← All blueprints

Spectral Graph Theory

11 11. Sobolev inequalities

In spectral geometry, Sobolev inequalities are among the main tools for controlling eigenfunctions. In this chapter we derive discrete versions of Sobolev inequalities for arbitrary graphs, based on a single combinatorial invariant, the isoperimetric dimension of a graph, and we use them to bound the eigenvalues of the normalized Laplacian \(\mathcal{L}\). Throughout, \(\operatorname {vol} X=\sum _{v\in X}d_v\) for \(X\subseteq V(G)\) and \(\operatorname {vol}G=\sum _{v}d_v\).

Lemma 382 Hölder’s inequality
#

For nonnegative reals \(f_i,g_i\) and exponents \(p,q{\gt}0\) with \(\frac1p+\frac1q=1\),

\[ \sum _i f_i g_i \le \Big(\sum _i f_i^{\, p}\Big)^{1/p}\Big(\sum _i g_i^{\, q}\Big)^{1/q}. \]

The case \(p=q=2\) is the Cauchy–Schwarz inequality.

Lemma 383 Superadditivity of \(x\mapsto x^{s}\) for \(s\le 1\)
#

For \(a,b\ge 0\) and \(0{\lt}s\le 1\) one has \(a^{s}+b^{s}\ge (a+b)^{s}\), and more generally \(\sum _i a_i^{s}\ge \big(\sum _i a_i\big)^{s}\).

Definition 384 Isoperimetric dimension, §11.1

A graph \(G\) has isoperimetric dimension \(\delta \) with isoperimetric constant \(c_\delta \) if for every subset \(X\subseteq V(G)\) the number of edges between \(X\) and its complement \(\bar X\), denoted \(|E(X,\bar X)|\), satisfies

\begin{equation} \label{eq:isoperimetric} |E(X,\bar X)|\ \ge \ c_\delta \, (\operatorname {vol}X)^{\frac{\delta -1}{\delta }},\qquad \text{whenever }\operatorname {vol}X\le \operatorname {vol}\bar X, \end{equation}
1

where \(c_\delta \) depends only on \(\delta \). Equivalently, the Sobolev constant \(c_\delta \) is

\[ c_\delta =\inf _{X:\, \operatorname {vol}X\le \operatorname {vol}\bar X}\ \frac{|E(X,\bar X)|}{(\operatorname {vol}X)^{\frac{\delta -1}{\delta }}}. \]
Definition 385 Sobolev constants \(s_{p,q}\), §11.1

For a function \(f:V(G)\to \mathbb {R}\) write \(\| \nabla f\| _p=\big(\sum _{u\sim v}|f(u)-f(v)|^{p}\big)^{1/p}\) and \(\| f\| _q=\big(\sum _{v}|f(v)|^{q}d_v\big)^{1/q}\). For \(p,q{\gt}0\) the Sobolev constant of \(G\) is

\[ s_{p,q}=\inf _{f\neq 0}\ \sup _{c\in \mathbb {R}}\ \frac{\| \nabla f\| _p}{\| f-c\| _q}. \]

The Cheeger constant of \(G\) equals \(h_G=s_{1,1}\), and the isoperimetric constant \(c_\delta \) is the case \(\delta =\infty \) of the Cheeger constant. The two Sobolev inequalities proved below (Corollary 389 and Theorem 391) concern \(s_{1,q}\) and \(s_{2,q}\) for \(q\) depending on \(\delta \).

Definition 386 Polynomial growth rate, §11.2

For a vertex \(v\) and an integer \(r\ge 0\) let \(N_r(v)=\{ u\in V(G):d(u,v)\le r\} \), where \(d(u,v)\) is the graph distance. The graph \(G\) has polynomial growth rate of type \((c,\delta )\) if

\[ \operatorname {vol}N_r(v)\ \ge \ c\, r^{\delta }\qquad \text{for all vertices }v,\ \text{whenever }\operatorname {vol}N_r(v)\le \tfrac 12\operatorname {vol}G. \]

A graph with isoperimetric dimension \(\delta \) and constant \(c_\delta \) has polynomial growth rate of some type \((c',\delta )\), but the converse fails.

Lemma 387 Summation by parts, (11.6)

For reals \(x_0,\dots ,x_{n+1}\) and \(y_0,\dots ,y_{n+1}\),

\[ \sum _{i=0}^{n}(x_{i+1}-x_i)\, y_i\ =\ \sum _{i=1}^{n}x_i\, (y_{i-1}-y_i)\ +\ x_{n+1}y_n-x_0y_0, \]

and the last two terms vanish if \(x_0=y_n=0\) or \(y_0=x_{n+1}=0\).

