arXiv is now an independent nonprofit! Learn more
License: CC BY 4.0
arXiv:2504.10089v3 [math.NA] 20 Aug 2026

Convergence Analysis of a Stochastic Interacting Particle-Field Algorithm for 3D Parabolic-Parabolic Keller-Segel SystemsThanks: To appear in Mathematics of Computation.

Boyi Hu Thanks: Department of Mathematics, The University of Hong Kong, Pokfulam Road, Hong Kong SAR, P.R.China. huby22@connect.hku.hk.    Zhongjian Wang Thanks: Division of Mathematical Sciences, School of Physical and Mathematical Sciences, Nanyang Technological University, 21 Nanyang Link, Singapore 637371. zhongjian.wang@ntu.edu.sg    Jack Xin Thanks: Corresponding author. Department of Mathematics, University of California at Irvine, Irvine, CA 92697, USA. jack.xin@uci.edu    Zhiwen Zhang Thanks: Corresponding author. Department of Mathematics, The University of Hong Kong, Pokfulam Road, Hong Kong SAR, P.R.China. Materials Innovation Institute for Life Sciences and Energy (MILES), HKU-SIRI, Shenzhen, 518045, P.R. China. zhangzw@hku.hk.
Abstract

Chemotaxis models describe the movement of organisms in response to chemical gradients. In this paper, we propose a stochastic interacting particle-field algorithm integrated with random batch approximation (SIPF-rr) for the three-dimensional (3D) parabolic-parabolic Keller-Segel (KS) system, also referred to as the fully parabolic KS system. The SIPF-rr method approximates the KS system by coupling particle-based density representations with a smooth field variable computed via spectral methods. By incorporating the random batch method (RBM), we bypass the mean-field limit and significantly reduce computational complexity. Under mild assumptions on the regularity of the original KS system and the boundedness of numerical approximations, we prove that the empirical measure associated with the SIPF-rr particle system converges in the 11-Wasserstein distance to the exact measure of the limiting McKean-Vlasov process with high probability. Finally, we present numerical experiments to validate the theoretical convergence results, and demonstrate the efficiency and robustness of the SIPF-rr method as a diagnostic tool for density concentration and potential finite-time singularity in 3D parabolic-parabolic KS systems. Specifically, the SIPF-rr method identifies the critical threshold of initial mass leading to blow-up under specific initial data configurations.

keywords
Fully parabolic Keller-Segel system, stochastic interacting particle-field (SIPF) algorithm, random batch method, 3D simulations, convergence analysis.
runningheads: Convergence Analysis of SIPF algorithm for Keller-Segel Systems / Boyi Hu, Zhongjian Wang, Jack Xin, Zhiwen Zhang
MSC
35K51, 65C05, 65M12, 65M75, 65T50.

1 Introduction

Chemotaxis is a biological phenomenon involving the movement of organisms (e.g., bacteria) in response to signals (e.g., chemoattractants), which can be produced by the organisms themselves. Theoretical and mathematical modeling of this phenomenon was initiated by Patlak [40] and by Keller and Segel [26]. In this work, we focus on the fully parabolic KS system as follows:

ρt=(μρχρc),ϵct=Δcλ2c+ρ,𝐱Ωd,t[0,T],\begin{split}\rho_{t}&=\nabla\cdot(\mu\,\nabla\rho-\chi\,\rho\,\nabla c),\\ \epsilon\,c_{t}&=\Delta\,c-\lambda^{2}\,c+\rho,\qquad\mathbf{x}\in\Omega\subseteq\mathbb{R}^{d},\quad t\in[0,T],\end{split} (1)

where χ,μ\chi,\mu are positive constants, and ϵ,λ\epsilon,\lambda are non-negative constants. The model is called parabolic-elliptic if ϵ=0\epsilon=0, and fully parabolic if ϵ>0\epsilon>0. Here, ρ\rho denotes the density of active particles (bacteria), and cc represents the concentration of chemoattractants emitted by these bacteria. The model finds broad applications in biology, ecology, and medicine [41, 39, 2, 46].

The nonlinear, potentially singular behavior of KS equations, especially blow-up phenomena [35, 17, 1, 12], makes numerical methods essential. Mesh-based methods such as finite difference [5, 42, 11], finite element [43, 9, 44], and finite volume schemes [13, 6, 54] are widely employed. Recent advances include local discontinuous Galerkin methods with optimal convergence rates [30] and semi-discrete symmetrization-based schemes that avoid nonlinear solvers [31]. Despite their success, challenges remain in ensuring stability, convergence, and the effective handling of singularities, making numerical studies on the KS model an active and evolving field of research.

In addition to mesh-based methods, particle-based approaches have also been developed to address the challenges posed by the KS system, providing a complementary perspective. Stevens [48] developed an NN-particle system and established its convergence for the fully parabolic case. Haškovec and Schmeiser [16] proposed a convergent regularized particle system for the 2D parabolic-elliptic KS model. Moreover, Godinho and Quininao [14] showed well-posedness results and the propagation of chaos property in a subcritical KS equation. Craig and Bertozzi [7] proved the convergence of a blob method for the related aggregation equation. Liu and Yang [32] introduced a random particle blob method with a mollified kernel for the parabolic-elliptic case, proving its convergence when the macroscopic mean-field equation possesses a global weak solution [33]. Due to the potential finite-time blow-up of the KS system, deriving a priori error estimates inherently requires sufficient regularity of the exact solution. Consequently, theoretical convergence analyses in the literature are typically restricted to the pre-blow-up regime [48, 30, 32, 54]. For the singular regime, numerical studies primarily focus on structural preservation (e.g., positivity and energy dissipation) and empirical robustness [6, 10, 45].

In [51], we proposed a novel stochastic interacting particle-field (SIPF) algorithm for the fully parabolic KS system (1) in 3D. The SIPF method approximates the KS solution ρ\rho as the empirical measure of particles (see Eq. (2)) coupled with a smoother field variable cc computed using the spectral method (see Eq. (3)). The algorithm employs an implicit Euler discretization and a one-step recursion based on Green’s function, which efficiently captures concentration behavior and potentially finite-time blow-up with only dozens of Fourier modes in each dimension.

Despite the algorithm’s numerical efficiency and stability as observed in [51], a complete convergence analysis remains open. This paper fills that gap by establishing and validating convergence estimates for the SIPF-rr method, a random-batch variant designed to reduce computational cost. Our main result, presented in Theorem 4, establishes the convergence of the solution (ρ~,c~)(\widetilde{\rho},\widetilde{c}) obtained from the SIPF-rr method to the exact solution (ρ,c)(\rho,c) in the pre-blow-up regime under mild assumptions. Specifically, the 11-Wasserstein distance between the SIPF-rr and exact density distributions, denoted as 𝒲1(ρ~t,ρt)\mathcal{W}_{1}(\widetilde{\rho}_{t},\rho_{t}), depends on the time step δt\delta t, the number of Fourier modes HH, and the number of particles PP, and scales as 𝒪(1H2+1P+δt)\mathcal{O}\left(\frac{1}{H^{2}}+\sqrt{\frac{1}{P}}+\sqrt{\delta t}\right). The proof, detailed in Section 3, builds upon key lemmas that quantify single-step update errors and analyze their propagation over time. Furthermore, we demonstrate that the spectral truncation inherently regularizes the singular kernel, ensuring the unconditional stability of the algorithm even as the exact solution approaches the finite-time singularity.

The error between the chemical concentration c~\widetilde{c} in the SIPF-rr method and the exact solution cc originates from Fourier truncation and implicit Euler time discretization. Under appropriate regularity assumptions on cc, the truncation error decays as the number of Fourier modes HH increases. The temporal error can be expressed through differences between Fourier coefficients of c~\widetilde{c} and cc, which further couples with the L2L^{2} particle trajectory error 𝔼(X~tXtL2)\mathbb{E}(\|\widetilde{X}_{t}-X_{t}\|_{L^{2}}) accumulated from the preceding step. Here, XtX_{t} and X~t\widetilde{X}_{t} are the exact dynamics of the system associated with the density and its numerical approximations obtained via our methods, respectively.

At the numerical discretization level, the RBM [24, 25, 23, 4] is incorporated into the SIPF-rr method, which renders the particles fully independent and identically distributed (i.i.d.), thereby effectively circumventing the need to address the propagation of chaos [33]. At each time step, small random batches of particles are selected with replacement for particle interactions. In the error estimate between c~\widetilde{c} and cc, the i.i.d. property, together with the complex-valued mean value theorem [34] and Bernstein’s inequality, bound the deviations of the empirical mean of particle trajectories from its expectation. This deviation accounts for the uncertainty described in Theorem 4. The numerical experiments in Section 4 further demonstrate that, with the introduction of the RBM, the numerical examples maintain a high level of precision.

The error between XtX_{t} and X~t\widetilde{X}_{t} is influenced by the gradient of cc and c~\widetilde{c}, reflecting the dependence of the particle trajectories on the interaction potential. Using Parseval’s identity [27], we relate c~cL2\|\nabla\widetilde{c}-\nabla c\|_{L^{2}} to Fourier-coefficient errors, establishing two coupled recursive inequalities between c~cL2\|\nabla\widetilde{c}-\nabla c\|_{L^{2}} and 𝔼(X~tXtL2)\mathbb{E}(\|\widetilde{X}_{t}-X_{t}\|_{L^{2}}). The first inequality relates c~cL2\|\nabla\widetilde{c}-\nabla c\|_{L^{2}} to 𝔼(X~tXtL2)\mathbb{E}(\|\widetilde{X}_{t}-X_{t}\|_{L^{2}}); the second updates 𝔼(X~tXtL2)\mathbb{E}(\|\widetilde{X}_{t}-X_{t}\|_{L^{2}}) by incorporating a c~cL2\|\nabla\widetilde{c}-\nabla c\|_{L^{2}} term (see Eqs. (70)-(71)). Substituting and decoupling these aforementioned recursive inequalities yields a bound for 𝔼(X~tXtL2)\mathbb{E}(\|\widetilde{X}_{t}-X_{t}\|_{L^{2}}) that depends only on earlier errors. Through the natural coupling γt=Law(X~t,Xt)\gamma_{t}=\text{Law}(\widetilde{X}_{t},X_{t}), this L2L^{2} error estimate translates to the 11-Wasserstein distance 𝒲1(ρ~t,ρt)\mathcal{W}_{1}(\widetilde{\rho}_{t},\rho_{t}). Applying the discrete Gronwall inequality together with Fourier-coefficient error estimates leads to the global convergence result in Theorem 4.

In addition to the theoretical analysis, Section 4 provides comprehensive numerical experiments to validate the performance of the SIPF-rr method. These experiments verify the predicted convergence rates, confirm the validity of the regularity assumptions, and explore the complex dynamics of the KS system. A particular focus is placed on identifying the initial-data-dependent threshold for finite-time blow-up in three dimensions under specific initial data configurations. In Subsection 4.2, the SIPF-rr algorithm identifies blow-up regimes and captures severe concentration toward finite-time singularities (see Figures 4 and 6). Notably, the method remains robust in detecting blow-up even under practical discretization parameters (HH, PP, δt\delta t) that are much coarser than those required for the convergence analysis (see Figure 5).

The rest of the paper is organized as follows. In Section 2, we review the SIPF method for solving the fully parabolic KS system and present the derivation of the SIPF-rr method, which incorporates the RBM to compute the particle interaction. In Section 3, under certain assumptions, we provide a detailed convergence analysis of the SIPF-rr method, which decomposes the proof into several lemmas. In Section 4, we present numerical results to validate the necessity of the assumptions, demonstrate the accuracy of the SIPF-rr method, and confirm the theoretical convergence rate derived in our analysis. Section 5 discusses extensions of the SIPF-rr framework to a wider class of mathematical-biological chemotaxis models, and Section 6 concludes the paper.

2 Derivation of the SIPF-rr Method

In this section, we present the SIPF-rr method for solving the fully parabolic KS model. Since the evolutionary phenomena of interest occur in the interior of the domain, we restrict the system (1) to a large domain Ω=[L/2,L/2]3\Omega=[-L/2,L/2]^{3} and assume Dirichlet boundary conditions for particle density ρ\rho and Neumann boundary conditions for chemical concentration cc.

Throughout this section, we use the standard notation ρ\rho, cc, etc., to represent the exact solutions of the fully parabolic KS model. For the variables computed or approximated using the SIPF-rr algorithm, we instead use the notations ρ~\widetilde{\rho}, c~\widetilde{c}, etc.

As a numerical algorithm, we assume that the temporal domain [0,T][0,T] is partitioned by {tn}n=0:nT\{t_{n}\}_{n=0:n_{T}} with t0=0t_{0}=0 and tnT=Tt_{n_{T}}=T. We approximate the density ρ~\widetilde{\rho} at t=tnt=t_{n} by empirical particles {X~tnp}p=1:P\{\widetilde{X}^{p}_{t_{n}}\}_{p=1:P}, i.e.,

ρ~tn=M0Pp=1Pδ(xX~tnp),P1,\displaystyle\widetilde{\rho}_{t_{n}}={\frac{M_{0}}{P}}\,\sum_{p=1}^{P}\delta(x-\widetilde{X}^{p}_{t_{n}}),\;\;P\gg 1, (2)

where M0M_{0} denotes the conserved total mass (integral of ρ\rho over the domain Ω\Omega). For the chemical concentration c~\widetilde{c}, we adopt a Fourier basis approximation. Its unnormalized Fourier coefficients are defined by

α~t;𝐣:=𝐣[c~(,t)]=Ωc~(𝐱,t)eiω𝐣𝐱d𝐱,ω𝐣=2πL𝐣.\widetilde{\alpha}_{t;\mathbf{j}}:=\mathcal{F}_{\mathbf{j}}[\widetilde{c}(\cdot,t)]=\int_{\Omega}\widetilde{c}(\mathbf{x},t)e^{-i\omega_{\mathbf{j}}\cdot\mathbf{x}}\,d\mathbf{x},\qquad\omega_{\mathbf{j}}=\frac{2\pi}{L}\mathbf{j}.

