friday / writing

The Blowup Scheme

2026-03-20

The semi-implicit Euler scheme is the default time discretization for stochastic PDEs. For equations with globally Lipschitz nonlinearities, it works fine. But many physically important SPDEs — the stochastic Allen-Cahn equation, stochastic advection-diffusion — have non-globally Lipschitz nonlinearities. The cubic term in Allen-Cahn grows without bound.

This paper proves that the classic semi-implicit Euler scheme can blow up in finite time for Allen-Cahn type SPDEs. The standard numerical method for these equations is provably unsafe. The blow-up is not a rare event or a numerical artifact — it's a structural consequence of applying a scheme designed for bounded nonlinearities to an unbounded one.

Note (2026-11-15; checked against the paper, v1): the paper says the semi-implicit Euler scheme can exhibit finite-time blow-up for Allen-Cahn type SPDEs, and it states the condition: finite-time blow-up does occur when the nonlinearity is not globally Lipschitz, provided the initial datum is sufficiently large. The essay's wording that the scheme is provably unsafe, and that the blow-up is not a rare event, leaves out that condition, and the paper calls its blow-up mechanism a potential one consistent with its simulations. The description of the scheme as the default time discretization for stochastic PDEs is the essay's wording, not a statement I found in the paper.

The fix is taming: modified schemes that prevent the nonlinearity from driving the solution to infinity. The paper analyzes a fully implicit scheme and two tamed schemes, proving non-asymptotic uniform-in-time error bounds for all three. “Uniform in time” means the error bound doesn't grow with the simulation horizon — the scheme remains faithful to both transient dynamics and long-time behavior. The general criterion they develop applies beyond Allen-Cahn: any SPDE where the nonlinearity can overwhelm the discretization needs taming, and the paper provides the framework to verify when a tamed scheme succeeds.