Structural identifiability of partially-observed stochastic processes: from single-particle trajectories to total particle density data

summary

The gist

Structural identifiability analysis for partially-observed stochastic processes develops a methodology to distinguish parameter identifiability based on whether data comes from single-particle…

In short

The study develops a method to check if parameters in stochastic models can be uniquely determined when data is incomplete. It compares identifiability from single-particle trajectories versus total particle density measurements. The findings show that while parameters are identifiable from many individual paths, they are only locally identifiable from aggregate density data, depending critically on the type of observation available.

Key concepts

Structural Identifiability
This determines if the parameters within a mathematical model can be uniquely recovered from the idealised data generated by that model. It is a prerequisite for estimating parameters in real-world scenarios.
Single-Particle Trajectories
This involves tracking the path of individual particles over time. The authors use this type of data to show that all model parameters, including initial conditions, are structurally identifiable when an infinite number of these trajectories are available.
Total Particle Density Data
This involves measuring the collective evolution or distribution of all particles at a given time. This data is less informative because it only describes the overall behavior and cannot distinguish between the dynamics of individual particles or their internal states.

Terminology used across episodes

This episode discusses

The paper

Structural identifiability of partially-observed stochastic processes: from single-particle trajectories to total particle density data · Read on arXiv

Arianna Ceccarelli, Alexander P. Browning, Ruth E. Baker

Mathematical Institute, University of Oxford · School of Mathematics and Statistics, University of Melbourne

The increasing availability of experimental data has intensified interest in calibrating stochastic models, raising fundamental questions about parameter identifiability. Structural identifiability determines whether parameters can be uniquely recovered from idealised data, a prerequisite to allow for parameter estimation in real-world scenarios. However, there is a lack of methodologies generally applicable to analyse the structural identifiability of such stochastic processes. Therefore, we develop a methodology for a class of spatio-temporal stochastic processes. We investigate how structural identifiability depends on the type of available data, distinguishing between single-particle trajectories and total particle density measurements. For trajectory data, we use a particle-based description that explicitly represents single-particle dynamics. For population-level data, we derive a partial differential equation representation, that describes the evolution of total particle density, and apply a differential algebra approach, common to ordinary differential equation analysis. We further introduce Taylor-series-expansion and Fourier-transform approaches to study the initial condition. We apply our methodology to an example model with a specified class of initial conditions. We show that the model is globally structurally identifiable given infinitely many, infinitely long single-particle trajectories, but only locally structurally identifiable from total particle density data, demonstrating that parameter identifiability depends on the type of data available.

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: "Structural identifiability of partially-observed stochastic processes".

Jane: Here's the gist of the paper. Structural identifiability analysis for partially-observed stochastic processes develops a methodology to distinguish parameter identifiability based on whether data comes from single-particle trajectories or total particle density measurements The paper demonstrates that structural identifiability depends critically on the type of data available, showing that a model is structurally identifiable from single-particle trajectory data but not from total particle density data Data Modality Comparison The study investigates how structural identifiability depends on the type of available data, distinguishing between single-particle trajectories and total particle density measurements For trajectory data, a particle-based model description is used that explicitly represents single-particle dynamics Conversely, for population-level data, a partial differential equation model representation describes the evolution of total particle density The authors note that from an infinite number of single-particle trajectories, one can extract total particle density data by obtaining and summing the distributions of particle locations over time for each state However, total particle density data only describe the collective evolution and do not allow distinguishing between individual behaviors and internal states Identifiability from Trajectory Data When measuring single-particle trajectories, the model parameters are shown to be structurally identifiable up to state relabelling using a countably infinite number of single-particle trajectories Specifically, all model parameters including the initial condition are structurally identifiable when using infinitely long single-particle trajectories The authors show that the coefficients c1,..., c15 are structurally identifiable for any nonconstant function f(x) when analyzing the Taylor series expansion of total particle density about t = 0 Identifiability from Density Data When assessing structural identifiability using total particle density data, the differential algebra approach is applied to reduce the PDE system to input-output equations The coefficients c1,..., c8 are structurally identifiable when written in a monic form because they uniquely determine the observable behavior However, the probability parameters p12, p21, and p31 are not fully identifiable from these input-output equations by the Implicit Function Theorem Incorporating Initial Conditions The initial condition influences structural identifiability because it can contain information about model parameters not captured by the input-output equations The authors propose a novel Taylor-series-expansion approach to incorporate the initial condition into the structural identifiability analysis of PDE models This expansion yields additional identifiable parameter combinations, such as k4, when two states share the same velocity The paper concludes that the model parameters are structurally identifiable measuring single-particle trajectories but only locally identifiable measuring the total particle density evolution The authors also show that parameter equivalences arising under a certain initial condition may not hold under alternative initial conditions, emphasizing the link between experimental design and identifiability analysis The model parameters are globally structurally identifiable from single-particle trajectories and locally structurally identifiable from the total particle density when velocities are all distinct Future Directions Future work may extend this methodology to chemical reaction models with single-particle trajectory and total particle density data, particularly with relevant spatial components The authors suggest that experimental techniques might only detect stationary or slow-moving particles, leading to analogous identifiability properties for such data They also propose that varying the initial condition could be a practical strategy to improve parameter identifiability The authors also note that when two states share the same velocity, both types of data lead to structural non-identifiability Supplementary Information The supplementary information provides detailed analysis for specific cases, such as structural identifiability measuring single-particle trajectories if two states share the same velocity, and structural identifiability measuring the total particle density if two states share the same velocity The supplementary material also shows examples of equivalent model parametrisations and the impact of the initial condition with velocities all distinct The authors show that for model A with a = w, the only equivalent model is model B The results suggest that varying the initial condition may be used as a practical strategy to identify the model parameters The authors conclude that they developed a general methodology to analyse the structural identifiability of spatio-temporal stochastic processes Declarations The authors declare that they have no conflict of interest For the purpose of open access, the author has applied a CC BY public copyright licence to any author accepted manuscript arising from this submission The Python files, Mathematica notebook and Apple Grapher files are available on GitHub in the repository The CRediT author statement details the contributions of each author References The paper cites numerous references covering topics like parameter identifiability, structural identifiability, and specific model types such as Markov models and ODEs Key references include

