You are currently browsing the tag archive for the ‘spectral gap’ tag.

In the previous set of notes we introduced the notion of expansion in arbitrary {k}-regular graphs. For the rest of the course, we will now focus attention primarily to a special type of {k}-regular graph, namely a Cayley graph.

Definition 1 (Cayley graph) Let {G = (G,\cdot)} be a group, and let {S} be a finite subset of {G}. We assume that {S} is symmetric (thus {s^{-1} \in S} whenever {s \in S}) and does not contain the identity {1} (this is to avoid loops). Then the (right-invariant) Cayley graph {Cay(G,S)} is defined to be the graph with vertex set {G} and edge set {\{ \{sx,x\}: x \in G, s \in S \}}, thus each vertex {x \in G} is connected to the {|S|} elements {sx} for {s \in S}, and so {Cay(G,S)} is a {|S|}-regular graph.

Example 2 The graph in Exercise 3 of Notes 1 is the Cayley graph on {{\bf Z}/N{\bf Z}} with generators {S = \{-1,+1\}}.

Remark 3 We call the above Cayley graphs right-invariant because every right translation {x\mapsto xg} on {G} is a graph automorphism of {Cay(G,S)}. This group of automorphisms acts transitively on the vertex set of the Cayley graph. One can thus view a Cayley graph as a homogeneous space of {G}, as it “looks the same” from every vertex. One could of course also consider left-invariant Cayley graphs, in which {x} is connected to {xs} rather than {sx}. However, the two such graphs are isomorphic using the inverse map {x \mapsto x^{-1}}, so we may without loss of generality restrict our attention throughout to left Cayley graphs.

Remark 4 For minor technical reasons, it will be convenient later on to allow {S} to contain the identity and to come with multiplicity (i.e. it will be a multiset rather than a set). If one does so, of course, the resulting Cayley graph will now contain some loops and multiple edges.
For the purposes of building expander families, we would of course want the underlying group {G} to be finite. However, it will be convenient at various times to “lift” a finite Cayley graph up to an infinite one, and so we permit {G} to be infinite in our definition of a Cayley graph.

We will also sometimes consider a generalisation of a Cayley graph, known as a Schreier graph:

Definition 5 (Schreier graph) Let {G} be a finite group that acts (on the left) on a space {X}, thus there is a map {(g,x) \mapsto gx} from {G \times X} to {X} such that {1x = x} and {(gh)x = g(hx)} for all {g,h \in G} and {x \in X}. Let {S} be a symmetric subset of {G} which acts freely on {X} in the sense that {sx \neq x} for all {s \in S} and {x \in X}, and {sx \neq s'x} for all distinct {s,s' \in S} and {x \in X}. Then the Schreier graph {Sch(X,S)} is defined to be the graph with vertex set {X} and edge set {\{ \{sx,x\}: x \in X, s \in S \}}.

Example 6 Every Cayley graph {Cay(G,S)} is also a Schreier graph {Sch(G,S)}, using the obvious left-action of {G} on itself. The {k}-regular graphs formed from {l} permutations {\pi_1,\ldots,\pi_l \in S_n} that were studied in the previous set of notes is also a Schreier graph provided that {\pi_i(v) \neq v, \pi_i^{-1}(v), \pi_j(v)} for all distinct {1 \leq i,j \leq l}, with the underlying group being the permutation group {S_n} (which acts on the vertex set {X = \{1,\ldots,n\}} in the obvious manner), and {S := \{\pi_1,\ldots,\pi_l,\pi_1^{-1},\ldots,\pi_l^{-1}\}}.

Exercise 7 If {k} is an even integer, show that every {k}-regular graph is a Schreier graph involving a set {S} of generators of cardinality {k/2}. (Hint: you may assume without proof Petersen’s 2-factor theorem, which asserts that every {k}-regular graph with {k} even can be decomposed into {k/2} edge-disjoint {2}-regular graphs. Now use the previous example.)

We return now to Cayley graphs. It is easy to characterise qualitative expansion properties of Cayley graphs:

Exercise 8 (Qualitative expansion) Let {Cay(G,S)} be a finite Cayley graph.

  • (i) Show that {Cay(G,S)} is a one-sided {\varepsilon}-expander for {G} for some {\varepsilon>0} if and only if {S} generates {G}.
  • (ii) Show that {Cay(G,S)} is a two-sided {\varepsilon}-expander for {G} for some {\varepsilon>0} if and only if {S} generates {G}, and furthermore {S} intersects each index {2} subgroup of {G}.

