Expander Graph Learning Notes

In graph theory, an expander graph is a sparse graph that has strong connectivity properties, quantified using vertex, edge or spectral expansion. Expander constructions have spawned research in pure and applied mathematics, with several applications to complexity theory, design of robust computer networks, and the theory of error-correcting codes.

Introduction to Expander Graphs

Measures of Expansion

First, there are two combinatorial definitions of expansion.

Definition (Edge Expansion): A DD-regular graph GG is a (K,ϵ)(K,\epsilon) edge expander if for all sets SS of at most KK vertices, the cut size e(S,Sˉ)e(S,\bar{S}) is at least ϵSD\epsilon |S|D.

In addition, we define h(G)h(G) by:

h(G):=minSV and SV/2e(S,Sˉ)dSh(G):=\min_{S\subset V\text{ and }|S|\le |V|/2} \dfrac{e(S,\bar{S})}{d|S|}

Definition (Vertex Expansion): A graph GG is a (K,A)(K,A) vertex expander if for all sets SS of at most KK vertices, the neighborhood N(S):={uvS s.t. (u,v)E}N(S) := \{u\mid ∃v ∈ S\text{ s.t. }(u,v) ∈ E\} is of size at least ASA · |S|.

And then, there is another definition of expansion.

Definition (Spectral Expansion): For γ[0,1]\gamma \in [0,1], we say that a regular graph GG has spectral expansion γ\gamma if γ(G):=1λ(G)γ\gamma(G):=1-\lambda(G)\ge \gamma, where λ(G)\lambda(G) is the second eigenvalue of the normalized adjacency matrix M(G)M(G).

Note that by Rayleigh theorem, we have:

λ(G)=maxπ0πM1π1=maxx1xMx\lambda(G)=\max_{\pi\ne \bm 0}\dfrac{\|\pi M-\bm{1}\|}{\|\pi -\bm{1}\|}=\max_{x \bot \bm{1}}\dfrac{\|x M\|}{\|x\|}

Some Relations between Different Definitions

Expander Mixing Lemma

We will prove a famous result known as the expander mixing lemma, which gives a relation between a graph’s spectral expansion and a certain regularity or spreadness condition of its edges that is a slightly more general statement than edge-expansion itself.

Theorem (Expander Mixing Lemma): Let GG be a DD-regular, NN vertex graph with spectral expansion 1λ1-\lambda. Then for all sets of vertices S,TS,T, we have:

e(S,T)NDμ(S)μ(T)λμ(S)(1μ(S))μ(T)(1μ(T))\left | \dfrac{e(S,T)}{ND}-\mu(S)\mu(T) \right |\le \lambda\sqrt{\mu(S)(1-\mu(S))\mu(T)(1-\mu(T))}

λμ(S)μ(T)λ\le \lambda\sqrt{\mu(S)\mu(T)}\le \lambda

where μ(H)(HV(G))\mu(H)\,(H\subset V(G)) denote the denstity of HH, H/N|H|/N.

Proof: Let 1S,1T\bm{1}_S,\bm{1}_T be the characteristic row vector of SS and TT, i.e. (1S)i=[iS](\bm{1}_S)_i=[i\in S].

By using orthogonal decompostion, we can write 1S\bm{1}_S as: 1S=1S,11,11+1S=1i(1S)i+1S=μ(S)N1+1S\bm{1}_S=\dfrac{\langle \bm{1}_S,\bm{1}\rangle}{\langle \bm{1},\bm{1}\rangle}\bm{1}+\bm{1}_S^\bot=\bm{1}\sum_i(\bm{1}_S)_i+\bm{1}_S^\bot=\mu(S)N\bm{1}+\bm{1}_S^\bot. Similarly, 1T=μ(T)N+1T\bm{1}_T=\mu(T)N+\bm{1}_T^\bot.

Then:

e(S,T)ND=1N(μ(S)N1+1S)M(μ(T)N1+1T)T\dfrac{e(S,T)}{ND}=\dfrac 1N (\mu(S)N\bm{1}+\bm{1}_S^\bot)M(\mu(T)N\bm{1}+\bm{1}_T^\bot)^T

Notice that 1M=1,MuT=1T,11T=1/N\bm{1}M=\bm{1},Mu^T=\bm{1}^T,\bm{1}\bm{1}^T=1/N. Thus:

e(S,T)NDμ(S)μ(T)=1SM(1T)TN1SM1TNλ1S1TN\left| \dfrac{e(S,T)}{ND}-\mu(S)\mu(T) \right|=\left|\dfrac{\bm{1}_S^\bot M(\bm{1}_T^\bot)^T}N\right|\le \dfrac{\|\bm{1}_S^\bot M\|\|\bm{1}_T^\bot\|}N\le \dfrac{\lambda\|\bm{1}_S^\bot\|\|\bm{1}_T^\bot\|}N

Notice that μ(S)N=1S2=μ(S)N12+1S2=μ2(S)N+1S2\mu(S)N=\|\bm{1}_S\|^2=\|\mu(S)N\bm{1}\|^2+\|\bm{1}_S^\bot\|^2=\mu^2(S)N+\|\bm{1}_S^\bot\|^2. Thus we have: 1S=μ(S)(1μ(S))\|\bm{1}_S^\bot\|=\sqrt{\mu(S)(1-\mu(S))}. Similarly, 1T=μ(T)(1μ(T))\|\bm{1}_T^\bot\|=\sqrt{\mu(T)(1-\mu(T))}.