1: Preston et al.,

2: Wieland et al., and various works on differential algebra approaches for ODE models The paper also cites works related to specific software tools like SIAN, GenSSI, and DAISY Supplementary Information Details The supplementary information includes derivations for the PDE system from the particle-based model It also contains the derivation of the stationary distribution in terms of switching rates and probability parameters The analysis of structural identifiability when two states share the same velocity is detailed in Section S3 and S4 The Taylor-series-expansion approach for incorporating the initial condition is presented in Section 4.2 The analysis of the rank of the matrix F when all velocities are distinct is proven in Section S6 The paper concludes by showing that for Gaussian functions, the matrix F has full rank, and thus coefficients are structurally identifiable The authors show that for model A with a = w, the only equivalent model is model B The results suggest that varying the initial condition could be used as a practical strategy to identify the model parameters The authors also show that for model A with a =

1, 0, 0: T there is exactly one equivalent model (model D) The paper demonstrates that the total density evolution is different between models A and C despite having equal input-output coefficients c1,..., c9 with initial condition a = w in each case The authors conclude that the model parameters are globally structurally identifiable from single-particle trajectories and locally structurally identifiable from the total particle density when velocities are all distinct References The paper cites

1: Preston et al.,

2: Wieland et al.,

3: Walter, E.,

4: Cobelli et al., and many others related to identifiability analysis of dynamical systems and stochastic processes The references also include works on specific software tools like SIAN, GenSSI, and DAISY The paper cites works related to specific model types such as Markov models and ODEs It also includes references for mathematical analysis techniques like differential algebra and Taylor series expansions The references cover both biological systems and physical transport phenomena The paper cites works related to continuous-time Markov chains and Fokker-Planck equations The references also include works on numerical methods for solving PDEs and ODEs The references cover both mathematical analysis techniques and specific model applications in biology and physics Supplementary Information The supplementary information contains derivations for the PDE system from the particle-based model It also contains the derivation of the stationary distribution in terms of switching rates and probability parameters The analysis of structural identifiability when two states share the same velocity is detailed in Section S3 and S4 The Taylor-series-expansion approach for incorporating the initial condition is presented in Section 4.2 The analysis of the rank of the matrix F when all velocities are distinct is proven in Section S6 The authors show that for Gaussian functions, the matrix F has full rank, and thus coefficients are structurally identifiable The supplementary information also shows numerical solutions for the PDE model with periodic boundary conditions The paper concludes by showing that for Gaussian functions, the matrix F has full rank, and thus the coefficients are structurally identifiable The authors also show that for model A with a = w, the only equivalent model is model B The results suggest that varying the initial condition could be used as a practical strategy to identify the model parameters The authors also show that for model A with a =

