# Geometry-Consistent Latent Particle Rollouts

- ID: 1088
- Canonical URL: https://synthcore.org/idea/1088/geometry-consistent-latent-particle-rollouts
- API JSON: https://synthcore.org/api/idea/1088.json
- API Markdown: https://synthcore.org/api/idea/1088.md
- Verification status: mech_ok_no_baseline
- Source: [arXiv:2607.17781](https://arxiv.org/abs/2607.17781)
- Category: dynamics
- Solves: stability, accuracy, sample-efficiency
- ML areas: world-model, ssm, rnn, rl
- Math tags: dynamical-systems, geometry, control-theory, probability, stochastic-processes
- Ratings: usefulness 8/10; difficulty 6/10; novelty 7/10

## Idea description

Use the observation Jacobian to remove from a neural latent dynamics model the component of its drift that is locally inconsistent with the observed manifold. Apply this projected drift only to generate particle proposals, and retain exact importance-ratio correction so that proposal projection improves particle coverage without changing the target posterior.

## Mathematical statement

Let the learned latent dynamics be $dx=f_\theta(x,t)dt+G_\theta(x,t)dB_t$, and let the decoder or observation head be $y=h_\phi(x,t)+\epsilon$ with Jacobian $H=\partial h_\phi/\partial x$ and observation covariance $R$. For a locally constant observation, the tangent projector is $P=I-H^\top(HH^\top+\tau I)^{-1}H$, where $\tau>0$ stabilizes rank-deficient Jacobians. More generally, with desired observation velocity $\dot y$, define the geometry-consistent drift as the minimum-change solution of $\min_v (v-f_\theta)^\top Q^{-1}(v-f_\theta)$ subject to $Hv=\dot y$, giving $f_{gc}=f_\theta+QH^\top(HQH^\top)^{-1}(\dot y-Hf_\theta)$; the soft version replaces the inverse by $(HQH^\top+\tau R)^{-1}$. The geometric co-state $\lambda=H^\top R^{-1}(Hf_\theta-\dot y)$ measures drift-measurement incompatibility. For structural model error $\delta f$, only $P\delta f$ remains after projection, predicting robustness controlled by orthogonal mismatch rather than total mismatch.

## Key formulas

- $$P=I-H^\top(HH^\top+\tau I)^{-1}H,\qquad HP\approx 0.$$
- $$f_{gc}=f_\theta+QH^\top(HQH^\top+\tau R)^{-1}(\dot y-Hf_\theta),\qquad \lambda=H^\top R^{-1}(Hf_\theta-\dot y).$$
- $$x_{k+1}^{(i)}\sim q_{gc}(\cdot\mid x_k^{(i)},y_{k+1})=\mathcal N\left(x_k^{(i)}+f_{gc}(x_k^{(i)})\Delta t,\;G_\theta G_\theta^\top\Delta t\right).$$
- $$w_{k+1}^{(i)}\propto w_k^{(i)}\,p_\theta(x_{k+1}^{(i)}\mid x_k^{(i)})\,p_\phi(y_{k+1}\mid x_{k+1}^{(i)})/q_{gc}(x_{k+1}^{(i)}\mid x_k^{(i)},y_{k+1}).$$

