28
1 Primal-dual Approximation Algorithms for Integral Flow and Multicut in Trees Garg, N., Vazirani, V. V. and Yannakakis, M., Algorithmica, Vol. 18, 1997, pp. 3-20. Advisor: R. C. T. Lee Speaker: F. L. Lin National Chi Nan University

1 Primal-dual Approximation Algorithms for Integral Flow and Multicut in Trees Garg, N., Vazirani, V. V. and Yannakakis, M., Algorithmica, Vol. 18, 1997,

Embed Size (px)

Citation preview

Page 1: 1 Primal-dual Approximation Algorithms for Integral Flow and Multicut in Trees Garg, N., Vazirani, V. V. and Yannakakis, M., Algorithmica, Vol. 18, 1997,

1

Primal-dual Approximation Algorithms for Integral Flow and

Multicut in Trees

Garg, N., Vazirani, V. V. and Yannakakis, M., Algorithmica, Vol. 18, 1997, pp. 3-20.

Advisor: R. C. T. Lee

Speaker: F. L. Lin

National Chi Nan University

Page 2: 1 Primal-dual Approximation Algorithms for Integral Flow and Multicut in Trees Garg, N., Vazirani, V. V. and Yannakakis, M., Algorithmica, Vol. 18, 1997,

2

Minimum Multicut Problem

Given a graph G=(V, E) with a positive capacity ce on every edge , and a list of vertex pairs, {(s1, t1), …, (sk, tk)}, find a minimum weight set of edges separating each pair of vertices in the list.

– est denotes the edge st.

Ee

Vertex pairs ={(a, f), (c, f), (h, d)}

Multicut = {eah, ebf, eef}

a

h

g

c

d

f

e

b5

7

1

12

3

1 2

1a

h

g

c

d

f

e

b5

7

1

12

3

1 2

1

Page 3: 1 Primal-dual Approximation Algorithms for Integral Flow and Multicut in Trees Garg, N., Vazirani, V. V. and Yannakakis, M., Algorithmica, Vol. 18, 1997,

3

• The minimum multicut problem is NP-hard even if it is restricted to trees of height 1 and unit capacity edges.

• This paper deals with undirected trees.

– denotes the unique path between si and ti in the tree.

vertex pairs ={(b, e), (c, d), (a, f)}pbe= b→g→h→j →i →epcd= c→g→h→j →i →dpaf= a→h→j→r →f

r

j f

h i

a g d e

b c

4 4

3 3

4 2

1 1

1 1

iitsp

Page 4: 1 Primal-dual Approximation Algorithms for Integral Flow and Multicut in Trees Garg, N., Vazirani, V. V. and Yannakakis, M., Algorithmica, Vol. 18, 1997,

4

• Since the minimum multicut problem is NP-hard, we have to find an approximation algorithm for it.

• This paper used the primal-dual approach. But we found that we could give the algorithm without mentioning the primal-dual approach which is quite hard to understand and confusing.

Page 5: 1 Primal-dual Approximation Algorithms for Integral Flow and Multicut in Trees Garg, N., Vazirani, V. V. and Yannakakis, M., Algorithmica, Vol. 18, 1997,

5

• For this problem, we can compute a lower bound of our solutions immediately.

• This lower bound is obtained by reducing the cost of each edge in the path and there will be at least one zero cost in every path.

• Let . We observe that is a lower bound of cutting the path .– For example, fac = min{4, 2, 1} =1 which is a lower bound

of cutting pac.

• ce = ce - , .

• Let Q be the set of edges whose

cost, ce, becomes zero in this step.

iitsp

iitsf

iitsp

h

a g

b c

4 2

1 1

iiii tsets pecf ,min

iitsf iitspe

Page 6: 1 Primal-dual Approximation Algorithms for Integral Flow and Multicut in Trees Garg, N., Vazirani, V. V. and Yannakakis, M., Algorithmica, Vol. 18, 1997,

6

• For example– fac=min{4, 2, 1}=1, the lower bound for cutting the path pac is 1.

– Q={egc}

– fab=min{3,1,1}=1, the lower bound for cutting the path pab is 1.

– fac+fab=2, the lower bound for cutting the path pac and pab is 2.

– Q={egc, ehg, egb}

h

a g

b c

4 2

1 1

vertex pairs ={(a, c), (a, b)}pac= a→h→g→cpab= a→h→g→b

h

a g

b c

3 1

1 0

h

a g

b c

4-1 2-1

1 1-1

h

a g

b c

3 1

1 0

h

a g

b c

3-1 1-1

1-1 0

h

a g

b c

2 0

0 0

Page 7: 1 Primal-dual Approximation Algorithms for Integral Flow and Multicut in Trees Garg, N., Vazirani, V. V. and Yannakakis, M., Algorithmica, Vol. 18, 1997,

7

• Because the graph is a tree, we may omit redundant edges from the multicut.

• When some pairs have the same least common ancestor, v, and the edges are the ancestor and the child in Q, the path passes through the child and the path must pass through the ancestor. We only retain the ancestor.

• Let frontier (v) = Q – the edge that is the descendant of other edges in Q.

Claim

The union of all frontiers is a multicut.

Page 8: 1 Primal-dual Approximation Algorithms for Integral Flow and Multicut in Trees Garg, N., Vazirani, V. V. and Yannakakis, M., Algorithmica, Vol. 18, 1997,

8

• For example, pairs (a, c) and (a, b) have the same least common ancestor, h, and ehg is the ancestor of egb and egc. The edge egb and egc are omitted. frontier(h)={ehg}.

0

0h

a g

b c

2

0

vertex pairs = {(a, c), (a, b)}

pac= a→h→g→cpab= a→h→g→b

Q={egc, ehg, egb}

frontier(h)={ehg}

Page 9: 1 Primal-dual Approximation Algorithms for Integral Flow and Multicut in Trees Garg, N., Vazirani, V. V. and Yannakakis, M., Algorithmica, Vol. 18, 1997,

9

• We move down the tree one level at a time and join frontiers to build the multicut.

• Considering vertex v, we include an edge only if no edge along the path from e to v is already included in the multicut.

• Let M be the set of edges picked.

• For example, suppose that M={efg} and consider frontier(f), efg is in the path from egd to f, egd is not picked to M.

LemmaM is a multicut.

vfrontiere

vertex pairs = {(b, d),(a, c)}pbd= b→f→g→dpac= a→e→f→g→cfrontier(e) = {efg}M = {efg}frontier(f) = {egd}

e

a f

b g

c d

0

02

1

2 2

Page 10: 1 Primal-dual Approximation Algorithms for Integral Flow and Multicut in Trees Garg, N., Vazirani, V. V. and Yannakakis, M., Algorithmica, Vol. 18, 1997,

10

Algorithm multicut_integral-flow(r)

Input : a tree T = (V, E) with a positive capacity ce on every edge , and a list of vertex pairs, {(s1, t1), …, (sk, tk)}.

1. for current_level = max_level downto 0 dofor all current_level do

1.1 for each pair (si, ti) such that its least common ancestor is at v do

denotes the unique path between si and ti in the tree.

= min{ ce}, .

ce = ce- , .

1.2 Let Q be the set of edges such that ci=0 in this step. frontier (v) = Q – the edge that is the descendant of other edges in Q.

2. 2.1 2.2 for current_level = 0 to max_level do

for all current_level dofor all do If no edge is on the path from e to v then

3. return (Multicut, F)

v

v vfrontiere

iitsf iitspeiitsf

Ee

iitspeiitsp

Multicut

eMulticutMulticut

Page 11: 1 Primal-dual Approximation Algorithms for Integral Flow and Multicut in Trees Garg, N., Vazirani, V. V. and Yannakakis, M., Algorithmica, Vol. 18, 1997,

11

1

33

2

1

• Example

vertex pairs ={(b, e), (c, d), (a, f)}pbe= b→g→h→j →i →epcd= c→g→h→j →i →dpaf= a→h→j→r →f

r

j f

h i

a g d e

b c

4 4

4

1

1

Pass 1.1The first node which is a least common ancestor of some vertex pair is j.j = lca(b, e)fbe = min{1, 2, 3, 3, 1}=1Subtract 1 from all edges in pbe.

