1: \begin{abstract}

2: 

3: Monte Carlo sampling of nonlinear state-space models is particularly difficult

4: in circumstances where the transition density is not of closed form. This is

5: common in the physical and biological sciences, where real phenomena demand

6: modelling that accurately reproduces nonlinearity, chaos, mass conservation,

7: multiple potentials and other behaviours that do not necessarily yield

8: convenient analytical forms. In this work, we explore a new idea in the design

9: of samplers for such models, collapsing out the state variables of a

10: state-space model to leave only its noise terms. We then consider the design

11: of proposal distributions over these noise terms, exploiting the independence

12: and simple parametric forms prescribed to them in the prior structure of the

13: model.

14: 

15: The chosen vehicle for joint state and parameter estimation in these collapsed

16: state-space models is the particle marginal Metropolis-Hastings (PMMH)

17: sampler, from the family of particle Markov chain Monte Carlo (PMCMC). An

18: auxiliary particle filter is used in the inner loop, with proposal

19: distributions over noise terms tuned using the unscented Kalman filter. We

20: conduct a thorough empirical investigation of the PMMH sampler in this

21: context, including a look at the improvement obtained in acceptance rates,

22: convergence rates and variability of the likelihood estimator. Case studies

23: from the domain of marine biogeochemistry are used to demonstrate the

24: ideas. We believe that these cases, exhibiting mass conservation and mild

25: chaotic behaviour, are particularly challenging applications on which to

26: exercise the PMCMC methodology.

27: \end{abstract}

28: