primal and the dual as parametric problems, but some data is heldconstant.
Uses of the Dual in Mathematical Programing
If the dual is easier to evaluate, we can use it to obtain a bound on theprimal optimal value.
We can also use the dual to perform sensitivity analysis on theparameterized primal input data.
Finally, we can also use the dual to warm start solution procedure basedon evaluation of the dual.
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 4 / 56
lehigh-logo
Duality in Integer Programming
We will initially be interested in the mixed integer linear program (MILP)instance
zIP = minx∈S
cx, (P)
where,c ∈ Rn, S= x ∈ Zr+ × Rn−r
+ | Ax = b with A ∈ Qm×n, b ∈ Rm.
We call this instance thebase primal instance.
To construct a dual, we need a parameterized version of this instance.
For reasons that will become clear, the most relevant parameterization is of theright-hand side.
Thevalue function(or primal function) of the base primal instance (P) is
z(d) = minx∈S(d)
cx,
where for a givend ∈ Rm, S(d) = x ∈ Zr+ × Rn−r
+ | Ax = d.
We letz(d) = ∞ if d ∈ Ω = d ∈ Rm | S(d) = ∅.
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 5 / 56
lehigh-logo
Example: Value Function
zIP = min 12x1 + 2x3 + x4
s.t x1 −32x2 + x3 − x4 = b and
x1, x2 ∈ Z+, x3, x4 ∈ R+.
0
z(d)
d1-1-2-3 3 42-4
−
32 −
12−
52−
72
52
32
12
12
32
52
72
1
2
3
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 6 / 56
lehigh-logo
Dual Functions
A dual functionF : Rm → R is one that satisfiesF(d) ≤ z(d) for all d ∈ Rm.
How to select such a function?
We choose may choose one that is easy to construct/evaluate and/or for whichF(b) ≈ z(b).
This results in thebase dual instance
zD = maxF(b) : F(d) ≤ z(d), d ∈ Rm, F ∈ Υm
whereΥm ⊆ f | f : Rm→R
We callF∗ strongfor this instance ifF∗ is afeasibledual function andF∗(b) = z(b).
This dual instance always has a solutionF∗ that is strong if the value function isbounded andΥm ≡ f | f : Rm→R. Why?
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 7 / 56
lehigh-logo
The LP Relaxation Dual Function
It is easy to obtain a feasible dual function for any MILP.
Consider the value function of the LP relaxation of the primal problem:
FLP(d) = maxv∈Rm
vd : vA≤ c.
By linear programming duality theory, we haveFLP(d) ≤ z(d) for all d ∈ Rm.
Of course,FLP is not necessarily strong.
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 8 / 56
lehigh-logo
Example: LP Dual Function
FLP(d) = min vd,s.t 0≥ v ≥ − 1
2, andv ∈ R,
which can be written explicitly as
FLP(d) =
0, d ≤ 0− 1
2d, d > 0.
0d
1-1-2-3 3 42-4−
32 −
12−
52−
72
52
32
12
12
32
52
72
1
2
3
z(d)FLP(d)
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 9 / 56
lehigh-logo
The Subadditive Dual
By considering that
F(d) ≤ z(d), d ∈ Rm ⇐⇒ F(d) ≤ cx , x ∈ S(d), d ∈ Rm
⇐⇒ F(Ax) ≤ cx , x ∈ Zn+,
the generalized dual problem can be rewritten as
zD = maxF(b) : F(Ax) ≤ cx, x ∈ Zr+ × Rn−r
+ , F ∈ Υm.
Can we further restrictΥm and still guarantee a strong dual solution?
The class of linear functions? NO!
The class of convex functions? NO!
The class of sudadditive functions? YES!
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 10 / 56
lehigh-logo
The Subadditive Dual
Let a functionF be defined over a domainV. ThenF is subadditive ifF(v1) + F(v2) ≥ F(v1 + v2)∀v1, v2, v1 + v2 ∈ V.
Note that the value functionz is subadditive overΩ. Why?
If Υm ≡ Γm ≡ F is subadditive| F : Rm→R, F(0) = 0, we can rewrite thedual problem above as thesubadditive dual
zD = max F(b)
F(aj) ≤ cj j = 1, ..., r,
F(aj) ≤ cj j = r + 1, ..., n, and
F ∈ Γm,
where the functionF is defined by
F(d) = lim supδ→0+
F(δd)
δ∀d ∈ Rm.
Here,F is theupper d-directional derivativeof F at zero.
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 11 / 56
lehigh-logo
Example: Upper D-directional Derivative
The upper d-directional derivative can be interpreted as the slope of the valuefunction in directiond at 0.
For the example, we have
0d
1-1-2-3 3 42-4−
32 −
12−
52−
72
52
32
12
12
32
52
72
1
2
3
z(d)z(d)
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 12 / 56
lehigh-logo
Weak Duality
Weak Duality Theorem
Let x be a feasible solution to the primal problem and letF be a feasible solutionto the subadditive dual. Then,F(b) ≤ cx.
Proof.
Corollary
For the primal problem and its subadditive dual:1 If the primal problem (resp., the dual) is unbounded then thedual problem
(resp., the primal) is infeasible.2 If the primal problem (resp., the dual) is infeasible, then the dual problem
(resp., the primal) is infeasible or unbounded.
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 13 / 56
lehigh-logo
Strong Duality
Strong Duality Theorem
If the primal problem (resp., the dual) has a finite optimum, then so does thesubadditive dual problem (resp., the primal) and they are equal.
Outline of the Proof. Show that the value functionzor an extension toz is a feasibledual function.
Note thatzsatisfies the dual constraints.
Ω ≡ Rm: z∈ Γm.
Ω ⊂ Rm: ∃ ze ∈ Γm with ze(d) = z(d) ∀ d ∈ Ω andze(d) < ∞ ∀d ∈ Rm.
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 14 / 56
lehigh-logo
Example: Subadditive Dual
For our IP instance, the subadditive dual problem is
max F(b)F(1) ≤ 1
2F(− 3
2) ≤ 0F(1) ≤ 2
F(−1) ≤ 1F ∈ Γ1.
.
and we have the following feasible dual functions:1 F1(d) = d
2 is an optimal dual function forb ∈ 0, 1, 2, ....
2 F2(d) = 0 is an optimal function forb ∈ ...,−3,− 32, 0.
3 F3(d) = max 12⌈d− ⌈⌈d⌉−d⌉
4 ⌉, 2d − 32⌈d− ⌈⌈d⌉−d⌉
4 ⌉ is an optimal function forb ∈ [0, 1
4] ∪ [1, 54] ∪ [2, 9
4] ∪ ....
4 F4(d) = max 32⌈
2d3 −
2⌈⌈ 2d3 ⌉− 2d
3 ⌉
3 ⌉ − d,− 34⌈
2d3 −
2⌈⌈ 2d3 ⌉− 2d
3 ⌉
3 ⌉ + d2 is an optimal
function forb ∈ ... ∪ [− 72,−3] ∪ [−2,− 3
2] ∪ [− 12, 0]
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 15 / 56
lehigh-logo
Example: Feasible Dual Functions
0d
1-1-2-3 3 42-4−
32 −
12−
52−
72
52
32
12
12
32
52
72
1
2
3
z(d)F(d)
Notice how different dual solutions are optimal for some right-hand sides andnot for others.
Only the value function is optimal for all right-hand sides.
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 16 / 56
lehigh-logo
Farkas’ Lemma (Pure Integer)
For the primal problem, exactly one of the following holds:1 S 6= ∅
2 There is anF ∈ Γm with F(aj) ≥ 0, j = 1, ..., n, andF(b) < 0.
Proof. Let c = 0 and apply strong duality theorem to subadditive dual.
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 17 / 56
lehigh-logo
Complementary Slackness (Pure Integer)
For a given right-hand sideb, let x∗ andF∗ be feasible solutions to the primaland the subadditive dual problems, respectively. Thenx∗ andF∗ are optimalsolutions if and only if
1 x∗j (cj − F∗(aj)) = 0, j = 1, ..., n and
2 F∗(b) =∑n
j=1 F∗(aj)x∗j .
Proof. For an optimal pair we have
F∗(b) = F∗(Ax∗) =n
∑
j=1
F∗(aj)x∗j = cx∗.
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 18 / 56
lehigh-logo
Example: Pure Integer Case
If we require integrality of all variables in our previous example, then the valuefunction becomes
0d
1-1-2-3 3 42-4−
32 −
12−
52−
72
52
32
12
12
32
52
72
1
2
3
z(d)
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 19 / 56
lehigh-logo
Constructing Dual Functions
Explicit constructionThe Value FunctionGenerating Functions
RelaxationsLagrangian RelaxationQuadratic Lagrangian RelaxationCorrected Linear Dual Functions
Primal Solution AlgorithmsCutting Plane MethodBranch-and-Bound MethodBranch-and-Cut Method
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 20 / 56
lehigh-logo
Properties of the Value Function
It is subadditive overΩ.