The corresponding inverse Fourier representation is

c~(𝐱,t)=1L3𝐣α~t;𝐣exp(i2πj1x1/L)exp(i2πj2x2/L)exp(i2πj3x3/L),\widetilde{c}(\mathbf{x},t)=\frac{1}{L^{3}}\sum_{\mathbf{j}\in\mathcal{H}}\,\widetilde{\alpha}_{t;\mathbf{j}}\,\exp(i2\pi j_{1}\,x_{1}/L)\exp(i2\pi j_{2}\,x_{2}/L)\exp(i2\pi j_{3}\,x_{3}/L), (3)

where \mathcal{H} denotes the index set, i.e.,

={𝐣3:|j1|,|j2|,|j3|H2},\displaystyle\mathcal{H}=\{\mathbf{j}\in\mathbb{Z}^{3}:|j_{1}|,|j_{2}|,|j_{3}|\leq\frac{H}{2}\}, (4)

and i=1i=\sqrt{-1}. Defining αt;𝐣:=𝐣[c(,t)]\alpha_{t;\mathbf{j}}:=\mathcal{F}_{\mathbf{j}}[c(\cdot,t)] in the same way, the exact solution c(𝐱,t)c(\mathbf{x},t) can also be approximated by a truncated spatial Fourier series expansion as follows:

c(𝐱,t)1L3𝐣αt;𝐣exp(i2πj1x1/L)exp(i2πj2x2/L)exp(i2πj3x3/L).c(\mathbf{x},t)\approx\frac{1}{L^{3}}\sum_{\mathbf{j}\in\mathcal{H}}\,\alpha_{t;\mathbf{j}}\,\exp(i2\pi j_{1}\,x_{1}/L)\exp(i2\pi j_{2}\,x_{2}/L)\exp(i2\pi j_{3}\,x_{3}/L). (5)
Remark 1.

The choice of the Fourier basis over Hermite polynomials for approximating the chemical concentration is based on the fact that, since the blow-up phenomenon is localized in the domain interior, periodic boundary conditions effectively emulate an infinite spatial domain under this configuration. When the spatial location of the singularity remains distant from domain boundaries, its interaction with these artificial edges becomes negligible.

Then at t0=0t_{0}=0, we generate PP empirical samples {X~0p}p=1:P\{\widetilde{X}^{p}_{0}\}_{p=1:P} according to the initial condition of ρ~0\widetilde{\rho}_{0} and set up α~0;𝐣\widetilde{\alpha}_{0;\mathbf{j}} using the Fourier series of c~0\widetilde{c}_{0}. For ease of presenting our algorithm, with a slight abuse of notation, we use ρ~n=M0Pp=1Pδ(xX~np)\widetilde{\rho}_{n}={\frac{M_{0}}{P}}\,\sum_{p=1}^{P}\delta(x-\widetilde{X}^{p}_{n}), and

c~n=1L3𝐣α~n;𝐣exp(i2πj1x1/L)exp(i2πj2x2/L)exp(i2πj3x3/L)\widetilde{c}_{n}=\frac{1}{L^{3}}\sum_{\mathbf{j}\in\mathcal{H}}\,\widetilde{\alpha}_{n;\mathbf{j}}\,\exp(i2\pi j_{1}\,x_{1}/L)\exp(i2\pi j_{2}\,x_{2}/L)\exp(i2\pi j_{3}\,x_{3}/L) (6)

to represent density ρ~\widetilde{\rho} and chemical concentration c~\widetilde{c} at time tnt_{n}.

Considering the time-stepping for the system (1) from tnt_{n} to tn+1t_{n+1}, with ρ~n\widetilde{\rho}_{n} and c~n1\widetilde{c}_{n-1} known, our algorithm, inspired by the operator splitting technique, consists of two sub-steps: updating chemical concentration c~\widetilde{c} and updating organism density ρ~\widetilde{\rho}.

Updating chemical concentration c~\widetilde{c}

Let δt=tn+1tn>0\delta t=t_{n+1}-t_{n}>0 be the time step. We discretize the c~\widetilde{c} equation of (1) in time by an implicit Euler scheme

ϵ(c~nc~n1)/δt=(Δλ2)c~n+ρ~n.\displaystyle\epsilon\,(\widetilde{c}_{n}-\widetilde{c}_{n-1})/\delta t=(\Delta-\lambda^{2})\,\widetilde{c}_{n}+\widetilde{\rho}_{n}. (7)

From Eq. (7), we obtain the explicit formula for c~n\widetilde{c}_{n} as follows:

(Δλ2ϵ/δt)c~n=ϵc~n1/δtρ~n.\displaystyle(\Delta-\lambda^{2}-\epsilon/\delta t)\,\widetilde{c}_{n}=-\epsilon\,\widetilde{c}_{n-1}/\delta t-\widetilde{\rho}_{n}. (8)

It follows that

c~n=c~(𝐱,tn)\displaystyle\widetilde{c}_{n}=\widetilde{c}(\mathbf{x},t_{n}) =𝒦ϵ,δt(ϵc~n1/δt+ρ~n)=𝒦ϵ,δt(ϵc~(𝐱,tn1)/δt+ρ~(𝐱,tn)),\displaystyle=-\mathcal{K}_{\epsilon,\delta t}\ast(\epsilon\,\widetilde{c}_{n-1}/\delta t+\widetilde{\rho}_{n})=-\mathcal{K}_{\epsilon,\delta t}\ast(\epsilon\,\widetilde{c}(\mathbf{x},t_{n-1})/\delta t+\widetilde{\rho}(\mathbf{x},t_{n})), (9)

where 𝒦ϵ,δt\mathcal{K}_{\epsilon,\delta t} is the Green’s function of the operator Δλ2ϵ/δt\Delta-\lambda^{2}-\epsilon/\delta t and \ast represents an approximation of spatial convolution that differs from the continuous setup, as c~\widetilde{c} is computed using truncated Fourier basis functions and ρ~\widetilde{\rho} is given by a discrete particle representation. Unless otherwise stated, all subsequent norms \|\cdot\| will refer to the L2L^{2} norms. In the case of 3\mathbb{R}^{3}, the Green’s function 𝒦ϵ,δt\mathcal{K}_{\epsilon,\delta t} reads as follows:

𝒦ϵ,δt=𝒦ϵ,δt(𝐱)=exp{β𝐱}4π𝐱,β=λ2+ϵ/δt.\displaystyle\mathcal{K}_{\epsilon,\delta t}=\mathcal{K}_{\epsilon,\delta t}(\mathbf{x})=-\frac{\exp\{-\beta\|\mathbf{x}\|\}}{4\pi\|\mathbf{x}\|},\quad\beta=\sqrt{\lambda^{2}+\epsilon/\delta t}. (10)

Green’s function admits a closed-form Fourier transform,

𝒦ϵ,δt(ω)=1ω2+β2.\displaystyle\mathcal{F}\mathcal{K}_{\epsilon,\delta t}(\mathbf{\omega})=-\frac{1}{\|\mathbf{\omega}\|^{2}+\beta^{2}}. (11)

For the term 𝒦ϵ,δtc~n1-\mathcal{K}_{\epsilon,\delta t}\ast\widetilde{c}_{n-1} in Eq. (9), by Eq. (11) it is equivalent to modify Fourier coefficients α~𝐣\widetilde{\alpha}_{\mathbf{j}} to α~𝐣/(4π2j12/L2+4π2j22/L2+4π2j32/L2+β2)\widetilde{\alpha}_{\mathbf{j}}/(4\pi^{2}j_{1}^{2}/L^{2}+4\pi^{2}j_{2}^{2}/L^{2}+4\pi^{2}j_{3}^{2}/L^{2}+\beta^{2}).

For the second term 𝒦ϵ,δtρ~\mathcal{K}_{\epsilon,\delta t}\ast\widetilde{\rho}, we first approximate 𝒦ϵ,δt\mathcal{K}_{\epsilon,\delta t} using a cosine series expansion. Then, we use the particle representation of ρ~\widetilde{\rho} given in Eq. (2) to derive

(𝒦ϵ,δtρ~)𝐣M0Pp=1Pexp(i2πj1X~pn;1/Li2πj2X~pn;2/Li2πj3X~pn;3/L)(1)j1+j2+j34π2j12/L2+4π2j22/L2+4π2j32/L2+β2,\displaystyle(\mathcal{K}_{\epsilon,\delta t}\ast\widetilde{\rho})_{\mathbf{j}}\approx\frac{M_{0}}{P}\sum_{p=1}^{P}\!\frac{\exp(-i2\pi j_{1}\widetilde{X}^{p}_{n;1}/L\!-\!i2\pi j_{2}\widetilde{X}^{p}_{n;2}/L\!-\!i2\pi j_{3}\widetilde{X}^{p}_{n;3}/L)(-1)^{j_{1}+j_{2}+j_{3}}}{4\pi^{2}j_{1}^{2}/L^{2}+4\pi^{2}j_{2}^{2}/L^{2}+4\pi^{2}j_{3}^{2}/L^{2}+\beta^{2}}, (12)

where the factor (1)j1+j2+j3(-1)^{j_{1}+j_{2}+j_{3}} arises from the redefinition of the Fourier computational domain from [0,L]3[0,L]^{3} to [L/2,L/2]3[-L/2,L/2]^{3}.

Finally, we summarize the one-step update of the Fourier coefficients of the chemical concentration c~\widetilde{c} in Algorithm 1, which follows the same procedure as in the original SIPF method [51].

Algorithm 1 One-step update of chemical concentration in the SIPF-rr method
0:  Distribution ρ~n\widetilde{\rho}_{n} represented by empirical samples X~n\widetilde{X}_{n}, initial concentration c~n1\widetilde{c}_{n-1} represented by Fourier coefficients α~n1\widetilde{\alpha}_{n-1}.
1:for 𝐣\mathbf{j}\in\mathcal{H} do
2:   α~n;𝐣ϵα~n1;𝐣δt(4π2j12/L2+4π2j22/L2+4π2j32/L2+β2)\widetilde{\alpha}_{n;\mathbf{j}}\leftarrow\dfrac{\epsilon\widetilde{\alpha}_{n-1;\mathbf{j}}}{\delta t(4\pi^{2}j_{1}^{2}/L^{2}+4\pi^{2}j_{2}^{2}/L^{2}+4\pi^{2}j_{3}^{2}/L^{2}+\beta^{2})}
3:   F𝐣0F_{\mathbf{j}}\leftarrow 0
4:   for p=1p=1 to PP do
5:    F𝐣F𝐣+exp(i2πj1X~n;1p/Li2πj2X~n;2p/Li2πj3X~n;3p/L)F_{\mathbf{j}}\leftarrow F_{\mathbf{j}}+\exp(-i2\pi j_{1}\widetilde{X}^{p}_{n;1}/L-i2\pi j_{2}\widetilde{X}^{p}_{n;2}/L-i2\pi j_{3}\widetilde{X}^{p}_{n;3}/L)
6:   end for
7:   F𝐣F𝐣(1)j1+j2+j34π2j12/L2+4π2j22/L2+4π2j32/L2+β2M0PF_{\mathbf{j}}\leftarrow F_{\mathbf{j}}\cdot\dfrac{(-1)^{j_{1}+j_{2}+j_{3}}}{4\pi^{2}j_{1}^{2}/L^{2}+4\pi^{2}j_{2}^{2}/L^{2}+4\pi^{2}j_{3}^{2}/L^{2}+\beta^{2}}\cdot\dfrac{M_{0}}{P}
8:end for
9:α~nα~n+F\widetilde{\alpha}_{n}\leftarrow\widetilde{\alpha}_{n}+F
9:  Updated chemical concentration field from c~n1\widetilde{c}_{n-1} to c~n\widetilde{c}_{n} via α~n\widetilde{\alpha}_{n}.
Updating density of active particles ρ~\widetilde{\rho}

In the one-step update of density ρ~n\widetilde{\rho}_{n} represented by particles {X~np}p=1:P\{\widetilde{X}_{n}^{p}\}_{p=1:P}, we apply the Euler-Maruyama scheme to solve the stochastic differential equation (SDE)

X~n+1p=X~np+χ𝐱c~(X~np,tn)δt+2μδtNnp,\displaystyle\widetilde{X}^{p}_{n+1}=\widetilde{X}^{p}_{n}+\chi\nabla_{\mathbf{x}}\widetilde{c}(\widetilde{X}^{p}_{n},t_{n})\delta t+\sqrt{2\,\mu\,\delta t}\,N^{p}_{n}, (13)

where the variables NnpN^{p}_{n} are i.i.d. standard normal random variables corresponding to the Brownian paths in the SDE formulation. For n>1n>1, substituting Eq. (9) in Eq. (13) gives:

X~n+1p=X~npχ𝐱𝒦ϵ,δt(ϵc~n1(𝐱)/δt+ρ~n(𝐱))|𝐱=X~npδt+2μδtNnp,\displaystyle\widetilde{X}^{p}_{n+1}=\widetilde{X}^{p}_{n}-\chi\nabla_{\mathbf{x}}\mathcal{K}_{\epsilon,\delta t}\ast(\epsilon\,\widetilde{c}_{n-1}(\mathbf{x})/\delta t+\widetilde{\rho}_{n}(\mathbf{x}))|_{\mathbf{x}=\widetilde{X}^{p}_{n}}\delta t+\sqrt{2\,\mu\,\delta t}\,N^{p}_{n}, (14)

from which ρ~n+1(𝐱)\widetilde{\rho}_{n+1}(\mathbf{x}) is constructed via Eq. (2).

In this particle formulation, the computation of the spatial convolution differs slightly from that in the update of c~\widetilde{c} (i.e., Eq. (9)).

For the term 𝐱𝒦ϵ,δtc~n1(X~np)\nabla_{\mathbf{x}}\mathcal{K}_{\epsilon,\delta t}\ast\,\widetilde{c}_{n-1}(\widetilde{X}^{p}_{n}), to avoid the singular points of 𝐱𝒦ϵ,δt\nabla_{\mathbf{x}}\mathcal{K}_{\epsilon,\delta t}, we evaluate the integral with quadrature points that are away from 00. Precisely, denote the standard quadrature point in Ω\Omega as

x𝐣=(j1L/H,j2L/H,j3L/H),x_{\mathbf{j}}=(j_{1}\,L/H,j_{2}\,L/H,j_{3}\,L/H), (15)

where j1j_{1}, j2j_{2}, j3j_{3} are integers ranging from H/2-H/2 to H/21H/2-1. When computing 𝐱𝒦ϵ,δtc~n1(X~np)\nabla_{\mathbf{x}}\mathcal{K}_{\epsilon,\delta t}\ast\,\widetilde{c}_{n-1}(\widetilde{X}^{p}_{n}), we evaluate 𝐱𝒦ϵ,δt\nabla_{\mathbf{x}}\mathcal{K}_{\epsilon,\delta t} at {X~np+X¯npx𝐣}𝐣\{\widetilde{X}^{p}_{n}+\bar{X}^{p}_{n}-x_{\mathbf{j}}\}_{\mathbf{j}}, where a small spatial shift is defined as X¯np=L2H+X~npL/HLHX~np\bar{X}^{p}_{n}=\frac{L}{2H}+\lfloor\frac{\widetilde{X}^{p}_{n}}{L/H}\rfloor\frac{L}{H}-\widetilde{X}_{n}^{p} and c~\widetilde{c} at {x𝐣X¯np}𝐣\{x_{\mathbf{j}}-\bar{X}^{p}_{n}\}_{\mathbf{j}} correspondingly. The latter is computed by inverse Fourier transform of the shifted coefficients, with α~𝐣\widetilde{\alpha}_{\mathbf{j}} modified to α~𝐣exp(i2πj1X¯n;1p/Li2πj2X¯n;2p/Li2πj3X¯n;3p/L)\widetilde{\alpha}_{\mathbf{j}}\exp(-i2\pi j_{1}\bar{X}^{p}_{n;1}/L-i2\pi j_{2}\bar{X}^{p}_{n;2}/L-i2\pi j_{3}\bar{X}^{p}_{n;3}/L), where (X¯n;ip)(\bar{X}^{p}_{n;i}) denotes the ii-th component of X¯np\bar{X}^{p}_{n}.

Motivated by mini-batch sampling [15, 47, 49, 50, 52, 53] and the random batch method (RBM) [24, 25, 4, 23], for each particle X~np\widetilde{X}_{n}^{p}, we choose a small batch CpC_{p} of size RR randomly with replacement. We restrict the interaction of X~np\widetilde{X}_{n}^{p} to particles within this batch; specifically, we approximate the interaction term χδt𝐱𝒦ϵ,δtρ~(X~np,tn)\chi\delta t\nabla_{\mathbf{x}}\mathcal{K}_{\epsilon,\delta t}\ast\widetilde{\rho}(\widetilde{X}^{p}_{n},t_{n}) by the random batch sum sCp,spχM0δtR𝐱𝒦ϵ,δt(X~npX~ns)\sum_{s\in C_{p},s\neq p}\frac{\chi M_{0}\delta t}{R}\nabla_{\mathbf{x}}\mathcal{K}_{\epsilon,\delta t}(\widetilde{X}^{p}_{n}-\widetilde{X}^{s}_{n}).

We summarize the one-step update (for n>1n>1) of the density in the SIPF-rr method as in Algorithm 2.

Algorithm 2 One-step update of density in the SIPF-rr method
0:  Distribution ρ~n\widetilde{\rho}_{n} represented by empirical samples X~n\widetilde{X}_{n}, concentration c~n1\widetilde{c}_{n-1} represented by Fourier coefficients α~n1\widetilde{\alpha}_{n-1}.
1:for p=1p=1 to PP do
2:   X~n+1pX~np+2μδtNnp\widetilde{X}_{n+1}^{p}\leftarrow\widetilde{X}_{n}^{p}+\sqrt{2\mu\delta t}N^{p}_{n} {NnpN^{p}_{n} is a standard normal random variable}
3:   CpC_{p}\leftarrow random subset of {1,,P}\{1,\dots,P\} with replacement, size RR
4:   X~n+1pX~n+1psCp,spχM0δtR𝐱𝒦ϵ,δt(X~npX~ns)\widetilde{X}_{n+1}^{p}\leftarrow\widetilde{X}_{n+1}^{p}-\sum_{s\in C_{p},s\neq p}\frac{\chi M_{0}\delta t}{R}\nabla_{\mathbf{x}}\mathcal{K}_{\epsilon,\delta t}(\widetilde{X}^{p}_{n}-\widetilde{X}^{s}_{n})
5:   X¯npL2H+X~npL/HLHX~np\bar{X}^{p}_{n}\leftarrow\frac{L}{2H}+\lfloor\frac{\widetilde{X}^{p}_{n}}{L/H}\rfloor\frac{L}{H}-\widetilde{X}_{n}^{p}
6:   for (𝐣)(\mathbf{j})\in\mathcal{H} do
7:    F𝐣𝐱𝒦ϵ,δt(X~np+X¯npx𝐣)F_{\mathbf{j}}\leftarrow\nabla_{\mathbf{x}}\mathcal{K}_{\epsilon,\delta t}(\widetilde{X}^{p}_{n}+\bar{X}_{n}^{p}-x_{\mathbf{j}})  {x𝐣x_{\mathbf{j}} from Eq. (15)}
8:    G𝐣α~n1;𝐣exp(i2πj1X¯n;1p/Li2πj2X¯n;2p/Li2πj3X¯n;3p/L)G_{\mathbf{j}}\leftarrow\widetilde{\alpha}_{n-1;\mathbf{j}}\exp(-i2\pi j_{1}\bar{X}^{p}_{n;1}/L-i2\pi j_{2}\bar{X}^{p}_{n;2}/L-i2\pi j_{3}\bar{X}^{p}_{n;3}/L)
9:   end for
10:   GˇiFFT(G)\check{G}\leftarrow\text{iFFT}(G)
11:   X~n+1pX~n+1pϵχ(F,Gˇ)L3H3\widetilde{X}_{n+1}^{p}\leftarrow\widetilde{X}_{n+1}^{p}-\epsilon\chi(F,\check{G})\frac{L^{3}}{H^{3}}  {(,)L3H3(\cdot,\cdot)\frac{L^{3}}{H^{3}} denotes an inner product corresponding to L2(Ω)L^{2}(\Omega) quadrature}
12:end for
12:  Updated distribution ρ~n+1\widetilde{\rho}_{n+1} represented by X~n+1\widetilde{X}_{n+1}.
Remark 2.

The spatial shift X¯np\bar{X}_{n}^{p} acts as a mild perturbation to avoid the singular attractive force induced by the Green’s function 𝒦ϵ,δt\mathcal{K}_{\epsilon,\delta t}. While it may technically introduce bias, due to the even structure of 𝒦ϵ,δt\mathcal{K}_{\epsilon,\delta t} and the regularity of c~\widetilde{c}, the influence of this shift is small when evaluating (F,Gˇ)(F,\check{G}). Consequently, particle aggregation remains driven by the intrinsic nonlinear dynamics rather than the grid structure. Combined with the stochasticity of the random batch method, this ensures that no systematic directional bias is introduced, which allows for robust singularity detection independent of specific grid points.

Combining Eq. (9) and Eq. (14), we conclude that the recursion from
({X~np}p=1:P,ρ~n(𝐱),c~n1(𝐱))(\{\widetilde{X}^{p}_{n}\}_{p=1:P},\widetilde{\rho}_{n}(\mathbf{x}),\widetilde{c}_{n-1}(\mathbf{x})) to ({X~n+1p}p=1:P,ρ~n+1(𝐱),c~n(𝐱))(\{\widetilde{X}^{p}_{n+1}\}_{p=1:P},\widetilde{\rho}_{n+1}(\mathbf{x}),\widetilde{c}_{n}(\mathbf{x})) is thus fully defined. We summarize the SIPF-rr method in the following Algorithm 3.

Algorithm 3 Stochastic Interacting Particle-Field Method
0:  Initial distribution ρ0\rho_{0}, initial concentration c0c_{0}.
1:  Generate PP i.i.d. samples according to distribution ρ0\rho_{0}: X~01,X~02,,X~0P\widetilde{X}_{0}^{1},\widetilde{X}_{0}^{2},\dots,\widetilde{X}_{0}^{P}.
2:for p=1p=1 to PP do
3:   Compute X~1p\widetilde{X}^{p}_{1} by Eq. (13), with c1=c0c_{-1}=c_{0}.
4:end for
5:  Compute c~1\widetilde{c}_{1} by Alg. 1 with c0c_{0} and ρ~1=p=1PM0PδX~1p\widetilde{\rho}_{1}=\sum_{p=1}^{P}\frac{M_{0}}{P}\delta_{\widetilde{X}^{p}_{1}}.
6:for n=2n=2 to N=T/δtN=\lfloor T/\delta t\rfloor do
7:   Compute X~n\widetilde{X}_{n} by Alg. 2 with ρ~n1\widetilde{\rho}_{n-1} and c~n2\widetilde{c}_{n-2}.
8:   Compute c~n\widetilde{c}_{n} by Alg. 1 with c~n1\widetilde{c}_{n-1} and ρ~n=p=1PM0PδX~np\widetilde{\rho}_{n}=\sum_{p=1}^{P}\frac{M_{0}}{P}\delta_{\widetilde{X}^{p}_{n}}.
9:end for
9:  Final particle distribution ρ~N\widetilde{\rho}_{N} and concentration field c~N\widetilde{c}_{N}.
Computational Complexity

We briefly analyze the complexity of the proposed SIPF-rr method. The memory usage is 𝒪(P+||)\mathcal{O}(P+|\mathcal{H}|), where |||\mathcal{H}| denotes the total number of Fourier modes. Regarding the computational cost, the original SIPF method typically requires 𝒪(P2)\mathcal{O}(P^{2}) operations for pairwise particle interactions. In contrast, by incorporating the RBM, the interaction cost in the SIPF-rr method is reduced to 𝒪(PR)\mathcal{O}(PR), where RR is the batch size.

The overall per-step complexity of Algorithm 3 is 𝒪(PR+P||log(||))\mathcal{O}(PR+P|\mathcal{H}|\log(|\mathcal{H}|)). Crucially, the error analysis in Theorem 4 reveals that the error introduced by the random batch approximation scales only as 𝒪(δt/R)\mathcal{O}(\sqrt{\delta t/R}). This allows us to choose RPR\ll P (e.g., R=100R=100 vs. P=104P=10^{4}) to achieve a substantial speedup while maintaining the desired accuracy.

Particle-wise Independence due to RBM

In the above derivation, {X~np}p=1:P\{\widetilde{X}^{p}_{n}\}_{p=1:P} are i.i.d. samples with distribution ρ~n\widetilde{\rho}_{n} and independent of c~n1\widetilde{c}_{n-1}. The one-step trajectories follow the discrete-time rule:

X~tn+1=X~tn+χc~(X~tn,tn)δt+tntn+12μdWs,\widetilde{X}_{t_{n+1}}=\widetilde{X}_{t_{n}}+\chi\nabla\widetilde{c}(\widetilde{X}_{t_{n}},t_{n})\delta t+\int_{t_{n}}^{t_{n+1}}\sqrt{2\mu}\,dW_{s}, (16)

where c~\nabla\widetilde{c} is computed via Eq. (9), and WsW_{s} denotes the Brownian motion. It is worth noting that, for the updated position of the pp-th particle X~n+1p\widetilde{X}_{n+1}^{p} by Eq. (13), the interaction term, 𝐱𝒦ϵ,δtρ~(X~np,tn)\nabla_{\mathbf{x}}\mathcal{K}_{\epsilon,\delta t}\ast\,\widetilde{\rho}(\widetilde{X}^{p}_{n},t_{n}) is computed by sCp,spχM0δtR𝐱𝒦ϵ,δt(X~npX~ns)\sum_{s\in C_{p},s\neq p}\frac{\chi M_{0}\delta t}{R}\nabla_{\mathbf{x}}\mathcal{K}_{\epsilon,\delta t}(\widetilde{X}^{p}_{n}-\widetilde{X}^{s}_{n}), where the selection of CpC_{p} is independent of X~np\widetilde{X}^{p}_{n} and hence {X~ns}sCp\{\widetilde{X}^{s}_{n}\}_{s\in C_{p}} can be viewed as i.i.d. samples of ρ~n\widetilde{\rho}_{n} independent of c~n1\widetilde{c}_{n-1} and X~np\widetilde{X}^{p}_{n}. Together with the independent Brownian motion term, we can deduce the independence of {X~n+1p}p=1:P\{\widetilde{X}_{n+1}^{p}\}_{p=1:P}.

Correspondingly, we denote the exact dynamics of the system by XtX_{t}, a ρ(,t)\rho(\cdot,t)-distributed random variable evolving continuously in time:

Xt=Xt0+χt0tc(Xs,s)𝑑s+t0t2μdWs,Xt0=X~t0,X_{t}=X_{t_{0}}+\chi\int_{t_{0}}^{t}\nabla c(X_{s},s)\,ds+\int_{t_{0}}^{t}\sqrt{2\mu}\,dW_{s},\quad{X}_{t_{0}}=\widetilde{X}_{t_{0}}, (17)

where c(,s)c(\cdot,s) is the exact concentration field, and the integral describes how the gradient field evolves in continuous time. Both processes share the same Brownian motion WsW_{s}, indicating that both processes are driven by the same source of randomness.

3 L2L^{2} Convergence of the SIPF-rr Method to Smooth Solutions

We now prove the convergence of the SIPF-rr method to classical solutions of the 3D parabolic-parabolic Keller-Segel equations. To ensure the validity of the following analysis, we introduce a set of assumptions that impose structure on the concentration fields and their gradients.

3.1 Assumptions and Theorem Statement

To establish the convergence analysis of the SIPF-rr method, we first introduce assumptions that ensure the boundedness of approximation errors for particles and gradients of the chemical concentration at any time within the temporal computational domain [0,T][0,T]. Specifically, we make the following assumptions.

Assumption 1.

There exist constants M1,M2>0M_{1},M_{2}>0 such that for all t[0,T]t\in[0,T] and 𝐱3\mathbf{x}\in\mathbb{R}^{3},

X~tXt\displaystyle\|\widetilde{X}_{t}-X_{t}\| M1,\displaystyle\leq M_{1}, (18)
c~(𝐱,t)c(𝐱,t)\displaystyle\|\nabla\widetilde{c}(\mathbf{x},t)-\nabla c(\mathbf{x},t)\| M2.\displaystyle\leq M_{2}. (19)

We note that this assumption only requires the errors to be bounded, but does not require them to converge to zero. The convergence of these errors to zero will be established in the subsequent analysis. Furthermore, the boundedness condition in Eq. (18) can be derived from Eqs. (16)-(17) and Assumption 2(c); Eq. (19) follows directly from the uniform bound in Assumption 2(c) detailed later.

Assumption 2.

Suppose both c~\nabla\widetilde{c} and c\nabla c satisfy Lipschitz continuity conditions in space and time, along with regularity and boundedness properties as follows:

(a) (Spatial Lipschitz Continuity) There exists a constant K>0K>0, depending on the regularity of c~\nabla\widetilde{c} and c\nabla c, as well as the parameters ϵ\epsilon and λ\lambda in the system (1), such that for all t[0,T]t\in[0,T] and 𝐱,𝐲3\mathbf{x},\mathbf{y}\in\mathbb{R}^{3},

max(c~(𝐱,t)c~(𝐲,t),c(𝐱,t)c(𝐲,t))K𝐱𝐲.\max\big(\|\nabla\widetilde{c}(\mathbf{x},t)-\nabla\widetilde{c}(\mathbf{y},t)\|,\,\|\nabla c(\mathbf{x},t)-\nabla c(\mathbf{y},t)\|\big)\leq K\|\mathbf{x}-\mathbf{y}\|. (20)

This implies that the second derivatives (Hessian entries) 2c~(𝐱,t)\nabla^{2}\widetilde{c}(\mathbf{x},t) exist almost everywhere and satisfy:

sup𝐱3,t[0,T](2c~(𝐱,t))K.\sup\limits_{\mathbf{x}\in\mathbb{R}^{3},t\in[0,T]}\big(\|\nabla^{2}\widetilde{c}(\mathbf{x},t)\|\big)\leq K. (21)

(b) (Temporal Lipschitz Continuity) There exists a constant K1>0K_{1}>0, depending on the regularity of c\nabla c and the parameters ϵ\epsilon and λ\lambda in the system (1), such that for any t1,t2[0,T]t_{1},t_{2}\in[0,T] and 𝐱3\mathbf{x}\in\mathbb{R}^{3},

c(𝐱,t1)c(𝐱,t2)K1|t1t2|.\|\nabla c(\mathbf{x},t_{1})-\nabla c(\mathbf{x},t_{2})\|\leq K_{1}|t_{1}-t_{2}|. (22)

(c) (Uniform Boundedness) There exists a constant M3>0M_{3}>0, depending on the regularity of c\nabla c and the parameters ϵ\epsilon and λ\lambda in the system (1), such that for all t[0,T]t\in[0,T] and 𝐱3\mathbf{x}\in\mathbb{R}^{3}:

max(c(𝐱,t),c~(𝐱,t))M3.\max\big(\|\nabla c(\mathbf{x},t)\|,\|\nabla\widetilde{c}(\mathbf{x},t)\|\big)\leq M_{3}. (23)

(d) (Regularity of Time Derivatives) The exact solution c(𝐱,t)c(\mathbf{x},t) is assumed to be sufficiently smooth in time and space such that both t2c(𝐱,t)\partial_{t}^{2}c(\mathbf{x},t) and t2c(𝐱,t)\nabla\partial_{t}^{2}c(\mathbf{x},t) are bounded. There exists a constant K2>0K_{2}>0 such that for all t[0,T]t\in[0,T]:

max(sup𝐱3t2c(𝐱,t),sup𝐱3t2c(𝐱,t))K2.\max\left(\sup_{\mathbf{x}\in\mathbb{R}^{3}}\|\partial_{t}^{2}c(\mathbf{x},t)\|,\sup_{\mathbf{x}\in\mathbb{R}^{3}}\|\nabla\partial_{t}^{2}c(\mathbf{x},t)\|\right)\leq K_{2}. (24)

Remark 3.

These regularity and boundedness conditions reflect inherent properties of the SIPF-rr method rather than external constraints on the KS system. As analyzed in Subsection 3.4, the discrete Fourier representation and Brownian diffusion naturally impose spectral regularity and suppress singular clustering, ensuring c~\nabla\widetilde{c} remains well-behaved throughout the computation. The boundedness conditions in Assumption 2 are validated through numerical simulations in Subsection 4.1.1.

Assumptions 1 and 2 characterize a regular regime where the exact solution remains smooth over the time interval [0,T][0,T]. We now state our main theorem, which quantifies the convergence of the SIPF-rr method on [0,T][0,T].

Theorem 4.

Suppose that the exact solutions (ρ,c)(\rho,c) and the numerical solutions (ρ~,c~)(\widetilde{\rho},\widetilde{c}) obtained by the SIPF-rr method in 3\mathbb{R}^{3} satisfy Assumptions 1 and 2 uniformly for all t[0,T]t\in[0,T]. Let HH, PP, RR, and δt\delta t denote the number of Fourier modes, the number of particles, the batch size, and the uniform time step in the SIPF-rr method, respectively. Then, for all discrete time levels tn=nδtTt_{n}=n\delta t\leq T, the following error estimates hold with high probability:

For all n{0,1,,T/δt}n\in\{0,1,\dots,\lfloor T/\delta t\rfloor\}, the 11-Wasserstein distance (defined in Eq. (78)) between ρ~tn\widetilde{\rho}_{t_{n}} and ρtn\rho_{t_{n}} satisfies 𝒲1(ρ~tn,ρtn)=𝒪(1H2+1P+δt)\mathcal{W}_{1}(\widetilde{\rho}_{t_{n}},\rho_{t_{n}})=\mathcal{O}\left(\frac{1}{H^{2}}+\sqrt{\frac{1}{P}}+\sqrt{\delta t}\right), and the maximum error in the truncated Fourier coefficients of c~tn\widetilde{c}_{t_{n}} and ctnc_{t_{n}} satisfies that max𝐣|α~tn;𝐣αtn;𝐣|=𝒪(1H+HP+Hδt)\max_{\mathbf{j}\in\mathcal{H}}|\widetilde{\alpha}_{t_{n};\mathbf{j}}-\alpha_{t_{n};\mathbf{j}}|=\mathcal{O}\left(\frac{1}{H}+\frac{H}{\sqrt{P}}+H\delta t\right). More specifically, for all n{0,1,,T/δt}n\in\{0,1,\dots,\lfloor T/\delta t\rfloor\}, the errors are bounded with high probability by:

𝒲1(ρ~tn,ρtn)\displaystyle\mathcal{W}_{1}(\widetilde{\rho}_{t_{n}},\rho_{t_{n}}) S0H2+S1δt+S2P+S3δtR,\displaystyle\leq\frac{S_{0}}{H^{2}}+S_{1}\sqrt{\delta t}+\frac{S_{2}}{\sqrt{P}}+\frac{S_{3}\sqrt{\delta t}}{\sqrt{R}}, (25)
max𝐣|α~tn;𝐣αtn;𝐣|\displaystyle\max_{\mathbf{j}\in\mathcal{H}}|\widetilde{\alpha}_{t_{n};\mathbf{j}}-\alpha_{t_{n};\mathbf{j}}| S4H+S5Hδt+S6HP+S7HδtR,\displaystyle\leq\frac{S_{4}}{H}+S_{5}H\delta t+\frac{S_{6}H}{\sqrt{P}}+\frac{S_{7}H\delta t}{\sqrt{R}},

where SiS_{i}, i=0,,7i=0,\dots,7, denote positive constants specified in Eqs. (79)-(80).

A direct consequence of Theorem 4 is that the SIPF-rr solution converges to the exact solution as δt0\delta t\to 0 and H,PH,P\to\infty. Specifically, the density ρ~tn\widetilde{\rho}_{t_{n}} converges in the 1-Wasserstein metric even without imposing specific scaling relations between the discretization parameters, as all error terms in the first inequality of Eq. (25) naturally vanish. The convergence of the chemical concentration c~tn\widetilde{c}_{t_{n}} requires the scaling conditions H/P0H/\sqrt{P}\to 0 and Hδt0H\delta t\to 0 to control the statistical and discretization errors, respectively.

Remark 5 (Practical convergence and scaling).

Theorem 4 establishes error estimates up to time TT, the maximal time at which Assumptions 1-2 hold. Beyond TT, although the theoretical bounds no longer apply, the Fourier spectral cutoff HH and the parabolic structure of the equation for cc act as a natural low-pass filter that prevents the numerical solution c~\widetilde{c} from blowing up. Note that this numerical upper bound grows with HH. By repeating the experiment with increasing values of HH, we obtain a reliable detection of the time and location of the singularity.

Regarding parameter scaling, the stochastic error term 𝒪(δt/R)\mathcal{O}(\sqrt{\delta t/R}) is negligible relative to the leading-order discretization errors. Hence, a moderate batch size (e.g., R=100R=100) is sufficient in practice, as the total error is dominated by HH and δt\delta t. Furthermore, while the theoretical gradient error bound scales with HH due to differentiation, the particle density error benefits from the smoothing effect of time integration. Numerical experiments in Section 4 verify that moderate parameters (H=24H=24, P=104P=10^{4}, and δt=104\delta t=10^{-4}) are sufficient to guarantee convergence and blow-up detection for our numerical examples, which demonstrates the efficiency of the proposed method.

3.2 Several lemmas

The result of Theorem 4 relies on the following lemmas concerning the change in single-step update error of the SIPF-rr method. We first prove several lemmas in this subsection and leave the complete proof of Theorem 4 to the next subsection.

From Eqs. (3)-(5), the error between c~\widetilde{c} and cc can be decomposed into two components: the error in their Fourier coefficients and the truncation error of cc. As the number of Fourier modes HH tends to infinity, and given the smoothness of cc, the truncation error becomes negligible and can be omitted from the analysis. We now focus on the error analysis between the Fourier coefficients α~𝐣\widetilde{\alpha}_{\mathbf{j}} and α𝐣\alpha_{\mathbf{j}} of c~\widetilde{c} and cc, as presented in the following lemma.

Lemma 6.

For all n+n\in\mathbb{N_{+}} and all 𝐣\mathbf{j}\in\mathcal{H} (the same index set as in Eq. (4)), under Assumptions 1 and 2, the following inequality holds with high probability:

|α~tn;𝐣αtn;𝐣|\displaystyle|\widetilde{\alpha}_{t_{n};\mathbf{j}}-\alpha_{t_{n};\mathbf{j}}|\leq |α~tn1;𝐣αtn1;𝐣|+C1δt2+C2ω𝐣δtP+C3ω𝐣δt𝔼[X~tnXtn],\displaystyle|\widetilde{\alpha}_{t_{n-1};\mathbf{j}}-\alpha_{t_{n-1};\mathbf{j}}|+C_{1}\delta t^{2}+\frac{C_{2}\|\omega_{\mathbf{j}}\|\delta t}{\sqrt{P}}+C_{3}\|\omega_{\mathbf{j}}\|\delta t\mathbb{E}[\|\widetilde{X}_{t_{n}}-X_{t_{n}}\|],

where C1,C2,C3C_{1},C_{2},C_{3} are constants, and tn=nδtt_{n}=n\delta t.

