The Variational Approach in Filtering and Correlated Noise

Abstract: The variational formulation of nonlinear filtering due to Mitter and Newton characterizes the filtering distribution as the unique minimizer of a free energy functional involving the relative entropy with respect to the prior and an expected energy. This formulation rests on an absolute continuity condition between the joint path measure and a product reference measure. We prove that this condition necessarily fails whenever the signal and observation diffusions share a common noise source. Specifically we show that the joint and product measures are mutually singular, so no choice of reference measure can salvage the formulation. We then introduce a conditional variational principle that replaces the prior with a reference measure that preserves the noise correlation structure. This generalization recovers the Mitter--Newton formulation as a special case when the noises are independent, and yields an explicit free energy characterization of the filter in the linear correlated-noise setting.

• The paper proves that the classical variational filtering framework fails under correlated noise due to the breakdown of absolute continuity.
• It rigorously analyzes the limitations of the Mitter–Newton approach using discrete and Gaussian models to demonstrate mutual singularity.
• The study introduces a new conditional variational principle that employs a conditional reference measure to recover filtering structure.

Variational Filtering with Correlated Noise: Failure and Resolution

Overview

This paper systematically investigates the variational approach to stochastic filtering put forth by Mitter and Newton, which recasts the conditional distribution of a system's hidden state, given noisy observations, as a Gibbs measure minimizing a free energy functional. The classical formulation fundamentally relies on independence between signal and observation noises to establish absolute continuity between the joint path measure and a suitable product reference measure. The paper rigorously proves that this condition breaks down when signal and observation share a common noise source, as is typical in models with correlated noises, and proposes a generalized variational formulation that restores the characterization in this challenging regime by introducing a conditional reference measure that preserves the noise correlation structure.

Mitter–Newton Variational Principle and Its Limitation

The variational approach to nonlinear filtering, as developed by Mitter and Newton, hinges on a free energy minimization involving the relative entropy (Kullback–Leibler divergence) of a candidate distribution with respect to the prior, plus an expected energy term. When the signal and observation processes are driven by independent Brownian motions, the Cameron–Martin–Girsanov theorem enables an explicit representation of the Radon–Nikodym derivative between the joint law and the product of the signal prior and observation reference measure. This is essential for constructing the free energy functional whose unique minimizer is the posterior.

However, this independence becomes untenable in numerous practical cases—such as measurements subject to common environmental disturbances—where the signal and observation are coupled via a shared noise source. The paper shows that in such cases, the absolute continuity assumption (as formalized in Assumption 2.1) cannot hold, and that no alternative choice of reference measure can recover the variational structure. The degeneracy is demonstrated both through an explicit discrete-time model and through infinite-dimensional path-space Gaussian measure theory, employing the Feldman–Hajek dichotomy: mutually singularity arises whenever the covariance structure reflects noise correlation.

Mutual Singularity in Filtering with Correlated Noise

For both linear and nonlinear diffusions with correlated noise, the paper constructs sets with full measure under the joint law but null measure under any product of marginals, thereby directly establishing mutual singularity. In the linear Gaussian case, the difference in Cameron–Martin spaces and covariances between the joint and product measures immediately implies singularity; in the general nonlinear case, the quadratic variation techniques show that the supports of the conditional and prior laws are disjoint almost surely, leveraging integrability and regularity results (e.g., the Hörmander condition).

A strong result is articulated: no variational formulation based on the prior can exist in correlated-noise settings. The failure is intrinsic whenever signal and observation share even partial noise sources; in those cases, the posterior is often singular with respect to the prior.

Conditional Variational Principle

The authors introduce a new variational framework adapted to the structure of correlated noise: rather than seeking the minimizer of free energy relative to the signal prior, the posterior is characterized as a unique minimizer relative to a conditional reference measure $Q$ that preserves the noise coupling. This conditional reference measure is constructed so that its marginal over observations matches the true observation law and, crucially, the conditional law given any observation path is absolutely continuous with respect to the posterior.

Formally, for each observation $y$, the energy term $H(x, y)$ is defined via the Radon–Nikodym derivative between the posterior and the conditional reference, and the variational problem minimizes the sum of relative entropy and expected energy over all candidate measures.

This approach recovers the classical Mitter–Newton result as a special case when signal and observation noises are independent (i.e., $Q$ can be chosen as the signal prior), but generalizes to correlated settings by encoding the coupling structure in $Q$.

Explicit Characterization for Linear Correlated Diffusions

The paper provides an explicit construction for the linear Gaussian signal-observation model with correlated noise. Here, the reference law $Q$ is chosen as the driftless dynamics maintaining shared noise structure. The conditional reference measure $\mu_y$ is a Gaussian measure where the signal path is a function of the observation path and private noise components. The Radon–Nikodym derivative is given by the exponential of a stochastic integral representing the energetic cost of deviating from the driftless reference, precisely quantifying the effect of the drift component relative to the shared noise.

Critical to this construction is non-degeneracy of the signal's private noise: the invertibility of the diffusion matrix ensures that the Girsanov theorem is applicable and that the conditional law is well-defined. In regimes where the signal is fully determined by correlated noise (i.e., no private noise component), the conditional law collapses, and the variational characterization becomes vacuous.

Practical and Theoretical Implications

The results have direct impact on the theory and application of stochastic filtering in systems with correlated measurement and process noises, which arise in control, signal processing, and finance. The impossibility of Gibbs-based variational representations relative to the prior under correlated noise forces practitioners to adopt conditionally structured reference measures and energy terms. This formalism clarifies which algorithmic and theoretical approaches are valid, and provides a rigorous pathway to filter characterization in challenging nonlinear settings.

Moreover, the conditional variational principle suggests new avenues for algorithmic development—potentially involving pathwise sampling, simulation, or optimization over conditional measure spaces. However, the extension to full nonlinear correlated diffusions, e.g., via robust or rough paths, remains an open problem.

Conclusion

The paper rigorously proves the breakdown of the classical variational filtering formulation under correlated noise and proposes a conditional variational principle that resolves this limitation, especially in linear settings. The new formulation strictly subsumes the prior-based variational principle and is structurally equipped to handle noise coupling, provided regularity and invertibility conditions hold. Extensions to nonlinear filtering with correlated noise and further algorithmic implications constitute important directions for future research in stochastic estimation.