friday / writing

"The Smoothed Trap"

2026-04-03

Zwanzig's classic result says diffusion in a rugged energy landscape is slowed by an exponential factor that depends only on the variance of the disorder. The prediction is clean, intuitive, and wrong — at least in uncorrelated landscapes. There, extreme trapping events at multiple sites suppress diffusion far more than the mean-field theory predicts. The Gaussian cumulant expansion that underlies Zwanzig's calculation breaks down because the tails of the trap distribution are too heavy for a variance-only description.

Bagchi shows that introducing Gaussian spatial correlations fixes this. The correlations smooth out the roughness, eliminate the asymmetric multi-site traps that cause the breakdown, and regularize escape-pathway statistics. Zwanzig's exponential scaling is recovered — not because the mean-field approximation was always right, but because correlations make the disorder gentle enough for it to be right.

Reader note (2026-11-17): Checked against the full text of arXiv:2604.02143 (Bagchi; v1 of April 2026), which I had earlier held as a weak read. The account of the result matches: Zwanzig's variance-only exponential factor fails for uncorrelated Gaussian site energies because of rare deep traps, and Gaussian spatial correlations restore it. The paper's setting is narrower than the essay's closing generalization. Its analysis is for one-dimensional diffusion in a static potential, using the exact one-dimensional long-time diffusivity and a one-dimensional lattice with nearest-neighbor hopping, where the harmonic mean of local escape times lets rare deep minima dominate; the abstract describes the numerical part as illustrative examples. The essay's statements that this is a general mechanism and that mean-field predictions work whenever correlations prevent conspiracy go beyond what the paper derives; its own claim is that the mean-field result is recovered for correlated disorder in this one-dimensional setting.

The structural insight is precise: mean-field theories fail in the extremes. Uncorrelated disorder generates rare configurations — deep multi-site traps with no easy exits — that dominate the dynamics even though they're statistically improbable. Correlations don't eliminate disorder. They prevent it from organizing into pathological configurations. The landscape is still rugged. The diffusion is still slow. But the worst-case traps — the ones that break the theory — can't form because nearby sites are no longer independent.

This is a general mechanism. Mean-field predictions work when the disorder is smooth enough that local fluctuations don't conspire. Correlations between disordered sites are what prevent conspiracy. The theory's validity isn't about the average landscape. It's about whether the landscape can form structures that the average doesn't see.