Disorder, entropy and harmonic functions (Q888529): Difference between revisions

The paper under review studies harmonic functions in random environments with particular emphasis on the case of the infinite cluster of supercritical percolation on \(Z^d\). This is related to the early work of \textit{S.-t. Yau} [Commun. Pure Appl. Math. 28, 201--228 (1975; Zbl 0291.31002)] on the Liouville property for positive harmonic functions on complete manifolds with nonnegative Ricci curvature. Yau conjectured that the space of polynomial growth harmonic functions of fixed order is always finite dimensional in open manifolds with nonnegative Ricci curvature. \textit{T. H. Colding} and \textit{W. P. Minicozzi II} proved this conjecture in [Ann. Math. (2) 146, No. 3, 725--747 (1997; Zbl 0928.53030)]. The paper under review concerns the case of random graphs, to which classical geometric analysis cannot be applied directly, and the random environment is not regular at the microscopic scale. Theorem 1 of the paper states that the infinite cluster \(\omega\) has no nonconstant sublinear harmonic functions with probability one for \(d\geq 2\) and \(p>p_c(d)\), where \(p_c(d) \in (0, 1)\) is the critical probability at which a unique cluster splits from no cluster almost surely. This implies that the conjecture of \textit{N. Berger} and \textit{M. Biskup} [Probab. Theory Relat. Fields 137, No. 1--2, 83--120 (2007; Zbl 1107.60066)] that the corrector is unique, is true. The proof of the main result follows by adapting \textit{A. Avez}'s entropy argument [in: Differ. Geom. Relativ., Vol. Honour A. Lichnerowicz 60th Birthday, 27--32 (1976; Zbl 0345.31004)] that showed that a Cayley graph satisfies the Liouville property. The authors define an environment as a random rooted Markov chain, and a random stationary graph as a particular type of reversible Markov chain. A function \(h: V\to\mathbb R\) is harmonic if \(h(X_n)\) is a martingale, i.e., \(h(x) \sum_y P(x, y) h(y)\) for all \(x\). In [J. Am. Math. Soc. 23, No. 3, 815--829 (2010; Zbl 1246.20038)], \textit{B. Kleiner} proved the Gromov theorem that any group has a nontrival linearly growing harmonic function, and that on groups with polynomial growth, the dimension of the space of polynomially growing harmonic functions is finite. \textit{Y. Shalom} and \textit{T. Tao} [Geom. Funct. Anal. 20, No. 6, 1502--1547 (2010; Zbl 1262.20044)] gave a quantitative version of Kleiner's proof of the Gromov theorem. The main result of the paper, Theorem 3, states that if a stationary environment \((P, V, \rho)\) satisfies (DB) (diffusive or subdiffusive behavior, \(E(d(\rho, X_n)^2)\leq Cn\) for every \(n\)), then there are no nonconstant sublinear harmonic functions for almost every environment. Let \((G, \nu, \rho)\) be a rooted weighted graph. If there exists \(0< C_{VD}<\infty\) such that for every \(\lambda <\infty\) there exists \(n_0\in N\) such that for all \(n> n_0\) and every \(x\in B_{\rho}(\lambda n)\) \[ \nu (B_x(2n))\leq C_{VD}\nu (B_x(n), \] then \((G, \nu, \rho)\) satisfies the \textit{anchored volume doubling property} \((VD)_G\), where \(\nu (B)\) is the total weight of the edges in the ball \(B\). If for every \(f: B_x(2n)\to\mathbb R\) \[ \sum_{y\in B_x(n)} (f(y) - \overline{f}_{B_x(n)})^2 \nu (y) \leq C_P n^2 \sum_{y, z\in E(B_x(2n))} |f(y)- f(z)|^2 \nu (y, z), \] then \((G, \nu, \rho)\) satisfies the \textit{anchored Poincaré inequality} \((P)_G\), where \[ \overline{f}_{B_x(n)}= \frac{1}{\nu (B_x(n))} \sum_{y\in B_x(n)} f(y)\nu (y). \] Theorem 4 of the paper states that for every \(k>0\), the space of harmonic functions with \(|h(x)|\leq C d(\rho , x)^k\) for all \(x\) far enough from \(\rho\), is finite-dimensional for rooted weighted graphs with \((VD)_G\) and \((P)_G\). This is exactly the discrete anchored version of Yau's conjecture except for the assumption of the anchored Poincaré inequality (this is a consequence of manifolds with nonnegative Ricci curvature). For random graphs, the dimension in Theorem 4 depends only on \(C_{VD}\) and \(C_P\), which are not bounded if these constants are not random. In supercritical percolation, there is a constant depending only on the dimension \(d\) and the probability \(p\) such that \(C_{VD} \leq A\) and \(C_P\leq A\) almost surely. Theorem 5 deals with the unique infinite component of percolation on \(Z^d\) with \(d\geq 2\) and \(p> p_c(d)\), where the dimension of the vector space of harmonic functions with at most linear growth on the component is equal to \(d+1\) almost surely. Theorem 6 states that the heat kernel estimate for the infinite cluster of percolation \(\omega\) on \(Z^d\) holds by the entropy techniques developed in the proof of Theorem 1 with bounds obtaind