← All blueprints

Expander Graphs and Their Applications

1 The magical mystery tour

Definition 1 Hamming distance
#

For words \(u,v\in \Sigma ^n\) over an alphabet \(\Sigma \), the Hamming distance \(d_H(u,v)\) is the number of coordinates \(i\) with \(u_i\neq v_i\).

Theorem 2 Hall’s marriage theorem
#

Let \(G=(L\cup R,E)\) be a bipartite graph. There is a matching saturating \(L\) iff \(|\Gamma (S)|\ge |S|\) for every \(S\subseteq L\).

Lemma 3 Union bound
#

For finite events \(A_1,\dots ,A_m\), \(\Pr [\bigcup _i A_i]\le \sum _i\Pr [A_i]\).

Lemma 4 Binomial coefficient bound
#

For integers \(1\le k\le n\), \(\binom nk\le (en/k)^k\).

Lemma 5 Chernoff bound for the binomial distribution
#

If \(Z\sim \mathrm{Binomial}(N,p)\) then for \(0{\lt}\Delta {\lt}1\), \(\Pr [|Z-\mathbb {E}Z|\ge \Delta \mathbb {E}Z]{\lt}2\exp (-Np\Delta ^2/3)\).

Problem 6 1.1, Hardness of linear transformations
#

Let \(A\) be an \(n\times n\) matrix over a field \(\mathcal F\). What is the least number of gates in an arithmetic circuit computing \(x\mapsto Ax\), where each gate has two inputs \(x,y\), two field coefficients \(a,b\), and output \(ax+by\)?

Definition 7 1.2, Super regular matrix
#

A matrix \(A\) is super regular if every square submatrix of \(A\) has full rank.

Definition 8 1.3, Super concentrator
#

Let \(G=(V,E)\) and let \(I,O\subseteq V\) each have \(n\) vertices (inputs and outputs). \(G\) is a super concentrator if for every \(k\) and every \(S\subseteq I\), \(T\subseteq O\) with \(|S|=|T|=k\) there are \(k\) vertex-disjoint paths from \(S\) to \(T\).

Problem 9 1.4, Communication over a noisy channel
#

Alice sends a \(k\)-bit message to Bob over a channel that may flip a fraction \(p\) of bits. What is the smallest \(n\) so that Bob can always recover the message?

Definition 10 1.5, Rate and distance of a code

For a code \(C\subseteq \{ 0,1\} ^n\), the rate and (normalized) distance are \(R=\frac{\log _2|C|}{n}\) and \(\delta =\frac{\min _{c_1\neq c_2\in C}d_H(c_1,c_2)}{n}\).

Problem 11 1.6, Refined communication problem

Do there exist arbitrarily large codes \(\{ C_k\} \), \(|C_k|=2^k\), with \(R(C_k)\ge R_0\) and \(\delta (C_k)\ge \delta _0\) for absolute constants \(R_0,\delta _0{\gt}0\), that are explicit and efficiently encodable/decodable?

Definition 12 \(\mathbf{RP}\) and one-sided error
#

\(\mathcal L\subseteq \{ 0,1\} ^*\) is in \(\mathbf{RP}\) if a randomized polynomial-time \(A\) satisfies: \(x\in \mathcal L\Rightarrow A(x,r)=1\) for all \(r\); \(x\notin \mathcal L\Rightarrow \Pr _r[A(x,r)=1]{\lt}1\) (fixed below \(1/16\)), with \(r\) uniform of length \(k=\mathrm{poly}(|x|)\).

Problem 13 1.7, Saving random bits

Given a one-sided-error randomized membership algorithm for \(\mathcal L\), how many random bits suffice to reduce the error to \(\le \epsilon \) on every input?

Definition 14 1.8, Magical graph
#

A bipartite graph \(G=(L,R,E)\) with \(|L|=n\), \(|R|=m\), every left vertex of degree \(d\), and neighbourhood map \(\Gamma \), is an \((n,m;d)\)-magical graph if: (1) \(|\Gamma (S)|\ge \tfrac {5d}8|S|\) for \(|S|\le \tfrac n{10d}\); and (2) \(|\Gamma (S)|\ge |S|\) for \(\tfrac n{10d}{\lt}|S|\le \tfrac n2\).

Lemma 15 1.9, Magical graphs exist

There is \(n_0\) so that for every \(d\ge 32\), \(n\ge n_0\), \(m\ge 3n/4\), an \((n,m;d)\)-magical graph exists.

Proof

Let \(G\) be random bipartite, each left vertex joined to a uniform \(d\)-set on the right. Property (1): for \(S\), \(s=|S|\le \tfrac n{10d}\) and \(T\subseteq R\), \(t{\lt}\tfrac {5ds}8\), let \(X_{S,T}\) indicate all edges from \(S\) land in \(T\); \(\Pr [X_{S,T}=1]=(t/m)^{sd}\). By the union bound and \(\binom nk\le (en/k)^k\),

\[ \Pr [\textstyle \sum X_{S,T}{\gt}0]\le \sum _{s\le n/10d}\binom ns\binom m{5ds/8} \bigl(\tfrac {5ds}{8m}\bigr)^{sd}\le \sum _s\bigl(\tfrac {ne}s\bigr)^s \bigl(\tfrac {8me}{5ds}\bigr)^{5ds/8}\bigl(\tfrac {5ds}{8m}\bigr)^{sd}{\lt}\tfrac 1{10}, \]

the \(s\)-th term being \(\le 20^{-s}\) for \(d\ge 32\). Property (2): for \(\tfrac n{10d}{\lt}s\le \tfrac n2\), \(t{\lt}s\), with \(Y_{S,T}\) the analogue, \(\Pr [\sum Y_{S,T}{\gt}0]\le \sum _s\binom ns\binom ms(s/m)^{sd} \le \sum _s[(ne/s)(me/s)(s/m)^d]^s{\lt}\tfrac 1{10}\), each bracket \(\le 10^{-4}\). The total failure probability is \({\lt}1\), so a magical (indeed most) graph exists.

