Definition 2 (Eigenvalue Expansion). We say a d-regular graph is a λ eigenvalue expander if

Similar documents
Sparsification using Regular and Weighted. Graphs

Lecture 9: Expanders Part 2, Extractors

Notes for Lecture 11

Machine Learning for Data Science (CS 4786)

(average number of points per unit length). Note that Equation (9B1) does not depend on the

1.3 Convergence Theorems of Fourier Series. k k k k. N N k 1. With this in mind, we state (without proof) the convergence of Fourier series.

Machine Learning for Data Science (CS 4786)

If a subset E of R contains no open interval, is it of zero measure? For instance, is the set of irrationals in [0, 1] is of measure zero?

Modular orientations of random and quasi-random regular graphs

Math Solutions to homework 6

Lecture 2. The Lovász Local Lemma

# fixed points of g. Tree to string. Repeatedly select the leaf with the smallest label, write down the label of its neighbour and remove the leaf.

Chapter 2 Transformations and Expectations

Algorithms in The Real World Fall 2002 Homework Assignment 2 Solutions

Math 778S Spectral Graph Theory Handout #3: Eigenvalues of Adjacency Matrix

1 Review and Overview

Week 5-6: The Binomial Coefficients

Chapter 3. Strong convergence. 3.1 Definition of almost sure convergence

Fall 2013 MTH431/531 Real analysis Section Notes

Math 508 Exam 2 Jerry L. Kazdan December 9, :00 10:20

A Proof of Birkhoff s Ergodic Theorem

Sequences and Series of Functions

Introduction to Probability. Ariel Yadin

Distribution of Random Samples & Limit theorems

Modern Discrete Probability Spectral Techniques

Chapter 6 Infinite Series

AN INTRODUCTION TO SPECTRAL GRAPH THEORY

ECE 6980 An Algorithmic and Information-Theoretic Toolbox for Massive Data

The structure of Fourier series

ACO Comprehensive Exam 9 October 2007 Student code A. 1. Graph Theory

Math 61CM - Solutions to homework 3

Lecture Notes for Analysis Class

A class of spectral bounds for Max k-cut

Ma 4121: Introduction to Lebesgue Integration Solutions to Homework Assignment 5

Combinatorics and Newton s theorem

f n (x) f m (x) < ɛ/3 for all x A. By continuity of f n and f m we can find δ > 0 such that d(x, x 0 ) < δ implies that

Machine Learning Brett Bernstein

Solutions to Math 347 Practice Problems for the final

The Choice Number of Random Bipartite Graphs

36-755, Fall 2017 Homework 5 Solution Due Wed Nov 15 by 5:00pm in Jisu s mailbox

Convergence of random variables. (telegram style notes) P.J.C. Spreij

Infinite Sequences and Series

Lecture 6: Integration and the Mean Value Theorem

Lesson 10: Limits and Continuity

Lecture #3. Math tools covered today

Frequentist Inference

(A sequence also can be thought of as the list of function values attained for a function f :ℵ X, where f (n) = x n for n 1.) x 1 x N +k x N +4 x 3

1+x 1 + α+x. x = 2(α x2 ) 1+x

5 Birkhoff s Ergodic Theorem

MATH301 Real Analysis (2008 Fall) Tutorial Note #7. k=1 f k (x) converges pointwise to S(x) on E if and

The Random Walk For Dummies

Disjoint Systems. Abstract

Infinite Products Associated with Counting Blocks in Binary Strings

Math 216A Notes, Week 5

It is often useful to approximate complicated functions using simpler ones. We consider the task of approximating a function by a polynomial.

Lecture 19: Convergence

Lovász Local Lemma a new tool to asymptotic enumeration?

Theorem 3. A subset S of a topological space X is compact if and only if every open cover of S by open sets in X has a finite subcover.

Brief Review of Functions of Several Variables

Homework 3. = k 1. Let S be a set of n elements, and let a, b, c be distinct elements of S. The number of k-subsets of S is

Application to Random Graphs

SPECTRAL THEOREM AND APPLICATIONS

Commutativity in Permutation Groups

PAPER : IIT-JAM 2010

The Brunn-Minkowski Theorem and Influences of Boolean Variables

On Random Line Segments in the Unit Square

Homework 9. (n + 1)! = 1 1

Lecture 23 Rearrangement Inequality

Riesz-Fischer Sequences and Lower Frame Bounds

MASSACHUSETTS INSTITUTE OF TECHNOLOGY 6.436J/15.085J Fall 2008 Lecture 19 11/17/2008 LAWS OF LARGE NUMBERS II THE STRONG LAW OF LARGE NUMBERS

Introduction to expander graphs

ANSWERS TO MIDTERM EXAM # 2

Definitions and Theorems. where x are the decision variables. c, b, and a are constant coefficients.

The Boolean Ring of Intervals

Proposition 2.1. There are an infinite number of primes of the form p = 4n 1. Proof. Suppose there are only a finite number of such primes, say

(VII.A) Review of Orthogonality

Fundamental Theorem of Algebra. Yvonne Lai March 2010

Lecture 19. sup y 1,..., yn B d n

Lecture 6 Testing Nonlinear Restrictions 1. The previous lectures prepare us for the tests of nonlinear restrictions of the form:

Math 220B Final Exam Solutions March 18, 2002

Lecture 12: November 13, 2018

Seunghee Ye Ma 8: Week 5 Oct 28

A Hadamard-type lower bound for symmetric diagonally dominant positive matrices

Lecture 1:Limits, Sequences and Series

McGill University Math 354: Honors Analysis 3 Fall 2012 Solutions to selected problems

Inhomogeneous Poisson process

Counting trees in graphs

A generalization of the Leibniz rule for derivatives

Lecture 7: Properties of Random Samples

4 The Sperner property.

1 Adiabatic and diabatic representations

Math 155 (Lecture 3)

Analytic Number Theory Solutions

CS 336. of n 1 objects with order unimportant but repetition allowed.

Optimally Sparse SVMs

Math 104: Homework 2 solutions

Linear Regression Demystified

University of Colorado Denver Dept. Math. & Stat. Sciences Applied Analysis Preliminary Exam 13 January 2012, 10:00 am 2:00 pm. Good luck!

HOMEWORK 2 SOLUTIONS

lim za n n = z lim a n n.

Transcription:

Expaer Graphs Graph Theory (Fall 011) Rutgers Uiversity Swastik Kopparty Throughout these otes G is a -regular graph 1 The Spectrum Let A G be the ajacecy matrix of G Let λ 1 λ λ be the eigevalues of A G Sometimes we will also be itereste i the Laplacia matrix of G This is efie to be L G = D A G, where D is the iagoal matrix where D vv equals the egree of the vertex v For -regular graphs, L G = I A G, a hece the eigevalues of L G are λ 1, λ,, λ Lemma 1 λ 1 = λ = λ 3 = = λ k = if a oly if G has at least k coecte compoets Proof For the first part, verify that the vector 1 is a eigevector with eigevalue For the seco part, let C 1,, C r be all the coecte compoets of G The 1 C1,, 1 Cr are all mutually orthogoal eigevectors with eigevalue if G has at least k coecte compoets, the λ = = λ r = Suppose there is a eigevector f ot i the spa of 1 C1,, 1 Cr with eigevalue Let C i be a compoet o which that eigevector is ot costat Let v C i be a vertex such that f(v) is maximum Sice f has eigevalue, f(v) = u Γ(v) f(u), a by maximality this implies that f(u) = f(v) for each u Γ(v) Repeatig this, we coclue that f must be costat o the compoet C i, a cotraictio I particular, coecte graphs have λ < We ow stuy the otio of expasio, where λ is less tha by a sigificat amout; this ca be thought of as a strog form of coecteess Expasio Defiitio (Eigevalue Expasio) We say a -regular graph is a λ eigevalue expaer if λ λ We say a -regular graph is a λ absolute eigevalue expaer if λ, λ λ If we have a family of graphs with teig to a with costat, we iformally call these graphs expaer graphs if they are all λ eigevalue expaers for some costat λ < Similarly we call these graphs absolute expaer graphs if they are all λ absolute eigevalue expaers for some costat λ < 1

