Adaptive Finite State Projection with Quantile-Based Pruning for Solving the Chemical Master Equation

TL;DR

The paper addresses the computational challenge of solving the Chemical Master Equation (CME) for stochastic biochemical networks by introducing an adaptive Finite State Projection (FSP) framework with quantile-based pruning. The method combines time stepping with dynamic state-space truncation and Krylov subspace techniques to approximate matrix exponentials, while enforcing rigorous error control that links pruned probability mass to the global solution error. The authors derive bounds showing pruning error remains bounded in the $\ell^1$-norm and does not amplify through time evolution, and they provide an explicit error allocation strategy balancing pruning and time-stepping errors. They demonstrate effectiveness on Lotka–Volterra, Michaelis–Menten, and stochastic toggle-switch models, achieving accurate mean trajectories and full probability distributions with substantially smaller state spaces and computational cost than conventional approaches. This work delivers a practical, scalable framework for analyzing high-dimensional stochastic networks with evolving state spaces, offering precise error guarantees and enabling efficient exploration of complex dynamical regimes.

Abstract

We present an adaptive Finite State Projection (FSP) method for efficiently solving the Chemical Master Equation (CME) with rigorous error control. Our approach integrates time-stepping with dynamic state-space truncation, balancing accuracy and computational cost. Krylov subspace methods approximate the matrix exponential, while quantile-based pruning controls state-space growth by removing low-probability states. Theoretical error bounds ensure that the truncation error remains bounded by the pruned mass at each step, which is user-controlled, and does not propagate forward in time. Numerical experiments on biochemical systems, including the Lotka-Volterra and Michaelis-Menten and bi-stable toggle switch models.

Figures (6)

• Figure 1: Partitioned master equation transitions between the truncated set $S$ and the remainder $R$. Here, $\mathbf{A}_{SS}$ and $\mathbf{A}_{RR}$ represent transitions within $S$ and $R$, respectively, while $\mathbf{A}_{SR}$ and $\mathbf{A}_{RS}$ represent transitions between them.
• Figure 1: Comparison of mean trajectories for the Lotka--Volterra model: FSP with quantile-based pruning vs. SSA (1000 realizations). The two methods produce nearly identical results.
• Figure 1: Local truncation error (top row) and the size of the truncated state space (bottom row) for the three benchmark models. The results confirm that (i) the local error remains bounded by $2m$, where $m$ is the total mass pruned at each step, and (ii) the state space dynamically adapts to the evolving probability distribution, efficiently capturing relevant states while discarding negligible ones.
• Figure 2: A single iteration of the adaptive FSP procedure for the Lotka--Volterra model. The top row displays the active states at each stage (before expansion, after time evolution, and after pruning), while the bottom row shows the corresponding probability distribution $p(x,t)$.
• Figure 3: Comparison of mean trajectories for the Michaelis--Menten system: FSP with quantile-based pruning vs. SSA (1000 realizations). The FSP solution closely matches the SSA ensemble average, confirming its accuracy.

Theorems & Definitions (22)

• Definition 2.1: Chemical Reaction Network
• Definition 2.2: Stoichiometric Matrix
• Definition 2.3: System State
• Definition 2.4: State Space
• Definition 2.5: Probability Distribution over State Space
• Definition 2.6: Chemical Master Equation (CME)
• Definition 2.7: Finite State Projection
• Remark 3.1: Handling Ties
• Remark 3.2: Renormalization
• Remark 3.3