Post on 24-Dec-2015
Expander graphs – applications and combinatorial constructions
Avi WigdersonIAS, Princeton
[Hoory, Linial, W. 2006] “Expander graphs and applications”Bulletin of the AMS.
Applications
in Math & CS
Applications of Expanders
In CS
• Derandomization
• Circuit Complexity
• Error Correcting Codes
• Data Structures
• …
• Computational Information
• Approximate Counting
• Communication & Sorting Networks
Applications of Expanders
In Pure Math
• Graph Theory - …
• Group Theory – generating random group elements [Babai,Lubotzky-Pak]
• Number Theory Thin Sets [Ajtai-Iwaniec-Komlos-Pintz-Szemeredi] -Sieve method [Bourgain-Gamburd-Sarnak]
- Distribution of integer points on spheres [Venkatesh]
• Measure Theory – Ruziewicz Problem [Drinfeld, Lubotzky-Phillips-Sarnak], F-spaces [Kalton-Rogers]
• Topology – expanding manifolds [Brooks]
- Baum-Connes Conjecture [Gromov]
Expander graphs:
Definition and basic properties
Expanding Graphs - Properties
• Combinatorial: no small cuts, high connectivity• Probabilistic: rapid convergence of random walk• Algebraic: small second eigenvalue
Theorem. [Cheeger, Buser, Tanner, Alon-Milman, Alon, Jerrum-Sinclair,…]: All properties are equivalent!
Expanding Graphs - Properties
• Combinatorial: no small cuts, high connectivity
• Geometric: high isoperimetry
G(V,E)
V vertices, E edges
|V|=n ( ∞ )
d-regular (d fixed)
d
S |S|< n/2
|E(S,Sc)| > α|S|d (what we expect in a random graph)
α constant
S
Expanding Graphs - Properties
• Probabilistic: rapid convergence of random walk
G(V,E)
d-regular
v1, v2, v3,…, vt,…
vk+1 a random neighbor of vk
vt converges to the uniform distribution
in O(log n) steps (as fast as possible)
Expanding Graphs - Properties
• Algebraic: small second eigenvalue
G(V,E) V
V AG
AG(u,v) =
normalized adjacency matrix
(random walk matrix)
0 (u,v) E1/d (u,v) E
1 = 1 ≥ 2 ≥ … ≥ n ≥ -1
(G) = maxi>1 |i| =
max { AG v : v =1, vu }
(G) ≤ < 1
1-(G) “spectral gap”
Expanders - Definition
Undirected, regular (multi)graphs.
G is [n,d]-graph: n vertices, d-regular.
Theorem: An [n,d]-graph G is connected iff (G) <1.
Indeed, if G is (non-bip) connected than (G) <1- 1/dn2
G is [n,d, ]-graph: (G) . G expander if <1.
Definition: An infinite family {Gi} of [ni,d, ]-graphs is an expander family if for all i <1 .
Random walk convergence (algebraic exp. fast mixing)
G(V,E) V
V AG
AG(u,v) =0 (u,v) E1/d (u,v) E
1 = 1 ≥ 2 ≥ … ≥ n eigenvalues
u = v1 v2 … vn eigenvectors
p: any probability distribution on V
p(t)= (AG)tp: distribution after t steps of a random walk
p(t) – u1 n p(t) – u = n (AG)t (p–u) n ((G))t
Converges in O(log n) steps. Corollary: Diam (G) < O(log n).
Expander mixing lemma [Alon-Chung]
(algebraic exp. combinatorial exp.)G(V,E) V
V AG
AG(u,v) =0 (u,v) E1/d (u,v) E
1 = 1 ≥ 2 ≥ … ≥ n eigenvalues
u = v1 v2 … vn eigenvectors
S V 1S = i ivi 1 = |S|/n
T V 1T = i ivi 1 = |T|/n
1SAG1T = i iii = |S||T|/n + i>1 iiI
||E(S,T)|- d|S||T|/n | d(G)i>1ii d(G) |S||T| d(G)n
Expected cut size:like a random graph AG J/n
Existence, explicit construction and basic parameters
Existence of sparse expanders
Theorem [Pinsker] Most 3-regular graphs are expanders.
Proof: The probabilistic method (an early application).
Take a random 3-regular graph G(V,E) with |V|=n.
Fix a set S V, |S| = s < n/2.
Pr[|E(S,Sc)| < .1s] < (s/n)3s
Pr[ G is not expanding ] < s ( ) (s/n)3s < 1/2
Challenge: Explicit (small degree) expanders!
n s
Explicit constructionG(V,E) V
V AG
AG(u,v) =0 (u,v) E1/d (u,v) E{Gi(Vi,Ei)} infinite sequence of graphs
Weakly explicit:
AGi can be constructed in poly (|Vi|) time
Strongly explicit:
AGi can be constructed in polylog (|Vi|) time:
A poly(n) algorithm, given i, u Vi, lists all v Vi s.t. (u,v) Ei
How small can (G) be?
Amplification: Consider Gk
G is [n,d, ]-graph (with <1) then
Gk is [n,dk, k ]-graph
So increasing d we can make (G) = 1/dc for some c>0
Theorem: [Alon-Boppana] An infinite family {Gi} of [ni,d]-graphs must have (G) > (2 (d-1))/d 1/d
Proof: (of a weaker statement). Assume d < n/2.
n/d = Tr[AGAGt] = i i
2 < 1+(n-1)(G)2
1/2d < (G)
V
V AG
Basic consequences: G [n,d,]-graph
Theorem. If S,T V, (u,v) E a random edge, then
| Pr[u S and v T] - (S) (T) | <
Cor 1: A random neighbor of a random vertex in S will land in S with probability < (S)
Cor 2: Every set of size > n contains an edge.
Chromatic number (G) > 1/
Graphs of large girth and chromatic number
Cor 3: Removing any fraction < of the edges leaves a connected component of 1-O() of the vertices.
Fault-tolerant computation
Infection Processes: G [n,d,]-graph, <1/4
Cor 4: Every set S of size s < n/2 contains at most s/2 vertices with > 2s neighbors in S
Infection process 1: Adversary infects I0, |I0| n/4.
I0=S0, S1, S2, …St,… are defined by:
v St+1 iff a majority of its neighbors are in St.
Fact: St= for t > log n [infection dies out]
Proof: |Si+1|≤|Si|/2
Infection process 2: Adversary picks I0, I1,… , |It| n/4.
I0=R0, R1, R2, …Rt,… are defined by Rt = St It
Fact: |Rt| n/2 for all t [infection never spreads]
1
0 1
100
Reliable circuits from unreliable components [von Neumann]
X2 X3
f
VV
VV
V
X1
V
Given, a circuit C for f of size s
Every gate fails with prob p < 1/10
Construct C’ for C’(x)=f(x) whp.
Possible? With small s’?
Reliable circuits from unreliable components [von Neumann]
X2 X3
f
VV
V
V
V
X1
V
Given, a circuit C for f of size s
Every gate fails with prob p < 1/10
Construct C’ for C’(x)=f(x) whp.
Possible? With small s’?
- Add Identity gatesII I
I
I
I
1
Reliable circuits from unreliable components [von Neumann]
f
Given, a circuit C for f of size s
Every gate fails with prob p < 1/10
Construct C’ for C’(x)=f(x) whp.
Possible? With small s’?
-Add Identity gates
-Replicate circuit
-Reduce errors
X2 X3
VV
V
V
V
X1
V
II I
I
I
I
X2 X3
VV
V
V
V
X1
V
II I
I
I
I
X2 X3
VV
V
V
V
X1
V
II I
I
I
I
X2 X3
VV
V
V
V
X1
V
II I
I
I
I
Reliable circuits from unreliable components[von Neumann, Dobrushin-Ortyukov, Pippenger]
Given, a circuit C for f of size s
Every gate fails with prob p < 1/10
Construct C’ for C’(x)=f(x) whp.
Possible? With small s’?
Majority “expanders”
of size O(log s)
Analysis:
Infection
Process 2
1
f
X2 X3
VV
V
V
V
X1
V
M
X2 X3
VV
V
V
V
X1
V
M
X2 X3
VV
V
V
V
X1
V
M
X2 X3
VV
V
V
V
X1
V
M
MMMM MMMM
MMMM MMMMMMMM
Bipartite expanders
(unbalanced) Bipartite expanders
G(V,E)
V’
V’’
u
w
u’ w’
u” w”
V’’’
H(V’,V’’’;E*) |V’|=n, |V’’’|=(2/3)n, degree=4dSV’ |S| n/2, |NV’’’ (S)| |S| S has a perfect matching to V’’’
SV’ |S| n/2, |NV’’ (S)| (3/2)|S|
Concentrators Cn [Bassalygo,Pinsker]
Cn
SV’ |S| n/2, S has a perfect matching to V’’
V’
V’’
n
2n/3
Networks
- Fault-tolerance- Routing- Distributed computing- Sorting
Superconcentrators [Valiant]
|I|=|O|=n
k n SCn I’ I O’ O|I’|=|O’|=kThere are k vertex disjoint paths I’ to O’
I
O
How many edges are needed for SCn ?