The above then implies:

e(S,T)NDμ(S)μ(T)λμ(S)(1μ(S))μ(T)(1μ(T))\left | \dfrac{e(S,T)}{ND}-\mu(S)\mu(T) \right |\le \lambda\sqrt{\mu(S)(1-\mu(S))\mu(T)(1-\mu(T))}

λμ(S)μ(T)λ\le \lambda\sqrt{\mu(S)\mu(T)}\le \lambda

as desired. \Box

Expander mixing lemma also has its converse. We present here without proving it.

Theorem (Converse to Expander Mixing Lemma): Let GG be a DD-regular, NN-vertex graph. Suppose that for all pairs of disjoint vertex sets S,TS,T, we have e(S,T)/(ND)μ(S)μ(T)θμ(S)μ(T)|e(S,T)/(ND) − \mu(S)\mu(T)| ≤θ\sqrt{\mu(S)\mu(T)} for some θ[0,1]θ ∈ [0,1]. Then λ(G)=O(θlog(1/θ))λ(G) = O(θ \log(1/θ)).

Cheeger Inequaltiy

Theorem (Cheeger Inequality): Let GG be a DD-regular graph, then:

1λ(G)2h(G)2(1λ(G))\dfrac{1 − λ(G)}2≤ h(G) ≤\sqrt{2(1 − λ(G))}

Proof: (鸽子一会)\Box

Random Walks on Expanders

Due to the good struction of expanders, we can use the random walk on expanders to reduce the randomness needed by random algorithms.

Hitting Property of Expander Walks: If GG is a regular graph with spectral expansion 1λ1 − λ, then for any BV(G)B \subset V(G) of density μ\mu, the probability that a random walk (V1,,Vt)(V_1,\ldots,V_t) of t1t − 1 steps in GG starting in a uniformly random vertex V1V_1 always remains in BB is

Pr[i=1tViB](μ+λ(1μ))t\Pr\left[\bigwedge_{i=1}^t V_i\in B \right]\le (\mu+\lambda(1-\mu))^t

Proof: Firstly, we prove a lemma:

Lemma: The probability that the random walk stays entirely within BB is precisely 1P(MP)t11\|\bm 1P(MP)^{t−1}\|_1, where PP is defined as follow:

Pij={[iB]i=j0ijP_{ij}=\begin{cases} [i\in B] & i=j \\ 0 & i\ne j \end{cases}

Proof of Lemma: We can easily prove this lemma by induction on tt. \blacksquare

And there becomes anther lemma:

Lemma: PMPμ+λ(1μ)\|PMP\|\le \mu+\lambda(1-\mu), where \|\cdot \| denotes the spectral norm of matrices.

Proof of Lemma: We decompose MM into two parts:

M=γJ+λEM=\gamma J+\lambda E

Then we have:

PMP=P(γJ+λE)PγPJP+λPEPγPJP+λ\|PMP\|=\|P(\gamma J+\lambda E)P\| \\ \le \gamma \|PJP\|+\lambda \|PEP\|\le \gamma \|PJP\|+\lambda

Noet that the last inequality uses the fact that E1\|E\|\le 1.

For all xRNx\in \mathbb R^N, let yy be xPxP, then xPJP=yJP=(iyi)1PxPJP=yJP=\left(\sum_i y_i \right)\bm 1 P.

Since:

xPJP2iyi1P2μNy2μNμx\|xPJP\|_2\le \left|\sum_i y_i \right|\|\bm 1P\|_2\le \sqrt{\mu N}\|y\|_2\sqrt{\dfrac\mu N}\le \mu \|x\|

Noet that the last inequality uses the fact that y2x2\|y\|_2\le \|x\|_2.