Proof

Expand the left side as \(\sum _{i=0}^{n}x_{i+1}y_i-\sum _{i=0}^{n}x_i y_i\) and reindex the first sum by \(i\mapsto i-1\) to obtain \(\sum _{i=1}^{n+1}x_i y_{i-1}-\sum _{i=0}^{n}x_i y_i=\sum _{i=1}^{n}x_i(y_{i-1}-y_i)+x_{n+1}y_n-x_0y_0\). The stated vanishing is immediate.

In a connected graph \(G\) with isoperimetric dimension \(\delta {\gt}1\) and isoperimetric constant \(c_\delta \), for an arbitrary function \(f:V(G)\to \mathbb {R}\) let \(m\) denote the largest value such that

\[ \sum _{v:\, f(v){\lt}m}d_v\ \le \ \sum _{u:\, f(u)\ge m}d_u. \]

Then

\[ \sum _{u\sim v}|f(u)-f(v)|\ \ge \ c_1\Big(\sum _{v}|f(v)-m|^{\frac{\delta }{\delta -1}}d_v\Big)^{\frac{\delta -1}{\delta }},\qquad c_1=c_\delta \, \frac{\delta -1}{\delta }. \]
Proof

Label the vertices so that \(f(v_1)\ge f(v_2)\ge \cdots \ge f(v_n)\), and abbreviate \(f(i)=f(v_i)\), \(d_i=d_{v_i}\). For \(1\le i\le n-1\) let \(A_i=\{ \{ v_j,v_k\} \in E(G):j\le i{\lt}k\} \) be the set of edges of the cut \((\{ v_1,\dots ,v_i\} ,\{ v_{i+1},\dots ,v_n\} )\) and put \(a_i=|A_i|\). Set

\[ S_i^{-}=\sum _{j{\lt}i}d_j,\qquad S_i^{+}=\sum _{j\ge i}d_j,\qquad S_i=\min \{ S_i^{-},S_i^{+}\} , \]

so that \(S_i=S_i^{+}\) when \(f(i)\ge m\) and \(S_i=S_i^{-}\) when \(f(i){\lt}m\), and \(S_i\) is the volume of the smaller side of the cut \(A_i\). By 1,

\[ a_i\ \ge \ c_\delta \, S_i^{\frac{\delta -1}{\delta }}. \]

Use the convention \(S_0=S_{n+1}=0\), set \(h(i)=f(i)-m\), and let \(i_0\) satisfy \(f(i_0)=m\), so \(h(i)\ge 0\) for \(i\le i_0\) and \(h(i){\lt}0\) for \(i{\gt}i_0\). Since \(f\) is monotone, applying Lemma 387,

\[ \begin{aligned} \sum _{u\sim v}|h(u)-h(v)| & =\sum _i a_i\, |h(i-1)-h(i)| \ \ge \ c_\delta \sum _i S_i^{\frac{\delta -1}{\delta }}\, |h(i-1)-h(i)|\\ & \ge \ c_\delta \sum _{i\le i_0}|h(i)|\Big(S_{i+1}^{\frac{\delta -1}{\delta }}-S_i^{\frac{\delta -1}{\delta }}\Big) +c_\delta \sum _{i{\gt}i_0}|h(i)|\Big(S_i^{\frac{\delta -1}{\delta }}-S_{i+1}^{\frac{\delta -1}{\delta }}\Big)\\ & =:W_1+W_2. \end{aligned} \]

We treat \(W_1\); \(W_2\) is symmetric. Writing \(S_{i+1}=S_i+d_i\) on this side and using \((S_i+d_i)^{\frac{\delta -1}{\delta }}-S_i^{\frac{\delta -1}{\delta }}\ge \frac{\delta -1}{\delta }\, d_i\, S_i^{-1/\delta }\),

\[ \begin{aligned} W_1& =c_\delta \sum _{i\le i_0}h(i)\Big((S_i+d_i)^{\frac{\delta -1}{\delta }}-S_i^{\frac{\delta -1}{\delta }}\Big) \ \ge \ c_\delta \, \frac{\delta -1}{\delta }\sum _{i\le i_0}\frac{h(i)\, d_i}{S_i^{1/\delta }}\\ & =c_\delta \, \frac{\delta -1}{\delta }\sum _{i\le i_0}\frac{h(i)^{\frac{\delta }{\delta -1}}d_i}{\big(h(i)^{\frac{\delta }{\delta -1}}S_i\big)^{1/\delta }} \ \ge \ c_\delta \, \frac{\delta -1}{\delta }\, \frac{\sum _{i\le i_0}|h(i)|^{\frac{\delta }{\delta -1}}d_i}{\big(\sum _{i\le i_0}|h(i)|^{\frac{\delta }{\delta -1}}d_i\big)^{1/\delta }}\\ & =c_\delta \, \frac{\delta -1}{\delta }\Big(\sum _{i\le i_0}|h(i)|^{\frac{\delta }{\delta -1}}d_i\Big)^{\frac{\delta -1}{\delta }}, \end{aligned} \]

