A perturbation theory for Feynman-Kac models and their particle interpretations

arXiv:2610.21587 · stat.CO, math.PR · Submitted 2026-10-15 · 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: "A perturbation theory for Feynman-Kac models and their particle interpretations".

Jane: Here's the gist of the paper. A perturbation theory is developed for Feynman–Kac models, describing how local errors introduced at each step of the Feynman–Kac recursion propagate over time.

Tom: Under general conditions on the potential functions and the Markov transitions, which hold for a large class of state-space models, it is proven that the global error for any time horizon T at least 1 is bounded by C delta T (1 + delta-1T), where delta T is the largest local error.

Jane: This bound holds both for the prediction and the updated models.

Tom: These conditions require no Lyapunov function and no stability of the Markov transitions, and allow for non-compact state spaces, with confinement provided by potential functions.

Paper summary: Tom: So we just talked about the general idea of bounding the global error over time using that C times Delta T log(one plus inverse Delta T) term in "A perturbation theory for Feynman-Kac models and their particle interpretations."

Jane: That bound applies to both the prediction and updated models, which is a nice piece of consistency for whatever we're modeling.

Lu: The paper sets up the groundwork by defining these primitive conditions on the model, specifically condition A0, which ensures there's a sequence Wt such that the transition operator Qt plus one applied to something in Wt plus one is positive for any starting point x in Wt Meng That condition is what allows them to define the mappings Psi t and Phi t plus one on PW t E, where PW t E is this set of measures where something positive happens on Wt Jane Essentially, they're defining how these measures evolve through time using these operators which are built from the transition kernel Gt and the operator Mt plus one Tom The key takeaway here is that as long as you have those primitive conditions holding for a large class of state-space models, you get this error bound regardless of whether you use total variation or Kolmogorov norm estimates Lu They also mention that these perturbation bounds can be extended to more general Markov transitions and stochastic differential equations under some relaxed conditions, like P1' and P2' Jane This means the machinery they developed isn't just for one specific type of model; it has broader applicability in handling model misspecification too.

Meng: When I think about deploying this, I worry about the complexity of checking those primitive conditions on real traffic data, but having a bound that depends on Delta T is better than nothing Tom It’s still an abstract analysis for now, but it gives us a framework to expect how errors behave in these recursive processes.

Conclusion: Jane: So we’ve covered the main points of "A perturbation theory for Feynman-Kac models and their particle interpretations," which is about getting time-uniform bounds on global error using that Delta T log(one plus inverse Delta T) factor.

Tom: The title itself suggests this is a perturbation approach, which means we're analyzing small changes or local errors at each step of the Feynman-Kac recursion as they build up over time.

Lu: What this paper really gives us is a robust mathematical tool that works for a wide variety of state-space models without needing things like Lyapunov functions or stable Markov transitions Jane It confirms that for many complex systems, you can still get useful error bounds by just looking at the local error size, Delta T Meng It applies this to concrete algorithms like Kalman filters and Euler schemes, showing how the error scales based on step sizes or other parameters Tom And in terms of what it changes for us as people who use these models, it provides a foundation for expecting certain error behaviors when we run things at scale Lu Ultimately, this theory gives us confidence that even if the system isn't perfectly stable or your observations are not perfect, the errors introduced locally will have a predictable growth rate over time Jane We're looking at how this affects our ability to deploy these methods reliably in real-world scenarios where things aren't perfectly clean.

Centre de Recherche Inria Bordeaux Sud-Ouest · University of Bristol

stat.CO, math.PR

Submitted: 2026-10-15

Updated: 2026-11-12

Comments: 84 pages (including a 33-page Appendix)

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 98/100

The gist: A perturbation theory is developed for Feynman–Kac models, describing how local errors introduced at each step of the Feynman–Kac recursion propagate over time.

Key concepts

Perturbation Theory
A mathematical method used to describe how small changes or errors introduced at each step of a recursive process (like the Feynman–Kac recursion) propagate and grow over time. It allows researchers to bound the total accumulated error based on the initial local errors.
Feynman–Kac Models
A class of models used in stochastic processes that are related to solutions of partial differential equations. They are fundamental in finance and physics, often involving Markov transitions and potential functions, which the theory analyzes for error propagation.
Kolmogorov Norm (K)
A specific mathematical norm used to measure the distance or error between probability measures (like those describing particle distributions). The paper introduces a dual norm, $|| u||_{VE}$, which is equivalent to this Kolmogorov norm under certain conditions, helping to bound errors in these measures.

Terminology

Summary

A perturbation theory is developed for Feynman–Kac models, describing how local errors introduced at each step of the Feynman–Kac recursion propagate over time. p. 0 Under general conditions on the potential functions and the Markov transitions, which hold for a large class of state-space models, it is proven that the global error for any time horizon T at least 1 is bounded by C delta T (1 + delta-1T), where delta T is the largest local error. p. 0 This bound holds both for the prediction and the updated models. p. 0 These conditions require no Lyapunov function and no stability of the Markov transitions, and allow for non-compact state spaces, with confinement provided by potential functions. p. 0 The perturbation bounds hold in total variation and in the Kolmogorov norm; for the latter, it extends the Hardy–Krause variation and the Koksma–Hlawka inequality to unbounded domains. p. 0 The theory is applied to derive time-uniform bounds for linearized Kalman filters, Euler schemes for partially observed diffusions and nudged state-space models, as well as high-probability error bounds, uniform over whole trajectories, for particle filters and mean field particle methods such as sequential Monte Carlo and diffusion Monte Carlo algorithms. p. 0

Theorem A states that under certain conditions on the potential functions (Section 4. p. 43), there exists a constant C eta 1 such that for any sequence mu = (mu t) t at least 1 with mu t in P W t(E), and any T at least 1, the estimate 1 at most t at most T mu t - eta t, 1 at most t at most T mu b t - eta b t at most C eta 1 delta T, mu, | times|.