We will however be interested in more quantitative expansion properties, in which the expansion constant {\varepsilon} is independent of the size of the Cayley graph, so that one can construct non-trivial expander families {Cay(G_n,S_n)} of Cayley graphs.
One can analyse the expansion of Cayley graphs in a number of ways. For instance, by taking the edge expansion viewpoint, one can study Cayley graphs combinatorially, using the product set operation

\displaystyle  A \cdot B := \{ab: a \in A, b \in B \}

of subsets of {G}.

Exercise 9 (Combinatorial description of expansion) Let {Cay(G_n,S_n)} be a family of finite {k}-regular Cayley graphs. Show that {Cay(G_n,S_n)} is a one-sided expander family if and only if there is a constant {c>0} independent of {n} such that {|E_n \cup E_n S_n| \geq (1+c) |E_n|} for all sufficiently large {n} and all subsets {E_n} of {G_n} with {|E_n| \leq |G_n|/2}.

One can also give a combinatorial description of two-sided expansion, but it is more complicated and we will not use it here.

Exercise 10 (Abelian groups do not expand) Let {Cay(G_n,S_n)} be a family of finite {k}-regular Cayley graphs, with the {G_n} all abelian, and the {S_n} generating {G_n}. Show that {Cay(G_n,S_n)} are a one-sided expander family if and only if the Cayley graphs have bounded cardinality (i.e. {\sup_n |G_n| < \infty}). (Hint: assume for contradiction that {Cay(G_n,S_n)} is a one-sided expander family with {|G_n| \rightarrow \infty}, and show by two different arguments that {\sup_n |S_n^m|} grows at least exponentially in {m} and also at most polynomially in {m}, giving the desired contradiction.)

The left-invariant nature of Cayley graphs also suggests that such graphs can be profitably analysed using some sort of Fourier analysis; as the underlying symmetry group is not necessarily abelian, one should use the Fourier analysis of non-abelian groups, which is better known as (unitary) representation theory. The Fourier-analytic nature of Cayley graphs can be highlighted by recalling the operation of convolution of two functions {f, g \in \ell^2(G)}, defined by the formula

\displaystyle  f * g(x) := \sum_{y \in G} f(y) g(y^{-1} x) = \sum_{y \in G} f(x y^{-1}) g(y).

This convolution operation is bilinear and associative (at least when one imposes a suitable decay condition on the functions, such as compact support), but is not commutative unless {G} is abelian. (If one is more algebraically minded, one can also identify {\ell^2(G)} (when {G} is finite, at least) with the group algebra {{\bf C} G}, in which case convolution is simply the multiplication operation in this algebra.) The adjacency operator {A} on a Cayley graph {Cay(G,S)} can then be viewed as a convolution

\displaystyle  Af = |S| \mu * f,

where {\mu} is the probability density

\displaystyle  \mu := \frac{1}{|S|} \sum_{s \in S} \delta_s \ \ \ \ \ (1)

where {\delta_s} is the Kronecker delta function on {s}. Using the spectral definition of expansion, we thus see that {Cay(G,S)} is a one-sided expander if and only if

\displaystyle  \langle f, \mu*f \rangle \leq (1-\varepsilon) \|f\|_{\ell^2(G)}^2 \ \ \ \ \ (2)

whenever {f \in \ell^2(G)} is orthogonal to the constant function {1}, and is a two-sided expander if

\displaystyle  \| \mu*f \|_{\ell^2(G)} \leq (1-\varepsilon) \|f\|_{\ell^2(G)} \ \ \ \ \ (3)

whenever {f \in \ell^2(G)} is orthogonal to the constant function {1}.
We remark that the above spectral definition of expansion can be easily extended to symmetric sets {S} which contain the identity or have multiplicity (i.e. are multisets). (We retain symmetry, though, in order to keep the operation of convolution by {\mu} self-adjoint.) In particular, one can say (with some slight abuse of notation) that a set of elements {s_1,\ldots,s_l} of {G} (possibly with repetition, and possibly with some elements equalling the identity) generates a one-sided or two-sided {\varepsilon}-expander if the associated symmetric probability density

\displaystyle  \mu := \frac{1}{2l} \sum_{i=1}^l \delta_{s_i} + \delta_{s_i^{-1}}

obeys either (2) or (3).
We saw in the last set of notes that expansion can be characterised in terms of random walks. One can of course specialise this characterisation to the Cayley graph case:

Exercise 11 (Random walk description of expansion) Let {Cay(G_n,S_n)} be a family of finite {k}-regular Cayley graphs, and let {\mu_n} be the associated probability density functions. Let {A > 1/2} be a constant.