Havig mae this efiitio, we ow show that families of expaer graphs (a absolute expaer graphs) exist Theorem 3 For every 3, there exists λ < such that for all sufficietly large, there exists a -vertex -regular λ absolute eigevalue expaer graph The proof is eferre to the e of these otes We first see why this otio is useful 3 Properties of Expaers May combiatorial properties of a graph ca be expresse i terms of the eigevalues of the ajacecy matrix I the case of eigevalue expaers, this coectio becomes very clea a powerful Let G = (V G, E G ) be a -regular -vertex graph with eigevalues = λ 1 λ λ Let v 1 = 1 1, v,, v be a orthoormal basis cosistig of eigevectors for λ 1,, λ respectively Let v : V G R The v ca be writte as a 1 v 1 + + a v, where a i = v, v i We have v = i=1 a i Note that v, L G v = x V G v(x) v, A G v = x V G y Γ(x) x V G y Γ(x) v(x)v(y), v(x)v(y) = 1 x V G y Γ(x) (v(x) v(y)) We ca also express these quatities i terms of the eigevalues We have A G v = a i λ i v i a L G v = a i ( λ i )v i Therefore v, A G v = a i λ i a v, L G v = a i ( λ i) Utilizig such expressios, we ca erive a large umber of combiatorial properties of expaer graphs Theorem 4 (Ege Expasio) Suppose G is a λ eigevalue expaer The for every S V G with S / we have e(s, S c ) λ S Proof We first wat to express the quatity e(s, S c ) i terms of the ajacecy matrix Let 1 S : V G R be the iicator fuctio of the set S The e(s, S c ) = 1 S, A G (1 1 S ) = S 1 S, A G 1 S Let S = i=1 a iv i We have a 1 = 1 S, 1 1 = S Further, we have a i = 1 S = S

So 1 S, A G 1 S = a i v i, a i λ i v i = a i λ i = S + a i λ i S S + λ( S ) = S ( λ) + λ S e(s, S c ) = ( λ) ) ( S S λ S Theorem 5 (Expaer Mixig Lemma) Suppose G is a λ absolute eigevalue expaer The for every S, T V G, we have e(s, T ) S T λ S T Proof Let 1 S = a k v i a 1 T = b i v i We have a 1 = S a a i = S, a similarly b 1 = T a b i = T The quatity of iterest e(s, T ) ca be expresse spectrally as follows: e(s, T ) = 1 S, A G 1 T = i a i v i, j b j λ j v j = i,j a i b j λ j v i, v j = i a i b i λ i = a 1 b 1 + = S T a i b i λ i + a i b i λ i The absolute value of seco term i this expressio ca be boue (usig the Cauchy-Schwarz iequality) by: ( ) 1/ ( ) 1/ λ b i λ S T This completes the proof a i 3

Theorem 6 (Vertex Expasio) Suppose G is a λ eigevalue expaer Let S V G The the eighborhoo of S is large: 1 Γ(S) S ( ) λ + 1 λ S I this theorem, if λ < (1 Ω(1)), a S < (1 Ω(1)), the Γ(S) = S (1 + Ω(1)), which is calle vertex expasio Proof Let 1 S = i a iv i The a 1 = S a a i = S Γ(S) is ot a quatity that ca be capture precisely i terms of the spectrum of G (ulike e(s, S c )) However it ca be lower boue usig the followig observatio: If f is a fuctio, the supp(f) f 1 f We apply this to f = A 1 S The supp(f) clearly equals Γ(S) We ow compute f 1 a f f 1 equals S f = A1 S = i a i λ i v i ( ( S + λ S 1 S )) 1/ Γ(S) S S + λ S S S = S + λ ( S ) 1 = S ( ) λ + 1 λ S This completes the proof Theorem 7 (Rapi Mixig of Raom Walks) Suppose G is a λ absolute eigevalue expaer Let v 0 V G be ay vertex Let µ be the probability istributio of the k-th step of the raom walk o G startig at v 0 The the statistical istace of µ from uiform is boue as follows: µ U 1 ( ) λ k, where U is the uiform istributio o V G Proof Let P be the matrix 1 A G If π R V G is a probability istributio o the vertices of G, the P π is the probability istributio of the vertex v, obtaie by pickig a vertex u accorig to π a the lettig v be a raom eighbor of that vertex Let π 0 be the probability istributio supporte o v 0 The probability istributio µ of the k-th step of the raom walk equals P k π 0 4