where the middle inequality uses \(\big(h(i)^{\frac{\delta }{\delta -1}}S_i\big)^{1/\delta }\le \big(\sum _{i\le i_0}|h(i)|^{\frac{\delta }{\delta -1}}d_i\big)^{1/\delta }\). Combining with the analogous bound for \(W_2\) and applying Lemma 383 (superadditivity of \(x\mapsto x^{\frac{\delta -1}{\delta }}\)),

\[ W_1+W_2\ \ge \ c_\delta \, \frac{\delta -1}{\delta }\Big(\big(\! \sum _{i\le i_0}\! |h(i)|^{\frac{\delta }{\delta -1}}d_i\big)^{\frac{\delta -1}{\delta }}+\big(\! \sum _{i{\gt} i_0}\! |h(i)|^{\frac{\delta }{\delta -1}}d_i\big)^{\frac{\delta -1}{\delta }}\Big) \ \ge \ c_\delta \, \frac{\delta -1}{\delta }\Big(\sum _i|h(i)|^{\frac{\delta }{\delta -1}}d_i\Big)^{\frac{\delta -1}{\delta }}, \]

which is the claim with \(c_1=c_\delta \frac{\delta -1}{\delta }\).

Corollary 389 Corollary 11.2

In a connected graph \(G\) with isoperimetric dimension \(\delta \) and isoperimetric constant \(c_\delta \), for an arbitrary \(f:V(G)\to \mathbb {R}\),

\[ \sum _{u\sim v}|f(u)-f(v)|\ \ge \ c_1\, \min _{m}\Big(\sum _{v}|f(v)-m|^{\frac{\delta }{\delta -1}}d_v\Big)^{\frac{\delta -1}{\delta }},\qquad c_1=c_\delta \, \frac{\delta -1}{\delta }. \]
Proof

Let \(m_0\) be the value of \(m\) produced in Theorem 388. Since the minimum over \(m\) is at most the value at \(m=m_0\),

\[ \sum _{u\sim v}|f(u)-f(v)|\ \ge \ c_1\Big(\sum _v|f(v)-m_0|^{\frac{\delta }{\delta -1}}d_v\Big)^{\frac{\delta -1}{\delta }}\ \ge \ c_1\min _m\Big(\sum _v|f(v)-m|^{\frac{\delta }{\delta -1}}d_v\Big)^{\frac{\delta -1}{\delta }}.\qedhere \]
Corollary 390 Corollary 11.3

In a connected graph \(G\) with isoperimetric dimension \(\delta \) and isoperimetric constant \(c_\delta \), for \(f:V(G)\to \mathbb {R}\), a vertex \(w\), and \(m\) as in Theorem 388, define the truncated function

\[ f_w(v)=\begin{cases} \max \{ f(v),f(w)\} & \text{if }f(w)\ge m,\\[2pt]\min \{ f(v),f(w)\} & \text{if }f(w){\lt}m,\end{cases} \]

the boundary count

\[ a_w=\begin{cases} \big|\{ \{ u,v\} \in E(G):f(u){\gt}f(w)\ge f(v)\} \big|& \text{if }f(w)\ge m,\\[2pt]\big|\{ \{ u,v\} \in E(G):f(u)\ge f(w){\gt}f(v)\} \big|& \text{if }f(w){\lt}m,\end{cases} \]

and the level set \(S_w=\{ v:f(v)\ge f(w)\} \) if \(f(w)\ge m\), resp. \(S_w=\{ u:f(u)\le f(w)\} \) if \(f(w){\lt}m\). Then

\[ \sum _{u\sim v}|f_w(u)-f_w(v)|+a_w\, |f(w)-m|\ \ge \ c_1\Big(\sum _{v\in S_w}|f(v)-m|^{\frac{\delta }{\delta -1}}d_v\Big)^{\frac{\delta -1}{\delta }},\qquad c_1=c_\delta \, \frac{\delta -1}{\delta }. \]
Proof

This follows from the proof of Theorem 388 applied to the truncated function \(f_w\). In the ordering of that proof, \(f_w\) is constant equal to \(f(w)\) outside \(S_w\), so all cuts \(A_i\) with \(f(i)\) strictly between \(m\) and \(f(w)\) contribute nothing, and for \(f(w)\ge m\)

\[ \sum _{u\sim v}|f_w(u)-f_w(v)|=\sum _{i:\, f(i)\ge f(w)}a_i\, (h(i-1)-h(i))\ \ge \ \sum _{i:\, h(i)\ge h(w)}|h(i)|\, (a_i-a_{i-1})\ -\ a_w|h(w)|, \]

by Lemma 387. Adding \(a_w|f(w)-m|\) and running the \(W_1\)-estimate of Theorem 388 over the vertices of \(S_w\) yields the stated lower bound.

