TL;DR
The authors introduce the ppFPF (point-process Feedback Particle Filter), a novel unweighted particle filter designed for tracking hidden states that live on curved surfaces (manifolds) and are observed through discrete events. By replacing simple "constant gain" translations with exact, geometry-aware transformations, the ppFPF maintains accurate estimates of both the state and its uncertainty, even in complex non-linear scenarios.
What Problem Does This Paper Address?
In many fields—from robotics and satellite tracking to neuroscience—we need to estimate a "hidden state" (like the orientation of a drone or the firing rate of a neuron) based on noisy observations. This is called filtering.
Two major challenges often arise:
- Manifold-valued states: The state doesn't live in a flat "Euclidean" space but on a manifold (e.g., a circle, a sphere, or a rotation group). In these spaces, standard operations like "adding a vector" are ill-defined.
- Point process observations: Instead of continuous noisy data, observations arrive as discrete events (counts or "spikes"). When an event occurs, the probability distribution of the hidden state changes abruptly.
Existing unweighted filters, like the EKSPF, typically move all particles by the same "constant" amount when an event occurs. This is problematic because it ignores the geometry of the manifold and fails to capture how an event might shrink the uncertainty (variance) of the estimate.
Methods
The paper derives the ppFPF, taking inspiration from the Feedback Particle Filter (FPF) used for continuous observations.
Key technical components include:
- Unweighted Particles: Unlike traditional bootstrap filters that assign "weights" to particles (which often leads to "particle collapse" in high dimensions), the ppFPF transforms the particles themselves so they always represent the correct distribution.
- Intrinsic Dynamics: The filter is "parametrization-invariant." Whether you describe a circle using degrees or radians, the filter's behavior remains the same because it is defined using the manifold's intrinsic geometry (Lie derivatives and differential forms).
- Two-Part Evolution:
- Drift (Between events): Particles evolve according to the hidden state's natural dynamics plus a control term that accounts for the lack of events.
- Jump (At an event): When an event is observed, particles undergo an instantaneous transformation. The authors use a log-homotopy particle flow, which moves particles along a path defined by a partial differential equation (PDE) to match the new, post-event distribution.
Datasets
The paper evaluates the algorithm using synthetic benchmarks designed to highlight geometric and statistical challenges:
- 1D Diffusion on $\mathbb{R}$: A state moving on a real line with an exponential observation intensity.
- Diffusion on $S^1$ (The Unit Circle): A state moving on a circle with multiple independent Poisson observation processes. This test specifically demonstrates how coordinate-dependent filters (like EKSPF) fail on manifolds.
Key Insights and Ideas
- The "Constant Gain" Trap: The authors show that assuming a constant gain (moving all particles equally) is not just a geometric error on manifolds; it's a statistical one. It correctly updates the "mean" but fails to update the "variance," leading to overconfident or inaccurate estimates.
- Poisson Equations for Control: The optimal way to move particles is found by solving equations analogous to the weighted Poisson equation. This links the filtering problem to well-studied problems in physics and optimal transport.
- Exactness: The ppFPF is "asymptotically exact," meaning that as the number of particles increases, the filter's estimate converges perfectly to the true mathematical solution of the filtering problem.
Key Results
- Uncertainty Quantification (UQ): In the 1D real-line simulation, the ppFPF matched the accuracy of the gold-standard Bootstrap Particle Filter (BPF) and significantly outperformed the EKSPF and Assumed Density Filter (ADF) in correctly estimating the variance (uncertainty).
- Manifold Performance: On the unit circle ($S^1$), the ppFPF maintained high accuracy, while the EKSPF's performance was "compromised." Because the EKSPF relies on a coordinate chart, its "translations" became distorted by the circle's curvature, leading to much higher errors.
- Computational Framework: The authors demonstrated that the complex jump update can be implemented using existing approximation algorithms (like the differential loss RKHS method), making the theoretical "exactness" practically achievable.
Limitations and Future Work
- Computational Complexity: Solving the PDEs required for the gain estimation is more computationally intensive than the simple updates used in approximate filters.
- Gain Estimation: The filter's accuracy depends on how well the "gain" (the vector field that moves the particles) can be estimated from a finite number of samples.
- High-Dimensional Scaling: While unweighted filters are theoretically better for high-dimensional problems, this paper focuses on 1D and 2D manifolds. Future work is needed to see how these specific PDE-based updates scale to very high-dimensional state spaces.
Summary updated