WorksMathematics
Convergence of replicator-mutator particle systems via bounded Wasserstein distance
Abstract
Interacting particle systems with selection and mutation mechanisms arise in biological modeling, evolutionary optimization, and particle filtering methods. In this work we will consider the paradigmatic replicator-mutator PDE, which combines a diffusive component with a source term that depends on an external potential. We derive quantitative error estimates for the distance between the empirical measure of an unweighted particle system and the solution of the PDE. The proof uses a Wasserstein-1 distance with a bounded ground cost, which allows us to treat unbounded potentials on an unbounded state space. The key tools are the optimal coupling of newly resampled particles and, for the time-continuous limit, the use of a discrete-time killing resampling.
Introduction and main result
Complex systems where different interacting particles undergo birth and death dynamics appear in several contexts. In biology, they describe the evolution of a population of cells which reproduce and mutate [23]. In Bayesian statistics, they constitute a large class of particle filter algorithms, particularly effective for multi-modal targets, under the name Sequential Monte Carlo [12]. In optimization, such dynamics is at the basis of genetic algorithms [20], a popular heuristic involving particle exploration to find global minimizers of non-differentiable functions.
All such systems can be described at two different levels. One is the microscopic level of the finite particle systems. One is the statistical level of an evolving probability distribution typically obtained by taking the many particle limit . In this paper we are interested in rigorously studying the relation between these two descriptions with quantitative approximation errors. We focus on maybe the simplest of such models, that is, the Replicator-Mutator (RM) PDE which describes particles undergoing birth or death according to a fitness potential , and diffusion with strength :
Algorithm 1: Replicator-mutator particle dynamics
Input: Potential function $V \ge 0$, population size $N$, initial distribution $\rho_0 \in \mathcal{P}(\mathbb{R}^d)$, diffusion coefficient $\sigma \ge 0$, time discretization $\tau \in (0,1]$, time horizon $T > 0$.
Particle initialization $X_0^i \sim \rho_0$ i.i.d., $i=1,\ldots,N$
$n=0$
while $n\tau \le T$ do
Compute weights $w_n^i = e^{-\tau V(X_n^i)}/\sum_{j=1}^N e^{-\tau V(X_n^j)}$, $i=1,\ldots,N$
for all $i=1,\ldots,N$ do
Sample $\xi_n^i \sim \mathcal{N}(0,I_d)$
with probability $1-e^{-\tau V(X_n^i)}$:
sample $J_n^i \sim \operatorname{Categorical}(w_n^1,\ldots,w_n^N)$
$X_{n+1}^i = X_n^{J_n^i}+\sigma\sqrt{\tau}\xi_n^i$
else:
$X_{n+1}^i = X_n^i+\sigma\sqrt{\tau}\xi_n^i$
end for
$n=n+1$
end while
The constant ensures the total mass remains constant during the evolution. The RM appears in mathematical biology for the description of cell dynamics [23], in machine learning as the grazing collision limit of collective hyperparameter learning dynamics [Borghi et al., 2026]. It can also be regarded as the time-continuous limit of a bootstrap particle filter [6]. From a variational point of view, while (1) is not a gradient flow, it is comprised of two gradients: the Laplacian corresponds to the -Wasserstein gradient of the log-entropy, while the source term corresponds to the Fisher–Rao gradient of the external energy . Interestingly, the two energies together form the KL divergence to , up to a constant.
Contribution
We are interested in connecting (1) and the simple time-discrete particle discretization described by Algorithm 1. At every step , particles may undergo independent resampling according to their potential value following a discrete-time killing resampling mechanism [8], also known as -bootstrap interaction [28]. Gaussian increments are added at every step to discretize the Laplacian operator. Let be the time-step. We will show that, as the number of particles increases, the particle distribution converges to the splitting scheme made by alternating distribution reweighting and convolution with the heat kernel
leading to the time discrete update
To study the convergence of the particle system we consider a Wasserstein-1 distance with bounded ground cost:
Note that, by the Kantorovich–Rubinstein duality [29], it corresponds to the Bounded Lipschitz norm between measures (see (3)). Since , one can use the usual Wasserstein concentration inequalities [17] for i.i.d. particles for the -Wasserstein distance
with the rate given by
when . Our objective is to show that the rate is preserved at later times, and that the time-discretization introduces an error of order .
Assumption 1.1. Let , , such that there exist
Theorem 1.2. Given a time horizon , let be a measure solution to (1) (see Definition 4.1) with finite -th moments, . Let be the empirical measure of the particle system constructed according to Algorithm 1 with time step .