Page 12: 1 Primal-dual Approximation Algorithms for Integral Flow and Multicut in Trees Garg, N., Vazirani, V. V. and Yannakakis, M., Algorithmica, Vol. 18, 1997,

12

Pass 1.1Consider the pair (c, d) whose lca is j.j = lca(c, d)fcd = min{1, 1, 2, 2, 1}=1Subtract 1 from all edges in pcd.The lower bound of cutting the paths pbe and pcd is fbe+ fcd=2.1

0

22

1

0

• Example

r

j f

h i

a g d e

b c

4 4

4 1

fbe = 1fcd = 1

vertex pairs ={(b, e), (c, d), (a, f)}pbe= b→g→h→j →i →epcd= c→g→h→j →i →dpaf= a→h→j→r →f

Page 13: 1 Primal-dual Approximation Algorithms for Integral Flow and Multicut in Trees Garg, N., Vazirani, V. V. and Yannakakis, M., Algorithmica, Vol. 18, 1997,

13

0

0

0

11

0

0

• Example

r

j f

h i

a g d e

b c

4 4

4

vertex pairs ={(b, e), (c, d), (a, f)}pbe= b→g→h→j →i →epcd= c→g→h→j →i →dpaf= a→h→j→r →f

Pass 1.2Let Q be the set of edges such that ci=0.Q = {egb, eie , ehg , egc , eid}.

Pairs (b,e) and (c,d) have the same least common ancestor, j, and ehg is the ancestor of egb and egc in Q. egb and egc are not added to frontier(j). frontier(j) = Q - {egb, egc} = {eie, ehg, eid}.

fbe = 1fcd = 1

Page 14: 1 Primal-dual Approximation Algorithms for Integral Flow and Multicut in Trees Garg, N., Vazirani, V. V. and Yannakakis, M., Algorithmica, Vol. 18, 1997,

14

4

4 4

0

0

0

11

0

0

• Example

r

j f

h i

a g d e

b c

vertex pairs ={(b, e), (c, d), (a, f)}pbe= b→g→h→j →i →epcd= c→g→h→j →i →dpaf= a→h→j→r →f

Pass 1.1Consider the pair (a, f) whose lca is r. r = lca(a, f)faf = min {4, 1, 4, 4}=1Minus 1 from all edges in paf. The lower bound of cutting the path pbe, pcd and paf is fbe+ fcd + faf=3.

fbe = 1fcd = 1faf = 1frontier(j)={eie, ehg, eid}

Page 15: 1 Primal-dual Approximation Algorithms for Integral Flow and Multicut in Trees Garg, N., Vazirani, V. V. and Yannakakis, M., Algorithmica, Vol. 18, 1997,

15

3

3 3

0

0

0

10

0

0

• Example

r

j f

h i

a g d e

b c

vertex pairs ={(b, e), (c, d), (a, f)}pbe= b→g→h→j →i →epcd= c→g→h→j →i →dpaf= a→h→j→r →f

Pass 1.2In this step, Q = {ejh} and frontier(j) = Q - { } = {ejh}.

frontier(j)={eie, ehg, eid}frontier(r)={ejh}.

Page 16: 1 Primal-dual Approximation Algorithms for Integral Flow and Multicut in Trees Garg, N., Vazirani, V. V. and Yannakakis, M., Algorithmica, Vol. 18, 1997,

16

3

3 3

0

0

0

10

0

0

• Example

r

j f

h i

a g d e

b c

vertex pairs ={(b, e), (c, d), (a, f)}pbe= b→g→h→j →i →epcd= c→g→h→j →i →dpaf= a→h→j→r →f

Pass 2.

Consider the vertex up to down to build the multicut.

Multicut = { }.

The first node of least common ancestor is r.

frontier(r) = {ejh}.

We check ejh. No edge in Multicut is on the path from ejh to r, ejh is included in Multicut.