[V] This number is a Circuit lower bound forDisc. Fourier Transform
Superconcentrators & circuits [Valiant]
[V] |E(SCn)| is a circuit lower bound forlinear circuits computingany matrix A with allminors nonsingular y = Ax
k n I’ I O’ O|I’|=|O’|=k there are k Vertex disjoint paths I’ to O’
Otherwise, by Menger’sTheorem, (k-1)-cut, soRank(AI’,O’) <k
I
O
X1 X2 X3 Xn
Y1 Y2 Y3 Yn
Superconcentrators [Valiant]
Theorem: [Valiant] Linear size SCn
Proof: [Pippenger]
•|I’|=|O’| n/2 recursion(2) |I’|=|O’|> n/2Matching & pigeonholeReduce to case (1)
|E(SCn)|== |E(SC2n/3)|+2|E(Cn)|+n= |E(SC2n/3)|+O(n) = O(n)
I
O
SCn
SC2n/3
Cn
Cn
Mn
Distributed routing [Sh,PY,Up,ALM,AC…]
Gn inputs, n outputs, many disjoint paths- Permutation networks- Non-blocking network- Wide-sense non-blocking networks- on-line routable networks- ….Building block G - some expander
Theorem: [Alon-Capalbo]Let G be a sufficiently strong expander.Given (s1,t1),(s2,t2),…(sk,tk), k < n/(log n),one can efficiently find (si,ti) edge-disjointpaths between them, on-line!
Sorting networks [Ajtai-Komlos-Szemeredi]
n inputs (real numbers), n outputs (sorted)
Many sorting algorithms of O(n log n) comparisonsMany sorting networks of O(n log2 n) comparators
Thm: [AKS] Explicit network with O(n log n) comparatorsProof: Extremely sophisticated use & analysis of expanders
Cor: Monotone Boolean formula for majority(derandomizing a probabilistic existence proof of Valiant)
Derandomization
Deterministic error reduction
Algx
r
{0,1}n
random
strings
Thm [Chernoff] r1 r2…. rk independent (kn random bits) Thm [AKS] r1 r2…. rk random path (n+ O(k) random bits)
Algx
rk
Algx
r1
Majority
G [2n,d, 1/8]-graphG explicit! Bx
Pr[error] < 1/3
then Pr[error] = Pr[|{r1 r2…. rk }Bx}| > k/2] < exp(-k)
|Bx|<2n/3
Metric embeddings
Metric embeddings (into l2)
Def: A metric space (X,d) embeds with distortion
into l2 if f : X l2 such that for all x,y
d(x,y) f(x)-f(y) d(x,y)
Theorem: [Bourgain] Every n-point metric space has a O(log n) embedding into l2
Theorem: [Linial-London-Rabinovich] This is tight! Let (X,d) be the distance metric of an [n,d]-expander G.
Proof: f,(AG-J/n)f (G) f2 ( 2ab = a2+b2-(a-b)2 )
(1-(G))Ex,y [(f(x)-f(y))2] Ex~y [(f(x)-f(y))2] (Poincare
inequality)
(clog n)2
2
All pairs Neighbors
Metric embeddings (into l2)
Def: A metric space (X,d) has a coarse embedding into l2 if f : X l2 and increasing, unbounded functions ,:RR such that for all x,y
(d(x,y)) f(x)-f(y) 2 (d(x,y))
Theorem: [Gromov] There exists a finitely generated, finitely presented group, whose Cayley graph metric has no coarse embedding into l2
Proof: Uses an infinite sequence of Cayley expanders…
Comment: Relevant to the Novikov & Baum-Connes conjectures
Nonlinear spectral gaps and metric embeddings into convex
spacesPoincare inequality: G any expander. f : V l2
Ex,y [f(x)-f(y)2] Ex~y [f(x)-f(y) 2] = 1/(1-(G))
Theorem: [Matousek] G any expander. p f : V lp
Ex,y [f(x)-f(y) ] pEx~y [f(x)-f(y) ]
Theorem: [Lafforgue, Mendel-Naor] Construct explicit
G (super-expander) K K f : V lp
Ex,y [f(x)-f(y) ] K Ex~y [f(x)-f(y) ]
Theorem: [Kasparov-Yu, Gromov] Such family of constant degree expanders give rise to a metric space X with no coarse embedding into any uniformly convex space.
