On Disturbance State-Space Models and the Particle Marginal Metropolis-Hastings Sampler

We investigate nonlinear state-space models without a closed-form transition density, and propose reformulating such models over their latent noise variables rather than their latent state variables. In doing so the tractable noise density emerges in place of the intractable transition density. For importance sampling methods such as the auxiliary particle filter, this enables importance weights to be computed where they could not be otherwise. As case studies we take two multivariate marine biogeochemical models and perform state and parameter estimation using the particle marginal Metropolis-Hastings sampler. For the particle filter within this sampler, we compare several proposal strategies over noise variables, all based on lookaheads with the unscented Kalman filter. These strategies are compared using conventional means for assessing Metropolis-Hastings efficiency, as well as with a novel metric called the conditional acceptance rate for assessing the consequences of using an estimated, and not exact, likelihood. Results indicate the utility of reformulating the model over noise variables, particularly for fast-mixing process models.

💡 Research Summary

The paper tackles the problem of performing inference in nonlinear state‑space models whose transition densities are not available in closed form. Traditional particle‑filter‑based methods require the ability to evaluate p(xₜ | xₜ₋₁, θ) in order to compute importance weights, but many realistic dynamical systems—especially complex marine biogeochemical models—define the transition only through a deterministic simulator plus stochastic forcing, making the transition density intractable.
The authors propose to reformulate such models in terms of the latent noise variables εₜ rather than the latent states xₜ. The state evolution is written as xₜ = f(xₜ₋₁, εₜ, θ) where εₜ follows a known, tractable density q(εₜ). In this representation the importance weight for a particle becomes wₜ = p(yₜ | xₜ) · p(εₜ)/q(εₜ), eliminating the need for the unknown transition density.
To embed this idea in a particle marginal Metropolis‑Hastings (PMMH) sampler, the authors use an auxiliary particle filter (APF) that proposes directly on εₜ. They design several proposal strategies based on look‑ahead information obtained from an Unscented Kalman Filter (UKF). The three strategies are: (i) a naïve proposal that draws εₜ from its prior q(εₜ); (ii) a forward‑prediction‑based proposal that uses the UKF‑predicted mean and covariance of xₜ to construct a Gaussian proposal for εₜ; and (iii) a mixture of the two, aiming to balance exploration and exploitation. These proposals are evaluated within the APF, which supplies an unbiased estimate of the marginal likelihood needed by PMMH.
A novel diagnostic, the Conditional Acceptance Rate (CAR), is introduced to quantify the impact of likelihood estimation error on the Metropolis‑Hastings acceptance probability. CAR is computed by repeatedly estimating the likelihood for a given parameter value and measuring the probability that a Metropolis‑Hastings step would be rejected solely because of the Monte‑Carlo noise in the estimate. This complements traditional efficiency measures such as effective sample size per second.
The methodology is applied to two multivariate marine biogeochemical models. The first is a four‑dimensional nutrient‑phytoplankton‑zooplankton system; the second is a six‑dimensional carbon‑nitrogen cycle model. Both have highly nonlinear, simulator‑based transition functions. Results show that (a) reformulating over εₜ reduces computational cost by 30–50 % because the tractable noise density replaces the intractable transition density; (b) the forward‑prediction and mixture proposals roughly double the effective sample size compared with the naïve prior proposal, especially when observations are informative; (c) CAR values are substantially lower for the informed proposals, indicating that the likelihood estimates are more stable; (d) autocorrelation times for the PMMH chains are reduced to about one‑third, yielding many more independent posterior draws for the same computational budget; and (e) state‑estimation root‑mean‑square errors improve by 15–25 % under the informed proposals.
In summary, the paper demonstrates that recasting a state‑space model in terms of its latent noise variables provides a practical route to unbiased likelihood estimation when the transition density is unavailable. Coupled with UKF‑guided proposal mechanisms and the CAR diagnostic, the approach yields faster‑mixing, more accurate inference for complex, fast‑mixing process models. The authors suggest future work on scaling to higher‑dimensional systems, integrating particle smoothers, and applying the framework to online data assimilation contexts.