Proof.

We first write the frequency ω𝐣=(2πj1L,2πj2L,2πj3L)\omega_{\mathbf{j}}=\left(\frac{2\pi j_{1}}{L},\frac{2\pi j_{2}}{L},\frac{2\pi j_{3}}{L}\right), and define the notation Z𝐣=(ω𝐣2+λ2)δtϵZ_{\mathbf{j}}=\left(\|\omega_{\mathbf{j}}\|^{2}+\lambda^{2}\right)\cdot\frac{\delta t}{\epsilon}, which satisfies Z𝐣0Z_{\mathbf{j}}\to 0 as δt0\delta t\to 0. Recall the update formula for the numerical approximation of the Fourier coefficients α~tn;𝐣\widetilde{\alpha}_{t_{n};\mathbf{j}} (from Eq. (9)):

(1+Z𝐣)α~tn;𝐣=α~tn1;𝐣+δtϵ𝐣[ρ~(𝐱,tn)],(1+Z_{\mathbf{j}})\widetilde{\alpha}_{t_{n};\mathbf{j}}=\widetilde{\alpha}_{t_{n-1};\mathbf{j}}+\frac{\delta t}{\epsilon}\mathcal{F}_{\mathbf{j}}[\widetilde{\rho}(\mathbf{x},t_{n})], (26)

where 𝐣[ρ~(𝐱,tn)]=M0Pp=1Peiω𝐣X~ptn\mathcal{F}_{\mathbf{j}}[\widetilde{\rho}(\mathbf{x},t_{n})]=\frac{M_{0}}{P}\sum_{p=1}^{P}e^{-i\mathbf{\omega}_{\mathbf{j}}\cdot\widetilde{X}^{p}_{t_{n}}} represents the Fourier coefficient of ρ~(𝐱,tn)\widetilde{\rho}(\mathbf{x},t_{n}) at the frequency ω𝐣\omega_{\mathbf{j}}.

The exact solution c(𝐱,t)c(\mathbf{x},t) satisfies the continuous equation:

ϵct=Δcλ2c+ρ.\epsilon\frac{\partial c}{\partial t}=\Delta c-\lambda^{2}c+\rho. (27)

Integrating this equation from tn1t_{n-1} to tnt_{n} and applying the Taylor expansion with integral remainder to the time derivative term yields c(tn1)=c(tn)δttc(tn)+tn1tn(stn1)t2c(s)𝑑sc(t_{n-1})=c(t_{n})-\delta t\partial_{t}c(t_{n})+\int_{t_{n-1}}^{t_{n}}(s-t_{n-1})\partial_{t}^{2}c(s)\,ds. This allows us to express the exact solution in a form compatible with the implicit Euler scheme:

c(tn)c(tn1)δt=1ϵ(Δc(tn)λ2c(tn)+ρ(tn))+Rn(𝐱).\frac{c(t_{n})-c(t_{n-1})}{\delta t}=\frac{1}{\epsilon}(\Delta c(t_{n})-\lambda^{2}c(t_{n})+\rho(t_{n}))+R_{n}(\mathbf{x}). (28)

By Assumption 2(d), t2c\|\partial_{t}^{2}c\| is bounded by K2K_{2}. The temporal truncation error Rn(𝐱)R_{n}(\mathbf{x}) satisfies

Rnδt2sups[tn1,tn]t2c(,s)C1δt,\|R_{n}\|\leq\frac{\delta t}{2}\,\sup_{s\in[t_{n-1},t_{n}]}\|\partial_{t}^{2}c(\cdot,s)\|\leq C_{1}\,\delta t, (29)

where C1=K22C_{1}=\frac{K_{2}}{2} is a constant. Taking the Fourier transform of the discrete relation for the exact solution and rearranging terms, we get

(1+Z𝐣)αtn;𝐣=αtn1;𝐣+δtϵ𝐣[ρ(tn)]+δt𝐣[Rn].(1+Z_{\mathbf{j}})\alpha_{t_{n};\mathbf{j}}=\alpha_{t_{n-1};\mathbf{j}}+\frac{\delta t}{\epsilon}\mathcal{F}_{\mathbf{j}}[\rho(t_{n})]+\delta t\mathcal{F}_{\mathbf{j}}[R_{n}]. (30)

Combining Eq. (30) and Eq. (26), we obtain

|αtn;𝐣α~tn;𝐣|=\displaystyle|\alpha_{t_{n};\mathbf{j}}\!-\!\widetilde{\alpha}_{t_{n};\mathbf{j}}|= 11+Z𝐣|(αtn1;𝐣α~tn1;𝐣)+δtϵ(𝐣[ρ(tn)]𝐣[ρ~n])+δt𝐣[Rn]|\displaystyle\frac{1}{1+Z_{\mathbf{j}}}\left|\left(\alpha_{t_{n-1};\mathbf{j}}-\widetilde{\alpha}_{t_{n-1};\mathbf{j}}\right)\!+\!\frac{\delta t}{\epsilon}\left(\mathcal{F}_{\mathbf{j}}[\rho(t_{n})]-\mathcal{F}_{\mathbf{j}}[\widetilde{\rho}_{n}]\right)\!+\!\delta t\mathcal{F}_{\mathbf{j}}[R_{n}]\right|
\displaystyle\leq |αtn1;𝐣α~tn1;𝐣|+δtϵ|𝐣[ρ(tn)]𝐣[ρ~n]|+δt|𝐣[Rn]|.\displaystyle|\alpha_{t_{n-1};\mathbf{j}}-\widetilde{\alpha}_{t_{n-1};\mathbf{j}}|+\frac{\delta t}{\epsilon}|\mathcal{F}_{\mathbf{j}}[\rho(t_{n})]-\mathcal{F}_{\mathbf{j}}[\widetilde{\rho}_{n}]|+\delta t|\mathcal{F}_{\mathbf{j}}[R_{n}]|. (31)

To bound the term |𝐣[ρ(tn)]𝐣[ρ~n]||\mathcal{F}_{\mathbf{j}}[\rho(t_{n})]-\mathcal{F}_{\mathbf{j}}[\widetilde{\rho}_{n}]|, we first state a generalization of the mean value theorem to complex-valued functions.

Let GG be an open subset of 3\mathbb{R}^{3}, and let f:Gf:G\to\mathbb{C} be a continuously differentiable function on GG. Fix points 𝐱,𝐲G\mathbf{x},\mathbf{y}\in G such that the line segment connecting 𝐱\mathbf{x} and 𝐲\mathbf{y} lies entirely within GG. There exists c1,c2(0,1)c_{1},c_{2}\in(0,1) such that

f(𝐲)f(𝐱)=Re(f((1c1)𝐱+c1𝐲)(𝐲𝐱))+iIm(f((1c2)𝐱+c2𝐲)(𝐲𝐱)).f(\mathbf{y})-f(\mathbf{x})=\operatorname{Re}\Big(\nabla f((1-c_{1})\mathbf{x}+c_{1}\mathbf{y})(\mathbf{y}-\mathbf{x})\Big)+i\operatorname{Im}\Big(\nabla f((1-c_{2})\mathbf{x}+c_{2}\mathbf{y})(\mathbf{y}-\mathbf{x})\Big). (32)

The proof of (32) is direct. First, we define the function

g(t)=f((1t)𝐱+t𝐲),t[0,1].g(t)=f((1-t)\mathbf{x}+t\mathbf{y}),\quad t\in[0,1]. (33)

Since gg is also a continuously differentiable function, the mean value theorem implies that there exist points c1,c2(0,1)c_{1},c_{2}\in(0,1) such that

Re(g(c1))=Re(g(1)g(0)),Im(g(c2))=Im(g(1)g(0)),\operatorname{Re}(g^{\prime}(c_{1}))=\operatorname{Re}(g(1)-g(0)),\quad\operatorname{Im}(g^{\prime}(c_{2}))=\operatorname{Im}(g(1)-g(0)), (34)

which implies Eq. (32). Applying this result to f(𝐱)=eiω𝐣𝐱f(\mathbf{x})=e^{-i\mathbf{\omega}_{\mathbf{j}}\mathbf{x}}, we obtain:

|eiω𝐣X~ptneiω𝐣Xptn|\displaystyle|e^{-i\mathbf{\omega}_{\mathbf{j}}\cdot\widetilde{X}^{p}_{t_{n}}}-e^{-i\mathbf{\omega}_{\mathbf{j}}\cdot X^{p}_{t_{n}}}| |ω𝐣sin(ω𝐣((1c1)X~tnp+c1Xtnp))\displaystyle\leq\|\omega_{\mathbf{j}}\cdot\sin(\omega_{\mathbf{j}}((1-c_{1})\widetilde{X}^{p}_{t_{n}}+c_{1}X^{p}_{t_{n}}))
+iω𝐣cos(ω𝐣((1c2)X~tnp+c2Xtnp))X~tnpXtnp\displaystyle+i\omega_{\mathbf{j}}\cdot\cos(\omega_{\mathbf{j}}((1-c_{2})\widetilde{X}^{p}_{t_{n}}+c_{2}X^{p}_{t_{n}}))\|\cdot\|\widetilde{X}^{p}_{t_{n}}-X^{p}_{t_{n}}\|
2ω𝐣X~tnpXtnp.\displaystyle\leq\sqrt{2}\|\omega_{\mathbf{j}}\|\cdot\|\widetilde{X}^{p}_{t_{n}}-X^{p}_{t_{n}}\|. (35)

Based on Eq. (3.2), we have

|𝐣[ρ(tn)]𝐣[ρ~n]|M0Pp=1P(eiω𝐣Xtnpeiω𝐣X~tnp)2M0ω𝐣p=1PX~tnpXtnpP.\displaystyle|\mathcal{F}_{\mathbf{j}}[\rho(t_{n})]\!-\!\mathcal{F}_{\mathbf{j}}[\widetilde{\rho}_{n}]|\leq\left\|\frac{M_{0}}{P}\!\sum_{p=1}^{P}(e^{-i\mathbf{\omega}_{\mathbf{j}}X^{p}_{t_{n}}}\!-\!e^{-i\mathbf{\omega}_{\mathbf{j}}\widetilde{X}^{p}_{t_{n}}})\right\|\!\leq\!\sqrt{2}M_{0}\|\omega_{\mathbf{j}}\|\!\sum_{p=1}^{P}\!\frac{\|\widetilde{X}^{p}_{t_{n}}\!-\!X^{p}_{t_{n}}\|}{P}. (36)

Let Yp=X~tnpXtnpY_{p}=\|\widetilde{X}_{t_{n}}^{p}-X_{t_{n}}^{p}\|, where {Yp}p=1P\{Y_{p}\}_{p=1}^{P} are i.i.d. random variables. This follows from the fact that the particles {Xtnp}p=1P\{X_{t_{n}}^{p}\}_{p=1}^{P} and {X~tnp}p=1P\{\widetilde{X}_{t_{n}}^{p}\}_{p=1}^{P} are separately i.i.d. Specifically, the i.i.d. property of {X~tnp}p=1P\{\widetilde{X}_{t_{n}}^{p}\}_{p=1}^{P} is ensured by the RBM described in Alg. 2. Based on Assumption 1, YpY_{p} is bounded. The empirical mean is defined as Y¯P=1Pp=1PYp\bar{Y}_{P}=\frac{1}{P}\sum_{p=1}^{P}Y_{p} and the expectation of YpY_{p} is μ=𝔼[Yp]=𝔼[X~tnXtn]\mu=\mathbb{E}[Y_{p}]=\mathbb{E}[\|\widetilde{X}_{t_{n}}-X_{t_{n}}\|].

According to Bernstein’s inequality, for i.i.d. random variables Y1,Y2,,YPY_{1},Y_{2},\dots,Y_{P} with |Ypμ|M1|Y_{p}-\mu|\leq M_{1} (from Assumption 1) almost surely, the probability that the empirical mean deviates from the expectation is bounded as

(|Y¯Pμ|η)2exp(Pη22σ2+2M1η3),\mathbb{P}\left(|\bar{Y}_{P}-\mu|\geq\eta\right)\leq 2\exp\left(-\frac{P\eta^{2}}{2\sigma^{2}+\frac{2M_{1}\eta}{3}}\right), (37)

where σ2=𝔼[(Ypμ)2]M12\sigma^{2}=\mathbb{E}[(Y_{p}-\mu)^{2}]\leq M_{1}^{2}. With high probability (e.g., 1δ01-\delta_{0} for very small δ0>0\delta_{0}>0), the following estimate holds

|Y¯Pμ|2σ2ln(2/δ0)P+2M1ln(2/δ0)3P.|\bar{Y}_{P}-\mu|\leq\sqrt{\frac{2\sigma^{2}\ln(2/\delta_{0})}{P}}+\frac{2M_{1}\ln(2/\delta_{0})}{3P}. (38)

This implies that, with probability 1δ01-\delta_{0},

|𝐣[ρ(tn)]𝐣[ρ~n]|2M0ω𝐣(𝔼[X~tnXtn]+2σ2ln(2/δ0)P+2M1ln(2/δ0)3P).|\mathcal{F}_{\mathbf{j}}[\rho(t_{n})]\!-\!\mathcal{F}_{\mathbf{j}}[\widetilde{\rho}_{n}]|\leq\sqrt{2}M_{0}\|\omega_{\mathbf{j}}\|\left(\!\mathbb{E}[\|\widetilde{X}_{t_{n}}-X_{t_{n}}\|]\!+\!\sqrt{\frac{2\sigma^{2}\ln(2/\delta_{0})}{P}}\!+\!\frac{2M_{1}\ln(2/\delta_{0})}{3P}\!\right). (39)

Combining Eqs. (29)(3.2)(39) above, we conclude that, with high probability

|α~tn;𝐣αtn;𝐣|\displaystyle|\widetilde{\alpha}_{t_{n};\mathbf{j}}-\alpha_{t_{n};\mathbf{j}}|\leq |α~tn1;𝐣αtn1;𝐣|+C1δt2+C2ω𝐣δtP+C3ω𝐣δt𝔼[X~tnXtn],\displaystyle|\widetilde{\alpha}_{t_{n-1};\mathbf{j}}-\alpha_{t_{n-1};\mathbf{j}}|+C_{1}\delta t^{2}+\frac{C_{2}\|\omega_{\mathbf{j}}\|\delta t}{\sqrt{P}}+C_{3}\|\omega_{\mathbf{j}}\|\delta t\mathbb{E}[\|\widetilde{X}_{t_{n}}-X_{t_{n}}\|],

where C2=2ln(2/δ0)M02+2M0M1ln(2/δ0)3C_{2}=2\sqrt{\ln(2/\delta_{0})}M_{0}^{2}+\frac{2M_{0}M_{1}\ln(2/\delta_{0})}{3} and C3=2M0ϵC_{3}=\frac{\sqrt{2}M_{0}}{\epsilon} are constants.

