linear programming duality

36
CS6501: Topics in Learning and Game Theory (Fall 2019) Linear Programming Duality Instructor: Haifeng Xu Slides of this lecture is adapted from Shaddin Dughmi at https://www-bcf.usc.edu/~shaddin/cs675sp18/index.html

Upload: others

Post on 07-Apr-2022

9 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Linear Programming Duality

CS6501: Topics in Learning and Game Theory(Fall 2019)

Linear Programming Duality

Instructor: Haifeng Xu

Slides of this lecture is adapted from Shaddin Dughmi athttps://www-bcf.usc.edu/~shaddin/cs675sp18/index.html

Page 2: Linear Programming Duality

2

ร˜Recap and Weak Duality

ร˜Strong Duality and Its Proof

ร˜Consequence of Strong Duality

Outline

Page 3: Linear Programming Duality

3

Linear Program (LP)

minimize (or maximize) ๐‘" โ‹… ๐‘ฅsubject to ๐‘Ž& โ‹… ๐‘ฅ โ‰ค ๐‘& โˆ€๐‘– โˆˆ ๐ถ-

๐‘Ž& โ‹… ๐‘ฅ โ‰ฅ ๐‘& โˆ€๐‘– โˆˆ ๐ถ/๐‘Ž& โ‹… ๐‘ฅ = ๐‘& โˆ€๐‘– โˆˆ ๐ถ1

General form:

maximize ๐‘" โ‹… ๐‘ฅsubject to ๐‘Ž& โ‹… ๐‘ฅ โ‰ค ๐‘& โˆ€๐‘– = 1,โ‹ฏ ,๐‘š

๐‘ฅ6 โ‰ฅ 0 โˆ€๐‘— = 1,โ‹ฏ , ๐‘›

Standard form:

Page 4: Linear Programming Duality

4

Application: Optimal Production

ร˜ ๐‘› products, ๐‘š raw materials

ร˜Every unit of product ๐‘— uses ๐‘Ž&6 units of raw material ๐‘–

ร˜There are ๐‘& units of material ๐‘– availableร˜Product ๐‘— yields profit ๐‘6 per unit

ร˜Factory wants to maximize profit subject to available raw materials

Can be formulated as an LP in standard form

max ๐‘" โ‹… ๐‘ฅs.t. โˆ‘6;-< ๐‘Ž&6 ๐‘ฅ6 โ‰ค ๐‘&, โˆ€๐‘– โˆˆ [๐‘š]

๐‘ฅ6 โ‰ฅ 0, โˆ€๐‘— โˆˆ [๐‘›]

Page 5: Linear Programming Duality

5

Primal and Dual Linear Program

max ๐‘" โ‹… ๐‘ฅs.t. โˆ‘6;-< ๐‘Ž&6 ๐‘ฅ6 โ‰ค ๐‘&, โˆ€๐‘– โˆˆ [๐‘š]

๐‘ฅ6 โ‰ฅ 0, โˆ€๐‘— โˆˆ [๐‘›]

Primal LP Dual LP

min ๐‘" โ‹… ๐‘ฆs.t. โˆ‘&;-@ ๐‘Ž&6 ๐‘ฆ& โ‰ฅ ๐‘6, โˆ€๐‘— โˆˆ [๐‘›]

๐‘ฆ& โ‰ฅ 0, โˆ€๐‘– โˆˆ [๐‘š]

Dual LP corresponds to the buyerโ€™s optimization problem, as follows:ร˜Buyer wants to directly buy the raw material

ร˜Dual variable ๐‘ฆ& is buyerโ€™s proposed price per unit of raw material ๐‘–ร˜Dual price vector is feasible if factory is incentivized to sell materials

ร˜Buyer wants to spend as little as possible to buy raw materials

Economic Interpretation:

Page 6: Linear Programming Duality

6

Primal and Dual Linear Program

max ๐‘" โ‹… ๐‘ฅs.t. โˆ‘6;-< ๐‘Ž&6 ๐‘ฅ6 โ‰ค ๐‘&, โˆ€๐‘– โˆˆ [๐‘š]