Let π 0 = i a iv i Note that a 1 = 1, a i a i = x V G π 0 (x) = 1 The µ = P k π 0 = i ( ) k λi a i v i = U + ( ) k λi a i v i Therefore µ U = ( a i ( ) λ k µ U 1 ( ) ) k 1/ λi ( ) λ k The ext theorem says that absolute eigevalue expaers have o large iepeet sets a hece they have large chromatic umber Theorem 8 (Iepeet Sets/Chromatic Number) The largest iepeet set i G has size at λ most λ the chromatic umber of G is at least λ λ You will prove this i your homework 4 Limits o eigevalue expasio How goo a eigevalue expaer ca a -regular graph be? This questio is aswere by a theorem of Alo a Boppaa Theorem 9 (Alo-Boppaa, simpler versio) Let G be a -regular graph which is a λ absolute eigevalue expaer The λ 1 o(1) (where the o(1) term tes to 0 as the umber of vertices tes to ifiity) 1 is ot just ay ol umber, it has very eep origis The ifiite ajacecy matrix of the ifiite -regular tree (this is the ultimate expaer graph) has its spectral raius equal to 1 Proof The mai iea is to stuy the eigevalues through the trace of a high eve power of A G We have: Tr(A k G ) = i=1 λ k i k + ( 1)λ k O the other ha, Tr(A k G ) couts the umber of close walks of legth k i G It is easy to see that the umber of such walks is at least as large as times the umber of such walks i the ifiite -regular tree The latter ca be coute by elemets σ of [] ([ 1] β) k 1 5

with exactly k β s, havig the property that ay prefix of σ of legth t has at most t/ β s This ca be coute easily i terms of Catala umbers; we get that the total umber of close walks i G of legth k is at least: ( 1 k + 1 Takig k = ω(1) a simplifyig, we get: ( ) k ) ( 1) k 1 k ( ( ) 1 k k + ( 1)λ k ) ( 1) k 1 k + 1 k λ 1 o(1) The full theorem of Alo-Boppaa shows that if G is a λ eigevalue expaer (ot ecessarily a absolute eigevalue expaer), eve the λ 1 o(1) This is a more elicate fact a the proof is more serious 5 Combiatorial versios of eigevalue expasio Defie the ege expasio of the graph G, eote h(g), by: h(g) = mi S V G, S / e(s, S c ) S h(g) beig large meas that most eges iciet o vertices of S are betwee S a S c By Theorem 4, we have that if G is a λ eigevalue expaer, the h(g) λ I particular, if λ = Ω(1), the h(g) Ω(1) We ow see the coverse of this statemet: if h(g) Ω(1), the G is a λ eigevalue expaer for λ Ω(1) Theorem 10 (Cheeger-Alo-Milma) Suppose λ R satisfies The G is a λ eigevalue expaer h(g) ( λ) Proof Let f : V G R be a uit orm eigevector of the eigevalue λ {x,y} E G (f(x) f(y)) = f, Lf = λ We will use f to fi a set S V G, with S /, such that e(s, S c ) ( λ ) S (this will show that h(g) ( λ ), as esire) First write f = f + f, where f + a f are oegative value Let V + a V be their supports Note that for each x V +, Lf + (x) Lf(x) = ( λ )f(x) = λ f + (x), 6

a for each x V, Lf (x) Lf(x) = ( λ )f(x) = ( λ )f (x) f +, Lf + ( λ ) f + a f, Lf ( λ ) f Now at least oe of the fuctios f + a f has support of size at most / Let us call that fuctio g We just showe that g, Lg ( λ ) g We will ow fi the esire S as a subset of the support of g It will be through a elicate probabilistic argumet Pick a [0, max x VG g(x)] uiformly at raom Let T = {x V G g(x) a} (ote that T is a raom set, with 0 < T /) The for a ege {x, y} E G, the probability that it gets coute i e(t, T c ) equals the probability that a lies i betwee g(x) a g(y), which equals g(x) g(y) : E[e(T, T c )] = g(x) g(y) {x,y} E G = (g(x) g(y)) (g(x) + g(y)) {x,y} E G 1/ (g(x) g(y)) (g(x) + g(y)) {x,y} E G {x,y} E G g, Lg 1/ ( g ( λ ) g ) 1/ O the other ha, E[ T ] = x V G g (x) = g E[e(T, T c )] E[ T ] ( λ ) This implies (verify this!) that with positive probability, e(t,t c ) T ( λ ), as esire Absolute eigevalue expasio ca also be characterize i terms of combiatorial properties of a graph A fuametal theorem of Bilu a Liial roughly says that a graph is a absolute eigevalue expaer if a oly if it satisfies the coclusio of the expaer mixig lemma 1/ 7