  • Show that the {Cay(G_n,S_n)} are a two-sided expander family if and only if there exists a {C>0} such that for all sufficiently large {n}, one has {\| \mu_n^{*m} - \frac{1}{|G_n|} \|_{\ell^2(G_n)} \leq \frac{1}{|G_n|^A}} for some {m \leq C \log |G_n|}, where {\mu_n^{*m} := \mu_n * \ldots * \mu_n} denotes the convolution of {m} copies of {\mu_n}.
  • Show that the {Cay(G_n,S_n)} are a one-sided expander family if and only if there exists a {C>0} such that for all sufficiently large {n}, one has {\| (\frac{1}{2} \delta_1 + \frac{1}{2} \mu_n)^{*m} - \frac{1}{|G_n|} \|_{\ell^2(G_n)} \leq \frac{1}{|G_n|^A}} for some {m \leq C \log |G_n|}.

In this set of notes, we will connect expansion of Cayley graphs to an important property of certain infinite groups, known as Kazhdan’s property (T) (or property (T) for short). In 1973, Margulis exploited this property to create the first known explicit and deterministic examples of expanding Cayley graphs. As it turns out, property (T) is somewhat overpowered for this purpose; in particular, we now know that there are many families of Cayley graphs for which the associated infinite group does not obey property (T) (or weaker variants of this property, such as property {\tau}). In later notes we will therefore turn to other methods of creating Cayley graphs that do not rely on property (T). Nevertheless, property (T) is of substantial intrinsic interest, and also has many connections to other parts of mathematics than the theory of expander graphs, so it is worth spending some time to discuss it here.
The material here is based in part on this recent text on property (T) by Bekka, de la Harpe, and Valette (available online here).
Read the rest of this entry »

The objective of this course is to present a number of recent constructions of expander graphs, which are a type of sparse but “pseudorandom” graph of importance in computer science, the theory of random walks, geometric group theory, and in number theory. The subject of expander graphs and their applications is an immense one, and we will not possibly be able to cover it in full in this course. In particular, we will say almost nothing about the important applications of expander graphs to computer science, for instance in constructing good pseudorandom number generators, derandomising a probabilistic algorithm, constructing error correcting codes, or in building probabilistically checkable proofs. For such topics, I recommend the survey of Hoory-Linial-Wigderson. We will also only pass very lightly over the other applications of expander graphs, though if time permits I may discuss at the end of the course the application of expander graphs in finite groups such as {SL_2(F_p)} to certain sieving problems in analytic number theory, following the work of Bourgain, Gamburd, and Sarnak.

Instead of focusing on applications, this course will concern itself much more with the task of constructing expander graphs. This is a surprisingly non-trivial problem. On one hand, an easy application of the probabilistic method shows that a randomly chosen (large, regular, bounded-degree) graph will be an expander graph with very high probability, so expander graphs are extremely abundant. On the other hand, in many applications, one wants an expander graph that is more deterministic in nature (requiring either no or very few random choices to build), and of a more specialised form. For the applications to number theory or geometric group theory, it is of particular interest to determine the expansion properties of a very symmetric type of graph, namely a Cayley graph; we will also occasionally work with the more general concept of a Schreier graph. It turns out that such questions are related to deep properties of various groups {G} of Lie type (such as {SL_2({\bf R})} or {SL_2({\bf Z})}), such as Kazhdan’s property (T), the first nontrivial eigenvalue of a Laplacian on a symmetric space {G/\Gamma} associated to {G}, the quasirandomness of {G} (as measured by the size of irreducible representations), and the product theory of subsets of {G}. These properties are of intrinsic interest to many other fields of mathematics (e.g. ergodic theory, operator algebras, additive combinatorics, representation theory, finite group theory, number theory, etc.), and it is quite remarkable that a single problem – namely the construction of expander graphs – is so deeply connected with such a rich and diverse array of mathematical topics. (Perhaps this is because so many of these fields are all grappling with aspects of a single general problem in mathematics, namely when to determine whether a given mathematical object or process of interest “behaves pseudorandomly” or not, and how this is connected with the symmetry group of that object or process.)

(There are also other important constructions of expander graphs that are not related to Cayley or Schreier graphs, such as those graphs constructed by the zigzag product construction, but we will not discuss those types of graphs in this course, again referring the reader to the survey of Hoory, Linial, and Wigderson.)

Read the rest of this entry »