๐‘ฅ6 โ‰ฅ 0, โˆ€๐‘— โˆˆ [๐‘›]

Primal LP Dual LP

min ๐‘" โ‹… ๐‘ฆs.t. โˆ‘&;-@ ๐‘Ž&6 ๐‘ฆ& โ‰ฅ ๐‘6, โˆ€๐‘— โˆˆ [๐‘›]

๐‘ฆ& โ‰ฅ 0, โˆ€๐‘– โˆˆ [๐‘š]

Upperbound Interpretation:

Dual LP can be interpreted as finding best upperbound for the primalร˜ Multiplying each row ๐‘– of primal by ๐‘ฆ& and summing the constraints

ร˜ Goal: find the best such ๐‘ฆ to get the smallest upper bound

Page 7: Linear Programming Duality

7

ร˜ So far, mainly writing the Dual based on syntactic rules

ร˜ Next, will show Primal and Dual are inherently related

Page 8: Linear Programming Duality

8

Weak Duality

max ๐‘A โ‹… ๐‘ฅs.t. ๐ด๐‘ฅ โ‰ค ๐‘

๐‘ฅ โ‰ฅ 0

Primal LPmin ๐‘A โ‹… ๐‘ฆs.t. ๐ดA๐‘ฆ โ‰ฅ ๐‘

๐‘ฆ โ‰ฅ 0

Dual LP

Theorem [Weak Duality]: For any primal feasible ๐‘ฅ and dualfeasible ๐‘ฆ, we have ๐‘" โ‹… ๐‘ฅ โ‰ค ๐‘" โ‹… ๐‘ฆ

Corollary:ร˜ If primal is unbounded, dual is infeasibleร˜ If dual is unbounded, primal is infeasibleร˜ If primal and dual are both feasible, then

OPT(primal) โ‰ค OPT(dual)

obj value of dual

obj value of primal

Page 9: Linear Programming Duality

9

Weak Duality

max ๐‘A โ‹… ๐‘ฅs.t. ๐ด๐‘ฅ โ‰ค ๐‘

๐‘ฅ โ‰ฅ 0

Primal LPmin ๐‘A โ‹… ๐‘ฆs.t. ๐ดA๐‘ฆ โ‰ฅ ๐‘

๐‘ฆ โ‰ฅ 0

Dual LP

Theorem [Weak Duality]: For any primal feasible ๐‘ฅ and dualfeasible ๐‘ฆ, we have ๐‘" โ‹… ๐‘ฅ โ‰ค ๐‘" โ‹… ๐‘ฆ

Corollary: If ๐‘ฅ is primal feasible and ๐‘ฆ is dualfeasible, and ๐‘" โ‹… ๐‘ฅ = ๐‘" โ‹… ๐‘ฆ, then both are optimal.

obj value of dual

obj value of primal

Page 10: Linear Programming Duality

10

Interpretation of Weak Duality

Economic Interpretation: If prices of raw materials are set such that there is incentive to sell raw materials directly, then factoryโ€™s total revenue from sale of raw materials would exceed its profit from any production.

Upperbound Interpretation: The method of rescaling and summing rows of the Primal indeed givens an upper bound of the Primalโ€™s objective value (well, self-evidentโ€ฆ).

Page 11: Linear Programming Duality

11

Proof of Weak Duality

max ๐‘A โ‹… ๐‘ฅs.t. ๐ด๐‘ฅ โ‰ค ๐‘

๐‘ฅ โ‰ฅ 0

Primal LPmin ๐‘A โ‹… ๐‘ฆs.t. ๐ดA๐‘ฆ โ‰ฅ ๐‘

๐‘ฆ โ‰ฅ 0

Dual LP

๐‘ฆ" โ‹… ๐‘ โ‰ฅ ๐‘ฆ" โ‹… ๐ด๐‘ฅ = ๐‘ฅ" โ‹… ๐ด"๐‘ฆ โ‰ฅ ๐‘ฅ" โ‹… ๐‘

Page 12: Linear Programming Duality

12

ร˜Recap and Weak Duality

ร˜Strong Duality and Its Proof

ร˜Consequence of Strong Duality

Outline

Page 13: Linear Programming Duality

13

Strong Duality

Theorem [Strong Duality]: If either the primal or dual is feasibleand bounded, then so is the other and OPT(primal) = OPT(dual).

obj value of primal

obj value of dual

John von Neumann

โ€ฆ I thought there was nothing worth publishing until the Minimax Theorem was proved.

Page 14: Linear Programming Duality

14

Interpretation of Strong Duality

Economic Interpretation: There exist raw material prices such that the factory is indifferent between selling raw materials or products.

Upperbound Interpretation: The method of scaling and summing constraints yields a tight upperbound for the primal objective value.

Page 15: Linear Programming Duality

15

Proof of Strong Duality

Page 16: Linear Programming Duality

16

Projection Lemma

Weierstrassโ€™ Theorem: Let ๐‘ be a compact set, and let ๐‘“(๐‘ง) be acontinuous function on ๐‘ง. Then min{ ๐‘“(๐‘ง) โˆถ ๐‘ง โˆˆ ๐‘ } exists.

๐‘ง

๐‘“(๐‘ง)

Page 17: Linear Programming Duality

17

Projection Lemma

Weierstrassโ€™ Theorem: Let ๐‘ be a compact set, and let ๐‘“(๐‘ง) be acontinuous function on ๐‘ง. Then min{ ๐‘“(๐‘ง) โˆถ ๐‘ง โˆˆ ๐‘ } exists.

Projection Lemma: Let ๐‘ โŠ‚ โ„@ be a nonempty closed convex setand let ๐‘ฆ โˆ‰ ๐‘. Then there exists ๐‘งโˆ— โˆˆ ๐‘ with minimum ๐‘™/ distancefrom ๐‘ฆ. Moreover, โˆ€ ๐‘ง โˆˆ ๐‘ we have ๐‘ฆ โ€“ ๐‘งโˆ— "(๐‘ง โ€“ ๐‘งโˆ—) โ‰ค 0.

๐‘ฆ ๐‘งโˆ—

๐‘งProof: homework exercise

๐‘

Page 18: Linear Programming Duality

18

Separating Hyperplane Theorem

Theorem: Let ๐‘ โŠ‚ โ„@ be a nonempty closed convex set and let๐‘ฆ โˆ‰ ๐‘. Then there exists a hyperplane ๐›ผ" โ‹… ๐‘ง = ๐›ฝ that strictlyseparates ๐‘ฆ from ๐‘. That is, ๐›ผ" โ‹… ๐‘ง โ‰ฅ ๐›ฝ, โˆ€ ๐‘ง โˆˆ ๐‘ and ๐›ผ" โ‹… ๐‘ฆ < ๐›ฝ.

๐‘ฆ ๐‘งโˆ—

๐‘ง

Proof: choose ๐›ผ = ๐‘งโˆ— โˆ’ ๐‘ฆ and ๐›ฝ = ๐›ผ โ‹… ๐‘งโˆ— and use projection lemmaร˜ Homework exercise

๐›ผ" โ‹… ๐‘ง = ๐›ฝ

๐‘๐›ผ

Page 19: Linear Programming Duality

19

Farkasโ€™ LemmaFarkasโ€™ Lemma: Let ๐ด โˆˆ โ„@ร—< and ๐‘ โˆˆ โ„@, then exactly one ofthe following two statements holds:a) There exists ๐‘ฅ โˆˆ โ„< such that ๐ด๐‘ฅ = ๐‘ and ๐‘ฅ โ‰ฅ 0b) There exists y โˆˆ โ„@ such that ๐ด"๐‘ฆ โ‰ฅ 0 and ๐‘"๐‘ฆ < 0

Case a):