for continuous-time random walks for the infinite cluster of supercritical percolation. Section 2 recalls the entropy argument of Avez [loc. cit.], and the entropy of the random walk at times \(n\), \(m\) started at \(\rho\) conditionally on \((P, V, \rho)\) is given by \[ H_{n, m}(P, V, \rho) = H(X_n, X_m) = \sum_{x, y\in V}\phi (P_{\rho}(X_n = x, X_m = y)). \] When \(n=m\), simply denote \(H_{n, n}(P, V, \rho)=H_n(P, V, \rho)\). Theorem 8 states that for every \(n>0\), \(E(\Delta_n (\rho, X_1)^2) \leq 2 (H_n - H_{n-1})\) for stationary environments. The proof of Theorem 8 follows from two lemmas. The proof of Theorem 3 is reduced to the identity \(h(\rho)=h(X_1)\) a.s., for any sublinear harmonic function \(h\) and stationarity implies \(h(X_n)= h(X_{n+1})\). By the assumption, the Markov chain has annealed polynomial growth. The entropy is at most logarithmic, by Fatou's lemma and sublinearity of \(h\), for almost every environment and every harmonic and sublinear \(h\) on it, \(E_{\rho}(|h(\rho) - h(X_1)|)\leq c_2 \varepsilon^{1/2}\). If \(H_n/n\) converges to 0, then \(P\) has the Liouville property (has no nonconstant bounded harmonic functions almost surely), as the Liouville property follows from Theorem 3. Then several examples (Random conductance with uniform ellipticity, infinite cluster of percolation, centered random environments, balanced random environments, random environments with cut points, Poisson point process, graphical fractals, critical Galton-Watson trees, infinite incipient cluster, graph limits and UPQ) are presented at the end of Section 2. Section 3 is devoted to prove Theorem 4 based on the observation that macroscopic Poincaré inequality and volume growth estimates are sufficient. The authors follow the lines of Shalom and Tao [loc. cit.] and are inspired by an elegant proof of Kleiner [loc. cit.] utilizing spaces of harmonic functions with polynomial growth in a crucial way. They first prove a very general inequality (the reverse Poincaré inequality) in Proposition 12 for any graph. With Lemmas 13 and 14 as preparations, the proof of Theorem 4 follows from a proper covering of balls \(B_{\rho}(n)\) and compares two Gram determinants on vectors for \(n\) large enough to see the linearly independent family \(u_i\) of harmonic functions (and hence the finite dimensional property). Section 4 studies linearly growing harmonic functions on the infinite cluster of percolation. For almost every environment, the sequence \((h_n)|_K=h(nx)/n|_K\) of any harmonic function with linear growth is uniformly bounded and equicontinuous for every compact \(K\subset\mathbb R^d\) (Proposition 19), and is linear with Gaussian estimates and Lévy-Prokhorov distance (Lemma 20). The authors provide the proof of Theorem 5 by first applying Theorem 8 and Proposition 19 to extract the subsequence, and then repeat the argument of Theorem 3 to conclude the result. Section 5 gives the proof of Theorem 6, i.e., an upper bound for the discrete derivative of the heat kernel \(p_n(x, y)-p_n(x^{'}, y)\) for \(x\sim x^{'}\). Lemma 21 gives the \(L^2\)-type estimate for \(p_{2n}(x, y)-p_{2n-1}(x^{'}, y)\) for any graph. For supercritical percolation one has \(H_n - H_{n-1} \leq C/n\) (Lemma 22). The proof of Theorem 6 is given by the estimate of the heat kernel derivative and by applying Theorem 8 to bound \(E[\Delta_n^2]\) by \(2(H_n-H_{n-1})\), and using the Lemmas 21 and 22 to derive the conclusion. It is an interesting question whether (DB) follows from polynomial growth in the reversible case. Section 6 lists open questions of this paper. For instance, whether there is a critical minimal growth harmonic function such that the Liouville property does not hold. Question 1 asks if there exist nonconstant harmonic functions with polynomial growth on the UIPQ (uniform infinite planar quadrangulation); if so, what is the minimal growth (Question 2); can the space of harmonic functions with some prescribed polynomial growth on the UIPQ still be of finite dimension (Question 3)? Question 4 extends the similarity between \(\mathbb Z^d\) and the infinite cluster of percolation to their dimensions of spaces of harmonic functions with arbitrary polynomial growth. Question 5 asks if the dimensions of spaces of harmonic functions with a given growth are equal for \(G\) and \(\omega (G)\) (\textit{I. Benjamini} et al. studied this for non-amenable Cayley graphs [in: Random walks and discrete potential theory. Cortona 1997. Proceedings of the conference, Cortona, Italy, June 1997. Cambridge: Cambridge University Press; Roma: Istituto Nazionale di Alta Matematica Francesco Severi. 56--84 (1999; Zbl 0958.05121)]). Question 6 asks if the same holds for a random subgraph of \(Z^d\).