Thus:

PMPγμ+λ=μ+λ(1μ)\|PMP\|\le \gamma \mu+\lambda=\mu+\lambda(1-\mu)

as desired. \blacksquare

By combining two above lemmas, we can get that:

1P(MP)t11=1P(PMP)t11μN1P(PMP)t12μN1P2PMPt1μNμN(μ+λ(1μ))t1(μ+λ(1μ))t\|\bm 1 P(MP)^{t-1}\|_1=\|\bm 1 P(PMP)^{t-1} \|_1\le \sqrt{\mu N}\|\bm 1 P(PMP)^{t-1}\|_2 \\ \le \sqrt{\mu N}\|\bm 1P\|_2 \|PMP\|^{t-1}\le \sqrt{\mu N}\sqrt{\dfrac{\mu}N}(\mu +\lambda(1-\mu))^{t-1}\le (\mu+\lambda(1-\mu))^t

as desired. \Box

To reduce the randomness of a RP\textbf{RP} algorithm AA which uses mm random bits, the general approach is to consider an expander graph with vertex set ${0,1}^m, where each vertex is associated with a setting of the random bits. We will choose a uniformly random vertex V1V_1 and then do a random walk of length t1t − 1, visiting additional vertices V2,,VtV_2,\ldots,V_t. This requires mm random bits for the initial choice, and
logD\log D for each of the t1t − 1 steps. For every vertex ViV_i on the random walk, we will run the algorithm with ViV_i as the setting of the random coins.

Then by the theorem above, in order to reduce the error probability to 2k2^{-k}, the length of the random walk should be Θ(k)\Theta(k). Thus, the number of random bits is m+O(k)m+O(k).

The success of the error-reduction procedure for two-sided error algorithms is obtained by the following theorem.

Chernoff Bound for Expander Walk: Let GG be a regular graph with NN vertices and spectral expansion 1λ1 − λ, and let B[N]B\in [N] be a subset of vertices. Consider a random walk V1,,VtV_1,\ldots,V_t in GG from a uniform start vertex V1V_1. Then for any ϵ>0\epsilon > 0

Pr[1ti=1t[ViB]μ(B)ϵ]2e(1λ)ϵ2t/4\Pr\left[\left|\dfrac 1t \sum_{i=1}^t [V_i\in B] -\mu(B) \right| \ge \epsilon \right]\le 2\text{e}^{-(1-\lambda)\epsilon^2t/4}

Proof:

Corollary: If a language LL has a BPP\textbf{BPP} algorithm with error probability at most 1/31/3 that uses m(n)m(n) random bits on inputs of length nn, then for every polynomial k(n)k(n), LL has a BPP algorithm with error probability at most 2k(n)2^{−k(n)} that uses m(n)+O(k(n))m(n) + O(k(n)) random bits.

Optimal Expanders

How good can the expansion of a graph be?

For vertex expansion, the following theorem shows its limitation.

Theorem (Limits on Vertex Expansion):

(1) Show that if a DD-regular graph GG is a (K,A)(K,A) expander, then TDT_D is a (K,A)(K,A) expander, where TDT_D is the infinite DD-regular tree.

(2) Show that for every DND \in \mathbb N, there are infinitely many KNK \in \mathbb N such that TDT_D is not a (K,D1+2/K)(K,D − 1+2/K) expander.

(3) Deduce that for constant DND \in \mathbb N and α>0α > 0, if a DD-regular, NN-vertex graph GG is an (αN,A)(αN,A) vertex expander, then AD1+o(1)A ≤D − 1 + o(1), where the o(1)o(1) term vanishes as NN \to \infty (and D,αD, α are held constant).

For spectral expansion, the result is known as Alon–Boppana Bound.

Theorem (Alon–Boppana Bound): Let GG be a DD-regular graph, then:

λ(G)2D1Do(1)\lambda(G)\ge \dfrac{2\sqrt{D-1}}D-o(1)

where the o(1)o(1) term vanishes as NN\to \infty.

Proof: For a graph HH, let pl(H)p_l(H) denote the probability that if we choose a random vertex vv in HH and do a random walk of length 2l2l, we end back at vertex vv. And let TDT_D be the infinite DD-regular tree.

Obviously, we have: pl(G)pl(TD)p_l(G)\ge p_l(T_D).

A random walk of length 2l2l which end back at starting vertex can be seen as a properly parenthesized string of length 2l2l with at least D1D-1 types of parentheses.

Thus: pl(TD)Cl(D1)l/D2lp_l(T_D)\ge C_l(D-1)^l/D^{2l}, where ClC_l is the ll-th Catalan number.

In addition, we have: tr(M2l)=1+λ22l++λn2l1+(N1)λ2l(G)\text{tr}(M^{2l})=1+\lambda_2^{2l}+\cdots+\lambda_n^{2l}\le 1+(N-1)\lambda^{2l}(G) and: pl(G)=vGpl,v(G)/N=tr(M2l)/Np_l(G)=\sum_{v\in G}p_{l,v}(G)/N=\text{tr}(M^{2l})/N.

Thus we have: Npl(G)1+(N1)λ2l(G)Np_l(G)\le 1+(N-1)\lambda^{2l}(G).