Page 20: Linear Programming Duality

20

Farkasโ€™ LemmaFarkasโ€™ Lemma: Let ๐ด โˆˆ โ„@ร—< and ๐‘ โˆˆ โ„@, then exactly one ofthe following two statements holds:a) There exists ๐‘ฅ โˆˆ โ„< such that ๐ด๐‘ฅ = ๐‘ and ๐‘ฅ โ‰ฅ 0b) There exists y โˆˆ โ„@ such that ๐ด"๐‘ฆ โ‰ฅ 0 and ๐‘"๐‘ฆ < 0

Case a):

Page 21: Linear Programming Duality

21

Farkasโ€™ LemmaFarkasโ€™ Lemma: Let ๐ด โˆˆ โ„@ร—< and ๐‘ โˆˆ โ„@, then exactly one ofthe following two statements holds:a) There exists ๐‘ฅ โˆˆ โ„< such that ๐ด๐‘ฅ = ๐‘ and ๐‘ฅ โ‰ฅ 0b) There exists y โˆˆ โ„@ such that ๐ด"๐‘ฆ โ‰ฅ 0 and ๐‘"๐‘ฆ < 0

Case b):

Page 22: Linear Programming Duality

22

Farkasโ€™ Lemma

Geometric interpretation:

Farkasโ€™ Lemma: Let ๐ด โˆˆ โ„@ร—< and ๐‘ โˆˆ โ„@, then exactly one ofthe following two statements holds:a) There exists ๐‘ฅ โˆˆ โ„< such that ๐ด๐‘ฅ = ๐‘ and ๐‘ฅ โ‰ฅ 0b) There exists y โˆˆ โ„@ such that ๐ด"๐‘ฆ โ‰ฅ 0 and ๐‘"๐‘ฆ < 0

Z๐‘Ž-

Z๐‘Ž/

Z๐‘Ž6 is ๐‘—โ€™th column of ๐ด๐‘

a) ๐‘ is in the cone

Page 23: Linear Programming Duality

23

Farkasโ€™ Lemma

Geometric interpretation:

Farkasโ€™ Lemma: Let ๐ด โˆˆ โ„@ร—< and ๐‘ โˆˆ โ„@, then exactly one ofthe following two statements holds:a) There exists ๐‘ฅ โˆˆ โ„< such that ๐ด๐‘ฅ = ๐‘ and ๐‘ฅ โ‰ฅ 0b) There exists y โˆˆ โ„@ such that ๐ด"๐‘ฆ โ‰ฅ 0 and ๐‘"๐‘ฆ < 0

Z๐‘Ž-

Z๐‘Ž/

Z๐‘Ž6 is ๐‘—โ€™th column of ๐ด

๐‘a) ๐‘ is in the coneb) ๐‘ is not in the cone, and there exists a hyperplane with direction ๐‘ฆ

that separates ๐‘ from the cone

๐‘ฆ

Page 24: Linear Programming Duality

24

Farkasโ€™ Lemma

Proof: ร˜ Cannot both hold; Otherwise, yields contradiction as follows:

ร˜ Next, we prove if (a) does not hold, then (b) must holdโ€ข This implies the lemma

Farkasโ€™ Lemma: Let ๐ด โˆˆ โ„@ร—< and ๐‘ โˆˆ โ„@, then exactly one ofthe following two statements holds:a) There exists ๐‘ฅ โˆˆ โ„< such that ๐ด๐‘ฅ = ๐‘ and ๐‘ฅ โ‰ฅ 0b) There exists y โˆˆ โ„@ such that ๐ด"๐‘ฆ โ‰ฅ 0 and ๐‘"๐‘ฆ < 0

= ๐‘ฆ" โ‹… ๐ด๐‘ฅ = ๐‘ฆ" โ‹… ๐‘ < 0.0 โ‰ค (๐ด"๐‘ฆ)" โ‹… ๐‘ฅ

Page 25: Linear Programming Duality

25

Farkasโ€™ Lemma

ร˜Consider Z = {๐ด๐‘ฅ: ๐‘ฅ โ‰ฅ 0} so that ๐‘ is closed and convexร˜(a) does not hold โ‡” ๐‘ โˆ‰ ๐‘ร˜By separating hyperplane theorem, there exists hyperplane ๐›ผ โ‹… ๐‘ง = ๐›ฝ such that ๐›ผ" โ‹… ๐‘ง โ‰ฅ ๐›ฝ for all ๐‘ง โˆˆ ๐‘ and ๐›ผ" โ‹… ๐‘ < ๐›ฝ