Multicut = {ejh}.

frontier(j)={eie, ehg, eid}frontier(r)={ejh}

Page 17: 1 Primal-dual Approximation Algorithms for Integral Flow and Multicut in Trees Garg, N., Vazirani, V. V. and Yannakakis, M., Algorithmica, Vol. 18, 1997,

17

3

3 3

0

0

0

10

0

0

• Example

r

j f

h i

a g d e

b c

vertex pairs ={(b, e), (c, d), (a, f)}pbe= b→g→h→j →i →epcd= c→g→h→j →i →dpaf= a→h→j→r →f

Pass 2.

Consider the node j.

frontier(j) = {eie, ehg, eid}.

Check eie. No edge in Multicut is on the path from eie to j, eie is included in Multicut.

Multicut = {ejh, eie}.

Check ehg. ejh along the path from ehg to j is already included in Multicut, ehg is not selected.

Multicut = {ejh, eie}.

x

Page 18: 1 Primal-dual Approximation Algorithms for Integral Flow and Multicut in Trees Garg, N., Vazirani, V. V. and Yannakakis, M., Algorithmica, Vol. 18, 1997,

18

3

3 3

0

0

0

10

0

0

• Example

r

j f

h i

a g d e

b c

vertex pairs ={(b, e), (c, d), (a, f)}pbe= b→g→h→j →i →epcd= c→g→h→j →i →dpaf= a→h→j→r →f

Pass 2.

Continue to consider the node j.

frontier(j) = {eie, ehg, eid}.

Check eid. No edge in Multicut is on the path from eid to j, eid is included in Multicut.

Multicut = {ejh, eie, eid}.

Cost of Multicut = 3+1+1 = 5.

The optimal solution is actually ejh where cost is 3.

Page 19: 1 Primal-dual Approximation Algorithms for Integral Flow and Multicut in Trees Garg, N., Vazirani, V. V. and Yannakakis, M., Algorithmica, Vol. 18, 1997,

19

Lemma

Let si-ti be a path and let v be the least common ancestor of si, ti. Then Multucut contains at most one edge from the path si-v and one edge from the path ti-v.

• Whenever an edge e is chosen, its cost must be a summation of ’s. That is, for every ce in Multicut.

itis

iipei

tse fc:

iitsf

v

si ti

Page 20: 1 Primal-dual Approximation Algorithms for Integral Flow and Multicut in Trees Garg, N., Vazirani, V. V. and Yannakakis, M., Algorithmica, Vol. 18, 1997,

20

• The cost of Multicut is , then

, where l=|Multicut|.

• Thus, .

• From the algorithm, every corresponds to a path .

• From Lemma 1, we know that for every path , only two edges of this path can be chosen.

Multicute

ec

itisl

ii

itis

ii

itis

iipei

tspei

tspei

tsMulticute

e fffc:::

...21

kk tsktstsMulticute

e fafafac

...2211 21

iitsf

iitsp

iitsp

Page 21: 1 Primal-dual Approximation Algorithms for Integral Flow and Multicut in Trees Garg, N., Vazirani, V. V. and Yannakakis, M., Algorithmica, Vol. 18, 1997,

21

• Thus, the cost of every edge of path which is chosen contains only one .

• Finally, only two ‘s appear in .

• Thus,

• .

• Therefore the cost of Multicut is at most .

Multicutlffffcitis

ii

itisl

ii

itis

ii

itis

iipi

tspei

tspei

tspei

tsMulticute

e

,2...::: 21

k

i ts iif

12

iitsp

iitsf

iitsf Multicute

ec

.1 ,2 kiai

Page 22: 1 Primal-dual Approximation Algorithms for Integral Flow and Multicut in Trees Garg, N., Vazirani, V. V. and Yannakakis, M., Algorithmica, Vol. 18, 1997,

22

• The algorithm achieves approximation guarantees of factor 2 for the minimum multicut problem on tree.

