Abstract
Using Kalman techniques, it is possible to perform optimal estimation in linear Gaussian state-space models. Here, we address the case where the noise probability density functions are of unknown functional form. A flexible Bayesian nonparametric noise model based on Dirichlet process mixtures is introduced. Efficient Markov chain Monte Carlo and sequential Monte Carlo methods are then developed to perform optimal batch and sequential estimation in such contexts. The algorithms are applied to blind deconvolution and change point detection. Experimental results on synthetic and real data demonstrate the efficiency of this approach in various contexts.
Introduction
In engineering and signal processing, linear dynamic models are the standard tool for estimating hidden states from noisy observations. When the noise follows a known Gaussian distribution, the Kalman filter provides the optimal solution. However, real-world data frequently violates these assumptions, exhibiting impulsivity, multimodality, or heavy tails. In such scenarios, traditional filters become sub-optimal. This paper introduces a Bayesian nonparametric framework that utilizes Dirichlet Process Mixtures (DPM) to model noise of unknown functional form, allowing for robust state estimation without prior knowledge of the noise's probability density function.
The Dirichlet Process Mixture Framework
The core innovation lies in modeling the unknown noise distributions for both state transitions and observations as DPMs. Rather than selecting a specific distribution (such as a Student-t or a fixed Mixture of Gaussians), the DPM treats noise as an infinite mixture of Gaussians.
This approach offers several advantages:
- Automatic Complexity Scaling: The model determines the necessary number of mixture components based on the data itself.
- Polya Urn Representation: By utilizing the Polya Urn scheme, the authors simplify posterior updates. New noise samples either cluster with existing behaviors or initiate a new component, effectively "learning" the shape of the noise on the fly.
- Dynamic Hyperparameters: The scaling parameter $\alpha$, which controls the creation of new clusters, is treated as an unknown variable. This allows the model to be more exploratory in early stages and more stable as it converges.
Computational Efficiency via Rao-Blackwellization
To make inference tractable, the authors employ Rao-Blackwellization. Because the model remains linear-Gaussian once the mixture parameters are fixed, the hidden states can be analytically integrated out using standard Kalman filtering and smoothing steps. This reduces the search space for sampling algorithms, as they only need to explore the mixture indicators rather than the entire state space.
The paper develops two primary computational strategies:
- Offline (Batch) Estimation: A Gibbs sampler (MCMC) utilizing a backward-forward recursion. This approach maintains linear computational complexity relative to time ($O(T)$), making it suitable for large historical datasets.
- Online (Sequential) Estimation: A Rao-Blackwellized Particle Filter (RBPF) designed for real-time applications where data arrives sequentially.
Experimental Results
Synthetic Blind Deconvolution
The model was tested on a Bernoulli-Gaussian process—a signal characterized by occasional spikes—observed through a filter with additive noise. The DPM-based model (M1) achieved a Mean Squared Error (MSE) only 10% higher than an "oracle" model that possessed perfect knowledge of the complex noise distribution. This significantly outperformed standard filters relying on Gaussian assumptions.
Biomedical Time Series: Kidney Transplant Monitoring
The framework was applied to a classic benchmark: serum creatinine levels in patients following kidney transplants. The goal was to detect "change points" or sudden jumps in creatinine that indicate shifts in kidney health. The MCMC algorithm successfully identified three major jumps, while the online Particle Filter identified two, demonstrating the trade-off between the global perspective of batch processing and the immediacy of real-time filtering.
Discussion and Future Work
While the DPM approach provides unprecedented flexibility for noise modeling, it faces certain limitations. The Rao-Blackwellized Particle Filter can struggle with "static" parameters; once a cluster value is initialized, the filter finds it difficult to adjust that value later. Additionally, the current framework is restricted to linear models and assumes that the noise statistics, while unknown, remain static over the duration of the sequence.
Future research directions include extending this nonparametric approach to nonlinear dynamic models and developing "Time-Varying DPMs" to account for sensors or systems whose noise characteristics evolve over time.
Summary updated