A shift at higher roughness
The clearest numerical signal appeared at α = 0.7 and α = 0.9, the two higher roughness settings tested. As the Fourier cutoff—the model’s spatial-resolution setting—increased, the gradient norms rose, while the reported C1 norm of the nonlinear remainder, a measure of the model’s nonlinear-current contribution, fell sharply. At α = 0.7, the simulated profile was described as noise-only rather than hill-like.
At α = 0.7, the nonlinear remainder’s C1 norm fell from 0.385331 at 16 Fourier modes to 0.000310 at 1,048,576 modes. At α = 0.9, it fell from 0.216057 to 0.000028 over the same range, while the reported gradient norm rose from 3.426971 to 739.753947.
The authors interpret the combination of larger gradients, smaller nonlinear remainders and, at α = 0.7, a noise-only profile as consistent with rough-noise suppression of the nonlinear surface current. That interpretation applies to this specified numerical model; it is not a measurement of real thin-film growth.
The calculation behind the result
The paper studies a stochastic epitaxial thin-film growth model driven by singular, spatially rough additive noise. Fourier truncation regularizes the solution; spatial discretization uses a spectral Galerkin projection and time integration uses an exponential Euler scheme.
Analytically, the authors derive strong error estimates that expose how spatial cutoff, time step and nonlinear-current decay interact. They also establish convergence of the fully discrete scheme, in the stated limiting argument, to the limiting linear Ornstein–Uhlenbeck process.
The expected H1 energy of the truncated stochastic convolution grows logarithmically with the cutoff when α = 0, the classical space-time white-noise case, and like N^(2α) when 0 < α < 1. For the rough-noise case, the expected nonlinear-current quantity is bounded by a negative power of the cutoff, with the same type of bound applied to the exponential-Euler nonlinearity.
Where the picture stops
At α = 0.4, the expected transition did not appear over the tested resolutions. The gradient norm showed no systematic growth, and the nonlinear-remainder gradient error stabilized at approximately 0.45 instead of decaying. The authors say much larger Fourier cutoffs would be needed.
The numerical tables used the unit-square domain, T = 10, a time step of h = 0.001, noise intensity σ = 0.11 and δ = 0.02; the theoretical analysis used δ = 1. The simulations started from a zero initial condition, and table entries were means over 100 simulations. No confidence intervals or other variability estimates were reported.
These were finite-time, finite-resolution calculations of the specified periodic stochastic equation, not experimental tests. They do not establish that changing noise roughness alters growth in real thin films, validate the simulated patterns against measurements, or identify a universal threshold at which hill formation disappears.
For the authors, the higher-roughness cases show the predicted transition visibly, while α = 0.4 remains unresolved. Determining how much resolution lower roughness needs remains an open numerical question.
Paper data and sources
Original title: Numerical Study of a Surface Growth Model with Singular Noise
Authors: Dirk Blömker, David Buchberger, Johannes Rimmele
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text