graph limit theory: algorithms lászló lovász eötvös loránd university, budapest may 20121
TRANSCRIPT
Graph limit theory:
Algorithms
László Lovász
Eötvös Loránd University, Budapest
May 2012 1
May 2012 2
General questions about extremal graphs
- Which inequalities between subgraph densities are valid?
- Which graphs are extremal?
- Can all valid inequalities be proved using just Cauchy-Schwarz?
- Local vs. global extrema
May 2012 3
Algorithms for very large graphs
Parameter estimation: edge density, triangle density, maximum cut
Property testing: is the graph bipartite? triangle-free? perfect?
Computing a structure: find a maximum cut, regularity partition,...
4May 2009
The maximum cut problem
maximize
Applications: optimization, statistical mechanics…
NP-hard, even with 6% error HastadPolynomial-time computable with 13% error Goemans-Williamson
5May 2009
Max cut in dense graphs How to estimate the density?cut with many
edges
Sampling O(4) nodes the density of max cut in the induced subgraph
is closer than to the density of the max cut in the whole
graphwith large probability.
Alon-Fernandez de la Vega-Kannan-Karpinski
For a graph parameter f, the following are equivalent:
(i) f can be estimated from samples
(ii) (Gn) convergent f(Gn) convergent
(iii) f is unifomly continuous w.r.t. dX
Estimable graph parameters
A graph parameter f can be estimated from samples iff
(a) f(G(k)) is convergent as k, and
(b) If V(Gn)=V(Hn), d(Gn,Hn)0, then f(Gn)-f(Hn)0, and
(c) If |V(Gn)|, vV(Gn), then f(Gn)-f(Gn\v)0
Density of maximum cut is estimable.
Estimable graph parameters
Testable graph properties
P testable: there is a test property P’, such that
(a) for every graph G ∈ P and every k ≥ 1,
G(k,G) ∈ P′ with probability at least 2/3, and
(b) for every ε > 0 there is a k0 ≥ 1 such that
for every graph G with d1(G,P) > ε
and every k ≥ k0 we have G(k,G) ∈ P′
with probability at most 1/3.
P: graph property
Testable graph properties
Example: triangle-free
Removal Lemma:
’ if t(,G)<’, then we can delete n2 edges
to get a triangle-free graph.
Ruzsa - Szemerédi
G’: sampled induced subgraph
G’ not triangle-free G not triangle free
G’ triangle-free with high probability, G has few triangles
Testable graph properties
Every hereditary graph property is testable
Alon-Shapira
inherited by induced subgraphs
Nondeterministically testable graph properties
Divine help: coloring the nodes, orienting and coloring the edges
Q: property of directed, colored graphs
Q’={G’: GQ}; “shadow of Q”
G: directed, (edge)-colored graph
G’: forget orientation, delete some colors, forget coloring; “shadow of G”
P nondeterministically testable: P=Q’, where Q is a testable property of colored directed graphs.
Nondeterministically testable graph properties
Examples: maximum cut contains at least n2/100 edges
constains a spanning subgraph with a testable property P
we can delete n2/100 edges to get a graph with a testable property P
Nondeterministically testable graph properties
Every nondeterministically testable graph property is testable.
L-Vesztergombi
H1, H2, ...
in Q
Hn’=Gn
... J2, J1
Jn’=Fn
close to Q
G1, G2, ... ... F2, F1
in P
far from P
Pure existence proof
of an algorithm...
May 2012 14
Computing a structure: representative set
Representative set of nodes: bounded size, (almost) every node is “similar” to one of the nodes in the set
When are two nodes similar? Neighbors? Same neighborhood?
May 2012 15
sim( , ) : E E ( ) E ( )v u su vu w wvtwa a ad t as = -
This is a metric, computable in the sampling model
Similarity distance of nodes
st
v
wu
Representative set U:
for any two nodes in U, dsim(s,t) >
for most nodes, dsim(U,v)
May 2012 16
Voronoi diagram= weak regularity
partition
Representative set and regularity partition
May 2012 17
Representative set and regularity partition
If P = {S1, . . . , Sk} is a partition of V(G) such that d(G,GP) = , then we can select nodes viSi such that the average similarity distance from S = {v1, . . . , vk} is at most 4.
If SV and the average similarity distance from S is ,then the Voronoi cells of S form a partition P such that d(G,GP) 8.
L-Szegedy
18May 2009
Max cut in dense graphs What answer to expect?
- Cannot list for all nodes
For any given node, we want to tell on which side of the cut it lies
(after some preprocessing)
May 2012 19
Construct representative set U
How to compute a (weak) regularity partition?
Each node is in same class as closest representative.
May 2012 20
- Construct representative set U
- Compute pij = density between classes Vi and Vj
(use sampling)
- Compute max cut (U1,U2) in complete graph on U
with edge-weights pij
How to compute the maximum cut?
(Different algorithm implicit by Frieze-Kannan.)
Each node is on same side as closest representative.
May 2012 21
Bounded degree (d)
- We can sample a uniform random node a bounded number of times, and explore its neighborhood to a bounded depth.
Algorithms for bounded degree graphs
Maximum cut cannot be estimated in this model
(random d-regular graph vs. randombipartite d-regular graph)
May 2012 22
Algorithms for bounded degree graphs
Cellular automata (1970’s): network of finite automata
Distributed computing (1980’s): agent at every node, bounded time
Constant time algorithms (2000’s): bounded number of nodes sampled, explored at bounded depth
May 2012 23
Distributed computing model
May 2012 24
Algorithms for bounded degree graphs
(Almost) maximum matching can be computedin bounded time.
Nguyen-Onak
(Almost) maximum flow and (almost) minimum cut can be computed in bounded time.
Csoka
May 2012 25
Distributed computing model
May 2012 26
Algorithms for bounded degree graphs
(Almost) maximum matching can be computedin bounded time.
Nguyen-Onak
(Almost) maximum flow and (almost) minimum cut can be computed in bounded time.
Csóka
need local random bits
no random bits need global random
bits
May 2012 27
Thanks, that’s all!