The error estimate between c\nabla c and c~\nabla\widetilde{c} is more complex than that between cc and c~\widetilde{c}. To analyze this, we introduce an intermediate quantity c\nabla c^{\ast}. Using the frequency notation ω𝐣\omega_{\mathbf{j}} from Lemma 6, where ω𝐣=(2πj1L,2πj2L,2πj3L)\omega_{\mathbf{j}}=\left(\frac{2\pi j_{1}}{L},\frac{2\pi j_{2}}{L},\frac{2\pi j_{3}}{L}\right), we define

c(𝐱,tn):\displaystyle\nabla c^{\ast}(\mathbf{x},t_{n}): =1L3𝐣iω𝐣α~n;𝐣exp(iω𝐣𝐱)\displaystyle=\frac{1}{L^{3}}\sum_{\mathbf{j}\in\mathcal{H}}i\omega_{\mathbf{j}}\widetilde{\alpha}_{n;\mathbf{j}}\exp(i\omega_{\mathbf{j}}\mathbf{x})
=ϵδt𝐱𝒦ϵ,δt(𝐱𝐲)c~n1(𝐲)d𝐲q=1PM0P𝐱𝒦ϵ,δt(𝐱X~nq)\displaystyle=-\frac{\epsilon}{\delta t}\int\nabla_{\mathbf{x}}\mathcal{K}_{\epsilon,\delta t}(\mathbf{x}-\mathbf{y})\widetilde{c}_{n-1}(\mathbf{y})\,d\mathbf{y}-\sum_{q=1}^{P}\frac{M_{0}}{P}\nabla_{\mathbf{x}}\mathcal{K}_{\epsilon,\delta t}(\mathbf{x}-\widetilde{X}^{q}_{n})
=ϵδt𝐱𝒦ϵ,δt(𝐱+𝐱¯𝐲)c~n1(𝐲𝐱¯)d𝐲I1q=1PM0P𝐱𝒦ϵ,δt(𝐱X~nq)I2,\displaystyle=-\frac{\epsilon}{\delta t}\underbrace{\int\nabla_{\mathbf{x}}\mathcal{K}_{\epsilon,\delta t}(\mathbf{x}+\bar{\mathbf{x}}-\mathbf{y})\widetilde{c}_{n-1}(\mathbf{y}-\bar{\mathbf{x}})\,d\mathbf{y}}_{\mathclap{\textstyle I_{1}}}-\underbrace{\sum_{q=1}^{P}\frac{M_{0}}{P}\nabla_{\mathbf{x}}\mathcal{K}_{\epsilon,\delta t}(\mathbf{x}-\widetilde{X}^{q}_{n})}_{\mathclap{\textstyle I_{2}}}, (40)

where 𝐱¯=L2H+𝐱L/HLH𝐱\bar{\mathbf{x}}=\frac{L}{2H}+\lfloor\frac{\mathbf{x}}{L/H}\rfloor\frac{L}{H}-\mathbf{x}. From Alg. 2, it follows that

c~(𝐱,tn)=\displaystyle\nabla\widetilde{c}(\mathbf{x},t_{n})= 𝐱𝒦ϵ,δt(ϵc~n1(𝐱)/δt+ρ~n(𝐱))\displaystyle-\nabla_{\mathbf{x}}\mathcal{K}_{\epsilon,\delta t}\ast(\epsilon\,\widetilde{c}_{n-1}(\mathbf{x})/\delta t+\widetilde{\rho}_{n}(\mathbf{x}))
=\displaystyle= ϵδtL3H3𝐣𝐱𝒦ϵ,δt(𝐱+𝐱¯x𝐣)c~n1(x𝐣𝐱¯)I3\displaystyle-\frac{\epsilon}{\delta t}\underbrace{\frac{L^{3}}{H^{3}}\sum_{\mathbf{j}\in\mathcal{H}}\nabla_{\mathbf{x}}\mathcal{K}_{\epsilon,\delta t}(\mathbf{x}+\bar{\mathbf{x}}-x_{\mathbf{j}})\widetilde{c}_{n-1}(x_{\mathbf{j}}-\bar{\mathbf{x}})}_{\mathclap{\textstyle I_{3}}}
sCp,spM0R𝐱𝒦ϵ,δt(𝐱X~ns)I4,\displaystyle-\underbrace{\sum_{s\in C_{p},s\neq p}\frac{M_{0}}{R}\nabla_{\mathbf{x}}\mathcal{K}_{\epsilon,\delta t}(\mathbf{x}-\widetilde{X}^{s}_{n})}_{\mathclap{\textstyle I_{4}}}, (41)

where x𝐣=(j1LH,j2LH,j3LH)x_{\mathbf{j}}=(\frac{j_{1}L}{H},\frac{j_{2}L}{H},\frac{j_{3}L}{H}). The error between c\nabla c and c~\nabla\widetilde{c} can be estimated by

c(𝐱,tn)c~(𝐱,tn)c(𝐱,tn)c(𝐱,tn)+c(𝐱,tn)c~(𝐱,tn).\|\nabla c(\mathbf{x},t_{n})-\nabla\widetilde{c}(\mathbf{x},t_{n})\|\leq\|\nabla c(\mathbf{x},t_{n})-\nabla c^{\ast}(\mathbf{x},t_{n})\|+\|\nabla c^{\ast}(\mathbf{x},t_{n})-\nabla\widetilde{c}(\mathbf{x},t_{n})\|. (42)

To estimate the error between c~\nabla\widetilde{c} and c\nabla c^{\ast}, we divide the analysis into two parts:

c~(𝐱,tn)c(𝐱,tn)ϵδtI1I3+I2I4.\displaystyle\|\nabla\widetilde{c}(\mathbf{x},t_{n})-\nabla c^{\ast}(\mathbf{x},t_{n})\|\leq\frac{\epsilon}{\delta t}\|I_{1}-I_{3}\|+\|I_{2}-I_{4}\|. (43)

The first part, involving I1I_{1} and I3I_{3}, focuses on the spatial discretization error of the continuous convolution (𝐱Kϵ,δtc~n1)(\nabla_{\mathbf{x}}K_{\epsilon,\delta t}\ast\widetilde{c}_{n-1}) on the periodic domain Ω=[L/2,L/2]3\Omega=[-L/2,L/2]^{3}. Specifically, I1I_{1} represents the continuous integral, while I3I_{3} is its discrete approximation. Since c~n1\widetilde{c}_{n-1} is strictly represented by a truncated Fourier series with modes in \mathcal{H}, I3I_{3} can be viewed as a discrete Fourier projection of the continuous convolution. Thanks to the spatial shift that regularizes the singular gradient kernel on the grid, the quadrature error introduced by this discrete evaluation is of the same order as the Fourier truncation error, and is thus controlled by the spectral projection Π\Pi_{\mathcal{H}}.

To analyze the error introduced by this approximation, we rely on the following lemma.

Lemma 7.

For all n+n\in\mathbb{N}^{+}, under Assumption 2, based on the definitions of I1I_{1} and I3I_{3} given in Eq. (3.2) and Eq. (3.2), the following spatial discretization error bound holds:

ϵδtI1I3C4H2,\frac{\epsilon}{\delta t}\|I_{1}-I_{3}\|\leq\frac{C_{4}}{H^{2}}, (44)

where C4C_{4} is a constant, and tn=nδtt_{n}=n\delta t.

Proof.

Define the continuous convolution operator 𝒯f:=ϵδt𝐱Kϵ,δtf\mathcal{T}f:=\frac{\epsilon}{\delta t}\nabla_{\mathbf{x}}K_{\epsilon,\delta t}*f. Its Fourier multiplier is M(ω)=iω(ϵ/δt)ω2+λ2+ϵ/δtM(\omega)=\frac{-i\omega(\epsilon/\delta t)}{\|\omega\|^{2}+\lambda^{2}+\epsilon/\delta t}, which satisfies |M(ω)|ω|M(\omega)|\leq\|\omega\| for all δt>0\delta t>0. Thus, ϵδtI1=𝒯(c~n1)\frac{\epsilon}{\delta t}I_{1}=\mathcal{T}(\tilde{c}_{n-1}).

The discrete sum I3I_{3} evaluates the convolution on a uniform grid with spacing Δx=L/H\Delta x=L/H. The spatial shift x¯\bar{x} offsets the quadrature points by half a grid spacing, i.e., L2H\frac{L}{2H}. Since the kernel 𝐱Kϵ,δt\nabla_{\mathbf{x}}K_{\epsilon,\delta t} is radially antisymmetric, this half-grid shift regularizes the origin singularity and induces symmetric cancellation. Consequently, the spatial discretization error is dominated by the Fourier truncation error of the spectral projection Π\Pi_{\mathcal{H}}.

For the truncated modes outside \mathcal{H}, the frequency magnitude satisfies ωj>πHL\|\omega_{j}\|>\frac{\pi H}{L}. Applying Parseval’s identity yields the explicit projection error bound:

ϵδtI1I3𝒯(c~n1)Π𝒯(c~n1)1π2(LH)22𝒯(c~n1).\displaystyle\frac{\epsilon}{\delta t}\|I_{1}-I_{3}\|\leq\|\mathcal{T}(\tilde{c}_{n-1})-\Pi_{\mathcal{H}}\mathcal{T}(\tilde{c}_{n-1})\|\leq\frac{1}{\pi^{2}}\left(\frac{L}{H}\right)^{2}\|{\color[rgb]{0,0,0}\nabla^{2}\mathcal{T}(\tilde{c}_{n-1})}\|. (45)

Since |M(ω)|ω|M(\omega)|\leq\|\omega\|, the term 2𝒯(c~n1)\|{\color[rgb]{0,0,0}\nabla^{2}\mathcal{T}(\tilde{c}_{n-1})}\| is controlled by the higher-order spatial regularity of c~n1\tilde{c}_{n-1} stipulated in Assumption 2(a)(c). Combining these estimates, we obtain

ϵδtI1I3C4H2,\displaystyle\frac{\epsilon}{\delta t}\|I_{1}-I_{3}\|\leq\frac{C_{4}}{H^{2}}, (46)

where C4=1π2L2(K+M3)C_{4}=\frac{1}{\pi^{2}}L^{2}(K+M_{3}) is a constant.

We now estimate (𝐱𝒦ϵ,δtρ~n)(X~np)(\nabla_{\mathbf{x}}\mathcal{K}_{\epsilon,\delta t}\ast\,\widetilde{\rho}_{n})(\widetilde{X}^{p}_{n}) in c~\nabla\widetilde{c} and c\nabla c^{\ast}. Using the RBM in Alg. 2, we replace q=1,qpPM0P𝐱𝒦ϵ,δt(X~npX~nq)\sum_{q=1,q\not=p}^{P}\frac{M_{0}}{P}\nabla_{\mathbf{x}}\mathcal{K}_{\epsilon,\delta t}(\widetilde{X}^{p}_{n}-\widetilde{X}^{q}_{n}) with sCp,spM0R𝐱𝒦ϵ,δt(X~npX~ns)\sum_{s\in C_{p},s\neq p}\frac{M_{0}}{R}\nabla_{\mathbf{x}}\mathcal{K}_{\epsilon,\delta t}(\widetilde{X}^{p}_{n}-\widetilde{X}^{s}_{n}). We write

ζn,p:=q=1,qpPM0P𝐱𝒦ϵ,δt(X~npX~nq)sCp,spM0R𝐱𝒦ϵ,δt(X~npX~ns).\zeta_{n,p}:=\sum_{q=1,q\not=p}^{P}\frac{M_{0}}{P}\nabla_{\mathbf{x}}\mathcal{K}_{\epsilon,\delta t}(\widetilde{X}^{p}_{n}-\widetilde{X}^{q}_{n})-\sum_{s\in C_{p},s\neq p}\frac{M_{0}}{R}\nabla_{\mathbf{x}}\mathcal{K}_{\epsilon,\delta t}(\widetilde{X}^{p}_{n}-\widetilde{X}^{s}_{n}). (47)
Lemma 8.