For a graph \(G\) with isoperimetric dimension \(\delta {\gt}2\) and isoperimetric constant \(c_\delta \), any function \(f:V(G)\to \mathbb {R}\) satisfies

\[ \Big(\sum _{u\sim v}|f(u)-f(v)|^{2}\Big)^{1/2}\ \ge \ c_2\, \min _{m}\Big(\sum _{v}|f(v)-m|^{\gamma }d_v\Big)^{1/\gamma }, \]

where \(\gamma =\dfrac {2\delta }{\delta -2}\) and \(c_2=\sqrt{c_\delta }\, \dfrac {\delta -1}{2\delta }\).

Proof

Use the notation of Theorem 388: order \(f(1)\ge \cdots \ge f(n)\), set \(h(i)=f(i)-m\) with \(f(i_0)=m\), and write \(h_i(x)=h_{v_i}(x)\) for the truncation of Corollary 390 at \(w=v_i\). Representing the square of a gradient through its level cuts,

\[ \begin{aligned} \sum _{u\sim v}|h(u)-h(v)|^{2} & =\sum _i (h(i-1)-h(i))\! \! \sum _{\{ j,k\} \in A_i}\! \! |h(j)-h(k)|\\ & \ge \sum _{i\le i_0}(h(i-1)-h(i))\! \! \sum _{\{ j,k\} \in A_i}\! \! |h_i(j)-h_i(k)| +\sum _{i{\gt} i_0}(h(i-1)-h(i))\! \! \sum _{\{ j,k\} \in A_i}\! \! |h_i(j)-h_i(k)|\\ & =:P_1+P_2. \end{aligned} \]

We bound \(P_1\), assuming without loss of generality that all \(h(i){\gt}0\); \(P_2\) is symmetric. Applying summation by parts (Lemma 387) and then Corollary 390 to the inner sums,

\[ \begin{aligned} P_1& \ge \sum _{i\le i_0}h(i)\Big(\sum _{\{ j,k\} \in A_{i+1}}\! \! |h_{i+1}(j)-h_{i+1}(k)|-\sum _{\{ j,k\} \in A_i}\! \! |h_i(j)-h_i(k)|\Big)\\ & \ge \sum _{i\le i_0}h(i)\Big(a_{i+1}(h(i)-h(i+1))-\sum _{\{ j,k\} \in A_i\setminus A_{i+1}}\! \! |h_i(j)-h_i(k)|\Big) \ \ge \ \sum _{i\le i_0}h(i)a_{i+1}(h(i)-h(i+1))-P_1. \end{aligned} \]

Hence \(2P_1\ge \sum _{i\le i_0}h(i)a_{i+1}(h(i)-h(i+1))\). Applying Lemma 387 once more,

\[ \begin{aligned} 2P_1& \ge \sum _{i\le i_0}h(i)a_{i+1}(h(i)-h(i+1)) \ \ge \ \sum _{i\le i_0}h(i)\big(a_{i+1}h(i)-a_i h(i-1)\big)\\ & \ge \sum _{i\le i_0}h(i)\big((a_{i+1}-a_i)h(i)-a_i(h(i-1)-h(i))\big) \ \ge \ \sum _{i\le i_0}h(i)(a_{i+1}-a_i)h(i)-2P_1, \end{aligned} \]

so that

\[ P_1\ \ge \ \frac14\sum _{i\le i_0}h(i)(a_{i+1}-a_i)h(i). \]

Define

\[ T_i=\sum _{j\le i}a_j\big(h(j-1)-h(j)\big)+a_i|h(i)|,\qquad T_i'=\sum _{j{\lt}i}h(j)^{\frac{\delta }{\delta -1}}d_j. \]