Notice that Cl=(2ll)/(l+1)C_l=\dbinom {2l}l/(l+1). Then using the inequality proved above, we have:

1+(N1)λ2l(G)Npl(G)NCl(D1)l/D2l=N(2ll)(D1)l(l+1)D2l1+(N-1)\lambda^{2l}(G)\ge Np_l(G)\ge NC_l(D-1)^l/D^{2l}=N\dfrac{\binom{2l}l(D-1)^l}{(l+1)D^{2l}}

Thus:

λ(G)N(2l)!(l+1)(l!)22lD1D\lambda(G)\ge \sqrt[2l]{\dfrac{N(2l)!}{(l+1)(l!)^2}}\cdot \dfrac{\sqrt{D-1}}D

Let l=Nl=N, then use Stirling’s Approximation, we have:

λ(G)D1D2π2N(2N/e)2N(2πN(N/e)N)22NNN+12N\lambda(G)\ge \dfrac{\sqrt{D-1}}D\cdot \sqrt[2N]{\dfrac{\sqrt{2\pi \cdot 2N}(2N/\text{e})^{2N}}{(\sqrt{2\pi N}(N/\text{e})^{N})^2}}\cdot \sqrt[2N]{\dfrac N{N+1}}

=2D1Do(1) (N)=\dfrac{2\sqrt{D-1}}D-o(1)\ (N\to \infty)

as desired. \Box

The optimal spectral expander is called as the Ramanujan graph.

Definition (Ramanujan Graphs): A dd-regular graph GG is called as a Ramnujan graph if λ(G)2D1D+o(1)\lambda(G)\le \dfrac{2\sqrt{D-1}}D+o(1).

We will discuss some explicit constructions of Ramnujan graphs in the next section.

Explicit Constructions

A random graph will be a good expander but we will not choose it as an expander. Some reasons we don’t choose it:

  • Don’t want to tolerate the error probability that the graph isn’t an expander, and checking it is NP-Hard\text{NP-Hard} problem.
  • Defeat the purpose that we want to reduce randomness.
  • Require exponentially large expander graphs, and thus we can’t even write down a randomly chosen expander.

Mildly Explicit: Construct a complete representation of the graph in time an poly(N)\text{poly}(N).

Fully Explicit: Given a mode u[N]u\in [N] and i[D]i\in [D], where DD is the degree of the expander, compute the iith nerghbor of uu in time poly(logN)\text{poly}(\log N). (more efficiently)

Goal: Devise a fully explicit construction of an infinite family {Gi}\{G_i \} of DD-regular graphs with spectral expansion at least γ\gamma, where DD and γ>0\gamma >0 are constants independent of ii.

Algebraic Constructions

Discrete torus expanders, pp-cycle with inverse chords.

Ramanujan Graphs: G=(V,E)G = (V,E) is a graph with vertex set V=Fq{}V = \mathbb F_q \cup \{\infty\}, the finite field of prime order qq s.t. q1(mod4)q \equiv 1 \pmod 4 plus one extra node representing infinity. The edges in this graph connect each node zz with all zz' of the form:

z=(a0+ia1)z+(a2+ia3)(a2+ia3)z+(a0ia1)z' = \dfrac{(a_0 + ia_1)z + (a_2 + ia_3)}{(−a_2 + ia_3)z + (a_0 − ia_1)}

for a0,a1,a2,a3Za_0, a_1, a_2, a_3 ∈ \mathbb Z such that a02+a12+a22+a32=pa^2_0 + a^2_1 + a^2_2 + a^2_3 = p, a0a_0 is odd and positive, and a1,a2,a3a_1, a_2, a_3 are even, for some fixed prime pqp \ne q such that p1(mod4)p \equiv 1 \pmod 4, qq is a square modulo pp, and iFqi \in \mathbb F_q such that i2=1(modq)i^2 = −1 \pmod q.

It’s an optimal spectral expander. (?, omit the proof)

Graph Operations

Definition: An (N,D,γ)(N,D,\gamma)-graph is a DD-regular graph on NN vertices with spectral expansion γ\gamma.

Squaring

Definition (Squaring of Graphs): If G=(V,E)G = (V,E) is a DD-regular graph, then G2=(V,E)G^2 = (V,E') is a D2D^2-regular graph on the same vertex set, where the (i,j)(i, j)th neighbor of a vertex xx is the jjth neighbor of the iith neighbor of xx.

Lemma: If GG is a (N,D,1λ)(N,D,1 − \lambda)-graph, then G2G^2 is a (N,D2,1λ2)(N,D^2,1 − λ^2)-graph.

