Scalable Bayesian inference for time-discretised Markov jump processes via exchangeable particle Gibbs

arXiv:2611.05232 · stat.CO · Submitted 2026-11-03 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.

Tom: We're looking at the paper: "Scalable Bayesian inference for time-discretised Markov jump processes via exchangeable particle Gibbs".

Jane: Here's the gist of the paper. Inference in stochastic reaction-network models, such as the susceptible-exposed-infectious-recovered (SEIR) epidemic model, is crucial for understanding the dynamics of interacting systems in epidemiology, ecology, and systems biology This paper focuses on improving the efficiency of particle Gibbs for inference on reaction networks by addressing the degeneracy problem The proposed method retains the exchangeability framework of xPG while adapting it to the structural and computational challenges specific to reaction networks The Core Methodology The paper develops a novel version of exchangeable particle Gibbs (xPG) tailored to discrete-state-space reaction networks, where randomness is driven by Poisson processes rather than Brownian motion This method is applied to the tau-leap discretisation of the MJPs The proposed method retains the exchangeability framework of xPG while adapting it to the structural and computational challenges specific to reaction networks Addressing Particle Degeneracy Particle degeneracy occurs when over successive resampling steps in the internal particle filter most of the ancestor trajectories get discarded leaving very few unique particles to represent the historical path of the system The paper introduces xPG, which is intended for discrete-time models with continuous state spaces The key idea of xPG applied to SDEs is to correlate both the initial state X0 and the Brownian motions of the proposed paths with the equivalent quantities in the reference path rather than simulating them independently Tuning and Efficiency Metrics The paper proposes a strategy for tuning two parameters: δ, which controls the correlation between the proposed paths and the reference path in the conditional particle filter (smaller δ =⇒ more correlated), and M, the number of alternative proposed paths The strategy targets an optimal acceptance rate for the worst mixing state, which is not only valid for xPG, but can also be used to find the optimal number of particles M for any particle Gibbs algorithm The sample efficiency is assessed using a metric defined as Eff = minl,d ESS(Xd tl) / M Comparative Performance The paper compares xPG and particle Gibbs (PG) on data simulated from SIS, SIR, and SEIR models with univariate and multivariate state spaces In the four coupled SIR models, xPGBS achieves the highest sample efficiency for the worst mixing component of the state for PGBS being approximately seven times higher than for PGBS For SEIR models, xPGBS is only slightly better for the lowest sample efficiency of the model parameters for PG (η), xPG is 21 times better, and xPGBS is 76 times better Backward Sampling Extension The paper describes a method for incorporating backward sampling into both particle Gibbs and exchangeable particle Gibbs yielding particle Gibbs with backward sampling (PGBS) and exchangeable particle Gibbs with backward sampling (xPGBS) This method is only practicable for reaction networks in which the reaction counts are uniquely determined by a given pair of states, such as the SIR and SEIR models The tuning heuristics suggest that the optimal value of M, which maximizes sample efficiency is Mˆ = E The corresponding optimal acceptance rate is αˆ:= exp(−1) ≈ 0.368 In practice, to maximise efficiency over (δ, M) we restrict attention to the curve αMδ ≈ αˆ using the following procedure The optimal value of M for xPGBS are substantially smaller than those for xPG However, both xPG and xPGBS attain their optima at relatively small values of δ For the SIR and SEIR models, the minimum ESS over all latent states and observation time points does not necessarily occur at the first observation time The tuning strategy may overestimate the number of particles required Further Research Directions An xPG method based on exact simulation of MJPs would remove the bias caused by discretisation and forced boundaries and is a promising direction for future research The paper concludes that xPG is also the first algorithm to make exact joint inference of the latent states and rate parameters of the five-dimensional autoregulatory model of Golightly and Wilkinson (2005) from a sequence of exact state observations feasible The paper is based on the tau-leap approximation of MJPs Smaller τ gives a more accurate approximation but increases computational cost Moreover, boundary conditions, such as all the states being non-negative must be enforced when applying xPG and xPGBS The paper is 2611.05232.

Tom: <ref:2611.05232#pg5> Inference in stochastic reaction-network models, such as the susceptible-exposed-infectious-recovered (SEIR) epidemic model, is crucial for understanding the dynamics of interacting systems in epidemiology, ecology, and systems biology <ref:2611.05232#pg2> This paper focuses on improving the efficiency of particle Gibbs for inference on reaction networks by addressing the degeneracy problem <ref:2611.05232#pg3> The proposed method retains the exchangeability framework of xPG while adapting it to the structural and computational challenges specific to reaction networks <ref:2611.05232#pg3> Reaction networks are commonly used to describe dynamic processes in chemistry, epidemiology, ecology and biology <ref:2611.05232#pg3> They can be represented by Markov jump processes (MJPs), which are continuous-time discrete-state-space stochastic processes <ref:2611.05232#pg3> The tau-leap discretisation of the MJPs is a popular approximation of the evolution of an MJP <ref:2611.05232#pg3> Mathematical Formulation The mathematical form of a time-homogeneous HMM in continuous time is given by X0 ∼ p0(· θ), X(tl−1,tl] (x