p p
p P
2 K
2 K
Constructions
Expansion of Finite GroupsG finite group, SG, symmetric. The Cayley graphCay(G;S) has xsx for all xG, sS.
Cay(Cn ; {-1,1}) Cay(F2n ; {e1,e2,
…,en})
(G) 1-1/n2 (G) 1-1/nBasic Q: for which G,S is Cay(G;S) expanding ?
Algebraic explicit constructions [Margulis,Gaber-Galil,Alon-Milman,Lubotzky-Philips-Sarnak,…Nikolov,Kassabov,..]
Theorem. [LPS] Cay(A,S) is an expander family.
Proof: “The mother group approach”:
Appeals to a property of SL2(Z) [Selberg’s 3/16 thm]
Strongly explicit: Say that we need n bits to describe a matrix M in SL2(p) . |V|=exp(n)
Computing the 4 neighbors of M requires poly(n) time!
A = SL2(p) : group 2 x 2 matrices of det 1 over Zp.
S = { M1 , M2 } : M1 = ( ) , M2 = ( ) 1 10 1
1 01 1
Algebraic Constructions (cont.)
Very explicit -- computing neighbourhoods in logspace
Gives optimal results Gn family of [n,d]-graphs-- Theorem. [AB] d(Gn) 2 (d-1)
--Theorem. [LPS,M] Explicit d(Gn) 2 (d-1) (Ramanujan graphs)Recent results:-- Theorem [KLN] All* finite simple groups expand.-- Theorem [H,BG] SL2(p) expands with most generators.-- Theorem [BGT] same for all Chevalley groups
Zigzag graph product
Combinatorial construction of expanders
Explicit Constructions (Combinatorial)-Zigzag Product [Reingold-Vadhan-W]
G an [n, m, ]-graph. H an [m, d, ]-graph.
Combinatorial construction of expanders.
H
v-cloud
Edges
Step in cloud
Step between clouds
Step In cloud
v uu-cloud
(v,k)
Thm. [RVW] G z H is an [nm,d2,+]-graph,
Definition. G z H has vertices {(v,k) : vG, kH}.
G z H is an expander iff G and H are.
Proof of the zigzag theorem
Proof: Information theoretic view of expanders.
When is G an expander? Random walk is entropy boost!
p, p’ distributions on V before and after a random step.
“Def”:G is an expander iff whenever Ent(v) << log n,
Ent(v’) > Ent(v) + (>0)
Ent0(p) = log |supp(p)|
Ent1(p) = Shannon’s ent.
Ent2(p) = log p2
p
v
v’
p’
Proof of the zigzag thm [Reingold-Vadhan-W]
H
v-cloud v u
u-cloud
Thm. [RVW] G, H expanders, then so is G z H
(u,c)
(u,d)
Want: Ent(u,d) > Ent(v,a) + (assuming Ent(v,a) << log nm)
Case 1: Ent(a|v) << log m Ent(b|v) > Ent(a|v) +
Ent(u,d) Ent(v,b) Ent(v,a) +
Case 2: Ent(a|v)= log m Ent(b|v)= log m Ent(v) << log n
Ent(u) Ent(v) + Ent(c|u) << log m
Ent(u,d) Ent(u,c) +
(v,b)
(v,a)
Mutuallyexclusive?
LinearAlgebra!
Iterative Construction of Expanders
G an [n,m,]-graph. H an [m,d,] -graph.
The construction:
Start with a constant size H a [d4,d,1/4]-graph.
• G1 = H 2
Theorem. [RVW] Gk is a [d4k, d2, ½]-graph.
Proof: Gk2 is a [d 4k,d 4, ¼]-graph.
H is a [d 4, d, ¼]-graph.
Gk+1 is a [d 4(k+1), d 2, ½]-graph.