By Corollary 390, \(T_i=\sum _{u\sim v}|f_{v_i}(u)-f_{v_i}(v)|+a_i(f(i)-m)\ge c_\delta \frac{\delta -1}{\delta }\, T_i'^{\, \frac{\delta -1}{\delta }}\), and directly from the definition \(T_i-T_{i-1}=(a_i-a_{i-1})h(i-1)\), so \((a_{i+1}-a_i)h(i)=T_{i+1}-T_i\). Using \(T_{i+1}'=T_i'+h(i)^{\frac{\delta }{\delta -1}}d_i\) and \((1+x)^{\frac{\delta -1}{\delta }}-1\ge \frac{\delta -1}{\delta }\, x\, (1+x)^{-1/\delta }\),

\[ \begin{aligned} P_1& \ge \frac14\sum _{i\le i_0}h(i)(T_{i+1}-T_i) \ \ge \ c_\delta \, \frac{\delta -1}{4\delta }\sum _{i\le i_0}h(i)\, T_i’^{\, \frac{\delta -1}{\delta }}\Big(\big(1+\tfrac {h(i)^{\frac{\delta }{\delta -1}}d_i}{T_i’}\big)^{\frac{\delta -1}{\delta }}-1\Big)\\ & \ge c_\delta \, \frac{(\delta -1)^2}{4\delta ^2}\sum _{i\le i_0}h(i)\, T_i’^{\, \frac{\delta -1}{\delta }}\, \frac{h(i)^{\frac{\delta }{\delta -1}}d_i}{T_i'} \ =\ c_\delta \, \frac{(\delta -1)^2}{4\delta ^2}\sum _{i\le i_0}\frac{h(i)^{\frac{2\delta -1}{\delta -1}}d_i}{T_i'^{\, 1/\delta }}\\ & \ge c_\delta \, \frac{(\delta -1)^2}{4\delta ^2}\, \frac{\sum _{i\le i_0}h(i)^{\frac{2\delta }{\delta -1}}d_i}{\big(\sum _{i{\lt}i_0}h(i)^{\frac{\delta }{\delta -1}}d_i\big)^{1/\delta }} \ \ge \ c_\delta \, \frac{(\delta -1)^2}{4\delta ^2}\Big(\sum _{i\le i_0}h(i)^{\gamma }d_i\Big)^{2/\gamma }, \end{aligned} \]

the last two steps using Hölder’s inequality (Lemma 382) and \(T_i'^{\, 1/\delta }\le \big(\sum _{i{\lt}i_0}h(i)^{\frac{\delta }{\delta -1}}d_i\big)^{1/\delta }\), with \(\gamma =\frac{2\delta }{\delta -2}\) so that \(2/\gamma =\frac{\delta -2}{\delta }\). Symmetrically \(P_2\ge c_\delta \frac{(\delta -1)^2}{4\delta ^2}\big(\sum _{i{\gt}i_0}|h(i)|^{\gamma }d_i\big)^{2/\gamma }\). Since \(\gamma \ge 2\), i.e. \(2/\gamma \le 1\), superadditivity (Lemma 383) gives

\[ \sum _{u\sim v}|h(u)-h(v)|^{2}=P_1+P_2\ \ge \ c_\delta \, \frac{(\delta -1)^2}{4\delta ^2}\Big(\sum _i|h(i)|^{\gamma }d_i\Big)^{2/\gamma }. \]

Taking square roots yields the inequality at this \(m\) with \(c_2=\sqrt{c_\delta }\, \frac{\delta -1}{2\delta }\); minimizing the right-hand side over \(m\) gives the stated form.

Assume \(\delta {\gt}2\) and set \(\gamma =\frac{2\delta }{\delta -2}\). Let \(H_t\) be the heat kernel of \(G\), fix a vertex \(x\), and let \(m\) be the value associated (as in Theorem 388) to the function \(y\mapsto H_{1/2}(x,y)/\sqrt{d_y}\). Then

\[ \Big(\sum _{y}\Big|\frac{H_{t/2}(x,y)}{\sqrt{d_y}}-m\Big|^{\gamma }d_y\Big)^{\frac{1}{\gamma -1}}\big(3\sqrt{d_x}\big)^{\frac{\gamma -2}{\gamma -1}}\ \ge \ \sum _{y}\Big(\frac{H_{t/2}(x,y)}{\sqrt{d_y}}-m\Big)^{2}d_y. \]
Proof

Apply Hölder’s inequality (Lemma 382) with \(\frac1p+\frac1q=1\), \(p=\gamma -1\), \(q=\frac{\gamma -1}{\gamma -2}\), and

\[ f_y=\Big|\frac{H_{t/2}(x,y)}{\sqrt{d_y}}-m\Big|^{\frac{2}{\gamma -1}}d_y^{\, \frac{1}{\gamma -1}},\qquad g_y=\Big|\frac{H_{t/2}(x,y)}{\sqrt{d_y}}-m\Big|^{\frac{\gamma -2}{\gamma -1}}d_y^{\, \frac{\gamma -2}{\gamma -1}}, \]

which gives

\[ \Big(\sum _{y}\Big|\tfrac {H_{t/2}(x,y)}{\sqrt{d_y}}-m\Big|^{\gamma }d_y\Big)^{\frac{1}{\gamma -1}}\Big(\sum _{y}\Big|\tfrac {H_{t/2}(x,y)}{\sqrt{d_y}}-m\Big|d_y\Big)^{\frac{\gamma -2}{\gamma -1}}\ \ge \ \sum _{y}\Big(\tfrac {H_{t/2}(x,y)}{\sqrt{d_y}}-m\Big)^{2}d_y. \]

