Probabilistic Exponential Integrators

Probabilistic solvers provide a flexible and efficient framework for simulation, uncertainty quantification, and inference in dynamical systems. However, like standard solvers, they suffer performance penalties for certain stiff systems, where small steps are required not for reasons of numerical accuracy but for the sake of stability. This issue is greatly alleviated in semi-linear problems by the probabilistic exponential integrators developed in this paper. By including the fast, linear dynamics in the prior, we arrive at a class of probabilistic integrators with favorable properties. Namely, they are proven to be L-stable, and in a certain case reduce to a classic exponential integrator -- with the added benefit of providing a probabilistic account of the numerical error. The method is also generalized to arbitrary non-linear systems by imposing piece-wise semi-linearity on the prior via Jacobians of the vector field at the previous estimates, resulting in probabilistic exponential Rosenbrock methods. We evaluate the proposed methods on multiple stiff differential equations and demonstrate their improved stability and efficiency over established probabilistic solvers. The present contribution thus expands the range of problems that can be effectively tackled within probabilistic numerics.

Paper

Similar papers

Peer review

Reviewer Lu6G8/10 · confidence 4/52023-07-01

Summary

This paper proposes a probabilistic integrator for solving stiff semi-linear ODEs that decouples linear and nonlinear components for improved accuracy. The approach uses an integrated Ornstein-Uhlenbeck process (IOUP) prior and posterior inference extends classical exponential integrator to exactly solve the linear components. Nonlinearities are approximated via local linearization similar to the extended Kalman filter (EKF). The authors present L-stability results and address computation through proposed approximations of the Jacobian that capture only the linear dynamics.

Strengths

This is a solid paper; well-written, well-motivated, nice execution. The results aren’t groundbreaking when compared to baseline, but there are clear benefits of the proposed method, particularly when nonlinearities are less significant. Furthermore, results are presented honestly and limitations are clearly identified.

Weaknesses

The paper fails to highlight any benefit of treating the initial value problem (IVP) as a Bayesian inference task. The posterior distribution *should* characterize uncertainty over the solution, and thus it is possible to quantify and report confidence in the solution. It is further possible to scrutinize the appropriateness of the IOUP as a prior. The paper does not address this, but instead prioritizes numerical accuracy and execution time. While these metrics are certainly important, they are not the primary motivation of probabilistic numerics. This is not a critical flaw, but perhaps a missed opportunity to demonstrate the benefit of a Bayesian approach to a numerical method. Unsurprisingly, the proposed method is fairly sensitive to nonlinearity of the ODEs. The experiment in Sec. 4.1 demonstrates that accuracy degrades as the quadratic term becomes more significant. In fact, the proposed method seems to be more sensitive to this than baseline IWP methods, albeit with uniformly lower error. For the simple ODE in Sec. 4.1 all methods achieve very low error (on the order of $10^{-10}$) so the observed improvement with the IOUP is marginal unless the ODE is highly linear. Some detailed comments: * Fig. 5 should reference subfigures (e.g. (a) is never explicitly mentioned). * L268: Typo "Since th problem..."

Questions

* The authors suggest that probabilistic integrators may be preferable for expensive-to-evaluate functions (L259). How? Can the method avoid explicit evaluation via prediction? * The EKF can be quite unreliable in even simple nonlinear dynamical systems, leading to preferred approaches such as the UKF. Would not such an approach be preferable here to avoid the local linear approximations? What is the main challenge in a more accurate approximation?

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

4 excellent

Contribution

3 good

Limitations

Limitations are clearly described in Sec. 5

Reviewer V9iy7/10 · confidence 1/52023-07-06

Summary

This paper presents probabilistic exponential integrators as a solver for stiff semi-linear ODEs.Their solver can also be extended to solve general non-linear ODEs with iterative re-linearization. The proposed method is shown to be L-stable in theory and empirically more stable than existing methods.

Strengths

The paper is well written. It tackles the problem of solving stiff systems, which is a great challenge in probabilistic numerics. The authors exploit the properties in semi-linear ODEs and develop a probabilistic solver that integrates the fast, linear dynamics into the prior model. The solver is also extended to general nonlinear ODEs via iterative linearization. Both theoretical and empirical analysis are sound.

Weaknesses

The main difficulty of conventional of stiff ODEs solvers is its sensitivity to step size: small step size is required for stability. Proposition 3 shows the proposed solver is L-stable. However, the experiments mainly focus on final errors across various step sizes. A detailed investigation of stability (boundedness of solutions?) behaviors of solvers under large step size may help readers better understand the proposed approach.

Questions

See Weakness

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

1: Your assessment is an educated guess. The submission is not in your area or the submission was difficult to understand. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

See Weakness

Reviewer vChc6/10 · confidence 3/52023-07-17

Summary

This paper proposed probabilistic exponential integrators, which is a new class of probabilistic solvers for stiff semi-linear ODEs. More specifically, the integrated OU process is introduced for a functional prior that directly incorporate the linear part of the dynamics. As a result, the proposed methods can be more stable than their non-exponential probabilistic counterparts. In addition, the authors also provide an extension for general non-linear systems, which can be viewed as a probablistic solver based on exponential Rosenbrock-type methods.

Strengths

1. The paper is overall clearly written and nicely organized. 2. The IOUP prior is neat. Also, some theoretical results of the proposed probabilistic exponential integrator have been established (e.g., equivalence to the classic exponential trapezoidal rule and L-stability).

Weaknesses

1. The proposed methods are only for semi-linear ODEs, while other baselines can be applied to more general cases. 2. The IOUP prior is more expensive than the IWP prior. 3. The Kalman filter/smoother based solver has cubic scaling in the ODE dimension. 4. It seems that the advantage of the method is more significant when the step size is large (error is large), and becomes negligible when the step size is small. Note that 2 and 3 have been adimitted by the authors.

Questions

1. What is the overhead computational time for computing the matrix exponentials? Is it contained in the runtime? 2. The proposed method can be more accurate when the linear part is more dominant. Since EK1 also introduces a linear approximation, would it perform better when combined with the IOUP prior?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Yes, they do.

Reviewer goRD5/10 · confidence 4/52023-07-24

Summary

The paper studies classical exponential integrators using the framework of probabilistic numerics, where Bayesian formalism is used to study numerical approximations of deterministic dynamical systems. The main result seems to be Proposition 1, which establishes exponential stability (called L-Stability for some reason) property for the exponential integrator.

Strengths

The paper reasonably well introduces probabilistic numerics and studies this in the context of ODEs and tackles important notion of stability of numerical approximation.

Weaknesses

The paper transfers well-understood properties of numerical approximations of ODEs using the formalism of probabilistic numerics. In my view, this is not a sufficiently novel contribution.

Questions

- Calling the Ornstein-Uhlenbeck process a Gauss-Markov in eqn 3 is overkill. - Notation in eqn 5 of normal distribution not clear what. Why three parameters? - Section 2.2 I don't see why \mathcal I [y] = 0 when y is a solution to ODE. This says that y_t-f(y_t,t)=0 not \frac{d}{dt} y_t -f(y_t,t)=0 - Explain the role of Z in 9b - Paper seems to treat ODEs, but some numerical examples are derived for PDEs. Of course, semi-group theory for PDEs links with exponential integrators, but this seems to go beyond this paper.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

Methodology applies to a particular class of ODES

Reviewer Lu6G2023-08-11

Thanks. Your responses cleared up a couple points. I will keep my scores as-is and am willing to argue in favor of the paper.

Reviewer vChc2023-08-16

Thanks for the reply

Thanks for the response. I will keep my score.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC