[Paper Review] The Kalman Like Particle Filter : Optimal Estimation With Quantized Innovations/Measurements

Ravi Teja Sukhavasi, Babak Hassibi|arXiv (Cornell University)|Sep 5, 2009

Target Tracking and Data Fusion in Sensor Networks15 references4 citations

TL;DR

This paper proposes the Kalman Like Particle Filter (KLPF), a novel estimation algorithm for linear systems with quantized measurements, leveraging a generalized closed skew-normal (GCSN) state distribution to achieve near-optimal performance with significantly fewer particles than standard particle filters. The method combines Kalman filter structure with particle filtering by exploiting the conditional state density's analytical form, enabling efficient and accurate estimation in sensor networks and validating the classical separation principle for LQG control.

ABSTRACT

We study the problem of optimal estimation and control of linear systems using quantized measurements, with a focus on applications over sensor networks. We show that the state conditioned on a causal quantization of the measurements can be expressed as the sum of a Gaussian random vector and a certain truncated Gaussian vector. This structure bears close resemblance to the full information Kalman filter and so allows us to effectively combine the Kalman structure with a particle filter to recursively compute the state estimate. We call the resulting filter the Kalman like particle filter (KLPF) and observe that it delivers close to optimal performance using far fewer particles than that of a particle filter directly applied to the original problem. We show that the conditional state density follows a, so called, generalized closed skew-normal (GCSN) distribution. We further show that for such systems the classical separation property between control and estimation holds and that the certainty equivalent control law is LQG optimal.

Motivation & Objective

• To address the challenge of optimal state estimation in linear systems when measurements are quantized, particularly in bandwidth-constrained sensor networks.
• To develop a computationally efficient filtering method that outperforms standard particle filters in terms of particle count while maintaining high accuracy.
• To characterize the conditional state density under quantized innovations and show it follows a generalized closed skew-normal (GCSN) distribution.
• To establish that the classical separation principle holds for LQG control with quantized measurements, enabling optimal control design.
• To demonstrate through simulations that KLPF stabilizes unstable systems and achieves performance close to the optimal filter with minimal particles.

Proposed method

• The KLPF is derived by recognizing that the state conditioned on causal quantized measurements decomposes into a Gaussian vector and a truncated Gaussian vector, leading to a generalized closed skew-normal (GCSN) distribution.
• The filter uses this GCSN structure to recursively compute the posterior mean and covariance, combining Kalman-like update equations with particle filtering for non-Gaussian components.
• It applies particle filtering to the non-Gaussian part of the state distribution, significantly reducing the number of particles required compared to filtering the full state directly.
• The method is general and applies to any causal quantization scheme, including sign-of-innovation and multiple-level quantization.
• Theoretical analysis shows that the optimal filter's error covariance is upper bounded by a modified Riccati recursion, though this bound is not always tight.
• The separation principle is proven to hold, allowing the use of certainty equivalent control for LQG optimality in systems with quantized measurements.

Experimental results

Research questions

• RQ1 Can the conditional state density under quantized measurements be analytically characterized, and does it deviate from Gaussianity?
• RQ2 Can a particle filter be designed to exploit the structure of the state distribution under quantized innovations to reduce particle count while maintaining accuracy?
• RQ3 Does the classical separation principle between estimation and control hold in systems with quantized measurements?
• RQ4 Can the KLPF stabilize unstable linear systems using only a small number of particles?
• RQ5 Is the modified Riccati recursion an upper bound on the optimal filter's error covariance?

Key findings

• The conditional state density given quantized innovations follows a generalized closed skew-normal (GCSN) distribution, which is non-Gaussian and captures the skewness induced by quantization.
• The KLPF achieves near-optimal estimation performance with dramatically fewer particles than standard particle filters—e.g., reducing from 500 to 50 particles when adding just one bit of quantization.
• In Example 1, the 1-bit MLQ-KF and MLQ-KF diverge, but KLPF with 50 particles achieves optimal performance, demonstrating robustness to quantization.
• In Example 4, KLPF successfully stabilizes an unstable system (with F having eigenvalues >1) over 1000 time steps using only 100 particles, with no divergence in 1000 Monte Carlo trials.
• The modified Riccati recursion does not always bound the optimal filter's error covariance, contradicting earlier assumptions and highlighting the need for tighter error bounds.
• The separation principle holds: the certainty equivalent control law is LQG optimal, enabling optimal control design using only the KLPF state estimate.