ร˜Note 0 โˆˆ ๐‘, therefore ๐›ฝ โ‰ค ๐›ผ" โ‹… 0 = 0 and thus ๐›ผ" โ‹… ๐‘ < 0ร˜๐›ผ"๐ด๐‘ฅ โ‰ฅ ๐›ฝ for any ๐‘ฅ โ‰ฅ 0 implies ๐›ผ"๐ด โ‰ฅ 0 since ๐‘ฅ can be arbitrary

largeร˜Letting ๐›ผ be our ๐‘ฆ yields the lemma

Farkasโ€™ Lemma: Let ๐ด โˆˆ โ„@ร—< and ๐‘ โˆˆ โ„@, then exactly one ofthe following two statements holds:a) There exists ๐‘ฅ โˆˆ โ„< such that ๐ด๐‘ฅ = ๐‘ and ๐‘ฅ โ‰ฅ 0b) There exists y โˆˆ โ„@ such that ๐ด"๐‘ฆ โ‰ฅ 0 and ๐‘"๐‘ฆ < 0

Claim: if (a) does not hold, then (b) must hold.

Page 26: Linear Programming Duality

26

An Alternative of Farkasโ€™ LemmaFollowing corollary of Farkasโ€™ lemma is more convenient for our proof

Corollary: Exactly one of the following systems holds:

โˆƒ ๐‘ฅ โˆˆ โ„<, s.t.๐ด โ‹… ๐‘ฅ โ‰ค ๐‘๐‘ฅ โ‰ฅ 0

โˆƒ ๐‘ฆ โˆˆ โ„@, s.t.๐ดA โ‹… ๐‘ฆ โ‰ฅ 0๐‘A โ‹… ๐‘ฆ < 0๐‘ฆ โ‰ฅ 0

Compare to the original version

โˆƒ ๐‘ฅ โˆˆ โ„<, s.t.๐ด โ‹… ๐‘ฅ = ๐‘๐‘ฅ โ‰ฅ 0

โˆƒ ๐‘ฆ โˆˆ โ„@, s.t.๐ดA โ‹… ๐‘ฆ โ‰ฅ 0๐‘A โ‹… ๐‘ฆ < 0

Page 27: Linear Programming Duality

27

An Alternative of Farkasโ€™ Lemma

Corollary: Exactly one of the following systems holds:

โˆƒ ๐‘ฅ โˆˆ โ„<, s.t.๐ด โ‹… ๐‘ฅ โ‰ค ๐‘๐‘ฅ โ‰ฅ 0

โˆƒ ๐‘ฆ โˆˆ โ„@, s.t.๐ดA โ‹… ๐‘ฆ โ‰ฅ 0๐‘A โ‹… ๐‘ฆ < 0๐‘ฆ โ‰ฅ 0

Proof: Apply Fakasโ€™ lemma to the following linear systems

โˆƒ ๐‘ฅ โˆˆ โ„<, s.t.๐ด โ‹… ๐‘ฅ + ๐ผ โ‹… ๐‘  = ๐‘๐‘ฅ, ๐‘  โ‰ฅ 0

โˆƒ ๐‘ฆ โˆˆ โ„@, s.t.๐ดA โ‹… ๐‘ฆ โ‰ฅ 0๐ผ โ‹… ๐‘ฆ โ‰ฅ 0๐‘A โ‹… ๐‘ฆ < 0

Following corollary of Farkasโ€™ lemma is more convenient for our proof

Page 28: Linear Programming Duality

28

Proof of Strong Duality

Proofร˜Dual of the dual is primal; so w.l.o.g assume primal is feasible and

bounded

ร˜Weak duality yields OPT(primal) โ‰ค OPT(dual) ร˜Next we prove the converse, i.e., OPT(primal) โ‰ฅ OPT(dual)

max ๐‘A โ‹… ๐‘ฅs.t. ๐ด๐‘ฅ โ‰ค ๐‘

๐‘ฅ โ‰ฅ 0

Primal LPmin ๐‘A โ‹… ๐‘ฆs.t. ๐ดA๐‘ฆ โ‰ฅ ๐‘

๐‘ฆ โ‰ฅ 0

Dual LP

Theorem [Strong Duality]: If either the primal or dual is feasibleand bounded, then so is the other and OPT(primal) = OPT(dual).

Page 29: Linear Programming Duality

29

Proof of Strong Duality

ร˜We prove if OPT(primal)< ๐›ฝ for some ๐›ฝ, then OPT(dual)< ๐›ฝร˜Apply Farkasโ€™ lemma to the following linear system

max ๐‘A โ‹… ๐‘ฅs.t. ๐ด๐‘ฅ โ‰ค ๐‘

๐‘ฅ โ‰ฅ 0

Primal LPmin ๐‘A โ‹… ๐‘ฆs.t. ๐ดA๐‘ฆ โ‰ฅ ๐‘

๐‘ฆ โ‰ฅ 0

Dual LP

โˆƒ๐‘ฅ โˆˆ โ„< such that๐ด๐‘ฅ โ‰ค ๐‘โˆ’๐‘A โ‹… ๐‘ฅ โ‰ค โˆ’๐›ฝ๐‘ฅ โ‰ฅ 0

โˆƒ๐‘ฆ โˆˆ โ„@ and ๐‘ง โˆˆ โ„๐ดA๐‘ฆ โˆ’ ๐‘๐‘ง โ‰ฅ 0๐‘"๐‘ฆ โˆ’ ๐›ฝ๐‘ง < 0๐‘ฆ, ๐‘ง โ‰ฅ 0

ร˜By assumption, the first system is infeasible, so the second must holdโ€ข If ๐‘ง > 0, can rescale (๐‘ฆ, ๐‘ง) to make ๐‘ง = 1, yielding OPT(dual)< ๐›ฝโ€ข If ๐‘ง = 0, then system ๐ดA๐‘ฆ โ‰ฅ 0, ๐‘"๐‘ฆ < 0, ๐‘ฆ โ‰ฅ 0 feasible. Farkasโ€™ lemma implies

that system ๐ด๐‘ฅ โ‰ค ๐‘, ๐‘ฅ โ‰ฅ 0 is infeasible, contradicting theorem assumption.

Page 30: Linear Programming Duality

30

ร˜Recap and Weak Duality

ร˜Strong Duality and Its Proof

ร˜Consequence of Strong Duality

Outline

Page 31: Linear Programming Duality

31

Complementary Slackness

max ๐‘A โ‹… ๐‘ฅs.t. ๐ด๐‘ฅ โ‰ค ๐‘

๐‘ฅ โ‰ฅ 0

Primal LPmin ๐‘A โ‹… ๐‘ฆs.t. ๐ดA๐‘ฆ โ‰ฅ ๐‘

๐‘ฆ โ‰ฅ 0

Dual LP

ร˜ ๐‘ & = ๐‘ โˆ’ ๐ด๐‘ฅ & is the ๐‘–โ€™th primal slack variableร˜ ๐‘ก6 = ๐ด"๐‘ฆ โˆ’ ๐‘ 6 is the ๐‘—โ€™th dual slack variable

Complementary Slackness:๐‘ฅ and ๐‘ฆ are optimal if and only if they are feasible andร˜ ๐‘ฅ6๐‘ก6 = 0 for all j = 1,โ‹ฏ ,๐‘šร˜ ๐‘ฆ&๐‘ & = 0 for all ๐‘– = 1,โ‹ฏ , ๐‘›

Remark: can be used to recover optimal solution of the primal from optimal solution of the dual (very useful in optimization).