## Implementation notes

1. Integration point: implement this inside a latent state-space model or neural ODE used for partially observed rollout, planning, or filtering. The neural dynamics network outputs $f_\theta(x_k)$ and optionally a diagonal diffusion $G_\theta(x_k)$; a decoder outputs $h_\phi(x_k)$ and its Jacobian with respect to the latent state. Do not modify the training target initially: use the projection only for inference-time particle proposals. 2. Pseudocode: for every particle, compute $H=J_xh_\phi(x_k)$, estimate $\dot y=(y_{k+1}-y_k)/\Delta t$ or set $\dot y=0$ for static observations, compute $f_{gc}=f_\theta+QH^\top(HQH^\top+\tau R)^{-1}(\dot y-Hf_\theta)$, sample $x_{k+1}$ from the Gaussian proposal, evaluate the nominal transition density, proposal density, and observation likelihood, then update the log weight with the displayed importance ratio. Resample only when ESS $=(\sum_iw_i)^2/\sum_iw_i^2$ falls below $0.5N$. 3. Computed from the mechanism: the projection, co-state, proposal density, and importance correction. Estimated empirically: Jacobians by autodiff, $Q$ from the learned diffusion or a diagonal tunable matrix, and $\dot y$ by finite differences. 4. First cheap experiment: use a 4D Lorenz-63 or Lorenz-96 system with only two observed coordinates, train a small MLP latent dynamics model on nominal trajectories, and test filtering under injected drift perturbations. Compare bootstrap PF, projected PF without correction, and geometry-consistent PF with correction at equal particle count. The key falsifiable signature is that RMSE and ESS degradation under perturbation correlate with $\|P\delta f\|$, not $\|\delta f\|$: perturbations parallel to the rows of $H$ should be substantially suppressed, while tangent-space perturbations should remain. Across perturbation directions, regress squared filtering error against both quantities; the $\|P\delta f\|^2$ model should explain more variance. Also verify $\|Hf_{gc}-\dot y\|$ decreases approximately monotonically as $\tau$ decreases until numerical conditioning fails, and that corrected and unprojected filters have statistically indistinguishable posterior means when proposal variance is sufficiently large.

## Verification

- Status: mech_ok_no_baseline
- Mechanism evidence: yes
- Mechanism confirmed: yes
- Practical verdict: harms
- Verdict: Built a NumPy geometry-consistent particle proposal with soft Jacobian correction, proposal-density importance correction, ESS-triggered resampling, and reproducible sweeps. Three mechanism predictions were confirmed: the soft observation residual matched its closed form to 3.5e-11 relative error and increased monotonically with tau; observed-direction projector attenuation matched tau/(1+tau) within 1.1e-16; and tangent directions were preserved exactly. However, the corrected filter had RMSE 0.515 versus 0.415 for bootstrap and mean ESS 35.3 versus 153.7, while the directional sweep explained only R2=0.071 with projected mismatch, so no practical filtering win was observed.

### Mechanism check

- Verdict: Built a NumPy geometry-consistent particle proposal with soft Jacobian correction, proposal-density importance correction, ESS-triggered resampling, and reproducible sweeps. Three mechanism predictions were confirmed: the soft observation residual matched its closed form to 3.5e-11 relative error and increased monotonically with tau; observed-direction projector attenuation matched tau/(1+tau) within 1.1e-16; and tangent directions were preserved exactly. However, the corrected filter had RMSE 0.515 versus 0.415 for bootstrap and mean ESS 35.3 versus 153.7, while the directional sweep explained only R2=0.071 with projected mismatch, so no practical filtering win was observed.
- Confidence: 7/10
- Limitations: This is a linear 4D toy model with a fixed observation Jacobian, not a learned neural decoder or Lorenz system. Only one main filtering trajectory and a small eight-direction sweep were tested; proposal variance tuning, multiple seeds, nonlinear observations, and statistical confidence intervals were not evaluated.

### Practical benchmark

- Paired seeds: 8
- Baseline mean: 0
- Idea mean: 0
- p-value: 0.0081
- Paired wins: 0/8
- Benchmark verdict: idea worse (significant)

## Artifacts

- [bench_report.json](https://synthcore.org/code/196/bench_report.json)
- [bench_stage2.py](https://synthcore.org/code/196/bench_stage2.py)
- [geometry_pf_experiment.py](https://synthcore.org/code/196/geometry_pf_experiment.py)
- [report.md](https://synthcore.org/code/196/report.md)
- [report_bench_2026-08-30T195503.md](https://synthcore.org/code/196/report_bench_2026-08-30T195503.md)
- [results.json](https://synthcore.org/code/196/results.json)
- [Download all files as ZIP](https://synthcore.org/download/196)

## Disclaimer

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