It is piecewise polyhedral.
For an ILP, it can be obtained by a finite number of limited operations onelements of the RHS:
(i) rational multiplication(ii) nonnegative combination(iii) rounding
Chvatal fcns.
(iv) taking the minimum
Gomory fcns.
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 21 / 56
lehigh-logo
The Value Function for MILPs
There is a one-to-one correspondence between ILP instancesand Gomoryfunctions.
TheJeroslow Formulashows that the value function of a MILP can also becomputed by taking the minimum of different values of a single Gomoryfunction with a correction term to account for the continuous variables.
The value function of the earlier example is
z(d) = min
32 max
⌈
⌊2d⌋3
⌉
,⌈
⌊2d⌋2
⌉
+ 3⌈2d⌉2 + 2d,
32 max
⌈
⌈2d⌉3
⌉
,⌈
⌈2d⌉2
⌉
− d,
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 22 / 56
lehigh-logo
Jeroslow Formula
Let the setE consist of the index sets of dual feasible bases of the linearprogram
min1M
cCxC :1M
ACxC = b, x ≥ 0
whereM ∈ Z+ such that for anyE ∈ E , MA−1E aj ∈ Zm for all j ∈ I .
Theorem (Jeroslow Formula)There is ag ∈ G m such that
z(d) = minE∈E
g(⌊d⌋E) + vE(d− ⌊d⌋E) ∀d ∈ Rm with S(d) 6= ∅,
where forE ∈ E , ⌊d⌋E = AE⌊A−1E d⌋ andvE is the corresponding basic feasible
solution.
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 23 / 56
lehigh-logo
The Single Constraint Case
Let us now consider an MILP with a single constraint for the purposes ofillustration:
minx∈S
cx, (P1)
c ∈ Rn, S= x ∈ Zr+ × Rn−r
+ | a′x = b with a ∈ Qn, b ∈ R.
Thevalue functionof (P) is
z(d) = minx∈S(d)
cx,
where for a givend ∈ R, S(d) = x ∈ Zr+ × Rn−r
+ | a′x = d.
Assumptions: LetI = 1, . . . , r, C = r + 1, . . . , n, N = I ∪ C.
z(0) = 0 =⇒ z : R → R ∪ +∞,N+ = i ∈ N | ai > 0 6= ∅ andN− = i ∈ N | ai < 0 6= ∅,r < n, that is,|C| ≥ 1 =⇒ z : R → R.
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 24 / 56
lehigh-logo
Example
min 3x1 + 72x2 + 3x3 + 6x4 + 7x5 + 5x6
s.t 6x1 + 5x2 − 4x3 + 2x4 − 7x5 + x6 = b andx1, x2, x3 ∈ Z+, x4, x5, x6 ∈ R+.
(SP)
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28-2-4-6-8-10-12-14-16
2
4
6
8
10
12
14
16
18
d
z
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 25 / 56
lehigh-logo
Example (cont’d)
η = 12, ζ = − 3
4, ηC = 3 andζC = −1:
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28-2-4-6-8-10-12-14-16
2
4
6
8
10
12
14
16
18
d
FL
FU
η
z
ζC
ζ
ηC
η = ηC⇐⇒z(d) = FU(d) = FL(d) ∀d ∈ R+
ζ = ζC⇐⇒z(d) = FU(d) = FL(d) ∀d ∈ R−
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 26 / 56
lehigh-logo
Observations
Considerd+U , d−
U , d+L , d−
L :
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28-2-4-6-8-10-12-14-16
2
4
6
8
10
12
14
16
18
dd−U
d−L d+U d+L2d+L 3d+L 4d+L2d−L2d−L
The relation betweenFU and the linear segments ofz: ηC, ζC
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 27 / 56
lehigh-logo
Redundant Variables
Let T ⊆ C be such that
t+ ∈ T if and only if ηC < ∞ andηC =ct+
at+and similarly,
t− ∈ T if and only if ζC > −∞ andζC =ct−
at−.
and define
ν(d) = min cI xI + cTxT
s.t. aI xI + aTxT = dxI ∈ ZI
+, xT ∈ RT+
Then
ν(d) = z(d) for all d ∈ R.
The variables inC\T areredundant.
zcan be represented withat most 2 continuous variables.
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 28 / 56
lehigh-logo
Example
min x1 − 3/4x2 + 3/4x3
s.t 5/4x1 − x2 + 1/2x3 = b, x1, x2 ∈ Z+, x3 ∈ R+.
ηc= 3/2, ζC
= −∞.
-1
-2
12
1
32
2
52
−
12−
32 -1−
52 -2-3−
72 11
2 2 352
32
72
−
12
−
32
−
52
For each discontinuous pointdi , we havedi − (5/4yi1 − yi
2) = 0 and each linear segment hasthe slope ofηC = 3/2.
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 29 / 56
lehigh-logo
Jeroslow Formula
Let M ∈ Z+ be such that for anyt ∈ T, Maj
at∈ Z for all j ∈ I .
Then there is a Gomory functiong such that
z(d) = mint∈T
g(⌊d⌋t) +ct
at(d− ⌊d⌋t), ⌊d⌋t =
at
M
⌊
Mdat
⌋
, ∀d ∈ R
Such a Gomory function can be obtained from the value function of a relatedPILP.
For t ∈ T, setting
ωt(d) = g(⌊d⌋t) +ct
at(d− ⌊d⌋t) ∀d ∈ R,
we can writez(d) = min
t∈Tωt(d) ∀d ∈ R
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 30 / 56
lehigh-logo
Piecewise Linearity and Continuity
For t ∈ T, ωt is piecewise linear with finitely many linear segments on anyclosed interval and each of those linear segments has a slopeof ηC if t = t+ orζC if t = t−.
ωt+ is continuous from the right, ωt− is continuous from the left.
ωt+ andωt− are bothlower-semicontinuous.
Theoremz is piecewise-linearwith finitelymany linear segments on any closed intervaland each of those linear segments has a slope ofηC or ζC.
(Meyer 1975)z is lower-semicontinuous.
ηC < ∞ if and only ifz is continuous from the right.
ζC > −∞ if and only ifz is continuous from the left.
BothηC andζC arefinite if and only ifz is continuous everywhere.
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 31 / 56
lehigh-logo
Maximal Subadditive Extension
Let f : [0, h]→R, h > 0 besubadditiveandf (0) = 0.
Themaximal subadditive extensionof f from [0, h] to R+ is
fS(d) =
f (d) if d ∈ [0, h]
infC∈C(d)
∑
ρ∈C
f (ρ) if d > h ,
C(d) is the set of all finite collectionsρ1, ..., ρR such thatρi ∈ [0, h], i = 1, ..., Rand
PRi=1 ρi = d.
Each collectionρ1, ..., ρR is called anh-partitionof d.We can also extend a subadditive functionf : [h, 0]→R, h < 0 to R
−similarly.
(Bruckner 1960)fS is subadditive and ifg is any other subadditive extension offfrom [0, h] to R+, theng ≤ fS (maximality).
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 32 / 56
lehigh-logo
Extending the Value Function
Suppose we usez itself as theseedfunction.
Observe that we can change the“ inf” to “min” :
Lemma
Let the functionf : [0, h]→R be defined byf (d) = z(d) ∀d ∈ [0, h]. Then,
fS(d) =
z(d) if d ∈ [0, h]
minC∈C(d)
∑
ρ∈C
z(ρ) if d > h .
For anyh > 0, z(d) ≤ fS(d) ∀d ∈ R+.
Observe that ford ∈ R+, fS(d)→z(d) while h→∞.
Is there anh < ∞ such thatfS(d) = z(d) ∀d ∈ R+?
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 33 / 56
lehigh-logo
Extending the Value Function (cont.)
Yes! For large enoughh, maximal extension produces the value function itself.
Theorem
Letdr = maxai | i ∈ N anddl = minai | i ∈ N and let the functionsfr andfl bethe maximal subadditive extensions ofz from the intervals[0, dr ] and[dl , 0] to R+ andR−, respectively. Let
F(d) =
fr(d) d ∈ R+
fl(d) d ∈ R−
then,z = F.
Outline of the Proof.
z≤ F : By construction.
z≥ F : Using MILP duality,F is dual feasible.
In other words, the value function is completely encoded by the breakpoints in[dl , dr ]and 2 slopes.
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 34 / 56
lehigh-logo
General Procedure
We will construct the value function in two stepsConstruct the value function on[dl , dr ].Extend the value function to the entire real line from[dl , dr ].
AssumptionsWe assumeηC < ∞ andζC < ∞.We construct the value function overR+ only.These assmuptions are only needed to simplify the presentation.
Constructing the Value Function on[0, dr ]
If both ηC andζC arefinite, the value function iscontinuousand the slopes of thelinear segments alternate betweenηC andζC.For d1, d2 ∈ [0, dr ], if z(d1) andz(d2) are connected by a line with slopeηC or ζC,thenz is linear over[d1, d2] with the respective slope (subadditivity).With these observations, we can formulate a finite algorithm to evaluatez in [dl , dr ].
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 35 / 56
lehigh-logo
Example (cont’d)
dr = 6:
ηC —— ζC ——
0 2 4 6
2
4
6
8
Figure:Evaluatingz in [0, 6]
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 36 / 56
lehigh-logo
Example (cont’d)
dr = 6:
ηC —— ζC ——
0 2 4 6
2
4
6
8
Figure:Evaluatingz in [0, 6]
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 37 / 56
lehigh-logo
Example (cont’d)
dr = 6:
ηC —— ζC ——
0 2 4 6
2
4
6
8
Figure:Evaluatingz in [0, 6]
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 38 / 56
lehigh-logo
Example (cont’d)
dr = 6:
ηC —— ζC ——
0 2 4 6
2
4
6
8
Figure:Evaluatingz in [0, 6]
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 39 / 56
lehigh-logo
Example (cont’d)
dr = 6:
ηC —— ζC ——
0 2 4 6
2
4
6
8
Figure:Evaluatingz in [0, 6]
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 40 / 56
lehigh-logo
Example (cont’d)
Extending the value function of (SP) from[0, 12] to [0, 18]
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28-2-4-6-8-10-12-14-16
2
4
6
8
10
12
14
16
18
d
FL
FU
z
Υ
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 41 / 56
lehigh-logo
Example (cont’d)
Extending the value function of (SP) from[0, 12] to [0, 18]
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28-2-4-6-8-10-12-14-16
2
4
6
8
10
12
14
16
18
d
FL
FU
z
Υ
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 42 / 56
lehigh-logo
Example (cont’d)
Extending the value function of (SP) from[0, 18] to [0, 24]
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28-2-4-6-8-10-12-14-16
2
4
6
8
10
12
14
16
18
d
FL
FU
z
Υ
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 43 / 56
lehigh-logo
Example (cont’d)
Extending the value function of (SP) from[0, 18] to [0, 24]
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28-2-4-6-8-10-12-14-16
2
4
6
8
10
12
14
16
18
d
FL
FU
z
Υ
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 44 / 56
lehigh-logo
General Case
For E ∈ E , setting
ωE(d) = g(⌊d⌋E) + vE(d− ⌊d⌋E) ∀d ∈ Rm with S(d) 6= ∅,
we can writez(d) = min
E∈E
ωE(d) ∀d ∈ Rm with S(d) 6= ∅.
Many of our previous results can be extended to general case in the obvious way.
Similarly, we can use maximal subadditive extensions to construct the valuefunction.
However, an obvious combinatorial explosion occurs.
Therefore, we consider using single row relaxations to get asubadditiveapproximation.
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 45 / 56
lehigh-logo
Gomory’s Procedure
There is a Chvatal function that is optimal to the subadditive dual of an ILP withRHSb ∈ ΩIP andzIP(b) > −∞.
The procedure:In iterationk, we solve the following LP
zIP(b)k−1 = min cxs.t. Ax = b
∑nj=1 f i(aj)xj ≥ f i(b) i = 1, ..., k− 1
x ≥ 0
Thekth cut,k > 1, is dependent on the RHS and written as:
f k(d) =
⌈
m∑
i=1
λk−1i di +
k−1∑
i=1
λk−1m+i f
i(d)
⌉
where λk−1 = (λk−11 , ..., λk−1
m+k−1) ≥ 0
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 46 / 56
lehigh-logo
Gomory’s Procedure (cont.)
Assume thatb ∈ ΩIP, zIP(b) > −∞ and the algorithm terminates afterk + 1iterations.
If uk is the optimal dual solution to the LP in the final iteration, then
Fk(d) =
m∑
i=1
uki di +
k∑
i=1
ukm+i f
i(d),
is a Chvatal function withFk(b) = zIP(b) and furthermore, it is optimal to thesubadditive ILP dual problem.
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 47 / 56
lehigh-logo
Gomory’s Procedure (cont.)
