Expanders and the Mixing Lemma

Themistoklis Haris

Themistoklis Haris

This write-up will be about expander graphs, an indispensable tool in computer science. Expanders are widely used in error-correcting codes, complexity theory, hashing and computer networks.

Definitions

Let G=(V,E)G=(V,E) be a graph. Let S,TVS,T \subseteq V. We define E(S,T):={(u,v){u,v}E,uS,vT}E(S,T) := \{(u,v)\mid \{u,v\} \in E, u\in S, v\in T\}: this is the edges uvu\to v from SS to TT. Let e(S,T)=E(S,T)e(S,T) = |E(S,T)|. Note that if SS and TT have a non-empty intersection then we count the edges between SS and TT twice.

We define the edge boundary of a set SVS \subseteq V as S:=E(S,VS)\partial S := E(S,V\setminus S). The vertex boundary is the number of neighbors of SS that are not in SS.

Definition

A graph is a ε\varepsilon-edge expander if all subsets SVS \subset V such that SV/2|S| \leq |V|/2 satisfy:

h(S)=SE(S,V)εh(S) = \frac{|\partial S|}{|E(S,V)|} \geq \varepsilon

Intuitively, we can think of it as follows: a social network is a 1/21/2 expander if for all groups of people, the number of friends outside of the group is at least 1/21/2 the total number of friends of the group. We have a name for the largest ε\varepsilon this definition works for: the isoperimetric number, or Cheeger constant.

h(G):=infS[h(S)]h(G) := \text{inf}_{S}[h(S)]

An adjacent definition, an ε\varepsilon-vertex expander is a graph where all SVS \subset V with SV/2|S|\leq |V|/2 have N(S)εSN(S) \geq \varepsilon |S|. These two notions of expansion are combinatorial and well-motivated. Let's see an example

Examples

  • Consider KnK_n. We can prove that it is a 11-vertex expander (which is actually optimal) and a 12+O(1/n)\frac{1}{2} + O(1/n) edge expander. This is a good exercise that we will leave for the end.
  • Cycle graphs and Cayley graphs are interesting constructions of expanders.

Spectral Expansion

Beyond the combinatorial definitions, we can also talk about spectral expansion. Let GG be dd-regular, and consider its normalized adjacency matrix: AG=A/dA_G = A/d. Let λ1λ2λn\lambda_1\geq \lambda_2\geq\cdots\geq \lambda_n be the spectrum of AGA_G. By the Perron-Frobenius Theorem we can show that:

  • 1\vec{1} is an eigenvector of AGA_G with eigenvalue 1. This eigenvector is also called a Perron vector, because all its coordinates are strictly positive.
  • 1=λ1λn11 = \lambda_1 \geq \cdots \geq \lambda_n \geq -1.

Definition

GG is a λ\lambda-spectral expander if

λ(G):=max{λ2,λn}λ\lambda(G) := \max\{|\lambda_2|,|\lambda_n|\} \leq \lambda

We define the spectral gap of a graph as 1λ21-\lambda_2.

What does the spectrum of a graph have anything to do with its expansion properties? The answer is that they are closely related, as we will see in the following theorem:

Theorem

Expander Mixing Lemma
Let GG be a dd-regular graph with normalized adjacency matrix AGA_G whose spectrum is λ1λn\lambda_1 \geq \cdots \geq \lambda_n. Then, for all S,TVS,T \subseteq V, we have:

e(S,T)dSTnλ(G)dST\left|e(S,T)-\frac{d|S||T|}{n}\right|\leq \lambda(G)\cdot d\sqrt{|S|\cdot |T|}

Proof: We kick things off with an important observation:

Tip
1ST(dAG)1T=e(S,T)1_S^T (d A_G) 1_T = e(S,T)

This is because (dAG)1T(d A_G) 1_T gives for every vVv \in V the number of neighbors in TT, and subsequently multiplying by 1S1_S only keeps vSv \in S.

Now, let us consider an orthonormal eigenbasis {v1,...,vn}\{v_1,...,v_n\} for AGA_G. We can write 1S=αivi1_S = \sum \alpha_i v_i and 1T=βivi1_T = \sum \beta_i v_i and then

1ST(AG)1T=i=1nαiviT(i=1nλiviviT)i=1nβivi=i=1nλiαiβi1_S^T (A_G) 1_T = \sum\limits_{i=1}^n\alpha_i v_i^T\left(\sum\limits_{i=1}^n \lambda_i v_i v_i^T\right)\sum\limits_{i=1}^n\beta_i v_i = \sum\limits_{i=1}^n \lambda_i \alpha_i \beta_i

because of orthonormality. Now, note that λ1=1\lambda_1 = 1 and so α1=1ST1n=Sn\alpha_1 = 1_S^T \frac{\vec{1}}{\sqrt{n}} = \frac{|S|}{\sqrt{n}}. Similarly, βi=1TT1n=Tn\beta_i = 1_T^T \frac{\vec{1}}{\sqrt{n}} = \frac{|T|}{\sqrt{n}}. So:

1ST(AG)1T=STn+i=2nλiαiβi1_S^T (A_G)1_T = \frac{|S|\cdot |T|}{n} + \sum\limits_{i=2}^n \lambda_i \alpha_i\beta_i

Which implies, by the triangle inequality, that:

e(S,T)dSTn=i=2nλiαiβiλ(G)i=2nαiβiλ(G)ST\left|\frac{e(S,T)}{d}-\frac{|S|\cdot |T|}{n}\right| = \left|\sum\limits_{i=2}^n \lambda_i \alpha_i \beta_i\right| \leq \lambda(G)\sum\limits_{i=2}^n\left|\alpha_i\beta_i\right|\leq \lambda(G)\sqrt{|S|\cdot |T|}

which finally gives us the desired result. Note that S=αi2|S| = \sum\alpha_i^2 and T=βi2|T| = \sum\beta_i^2 and so by Cauchy-Schwarz we have:

(αiβi)2(αi2)(βi2)=ST\left(\sum\alpha_i\beta_i\right)^2 \leq \left(\sum \alpha_i^2\right)\cdot\left(\sum\beta_i^2\right) = |S|\cdot |T|

This finishes the proof of the Expander Mixing Lemma. Next, we shall see some applications of it.

Applications of the Expander Mixing Lemma

  1. Let α(G)\alpha(G) be the size of the largest independent set in a graph GG. If GG is a dd-regular λ\lambda-expander, then
α(G)λn\alpha(G) \leq \lambda n

To see why, consider any independent set SS and apply the expander mixing lemma with T=ST = S. We have that e(S,T)=0e(S,T) = 0, so:

S2nλS\frac{|S|^2}{n} \leq \lambda |S|

which gives us the desired relationship.

  1. If GG is a dd-regular λ\lambda-expander with λ<1/3\lambda < 1/3, then its diameter (the largest distance between any two vertices) satisfies:
Ω(logn/logd)diam(G)O(logn)\Omega(\log n/\log d) \leq \text{diam}(G) \leq O(\log n)