0,tl−1: , y1:l−1, θ) ∼ p(· xtl−1, θ), Yl (x

0,t: , y1:l-1, θ) ∼ g(· xtl, θ) <ref:2611.05232#pg3> The state updates according to Xt = Xt− + Γi where Γi = (bi1 − ai1,..., biu − aiu) T <ref:2611.05232#pg3> Particle Gibbs Mechanism In each iteration, it first generates M new paths based on the current reference path using the conditional particle filter <ref:2611.05232#pg3> Next, it updates the reference path by sampling a new path from these M + 1 paths including the current reference path <ref:2611.05232#pg3> Finally, it updates the model parameters according to their conditional distribution given the selected reference path <ref:2611.05232#pg3> Exchangeable Particle Gibbs Details The exchangeable particle Gibbs algorithm replaces the standard conditional particle filter with the exchangeable conditional particle filter <ref:2611.05232#pg4> The joint proposal density is defined as q˜(x(1:M)tx(0)t, x(0:M)t−τ) = q((M)(x(0:M)tx(0:M)t−τ))q(xtxt−τ) <ref:2611.05232#pg4> Backward Sampling Extension Backward sampling is an extension to particle Gibbs which seeks to counter particle degeneracy by allowing a point at time t to choose a different time t−1 ancestor <ref:2611.05232#pg3> The backward sampling step is performed immediately after step 6 in the particle Gibbs algorithm <ref:2611.05232#pg3> Tuning Strategy Summary The tuning strategy targets an optimal acceptance rate for the worst mixing state, which is not only valid for xPG, but can also be used to find the optimal number of particles M for any particle Gibbs algorithm <ref:2611.05232#pg4> The sample efficiency plots for the SIS models (Figure 1) show that picking δ and M following the tuning rule gives close to optimal sample efficiency <ref:2611.05232#pg4> Numerical Results Summary The numerical experiments show that xPG attains the maximum ESS for X4 at time zero when δ = 0.05, whereas xPG with δ = 0.5 fails to move X4 from its initial value <ref:2611.05232#pg4> The results suggest that for the autoregulatory model, xPG achieves higher sample efficiency than nMESA <ref:2611.05232#pg4> For the SIR and SEIR models, adding backward sampling to particle Gibbs improves sample efficiency, but much less than using xPGBS <ref:2611.05232#pg4> Discussion Summary The benefit of xPG over PG is most pronounced when the observations are highly informative <ref:2611.05232#pg3> In contrast, when the observations are less informative, standard particle Gibbs may already produce a reasonably large proportion of particles with high weights <ref:2611.05232#pg3> The paper is 2611.05232.

Jane: References Andrieu, C., A.

Tom: Doucet, and R.

Title and authors: Tom: So we're looking at this paper, "Scalable Bayesian inference for time-discretised Markov jump processes via exchangeable particle Gibbs." The authors are Yang, Chapman, and Sherlock. It's essentially about making particle Gibbs more efficient when dealing with these reaction networks.

Jane: That sounds like it deals with those complex systems in epidemiology or chemistry where you have discrete steps over time, right? What’s the big picture here for someone listening who isn't deep in stochastic processes?

Lu: It tackles the problem of particle degeneracy which is a major hurdle when using these methods on reaction networks because most of your generated paths just get thrown away after a few resampling steps.

Meng: Particle degeneracy sounds like when you try to track millions of possible paths, but only a few actually survive the filtering process. That's bad for running models at scale, right?

Lalam: Exactly. The paper introduces something called exchangeable particle Gibbs which tries to fix that by changing how it generates new paths relative to the historical path instead of treating them as completely independent random walks.

The paper's summary: Tom: So, what the authors are doing is developing this xPG method, which keeps the exchangeability structure but adapts it specifically for reaction networks where randomness comes from Poisson processes rather than those standard Brownian motions we usually see in other models.

Jane: That makes sense because reaction networks often deal with counts of individuals or molecules, which naturally lead to those Poisson distributions they mention. So, instead of assuming continuous movement like in some SDEs, they are working with discrete state spaces driven by jumps.

Lu: The core idea is to correlate the initial state and the Brownian motions of the proposed paths with equivalent quantities in a reference path, rather than simulating them completely independently. That correlation helps keep those important historical trajectories alive during resampling.

Meng: If you're running a production model, that sounds like an improvement because it means less wasted computation when you're trying to infer parameters for something like an SEIR model. It’s about improving the mixing of the algorithm itself.

Lalam: They also looked at backward sampling and showed that the exchangeable version with backward sampling, called xPGBS, is actually the most efficient of four algorithms they tested on those reaction networks.

The paper's improvements: Tom: The paper proposes a tuning strategy for two parameters, delta and M. Delta controls how correlated the proposed paths are with the reference path—smaller delta means more correlated paths—and M is the number of alternative proposed paths you generate in each run.

Jane: So they aren't just throwing a black box method at us; they’re giving us concrete ways to tune these settings to get better results, specifically targeting an optimal acceptance rate for the worst mixing state.

Lu: They found that this tuning strategy actually helps find the optimal number of particles M for any particle Gibbs algorithm, which is pretty powerful because it gives you a rule instead of just guessing.

Meng: From an engineering standpoint, if we can use this to find the right M without running through massive sweeps over all settings, that's a win for deployment. It means we can control the computational cost more predictably.

Lalam: They showed that when you follow this tuning rule, like picking delta and M based on those acceptance rates, you get close to optimal sample efficiency on data simulated from SIS, SIR, and SEIR models with both univariate and multivariate state spaces.

Conclusion: Tom: So to wrap up, the main thing is that xPGBS is the most efficient algorithm they tested for those reaction networks, achieving sample efficiency gains over standard particle Gibbs. They also found that for some models like SEIR, adding backward sampling to just particle Gibbs isn't as helpful as using the full xPGBS method.

Jane: What this means for us is that when you’re working with these types of discrete state-space models, especially those with reaction counts, you should look at xPGBS first if your goal is maximum efficiency.

Lu: They also pointed out some practical limitations. For instance, the optimal number of particles M they found for xPGBS are substantially smaller than what they found for standard particle Gibbs, and the best performance sometimes happens when delta is quite small.

Meng: And one thing they mentioned is that for SIR and SEIR models, the minimum efficiency doesn't always happen at time zero; it can occur at a later observation time point. So you can't just assume the first step is always the best spot to look for maximum efficiency.

Lalam: And finally, they conclude that this method is also the first one to allow exact joint inference of the latent states and rate parameters for a five-dimensional autoregulatory model using only exact state observations, which is a big deal for those complex kinetic models.

Lanya Yang, Lloyd A. C. Chapman, Chris Sherlock

Lancaster University

stat.CO

Submitted: 2026-11-03

Updated: 2026-11-03

Comments: 21 pages + 19 pages appendix. 6 figures and tables (+ 8 in appendix). Code and data available at: https://github.com/yanglanya980929/xPGTauLeap

Code: https://github.com/yanglanya980929/xPGTauLeap

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 83/100

The gist: Inference in stochastic reaction-network models, such as the susceptible-exposed-infectious-recovered (SEIR) epidemic model…

Key concepts

Particle Gibbs (PG)
A Monte Carlo method used to estimate the posterior distribution of model parameters by simulating a sequence of paths through the state space. It involves generating new paths based on current ones and updating parameters using these samples.
Exchangeable Particle Gibbs (xPG)
An adaptation of PG that improves efficiency by correlating proposed paths with a reference path. This method is tailored for discrete-state reaction networks where randomness is driven by Poisson processes, overcoming the degeneracy problem common in standard PG methods.
Particle Degeneracy
A problem in particle filtering where successive resampling steps discard most ancestor trajectories, leaving very few unique particles to represent the system's historical path. This severely limits the ability of standard MCMC methods to explore the full parameter space effectively.
Backward Sampling Extension (PGBS)
An extension that helps counter degeneracy by allowing a point at time t to choose a different ancestor from time t-1 instead of just using the immediate predecessor. This technique is effective for reaction networks where reaction counts are uniquely determined by state pairs.

Terminology

Summary

Inference in stochastic reaction-network models, such as the susceptible-exposed-infectious-recovered (SEIR) epidemic model, is crucial for understanding the dynamics of interacting systems in epidemiology, ecology, and systems biology This paper focuses on improving the efficiency of particle Gibbs for inference on reaction networks by addressing the degeneracy problem The proposed method retains the exchangeability framework of xPG while adapting it to the structural and computational challenges specific to reaction networks

The Core Methodology

The paper develops a novel version of exchangeable particle Gibbs (xPG) tailored to discrete-state-space reaction networks, where randomness is driven by Poisson processes rather than Brownian motion This method is applied to the tau-leap discretisation of the MJPs The proposed method retains the exchangeability framework of xPG while adapting it to the structural and computational challenges specific to reaction networks

Addressing Particle Degeneracy

Particle degeneracy occurs when over successive resampling steps in the internal particle filter most of the ancestor trajectories get discarded leaving very few unique particles to represent the historical path of the system The paper introduces xPG, which is intended for discrete-time models with continuous state spaces The key idea of xPG applied to SDEs is to correlate both the initial state X0 and the Brownian motions of the proposed paths with the equivalent quantities in the reference path rather than simulating them independently

Tuning and Efficiency Metrics

The paper proposes a strategy for tuning two parameters: δ, which controls the correlation between the proposed paths and the reference path in the conditional particle filter (smaller δ =⇒ more correlated), and M, the number of alternative proposed paths The strategy targets an optimal acceptance rate for the worst mixing state, which is not only valid for xPG, but can also be used to find the optimal number of particles M for any particle Gibbs algorithm The sample efficiency is assessed using a metric defined as Eff = minl,d ESS(Xd tl) / M

Comparative Performance

The paper compares xPG and particle Gibbs (PG) on data simulated from SIS, SIR, and SEIR models with univariate and multivariate state spaces In the four coupled SIR models, xPGBS achieves the highest sample efficiency for the worst mixing component of the state for PGBS being approximately seven times higher than for PGBS For SEIR models, xPGBS is only slightly better for the lowest sample efficiency of the model parameters for PG (η), xPG is 21 times better, and xPGBS is 76 times better

Backward Sampling Extension

The paper describes a method for incorporating backward sampling into both particle Gibbs and exchangeable particle Gibbs yielding particle Gibbs with backward sampling (PGBS) and exchangeable particle Gibbs with backward sampling (xPGBS) This method is only practicable for reaction networks in which the reaction counts are uniquely determined by a given pair of states, such as the SIR and SEIR models

Conclusion on Tuning

The tuning heuristics suggest that the optimal value of M, which maximizes sample efficiency is Mˆ = E The corresponding optimal acceptance rate is αˆ:= exp(−1) ≈ 0.368 In practice, to maximise efficiency over (δ, M) we restrict attention to the curve αMδ ≈ αˆ using the following procedure The optimal value of M for xPGBS are substantially smaller than those for xPG However, both xPG and xPGBS attain their optima at relatively small values of δ For the SIR and SEIR models, the minimum ESS over all latent states and observation time points does not necessarily occur at the first observation time The tuning strategy may overestimate the number of particles required

Further Research Directions

An xPG method based on exact simulation of MJPs would remove the bias caused by discretisation and forced boundaries and is a promising direction for future research The paper concludes that xPG is also the first algorithm to make exact joint inference of the latent states and rate parameters of the five-dimensional autoregulatory model of Golightly and Wilkinson (2005) from a sequence of exact state observations feasible The paper is based on the tau-leap approximation of MJPs Smaller τ gives a more accurate approximation but increases computational cost Moreover, boundary conditions, such as all the states being non-negative must be enforced when applying xPG and xPGBS The paper is 2611.05232.