Proof: Let MM be the random walk matrix for GG, and then M2M^2 is the random walk matrix for G2G^2. Thus: x1,xM2λxMλ2x\forall x\bot \bm{1}, \|xM^2 \|\le \lambda \|xM \|\le \lambda^2 \| x\|. \Box

Tensor

Definition (Tensor Product of Graphs): Let G1=(V1,E1)G_1 = (V_1,E_1) be D1D_1-regular and G2=(V2,E2)G_2 = (V_2,E_2) be D2D_2-regular. Then their tensor product is the D1D2D_1D_2-regular graph G1G2=(V1×V2,E)G_1 \otimes G_2 = (V_1 × V_2,E), where the (i1,i2)(i_1, i_2)th neighbor of a vertex (x1,x2)(x_1,x_2) is (y1,y2)(y_1,y_2), where yby_b is the ibi_bth neighbor of xbx_b in GbG_b.

A Few Comments on the Tensor of Matrix and Vector:

  • (xy)(AB)=(xA)(yB)(x ⊗ y)(A ⊗ B)=(xA) ⊗ (yB).
  • M1M2=(IN1M2)(M1IN2)=(M1IN2)(IN1M2)M_1 ⊗ M_2 = (I_{N_1} ⊗ M_2)(M_1 ⊗ I_{N_2} )=(M_1 ⊗ I_{N_2} )(I_{N_1} ⊗ M_2).

Lemma: If G1G_1 is an (N1,D1,γ1)(N_1,D_1, \gamma_1)-graph and G2G_2 is an (N2,D2,γ2)(N_2,D_2, \gamma_2)- graph, then G1G2G_1 ⊗ G_2 is an (N1N2,D1D2,min{γ1,γ2})(N_1N_2,D_1D_2,\min\{γ_1, γ_2\})-graph.

Proof: Let M1,M2M_1,M_2 be the random walk matrix for G1G_1 and G2G_2, and then MM1M2M\triangleq M_1\otimes M_2 is the random walk matrix for G1G2G_1 \otimes G_2.

x(RN1N2)1N1N2\forall x(\in \mathbb R^{N_1N_2})\bot \bm{1}_{N_1N_2}, decompose xx as x=x+xx = x^{\|} + x^\bot, where xx^{\|} is a multiple of 1N2\bm{1}_{N_2} on each cloud of size N2N_2 and xx^\bot is orthogonal to 1N2\bm{1}_{N_2} on each cloud.

Note that x,1N1N2=Seach cloudxS,1N2=0\langle x^\bot,\bm{1}_{N_1N_2} \rangle=\sum\limits_{S\in\text{each cloud}}\langle x^\bot|_S,\bm{1}_{N_2} \rangle=0. Thus, x,1N1N2=x,1N1N2x,1N1N2=0\langle x^\|,\bm{1}_{N_1N_2} \rangle=\langle x,\bm{1}_{N_1N_2} \rangle-\langle x^\bot,\bm{1}_{N_1N_2} \rangle=0, i.e. x1N1N2x^\| \bot \bm{1}_{N_1N_2}. Therefore, if we write xx^\bot as x=y1N2x^\bot=y\otimes \bm{1}_{N_2}, then y1N1y\bot \bm{1}_{N_1}.

For the first case, we have: xM=(y1N2)(M1M2)=(yM1)1N2x^{\|}M = (y ⊗ \bm{1}_{N_2} )(M_1 ⊗ M_2)=(yM_1) ⊗ \bm{1}_{N_2}. The expansion G1G_1 tells us that M1M_1 shrinks yy by a factor of λ1λ_1, and thus xM=N2yMλ1N2y=λ1x\|x^{\|}M\|=\sqrt{N_2}\|yM\|\le \lambda_1 \sqrt{N_2}\|y\| = λ_1 \|x^{\|}\|.

For another case, we have: xM=x(IN1M2)(M1IN2)x^\bot M = x^\bot (I_{N_1} ⊗ M_2)(M_1 ⊗ I_{N_2} ). The expansion of G2G_2 tells us that M2M_2 will shrink xx^\bot by a factor of λ2λ_2 on each cloud, and thus IN1M2I_{N_1} ⊗ M_2 will shrink xx^\bot by the same factor. The subsequent application of M1IN2M_1 ⊗ I_{N_2} cannot increase the length. Thus, xMλ2x\|x^\bot M\| ≤ λ_2\|x^\bot\|.

Finally, we argue that xMx^\|M and xMx^\bot M are orthogonal. Note that xM=(yM1)1N2x^\|M = (yM_1) ⊗ \bm{1}_{N_2} is a multiple of 1N2\bm{1}_{N_2} on every cloud. Thus it suffices to argue that xx^\bot remains orthogonal to 1N2\bm{1}_{N_2} on every cloud after we apply MM. Applying IN1M2I_{N_1} ⊗ M_2 retains this property and applying M1IN2M_1 ⊗ I_{N_2} retains this property because it assigns each cloud a linear combination of several other clouds.

