Multicut Lower Bounds via Network Coding Anna Blasiak Cornell University
Dec 15, 2015
MulticutGiven:• Directed graph G = (V,E)• Capacities on edges• k source-sink pairs
Find:• A min-cost subset of E
such that on removal all source-sink pairs are disconnected
s1 s2 s3
t1 t2 t3
Directed Multicut The State of the Art
• Õ(n11/23) - approx algorithm [Agarwal, Alon, Charikar ’07]
• 2Ω(log1-ε n) hardness [Chuzhoy, Khanna ‘09]
• Nothing non-trivial known in terms of k
Dual Problem:Maximum Multicommodity Flow
Given:• A directed graph G = (V,E)• Capacities on nodes• k source-sink pairs
Find:• A maximum total weight
set of fractional si-ti paths.
s1 s2 s3
t1 t2 t3
All approximation algorithms for multicut are based on the LP and compare to the maximum multicommodity flow.
Limited by large integrality gap: Ω(min((k, nδ)) [Saks, Samorodnitsky,Zosin ’04, Chuzhoy, Khanna ’09 ]
Better Lower Bound?Network Coding
• Good News:Coding Rate ≥ Flow Rate, can be a factor k larger
• Bad News:Multicut ≱ Coding Rate, can be a factor k smaller
Results
• Identify a property P of a network code such that any code satisfying P is a lower bound on the multicut.
• Show P is preserved under a graph product.• Main Corollary: – Improved and tight lower bound on the multicut
in the construction of Saks et al. and give a network code with rate = min multicut
Simplifying Assumptions
• A network is:• Undirected graph G = (V,E )• Capacity one for each node • Subsets of V: {Si}i ∈[k] , {Ti} i ∈[k] si - ti pairs, i ∈[k], connect to G: • si connects to v ∈Si • v ∈Ti connect to ti
Type of networks giving: Ω(min((k, nδ)) [Saks, Samorodnitsky,Zosin ’04, Chuzhoy, Khanna ’09 ]
Saks et al. Construction
• Begin with The n-path network Pn :v1 v2 v3 v4 v5 vn-1 vn
s1 t1Hypergrid(n, k ) is the k-fold strong product of Pn . It has nk nodes and k s-t pairs and flow rate nk-1Saks et al. show it has multicut at least has knk-1 THEOREMHypergrid(n,k) has a code with rate nk - (n-1)k that is a lower bound on the multicut.
S1 T1
Hypergrid(3,2) = P3 ☒P3
s2
t2
s1
(v2 , v1’ )(v1 , v2’ ) (v3 , v2’ )(v2 , v2’ )
(v2 , v3’ )t1
(v1 , v1’ ) (v3 , v1’ )
(v1 , v3’ ) (v3 , v3’ )
Coding Matrix s2
t2
s1
(v2 , v1’ )(v1 , v2’ ) (v3 , v2’
)(v2 , v2’ )(v2 , v3’ )
t1(v1 , v1’ ) (v3 , v1’
)
(v1 , v3’ ) (v3 , v3’ )
(v1 ,v1’ )(v 2 ,v1’ )
(v3 ,v1’ )
(v1 ,v2’ )(v2 ,v2’ )
(v3 ,v2’ )(v1 ,v3’ )
(v2 ,v3’ )(v3 ,v3’ )
a1 1 1 1 0 0 0 0 0 0b1 0 0 0 1 1 1 0 0 0c1 0 0 0 0 0 0 1 1 1a2 1 0 0 1 0 0 1 0 0b2 0 1 0 0 1 0 0 1 0• Column v describes the linear combination of messages sent by v.
• Column v is a linear combination of columns of predecessors of v.
• ti can decode messages from si.
Columns labeled
with v in V
• Rows labeled with messages• ai , bi , ci originate at si • rate of code = # of messages
a1+a2 b1
a1+b2
a1
c1
b1+b2c1+a
2c1+b2
a2, b2, c2
a1, b1, c1b1+a2
Entries in finite field
Coding Matrix as a Lower Bound
L gives a lower bound on the multicut:For M a minimum multicut, |M| = rank (IM) ≥ rank (LIM) ≥ p.
DEFINITIONA coding matrix L is p-certifiable if 1. For any multicut M, rank(LIM) ≥ p. 2. Column v of L is a linear combination of columns of
incoming sources and predecessors of v that form a clique.
|V|x|M| matrix, column v ∈ M is an indicator vector for v.
Main Theorem*
Given networks N1and N2 with coding matrices L1and L2, there is a coding matrix for N1 ☒N2 :
L = such that:
1. If L1and L2 have rates p1and p2 then L has rate p := n1p2 + n2 p1 - p1p2.2. If L1and L2 are p1and p2 certifiable then L is p-
certifiable.
In1 ⊗ L2L1 ⊗ In2
Saks et al. Construction
• Begin with The n-path network Pn :v1 v2 v3 v4 v5 vn-1 vn
s1 t1Hypergrid(n, k ) is the k-fold strong product of Pn . It has nk nodes and k s-t pairs and flow rate nk-1Saks et al. show it has multicut at least has knk-1 THEOREMHypergrid(n,k) has a code with rate nk -(n-1)k that is a lower bound on the multicut.
Pn is 1-certifiable.
Conclusions and Open Questions
• Our result: a certain type of network coding solution is a lower bound on directed multicut– Is there a more general class of network coding
solutions that is a lower bound?• Multicommodity flow can be far from the
multicut, what about the network coding rate?– Does there exist an α = o(k) s.t.
multicut ≤ α network coding rate?