For all n+n\in\mathbb{N_{+}}, p{1,2,,P}p\in\{1,2,\dots,P\}, we have the estimate as follows:

𝔼(ζn,p)M0M41R(11P),\displaystyle\mathbb{E}(\|\zeta_{n,p}\|)\leq M_{0}M_{4}\sqrt{\frac{1}{R}\left(1-\frac{1}{P}\right)}, (48)

where M4=maxqp𝐱𝒦ϵ,δt(X~npX~nq)M_{4}=\max_{q\neq p}\|\nabla_{\mathbf{x}}\mathcal{K}_{\epsilon,\delta t}(\widetilde{X}^{p}_{n}-\widetilde{X}^{q}_{n})\|, M0M_{0} is the conserved total mass, PP is the total number of particles, and RR is the batch size.

Proof.

Similar to Lemma 3.1 in [24], we rewrite

fp=sCp,spM0R𝐱𝒦ϵ,δt(X~npX~ns)=q=1,qpPM0R𝐱𝒦ϵ,δt(X~npX~nq)Nq,\displaystyle f_{p}=\sum_{s\in C_{p},s\neq p}\frac{M_{0}}{R}\nabla_{\mathbf{x}}\mathcal{K}_{\epsilon,\delta t}(\widetilde{X}^{p}_{n}-\widetilde{X}^{s}_{n})=\sum_{q=1,q\not=p}^{P}\frac{M_{0}}{R}\nabla_{\mathbf{x}}\mathcal{K}_{\epsilon,\delta t}(\widetilde{X}^{p}_{n}-\widetilde{X}^{q}_{n})N_{q}, (49)

where NqN_{q} denotes the number of times particle qq is selected in the batch CpC_{p}. Since the sampling is with replacement, NqN_{q} follows a Binomial distribution B(R,1/P)B(R,1/P) with 𝔼[Nq]=RP\mathbb{E}[N_{q}]=\frac{R}{P}, which indicates that 𝔼[ζn,p]=0\mathbb{E}[\zeta_{n,p}]=0.

𝔼fp2=M02R2q,r:qr,qp,rp𝐱𝒦ϵ,δt(X~pnX~qn)𝐱𝒦ϵ,δt(X~pnX~rn)𝔼[NqNr]+M02R2q=1,qpP𝐱𝒦ϵ,δt(X~pnX~qn)2𝔼[Nq2]=M02(R1)RP2q,r:qr,qp,rp𝐱𝒦ϵ,δt(X~pnX~qn)𝐱𝒦ϵ,δt(X~pnX~rn)+M02(P1+R)RP2q=1,qpP𝐱𝒦ϵ,δt(X~pnX~qn)2.\displaystyle\begin{split}\mathbb{E}\|f_{p}\|^{2}=&\frac{M_{0}^{2}}{R^{2}}\sum_{\begin{subarray}{c}q,r:\\ q\neq r,q\neq p,\\ r\neq p\end{subarray}}\|\nabla_{\mathbf{x}}\mathcal{K}_{\epsilon,\delta t}(\widetilde{X}^{p}_{n}-\widetilde{X}^{q}_{n})\cdot\nabla_{\mathbf{x}}\mathcal{K}_{\epsilon,\delta t}(\widetilde{X}^{p}_{n}-\widetilde{X}^{r}_{n})\|\mathbb{E}[N_{q}N_{r}]\\ &+\frac{M_{0}^{2}}{R^{2}}\sum_{q=1,q\neq p}^{P}\|\nabla_{\mathbf{x}}\mathcal{K}_{\epsilon,\delta t}(\widetilde{X}^{p}_{n}-\widetilde{X}^{q}_{n})\|^{2}\mathbb{E}[N_{q}^{2}]\\ =&\frac{M_{0}^{2}(R-1)}{RP^{2}}\sum_{\begin{subarray}{c}q,r:\\ q\neq r,q\neq p,\\ r\neq p\end{subarray}}\|\nabla_{\mathbf{x}}\mathcal{K}_{\epsilon,\delta t}(\widetilde{X}^{p}_{n}-\widetilde{X}^{q}_{n})\cdot\nabla_{\mathbf{x}}\mathcal{K}_{\epsilon,\delta t}(\widetilde{X}^{p}_{n}-\widetilde{X}^{r}_{n})\|\\ &+\frac{M_{0}^{2}(P-1+R)}{RP^{2}}\sum_{q=1,q\not=p}^{P}\|\nabla_{\mathbf{x}}\mathcal{K}_{\epsilon,\delta t}(\widetilde{X}^{p}_{n}-\widetilde{X}^{q}_{n})\|^{2}.\end{split}

Hence,

Var(ζn,p)=𝔼fp2𝔼fp2\displaystyle\text{Var}(\zeta_{n,p})=\mathbb{E}\|f_{p}\|^{2}-\|\mathbb{E}f_{p}\|^{2} =M021R(11P)1Pq=1,qpP𝐱𝒦ϵ,δt(X~npX~nq)2\displaystyle=M_{0}^{2}\frac{1}{R}\left(1-\frac{1}{P}\right)\frac{1}{P}\sum_{q=1,q\not=p}^{P}\|\nabla_{\mathbf{x}}\mathcal{K}_{\epsilon,\delta t}(\widetilde{X}^{p}_{n}-\widetilde{X}^{q}_{n})\|^{2}
M02RP2q=1,qpP𝐱𝒦ϵ,δt(X~npX~nq)2\displaystyle\quad-\frac{M_{0}^{2}}{RP^{2}}\left\|\sum_{q=1,q\not=p}^{P}\nabla_{\mathbf{x}}\mathcal{K}_{\epsilon,\delta t}(\widetilde{X}^{p}_{n}-\widetilde{X}^{q}_{n})\right\|^{2}
M021R(11P)1Pq=1,qpP𝐱𝒦ϵ,δt(X~npX~nq)2.\displaystyle\leq M_{0}^{2}\frac{1}{R}\left(1-\frac{1}{P}\right)\frac{1}{P}\sum_{q=1,q\not=p}^{P}\|\nabla_{\mathbf{x}}\mathcal{K}_{\epsilon,\delta t}(\widetilde{X}^{p}_{n}-\widetilde{X}^{q}_{n})\|^{2}. (50)

According to Jensen’s inequality, we obtain:

𝔼(ζn,p)𝔼(ζn,p2)=Var(ζn,p)M0M41R(11P),\displaystyle\mathbb{E}(\|\zeta_{n,p}\|)\leq\sqrt{\mathbb{E}(\|\zeta_{n,p}\|^{2})}=\sqrt{\text{Var}(\zeta_{n,p})}\leq M_{0}M_{4}\sqrt{\frac{1}{R}\left(1-\frac{1}{P}\right)}, (51)

where M4=maxqp𝐱𝒦ϵ,δt(X~npX~nq)M_{4}=\max_{q\neq p}\|\nabla_{\mathbf{x}}\mathcal{K}_{\epsilon,\delta t}(\widetilde{X}^{p}_{n}-\widetilde{X}^{q}_{n})\|. Since all particles are located at distinct positions in the SIPF-rr algorithm (X~npX~nq\widetilde{X}^{p}_{n}\neq\widetilde{X}^{q}_{n} for pqp\neq q), there exists a minimum separation distance dmin>0d_{\text{min}}>0 between any two distinct particles. Consequently, 𝐱𝒦ϵ,δt(X~npX~nq)\|\nabla_{\mathbf{x}}\mathcal{K}_{\epsilon,\delta t}(\widetilde{X}^{p}_{n}-\widetilde{X}^{q}_{n})\| is bounded for all pairs of particles. This ensures that M4M_{4}, which is the maximum of these kernel gradient norms, is finite.

3.3 Proof of the Main Theorem

Building on the lemmas established in the previous subsection, we now prove the main theorem in this section. For simplicity of notation in the proof below, we define:

an\displaystyle a_{n} :=𝔼(X~tnXtn),\displaystyle:=\mathbb{E}(\|\widetilde{X}_{t_{n}}-X_{t_{n}}\|), (52)
bn\displaystyle b_{n} :=𝔼(c(X~tn,tn)c(X~tn,tn)).\displaystyle:=\mathbb{E}(\|\nabla c(\widetilde{X}_{t_{n}},t_{n})-\nabla c^{\ast}(\widetilde{X}_{t_{n}},t_{n})\|). (53)

The following provides a bound on the error between X~tn+1\widetilde{X}_{t_{n+1}} and Xtn+1X_{t_{n+1}}.

Lemma 9.

For all n+n\in\mathbb{N_{+}}, under Assumption 2,

an+1=𝔼(X~tn+1Xtn+1)\displaystyle a_{n+1}=\mathbb{E}(\|\widetilde{X}_{t_{n+1}}-X_{t_{n+1}}\|)\!\leq (1+C5δt)an+χδtbn+C6δt32+C7δtH2+C8δtR,\displaystyle(1+C_{5}\delta t)a_{n}+\chi\delta tb_{n}+C_{6}\delta t^{\frac{3}{2}}+\frac{C_{7}\delta t}{H^{2}}+\frac{C_{8}\delta t}{\sqrt{R}}, (54)

where C5C_{5}, C6C_{6}, C7C_{7}, and C8C_{8} are constants, χ\chi is the chemotaxis coefficient in the system (1), and tn=nδtt_{n}=n\delta t.

Proof.

According to Eqs. (16)-(17), it follows that

𝔼(X~tn+1Xtn+1)\displaystyle\mathbb{E}(\|\widetilde{X}_{t_{n+1}}-X_{t_{n+1}}\|)\leq 𝔼(X~tnXtn)+χ𝔼(tntn+1c~(X~tn,tn)c(Xs,s)𝑑s)\displaystyle\mathbb{E}(\|\widetilde{X}_{t_{n}}-X_{t_{n}}\|)+\chi\mathbb{E}(\int_{t_{n}}^{t_{n+1}}\|\nabla\widetilde{c}(\widetilde{X}_{t_{n}},t_{n})-\nabla c(X_{s},s)\|\,ds)
=\displaystyle= 𝔼(X~tnXtn)+χtntn+1𝔼(c~(X~tn,tn)c(Xs,s))𝑑s,\displaystyle\mathbb{E}(\|\widetilde{X}_{t_{n}}-X_{t_{n}}\|)+\chi\int_{t_{n}}^{t_{n+1}}\mathbb{E}(\|\nabla\widetilde{c}(\widetilde{X}_{t_{n}},t_{n})-\nabla c(X_{s},s)\|)\,ds, (55)

by the triangle inequality and Tonelli’s theorem. To bound the integrand, we decompose the gradient difference using the triangle inequality:

c~(X~tn,tn)c(Xs,s)\displaystyle\|\nabla\widetilde{c}(\widetilde{X}_{t_{n}},t_{n})\!-\!\nabla c(X_{s},s)\|\!\leq c~(X~tn,tn)c(X~tn,tn)+c(X~tn,tn)c(Xtn,tn)\displaystyle\|\nabla\widetilde{c}(\widetilde{X}_{t_{n}},t_{n})\!-\!\nabla c(\widetilde{X}_{t_{n}},t_{n})\|\!+\!\|\nabla c(\widetilde{X}_{t_{n}},t_{n})\!-\!\nabla c(X_{t_{n}},t_{n})\|
+c(Xtn,tn)c(Xs,s).\displaystyle+\|\nabla c(X_{t_{n}},t_{n})-\nabla c(X_{s},s)\|. (56)

We now estimate the expectation of each term on the right-hand side of Eq. (3.3). Regarding the first term, recalling the error decomposition in Eqs. (42)-(43), along with the bounds from Lemmas 7 and 8, we have

𝔼(c~(X~tn,tn)c(X~tn,tn))\displaystyle\mathbb{E}(\|\nabla\widetilde{c}(\widetilde{X}_{t_{n}},t_{n})\!-\!\nabla c(\widetilde{X}_{t_{n}},t_{n})\|)\leq 𝔼(c~(X~tn,tn)c(X~tn,tn))\displaystyle\mathbb{E}(\|\nabla\widetilde{c}(\widetilde{X}_{t_{n}},t_{n})-\nabla c^{\ast}(\widetilde{X}_{t_{n}},t_{n})\|)
+𝔼(c(X~tn,tn)c(X~tn,tn))\displaystyle+\mathbb{E}(\|\nabla c^{\ast}(\widetilde{X}_{t_{n}},t_{n})-\nabla c(\widetilde{X}_{t_{n}},t_{n})\|)
\displaystyle\leq bn+C4H2+M0M41R(11P).\displaystyle b_{n}+\frac{C_{4}}{H^{2}}+M_{0}M_{4}\sqrt{\frac{1}{R}\left(1-\frac{1}{P}\right)}. (57)

For the second term, the spatial Lipschitz continuity in Assumption 2(a) yields

𝔼(c(X~tn,tn)c(Xtn,tn))K𝔼(X~tnXtn),\displaystyle\mathbb{E}(\|\nabla c(\widetilde{X}_{t_{n}},t_{n})\!-\!\nabla c(X_{t_{n}},t_{n})\|)\leq K\mathbb{E}(\|\widetilde{X}_{t_{n}}-X_{t_{n}}\|), (58)

where KK is the Lipschitz constant defined in Assumption 2(a). By the Lipschitz conditions in Assumption 2(a)(b), the exact dynamics in Eq. (17), and the property 𝔼(WsWtn)3(stn)\mathbb{E}(\|W_{s}-W_{t_{n}}\|)\leq\sqrt{3(s-t_{n})} of the 3D Brownian motion, the temporal variation of the exact gradient satisfies