It remains to prove \(\sum _{y}\big|\frac{H_{t/2}(x,y)}{\sqrt{d_y}}-m\big|d_y\le 3\sqrt{d_x}\). Put \(m'=\frac{\sqrt{d_x}}{\operatorname {vol}G}\). By Lemma 365, \(\sum _y H_{t/2}(x,y)\sqrt{d_y}=\sqrt{d_x}\), hence

\begin{equation} \label{eq:heat-centered} \sum _{y}\Big(\frac{H_{t/2}(x,y)}{\sqrt{d_y}}-m'\Big)d_y=\sum _y H_{t/2}(x,y)\sqrt{d_y}-m'\operatorname {vol}G=\sqrt{d_x}-m'\operatorname {vol}G=0. \end{equation}
6

Writing \(N_x^{-}=\{ y:H_{t/2}(x,y)/\sqrt{d_y}{\lt}m'\} \), equation 6 gives \(\sum _y|\frac{H_{t/2}(x,y)}{\sqrt{d_y}}-m'|d_y=2\sum _{y\in N_x^{-}}(m'-\frac{H_{t/2}(x,y)}{\sqrt{d_y}})d_y\le 2\sum _{y\in N_x^{-}}m'd_y=2\frac{\sqrt{d_x}}{\operatorname {vol}G}\sum _{y\in N_x^{-}}d_y\le 2\sqrt{d_x}\). Therefore, provided \(m'\ge m/2{\gt}0\),

\[ \sum _{y}\Big|\tfrac {H_{t/2}(x,y)}{\sqrt{d_y}}-m\Big|d_y\le \sum _y\Big|\tfrac {H_{t/2}(x,y)}{\sqrt{d_y}}-m'\Big|d_y+\sum _y|m'-m|d_y\le 2\sqrt{d_x}+m'\operatorname {vol}G=3\sqrt{d_x}. \]

Finally \(m'\ge m/2\): with \(M_x^{+}=\{ y:H_{1/2}(y,x)/\sqrt{d_y}\ge m\} \) (of volume \(\ge \operatorname {vol}G/2\) by the definition of \(m\)) and \(H\ge 0\),

\[ m'=\frac{1}{\operatorname {vol}G}\sum _y H_{1/2}(x,y)\sqrt{d_y}\ \ge \ \frac{1}{\operatorname {vol}G}\sum _{y\in M_x^{+}}H_{1/2}(x,y)\sqrt{d_y}\ \ge \ \frac{m\operatorname {vol}M_x^{+}}{\operatorname {vol}G}\ \ge \ \frac{m}{2}.\qedhere \]

For a graph \(G\) with isoperimetric dimension \(\delta {\gt}2\) and isoperimetric constant \(c_\delta \), the eigenvalues \(0=\lambda _0\le \lambda _1\le \cdots \le \lambda _{n-1}\) of \(\mathcal{L}\) satisfy

\begin{equation} \label{eq:heat-trace} \sum _{i\neq 0}e^{-\lambda _i t}\ \le \ \frac{c_3\operatorname {vol}G}{t^{\delta /2}}, \end{equation}
7

where \(c_3\) is a constant depending only on \(\delta \).

Proof

Set \(\gamma =\frac{2\delta }{\delta -2}\). By the semigroup property \(H_t(x,y)=\sum _z H_{t/2}(x,z)H_{t/2}(z,y)\) and symmetry, \(H_t(x,x)=\sum _y H_{t/2}(x,y)^2\). Differentiating and using that \(y\mapsto H_{t/2}(x,y)\) solves the heat equation,

\[ \frac{\partial }{\partial t}H_t(x,x)=2\sum _y H_{t/2}(x,y)\frac{\partial }{\partial t}H_{t/2}(x,y)=\sum _y H_{t/2}(x,y)\big(-\mathcal{L}H_{t/2}(y,x)\big)=-\sum _{y\sim z}\Big(\frac{H_{t/2}(y,x)}{\sqrt{d_y}}-\frac{H_{t/2}(z,x)}{\sqrt{d_z}}\Big)^2. \]