The above then implies:

xM2=xM2+xM2λ12x2+λ22x2max{λ1,λ2}2(x2+x2)=max{λ1,λ2}2x2\|xM\|^2=\|x^\| M\|^2+\|x^\bot M\|^2\le \lambda_1^2\|x^\|\|^2+\lambda_2^2\|x^\bot\|^2\\ \le \max\{\lambda_1,\lambda_2\}^2(\|x^\|\|^2+\|x^\bot\|^2)=\max\{\lambda_1,\lambda_2\}^2\|x\|^2

as desired. \Box

Zig–Zag Product

Definition (Zig–Zag Product): Let GG be a D1D_1-regular digraph on N1N_1 vertices, and HH a D2D_2-regular digraph on D1D_1 vertices. Then GHG Ⓩ H is a graph whose vertices are pairs (u,i)[N1]×[D1](u, i) \in [N_1] × [D_1]. For a,b[D2]a, b ∈ [D_2], the (a,b)(a, b)th neighbor of a vertex (u,i)(u, i) is the vertex (v,j)(v,j) computed as follows:

(1) Let ii' be the aath neighbor of ii in HH.

(2) Let vv be the ii'th neighbor of uu in GG, so e=(u,v)e = (u,v) is the ii'th edge leaving uu. Let jj' be such that ee is the jj'th edge entering vv in GG.

(3) Let jj be the bbth neighbor of jj' in HH.

Lemma: If GG is a (N1,D1,γ1)(N_1,D_1, γ_1)-graph, and HH is a (D1,D2,γ2)(D_1,D_2, γ_2)-graph then GHG Ⓩ H is a (N1D1,D22,γ=γ1γ22)(N_1D_1,D_2^2, γ = γ_1 · γ_2^2 )-graph. In particular, if γ1=1λ1γ_1 = 1 − λ_1 and γ2=1λ2γ_2 = 1 − λ_2, then γ=1λγ = 1 − λ for λλ1+2λ2λ ≤ λ_1 + 2λ_2.

Proof: (TODO)

Replacement Product

Definition (Replacement Product): Given a D1D_1-regular graph G1G_1 on N1N_1 vertices and a D2D_2-regular graph G2G_2 on D1D_1 vertices, consider the following graph G1G2G_1 ⓡ G_2 on vertex set [N1]×[D1][N_1] × [D_1]: vertex (u,i)(u, i) is connected to (v,j)(v,j) iff u=vu = v and (i,j)(i, j) is an edge in G2G_2, or vv is the ii’th neighbor of uu in G1G_1 and uu is the jjth neighbor of vv.

(1) Prove that there is a function gg such that if G1G_1 has spectral expansion γ1>0γ_1 > 0 and G2G_2 has spectral expansion γ2>0γ_2 > 0 (and both graphs are undirected), then G1G2G_1 ⓡ G_2 has spectral expansion g(γ1,γ2,D2)>0g(γ_1, γ_2,D_2) > 0.

(2) Show how to convert an explicit construction of constant degree spectral expanders into an explicit construction of degree 33 spectral expanders.

(3) Without using Theorem 4.14, prove an analogue of Part 1 for edge expansion. That is, there is a function h such that if G1 is an (N1/2,ε1) edge expander and G2 is a (D1/2,ε2) edge expander, then G1 $r G2 is a (N1D1/2,h(ε1,ε2,D2)) edge expander, where h(ε1,ε2,D2) > 0 if ε1,ε2 > 0. (Hint: given any set S of vertices of G1 $r G2, partition S into the clouds that are more than “half-full” and those that are not.)

(4) Prove that the functions g(γ1,γ2,D2)g(γ_1, γ_2,D_2) and h(ε1,ε2,D2)h(ε_1,ε_2,D_2) must depend on D2D_2, by showing that G1G2G_1 ⓡ G_2 cannot be a (N1D1/2,ε)(N_1D_1/2,ε) edge expander if ε>1/(D2+1)ε > 1/(D_2 + 1) and N12N_1 ≥ 2.

The Expander Construction

Construction of Mildly Explicit Expanders: Let HH be a (D4,D,7/8)(D^4,D,7/8)-graph, and define:

G1=H2Gt+1=Gt2HG_1=H^2\\ G_{t+1}=G_t^2 Ⓩ H

We can easily prove that for all tt, GtG_t is a (D4t,D2,1/2)(D^{4t},D^2,1/2)-graph.

To bound the time to compute neighbors in GtG_t, we let time(Gt)\text{time}(G_t) be the time to find the IIth neighbor of XX in Gt (X[D4t],I[D2])G_t\ (X\in [D^{4t}],I\in [D^2]).

