Stochastic trajectory generation moves past one dimension: SINN constrains joint distributions with projected densities, and joint-density error falls monotonically as projection frames are added on a coupled-oscillator benchmark
Synopsis
Statistics-informed neural networks (SINNs) had been demonstrated only on one-dimensional stochastic processes, and independently trained one-dimensional models cannot reproduce the joint distribution across components. The multidimensional SINN constrains that joint distribution through densities of several one-dimensional linear projections and balances multiple statistical objectives with adaptive loss weights; on a benchmark of two coupled Duffing oscillators, joint-density error falls monotonically as projection frames are added, and that error enters neither training nor model selection.
Interpretation
Trajectory generation for multidimensional stochastic processes no longer requires direct estimation of high-dimensional probability densities. SINN instead constrains the joint distribution across components with a set of one-dimensional projected densities along chosen directions, bringing joint-distribution information into the training objective in a computable form. Earlier SINN demonstrations covered one-dimensional non-Markovian processes, and independently trained one-dimensional SINNs cannot reproduce inter-component correlations and higher-order cross statistics, while direct multivariate density estimation becomes computationally difficult with dimension. On the two-coupled-Duffing benchmark, the relative joint-density error enters neither the training objective nor checkpoint selection, and it decreases monotonically at every coupling strength as projection frames go from the coordinate frame alone to the full frame set.
Multiple statistical objectives no longer rely on manually fixed weights. Each statistical family carries a trainable positive scale parameter optimized together with the network parameters, so the relative contribution of each family adjusts during training. In the multidimensional setting the number of statistical objectives grows substantially, and the families differ in optimization sensitivity and convergence behavior, making fixed weights hard to choose. Across 70 runs per configuration at coupling strengths 0, 0.5, 1, 2, 4, and 8, both balancing-enabled and balancing-disabled settings had 66 of 70 runs meet the validation criterion; with balancing, the median run-level spread of family validation-loss ratios at retained checkpoints is lower at every nonzero coupling strength.
Validation loss is no longer read as absolute accuracy. Reference normalization divides the statistical discrepancy between generated and validation ensembles by the finite-sample discrepancy between two independent target ensembles, yielding a dimensionless ratio. Statistical quantities are estimated from finite ensembles, so even two independent ensembles drawn from the same distribution produce nonzero discrepancies, and a validation loss cannot be interpreted in isolation. Ratios above 10 were not observed among the reference ensembles themselves, supporting 10 as a practical validation-failure threshold; in the FPU and lattice Lotka–Volterra applications, the distinction between validation-successful and validation-unsuccessful models largely persists under independent final evaluation.
Perspective
The demonstrated scope is limited to two- and three-dimensional target processes: the coupled-Duffing benchmark, coarse observables formed from two and three spatially separated positions on a 100-particle FPU chain, and prey and predator densities from kinetic Monte Carlo simulations of a lattice Lotka–Volterra model. The intended readers are researchers in simulation and surrogate modeling who need large ensembles of stochastic trajectories.
A finite set of projection directions does not uniquely determine the complete multivariate distribution, and temporal structure not adequately constrained by the training objective is not guaranteed to be reproduced; the two-dimensional direction set guarantees identifiability of joint cumulants only through order seven, and the three-dimensional set only through order three.