Example:Consider an integer program with two constraints. Suppose that thefollowing is the first constraint:
2x1 + −2x2 + x3 − x4 = 3
At the first iteration, Gomory’s procedure can be used to derive the valid inequality
⌈2/2⌉x1 + ⌈ − 2/2⌉x2 + ⌈1/2⌉x3 + ⌈ − 1/2⌉x4 ≥ ⌈3/2⌉
by scaling the constraint withλ = 1/2. In other words, we obtainx1 − x2 + x3 ≥ 2.Then the corresponding function is:
F1(d) = ⌈d1/2⌉ + 0d2 = ⌈d1/2⌉
By continuing Gomory’s procedure, one can then build up an optimal dual function inthis way.
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 48 / 56
lehigh-logo
Branch-and-Bound Method
Assume that the primal problem is solved to optimality.Let T be the set of leaf nodes of the search tree.Thus, we’ve solved the LP relaxation of the following problem at nodet ∈ T
zt(b) = min cxs.t x ∈ St(b)
,
whereSt(b) = Ax = b, x ≥ lt,−x ≥ −ut, x ∈ Zn andut, lt ∈ Zr are thebranching bounds applied to the integer variables.Let (vt, vt, vt) be
the dual feasible solution used to prune nodet, if t is feasibly pruneda dual feasible solution (that can be obtained from it parent) to nodet, if t isinfeasibly pruned
Then,
FBB(d) = mint∈T
vtd + vtlt − vtut
is an optimal solution to the generalized dual problem.
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 49 / 56
lehigh-logo
Branch-and-Cut Method
It is thus easy to get a strong dual function from branch-and-bound.
Note, however, that it’s not subadditive in general.
To obtain a subaditive function, we can include the variablebounds explicitly asconstraints, but then the function may not be strong.
For branch-and-cut, we have to take care of the cuts.Case 1: Do we know the subadditive representation of each cut?Case 2: Do we know the RHS dependency of each cut?Case 3: Otherwise, we can use some proximity results or the variable bounds.
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 50 / 56
lehigh-logo
Case 1
If we know the subadditive representation of each cut:At a nodet, we solve the LP relaxation of the following problem
zt(b) = min cxs.t Ax ≥ b
x ≥ lt
−x ≥ −gt
Htx ≥ ht
x ∈ Zr+ × Rn−r
+
wheregt, lt ∈ Rr are the branching bounds applied to the integer variables andHtx ≥ ht is the set of added cuts in the form
∑
j∈I
Ftk(a
kj )xj +
∑
j∈N\I
Ftk(a
kj )xj ≥ Ft
k(σk(b)) k = 1, ..., ν(t),
ν(t): the number of cuts generated so far,ak
j , j = 1, ..., n: the columns of the problem that thekth cut is constructed from,σk(b): is the mapping ofb to the RHS of the corresponding problem.
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 51 / 56
lehigh-logo
Case 1
Let T be the set of leaf nodes,ut, ut, ut andwt be the dual feasible solution used toprunet ∈ T. Then,
F(d) = mint∈T
utd + utlt − utgt +
ν(t)∑
k=1
wtkF
tk(σk(d))
is an optimal dual function, that is,z(b) = F(b).
Again, we obtain a subaddite function if the variables are bounded.
However, we may not know the subadditive representation of each cut.
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 52 / 56
lehigh-logo
Case 2
If we know the RHS dependencies of each cut:We know for
Gomory fractional cuts.
Knapsack cuts
Mixed-integer Gomory cuts
?
Then, we do the same analysis as before.
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 53 / 56
lehigh-logo
Case 3
In the absence of subadditive representations or RHS dependencies:For each nodet ∈ T, let ht be such that
htk =
n∑
j=1
htkjxj with xj =
ltj if htkj ≥ 0
gtj otherwise
, k = 1, ..., ν(t)
wherehtkj is thekth entry of columnht
j . Furthermore, define
ht =n
∑
j=1
htj xj with xj =
ltj if wthtj ≥ 0
gtj otherwise
.
Then the function
F(d) = mint∈T
utd + utlt − utgt + maxwtht, wtht
is a feasible dual solution and henceF(b) yields a lower bound.This approach is the easiest way and can be used for bounded MILPs (binaryprograms), however it is unlikely to get an effective dual feasible solution forgeneral MILPs.
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 54 / 56
lehigh-logo
Conclusions
It is possible to generalize the duality concepts that are familiar to us from linearprogramming.
However, it is extremely difficult to make any of it practical.
There are some isolated cases where this theory has been applied in practice, butthey are far and few between.
We have a number of projects aimed in this direction, so stay tuned...
Ralphs and Guzelsoy (Lehigh University) Duality for MILP 18 February 2009 55 / 56