Bourbaki

Exercise 16, § 2

Tag 02ZZ, permanent. Cite it as [Bourbaki, Tag 02ZZ].

Spectral Theories, IV, p. 323

We retain the notation of the preceding exercise, and we suppose that S and F are finite. We denote by hGh_G the Cheeger constant of the graph G, defined by

Card(E(V,S,S - V))

hG=h_G= inf,

∅≠V⊂S_{\emptyset\not=V\subset S} inf(Card(V), Card(S - V))

where E(V,W)E(V,W) denotes the set of arrows of G with one end in V and one in W.

a) One has ϱ(MG)=1\varrho (M_G) = 1 and λ=1\lambda = 1 is an eigenvalue of MGM_G of multiplicity 1. One has −1∈-1\in Sp(MG_G) if and only if G is bipartite (cf. TA, II, p. 219, exercise 2); in this case, the spectrum of MGM_G is invariant under λ↦−λ\lambda \mapsto -\lambda and the spectral multiplicity of λ\lambda is equal to that of −λ-\lambda for every λ∈C\lambda \in \mathbf{C}.

We denote by L02(S,ν)L^2_0(S, \nu ) the orthogonal complement in LC2(S,ν)L^2_{\mathbf{C}}(S, \nu ) of the union of the eigenspaces of MGM_G corresponding to the eigenvalues 1 and −1-1. The endomorphism MGM_G induces by passing to the subspaces an endomorphism of L02(S,ν)L^2_0(S, \nu ), denoted by MG,0M_{G,0}. One has ϱ(MG,0)<1\varrho (M_{G,0})<1. We denote by λ1(G)\lambda_1(G) the smallest nonzero eigenvalue of 1−MG1-M_G

b) Let v+=v_+= sup v(x)v(x) and v−=v_-= inf v(x)v(x). One has

λ1(G)⩽2v+2hG \lambda_1(G)\leqslant \frac{2v_+}{2}h_G
v− v_-

c) Let φ:S→R\varphi : S\rightarrow \mathbf{R} be a non-constant function on S with real values; let m=m= inf−1φ(x)_{-1}\varphi (x) and M = sup φ(x)\varphi (x). Let t0∈Rt_0\in \mathbf{R} be such that, for every t∈Rt\in \mathbf{R}, the set Wt=φ(]−∞,t0[)W_t=\varphi (]- \infty , t_0[) satisfies Card(Wt_t)⩽\leqslant Card(S)/2/2 if and only if t<t0t < t_0. Let ν\nu be a diffuse measure (INT, V, p. 67, Def. 5) supported on [m,M][m,M]. Put

A=12∑a(x,y)ν([φ(x),φ(y)]) A =\frac{1}{2}\sum a(x, y)\nu ([\varphi (x), \varphi (y)])
(x,y)∈S2(x,y)\in S^2

B=∑x∈Sν([t0,φ(x)])B =\sum_{x\in S}\nu ([t_0, \varphi (x)]), where, if a<ba < b are real numbers, we write ν([b,a])=ν([a,b])\nu ([b, a]) =\nu ([a, b]). There exists t∈Rt\in \mathbf{R} such that

Card(E(Wt,S_t,S - WtW_t) A

⩽ \leqslant

inf(Card(Wt_t), Card(S - WtW_t)) B

d) One has

hG⩽v+√2λ1(G)‾ h_G\leqslant v_+\surd\overline{2\lambda_1(G)}

(Apply the preceding question with φ\varphi an eigenfunction of MGM_G for the eigenvalue λ1(G)\lambda_1(G) and for a suitably chosen measure ν\nu.)

e) A family (Gj)j∈J(G_j)_{j\in J} of finite graphs Gj=(Sj,Fj,oj,tj,ij)G_j= (S_j,F_j, o_j, t_j, i_j) is said to be a family of expander graphs if the following conditions are satisfied:

(i) One has Card(Sj_j)→+∞\rightarrow +\infty with respect to the filter of complements of finite subsets of J;

(ii) There exists C⩾0C\geqslant 0 such that for every jj and every x∈Sjx\in S_j, the valency of GjG_j at xx is ⩽C\leqslant C;

(iii) There exists δ>0\delta >0 such that hGj⩾δh_{G_j}\geqslant \delta for every j∈Jj\in J.

Now prove that (Gj)(G_j) is a family of expander graphs if and only if (i), (ii) are satisfied, as well as the condition

(iii’) There exists δ>0\delta >0 such that λ1(Gj)⩾δ\lambda_1(G_j)\geqslant \delta for every j∈Jj\in J.

Solution

No solution has been written yet.

Cited by

Nothing.

Cites

Nothing.

Machine translation of the English, by gpt-5.4. Not checked by a person.