<ref:2611.05232#pg5>

Inference in stochastic reaction-network models, such as the susceptible-exposed-infectious-recovered (SEIR) epidemic model, is crucial for understanding the dynamics of interacting systems in epidemiology, ecology, and systems biology <ref:2611.05232#pg2> This paper focuses on improving the efficiency of particle Gibbs for inference on reaction networks by addressing the degeneracy problem <ref:2611. p. 0

Where the paper says it

Particle Markov chain Monte Carlo (PMCMC) offers a natural mechanism for inference on these hidden Markov models, but, even in moderate dimensional state spaces, particle degeneracy typically leads to poor mixing of the PMCMC algorithm. This paper focuses on improving the efficiency of particle Gibbs for inference on reaction networks by addressing the degeneracy problem.

05232#pg3> The proposed method retains the exchangeability framework of xPG while adapting it to the structural and computational challenges specific to reaction networks <ref:2611. p. 0
Where the paper says it

We focus on the τ -leap discretisation of the MJPs. The proposed method retains the exchangeability framework of xPG while adapting it to the structural and computational challenges specific to reaction networks.

05232#pg3>

Background Context

Reaction networks are commonly used to describe dynamic processes in chemistry, epidemiology, ecology and biology <ref:2611.05232#pg3> They can be represented by Markov jump processes (MJPs), which are continuous-time discrete-state-space stochastic processes <ref:2611. p. 1

Where the paper says it

A PREPRINT 1 Introduction Reaction networks are commonly used to describe dynamic processes in chemistry, epidemiology, ecology and biology. Their evolution can be represented by Markov jump processes (MJPs), which are continuous-time discrete-state stochastic processes.

05232#pg3> The tau-leap discretisation of the MJPs is a popular approximation of the evolution of an MJP <ref:2611. p. 4
Where the paper says it

An MJP may be simulated exactly using Gillespie’s direct method (Gillespie, 1977). We focus on a popular approximation of the evolution of an MJP: the tau-leap discretisation (Gillespie, 2001), given in Algorithm 1.

05232#pg3>

Mathematical Formulation

The mathematical form of a time-homogeneous HMM in continuous time is given by X0 ∼ p0(· θ), X(tl−1,tl] (x[0,tl−1], y1:l−1, θ) ∼ p(· xtl−1, θ), Yl (x[0,t], y1:l-1, θ) ∼ g(· xtl, θ) <ref:2611.05232#pg3> The state updates according to Xt = Xt− + Γi where Γi = (bi1 − ai1,. p. 3

Where the paper says it

If an event of reaction Ri occurs at time t, the change of the number of species Xj is bij − aij , j = 1, . The state updates according to Xt = Xt− + Γi , where Xt− denotes the state immediately before t, and Γ i = (bi1 − ai1, .

.., biu − aiu) T <ref:2611. p. 3
Where the paper says it

The state updates according to Xt = Xt− + Γi , where Xt− denotes the state immediately before t, and Γ i = (bi1 − ai1, . , biu − aiu) T .

05232#pg3>

Particle Gibbs Mechanism

In each iteration, it first generates M new paths based on the current reference path using the conditional particle filter <ref:2611.05232#pg3> Next, it updates the reference path by sampling a new path from these M + 1 paths including the current reference path <ref:2611. p. 5

Where the paper says it

These M paths are the proposals of the latent process. Next, it updates the reference path by sampling a new path from these M + 1 paths including the current reference path.

05232#pg3> Finally, it updates the model parameters according to their conditional distribution given the selected reference path <ref:2611. p. 5
Where the paper says it

Next, it updates the reference path by sampling a new path from these M + 1 paths including the current reference path. Finally, it updates the model parameters according to their conditional distribution given the selected reference path.

05232#pg3>

Exchangeable Particle Gibbs Details

The exchangeable particle Gibbs algorithm replaces the standard conditional particle filter with the exchangeable conditional particle filter <ref:2611.05232#pg4> The joint proposal density is defined as q˜(x(1:M)tx(0)t, x(0:M)t−τ) = q((M)(x(0:M)tx(0:M)t−τ))q(xtxt−τ) <ref:2611. p. 7

Where the paper says it

, x (M) t are sampled from a joint proposal density q˜ conditional on x (0:M) t−τ and x (0) t . The joint proposal density is defined as q˜(x (1:M) t |x (0) t , x (0:M) t−τ ) = q (M) (x (0:M) t |x (0:M) t−τ ) q(x (0) t |x (0) t−τ ) (6) where q (M) (·|x (0:M) t−τ ) is a joint conditional density being exchangeable for any permutation of the M + 1 pairs {(x (0) t , x (0) t−τ ), .

05232#pg4>

Backward sampling is an extension to particle Gibbs which seeks to counter particle degeneracy by allowing a point at time t to choose a different time t−1 ancestor <ref:2611.05232#pg3> The backward sampling step is performed immediately after step 6 in the particle Gibbs algorithm <ref:2611. p. 9

Where the paper says it

Let B˜ ℓ denote the index of the selected particle xℓ. Within each iteration of the particle Gibbs algorithm, the backward sampling step is performed immediately after step 6.

05232#pg3>

Tuning Strategy Summary

The tuning strategy targets an optimal acceptance rate for the worst mixing state, which is not only valid for xPG, but can also be used to find the optimal number of particles M for any particle Gibbs algorithm <ref:2611.05232#pg4> The sample efficiency plots for the SIS models (Figure 1) show that picking δ and M following the tuning rule gives close to optimal sample efficiency <ref:2611. p. 1

Where the paper says it

1 explains why we choose this to quantify sample efficiency. In Section 4, we compare the sample efficiency of xPG with PG on data simulated from SIS, susceptible-infectious-recovered (SIR), and susceptible-exposed-infectious-recovered (SEIR) models with univariate and multivariate state spaces, and data simulated from the gene autoregulatory model from Golightly and Wilkinson (2005). p. 1

Where the paper says it

1 explains why we choose this to quantify sample efficiency. In Section 4, we compare the sample efficiency of xPG with PG on data simulated from SIS, susceptible-infectious-recovered (SIR), and susceptible-exposed-infectious-recovered (SEIR) models with univariate and multivariate state spaces, and data simulated from the gene autoregulatory model from Golightly and Wilkinson (2005).

05232#pg4>

Numerical Results Summary

The numerical experiments show that xPG attains the maximum ESS for X4 at time zero when δ = 0.05, whereas xPG with δ = 0. p. 1

Where the paper says it

These approaches modify the conditional particle filter, also known as conditional SMC (CSMC), within particle Gibbs to mitigate the particle degeneracy issue. However, i-RW-CSMC, particle-MALA, and particle-mGRAD all require an analytical form of the transition density in the HMM, whereas xPG does not.

5 fails to move X4 from its initial value <ref:2611. p. 4
Where the paper says it

2: Set t = 0. Start from the initial value x0 = (x0,1, .

05232#pg4> The results suggest that for the autoregulatory model, xPG achieves higher sample efficiency than nMESA <ref:2611. p. 1
Where the paper says it

1 explains why we choose this to quantify sample efficiency. In Section 4, we compare the sample efficiency of xPG with PG on data simulated from SIS, susceptible-infectious-recovered (SIR), and susceptible-exposed-infectious-recovered (SEIR) models with univariate and multivariate state spaces, and data simulated from the gene autoregulatory model from Golightly and Wilkinson (2005).

05232#pg4> For the SIR and SEIR models, adding backward sampling to particle Gibbs improves sample efficiency, but much less than using xPGBS <ref:2611. p. 9
Where the paper says it

3 Backward Sampling for exchangeable particle Gibbs We now describe an approach for incorporating backward sampling into both particle Gibbs and exchangeable particle Gibbs for MJPs, yielding particle Gibbs with backward sampling (PGBS) and exchangeable particle Gibbs with backward sampling (xPGBS), respectively. This method is only practicable for reaction networks in which the reaction counts are uniquely determined by a given pair of states, such as the SIR and SEIR models.

05232#pg4>

Discussion Summary

The benefit of xPG over PG is most pronounced when the observations are highly informative <ref:2611.05232#pg3> In contrast, when the observations are less informative, standard particle Gibbs may already produce a reasonably large proportion of particles with high weights <ref:2611. p. 1

Where the paper says it

By correlating the proposed particles with those in the reference path, xPG generates particles with higher weights. As a result, it is less likely to resample the same particle excessively, thereby mitigating particle degeneracy.

05232#pg3> The paper is 2611. p. 0
Where the paper says it

Keywords: particle MCMC, particle Gibbs, Markov jump process, particle filter, reaction networks. arXiv:2611.

05232.

References

Andrieu, C., A. Doucet, and R. Holenstein (2010, May). p. 1

Where the paper says it

Particle Markov chain Monte Carlo (PMCMC) algorithms (Andrieu et al. , 2010) enable inference on HMMs.

Particle markov chain monte carlo methods. p. 0
Where the paper says it

These models are typically represented as Markov jump processes (MJPs), with discrete, noisy observations and intractable likelihoods. Particle Markov chain Monte Carlo (PMCMC) offers a natural mechanism for inference on these hidden Markov models, but, even in moderate dimensional state spaces, particle degeneracy typically leads to poor mixing of the PMCMC algorithm.

Journal of the Royal Statistical Society Series B: Statistical Methodology 72(3), 269–342 <ref:2611.05232#pg2> Corenflos, A. p. 1
Where the paper says it

In addition, they only address the particle degeneracy issue caused by long time series; the issue persists when the dimensionality of the state space is high (Finke and Thiery, 2023). Exchangeable particle Gibbs (xPG) (Malory, 2021), iterated random-walk conditional sequential Monte Carlo (i-RWCSMC) (Finke and Thiery, 2023) and Particle-MALA and Particle-mGRAD (Corenflos and Finke, 2024) have been introduced to improve the sample efficiency of particle Gibbs in moderate and high-dimensional state spaces.

and A. Finke (2024). p. 1
Where the paper says it

In addition, they only address the particle degeneracy issue caused by long time series; the issue persists when the dimensionality of the state space is high (Finke and Thiery, 2023). Exchangeable particle Gibbs (xPG) (Malory, 2021), iterated random-walk conditional sequential Monte Carlo (i-RWCSMC) (Finke and Thiery, 2023) and Particle-MALA and Particle-mGRAD (Corenflos and Finke, 2024) have been introduced to improve the sample efficiency of particle Gibbs in moderate and high-dimensional state spaces.

Particle-MALA and Particle-mGRAD: Gradient-based MCMC methods for high-dimensional state-space models. p. 1
Where the paper says it

In addition, they only address the particle degeneracy issue caused by long time series; the issue persists when the dimensionality of the state space is high (Finke and Thiery, 2023). Exchangeable particle Gibbs (xPG) (Malory, 2021), iterated random-walk conditional sequential Monte Carlo (i-RWCSMC) (Finke and Thiery, 2023) and Particle-MALA and Particle-mGRAD (Corenflos and Finke, 2024) have been introduced to improve the sample efficiency of particle Gibbs in moderate and high-dimensional state spaces.

arXiv preprint arXiv:2401. p. 0
Where the paper says it

Keywords: particle MCMC, particle Gibbs, Markov jump process, particle filter, reaction networks. arXiv:2611.

14868 <ref:2611. p. 0
Where the paper says it

Keywords: particle MCMC, particle Gibbs, Markov jump process, particle filter, reaction networks. arXiv:2611.

