Non-Divergence Form Difference Operators

• Non-divergence form difference operators are discrete analogues of elliptic operators that use directional second differences to capture key homogenization properties.
• They play a central role in addressing homogenization problems in random settings, achieving improved convergence rates by exploiting structural and reflection symmetries.
• The analysis employs a two-scale expansion with sharp error estimates, yielding convergence rates of O(R⁻³ᐟ²) in d=3 and O(R⁻² log R) in higher dimensions.

A non-divergence form difference operator is a discrete analogue of continuous non-divergence form elliptic operators, with a distinctive structure that separates it from divergence-form operators commonly encountered in the analysis of random walks and discrete PDEs. Non-divergence form difference operators are central in the study of homogenization problems on $\mathbb Z^d$, especially in the context of random environments where statistical invariance and ergodicity play crucial roles. Recent advances overcome natural homogenization rate barriers by exploiting structural symmetries present in i.i.d. settings.

1. Definition and Structure of Non-Divergence Form Difference Operators

Let $$U = \{ e \in \mathbb Z^d : |e| = 1 \} = \{ \pm e_1, \dots, \pm e_d \}$$ denote the set of standard nearest-neighbor directions. Let $$\omega : \mathbb Z^d \to \mathrm{Diag}(>0) \subset \mathbb R^{d \times d}$$ be a random environment assigning positive weights to each coordinate direction at each site. Define, for a function $u: \mathbb Z^d \rightarrow \mathbb R$:

• The forward difference in direction $e$: $\nabla_e u(x) = u(x+e) - u(x)$;
• The second difference in coordinate $i$: $\nabla_i^2 u(x) = u(x+e_i) + u(x-e_i) - 2u(x)$.

Define

$$a_i(x) = \frac{\omega_i(x)}{\sum_{k=1}^d \omega_k(x)}, \qquad a(x) = [a_1(x), \dots, a_d(x)].$$

Then, the non-divergence form difference operator (often denoted $L_\omega$) acts as

$$L_\omega u(x) = \sum_{y : |y-x|=1} \omega(x,y) [u(y) - u(x)] = \frac12 [a(x) \cdot \nabla^2 u(x)] = \sum_{i=1}^d a_i(x)\, \frac12 \nabla_i^2 u(x).$$

The environment is called balanced when the sum of drifts cancels: $\sum_{e \in U} e \omega(x,x+e) = 0$.

2. Dirichlet Problem and Homogenized Limit

For a given $R \gg 1$, let $B_R = \{ x \in \mathbb Z^d : |x| < R \}$ be the discrete ball of radius $R$ and $\mathring{B}_R$ its interior. Given smooth functions $f,g$ on the continuum unit ball $B_1$ and a bounded function $\psi(\omega)$ depending on the local environment at the origin, the Dirichlet problem reads: $$\text{(DP)}\quad \begin{cases} L_\omega u_R(x) = \frac{1}{R^2} f(x/R) \psi(\tau_x \omega), & x \in \mathring{B}_R, \ u_R(x) = g(x/R), & x \in \partial B_R. \end{cases}$$ As $R \rightarrow \infty$, the discrete solution $u_R(x)$ converges to $\bar{u}(x/R)$, where $\bar{u}$ solves the continuum homogenized equation: $$\text{(H)}\quad \begin{cases} \frac12 [\bar{a} D^2 \bar{u}(x)] = f(x)\, \bar{\psi}, & x \in B_1,\ \bar{u}(x) = g(x), & x \in \partial B_1, \end{cases}$$ with effective coefficients $\bar{a} = E_{\mathbb{Q}}[a]$, $\bar{\psi} = E_{\mathbb{Q}}[\psi]$ under the invariant measure $\mathbb{Q}$ (Guo et al., 4 Dec 2025).

3. Quantitative Homogenization in i.i.d., Balanced Environments

Consider the following hypotheses:

• (A1) The family $\{\omega(x)\}_{x \in \mathbb Z^d}$ is i.i.d.
• (A2) Uniform ellipticity: $a(x) \ge 2\kappa I$ for some $\kappa > 0$.
• (A3) $\psi$ is a bounded local function of $\omega(0)$.

For $d \ge 3$, there exists a choice of boundary extension such that $u_R$ and $\bar{u}$ satisfy, for a random prefactor $H = H(\omega)$ with stretched-exponential integrability,

$$\sup_{x\in B_R} |u_R(x) - \bar{u}(x/R)| \lesssim \begin{cases} R^{-3/2} \| \bar{u} \|_{C^6(\overline{B}_1)} H, & d=3, \ R^{-2} \log R\, \| \bar{u} \|_{C^6(\overline{B}_1)} H, & d \ge 4. \end{cases}$$

The convergence rates are thus $O(R^{-3/2})$ in $d=3$, and $O(R^{-2}\log R)$ in $d\ge 4$. This substantially improves upon the $O(R^{-1})$ rate, which is optimal under finite-range dependence or general ergodic environments (Guo et al., 4 Dec 2025).

4. Mechanisms for Improved Convergence Rates

In general, the error in finite-range dependent or ergodic media is $O(R^{-1})$, dictated by discretization effects or non-zero third-order homogenized tensors in the two-scale expansion. For i.i.d. environments, an additional reflection symmetry (the law of $\omega(\cdot)$ equals that of $\omega(-\cdot)$) ensures all third-order homogenized tensors $\overline\lambda_j^k$ and $\overline\eta_j$ vanish via a change-of-variables argument. In the formal two-scale expansion, the $O(R^{-1})$ corrector vanishes, making the leading error contribution come from higher-order terms, thus yielding the aforementioned improved rates (Guo et al., 4 Dec 2025).

5. Proof Strategy and Technical Innovations

The argument proceeds via a second-order two-scale expansion:

• First-order correctors $v^k$ solve $L_\omega v^k = a_k - \bar{a}_k$ and $\xi$ solves $L_\omega \xi = \psi - \bar{\psi}$.
• The expansion takes the form:

$$u_R(x) \approx \bar{u}(x/R) + \frac{1}{R} \sum_k v^k(x)\, \partial_{kk} \bar{u}(x/R) - \frac{1}{R^2} \xi(x) f(x/R) + \cdots$$

• Higher-order correctors $p_j^k$, $s_j$ address the remaining errors, with $$L_\omega p_j^k = \lambda_j^k - \overline{\lambda}_j^k$$, $L_\omega s_j = \eta_j - \overline\eta_j$, where the sources are nonlocal.
• Reflection symmetry enforces $\overline\lambda_j^k = \overline\eta_j = 0$, causing the $O(R^{-1})$ term to vanish.

Sharp control of $p_j^k$ and its gradients is nontrivial due to the highly nonlocal nature of their source terms:

• Represent $$p_j^k(x) = - \sum_{y \in \mathbb Z^d} G_R(x, y) [\lambda_j^k(y) - \overline\lambda_j^k]$$ using a localized Green's function $G_R(x, y)$ of $L_\omega$ with compact cut-off.
• Combine moment bounds and stationarity of correctors, large-scale Hölder/$C^{1,1}$-estimates for $G_R$, and an Efron–Stein sensitivity argument.

The resulting bounds are: $$\|p_j^k\|_{L^\infty(B_R)} \lesssim_H \begin{cases} R^{3/2}, & d=3, \ (\log R)^{1/2}, & d=4, \ 1, & d \ge 5, \end{cases} \qquad \|\nabla p_j^k\|_{L^\infty(B_R)} \lesssim_H \begin{cases} R^{1/2}, & d=3, \ \log R, & d \ge 4. \end{cases}$$ These bounds, substituted into the error analysis from the two-scale expansion, yield the stated convergence rates (Guo et al., 4 Dec 2025).

6. Optimality, Dimensional Dependence, and Related Results

In $d=2$, the first-order corrector exhibits superlinear growth, forcing an $O(R^{-1})$ error without improvement. In periodic one-degree-of-freedom media, an $O(R^{-2})$ rate has been observed, but in typical periodic environments non-vanishing third-order tensors cap the rate at $O(R^{-1})$. The reflection symmetry intrinsic to the i.i.d. law is essential for nullifying the third-order tensor; its absence in correlated or finite-range-dependent contexts precludes any improvement beyond $O(R^{-1})$.

These conclusions delineate the circumstances under which optimal and sub-optimal homogenization rates appear for non-divergence form difference operators in random environments, especially highlighting the role of environmental symmetries and the interaction between discrete and continuum scales (Guo et al., 4 Dec 2025).