. 1

optimalfk

i ts ii

.221

optimalfck

its

Multicutee ii

Page 23: 1 Primal-dual Approximation Algorithms for Integral Flow and Multicut in Trees Garg, N., Vazirani, V. V. and Yannakakis, M., Algorithmica, Vol. 18, 1997,

23

• This paper interpreted the dual program as specifying a multicommodity flow in T, with a separate commodity corresponding to each vertex pair (si, ti).

• will denote the amount of this commodity routed along the unique path from si to ti.

• For example:

vertex pairs ={(b, e), (c, d), (a, f)}pbe= b→g→h→j →i →epcd= c→g→h→j →i →dpaf= a→h→j→r →f

2

1

1

1

4

44

3

r

j f

h i

a g d e

b c

3

1

iitsf

Page 24: 1 Primal-dual Approximation Algorithms for Integral Flow and Multicut in Trees Garg, N., Vazirani, V. V. and Yannakakis, M., Algorithmica, Vol. 18, 1997,

24

• Let the value of which gets from the algorithm be the solution of the maximum integer multicommodity flow problem.

• The algorithm achieves approximation guarantees of factor 1/2 for the maximum integer multicommodity flow problem on trees.

.21

k

i tsMe

e iifcoptimal

kifiits 1,

.2

11

k

i ts iifoptimal

Page 25: 1 Primal-dual Approximation Algorithms for Integral Flow and Multicut in Trees Garg, N., Vazirani, V. V. and Yannakakis, M., Algorithmica, Vol. 18, 1997,

25

Theorem The algorithm achieves approximation guarantees

of factor 2 for the minimum multicut problem and factor 1/2 for the maximum integer multicommodity flow problem on trees.

Page 26: 1 Primal-dual Approximation Algorithms for Integral Flow and Multicut in Trees Garg, N., Vazirani, V. V. and Yannakakis, M., Algorithmica, Vol. 18, 1997,

26

Reference[ALMSS92] Proof verification and the hardness of approximation problems, Arora, S., Lund, C., Motwani, R., Sudan, M. and

Szegedy, M., Proceedings, 33rd IEEE Symposium on Foundations of Computer Science, 1992, p.p. 14-23.[BE81] A linear time approximation algorithm for the weighted vertex cover problem, Bar-Yehuda, R. and Even, S., J.

Algorithms, Vol. 2., 1981, p.p. 198-203.[BGLR93] Efficient probabilistically checkable proofs and applications to approximation, Bellare, M., Goldwasser, S., Lund,

C. and Russel, A., Proceedings 25th Annual ACM Symposium on Theory of Computing, 1993, p.p. 294-305.[B76] Graph and Hypergraphs, Berge, C., North-Holland, Amsterdam, 1976.[BW88] An almost linear time algorithm for graph realization, Bixby, R. E. and Wagner, D. K., Math. Oper. Res., Vol. 13,

1988, p.p. 99-123. [C77] Solution of a problem of multicommodity flows in a network (in Russian), Cherkasskij, B. V., Mat. Metody, Vol. 13,

1977, p.p. 143-151.[CR91] On the multiway cut polyhedron, Chopra, S. and Rao, M. R., Networks, Vol. 21, 1991, p.p. 51-89.[DJPSY94] The complexity of multiterminal cuts, Dahlhaus, E., Johnson, D. S., Papadimitriou, C. H., Seymour, P. D. and

Yannakakis, M., SIAM J. comput., Vol. 23, 1994, p.p. 864-894.[EIS76] On the complexity of timetable and multicommodity flow problems, Even, S., Itai, A. and Shamir, A., SIAM J.

Comput., Vol. 5, 1976, p.p. 691-703.[F91] Packing paths, circuits and cuts-a survey, Frank, A., Algorithms and Combinatorics, Vol. 9, 1991.[G83] An efficient reduction technique for degree-constrained subgraph and bidirected network flow problem, Gabow, H. N.,

Proceedings, 15th Annual ACM Symposium on Teory of Computing, 1983, p. p. 448-456.[GVY93] Approximate max-flow min-(multi)cut theorems and their applications, Garg, N., Vazirani, V. V. and Yannakakis,

M., Proceddings, 25th Annual ACM Symposium on Theory of computing, 1993, p. p. 689-707.[GVY94] Approximation algorithm for multiway cuts in node-weighted and directed graphs, Garg, N., Vazirani, V. V. and

Yannakakis, M., Proceedings 21st International Colloquium on Automata, Languages and Programming, 1994, p. p., 487-498.