If the potential satisfies 1.1, it holds
with the positive constant depending on and on the -th moment of .
The propagation of chaos result is obtained via the coupling method [7], by considering an auxiliary system of i.i.d. particles , with . Interestingly, the dependence on the dimension already appears in the coupling estimate:
(see Proposition 3.2) where mean-field type models, instead, typically exhibit an interaction error of order . This is due to the fact that in Algorithm 1 the interaction requires, at each step, a resampling procedure. This may explain the empirically observed struggle of discretizing particle filters in high dimension due to the weight collapse phenomenon [1, 24]. Convergence of the time-discretization is more standard and is obtained via the Feynman–Kac representation of solutions to (1), as in [14].
Related literature
Evolution of probabilities involving selection and mutation, either in the time-continuous form (1) or the time-discrete one (2), has been subject to extensive studies [11], with the prominent application being Sequential Monte Carlo methods [12]. The time-continuous particle system is sometimes called a Moran particle process or a branching-diffusion process, and the corresponding nonlinear evolution is called a Fleming Viot process [15]. What we call the external potential is sometimes referred to as a fitness function. The evolution can be studied via its Feynman–Kac formulation (see (4)) and, typically, particle approximation errors of the type
are obtained for a class of observables ; see, for instance, [10]. Note that, differently from Theorem 1.2, the expectation is taken with respect to one single observable. Intuitively, this corresponds to averaging the error among different runs and considering instead of a single instance .
In the more recent work [21], instead, convergence is shown with respect to the expected Wasserstein distance by considering evolution on the torus with rate , as here. The use of Wasserstein distance with bounded ground cost, as , has been used in [26] to study particle systems also evolving in a bounded domain, and undergoing resampling when hitting the boundary. In the work [Bo and Liao, 2022] the authors also consider a Bounded-Lipschitz (BL) distance, with general RM dynamics which also includes (1). The particle system for which propagation of chaos is established, though, is obtained via the Feynman–Kac representation of the evolution, and, practically, corresponds to a weighted particle discretization. In Algorithm 1, instead, we consider unweighted particles, which is computationally more desirable to avoid loss of diversity. In [27] the author also considers the BL distance to study the bootstrap particle filter, similar to Algorithm 1, and obtains non-quantitative convergence of the particle approximation for non-compact domains and possibly unbounded fitness. Recent work [19] also considers particle filters and establishes almost sure, quantitative, convergence for one-dimensional systems with quadratic potential. Long-time high probability bounds for particle filters are also studied in [13] for non-compact state spaces via total variation and Kolmogorov distance under confinement assumptions on the potential. The authors of [16] consider a jump-diffusion process, as in (1), for an unbounded state space, and use the -Wasserstein distance . The analysis assumes directly the resampling mechanism to be Lipschitz with respect to , which does not cover, for instance, quadratic .
Our contribution follows a different line of work on propagation of chaos results for collisional models in kinetic theory, [18, 9], where each particle interacts with a randomly sampled partner from the ensemble. While in Algorithm 1 there is no binary interaction, the technique of optimally coupling the newly sampled particle is central in our proof. More precisely, we follow the strategy developed in [Borghi and Pareschi, 2026] for studying Nambu-type Monte Carlo methods and in [3] for Genetic Algorithms, where first the limit is considered, and only afterward do we go to the time-continuous settings with .
Preliminary lemmas
Let us set the notation first. Given a Borel probability measure , we denote by its -th moment. We equip the space with the Wasserstein-1 distance associated with the cost . By Kantorovich–Rubinstein duality [29], corresponds to the Bounded Lipschitz norm between measures
where and . Also, note that we can restrict the class to the Lipschitz functions with since translations do not change the value of the integrals appearing in the dual formulation.
Random variables are assumed to be defined over an underlying probability space and we write if . For a measurable test function and a measure we write the duality pairing as .
Useful bounds
The nonlinear update defined in (2) admits a useful representation via a time discrete Feynman–Kac formula [11 Chapter 2]. Let and be an independent standard Brownian process and define
The law can be equivalently defined as
for any bounded test function .
Lemma 2.1. Let with . If satisfies Assumption 1.1, then for every there exists a constant such that
and
Proof. We use (4) with to obtain
Recall that for a Brownian process, it holds that . Therefore, we have for all
Since the exponential weights take values in , the numerator in (5) can be bounded as
with . By the quadratic upper bound on in Assumption 1.1, therefore we also have
For the denominator in (5), instead, we apply Jensen’s inequality to obtain
where is independent of and . Putting the two estimates together in (5) gives the first desired bound.
For the second bound we note that, by using the Feynman–Kac formula with ,
Note that since . By multiplying this ratio for , the product simplifies to
for which we already derived a lower bound in (6).
Thanks to the bound on the moments, we can also estimate the error in approximating the partition function from i.i.d. samples. For a bound of order at a single step, it suffices to assume , since grows at most quadratically.
Lemma 2.2. Let with and , i.i.d. with empirical measure . If satisfies Assumption 1.1, it holds that
Proof. Since and the samples are independent, we have
By Cauchy–Schwarz, then we have
where we also used with . Finally, since satisfies the quadratic upper bound of Assumption 1.1, we have
Stability of replicator operator
We collect another important result concerning the Lipschitz continuity of with respect to the bounded Wasserstein distance . A key aspect of the estimate is its asymmetry with respect to the two input measures, similarly to [25 Theorem 15], as only one of them appears in the Lipschitz constant through its partition function .
Lemma 2.3. Let satisfy Assumption 1.1 and let . It holds that
Proof. We will make use of the dual formulation of as the bounded Lipschitz norm (3). Let be such that and and consider
Recall . By performing some algebraic computations we can show
Note that since and translation by does not change the oscillations of the test function, we have . For the Lipschitz constant, we have and by Assumption 1.1 it holds that , which leads to
where we also used with . Therefore, we have that is a suitable test function and we have , from which the desired bound follows.
We note that, if , Jensen’s inequality gives and the constant is of order . Keeping in the estimate allows us to apply Lemma 2.1 directly when iterating over .
Stability of resampling probabilities
The last result we collect concerns the probability of a particle undergoing resampling. Specifically, given two particles , we ask ourselves what the probability is that they are simultaneously resampled, or not. From Algorithm 1, we have that an particle is resampled with probability
If we try to maximize the correspondence between particles and , we have:
with probability : both resampled
with probability : only one resampled
with probability : neither resampled
Lemma 2.4. Let the potential satisfy Assumption 1.1. For any , it holds that
Proof. Consider first the case . Since , it holds that
and the bound is satisfied. For , recall from Assumption 1.1 that almost everywhere. We have
where in the last step we used with . Let . The function agrees with at and and has gradient bounded by almost everywhere. Its Lipschitz bound therefore gives
Given we are considering the case , this inequality corresponds to the desired estimate.
Coupling and limit
Two particle systems
We show propagation of chaos by coupling the particle system generated via Algorithm 1 with a nonlinear system sharing the same random variables. The key coupling is the one of the newly sampled particles, which will depend on i.i.d. continuous random variables .
To this end, given a set of positive weights we introduce the index map
If , note that and so as in Algorithm 1.
Recall a particle gets resampled with probability . We make the activation also explicitly dependent on i.i.d. continuous random variables so that is resampled if .
Starting from i.i.d., the update procedure of Algorithm 1 can then be rewritten as
with the weights given by , .
To construct the nonlinear particle system we will re-use the random variables , , , and, most importantly, optimally couple the newly sampled random variables thanks to the following result, which we recall for completeness.
Lemma 3.1. [3 Lemma 3.1] For , let be the Wasserstein distance with a lower semi-continuous cost function , and let such that . Consider and weights such that and . Define the weighted empirical measure as , and recall from (7) the index-map .
There exists a measurable mapping
with the following property: if , then the pair is an optimal coupling between and , namely
For any , let be the map defined above with for the cost and . The nonlinear particle system is iteratively defined as and
Note that, by Lemma 3.1, conditionally on the particles at time , we have and these draws are conditionally independent since the are independent. The conditional law of each update then depends only on . If the are i.i.d. with law , the law after resampling is . Adding the independent Gaussian increments gives . Therefore, by induction, we have that
with being the time-discrete update , see (2).
Propagation of chaos
To show convergence as , we track the distance for and iterate the estimate via a time-discrete Grönwall argument.
Proposition 3.2. Let satisfy Assumption 1.1, and , . Let be the empirical measure associated with the algorithm’s particle system and the one associated with the i.i.d. system (9). For every fixed and all such that it holds
with depending only on , and the -th moment of .
Proof. Let the particles’ resampling probability be given by . Since the resampling probabilities of the particles are coupled through the same random variables , we have that, conditionally on the particles at time , the probabilities of both, only one, and neither being resampled are respectively
The common Gaussian increment does not change . When only one is resampled, we will use that the maximum transport cost to be paid is .
By applying Lemma 2.4 and optimality of the -coupling we have
Let us focus on the first term. Since and , we have
where in the last steps we have used , Lemma 2.2 for the difference between the partition functions and Lemma 2.3 for the stability of the replicator operator .
The Wasserstein distance term itself, instead, can be bounded by the triangular inequality as
by the concentration inequalities of i.i.d. particles with respect to the Wasserstein-1 distance [17]. Note that the Monte Carlo rate is faster than the concentration rate (or equal if ) and therefore can be absorbed in an term.
Continuing the estimate from (10) and collecting the two terms in the distance between the coupled particles, we obtain
where we used the bound from Lemma 2.1. To iterate the estimate, we add to both sides of (3.5). Since
we obtain
Since , each multiplier is at least one and every partial product appearing in the iteration is bounded by the full product. By Lemma 2.1, we have
Recall that , so the initial particle distance is zero. Iterating the estimate above and using therefore gives
Subtracting the sampling term from both sides, we obtain
We note that , and so this Monte Carlo rate can be absorbed in the Wasserstein concentration rate . By recalling the definition of and from Lemma 2.1, we obtain
where the last inequality follows from . Since , we have obtained the desired estimate.
Limit and proof of the main Theorem
We start by stating what we mean by a measure solution to the RM model (1).
Definition 4.1. Let and . We say is a weak solution to (1) with initial datum over if for all and
Time-discretization error
To show the RM splitting scheme introduces a first order error with respect to the time discretization step , we briefly adapt the proof of [14] to account for unbounded potentials.
Proposition 4.2. Let be a measure solution to (1) with , , and be iteratively defined by (2) with . If satisfies Assumption 1.1, it holds
with the positive constant depending on , , and the -th moment of .
Proof. We rely on the Feynman–Kac representation of both the time-continuous solution and its time-discrete approximation; see (4) and [11]. Recall with and an independent Brownian motion. We have for any bounded test function
By integrating by parts on each time interval and applying Itô’s formula, we obtain
To estimate each term, we use the fact that both and grow at most quadratically by Assumption 1.1. Since for , by using the Burkholder–Davis–Gundy inequality [22], we have
with constants that depend on and the growth assumptions on . Note that we had already derived bounds for the moments of in the proof of Lemma Lemma 2.1. Similarly, we also have the bound
Let us take now a test function from the dual formulation (3) of , that is, with and . We use a similar trick to that in Lemma Lemma 2.3 and assume, without loss of generality, that . By also applying the inequality for positive we have
Now, to bound the numerator, we use the computation we just performed, while to bound the denominator we use Lemma Lemma 2.1. Altogether, since , this leads to
The desired estimate is obtained by taking the supremum over and using the dual formulation (3).
Proof of Theorem 1.2
Proof of Theorem 1.2. Let us recall the different probability measures we have considered: is the empirical probability measure generated with Algorithm 1, is the time-discrete update (1.2), while $\bar{\rho}_n^N$ is the empirical measure of the particles (3.3) which are i.i.d. with law . Finally, is the solution to (1.1) at time .
By the triangular inequality, we have
By Proposition Proposition 3.2, the first term is of order . Since is made of i.i.d. particles, Fournier and Guillin, [17] holds, and so the second error is also of order given by Lemma Lemma 2.1. Finally, the last term is of order by Proposition Proposition 4.2. Altogether, the rate with respect to the number of particles and time discretization is .
Aknowledgments
The author was supported by the Advanced Grant NEITALG of the European Research Council Executive Agency (ERC) (grant agreement No. 101198055), and by the Munich Center for Machine Learning (MCML)
Funded by the European Union. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Council Executive Agency. Neither the European Union nor the granting authority can be held responsible for them. This project has received funding from the European Research Council (ERC) under the European Union’s Horizon Europe research and innovation programme (grant agreement No. 101198055, project acronym NEITALG).
References
- [1]Bickel, P., Li, B., and Bengtsson, T. (2008). Sharp failure rates for the bootstrap particle filter in high dimensions. In Publishing the limits of contemporary statistics: Contributions in honor of Jayanta K. Ghosh, volume 3, pages 318–330. Institute of Mathematical Statistics.DOI
- [3]Borghi, G. (2026). Chaos propagation in genetic algorithms: An optimal transport approach. Bulletin of the London Mathematical Society, 58(3):e70333.arxiv.org/abs/2601.14169
- [6]Candy, J. V. (2007). Bootstrap particle filtering. IEEE Signal Processing Magazine, 24(4):73–85.DOI
- [7]Chaintron, L.-P. and Diez, A. (2022). Propagation of chaos: A review of models, methods and applications. I. Models and methods. Kinetic and Related Models, 15(6):895–1015.DOI
- [8]Chopin, N., Singh, S. S., Soto, T., and Vihola, M. (2022). On resampling schemes for particle filters with weakly informative observations. The Annals of Statistics, 50(6):3197–3222.arxiv.org/abs/2203.10037
- [9]Cortez, R. and Fontbona, J. (2016). Quantitative propagation of chaos for generalized Kac particle systems. The Annals of Applied Probability, 26(2):892 – 916.DOI
- [10]Crisan, D. and Doucet, A. (2002). A survey of convergence results on particle filtering methods for practitioners. IEEE Transactions on Signal Processing, 50(3):736–746.
- [11]Del Moral, P. (2004). Feynman–Kac Formulae: Genealogical and Interacting Particle Systems with Applications. Probability and Its Applications. Springer, New York.DOI
- [12]Del Moral, P., Doucet, A., and Jasra, A. (2006). Sequential Monte Carlo samplers. Journal of the Royal Statistical Society Series B: Statistical Methodology, 68(3):411–436.
- [13]Del Moral, P. and Gerber, M. (2026). A perturbation theory for Feynman-Kac models and their particle interpretations. arXiv preprint arXiv:2610.21587.arxiv.org/abs/2610.21587
- [14]Del Moral, P., Jacob, P. E., Lee, A., Murray, L., and Peters, G. W. (2013). Feynman-Kac particle integration with geometric interacting jumps. Stochastic Analysis and Applications, 31(5):830–871.arxiv.org/abs/1211.7191
- [15]Del Moral, P. and Miclo, L. (2000). A Moran particle system approximation of Feynman–Kac formulae. Stochastic processes and their applications, 86(2):193–216.
- [16]Erny, X. (2022). Well-posedness and propagation of chaos for McKean–Vlasov equations with jumps and locally Lipschitz coefficients. Stochastic Processes and their Applications, 150:192–214.arxiv.org/abs/2102.06472
- [17]Fournier, N. and Guillin, A. (2015). On the rate of convergence in Wasserstein distance of the empirical measure. Probability Theory and Related Fields, 162(3-4):707–738.arxiv.org/abs/1312.2128
- [18]Fournier, N. and Mischler, S. (2016). Rate of convergence of the Nanbu particle system for hard potentials and Maxwell molecules. The Annals of Probability, 44(1):589 – 627.DOI
- [19]Gerber, M. (2026). Safety of particle filters: Some results on the time evolution of particle filter estimates. Electronic Journal of Probability, 31:1–24.arxiv.org/abs/2503.21334
- [20]Holland, J. H. (1992). Adaptation in natural and artificial systems: an introductory analysis with applications to biology, control, and artificial intelligence. MIT press.
- [21]Journel, L. and Monmarché, P. (2022). Convergence of a particle approximation for the quasi-stationary distribution of a diffusion process: Uniform estimates in a compact soft case. ESAIM: Probability and Statistics, 26:1–25.arxiv.org/abs/1910.05060
- [22]Kallenberg, O. (2021). Foundations of Modern Probability. Probability Theory and Stochastic Modelling. Springer International Publishing.
- [23]Perthame, B. (2007). Transport equations in biology. Springer.
- [24]Snyder, C., Bengtsson, T., Bickel, P., and Anderson, J. (2008). Obstacles to high-dimensional particle filtering. Monthly Weather Review, 136(12):4629–4640.DOI
- [25]Sprungk, B. (2020). On the local Lipschitz stability of Bayesian inverse problems. Inverse Problems, 36(5):055015.DOI
- [26]Tough, O. and Nolen, J. (2022). The Fleming-Viot process with McKean-Vlasov dynamics. Electronic Journal of Probability, 27:1–72.arxiv.org/abs/2011.11689
- [27]Van Handel, R. (2009). Uniform time average consistency of Monte Carlo particle filters. Stochastic Processes and their Applications, 119(11):3835–3861.DOI
- [28]Vergé, C., Dubarry, C., Del Moral, P., and Moulines, E. (2015). On parallel implementation of sequential Monte Carlo methods: the island particle model. Statistics and Computing, 25(2):243–260.
- [29]Villani, C. (2009). Optimal transport, volume 338 of Grundlehren der mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences]. Springer-Verlag, Berlin. Old and new.