The Uniform Bound

Numerical solvers for stochastic partial differential equations are typically evaluated by their convergence rate — how quickly the error shrinks as the time step decreases. But the standard convergence analysis is asymptotic: it says the error is O(Δt^p) as Δt → 0. It doesn’t say the error stays bounded as time → ∞.

For equations with long-lived dynamics — phase separation, pattern formation, metastable states — the long-time behavior is the point. A solver that converges for each finite time horizon but accumulates error over time misses the dynamics it was designed to capture.

This paper establishes non-asymptotic, uniform-in-time error bounds for SPDE solvers. The bounds guarantee that the numerical scheme captures both transient and long-term dynamics faithfully, with explicit error control that doesn’t degrade as time advances.

The need is concrete. For the stochastic Allen-Cahn equation — a canonical model of phase separation with non-globally Lipschitz nonlinearity — the classic semi-implicit Euler scheme can exhibit finite-time blow-up. The standard solver, applied to a standard equation, fails catastrophically. Tamed schemes and fully implicit schemes avoid the blow-up, and the authors prove uniform-in-time bounds for three such alternatives.

The analysis covers full space-time discretization, not just time-stepping. This matters because spatial discretization introduces its own errors that can interact with temporal ones in ways that standard separated analysis misses.

An error bound that degrades over time is a promise with an expiration date. A uniform bound is a guarantee.