[GW92] A general approximation technique for constrained forest problems, Goemans, M. X. and Williamson, D. P., SIAM J. comput., Vol. 24, 1995, p. p. 296-317.

[GW95] The primal-dual method for approximation algorithms and its application to network design problems. Goemans, M. X. and Williamson, D. P., Approximation Algorithms for NP-hard Problems, 1995, P. P. 144-191.

Page 27: 1 Primal-dual Approximation Algorithms for Integral Flow and Multicut in Trees Garg, N., Vazirani, V. V. and Yannakakis, M., Algorithmica, Vol. 18, 1997,

27

Reference[GLS88] Geometric Algorithms and Combinatorial Optimization, Grotschel, M., Lovasz, L. and Schrijver, A., Spring-Verlag,

Berlin, 1988.[H69] Integer Programming and Network Flows, Hu, T. C., Addison-Wesley, Reading, Ma, 1969.[K92] On the approximability of NP-complete optimization problems, Kann, V., Th. D. Thesis, Royal Institute of Technology,

Stockholm, 1992.[KARR90] Approximation through multicommodity flow, Klein, P., Agrawal, A. and Rao, S., Proceedings 31st IEEE

Symposium on Foundations of Computer Science, 1990, p. p. 726-737.[KP20] Tight integral duality gap in the Chinese postman problem, Korach, E. and Penn, M., Computer Science Department,

Israel Institute of Technology, Haifa, 1989.[L76] On some connectivity properties of eulerian graphs, Lovasz,. L., Acta Math,. Akad. Sci. Hungar., Vol. 28, 1976, p. p.

129-138.[LR88] An approximate max-flow min-cut theorm for uniform multicommodity flow problem with application to

approximation algorithm, Leighton, F. T. and Rao, S., Proceedings 29th Symposium on Foundations of computer Science, 1988, p. p. 422-431.

[LY93] On the hardness of approximating minimization problems, Lund, C. and Yannakakis, M., J. Assoc. comput. Mach., Vol. 41, No. 5, 1994, p. p. 960-981.

[M78] Uber die maximalzahl kantendisjunkter a-wege, Mader, W., Arch. Math., Vol. 30, 1978, p. p. 325-336.[PY91] Optimization approximation and complexiyt classes, Papadimitriou, C. H. and Yannakakis, M., J. Comput, System

sci,., Vol. 43, 1991, p. p. 425-440.[RS95] Graph minors XIII: The disjoint path problem, Rabertson, N. and Seymour, P. D., J. Combin. Theory Ser. B, Vol. 63,

1995, p. p. 65-110.[SS93] Integer multicommodity flows with reduced demands, Srivastav, A. and Stangier, P., Proceedings European

Symposium on algorithms, 1993, p. p. 360-372.[T60] An algorithm for determining whether a given binary matroid is graphic, Tutte, W. T., Proc. Amer. Math. Soc. Vol. 11,

1960, p. p. 905-917.[WGMV93] A primal-dual approximation algorithm for generalized steiner network problems, Williamson, D. P., Goemans,

M. X., Mihail, M. and Vazirani, V. V., Proceedings, 25th Annual ACM symposium on Theory of Computing, 1993, p. p. 708-717.

[YKCP83] Cutting and partitioning a graph after a fixed pattern, Yannakakis, M., Kanellakis, P. C., Cosmadakis, S. C. and Papadimitriou, C. H., In Automata, Languages and Programming, Vol. 154, 1983, p. p. 712-722.

Page 28: 1 Primal-dual Approximation Algorithms for Integral Flow and Multicut in Trees Garg, N., Vazirani, V. V. and Yannakakis, M., Algorithmica, Vol. 18, 1997,

28

Thank you.