Theorem. [RVW] G z H is an [nm,d2,+]-graph.
• Gk+1 = Gk2 z H
Weakly explicit construction.Strongly: (Gk Gk)2 z H
Consequences of the zigzag product
- Isoperimetric inequalities beating e-value bounds
[Reingold-Vadhan-W, Capalbo-Reingold-Vadhan-W]
- Connection with semi-direct product in groups
[Alon-Lubotzky-W]
- New expanding Cayley graphs for non-simple groups
[Meshulam-W] : Iterated group algebras
[Rozenman-Shalev-W] : Iterated wreath products
- SL=L : Escaping every maze deterministically [Reingold ’05]
- Super-expanders [Mendel-Naor]
- Monotone expanders [Dvir-W]
SL=L: escaping mazes and navigating
unknown terrains
Getting out of mazes / Navigating unknown terrains (without map & memory)
Theseus
Ariadne
Thm [Aleliunas-Karp-Lipton-Lovasz-Rackoff ’80] A random walk will visit every vertex in n2 steps (with probability >99% )
Only a local view (logspace)n–vertex maze/graph
Thm [Reingold ‘05] : SL=L:A deterministic walk, computable in Logspace, will visit every vertex. Uses ZigZag expanders
Crete1000BC
Mars2006
Expander from any connected graph [Reingold]
Analogy with the iterative construction
G an [n,m,]-graph.
H an [m,d,] -graph.
The construction:
Fix a constant size H a [d4,d,1/4]-graph.• G1 = H 2
Theorem. [RVW] Gk is a [d4k, d2, ½]-graph Proof: Gk
2 is a [d 4k,d 4, ¼]-graph.
H is a [d 4, d, ¼]-graph.
Gk+1 is a [d 4(k+1), d 2, ½]-graph.
Theorem. G z H is an [nm,d2,+]-graph.
• Gk+1 = Gk2 z H
G an [n,m, 1-]-graph.
H an [m,d,1/4] -graph. [nm,d2, 1-/2]-
graph.
H a [d10,d,1/4]-graph.• G1 = G
• Gk+1 = Gk5 z H
Gclog n is [nO(1), d2, ½]
Theorem [R] G1 is [n, d2, 1-1/n3]
Undirected connectivity in Logspace [R] Algorithm
- Input G=G1 an [n,d2]-graph
- Compute Gclog n
-Try all paths of length clog n from vertex 1.
Correctness
- Gi+1 is connected iff Gi is
- If G is connected than it is an [n,d2, 1-1/n3]-graph
- G1 connected Gclog n has diameter < clog n
Space bound
- Gi Gi+1 in constant space (squaring & zigzag are local)
- Gclog n from G1 requires O(log n) space
Zigzag graph product & Semi-direct group
product
Expansion in non-simple groups
Semi-direct Product of groups
A, B groups. B acts on A as automorphisms.
Let ab denote the action of b on a.
Definition. A B has elements {(a,b) : aA, bB}.
group mult (a’,b’ ) (a,b) = (a’ab , b’b)Connection: semi-direct product is a special case of zigzag
Assume <T> = B, <S> = A , S = sB (S is a single B-orbit)
Proof: By inspection (a,b)(1,t) = (a,bt) (Step in a cloud) (a,b)(s,1) = (asb,b) (Step between clouds)
Theorem [ALW] Cay(A B, TsT ) = Cay(A,S ) z Cay(B,T )
Theorem [ALW] Expansion is not a group property
Theorem [MW,RSW] Iterative construction of Cayley expanders
Example: =
A=F2m, the vector space,
S={e1, e2,…, em} , the unit vectors
B=Zm, the cyclic group,
T={1}, shift by 1
B acts on A by shifting coordinates.
S=e1B.
G =Cay(A,S), H = Cay(B,T), and
G z H = Cay(A B, {1 }e1 {1 } )
z
Is expansion a group property?
[Lubotzky-Weiss’93] Is there a group G, and two generating subsets |S1|,|S2|=O(1) such that Cay(G;S1) expands but Cay(G;S2) doesn’t ?
(call such G schizophrenic)
nonEx1: Cn - no S expands
nonEx2: SL2(p)-every S expands[Breuillard-Gamburd’09]
[Alon-Lubotzky-W’01] SL2(p)(F2)p+1
schizophrenic
[Kassabov’05] Symn schizophrenic
Expansion in Near-Abelian Groups
G group. [G;G] commutator subgroup of G
[G;G] = <{ xyx-1y-1 : x,y G }>
G= G0 > G1> … > Gk = Gk+1 Gi+1=[Gi;Gi]
G is k-step solvable if Gk=1.
Abelian groups are 1-step solvable
[Lubotzky-Weiss’93] If G is k-step solvable,
Cay(G;S) expanding, then |S| ≥ O(log(k)|G|)
[Meshulam-W’04] There exists k-step solvable Gk,
|Sk| ≤ O(log(k/2)|Gk|), and Cay(Gk;Sk) expanding.
loglog….log k times
Near-constant degree expanders for
near Abelian groups [Meshulam-W’04]
Iterate: G’= G FqG
Start with G1 = Z2
Get G1 , G2,…, Gk ,… |Gk+1|>exp (|Gk|)
S1 , S2,…, Sk ,… <Sk > = Gk |Sk+1|<poly (|Sk|)
- Cay(Gk, Sk) expanding
- |Sk| O(log(k/2)|Gk|) deg “approaching” constant
Theorem/Conjecture: Cay(G,S) expands, then G has at most exp(d) irreducible reps of dimension d.
Expanders from iterated wreath products [Rozenman-Shalev-W’04]
d fixed. Gk = Aut*(Tdk)
Odd automorphisms of a depth-k
uniform d-ary tree
Iterative: Gk+1 = Gk Ad
Thm: Gk expands with explicit O(1) generators
Ingredients: zigzag, equations in perfect groups[Nikolov], correlated random walks expand,…
0
1 32
d=3, k=2i A3
Beating eigenvalue expansion
Beating e-value expansion [WZ, RVW]
In the following a is a large constant.
Task: Construct an [n,d]-graph s.t. every two sets of size n/a are connected by an edge. Minimize dRamanujan graphs: d=(a2)
Random graphs: d=O(a log a)
Zig-zag graphs: [RVW] d=O(a(log a)O(1))
Uses zig-zag product on extractors!
Applications Sorting & selection in rounds, Superconcentrators,…
Lossless expanders [Capalbo-Reingold-Vadhan-W]
Task: Construct an [n,d]-graph in which every set S,
|S|<<n/d has > c|S| neighbors. Max c (vertex expansion)Upper bound: cd
Ramanujan graphs: [Kahale] c d/2
Random graphs: c (1-)d Lossless
Zig-zag graphs: [CRVW] c (1-)d Lossless
Use zig-zag product on conductors!
Extends to unbalanced bipartite graphs.
Applications (where the factor of 2 matters):Data structures, Network routing, Error-correcting codes
Error correcting codes
Error Correcting Codes [Shannon, Hamming]
C: {0,1}k {0,1}n C=Im(C)
Rate (C) = k/n Dist (C) = min dH(C(x),C(y))
C good if Rate (C) = (1), Dist (C) = (n)
Theorem: [Shannon ‘48] Good codes exist (prob. method)
Challenge: Find good, explicit, efficient codes.
- Many explicit algebraic constructions: [Hamming, BCH, Reed-Solomon, Reed-muller, Goppa,…]
- Combinatorial constructions [Gallager, Tanner, Luby-Mitzenmacher-Shokrollahi-Spielman, Sipser-Spielman..]
Thm: [Spielman] good, explicit, O(n) encoding & decoding
Inspiration: Superconcentrator construction!
Graph-based Codes [Gallager’60s]C: {0,1}k {0,1}n C=Im(C)
Rate (C) = k/n Dist (C) = min dH(C(x),C(y))
C good if Rate (C) = (1), Dist (C) = (n)
zC iff Pz=0 C is a linear code
LDPC: Low Density Parity Check (G has constant degree)Trivial Rate (C) k/n , Encoding time = O(n2)
G lossless Dist (C) = (n), Decoding time = O(n)
n
n-k
1 1 0 1 0 0 1 1 z
0 0 0 0 0 0 Pz + + + + + +
G
Decoding
Thm [CRVW] Can explicitly construct graphs: k=n/2,
bottom deg = 10, B[n], |B| n/200, |(B)| 9|B|
B = corrupted positions (|B| n/200)
B’ = set of corrupted positions after flip
Decoding algorithm [Sipser-Spielman]: while Pw0 flip all wi with i FLIP = { i : (i) has more 1’s than 0’s }
Claim [SS] : |B’| |B|/2
Proof: |B \ FLIP | |B|/4, |FLIP \ B | |B|/4
n
n-k
1 1 1 0 1 0 1 1 w
0 0 1 0 1 1 Pw + + + + + +
Rapidly mixing Markov chains:
Uniform generation and enumeration problems
Expansion of (exponential-size) graphs which arise naturally (as opposed to specially
designed)
Volumes of convex bodies
Algorithm: [Dyer-Frieze-Kannan ‘91]: Approx counting random sampling Random walk inside K. Rapidly mixing Markov chain. Techniques: Spectral gap isoperimetric inequality
Given (implicitly) a convex body K in Rd (d large!)(e.g. by a set of linear inequalities)Estimate volume (K)Comment: Computing volume(K) exactly is #P-complete
K
Statistical mechanicsExample: the dimer problemCount # of domino configurations?