First, to do zig-zag product, we should caculate (u,i)(u,i) and (a,b)(a,b). By properly labeling of vertices and edges, we can caculate them in poly(logNt)\text{poly}(\log N_t) time (e.g. vertex (u,i)(u,i) in GtG_t can be labelled as (u1)D2+i(u-1)D^2+i, and edge (a,b)(a,b) in GtG_t can be labelled as (a1)D+b(a-1)D+b).

And then, walk on HH only takes the constant time, while walk on Gt12G_{t-1}^2 takes 2time(Gt1)2\text{time}(G_{t-1}) time (coefficient 22 is due to squaring).

Therefore: time(Gt)=2time(Gt1)+poly(logNt)=2tpoly(logNt)=NtΘ(1)\text{time}(G_t)=2\text{time}(G_{t-1})+\text{poly}(\log N_t)=2^t\text{poly}(\log N_t)=N_t^{\Theta(1)}, where the last equality holds because Nt=D4tN_t = D^{4t} for a constant DD. Thus,
this construction is only mildly explicit.

Construction of Fully Explicit Expanders: Let HH be a (D8,D,7/8)(D^8,D,7/8)-graph, and define:

G1=H2Gt+1=(GtGt)2HG_1 = H^2\\ G_{t+1} = (G_t ⊗ G_t)^2 Ⓩ H

Obviously, we have: Nt=Nt12D8N_t=N_{t-1}^2D^8. Thus: Ntc2tN_t\approx c^{2^t}.

Therefore: time(Gt)=4time(Gt1)+poly(logNt)=4tpoly(logNt)=poly(logNt)\text{time}(G_t)=4\text{time}(G_{t-1})+\text{poly}(\log N_t)=4^t\text{poly}(\log N_t)=\text{poly}(\log N_t). This construction is fully explicit.

A Better Construction of Fully Explicit Expanders: Let HH be a (D8,D,7/8)(D^8,D,7/8)-graph, and define:

G1=H2Gt+1=(Gt/2Gt/2)2HG_1 = H^2\\ G_{t+1} = (G_{\lfloor t/2\rfloor} ⊗ G_{\lceil t/2\rceil})^2 Ⓩ H

Now Nt=D8tN_t = D^{8t}, and time(Gt)=O(time(Gt/2)+time(Gt/2))+poly(logNt)=poly(t)\text{time}(G_t)=O(\text{time}(G_{\lfloor t/2\rfloor})+\text{time}(G_{\lceil t/2\rceil}))+\text{poly}(\log N_t)=\text{poly}(t).

However, these constructions do not reach Alon–Boppana Bound, i.e. they are all not the optimal spectral expanders. Therefore here remains an open problem.

Open Problem: Give an explicit “combinatorial” construction of constant-degree expander graphs GG with λ(G)2D1/D+oN(1)λ(G) ≤ 2\sqrt{D − 1}/D+o_N(1) (or even λ(G)=O(1/D)λ(G) = O(1/\sqrt D)), where DD is the degree.

Margulis-Gabber-Galil’s Construction

(TODO)

Undirected S-T Connectivity in Deterministic Logspace

Algorithm (Undirected S-T Connectivity in L\text{L}):

Input: An undirected graph GG with NN edges and vertices ss and tt.

(1) Let HH be a fixed (D4,D,3/4)(D^4,D,3/4) graph for some constant DD.

(2) Reduce (G,s,t)(G,s, t) to (G0,s0,t0)(G_0,s_0, t_0), where G0G_0 is a D2D^2-regular graph in which every connected component is nonbipartite and s0s_0 and t0t_0 are connected in G0G_0 iff ss and tt are connected in GG.

(3) For k=1,...,l=O(logN)k = 1,...,l = O(\log N), define:

  • Let Gk=Gk12HG_k = G^2_{k−1} Ⓩ H.
  • Let sks_k and tkt_k be any two vertices in the “clouds” of GkG_k corresponding to sk1s_{k−1} and tk1t_{k−1}, respectively.

(4) Try all paths of length O(logN)O(\log N) in GlG_l from sls_l and accept if any of them visit tlt_l.

Some Problems

Problem (More Combinatorial Consequences of Spectral Expansion): Let GG be a DD-regular, connected, undirected, nonbipartite graph on NN vertices with spectral expansion γ=1λγ = 1 − λ. Prove that:

(1) The independence number α(G)α(G) is at most λN/(1+λ)λN/(1 + λ).

(2) The chromatic number χ(G)χ(G) is at least (1+λ)/λ(1 + λ)/λ.

(3) The diameter of GG is O(log1/λN)O(\log_{1/λ} N).

Proof:

(1) Let SS be the maximum independent set of GG, then e(S,S)e(S,S)=0.

By Expander Mixing Lemma, we have: μ2(S)λμ(S)(1μ(S))μ(S)λ/(1+λ)\mu^2(S)\le \lambda \mu(S)(1-\mu(S))\Rightarrow \mu(S)\le \lambda/(1+\lambda).

