Cheeger-harmonic functions in metric measure spaces revisited (Q2452452)

Following \textit{J. Cheeger}'s Rademacher-style differentiation theorem for Lipschitz functions on metric measure spaces [Geom. Funct. Anal. 9, No. 3, 428--517 (1999; Zbl 0942.58018)], the theory of second-order PDEs in divergence form can be extended to a wide class of abstract metric spaces. In particular, one can introduce a Laplacian \(L\) acting on a suitable Sobolev space \(W^{1,2}\) via the usual weak formulation. In this paper, solutions to \(Lu=0\) are termed `Cheeger-harmonic' functions. The main topic in this paper is the regularity of solutions to nonlinear equations of Poisson type involving the Cheeger Laplacian. As a consequence, existing results on the Lipschitz continuity of Cheeger-harmonic functions are improved. The relevant assumptions on the metric measure space \(X = (X,d,\mu)\) are conditions (i) through (iv) below. {\parindent=8mm \begin{itemize}\item[(i)] \(X\) is locally \(Q\)-doubling for some \(Q>1\), i.e., the inequality \[ 0 < \mu(B(x,R)) \leq C_1 (R/r)^Q \mu(B(x,r)) \] holds for every \(x \in X\) and \(0<r<R\leq R_0\), where \(R_0>0\) is fixed and \(C_1=C_1(R_0)\). (In other references this condition is sometimes known as local \(Q\)-homogeneity of the measure \(\mu\).) \item[(ii)] \(X\) satisfies the local \(L^2\)-Poincaré inequality \[ \frac1{\mu(B)} \int_{B} |u-u_{B}| \, d\mu \leq C_2 r \left( \frac1{\mu(2B)} \int_{2B} (\text{Lip } u)^2 \, d\mu \right)^{1/2}, \] for every ball \(B=B(x,r)\) with \(r<R_0\), where \(R_0>0\) is again fixed, \(2B=B(x,2r)\), \(u_A=\mu(A)^{-1}\int_A u \, d\mu\) denotes the \(\mu\)-average of \(u\) over a set \(A \subset X\) of positive measure, \[ (\text{Lip } u)(x) = \limsup_{r\to 0} r^{-1} \sup\big\{ |u(x)-u(y)| \, : \, d(x,y) \leq r \big\} \] denotes the upper pointwise Lipschitz constant of \(u\), and \(C_2=C_2(R_0)\). \end{itemize}} By Cheeger's results, these two assumptions suffice to guarantee that each Lipschitz function \(u:X\to{\mathbb R}\) is differentiable almost everywhere. Moreover, the Cheeger derivative \(u \mapsto Du\) can also be defined for functions in the Sobolev space \(W^{1,2}(X)\), i.e., the completion of locally Lipschitz functions under the norm \(\|u\|_{W^{1,2}} = \|u\|_{L^2} + \|\text{Lip } u\|_{L^2}\). Introduce the Dirichlet form \({\mathcal E}(u,v) = \int_X Du(x) \cdot Dv(x) \, d\mu(x)\). {\parindent=8mm \begin{itemize}\item[(iii)] The heat kernel \(p(t,x,y)\) for the semigroup \(({\mathbb T}_t)_{t\geq 0}\) associated to \({\mathcal E}\) is stochastically complete, i.e., \[ {\mathbb T}_t(x) = \int_X p(t,x,y) \, d\mu(y) = 1 \] for every \(x \in X\) and \(t>0\). \item[(iv)] \(X\) satisfies the heat semigroup curvature condition \[ {\mathbb T}_t(u^2)(x) - ({\mathbb T}_t(u)(x))^2 \leq (2t+c_\kappa(T)t^2) {\mathbb T}_t(|Du|^2) \] for each \(0<t<T\) and each \(u \in W^{1,2}(X)\), where \(c_\kappa:(0,\infty)\to[0,\infty)\) is a fixed nondecreasing function. \end{itemize}} Under the above four assumptions, Theorem 1.1 of the paper provides \(L^\infty\) estimates for \(|Du|\) for solutions to the equation \(Lu=-\lambda u\) on a ball \(B \subset X\) for \(L^\infty\) data \(\lambda\). In particular, this gives local Lipschitz regularity for Cheeger harmonic functions. This was previously known [\textit{P. Koskela} et al., J. Funct. Anal. 202, No. 1, 147--173 (2003; Zbl 1027.31006)] with the local \(Q\)-doubling assumption replaced by the stronger two-sided assumption of Ahlfors \(Q\)-regularity. Under the same four assumptions, Theorem 1.2 of the paper provides Yau-type gradient estimates for positive Cheeger-harmonic functions, i.e., estimates for \(|Du(x)|/u(x)\). A rich class of examples of spaces satisfying these assumptions are the metric measure spaces satisfying the \(RCD^*(K,N)\) hypothesis; see recent preprints of Ambrosio-Mondino-Savaré or Erbar-Kuwada-Sturm. This hypothesis, which combines the Lott-Villani-Sturm curvature dimension estimate \(CD(K,N)\) with the Ambrosio-Gigli-Savaré Riemannian Ricci curvature estimate \(RCD(K,\infty)\), implies all four of the preceding hypotheses by results of Sturm and T. Rajala and an argument due to Bakry. The proofs of Theorems 1.1 and 1.2 use Cacciopoli estimates for heat solutions, Moser's iteration theorem and comparison principles for Cheeger superharmonic functions.