Two-Filter Adaptive Gaussian Mixture Smoothing for Nonlinear Systems

Source: arXiv:2607.27151 · Published 2026-07-29 · By Benjamin Schneiderheinze, Andrea De Vittori, Keith A. LeGrand, Jill Bruer

TL;DR

This paper addresses the challenge of accurately estimating the states of nonlinear dynamical systems where the underlying state distributions become significantly non-Gaussian, such as during space object tracking under highly nonlinear orbital dynamics or extended measurement gaps. Traditional filtering approaches, including Gaussian mixture filters (GMFs), approximate complex distributions but can still yield uncertain or inaccurate estimates when measurements are sparse or delayed. Smoothing methods that leverage future measurement information can reduce estimation uncertainty but are difficult to apply due to theoretical and computational challenges, particularly in nonlinear and non-Gaussian contexts.

The authors propose a novel recursive Bayesian smoother for nonlinear systems based on a two-filter approach. This approach refines Gaussian mixture posteriors produced by a forward adaptive Gaussian mixture filter by recursively estimating a future measurement likelihood represented as an information-form Gaussian mixture in the state space. They integrate advanced Gaussian mixture filtering techniques such as splitting, merging, and recursive measurement updating to improve the approximation accuracy and computational efficiency of the backward filter. The method is evaluated on challenging space object tracking problems, including Molniya and Earth-Moon halo orbits, demonstrating significant reductions in estimation error and uncertainty compared to filtering alone.

Key findings

• The proposed two-filter adaptive Gaussian mixture smoother significantly reduces estimation error and uncertainty relative to the forward adaptive Gaussian mixture filter alone in nonlinear space object tracking scenarios (Molniya and Earth-Moon halo orbits).
• The smoother employs an information-form Gaussian mixture to represent the backward filter likelihood, allowing singular precision matrices and thus handling non-invertible measurement models.
• Incorporation of splitting in both forward and backward filters improves the accuracy of nonlinear likelihood approximations, mitigating errors from local linearizations over wide Gaussian components.
• A recursive Bayesian measurement update (based on multiple gradual updates) improves stability and accuracy of the backward filter measurement incorporation, using 20 update steps per measurement.
• New precision-matrix formulations for Gaussian merging and splitting enable handling degenerate mixtures encountered during smoothing.
• The backward filter likelihood is initialized by centering likelihood components on the means of the final filtered posterior, maintaining local accuracy of approximations.
• The discrete-time backward filter prediction step applies a Taylor expansion of the inverse dynamics flow, with process noise discretized via Van Loan’s method, enabling tractable computation of the backward likelihood.
• The method is demonstrated on nonlinear orbital dynamics with strong non-Gaussianity due to long observation gaps and chaotic behavior, a scenario where classical Kalman-based smoothers fail or degrade.

Threat model

n/a - This is a state estimation methodology paper focusing on nonlinear Bayesian filtering and smoothing for space object tracking rather than a security or adversarial threat scenario.

Methodology — deep read

The paper develops a two-filter Bayesian smoothing algorithm for nonlinear, non-Gaussian state estimation problems. The problem setting is a nonlinear discrete-time or discretized continuous-time stochastic dynamical system with Gaussian process noise and nonlinear noisy measurements.

1. Threat Model & Assumptions: The task is state estimation for nonlinear system dynamics with partial, noisy, and possibly sparse measurements, typical in space object tracking. No explicit adversary is modeled. The method assumes the full batch of measurements is available post hoc for smoothing.

2. Data: The smoother is evaluated on space object tracking simulations for Molniya and Earth-Moon halo orbits with nonlinear dynamics and measurement sparsity. Details on data size and splits are not specified.

3. Architecture / Algorithm: The approach refines forward filter posteriors (Gaussian mixtures from an adaptive Gaussian mixture EKF) by recursively computing a backward filter likelihood over the state space. This backward likelihood is represented as a mixture of likelihood components parameterized by centers and singular or semidefinite precision matrices (information form).

   • Backward filter update: A nonlinear measurement update is applied to each likelihood component via first-order Taylor expansions around each component center, analogous to an EKF update but extended to mixtures and likelihoods. To improve nonlinear measurement approximation, a Bayesian recursive update applies multiple measurement updates with inflated measurement noise and sequential relinearization.

   • Backward filter prediction: The backward likelihood components are propagated backward in time by applying a linearized inverse system dynamics flow to the component centers, updating the precision matrices accounting for process noise discretized using Van Loan's matrix exponential solution.

   • Adaptive Gaussian mixture operations: Splitting and merging are applied recursively in both forward and backward filters to maintain accuracy and control mixture complexity. Splitting directions are chosen by heuristics involving first and second-order nonlinearities.

