Lower Bounds for Distributed Dominating Sets

CPSC 689: Discrete Algorithms
for Mobile and Wireless Systems
Spring 2009
Prof. Jennifer Welch
Lecture 23
 Topic:
 Lower Bounds for Dominating Sets and Related
Problems
 Sources:
 Kuhn, Moscibroda and Wattenhofer, "What Cannot
Be Computed Locally!"
 MIT 6.885 Fall 2008 slides
Discrete Algs for Mobile Wireless Sys
2
Locality
 Locality means that nodes only have to
communicate (even indirectly) with nodes
that are close by
 Desirable property of a distributed
algorithm:
 local algorithms have (the possibility of) low
time complexity
 why bother far away nodes?
Discrete Algs for Mobile Wireless Sys
3
Locality
 k communication
rounds means
being restricted to
a locality radius k.
1 rounds
2
3
Discrete Algs for Mobile Wireless Sys
4
Locality
 Can we find local algorithms for various
distributed problems?
 means time complexity (number of rounds) is
independent of network size
 A few positive results, e.g.:
 Naor & Stockmeyer: studied a class of problems called
locally checkable labelings and showed there are nontrivial LCL problems that have local algorithms,
including a variant of dining philosophers
 What about negative results (lower bounds)?
 Linial: coloring on a ring takes (log*n) rounds
 What about for dominating set and related problems?
Discrete Algs for Mobile Wireless Sys
5
Minimum Vertex Cover
 Minimum Vertex Cover problem: Given a graph,
find smallest subset S of vertices (nodes) such
that every edge is "covered" by a node in S (at
least one endpoint is in S)
 NP-complete
 consider polynomial time approximation algorithms
Discrete Algs for Mobile Wireless Sys
6
Overview of [KMW] Results
 Any k-round MVC algorithm has an approximation
ratio that is (nc/k*k/k), where n is number of nodes
and c is a constant > 1/4
 To ensure that the approximation ratio is no more
than poly-log, k has to be at least
((log n / log log n)), which is not local
 Any k-round MVC algorithm has an approximation
ratio that is (1/k /k), where  is the maximum degree
 To ensure that the approximation ratio is no more
than poly-log, k has to be at least
(log  / log log ), which is not local
Discrete Algs for Mobile Wireless Sys
7
Some Special Case Graphs
 Consider a ring:




minimum VC consists of every other node
constant-time approx algorithm is to include every node
approx ratio w.r.t. n is 2
Generalize to a d-regular graph
 Consider a tree:
 minimum VC consists of every other node down each
branch
 constant-time approx algorithm is to include every nonleaf node
 approx ratio w.r.t. n is 2
Discrete Algs for Mobile Wireless Sys
8
Some More Special Case Graphs
 Consider graphs with constant max degree :
 constant time approx alg is to include every node
 approx ratio w.r.t.  is constant
 Consider graphs that contain nodes with high
degree (say, (n)):
 then diameter is small (say, O(1)), so in constant time,
an alg can learn the entire graph and choose exactly
which nodes to include
 approx ratio is 1
 To show non-locality property, need to consider
more complicated graphs…
Discrete Algs for Mobile Wireless Sys
9
Intuition for Locality-Based Lower
Bounds
 In k rounds of communication (time k), every node
can collect information about its k-neighborhood
 Hence, the solution of a node v in a distributed kround computation can only depend on the k-hop
neighborhood of v
 If two nodes u and v have the same k-hop
neighborhoods, they will make the same decision:
the execution of a k-round algorithm looks the
same to both nodes
Discrete Algs for Mobile Wireless Sys
10
Example for Locality-Based Lower
Bound
 How to prove such a lower bound?
 Let’s look at case k=2 to get the basic
intuition
 After 1 round, nodes know their neighbors
 After 2 rounds, nodes know the neighbors
of their neighbors
Discrete Algs for Mobile Wireless Sys
11
Two-Round Lower Bound
m3
…
…
m
…
…
m
…
m2
complete
m nodes
…
m2
m2-1
m nodes
…
same view
Discrete Algs for Mobile Wireless Sys
12
Hint of Proof
 Construct graph Gk for each k > 0 containing a
bipartite subgraph S with node set C0 U C1
 C0 has n0 nodes, each with d0 neighbors in C1
 C1 has n1 nodes, each with d1 neighbors in C0
 n1 = n0*d0/d1
C1, n1 = 8, d1 = 2
C0, n0 = 4, d0 = 4
Discrete Algs for Mobile Wireless Sys
13
Hint of Proof
 In a globally optimal solution, all edges of S (the
bipartite graph) can be covered by choosing all
nodes of C1, and none of C0, to be in the VC
 But in a local algorithm, decision can only be
made based on k-neighborhood
 Construct Gk so that two adjacent nodes (one in
C0 and one in C1) have the same k-neighborhood
and thus do the same thing (both join the VC)
 Since symmetry cannot be broken in only k
rounds, suboptimal local decisions are made and
a suboptimal approximation ratio achieved
Discrete Algs for Mobile Wireless Sys
14
Constructing Gk
 The heart of the paper is recursive
construction of Gk with high degree of
symmetry
 See Appendix for G3
 What do we do with Gk? Have to consider
what happens with node IDs
Discrete Algs for Mobile Wireless Sys
15
Handling Node IDs
 Assume random node ID assignment with IDs
from {1,…,N}
 If nodes u and v see same topology up to
distance k:
 Every possible ID assignment is equally probable
 Probability to see a particular ID assignment equal for u
and v
 u and v make the same decision with the same
probability
Discrete Algs for Mobile Wireless Sys
16
Handling Node IDs
 Deterministic algorithms: there exists a
node assignment for which solution is at
least as bad as expected value with
random IDs
 Randomized algorithms: Same bound
using Yao’s principle
Discrete Algs for Mobile Wireless Sys
17
Hints on Rest of Proof
 Lemma: Any (randomized or deterministic) kround distributed algorithm, when run on Gk, puts
at least half the nodes of C0 into the VC.
 Proof is based on constructed properties of Gk
and previous discussion about handling IDs.
 So approx ratio is at least (n0/2) / (n – n0), since
optimal solution does not need any node in C0
 Do some math to show that the construction of Gk
can be tweaked to ensure that n0 is sufficiently
large relative to n to show the claimed lower
bounds w.r.t. n and .
Discrete Algs for Mobile Wireless Sys
18
Relationship to Dominating Sets
 Theorem: Every (randomized or deterministic) kround distributed algorithm for MDS has same
asymptotic lower bounds on approx ratio as for
MVC: (nc/k*k/k) and (1/k/k).
 Proof: By reduction.
 Let A be a k-round alg for MDS with approx ratio R.
 Show how to use ADS as a subroutine in algorithm AVC
to approximate MVC with only a constant number of
extra rounds
 Analyze the resulting approx ratio for the MVC problem
Discrete Algs for Mobile Wireless Sys
19
Reduction
Here is algorithm AVC:
1. Suppose input graph for MVC is G'
2. Simulate another graph G
 see next slide
3. Call MDS approx alg ADS on G
4. ADS returns some set of nodes S
5. Return S as an approx MVC for G'
Discrete Algs for Mobile Wireless Sys
20
Reduction
 Transform G' into G:
a
b
d
c
a
ab
b
ad
bd
bc
d
cd
c
Discrete Algs for Mobile Wireless Sys
21
VC to DS
 Any VC of G' is a DS of G:
a
b
d
c
red nodes
cover all edges
a
ab
b
ad
bd
bc
d
cd
c
red nodes
cover all nodes
Discrete Algs for Mobile Wireless Sys
22
DS to VC
 Take any DS of G, replace any green node with a nongreen neighbor; result is a VC of G'
a
b
d
c
red nodes
cover all edges
a
ab
b
ad
bd
bc
d
cd
c
red nodes
cover all nodes
Discrete Algs for Mobile Wireless Sys
23
Relating Quality of Approximations
 Algorithm ADS returns S, a DS of G that is
at most R times as large as an optimal DS
of G
 Size of optimal DS of G is ≤ size of optimal
VC on G'
 since every VC of G' forms a DS of G
 Thus S is a VC of G' that is at most R times
as large as an optimal VC of G'
 By MVC lower bound R must be at least …
Discrete Algs for Mobile Wireless Sys
24
Summary: Lower Bound
 Lower bound shows that the time-approximation
trade-off of the existing algorithm is not too far off the
optimum (there still is a significant gap…)
 By a reduction, the time lower bound for polylog
approximations also holds for the apparently unrelated
problem of computing a maximal independent set
 Remark: The lower bound is obtained by using very
special graphs. This is definitely not how wireless
network graphs look! In fact, for special graph classes,
we can do better.
Discrete Algs for Mobile Wireless Sys
25