✗ Failed on benchmark 2026

Coefficient-Space Neural Uncertainty Filter

Implementation & benchmark of arXiv:2607.16504 — Polynomial Chaos Expansion Based Nonlinear Filtering of Stochastic Processes

Usefulness8/10

Difficulty6/10

Novelty7/10

Dynamics Accuracy Stability Speedup World model Rnn Rl Training dynamics Stochastic processes Probability Approximation theory Linear algebra Dynamical systems

Source paper: Polynomial Chaos Expansion Based Nonlinear Filtering of Stochastic Processes arXiv:2607.16504 ⓘ · analyzed Aug 30, 2026

AI-generated research hypothesis, automatically tested. Not peer-reviewed.

Idea description

Replace an EKF or a large particle ensemble inside a neural world model with a fixed-order polynomial chaos representation of the latent state distribution. The transition network is evaluated under quadrature or sampled chaos variables, and Galerkin projection produces the next uncertainty coefficients directly; a coefficient-wise LMMSE update then assimilates observations without backpropagating through resampling.

Formulas

$$d\mathbf{x}(t)=\mathbf{f}(\mathbf{x}(t),t)dt+\mathbf{G}(\mathbf{x}(t),t)d\mathbf{w}(t)$$

$$\mathbf{x}\approx\sum_{\alpha\in\mathcal{A}}\mathbf{c}_\alpha\Psi_\alpha(\boldsymbol{\xi}),\qquad \mathbb{E}[\Psi_\alpha\Psi_\beta]=\delta_{\alpha\beta},\qquad \operatorname{Cov}(\mathbf{x})\approx\sum_{\alpha\neq0}\mathbf{c}_\alpha\mathbf{c}_\alpha^\top$$

$$\mathbf{c}^+_\alpha=\mathbf{c}_\alpha+\Delta t\,\mathbb{E}[\mathbf{f}(\mathbf{x},t)\Psi_\alpha]+\sqrt{\Delta t}\,\mathbb{E}[\mathbf{G}(\mathbf{x},t)\boldsymbol{\eta}\Psi_\alpha]$$

$$\mathbf{K}=\mathbf{P}_{xy}\mathbf{P}_{yy}^{-1},\quad \mathbf{P}_{xy}=\sum_{\alpha\neq0}\mathbf{c}_\alpha\mathbf{d}_\alpha^\top,\quad \mathbf{P}_{yy}=\sum_{\alpha\neq0}\mathbf{d}_\alpha\mathbf{d}_\alpha^\top+\mathbf{R},\quad \mathbf{c}_0^+=\mathbf{c}_0^-+\mathbf{K}(\mathbf{y}_{obs}-\mathbf{d}_0^-),\quad \mathbf{c}_\alpha^+=\mathbf{c}_\alpha^- -\mathbf{K}\mathbf{d}_\alpha^-$$

Mathematical statement

The paper starts from the Itô SDE $d\mathbf{x}(t)=\mathbf{f}(\mathbf{x}(t),t)dt+\mathbf{G}(\mathbf{x}(t),t)d\mathbf{w}(t)$, where $\mathbf{x}(t)\in\mathbb{R}^{d_x}$ is the stochastic state, $\mathbf{f}$ is the drift, $\mathbf{G}$ is the diffusion matrix, and $\mathbf{w}(t)$ is Brownian motion. Represent each state component with an orthonormal polynomial chaos basis $\{\Psi_\alpha(\boldsymbol{\xi})\}_{\alpha\in\mathcal{A}}$, where $\boldsymbol{\xi}$ contains standardized latent randomness and $\mathbb{E}[\Psi_\alpha\Psi_\beta]=\delta_{\alpha\beta}$: $\mathbf{x}\approx\sum_{\alpha\in\mathcal{A}}\mathbf{c}_\alpha\Psi_\alpha$. Here $\mathbf{c}_\alpha\in\mathbb{R}^{d_x}$ is a coefficient and $\alpha$ is a multi-index; $\mathbf{c}_0=\mathbb{E}[\mathbf{x}]$, and for orthonormal bases $\operatorname{Cov}(\mathbf{x})\approx\sum_{\alpha\neq0}\mathbf{c}_\alpha\mathbf{c}_\alpha^\top$. For timestep $\Delta t$, Euler-Maruyama gives $\mathbf{x}^+=\mathbf{x}+\mathbf{f}(\mathbf{x},t)\Delta t+\mathbf{G}(\mathbf{x},t)\sqrt{\Delta t}\,\boldsymbol{\eta}$, with standard normal process-noise vector $\boldsymbol{\eta}$. Galerkin projection gives $\mathbf{c}^+_\alpha=\mathbb{E}[\mathbf{x}^+\Psi_\alpha]=\mathbf{c}_\alpha+\Delta t\,\mathbb{E}[\mathbf{f}(\mathbf{x},t)\Psi_\alpha]+\sqrt{\Delta t}\,\mathbb{E}[\mathbf{G}(\mathbf{x},t)\boldsymbol{\eta}\Psi_\alpha]$. For observation $\mathbf{y}=\mathbf{h}(\mathbf{x})+\boldsymbol{\varepsilon}$, define $\mathbf{d}_\alpha=\mathbb{E}[\mathbf{h}(\mathbf{x})\Psi_\alpha]$, $\mathbf{P}_{xy}=\sum_{\alpha\neq0}\mathbf{c}_\alpha\mathbf{d}_\alpha^\top$, and $\mathbf{P}_{yy}=\sum_{\alpha\neq0}\mathbf{d}_\alpha\mathbf{d}_\alpha^\top+\mathbf{R}$, where $\mathbf{R}=\operatorname{Cov}(\boldsymbol{\varepsilon})$. The LMMSE gain is $\mathbf{K}=\mathbf{P}_{xy}\mathbf{P}_{yy}^{-1}$, followed by $\mathbf{c}_0^+=\mathbf{c}_0^-+\mathbf{K}(\mathbf{y}_{obs}-\mathbf{d}_0^-)$ and $\mathbf{c}_\alpha^+=\mathbf{c}_\alpha^- -\mathbf{K}\mathbf{d}_\alpha^-$ for $\alpha\neq0$. Superscripts $-$ and $+$ denote pre- and post-observation coefficients.

Implementation notes

Integrate this at the latent-state interface of a small neural world model or recurrent estimator. Let the latent state have dimension $D$, choose total polynomial order $p=2$ or $3$, and let $M=\binom{q+p}{p}$ coefficients be used for $q$ independent Gaussian chaos variables. Store a tensor C of shape [M,B,D], with coefficient zero as the latent mean. Use a fixed tensor Xi of quadrature points or antithetic standard-normal samples and precompute the orthonormal Hermite matrix Psi[m,alpha]. At each transition, reconstruct X[m]=sum_alpha C[alpha]*Psi[m,alpha], evaluate the neural dynamics F=transition_net(X, action), and, if process noise is state dependent, evaluate G; form Xnext=X+dt*F+sqrt(dt)*G*Eta. Reproject with Cnext[alpha]=sum_m w[m]*Xnext[m]*Psi[m,alpha], which implements the Galerkin formula. For an observation, evaluate Y[m]=measurement_net(X[m]), compute Dcoef[alpha]=sum_m w[m]*Y[m]*Psi[m,alpha], construct Pxy and Pyy from the displayed sums, add diagonal jitter $\epsilon I$ to Pyy, solve for K rather than explicitly inverting it, and apply the coefficient-wise update. Backpropagate through reconstruction, transition evaluation, projection, and update; optionally train with negative log-likelihood and a coverage penalty. The math determines projection and LMMSE equations. Empirical choices are polynomial order, quadrature, noise covariance, and jitter. First test a 2D nonlinear pendulum or Lorenz-63 observation task with a GRU transition, comparing against a diagonal EKF and a 128-particle filter at matched transition-network evaluations. Measure one-step and long-horizon negative log-likelihood, RMSE, 90% interval coverage, coefficient-norm explosions, and wall-clock cost. Success is better calibration and accuracy than EKF with substantially fewer transition evaluations than particles.