Thus, α(G)=μ(S)Nα(G)=\mu(S)N is at most λN/(1+λ)λN/(1 + λ), as desired.

(2) Let c:V(G)[χ(G)]c:V(G)\to [\chi(G)] be the coloring mapping of GG.

Notice that for every 1iχ(G)1\le i\le \chi(G), c1(i)c^{-1}(i) is an independent set of GG.

Thus: N=i=1χ(G)c1(i)χ(G)λN/(1+λ)χ(G)(1+λ)/λN=\sum_{i=1}^{\chi(G)}|c^{-1}(i)|\le \chi(G) \cdot \lambda N/(1+\lambda)\Rightarrow \chi(G)\ge (1+\lambda)/\lambda, as desired.

(3) Let MM be the random walk matrix for GG, and it has eigenvalues λ1λ2λn|\lambda_1|\ge |\lambda_2|\ge \ldots \ge |\lambda_n|, where λ1=1\lambda_1=1 and λiλ(i>1)|\lambda_i|\le \lambda\,(i>1).

By using spectral decomposition, we can write MM as:

M=i=1NλiviviTM=\sum_{i=1}^N \lambda_i v_iv_i^T

where v1,,vnv_1,\ldots,v_n denote normalized orthonormal column eigenvectors with eigenvalues λ1,,λn\lambda_1,\ldots,\lambda_n.

For a positive integer mm, consider arbitrary entry of matrix MmM^m:

(Mm)r,s=i=1Nλim(viviT)r,s1Ni=2Nλim(vi)r(vi)s1Nλmi=2N(vi)r(vi)s1Nλmi=2N(vi)r2i=2N(vi)s2=1Nλm(1(v1)r2)(1(v1)s2)=1Nλm(11/N)\begin{aligned} (M^m)_{r,s} & =\sum_{i=1}^N \lambda_i^m (v_iv_i^T)_{r,s} \\ & \ge \dfrac 1N - \left|\sum_{i=2}^N \lambda_i^m (v_i)_r (v_i)_s \right| \\ & \ge \dfrac 1N - \lambda^m\left|\sum_{i=2}^N (v_i)_r (v_i)_s \right| \\ & \ge \dfrac 1N - \lambda^m\sqrt{\sum_{i=2}^N (v_i)_r^2 \sum_{i=2}^N (v_i)_s^2} \\ & = \dfrac 1N - \lambda^m\sqrt{(1-(v_1)_r^2)(1-(v_1)_s^2)} \\ & = \dfrac 1N -\lambda^m(1-1/N) \end{aligned}

Thus if we have m>log1/λ(N1)m>\log_{1/\lambda}(N-1), then 1r,sN,(Mm)r,s>0\forall 1\le r,s\le N,(M^m)_{r,s}>0, i.e. MmM^m has all entries positive.

Therefore the diameter of GG is at most log1/λ(N1)+1\lfloor \log_{1/\lambda}(N-1)+1 \rfloor, as desired. \Box

Problem (Error Reduction For Free): Show that if a problem has a BPP\textbf{BPP} algorithm with constant error probability, then it has a BPP\textbf{BPP} algorithm with error probability 1/n1/n that uses exactly the same
number of random bits.

Proof: Let A(x;r)A(x;r) be the BPP\textbf{BPP} algorithm of this problem, which r{0,1}mr\in \{0,1\}^m.

Then we can design an expander with 2m2^m vertices, and do a random walk on it. However, this method costs the extra ranom bits.

Notice that if we only want to reduce error probability less than 1/n1/n, the length of the random walk is only O(logn)O(\log n). Thus we can directly try all paths of length O(logn)O(\log n) in polynomial time by using some method such as depth first search. By this way, we only need to randomly choose the starting vertex, which only need mm random bits. \Box

Reference

[1] Pseudorandomness. Salil P. Vadhan, 2012. From Harvard University, site: https://people.seas.harvard.edu/~salil/pseudorandomness/

[2] Expander Graphs and Their Applications. Shlomo Hoory, Nathan Linial, Avi Wigderson, 2006. From Hebrew University of Jerusalem, site: https://www.cs.huji.ac.il/~nati/PAPERS/expander_survey.pdf

[3] Expander graphs and High-Dimensional Expanders. Shachar Lovett, Max Hopkins. From University of California San Diego, site: https://cseweb.ucsd.edu/classes/sp21/cse291-g/

[4] Computational Complexity: A Modern Approach. Sanjeev Arora, Boaz Barak, 2009.

[5] Chung, F. R. K. (1989). “Diameters and eigenvalues”. Journal of the American Mathematical Society. 2 (2): 187–196. doi:10.1090/S0894-0347-1989-0965008-X. ISSN 0894-0347.