05232#pg4> Doucet, A., M. K. Pitt, G. Deligiannidis, and R. Kohn (2015, 06). Efficient implementation of markov chain monte carlo when using an unbiased likelihood estimator. p. 0
Where the paper says it

These models are typically represented as Markov jump processes (MJPs), with discrete, noisy observations and intractable likelihoods. Particle Markov chain Monte Carlo (PMCMC) offers a natural mechanism for inference on these hidden Markov models, but, even in moderate dimensional state spaces, particle degeneracy typically leads to poor mixing of the PMCMC algorithm.

Biometrika 102(2), 295–313 <ref:2611. p. 0
Where the paper says it

Keywords: particle MCMC, particle Gibbs, Markov jump process, particle filter, reaction networks. arXiv:2611.

05232#pg2> Fearnhead, P., C. Nemeth, C. J. Oates, and C. Sherlock (2025a). p. 0
Where the paper says it

chapman4@lancaster. uk Chris Sherlock School of Mathematical Sciences Lancaster University c.

Background <ref:2611. p. 3
Where the paper says it

A PREPRINT 2 Background 2. 1 Hidden Markov model The Hidden Markov model (HMM) is a mathematical framework modelling cases where there are noisy and/or partial observations of an underlying Markov process.

05232#pg5> Fearnhead, P., C. Nemeth, C. J. Oates, and C. Sherlock (2025b). p. 0
Where the paper says it

chapman4@lancaster. uk Chris Sherlock School of Mathematical Sciences Lancaster University c.

Reversible MCMC and its scaling <ref:2611. p. 0
Where the paper says it

Keywords: particle MCMC, particle Gibbs, Markov jump process, particle filter, reaction networks. arXiv:2611.

05232#pg5> Gillespie, D. p. 1
Where the paper says it

In this paper we develop xPG for HMMs whose latent processes are Markov jump processes (MJPs). We focus on the tau-leap method (Gillespie, 1977), a time discretisation of the continuous-time MJP.

T. (1977, December). p. 1
Where the paper says it

In this paper we develop xPG for HMMs whose latent processes are Markov jump processes (MJPs). We focus on the tau-leap method (Gillespie, 1977), a time discretisation of the continuous-time MJP.

Exact stochastic simulation of coupled chemical reactions <ref:2611.05232#pg3> Gillespie, D. p. 1
Where the paper says it

In this paper we develop xPG for HMMs whose latent processes are Markov jump processes (MJPs). We focus on the tau-leap method (Gillespie, 1977), a time discretisation of the continuous-time MJP.

T. (2001). Approximate accelerated stochastic simulation of chemically reacting systems <ref:2611. p. 0
Where the paper says it

sherlock@lancaster. uk ABSTRACT Inference in stochastic reaction-network models, such as the susceptible-exposed-infectious-recovered (SEIR) epidemic model, is crucial for understanding the dynamics of interacting systems in epidemiology, ecology, and systems biology.

05232#pg3> Golightly, A. p. 1
Where the paper says it

1 explains why we choose this to quantify sample efficiency. In Section 4, we compare the sample efficiency of xPG with PG on data simulated from SIS, susceptible-infectious-recovered (SIR), and susceptible-exposed-infectious-recovered (SEIR) models with univariate and multivariate state spaces, and data simulated from the gene autoregulatory model from Golightly and Wilkinson (2005).

and D. J. Wilkinson (2005). Bayesian inference for stochastic kinetic models using a diffusion approximation <ref:2611. p. 1
Where the paper says it

Compared to ordinary differential equations or stochastic differential equations (SDEs), MJPs provide a more realistic description of the evolution of such systems when there is a small number of individuals (or molecules) of each species (or reactant). Given a sequence of noisy observations of the number of individuals of a subset of the species at a discrete set of times, we focus on performing Bayesian inference for the parameters which control the Markov jump process.

05232#pg5> Lindsten, F. p. 1
Where the paper says it

One approach that has been introduced to tackle particle degeneracy is particle Gibbs with backward simulation (PGBS) (Whiteley, 2010; Whiteley et al. , 2010; Lindsten and Schön, 2013).

, M. I. Jordan, and T. B. Schön (2014, 06). p. 1
Where the paper says it

One approach that has been introduced to tackle particle degeneracy is particle Gibbs with backward simulation (PGBS) (Whiteley, 2010; Whiteley et al. , 2010; Lindsten and Schön, 2013).

Particle gibbs with ancestor sampling <ref:2611. p. 1
Where the paper says it

, 2010; Lindsten and Schön, 2013). This strategy significantly improves the sample efficiency of particle Gibbs, and was extended to particle Gibbs with ancestor sampling (PGAS) by Lindsten et al.

05232#pg4> Lindsten, F. p. 1
Where the paper says it

One approach that has been introduced to tackle particle degeneracy is particle Gibbs with backward simulation (PGBS) (Whiteley, 2010; Whiteley et al. , 2010; Lindsten and Schön, 2013).

and T. B. Schön (2013). p. 1
Where the paper says it

One approach that has been introduced to tackle particle degeneracy is particle Gibbs with backward simulation (PGBS) (Whiteley, 2010; Whiteley et al. , 2010; Lindsten and Schön, 2013).

Backward simulation methods for Monte Carlo statistical inference <ref:2611. p. 0
Where the paper says it

These models are typically represented as Markov jump processes (MJPs), with discrete, noisy observations and intractable likelihoods. Particle Markov chain Monte Carlo (PMCMC) offers a natural mechanism for inference on these hidden Markov models, but, even in moderate dimensional state spaces, particle degeneracy typically leads to poor mixing of the PMCMC algorithm.

05232#pg4> Malory, S. p. 1
Where the paper says it

In addition, they only address the particle degeneracy issue caused by long time series; the issue persists when the dimensionality of the state space is high (Finke and Thiery, 2023). Exchangeable particle Gibbs (xPG) (Malory, 2021), iterated random-walk conditional sequential Monte Carlo (i-RWCSMC) (Finke and Thiery, 2023) and Particle-MALA and Particle-mGRAD (Corenflos and Finke, 2024) have been introduced to improve the sample efficiency of particle Gibbs in moderate and high-dimensional state spaces.

J. (2021). Bayesian inference for stochastic processes. p. 1
Where the paper says it

Compared to ordinary differential equations or stochastic differential equations (SDEs), MJPs provide a more realistic description of the evolution of such systems when there is a small number of individuals (or molecules) of each species (or reactant). Given a sequence of noisy observations of the number of individuals of a subset of the species at a discrete set of times, we focus on performing Bayesian inference for the parameters which control the Markov jump process.

Ph. D thesis, Lancaster University <ref:2611. p. 0
Where the paper says it

SCALABLE BAYESIAN INFERENCE FOR TIME-DISCRETISED MARKOV JUMP PROCESSES VIA EXCHANGEABLE PARTICLE GIBBS A PREPRINT Lanya Yang School of Mathematical Sciences Lancaster University l. yang28@lancaster.

05232#pg5> Roberts, G. O., A. Gelman, and W. R. Gilks (1997). Weak convergence and optimal scaling of random walk Metropolis algorithms <ref:2611. p. 1
Where the paper says it

In addition, they only address the particle degeneracy issue caused by long time series; the issue persists when the dimensionality of the state space is high (Finke and Thiery, 2023). Exchangeable particle Gibbs (xPG) (Malory, 2021), iterated random-walk conditional sequential Monte Carlo (i-RWCSMC) (Finke and Thiery, 2023) and Particle-MALA and Particle-mGRAD (Corenflos and Finke, 2024) have been introduced to improve the sample efficiency of particle Gibbs in moderate and high-dimensional state spaces.

05232#pg5> Roberts, G. O. and J. S. Rosenthal (2001). p. 4
Where the paper says it

An MJP may be simulated exactly using Gillespie’s direct method (Gillespie, 1977). We focus on a popular approximation of the evolution of an MJP: the tau-leap discretisation (Gillespie, 2001), given in Algorithm 1.

Optimal scaling for various Metropolis-Hastings algorithms <ref:2611. p. 0
Where the paper says it

Keywords: particle MCMC, particle Gibbs, Markov jump process, particle filter, reaction networks. arXiv:2611.

05232#pg5> Sherlock, C. p. 0
Where the paper says it

chapman4@lancaster. uk Chris Sherlock School of Mathematical Sciences Lancaster University c.

p. 0
Where the paper says it

chapman4@lancaster. uk Chris Sherlock School of Mathematical Sciences Lancaster University c.

p. 0
Where the paper says it

chapman4@lancaster. uk Chris Sherlock School of Mathematical Sciences Lancaster University c.

, P. Fearnhead, and G. O. Roberts (2010, May). p. 1
Where the paper says it

Particle Markov chain Monte Carlo (PMCMC) algorithms (Andrieu et al. , 2010) enable inference on HMMs.

The random walk metropolis: Linking theory and practice through a case study <ref:2611. p. 1
Where the paper says it

In addition, they only address the particle degeneracy issue caused by long time series; the issue persists when the dimensionality of the state space is high (Finke and Thiery, 2023). Exchangeable particle Gibbs (xPG) (Malory, 2021), iterated random-walk conditional sequential Monte Carlo (i-RWCSMC) (Finke and Thiery, 2023) and Particle-MALA and Particle-mGRAD (Corenflos and Finke, 2024) have been introduced to improve the sample efficiency of particle Gibbs in moderate and high-dimensional state spaces.

05232#pg5> Sherlock, C. and A. Golightly (2023). p. 1
Where the paper says it

However, PGBS and PGAS are only usable when the transition densities of the MJP are tractable. In addition, they only address the particle degeneracy issue caused by long time series; the issue persists when the dimensionality of the state space is high (Finke and Thiery, 2023).

Exact bayesian inference for discretely observed markov jump processes using finite rate matrices <ref:2611. p. 0
Where the paper says it

SCALABLE BAYESIAN INFERENCE FOR TIME-DISCRETISED MARKOV JUMP PROCESSES VIA EXCHANGEABLE PARTICLE GIBBS A PREPRINT Lanya Yang School of Mathematical Sciences Lancaster University l. yang28@lancaster.

05232#pg5> Sherlock, C. and G. Roberts (2009). Optimal scaling of the random walk metropolis on elliptically symmetric unimodal targets <ref:2611. p. 1
Where the paper says it

In addition, they only address the particle degeneracy issue caused by long time series; the issue persists when the dimensionality of the state space is high (Finke and Thiery, 2023). Exchangeable particle Gibbs (xPG) (Malory, 2021), iterated random-walk conditional sequential Monte Carlo (i-RWCSMC) (Finke and Thiery, 2023) and Particle-MALA and Particle-mGRAD (Corenflos and Finke, 2024) have been introduced to improve the sample efficiency of particle Gibbs in moderate and high-dimensional state spaces.

05232#pg5> Tjelmeland, H. p. 7
Where the paper says it

, Z(M) ) ⊤ with distribution N(µM+1, ΣM+1) where µM+1 = 0(M+1)d and Σ ∈ R (M+1)d×(M+1)d is the Kronecker product of Id and the (M + 1) × (M + 1) matrix with 1s on the diagonal and with all off-diagonal terms equal to ρ, the correlation parameter. At the centre of the mechanism is the methodology of Tjelmeland (2004) for generating correlated standard Gaussian random variables.

(2004). Using all Metropolis–Hastings proposals to estimate mean values. p. 7
Where the paper says it

This meant that the paths were exchangeable with marginals given by their prior distribution. Here, we describe methodology for making the initial values and the Poisson increments over each time step exchangeable between the reference path and the proposals.

Technical Report Statistics No. 4/2004, Norwegian University of Science and Technology, Trondheim, Norway <ref:2611. p. 0
Where the paper says it

Keywords: particle MCMC, particle Gibbs, Markov jump process, particle filter, reaction networks. arXiv:2611.

05232#pg4> Whiteley, N. p. 1
Where the paper says it

We explain particle degeneracy in more detail in Section 2. One approach that has been introduced to tackle particle degeneracy is particle Gibbs with backward simulation (PGBS) (Whiteley, 2010; Whiteley et al.

(2010). Discussion on particle markov chain monte carlo methods. p. 0
Where the paper says it

These models are typically represented as Markov jump processes (MJPs), with discrete, noisy observations and intractable likelihoods. Particle Markov chain Monte Carlo (PMCMC) offers a natural mechanism for inference on these hidden Markov models, but, even in moderate dimensional state spaces, particle degeneracy typically leads to poor mixing of the PMCMC algorithm.

Journal of the Royal Statistical Society: Series B 72(3), 306–307 <ref:2611. p. 0
Where the paper says it

Keywords: particle MCMC, particle Gibbs, Markov jump process, particle filter, reaction networks. arXiv:2611.

05232#pg2> Whiteley, N. p. 1
Where the paper says it

We explain particle degeneracy in more detail in Section 2. One approach that has been introduced to tackle particle degeneracy is particle Gibbs with backward simulation (PGBS) (Whiteley, 2010; Whiteley et al.

, C. Andrieu, and A. p. 1
Where the paper says it

In other words, we consider Hidden Markov models (HMMs) whose latent processes are MJPs. Particle Markov chain Monte Carlo (PMCMC) algorithms (Andrieu et al.

Doucet (2010). p. 1
Where the paper says it

Particle Markov chain Monte Carlo (PMCMC) algorithms (Andrieu et al. , 2010) enable inference on HMMs.

Efficient bayesian inference for switching state-space models using discrete particle markov chain monte carlo methods. p. 0
Where the paper says it

These models are typically represented as Markov jump processes (MJPs), with discrete, noisy observations and intractable likelihoods. Particle Markov chain Monte Carlo (PMCMC) offers a natural mechanism for inference on these hidden Markov models, but, even in moderate dimensional state spaces, particle degeneracy typically leads to poor mixing of the PMCMC algorithm.

Technical Report 10:04, Bristol Statistics Research Report <ref:2611. p. 0
Where the paper says it

Keywords: particle MCMC, particle Gibbs, Markov jump process, particle filter, reaction networks. arXiv:2611.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Particle Gibbs with backward sampling (PGBS) can be used for inference on reaction networks where reaction counts are uniquely determined by a given pair of states, such as SIR and SEIR models, to counter particle degeneracy. This improves sample efficiency for the latent states and model parameters compared to standard particle Gibbs.

  2. Exchangeable particle Gibbs (xPG) mitigates particle degeneracy in discrete-state-space Markov jump processes like those arising from the tau leap discretization method by correlating proposed paths with the reference path. This allows inference on reaction networks where transition densities are intractable, such as those involving Poisson increments rather than Brownian motion.

  3. A tuning strategy can be provided to maximize sample efficiency by targeting an optimal acceptance rate for the worst mixing state, which is not only valid for xPG but can also be used to find the optimal number of particles M for any particle Gibbs algorithm. This includes setting M proportional to L and tuning delta based on empirical acceptance rates.

  4. The improved AI system can perform exact joint inference of latent states and rate parameters for models like the five-dimensional autoregulatory model from exact state observations, achieving a minimum ESS of 1570 across all eight model parameters. This capability surpasses alternative Bayesian inference approaches like nMESA for this specific model class.

  5. The system can dynamically adjust the correlation parameter delta based on the number of dimensions and observation time points to find an optimal trade-off between computational cost and mixing efficiency, as evidenced by contour plots showing how maximum sample efficiency is achieved at progressively smaller values of delta as dimension increases.

Abstract

Inference in stochastic reaction-network models, such as the susceptible-exposed-infectious-recovered (SEIR) epidemic model, is crucial for understanding the dynamics of interacting systems in epidemiology, ecology, and systems biology. These models are typically represented as Markov jump processes (MJPs), with discrete, noisy observations and intractable likelihoods. Particle Markov chain Monte Carlo (PMCMC) offers a natural mechanism for inference on these hidden Markov models, but, even in moderate dimensional state spaces, particle degeneracy typically leads to poor mixing of the PMCMC algorithm. This paper focuses on improving the efficiency of particle Gibbs for inference on reaction networks by addressing the degeneracy problem. Building on recent work on the exchangeable particle Gibbs (xPG) sampler for diffusions, this paper develops a novel version of xPG tailored to discrete-state-space reaction networks, where randomness is driven by Poisson processes rather than Brownian motion. We focus on the τ-leap discretisation of the MJPs. The proposed method retains the exchangeability framework of xPG while adapting it to the structural and computational challenges specific to reaction networks. The new method is applied to several examples, including four coupled SEIR models, where it substantially improves mixing efficiency compared with standard particle Gibbs. In this example, xPG improves sample efficiency by more than an order of magnitude relative to standard particle Gibbs, making inference feasible in hours rather than weeks. As part of our tuning strategy, we provide a mechanism for tuning the number of particles that is applicable more generally, to any particle Gibbs algorithm.

Sources

Related papers