Diffusion Monte Carlo (DMC) represents Feynman–Kac expectations by branching particle systems; when the Feynman–Kac exponential contains a stochastic integral, the naïve fine-time-step limit has an exploding particle population. Hairer and Weare's ticketed variant (Algorithm TDMC) repairs this, and their companion analysis of the continuous-time limit proves convergence only in a one-dimensional toy setting (zero drift, unit diffusion coefficient, linear potential), identifying the limit as a new object, the Brownian fan: a recursive Poisson point process of Itô excursions. They heuristically derive a conjectural limit for general multidimensional diffusions with nonlinear potentials, with offspring intensity proportional to the squared gradient of the potential in the diffusion metric (prefactor 1/2), and they show that the standard generator-based identification route fails and that the limit has a dense set of exceptional times with infinitely many particles. The open problem is to establish or refute this conjectured convergence theory beyond the toy model. Even within the toy case the source paper records unresolved limitations in passing — the proven rate exponents below 1/32 are explicitly non-optimal, the generator-based identification route fails, and the integrability assumptions on the step law are expected to be weakenable.

Background

Diffusion Monte Carlo (DMC) is a Monte Carlo technique, used in contexts ranging from quantum chemistry to rare-event simulation and particle filtering, that represents expectations arising from Feynman–Kac formulae — expressions of the form E[exp⁡(∫0tG(Xs) ds) f(Xt)]E[\exp(\int_0^t G(X_s)\,ds)\,f(X_t)] for a Markov process XX — by an ensemble of weighted particles ("walkers") that undergo branching (birth and death) events. When the Feynman–Kac exponential contains a stochastic integral with respect to the driving noise, as happens in rare-event simulation and continuous-time filtering, the branching intensity per time step diverges as the step size ε\varepsilon tends to zero, so the naïve fine-time-step limit of DMC has an exponentially exploding particle count and no finite limiting object. Hairer and Weare introduced a modified algorithm, ticketed DMC (Algorithm TDMC), analysed in their companion paper: each particle carries a "ticket" θ∈(0,1)\theta\in(0,1); upon a proposed move from xx to yy one forms P=eV(x)−V(y)P=e^{V(x)-V(y)} for a potential VV; if P<θP<\theta the particle is killed, otherwise it is replaced by max⁡{⌊P+u⌋,1}≥1\max\{\lfloor P+u\rfloor,1\}\ge 1 particles at the new location, the first inheriting the rescaled ticket θ/P\theta/P and the others drawing fresh tickets uniformly from (P−1,1)(P^{-1},1). Increases of VV thus thin the population, while decreases of VV — the events in which the Feynman–Kac weight grows — spawn offspring.

The present problem originates in the second paper, which studies the continuous-time limit of TDMC as the time step ε→0\varepsilon\to 0. In the simplest possible situation — a one-dimensional random walk approximating a driftless Brownian motion with unit diffusion coefficient, biased by a linear potential — the authors prove (Theorem 5.1 in the arXiv version) that the empirical particle measures converge weakly in C([0,1],ℓp(M))C([0,1],\ell^p(\mathcal M)) for p∈(0,1]p\in(0,1], where M\mathcal M is a space of (position, tag) pairs and ℓp(M)\ell^p(\mathcal M) is the space of point configurations with finite pp-th moments, possibly of infinite total mass. The limiting Markov process is named the Brownian fan: loosely, each particle performs a Brownian motion until it hits a killing barrier determined by its tag, and gives rise to offspring whose trajectories are drawn from a Poisson point process built on Itô's Brownian excursion measure QQ (the σ-finite measure, standard in the excursion theory of Brownian motion, describing infinitesimal excursions away from a boundary point), with intensity a dt⊗Qa\,dt\otimes Q; this mechanism repeats recursively. Section 2.1 of the paper derives heuristically a conjectural generalization: for a multidimensional diffusion dyt=F(yt) dt+Σ(yt) dBtdy_t=F(y_t)\,dt+\Sigma(y_t)\,dB_t with smooth drift FF, diffusion coefficient Σ\Sigma, and smooth nonlinear potential VV, the offspring of a particle performing an excursion ww of lifetime τ\tau should form a Poisson point process of intensity ∫0τA(wt) Θt⋆Qwt dt\int_0^\tau A(w_t)\,\Theta_t^\star Q_{w_t}\,dt, where Θt\Theta_t is the time shift, Qx=lim⁡δ→0+δ−1Qx,V(x)+δQ_x=\lim_{\delta\to 0^+}\delta^{-1}Q_{x,V(x)+\delta} is an Itô-type excursion measure for the diffusion at the level set of VV, and the offspring intensity is

A(x)=12⟨∇V(x), Σ(x)Σ(x)⊤∇V(x)⟩.A(x)=\frac{1}{2}\big\langle\nabla V(x),\ \Sigma(x)\Sigma(x)^{\top}\nabla V(x)\big\rangle.

A naïve computation gives the wrong prefactor 1/41/4; the correction to 1/21/2 arises because the relevant offspring are born at distance ε\sqrt\varepsilon of the killing boundary, where the excursion-measure approximation is being used at a scale δ∼ε\delta\sim\sqrt\varepsilon outside its range of validity. The value 1/21/2 is proved only in the toy case.

Three features discovered in the toy case indicate that the general theory, if true, is subtle. First, the standard route to identifying a Markovian limit — convergence of generators — fails: computing the generator of the Brownian fan on multiplicative test functions F(μ)=exp⁡(⟨log⁡f,μ⟩)F(\mu)=\exp(\langle\log f,\mu\rangle), where ⟨g,μ⟩\langle g,\mu\rangle denotes integration of a function gg against the point measure μ\mu, yields a branching term (a/2)f′(x)(a/2)f'(x), whereas the rescaled one-step operators of the discrete algorithm converge to the same expression with a different constant involving ∫y2 ν(dy)\int y^2\,\nu(dy), where ν\nu is the step law; the discrete generators converge to a generator with wrong parameters, and natural observables such as the total particle count NtN_t do not even belong to the generator's domain. Second, although NtN_t has finite exponential moments at every deterministic time, there is a dense set of exceptional times at which Nt=∞N_t=\infty (indeed times at which the first generation of offspring alone is infinite), so the choice of state space and topology on possibly-infinite point configurations is delicate. Third, the moment assumptions on the step law ν\nu are nearly sharp: one-sided exponential integrability of ν\nu is necessary for any moment bounds on NtN_t, and even Gaussian step laws do not yield exponential moments of NtN_t uniformly in ε\varepsilon.

Related but distinct lines of work frame the problem. The Brownian fan is formally a special case of Hutzenthaler's virgin island model — a branching structure on trees of excursions — but with a constant branching rate that violates the standing assumption of that framework, and Hutzenthaler's convergence results for interacting diffusions concern a different limit (number of islands tending to infinity). Ferré and Stoltz proved weak error estimates of Talay–Tubaro type — expansions of the time-discretization bias in powers of the step size, in the tradition of the classical analysis of Euler discretizations of diffusions — for time-discretized Feynman–Kac semigroups, citing the Hairer–Weare convergence results as related work but addressing the discretization of the semigroup rather than the branching-particle limit of the algorithm. Alternative approaches such as fast randomized iteration (Lim and Weare) — a numerical-linear-algebra method that represents the Feynman–Kac action as a random sparse matrix–vector product with a controlled population size — and resampling-matrix formulations (Webber), which recast the resampling step of sequential Monte Carlo as matrix multiplication, avoid the branching explosion by abandoning the walker interpretation altogether.

Problem Statement

Establish, or disprove, the convergence of the ticketed diffusion Monte Carlo algorithm (Algorithm TDMC of Hairer and Weare) to its conjectured continuous-time limit beyond the one-dimensional linear-potential toy model. Precisely: consider the Euler-type discretization y(k+1)ε=ykε+εF(ykε)+ε Σ(ykε) ξk+1y_{(k+1)\varepsilon}=y_{k\varepsilon}+\varepsilon F(y_{k\varepsilon})+\sqrt\varepsilon\,\Sigma(y_{k\varepsilon})\,\xi_{k+1} of a multidimensional diffusion with smooth drift FF, diffusion coefficient Σ\Sigma, and smooth potential VV, where the step law ν\nu of the independent identically distributed noise vectors ξk\xi_k has mean zero and identity covariance; run Algorithm TDMC with time step ε\varepsilon on this chain and let μtε\mu^\varepsilon_t denote the resulting empirical particle measures. Prove or refute that, under a near-minimal integrability assumption on ν\nu, the processes με\mu^\varepsilon converge as ε→0\varepsilon\to 0, in a suitable path space of possibly-infinite point configurations, to the recursive Poisson point process of excursions conjectured in Section 2.1 of the source paper, whose offspring carry intensity

A(x)=12⟨∇V(x), Σ(x)Σ(x)⊤∇V(x)⟩,A(x)=\frac{1}{2}\big\langle\nabla V(x),\ \Sigma(x)\Sigma(x)^{\top}\nabla V(x)\big\rangle,

The convergence is to be established for a class of smooth drifts FF, diffusion matrices Σ\Sigma, and potentials VV with non-degenerate Σ\Sigma and ∇V≠0\nabla V\ne 0, so that the level-set excursion measure QxQ_x is well defined.

References

  1. Martin Hairer and Jonathan Weare, “The Brownian Fan”, Communications on Pure and Applied Mathematics 68(1), 1–60 (2015). DOI: 10.1002/cpa.21544; arXiv:1404.2928; https://doi.org/10.1002/cpa.21544
  2. Martin Hairer and Jonathan Weare, “Improved Diffusion Monte Carlo”, Communications on Pure and Applied Mathematics 67(12), 1995–2021 (2014). DOI: 10.1002/cpa.21526; arXiv:1207.2866; https://doi.org/10.1002/cpa.21526
  3. Grégoire Ferré and Gabriel Stoltz, “Error estimates on ergodic properties of discretized Feynman–Kac semigroups”, Numerische Mathematik 143(2), 261–313 (2019). DOI: 10.1007/s00211-019-01059-1; arXiv:1712.04013; https://doi.org/10.1007/s00211-019-01059-1
  4. Grégoire Ferré, “Théorie des grandes déviations en physique statistique : quelques aspects théoriques et numériques” (English title: “Large deviations theory in statistical physics: some theoretical and numerical aspects”), doctoral thesis, Université Paris-Est (2019). NNT: 2019PESC1035; DOI: 10.70675/ca209240zc41dz4386zaa48z09d0441ad7aa; https://pastel.hal.science/tel-02884212
  5. Martin Hutzenthaler, “The Virgin Island Model”, Electronic Journal of Probability 14, 1117–1161 (2009). DOI: 10.1214/ejp.v14-646; https://doi.org/10.1214/ejp.v14-646
  6. Martin Hutzenthaler, “Interacting diffusions and trees of excursions: convergence and comparison”, Electronic Journal of Probability 17, Paper No. 71 (2012). DOI: 10.1214/EJP.v17-2278; arXiv:1104.1060; https://doi.org/10.1214/EJP.v17-2278
  7. Nicolas Chopin and Omiros Papaspiliopoulos, “Advanced Topics and Open Problems”, in An Introduction to Sequential Monte Carlo, Springer Series in Statistics, Springer, 371–376 (2020). DOI: 10.1007/978-3-030-47845-2_19; https://doi.org/10.1007/978-3-030-47845-2_19
  8. Lek-Heng Lim and Jonathan Weare, “Fast randomized iteration: diffusion Monte Carlo through the lens of numerical linear algebra” (2015). arXiv:1508.06104; https://arxiv.org/abs/1508.06104
  9. Samuel M. Greene, Robert J. Webber, Jonathan Weare, and Timothy C. Berkelbach, “Beyond Walkers in Stochastic Quantum Chemistry: Reducing Error using Fast Randomized Iteration” (2019). arXiv:1905.00995; https://arxiv.org/abs/1905.00995
  10. Robert J. Webber, “Unifying Sequential Monte Carlo with Resampling Matrices” (2019). arXiv:1903.12583; https://arxiv.org/abs/1903.12583
  11. Zhizhang Wu, Renaud Raquépas, Jack Xin, and Zhiwen Zhang, “Computing large deviation rate functions of entropy production for diffusion processes by an interacting particle method” (2024). arXiv:2403.19223; https://arxiv.org/abs/2403.19223
  12. Bastiaan J. Braams, “A Diffusion Monte Carlo algorithm employing depth first traversal and a stack instead of a swarm” (2026). arXiv:2606.08946; https://arxiv.org/abs/2606.08946