Neural operator filter NOF skips Yau–Yau's offline PDE solves, cutting 2D runtime about 77.9% versus a particle filter
Synopsis
Yau–Yau nonlinear filtering splits state estimation into an offline stage that solves several Fokker–Planck–Kolmogorov equations and an online projection onto prescribed basis functions, and both stages are costly. The neural operator filter (NOF) instead uses a Fourier neural operator to learn the map from prior density to posterior density directly, removing both steps. On one- and two-dimensional cubic filtering benchmarks, NOF's mean-square error is comparable to a particle filter and the LGS method, while in the 2D case its average runtime is about 77.9% lower than the particle filter's and about 63.3% lower than LGS's.
Interpretation
In nonlinear filtering, the posterior density no longer has to come from repeated offline partial-differential-equation solves plus an online projection; a once-trained operator supplies it directly. Yau–Yau filtering originally solved the Fokker–Planck–Kolmogorov equations offline with a prescribed set of basis functions, then projected the prior density onto those basis functions online, so solving several equations offline was expensive and the online projection became costly under repeated filtering or in high dimensions. On the 1D cubic filtering benchmark, the three valid methods NOF, particle filter and LGS reach statistically comparable average mean-square error, with differences well within one standard deviation; the extended Kalman filter diverges for that cubic observation model and is excluded from the accuracy comparison.
The mechanism uses a Fourier neural operator as a surrogate for the forward map of the Fokker–Planck–Kolmogorov equation, reused inside the online filtering loop. A Fourier neural operator is a network that learns mappings between function spaces; it transforms the input into the Fourier domain, applies a learned kernel there, transforms back, and keeps only a finite number of Fourier modes. Training data combine prior–posterior density pairs from Monte Carlo simulations of the true dynamics with posteriors computed by a finite-difference Yau–Yau solver, and equation input–output pairs whose initial conditions are zero-mean Gaussian random fields with finite-difference reference solutions; the 1D example uses 235 trajectories on a 100-point spatial grid, and the 2D example uses 1397 input–output pairs plus 599 test trajectories.
The speed gain also holds in the 2D example, where accuracy is slightly below the particle filter. In the 2D cubic system the particle filter gives the lowest mean-square error, LGS comes next and NOF follows, a different ordering from the 1D example. In the 2D example NOF's average runtime is about 77.9% lower than the particle filter's and about 63.3% lower than LGS's; in the 1D example NOF takes only seconds per trial, about 83.1% less than the particle filter and 51.9% less than LGS.
Perspective
The demonstrated scope is a 1D cubic filter and a 2D system with linear drift and cubic observations, both strongly nonlinear; accuracy conclusions cover low- and medium-dimensional systems only, and high-dimensional settings are not tested. The theory gives a qualitative consistency estimate for a fixed number of observation steps, with deliberately non-sharp constants.
The theoretical error bound holds only for a fixed number of observation steps and its constants are not meant to be sharp; as the step size tends to zero the forward operator becomes near-identity and operator learning offers no benefit, so that limit is not of practical interest. In the 2D example NOF's mean-square error is higher than the particle filter's, so the speed-accuracy trade-off must be weighed per application.