Given G, count the number of perfect matchings GGlauber Dynamics (MCMC sampling the Gibbs distribution)Construct HG on configurations, with edges representing local changes (e.g. rotate adjacent parallel dominos). Then run a random walk for “sufficiently long” time.
Theorems:[Jerrum-Sinclair] poly(n) time convergence for-near-uniform perfect matching in dense & random graphs-sampling Gibbs dist in ferromagnetic Ising modelTechniques: coupling, conductance,…
v HG
approximately
Generating random group elements
Given: S={g1,g2,…gd} generators of a group G (of size n)Find: a near-uniform element x of G (gG: Pr[x=g] < /n )
Straight-line program (SLP) [Babai-Szemeredi]:x1,x2,…xt where every xi is either:•A generator gj or gj
-1
• xkxm or xm-1 for m,k < i
Theorem:[BS] G,S,g an SLP for g of length t=O(log2 n)Theorem:[Babai] G,S a near-uniform SLP* of t=O(log5 n)Theorem:[Cooperman,Dixon,Green] …. SLP* of t=O(log2 n)Proof: Cube(z1,z2,…zt) = {subwords of zt
-1…z2-1z1
-1 z1z2…zt )
x1,x2,…xr with xk+1 R Cube(x1,x2,…xk), with r=O(log n)
Techniques: local expansion, Arithmetic combinatorics
Extensions of expanders
Dimension ExpandersExpanderPermutations 1, 2, …, k: [n] [n] are an expander if for every subset S[n], |S| < n/2 i[k] s.t. |TiSS| < (1-)|S|
Dimension expander [Barak-Impagliazzo-Shpilka-W’01]Linear operators T1,T2, …,Tk: Fn Fn are (n,F)-dimension expander if subspace VFn with dim(V) < n/2 i[k] s.t. dim(TiVV) < (1-) dim(V)
Fact: k=O(1) random Ti’s suffice for every F,n.
[Lubotzky-Zelmanov’04] Construction for F=CWhat about finite fields?
Monotone Expandersf: [n] [n] partial monotone map: x<y and f(x),f(y) defined, then f(x)<f(y).
f1,f2, …,fk: [n] [n] are a k-monotone expander
if fi partial monotone and the (undirected) graph on [n] with edges (x,fi(x)) for all x,i, is an expander.
[Dvir-Shpilka] k-monotone exp 2k-dimension exp F,d Explicit (log n)-monotone expander[Dvir-W’09] Explicit (log*n)-monotone expander (zig-zag)[Bourgain’09] Explicit O(1)-monotone expander[Dvir-W’09] Existence Explicit reductionOpen: Prove that O(1)-mon exp exist!
Real Monotone Expanders [Bourgain’09]
Explicitly constructsf1,f2, …,fk: [0,1] [0,1] continuous, Lipshitz, monotone maps, such that for every S [0,1] with (S)< ½, there exists i[k] such that (Sf2(S)) < (1-) (S)
Monotone expanders on [n] – by discretization
M=( )SL2(R), xR, let fM(x) = (ax+b)/(cx+d)
Take ’-net of such Mi’s in an ’’-ball around I.
Proof:-Expansion in SU(2) [Bourgain-Gamburd] -Tits alternative [Bruillard] -…
a bc d
Open Problems
Open Problems
Characterize: Cayley expanders
Construct: Lossless expanders
Construct: Rate concentratorsof constant left degree
N3
N2
|X| N|Γ(X)||X|