Lemma 16 Magical graphs have perfect matchings

In an \((n,3n/4;d)\)-magical graph, \(|\Gamma (S)|\ge |S|\) for every \(S\subseteq L\) with \(|S|\le \tfrac n2\), so by Hall’s theorem there is a matching saturating any such \(S\).

Proof

Property (1) gives \(|\Gamma (S)|\ge \tfrac {5d}8|S|\ge |S|\) for \(|S|\le \tfrac n{10d}\); property (2) gives \(|\Gamma (S)|\ge |S|\) for \(\tfrac n{10d}{\lt}|S|\le \tfrac n2\). Thus Hall’s condition holds for all such \(S\).

Lemma 17 Unique-neighbour property

In an \((n,3n/4;d)\)-magical graph, every nonempty \(S\subseteq L\) with \(|S|\le \tfrac n{10d}\) has a vertex \(u\in R\) with \(|\Gamma (u)\cap S|=1\).

Proof

The number of edges between \(S\) and \(\Gamma (S)\) is \(d|S|\). By property (1) \(|\Gamma (S)|\ge \tfrac {5d}8|S|\), so the average number of \(S\)-neighbours of a vertex of \(\Gamma (S)\) is \(\le 8/5{\lt}2\); since each has \(\ge 1\), some has exactly one.

Theorem 18 1.3.1, Super concentrator with \(O(n)\) edges

There exist super concentrators with \(n\) inputs, \(n\) outputs and \(O(n)\) edges.

Proof

Recursively: for \(n{\lt}n_0\) use the complete bipartite graph (\(n^2\) edges). For \(n\ge n_0\) build \(C'\) from two copies \(G_1,G_2\) of an \((n,3n/4;d)\)-magical graph, a super concentrator \(C\) on \(R_1,R_2\) (size \(3n/4\), by induction), and a perfect matching between \(L_1,L_2\); inputs/outputs are \(L_1,L_2\). For \(|S|=|T|=k\le n/2\), Lemma 16 and Hall give matchings of \(S,T\) into \(R_1,R_2\); their images are joined by \(k\) disjoint paths in \(C\). For \(k{\gt}n/2\), route \(k-n/2\) pairs directly through the matching and reduce to \(k\le n/2\). The edge recursion \(e(n)\le 2nd+n+e(3n/4)\) (\(n{\gt}n_0\)), \(e(n)=n^2\) (\(n\le n_0\)) solves to \(e(n)\le Kn\).

Theorem 19 1.3.2, Good code from a magical graph

An \((n,3n/4;d)\)-magical graph yields a linear code \(C\subseteq \{ 0,1\} ^n\) with rate \(R\ge \tfrac 14\) and normalized distance \(\delta \ge \tfrac 1{10d}\).

Proof

Let \(A\) be the \(|R|\times |L|\) bipartite adjacency matrix over \(\mathrm{GF}(2)\) and \(C=\{ x:Ax=0\} \). Then \(\dim C\ge n-|R|=n/4\), so \(R\ge \tfrac 14\). Since \(C\) is linear, its distance is the minimum weight of a nonzero codeword; if a nonzero \(x\) had support \(S\) with \(|S|{\lt}\tfrac n{10d}\), then by Lemma 17 some coordinate of \(Ax\) is \(1\), so \(x\notin C\). Hence \(\delta \ge \tfrac 1{10d}\).

Theorem 20 1.3.3, Deterministic error amplification for \(\mathbf{RP}\)

For \(\mathcal L\in \mathbf{RP}\) with one-sided-error algorithm \(f\) using \(k\) bits, \(n=2^k\), failing on \(B\) with \(|B|\le n/16\): for every \(d\) there is an algorithm evaluating \(f\) exactly \(d\) times, using only the \(k\) random bits, with error \(\le \tfrac 1{10d}\).

Proof

Use an \((n,n;d)\)-magical graph \(G=(L,R,E)\), \(R\) identified with \(\{ 0,1\} ^k\). Sample \(r\in L\), let \(r_1,\dots ,r_d\) be its neighbours, output \(\bigwedge _i f(x,r_i)\). For \(x\in \mathcal L\) this is always \(1\). For \(x\notin \mathcal L\) failure means all \(r_i\in B\); let \(S=\{ r:\Gamma (r)\subseteq B\} \). If \(|S|{\gt}\tfrac n{10d}\) then by property (1) \(|B|\ge |\Gamma (S)|{\gt}\tfrac {5d}8\cdot \tfrac n{10d}=\tfrac n{16}\), contradiction. So \(|S|\le \tfrac n{10d}\) and \(\Pr [r\in S]\le \tfrac 1{10d}\).