4. Training Regime: Not applicable as this is a recursive Bayesian filtering and smoothing algorithm without learning.

5. Evaluation Protocol: The method is evaluated on nonlinear orbital tracking problems demonstrating reductions in estimation error and uncertainty compared to the forward filter. The exact metrics, baselines, and statistical tests are not detailed in the excerpt. No mention of cross-validation or adversarial tests.

6. Reproducibility: The paper states the forward filter is an adaptive Gaussian mixture EKF variant; appendices provide derivations. No explicit mention of code or dataset release. The datasets are simulated space orbital data.

One example end-to-end flow is: starting from forward filter Gaussian mixture posteriors, the backward filter initializes the backward likelihood at final time by centering likelihood components on filtered mixture means. It then recursively applies multiple-step measurement updates with linearization and Bayesian recursive updates for each backward time step. Finally, it predicts backward likelihoods using backward dynamics linearization and updates precision matrices incorporating discretized process noise, producing smoothed Gaussian mixture posteriors incorporating future measurement info.

Technical innovations

• First recursive Two-Filter Adaptive Gaussian Mixture smoother for highly nonlinear, non-Gaussian systems extending prior linear system methods [29].
• Representation of backward filter likelihood as an information-form Gaussian mixture with singular precision matrices enabling non-invertible measurement models.
• Incorporation of Gaussian splitting and merging in the backward filter pass to improve nonlinear likelihood approximations and computational tractability.
• Application of a Bayesian recursive measurement update for nonlinear likelihood terms, using multiple measurement steps with relinearization to improve stability and accuracy.

Datasets

• Simulated space object tracking data for Molniya orbits — size not specified — simulation
• Simulated cislunar Earth-Moon halo orbit tracking data — size not specified — simulation

Baselines vs proposed

• Forward adaptive Gaussian mixture EKF filter: baseline estimation error and uncertainty (exact metrics not given) vs proposed two-filter AGM smoother: significant reduction in estimation error and uncertainty shown on nonlinear space object tracking tasks (Molniya and Earth-Moon orbits)

Figures from the paper

Figures are reproduced from the source paper for academic discussion. Original copyright: the paper authors. See arXiv:2607.27151.

Fig 1: Smoothing visualization for space object tracking.

Fig 2 (page 3).

Fig 3 (page 3).

Fig 4 (page 3).

Fig 5 (page 3).

Fig 6 (page 3).

Fig 7 (page 3).

Fig 8 (page 3).

Limitations

• Evaluation limited to simulated space tracking scenarios; no real data or additional applications shown.
• No adversarial robustness or distribution shift assessments presented.
• Method relies on local linearization in updates and predictions, which can degrade if nonlinearities are extreme or mixture splitting is insufficient.
• Computational scalability as mixture size grows is managed via merging but remains a potential bottleneck for very high-dimensional systems.
• No quantitative ablation studies reported isolating contribution of splitting, recursive update, or mixture representation choices.
• No codebase or data release currently available to independently verify or extend results.

Open questions / follow-ons

• How does the smoother perform under real observational noise and model mismatch encountered in operational space tracking?
• Can the two-filter AGM approach be extended or adapted for very high-dimensional or multi-object tracking problems?
• What trade-offs in computational complexity vs accuracy arise with different mixture splitting heuristics and merging tolerances in backward filtering?
• How robust is the recursive Bayesian measurement update to strong nonlinearities or discontinuities in measurement models beyond Gaussian noise assumptions?

Why it matters for bot defense

While this paper does not directly address bot defense or CAPTCHA generation, the methodology of adaptively approximating highly nonlinear and non-Gaussian posterior distributions via Gaussian mixture smoothing is relevant to adversarial robustness research. Bot and CAPTCHA detection systems often operate under complex, time-sequential uncertainty with partial observability and measurement gaps. Techniques that improve nonlinear smoothing and uncertainty quantification could inspire advanced state or intent estimation methods in authentication systems. The recursive Bayesian update using information-form likelihood mixtures might inform adaptive anomaly detection where future event information is retrospectively available. Practitioners may glean novel strategies to combine forward filtering with backward smoothing to refine detection confidence and reduce false positives in bot identification pipelines, especially in dynamic user-behavior models featuring non-Gaussian noise or sparse observations.

Cite

@article{arxiv2607_27151,
  title={ Two-Filter Adaptive Gaussian Mixture Smoothing for Nonlinear Systems },
  author={ Benjamin Schneiderheinze and Andrea De Vittori and Keith A. LeGrand and Jill Bruer },
  journal={arXiv preprint arXiv:2607.27151},
  year={ 2026 },
  url={https://arxiv.org/abs/2607.27151}
}

Read the full paper

• arXiv:2607.27151 (abstract)
• arXiv:2607.27151 (PDF)