Homogenization limits and noncommuting scales

For a fixed microscopic period ε, a consistent numerical scheme may converge as H tends to zero. Separately, the continuum solution may homogenize as ε tends to zero. Combining these facts without a uniform numerical error estimate is invalid: a chosen path H(ε) may never resolve the coefficient sufficiently for the fixed-ε estimate to apply. The relevant quantity can be H/ε, an oversampling radius, or a quadrature resolution distinct from the coarse mesh. State which discretization and which norm are controlled before interchanging limits.

This issue survives even for smooth periodic coefficients with fixed ellipticity bounds. Point sampling can lock onto the same microscopic phase in every element while both H and ε shrink. The sampled operator then approximates a different material. A correct effective method may intentionally avoid resolving every microscopic cell, but it must replace unresolved sampling with a justified cell solve, averaging procedure, or asymptotic argument. Small mesh size in physical units alone does not certify consistency relative to a still smaller period.

Consider the unit interval with zero forcing and endpoint values zero and one. A constant sampled conductivity produces the same affine potential x whether its value is one, three, or √3. Potentials alone therefore fail to distinguish the incorrect sampled limits in this example. The flux and energy do distinguish them: constant conductivity c gives flux c and energy c/2. The continuum periodic flux tends to √3, whereas phase-locked midpoint sampling can produce one or three. Choosing the observable is essential to a valid counterexample.

The exact continuum resistance integral can be evaluated by the tangent half-angle substitution, giving the mean resistance 1/√3. Equivalently, one may integrate resistance in local harmonic basis functions. Either approach captures the series response that coefficient point sampling misses. Even a path with ε/H tending to zero leaves this aliasing mechanism intact when quadrature remains phase locked. That path is useful for representative cell averaging only when the averaging itself is performed accurately; a favorable ratio is not a substitute for a consistent algorithm.

For a parabolic coefficient A(x/ε,t/η), a formal first-order expansion u₀+εχ(x/ε,t/η)·∇u₀ has a fast time derivative of size ε/η and a leading spatial divergence of size 1/ε. They balance when η=ε². In a smooth periodic space-time setting, the corresponding formal cell equation is ∂sχp−divy[A(y,s)(p+∇yχp)]=0, with space-time periodicity and a normalization. This calculation identifies a distinguished regime; a rigorous convergence theorem must specify the scale relation, initial data, and parabolic assumptions. It does not prove that every other scale relation has the same cell problem.

A simultaneous high-contrast limit raises a separate concern. If the lower ellipticity bound tends to zero with ε, the uniform H¹ estimate used throughout this module can fail. Additional fields, memory effects, or other limits may then become relevant, but none should be asserted from the uniformly elliptic theory alone. The practical habit is to list every varying parameter, record which constants depend on it, and determine whether the planned limiting path keeps the proof valid. Joint scaling is part of the model, not an afterthought to a qualitative limit.