Applying Theorem 391 to \(y\mapsto H_{t/2}(y,x)/\sqrt{d_y}\),

\begin{equation} \label{eq:heat-diag-sobolev} \frac{\partial }{\partial t}H_t(x,x)\ \le \ -c_\delta \, \frac{(\delta -1)^2}{4\delta ^2}\Big(\sum _y\Big(\frac{H_{t/2}(y,x)}{\sqrt{d_y}}-m\Big)^{\gamma }d_y\Big)^{2/\gamma }. \end{equation}
8

Insert Lemma 392: since \(\big(\sum _y(\cdot )^{\gamma }d_y\big)^{2/\gamma }=\big[\big(\sum _y(\cdot )^{\gamma }d_y\big)^{\frac{1}{\gamma -1}}\big]^{\frac{2(\gamma -1)}{\gamma }}\ge \big[\sum _y(\frac{H_{t/2}}{\sqrt{d}}-m)^2 d_y\big]^{\frac{2(\gamma -1)}{\gamma }}(3\sqrt{d_x})^{-\frac{2(\gamma -2)}{\gamma }}\), and using \(\sum _y(\frac{H_{t/2}(x,y)}{\sqrt{d_y}}-m)^2 d_y\ge \sum _y(\frac{H_{t/2}(x,y)}{\sqrt{d_y}}-m')^2 d_y=H_t(x,x)-2m'\sqrt{d_x}+m'^2\operatorname {vol}G=H_t(x,x)-\frac{d_x}{\operatorname {vol}G}\) (by 6 and \(m'=\sqrt{d_x}/\operatorname {vol}G\)),

\[ \frac{\partial }{\partial t}H_t(x,x)\ \le \ -c_5\Big(H_t(x,x)-\frac{d_x}{\operatorname {vol}G}\Big)^{\frac{2(\gamma -1)}{\gamma }}\big(3\sqrt{d_x}\big)^{-\frac{2(\gamma -2)}{\gamma }},\qquad c_5=\Big(c_\delta \frac{(\delta -1)^2}{4\delta ^2}\Big)^{\frac{2(\gamma -1)}{\gamma }}. \]

Note \(1-\frac{2(\gamma -1)}{\gamma }=-\frac{2}{\delta }\) and \(\frac{2(\gamma -2)}{\gamma }=\frac{4}{\delta }\). Writing \(u(t)=H_t(x,x)-\frac{d_x}{\operatorname {vol}G}\ge 0\),

\[ \frac{\partial }{\partial t}u(t)^{-2/\delta }=-\frac{2}{\delta }\, u(t)^{-\frac{2(\gamma -1)}{\gamma }}\, \frac{\partial }{\partial t}H_t(x,x)\ \ge \ \frac{2c_5}{\delta }\big(3\sqrt{d_x}\big)^{-\frac{2(\gamma -2)}{\gamma }}. \]

Integrating from \(0\) to \(t\) and using \(u(0)=H_0(x,x)-\frac{d_x}{\operatorname {vol}G}=1-\frac{d_x}{\operatorname {vol}G}\le 1\),

\[ u(t)^{-2/\delta }\ \ge \ c_6\big(3\sqrt{d_x}\big)^{-\frac{2(\gamma -2)}{\gamma }}t,\qquad c_6=\frac{2c_5}{\delta }, \]

so that, since \(\frac{2(\gamma -2)}{\gamma }\cdot \frac{\delta }{2}=2\),

\[ H_t(x,x)-\frac{d_x}{\operatorname {vol}G}\ \le \ c_6^{-\delta /2}\big(3\sqrt{d_x}\big)^{2}\, t^{-\delta /2}\ =\ \frac{c_3\, d_x}{t^{\delta /2}},\qquad c_3=9\, c_6^{-\delta /2}. \]

Summing over \(x\) and using \(\sum _x d_x/\operatorname {vol}G=1\) gives \(\sum _x H_t(x,x)-1\le c_3\operatorname {vol}G/t^{\delta /2}\). Finally, by the spectral decomposition \(\sum _x H_t(x,x)=\sum _x\sum _i e^{-\lambda _i t}\phi _i^2(x)=\sum _i e^{-\lambda _i t}\), whence \(\sum _{i\neq 0}e^{-\lambda _i t}=\sum _i e^{-\lambda _i t}-1\le c_3\operatorname {vol}G/t^{\delta /2}\).

Let \(G\) have isoperimetric dimension \(\delta {\gt}2\) and isoperimetric constant \(c_\delta \). The \(k\)-th eigenvalue \(\lambda _k\) of \(\mathcal{L}\) satisfies

\[ \lambda _k\ \ge \ c_4\Big(\frac{k}{\operatorname {vol}G}\Big)^{2/\delta }, \]

where \(c_4=\frac{\delta }{2e}\, c_3^{-2/\delta }\) depends only on \(\delta \).

Proof

Since the eigenvalues are ordered, for every \(t{\gt}0\) Theorem 393 gives

\[ k\, e^{-\lambda _k t}\ \le \ \sum _{i=1}^{k}e^{-\lambda _i t}\ \le \ \sum _{i\neq 0}e^{-\lambda _i t}\ \le \ \frac{c_3\operatorname {vol}G}{t^{\delta /2}}. \]

Hence \(k\le c_3\operatorname {vol}G\cdot \inf _{t{\gt}0}\frac{e^{\lambda _k t}}{t^{\delta /2}}\). The function \(t\mapsto e^{\lambda _k t}/t^{\delta /2}\) is minimized at \(t=\frac{\delta }{2\lambda _k}\), giving value \(\big(\frac{2\lambda _k e}{\delta }\big)^{\delta /2}\), so \(k\le c_3\operatorname {vol}G\big(\frac{2\lambda _k e}{\delta }\big)^{\delta /2}\). Solving for \(\lambda _k\),

\[ \lambda _k\ \ge \ \frac{\delta }{2e}\Big(\frac{k}{c_3\operatorname {vol}G}\Big)^{2/\delta }\ =\ \frac{\delta }{2e}\, c_3^{-2/\delta }\Big(\frac{k}{\operatorname {vol}G}\Big)^{2/\delta }=c_4\Big(\frac{k}{\operatorname {vol}G}\Big)^{2/\delta }.\qedhere \]
Definition 395 Isoperimetric dimension of a weighted graph, §11.5

Let \(G\) be a weighted undirected graph with edge weights \(w_{u,v}=w_{v,u}\), \(d_v=\sum _u w_{u,v}\), and \(\operatorname {vol}X=\sum _{v\in X}d_v\). We say \(G\) has isoperimetric dimension \(\delta \) with isoperimetric constant \(c_\delta \) if

\[ \sum _{u\in X}\sum _{v\in \bar X}w_{u,v}\ \ge \ c_\delta \, (\operatorname {vol}X)^{1-\frac{1}{\delta }}\qquad \text{for all }X\subseteq V(G)\text{ with }\operatorname {vol}X\le \operatorname {vol}\bar X. \]

Let \(G\) be a weighted undirected graph with edge weights \(w_{u,v}\) having isoperimetric dimension \(\delta \) and isoperimetric constant \(c_\delta \). Let \(S\subseteq V(G)\) and let \(S^{*}\) be the set of edges with at least one endpoint in \(S\). Then any function \(f:S\cup \delta S\to \mathbb {R}\) satisfying either Dirichlet or Neumann boundary conditions satisfies

\[ \sum _{\{ u,v\} \in S^{*}}|f(u)-f(v)|\, w_{u,v}\ \ge \ c_1\, \min _{m}\Big(\sum _{v\in S}|f(v)-m|^{\frac{\delta }{\delta -1}}d_v\Big)^{\frac{\delta -1}{\delta }},\qquad c_1=c_\delta \, \frac{\delta -1}{\delta }. \]

Let \(G\) be a weighted undirected graph with edge weights \(w_{u,v}\) having isoperimetric dimension \(\delta {\gt}2\) and isoperimetric constant \(c_\delta \). Then any function \(f:V(G)\to \mathbb {R}\) satisfying either Dirichlet or Neumann boundary conditions satisfies

\[ \Big(\sum _{\{ u,v\} \in S^{*}}|f(u)-f(v)|^{2}w_{u,v}\Big)^{1/2}\ \ge \ c_2\, \min _{m}\Big(\sum _{v\in S}|f(v)-m|^{\gamma }d_v\Big)^{1/\gamma }, \]

where \(\gamma =\frac{2\delta }{\delta -2}\) and \(c_2=\sqrt{c_\delta }\, \frac{\delta -1}{2\delta }\).

For a weighted undirected graph \(G\) with isoperimetric dimension \(\delta {\gt}2\) and an induced subgraph \(S\), the Dirichlet or Neumann eigenvalues \(\lambda _i\) of \(S\) satisfy

\[ \sum _{i\neq 0}e^{-\lambda _i t}\ \le \ \frac{c_3\operatorname {vol}S}{t^{\delta /2}}, \]

where \(c_3\) is a constant depending only on \(\delta \).

Suppose a weighted undirected graph \(G\) has isoperimetric dimension \(\delta {\gt}2\) and isoperimetric constant \(c_\delta \), and let \(S\) be an induced subgraph. The \(k\)-th Dirichlet or Neumann eigenvalue \(\lambda _k\) of \(S\) satisfies

\[ \lambda _k\ \ge \ c_4\Big(\frac{k}{\operatorname {vol}G}\Big)^{2/\delta }, \]

where \(c_4\) is a constant depending only on \(\delta \).