Verification

Failed on benchmark

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Not reproduced ✗

Stage 2 · Practical benchmark result: No measurable effect

Methodology: 8 paired seeds, permutation test (p<0.05) against a learning-rate-tuned baseline. How verification works

Stage 1 — Mechanism check agent confidence 8/10

Built a fixed-order Hermite PCE filter with Gauss-Hermite Galerkin projection and coefficient-wise LMMSE observation updates. The mechanism checks passed: degree-2/3 truncation errors vanished at matching order (about 1e-15), Parseval variance loss matched exactly, and the LMMSE post-update variance identity differed by 1.2e-16. In the nonlinear scalar toy task, PCE was slightly worse than EKF: RMSE 0.2670 vs 0.2637, NLL 0.0756 vs 0.0541, and 90% coverage 0.8806 vs 0.8986; it also used 48 versus 1 transition evaluations per step.

Agent confidence: 8/10

Baseline: Diagonal EKF: RMSE=0.2637, NLL=0.05415, 90% coverage=0.8986, 1 transition evaluation/step
Idea: Order-3 PCE: RMSE=0.2670, NLL=0.07557, 90% coverage=0.8806, 48 quadrature transition evaluations/step; math checks: exactness error <=5.9e-15, LMMSE variance identity error=1.2e-16

Limitations: This MVP is scalar and uses a nonlinear toy filter rather than a neural world model or GRU. It does not compare against a particle filter, and additive process noise is approximated by variance inflation rather than represented with augmented independent chaos variables. Runtime and FLOP advantages were not established; the PCE quadrature implementation was substantially more expensive than the EKF in this setup.

How to run: python3 pce_filter_experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2607.16504) Sep 2, 2026

No significant difference vs. the tuned baseline (+0.6%, p=0.2379, wins 2 / 8 paired seeds).

Benchmark: Dynamics forecasting
Model: rnn_small
Paired seeds: 8
Baseline mean: 0.0014
Idea mean: 0.0014
Effect (Δ): 0 (+0.6%; negative = idea better)
Wins: 2 / 8 paired seeds
p-value: 0.2379 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±0.9%
Mechanism: Not confirmed ✗
Practical effect: No effect
Baseline tuning: swept over 3 configs

Limitations:

Only the registered built-in dynamics track was tested. The MVP uses a one-dimensional fixed perturbation chaos construction rather than a full augmented process-noise chaos state, and it evaluates coefficient uncertainty on model outputs rather than implementing sequential observation assimilation. No particle-filter or EKF comparison was included in the fixed bench report.

How to run: python3 bench_pce_dynamics.py

Verdict computed by deterministic test code from paired-seed statistics — not by the language model.

Stage 2 — Benchmark Aug 30, 2026 latest Failed ✗

Agent confidence: 10/10

On the registered built-in dynamics track with shared rnn_small architecture, the tuned baseline achieved mean test MSE 0.0014198302 and the coefficient-space PCE idea achieved 0.0014280338. The paired delta was +8.2036e-06 with permutation p=0.23785 and only 2/8 idea wins, yielding no measurable effect. The trained-model mechanism signature was also not confirmed: predicted coefficient RMS 0.11892035 versus observed residual RMS 0.01930083, ratio 6.1614.

Baseline: best lr=0.01; sweep means: lr=0.001 => 0.0042987791, lr=0.003 => 0.0034615941, lr=0.01 => 0.0014865802; full 8-seed mean=0.0014198302, std=0.0003983911.
Idea: order=2, sigma=0.08, best lr=0.01; 8-seed mean=0.0014280338, std=0.0003961076; paired delta=+0.0000082036, p=0.23785, wins=2/8; mechanism ratio=6.1614, coverage90=0.985, confirmed=false.

Benchmark result