Page 32: Linear Programming Duality

32

Economic Interpretation of Complementary Slackness: Given the optimal production and optimal raw material pricesร˜ It only produces products for which profit equals raw material

costร˜ A raw material is priced greater than 0 only if it is used up in

the optimal production

max ๐‘" โ‹… ๐‘ฅs.t. โˆ‘6;-< ๐‘Ž&6 ๐‘ฅ6 โ‰ค ๐‘&, โˆ€๐‘– โˆˆ [๐‘š]

๐‘ฅ6 โ‰ฅ 0, โˆ€๐‘— โˆˆ [๐‘›]

Primal LP Dual LP

min ๐‘" โ‹… ๐‘ฆs.t. โˆ‘&;-@ ๐‘Ž&6 ๐‘ฆ& โ‰ฅ ๐‘6, โˆ€๐‘— โˆˆ [๐‘›]

๐‘ฆ& โ‰ฅ 0, โˆ€๐‘– โˆˆ [๐‘š]

Page 33: Linear Programming Duality

33

Proof of Complementary Slackness

max ๐‘A โ‹… ๐‘ฅs.t. ๐ด๐‘ฅ โ‰ค ๐‘

๐‘ฅ โ‰ฅ 0

Primal LPmin ๐‘A โ‹… ๐‘ฆs.t. ๐ดA๐‘ฆ โ‰ฅ ๐‘

๐‘ฆ โ‰ฅ 0

Dual LP

Page 34: Linear Programming Duality

34

Proof of Complementary Slackness

ร˜ Add slack variables into both LPs

max ๐‘A โ‹… ๐‘ฅs.t. ๐ด๐‘ฅ + ๐‘  = ๐‘

๐‘ฅ, ๐‘  โ‰ฅ 0

Primal LPmin ๐‘A โ‹… ๐‘ฆs.t. ๐ดA๐‘ฆ โˆ’ ๐‘ก = ๐‘

๐‘ฆ, ๐‘ก โ‰ฅ 0

Dual LP

๐‘ฆ"๐‘ โˆ’ ๐‘ฅ"๐‘ = ๐‘ฆ" ๐ด๐‘ฅ + ๐‘  โˆ’ ๐‘ฅ" ๐ด"๐‘ฆ โˆ’ ๐‘ก = ๐‘ฆ"๐‘  + ๐‘ฅ"๐‘ก

Page 35: Linear Programming Duality

35

Proof of Complementary Slackness

ร˜ Add slack variables into both LPs

ร˜ For any feasible ๐‘ฅ, ๐‘ฆ, the gap between primal and dual objectivevalue is precisely the โ€œaggregated slacknessโ€ ๐‘ฆ"๐‘  + ๐‘ฅ"๐‘ก

ร˜ Strong duality implies ๐‘ฆ"๐‘  + ๐‘ฅ"๐‘ก = 0 for the optimal ๐‘ฅ, ๐‘ฆ.

ร˜ Since ๐‘ฅ, ๐‘ , ๐‘ฆ, ๐‘ก โ‰ฅ 0, we have ๐‘ฅ6๐‘ก6 = 0 for all j and ๐‘ฆ&๐‘ & = 0 for all ๐‘–.

max ๐‘A โ‹… ๐‘ฅs.t. ๐ด๐‘ฅ + ๐‘  = ๐‘

๐‘ฅ, ๐‘  โ‰ฅ 0

Primal LPmin ๐‘A โ‹… ๐‘ฆs.t. ๐ดA๐‘ฆ โˆ’ ๐‘ก = ๐‘

๐‘ฆ, ๐‘ก โ‰ฅ 0

Dual LP

๐‘ฆ"๐‘ โˆ’ ๐‘ฅ"๐‘ = ๐‘ฆ" ๐ด๐‘ฅ + ๐‘  โˆ’ ๐‘ฅ" ๐ด"๐‘ฆ โˆ’ ๐‘ก = ๐‘ฆ"๐‘  + ๐‘ฅ"๐‘ก

Page 36: Linear Programming Duality

Thank You

Haifeng Xu University of Virginia

[email protected]