1, 0, 0: T there is exactly one equivalent model (model D) The paper demonstrates that the total density evolution is different between models A and C despite having equal input-output coefficients c1,..., c9 with initial condition a = w in each case The authors conclude that the model parameters are globally structurally identifiable from single-particle trajectories and locally structurally identifiable from the total particle density when velocities are all distinct Supplementary Information The supplementary information contains derivations for the PDE system from the particle-based model It also contains the derivation of the stationary distribution in terms of switching rates and probability parameters The analysis of structural identifiability when two states share the same velocity is detailed in Section S3 and S4 The Taylor-series-expansion approach for incorporating the initial condition is presented in Section 4.2 The analysis of the rank of the matrix F when all velocities are distinct is proven in Section S6 The authors show that for Gaussian functions, the matrix F has full rank, and thus coefficients are structurally identifiable The supplementary information also shows numerical solutions for the PDE model with periodic boundary conditions The paper concludes by showing that for Gaussian functions, the matrix F has full rank, and thus coefficients are structurally identifiable The authors also show that for model A with a = w, the only equivalent model is model B The results suggest that varying the initial condition could be used as a practical strategy to identify the model parameters The authors also show that for model A with a =

1, 0, 0: T there is exactly one equivalent model (model D) The paper demonstrates that the total density evolution is different between models A and C despite having equal input-output coefficients c1,..., c9 with initial condition a = w in each case The authors conclude that the model parameters are globally structurally identifiable from single-particle trajectories and locally structurally identifiable from the total particle density when velocities are all distinct References The paper cites

1: Preston et al.,

2: Wieland et al.,

3: Walter, E.,

4: Cobelli et al.

Paper summary: Tom: Okay, so to wrap up the summary of "Structural identifiability of partially-observed stochastic processes: from single-particle trajectories to total particle density data", the authors show that structural identifiability depends critically on the type of data available.

Jane: And it claims that a model is structurally identifiable from single-particle trajectory data but only locally identifiable when you have total particle density measurements. That distinction is really key for understanding what information we actually get from different experiments.

Lu: For trajectory data, they use a particle-based description that explicitly represents the dynamics of each individual particle, while for population-level data, they use a partial differential equation representation describing the evolution of the total particle density.

Meng: So if I’m building something in production and I only have density measurements, this paper suggests my model might be structurally identifiable only locally. That means there are infinite sets of parameters that could produce the same overall density curve, which sounds like a headache for calibration.

Lalam: The authors also introduce techniques like Taylor-series expansion and Fourier-transform approaches to study the initial condition, showing how those things influence whether we can identify parameters in these PDE models.

Tom: So, what this means for the world is that when you are trying to build a model of something dynamic—say particle movement—you have to be very careful about how you measure it. You can get complete parameter identification from tracking individual particles, but density measurements only give you a partial picture of the collective evolution.

Jane: That’s the big picture for us; it suggests that experimental design needs to match what we want to know. If we only collect density data, we have to be accepting that some parameters will remain unidentifiable.

Lu: And they even show examples, using a specific model with a certain class of initial conditions, where the model is globally identifiable from infinite single-particle trajectories but only locally identifiable from total particle density data.

Meng: That’s interesting because it shows the theoretical limit based on the input. If you can't get that infinite set of trajectory data, you're stuck with local identification for density measurements. What kind of real-world systems are we talking about here?

Conclusion: Tom: So, concluding this paper "Structural identifiability of partially-observed stochastic processes: from single-particle trajectories to total particle density data" brings us back to the central finding that parameters are structurally identifiable when measuring single-particle trajectories but only locally identifiable when measuring the total particle density evolution.

Jane: This means that the type of experimental data you collect directly dictates what kind of parameters you can actually figure out about the underlying process. It’s a direct link between measurement and model structure.

Lu: The paper emphasizes that this dependence on data type is not just theoretical; it shows how different hidden-state dynamics can produce the same observed particle motion or density evolution, making structural identifiability non-trivial for partially-observed processes.

Meng: From a production standpoint, if we’re aiming for robust model calibration based only on density measurements, we have to be very cautious because those probability parameters might not be fully identifiable from the input output equations.

Lalam: They also point out that parameter equivalences arising under certain initial conditions might not hold if you change the initial condition, which highlights how sensitive the results are to experimental setup choices.

Tom: Overall, this work gives us a general methodology to analyze structural identifiability for these spatio-temporal stochastic processes, showing that what you measure matters profoundly when trying to understand the underlying physics or biology.

More episodes

← Home