Theorem B provides a time-uniform estimate in total variation distance when comparing two Feynman–Kac models, showing that t at least 1 eta t - eta t tv, t at least 1 eta b t - eta b t tv at most C eta 1 Q Q. p. 7

Theorem D provides high-probability error bounds for particle filters in the Kolmogorov norm: there exists a constant C eta 1 such that for any q in (0, 1], any N at least 1 and any T at least 1, with probability at least 1 - q, 1 at most t at most T eta N t - eta t K, 1 at most t at most T eta b N t - eta b t K at most C eta 1 epsilon N,T, q. p. 7

Section 6.2 details the analysis for linearized Kalman filters, showing that the error depends on the nonlinearity of F measured by kappa F, and provides time-uniform estimates based on this parameter. p. 6 Section 6.3 covers Euler approximations of partially observed diffusions, yielding bounds dependent on the step size l. p. 0 Section 6.4 analyzes nudged state-space models, deriving error bounds that depend on a parameter chi related to the nudging strength. p. 0 Theorem 7.3 provides high-probability trajectory bounds for particle filters under conditions (A3) and (B1). p. 7 The results show that the time-uniform estimates presented are new for linearized Kalman filters, Euler approximations of partially observed diffusions, and nudged state-space models. p. 0

The paper introduces a dual norm, times VE, which is equivalent to the Kolmogorov norm on P(E) under suitable conditions. p. 7 Theorem 3.9 establishes that mu 1 - mu 2 VE is a norm on M(E) satisfying mu 1 - mu 2 K at most mu 1 - mu 2 VE at most mu 1 - mu 2 K for any mu 1, mu 2 in P(E). p. 0 The paper explores the connection between total variation and Kolmogorov norms through the Koksma–Hlawka inequality, extending it to unbounded hyper-rectangles and introducing the norm times VE. p. 9 Theorem 3.4 shows that mu 1(f) - mu 2(f) at most VE(f) mu 1 - mu 2 K for any f in BHK(E). p. 5 Theorem 3.9 further establishes a norm on M(E), denoted nu VE, satisfying mu 1 - mu 2 VE at most mu 1 - mu 2 K for any mu 1, mu 2 in P(E). p. 1

The section on state-space models verifies the primitive conditions (A3) and (B1) for a class of models. p. 9 Proposition 5. p. 68 shows that under certain assumptions, these conditions hold for a representation where the potential functions are modified by terms related to the orthogonal projection onto observation spaces. p. 0 The analysis extends to more general Markov transitions and stochastic differential equations, showing that perturbation bounds hold under conditions (P1') and (P2'), which allow for model misspecification.

The paper demonstrates that for exponential potential functions and linear observation spaces, the assumptions (A3) and (B1) hold if condition (41) is satisfied, even when observations are unbounded. p. 0 The results of Section 4 apply to this representation. p. 4 For partially observed stochastic differential equations, conditions (D1)–(D3) ensure that the model satisfies (P2'), and perturbation bounds hold under the assumption that t at least 1 y t+1 - y t < infinity.

The paper concludes by showing that for particle filters, high-probability trajectory bounds exist in Theorem 7. p. 03, which depend on epsilon N,T, q = r/1 + (T /q)N. p. 4 The results apply to a broader class of algorithms including those based on stratified resampling and Baker resampling and mean field particle models defined by conditionally independent particles. p. 0

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. Error propagation control for general Feynman-Kac models: The system can derive global error bounds for any time horizon T, showing that global error is bounded by C deltaT log(1 + delta-1T), where deltaT is the largest local error, even under general conditions on potential functions and Markov transitions.

  2. Application to state-space models: The AI system can use these bounds to derive time-uniform error estimates for linearized Kalman filters, Euler schemes for partially observed diffusions, and nudged state-space models with model misspecification.

  3. High probability trajectory error bounds for particle filters: The system can compute high probability error bounds for particle filter estimates uniformly over whole trajectories, providing a bound of the form C eta1 ϑ epsilonN,T q.

  4. Comparison between models: The AI can compare two Feynman-Kac models by using Theorem B to establish time-uniform estimates between them based on the local discrepancy DeltaQ|Qstar.

  5. Robustness to model misspecification: The system can handle model misspecification in state-space representations by using nudged state-space models, controlling the dependence on the nudging parameter chi.

  6. Norm equivalence and metric selection: The system can choose between total variation and Kolmogorov norms for error measurement, leveraging the relationship established by Theorem 3.9 that nuVE is a norm equivalent to nuK on P(E).

  7. Handling unbounded domains: The system can apply these perturbation bounds to non-compact state spaces, as the conditions do not require Lyapunov functions or stability of Markov transitions, allowing for confinement provided by potential functions.

Sources

Related papers