Stochastic homogenization and ergodicity

Let a probability space carry a measurable measure-preserving action τx of the translation group. A stationary coefficient field is represented as a(x,ω)=a(0,τxω), with the appropriate jointly measurable realization. Assume a is symmetric and satisfies fixed positive lower and upper ellipticity bounds almost surely, uniformly in location. Ergodicity means every event invariant under all translations has probability zero or one. It does not mean that coefficients at separated locations are independent, nor does it provide a numerical speed for spatial averaging.

For this stationary ergodic uniformly elliptic scalar divergence-form setting, a deterministic homogenized matrix governs almost every realization. On a fixed bounded Lipschitz physical domain, the zero-Dirichlet solutions for a(x/ε,ω) and fixed H⁻¹ load converge weakly in H¹ and strongly in L² to the homogenized solution. Without ergodicity, effective coefficients can remain random and depend on the invariant information in the environment. Qualitative convergence alone supplies neither a central limit theorem nor a universal rate in ε.

A stochastic corrector solves −div a(p+∇φp)=0 in the whole space in the weak sense. Its gradient is stationary, has mean zero, and has finite second moment. The potential is defined only up to an additive constant. In dimensions where point evaluation is unavailable for H¹ functions, normalization should use a spatial average rather than a claimed value at the origin. The potential itself need not be a stationary square-integrable random field; treating that stronger property as automatic loses important low-dimensional behavior.

The useful large-scale property is sublinearity after subtracting a spatial mean: the mean-square oscillation of φp on a large ball, divided by the square of its radius, tends to zero almost surely. It ensures the rescaled potential correction becomes negligible at the macroscopic scale while its gradient still modifies the flux. Constructing this object involves the stationary potential-field space and ergodic arguments. The formula is not obtained by imposing periodic boundary conditions on one arbitrary random sample and declaring that finite sample exact.

In one dimension, a fixed flux produces local slope proportional to resistance r=1/a. For a stationary ergodic field, the spatial mean of resistance converges almost surely to its expectation, yielding the reciprocal of that expectation as the effective conductivity. If the environment has invariant random information, the spatial limit instead involves the corresponding conditional expectation. The arithmetic expectation of conductivity is generally the wrong series response, even when the random law has a very simple description.

For independent cell values on an integer lattice, stationarity initially refers to integer shifts. Adding an independent uniformly distributed origin phase in a unit cell produces a field stationary under all real translations; its translation action is ergodic when the underlying cell process is ergodic. This detail matters when applying a theorem formulated for continuous translations. Random finite-cell approximations can be useful computationally, but representativeness, boundary effects, and ensemble uncertainty must be analyzed separately from the existence of an infinite-volume effective coefficient.