𝔼(c(Xtn,tn)c(Xs,s))(χKM3+K1)(stn)+K6μstn.\displaystyle\mathbb{E}(\|\nabla c(X_{t_{n}},t_{n})-\nabla c(X_{s},s)\|)\leq(\chi KM_{3}+K_{1})(s-t_{n})+K\sqrt{6\mu}\sqrt{s-t_{n}}. (59)

Substituting the bounds Eqs. (3.3)-(59) into the integral and evaluating over the time interval [tn,tn+1][t_{n},t_{n+1}] of length δt\delta t, we can obtain the estimate as follows:

an+1=\displaystyle a_{n+1}= 𝔼(X~tn+1Xtn+1)\displaystyle\mathbb{E}(\|\widetilde{X}_{t_{n+1}}-X_{t_{n+1}}\|)
\displaystyle\leq (1+χKδt)an+χδt(bn+C4H2+M0M41R(11P))+C6δt32\displaystyle(1+\chi K\delta t)a_{n}+\chi\delta t\left(b_{n}+\frac{C_{4}}{H^{2}}+M_{0}M_{4}\sqrt{\frac{1}{R}\left(1-\frac{1}{P}\right)}\right)+C_{6}\delta t^{\frac{3}{2}}
\displaystyle\leq (1+C5δt)an+χδtbn+C6δt32+C7δtH2+C8δtR,\displaystyle(1+C_{5}\delta t)a_{n}+\chi\delta tb_{n}+C_{6}\delta t^{\frac{3}{2}}+\frac{C_{7}\delta t}{H^{2}}+\frac{C_{8}\delta t}{\sqrt{R}}, (60)

where the constants are defined as C5=χKC_{5}=\chi K, C6=χ(KχM3+K1)2+2χK6μ3C_{6}=\frac{\chi(K\chi M_{3}+K_{1})}{2}+\frac{2\chi K\sqrt{6\mu}}{3}, C7=χC4C_{7}=\chi C_{4}, and C8=χM0M4C_{8}=\chi M_{0}M_{4}. This concludes the proof.

Now we analyze the error between c\nabla c^{\ast} and c\nabla c as follows.

We next estimate the error between c\nabla c^{\ast} and c\nabla c.

Lemma 10.

For all n+n\in\mathbb{N}_{+}, under Assumption 2, with high probability,

bn=𝔼(c(X~tn,tn)c(X~tn,tn))k=1nC11δtnk+1ak+C10P+C9δt,\displaystyle b_{n}=\mathbb{E}\bigl(\|\nabla c(\widetilde{X}_{t_{n}},t_{n})-\nabla c^{\ast}(\widetilde{X}_{t_{n}},t_{n})\|\bigr)\leq\sum_{k=1}^{n}\frac{C_{11}\sqrt{\delta t}}{\sqrt{n-k+1}}a_{k}+\frac{C_{10}}{\sqrt{P}}+C_{9}\delta t, (61)

where C9C_{9}, C10C_{10}, and C11C_{11} are positive constants independent of HH, PP, and δt\delta t.

Proof.

Subtracting Eq. (26) from Eq. (30), multiplying by iω𝐣i\omega_{\mathbf{j}}, and using the notation Z𝐣Z_{\mathbf{j}} defined in the proof of Lemma 6 give

iω𝐣(αtn;𝐣α~tn;𝐣)=iω𝐣1+Z𝐣[αtn1;𝐣α~tn1;𝐣+δtϵ(𝐣[ρ(tn)]𝐣[ρ~n])+δt𝐣[Rn]].\displaystyle i\omega_{\mathbf{j}}\bigl(\alpha_{t_{n};\mathbf{j}}-\widetilde{\alpha}_{t_{n};\mathbf{j}}\bigr)=\frac{i\omega_{\mathbf{j}}}{1+Z_{\mathbf{j}}}\Bigl[\alpha_{t_{n-1};\mathbf{j}}-\widetilde{\alpha}_{t_{n-1};\mathbf{j}}+\frac{\delta t}{\epsilon}\bigl(\mathcal{F}_{\mathbf{j}}[\rho(t_{n})]-\mathcal{F}_{\mathbf{j}}[\widetilde{\rho}_{n}]\bigr)+\delta t\,\mathcal{F}_{\mathbf{j}}[R_{n}]\Bigr]. (62)

Since the initial chemical fields coincide, iteration of Eq. (62) yields

iω𝐣(αtn;𝐣α~tn;𝐣)=k=1niω𝐣(1+Z𝐣)nk+1[δtϵ(𝐣[ρk]𝐣[ρ~k])+δt𝐣[Rk]].\displaystyle i\omega_{\mathbf{j}}\bigl(\alpha_{t_{n};\mathbf{j}}-\widetilde{\alpha}_{t_{n};\mathbf{j}}\bigr)=\sum_{k=1}^{n}\frac{i\omega_{\mathbf{j}}}{(1+Z_{\mathbf{j}})^{n-k+1}}\Bigl[\frac{\delta t}{\epsilon}\bigl(\mathcal{F}_{\mathbf{j}}[\rho_{k}]-\mathcal{F}_{\mathbf{j}}[\widetilde{\rho}_{k}]\bigr)+\delta t\,\mathcal{F}_{\mathbf{j}}[R_{k}]\Bigr]. (63)

By the inverse Fourier representation in Eqs. (3)–(5) and the definition of bnb_{n},

bn=𝔼[1L3𝐣iω𝐣(αtn;𝐣α~tn;𝐣)eiω𝐣X~tn].b_{n}=\mathbb{E}\!\left[\left\|\frac{1}{L^{3}}\sum_{\mathbf{j}\in\mathcal{H}}i\omega_{\mathbf{j}}\bigl(\alpha_{t_{n};\mathbf{j}}-\widetilde{\alpha}_{t_{n};\mathbf{j}}\bigr)e^{i\omega_{\mathbf{j}}\cdot\widetilde{X}_{t_{n}}}\right\|\right]. (64)

By Eq. (39) in Lemma 6, with the definitions of aka_{k}, C2C_{2}, and C3C_{3}, we have, with high probability,

1ϵ|𝐣[ρk]𝐣[ρ~k]|ω𝐣(C3ak+C2P).\frac{1}{\epsilon}\left|\mathcal{F}_{\mathbf{j}}[\rho_{k}]-\mathcal{F}_{\mathbf{j}}[\widetilde{\rho}_{k}]\right|\leq\|\omega_{\mathbf{j}}\|\left(C_{3}a_{k}+\frac{C_{2}}{\sqrt{P}}\right). (65)

Substituting the density-source term in Eq. (63) into the inverse Fourier representation in Eqs. (3)–(5) and evaluating the resulting gradient at X~tn\widetilde{X}_{t_{n}} gives its contribution to bnb_{n}. Applying Eq. (65) together with the decay of the resolvent factor gives, for m=nk+1m=n-k+1,

𝔼[1L3𝐣iω𝐣δt/ϵ(1+Z𝐣)m(𝐣[ρk]𝐣[ρ~k])eiω𝐣X~tn]\displaystyle\mathbb{E}\!\left[\left\|\frac{1}{L^{3}}\sum_{\mathbf{j}\in\mathcal{H}}\frac{i\omega_{\mathbf{j}}\delta t/\epsilon}{(1+Z_{\mathbf{j}})^{m}}\bigl(\mathcal{F}_{\mathbf{j}}[\rho_{k}]-\mathcal{F}_{\mathbf{j}}[\widetilde{\rho}_{k}]\bigr)e^{i\omega_{\mathbf{j}}\cdot\widetilde{X}_{t_{n}}}\right\|\right]
δtm(1L3+L4π2+L720)1/2[C3ak+1P(C2+2M0ϵ2ln(4/δ0))].\displaystyle\qquad\leq\frac{\sqrt{\delta t}}{\sqrt{m}}\left(\frac{1}{L^{3}}+\frac{L}{4\pi^{2}}+\frac{L}{720}\right)^{1/2}\left[C_{3}a_{k}+\frac{1}{\sqrt{P}}\left(C_{2}+\frac{2M_{0}}{\epsilon}\sqrt{2\ln(4/\delta_{0})}\right)\right]. (66)

The Fourier sum in Eq. (3.3) can be estimated by grouping the indices according to =𝐣\ell=\|\mathbf{j}\|_{\infty}. For 1\ell\geq 1, the shell 𝐣=\|\mathbf{j}\|_{\infty}=\ell contains 242+224\ell^{2}+2 indices and ω𝐣24π22/L2\|\omega_{\mathbf{j}}\|^{2}\geq 4\pi^{2}\ell^{2}/L^{2}. Hence

1L3𝐣31(1+ω𝐣2)2\displaystyle\frac{1}{L^{3}}\sum_{\mathbf{j}\in\mathbb{Z}^{3}}\frac{1}{(1+\|\omega_{\mathbf{j}}\|^{2})^{2}} 1L3+L16π4=1(242+24)=1L3+L4π2+L720.\displaystyle\leq\frac{1}{L^{3}}+\frac{L}{16\pi^{4}}\sum_{\ell=1}^{\infty}\left(\frac{24}{\ell^{2}}+\frac{2}{\ell^{4}}\right)=\frac{1}{L^{3}}+\frac{L}{4\pi^{2}}+\frac{L}{720}. (67)

Thus, the series converges. Summing Eq. (3.3) over kk and using δtm=1nm1/22T\sqrt{\delta t}\sum_{m=1}^{n}m^{-1/2}\leq 2\sqrt{T} gives

k=1n𝔼[1L3𝐣iω𝐣δt/ϵ(1+Z𝐣)nk+1(𝐣[ρk]𝐣[ρ~k])eiω𝐣X~tn]k=1nC11δtnk+1ak+C10P.\displaystyle\sum_{k=1}^{n}\mathbb{E}\!\left[\left\|\frac{1}{L^{3}}\!\sum_{\mathbf{j}\in\mathcal{H}}\frac{i\omega_{\mathbf{j}}\delta t/\epsilon}{(1+Z_{\mathbf{j}})^{n-k+1}}\bigl(\mathcal{F}_{\mathbf{j}}[\rho_{k}]-\mathcal{F}_{\mathbf{j}}[\widetilde{\rho}_{k}]\bigr)e^{i\omega_{\mathbf{j}}\cdot\widetilde{X}_{t_{n}}}\right\|\right]\!\!\leq\!\sum_{k=1}^{n}\frac{C_{11}\sqrt{\delta t}}{\sqrt{n-k+1}}a_{k}\!+\!\frac{C_{10}}{\sqrt{P}}. (68)

where C10=2T(C2+2M0ϵ2ln(4/δ0))(L3+L/(4π2)+L/720)1/2C_{10}=2\sqrt{T}\left(C_{2}+\frac{2M_{0}}{\epsilon}\sqrt{2\ln(4/\delta_{0})}\right)(L^{-3}+L/(4\pi^{2})+L/720)^{1/2} and C11=C3(L3+L/(4π2)+L/720)1/2C_{11}=C_{3}(L^{-3}+L/(4\pi^{2})+L/720)^{1/2}.

For the temporal truncation term, Assumption 2(d), Eq. (29), and the same resolvent estimate yield

k=1n𝔼[1L3𝐣iω𝐣δt𝐣[Rk](1+Z𝐣)nk+1eiω𝐣X~tn]C9δt,\sum_{k=1}^{n}\mathbb{E}\!\left[\left\|\frac{1}{L^{3}}\sum_{\mathbf{j}\in\mathcal{H}}\frac{i\omega_{\mathbf{j}}\delta t\,\mathcal{F}_{\mathbf{j}}[R_{k}]}{(1+Z_{\mathbf{j}})^{n-k+1}}e^{i\omega_{\mathbf{j}}\cdot\widetilde{X}_{t_{n}}}\right\|\right]\leq C_{9}\delta t, (69)

where C9=K2ϵT/2C_{9}=K_{2}\sqrt{\epsilon T/2}. Combining Eqs. (63), (68), and (69) proves Eq. (61).

Finally, we are ready to prove Theorem 4.

Proof of Theorem 4. From Lemma 9 and Lemma 10, we obtain the system of inequalities that couples ana_{n} and bnb_{n} defined in Eqs. (52)-(53) as follows:

an+1\displaystyle a_{n+1}\leq (1+C5δt)an+χδtbn+C6δt32+C7δtH2+C8δtR,\displaystyle(1+C_{5}\delta t)a_{n}+\chi\delta tb_{n}+C_{6}\delta t^{\frac{3}{2}}+\frac{C_{7}\delta t}{H^{2}}+\frac{C_{8}\delta t}{\sqrt{R}}, (70)
bn\displaystyle b_{n}\leq k=1nC11δtnk+1ak+C101P+C9δt.\displaystyle\sum_{k=1}^{n}\frac{C_{11}\sqrt{\delta t}}{\sqrt{n-k+1}}a_{k}+C_{10}\sqrt{\frac{1}{P}}+C_{9}\delta t. (71)

From this coupled system, we can derive a general bound for ana_{n}. To be specific, substituting Eq. (71) into Eq. (70), we iteratively propagate and simplify the inequality to derive:

an+1(1+C5δt)an+k=1nχC11δt32nk+1ak+δt(Λ+C8R),\displaystyle a_{n+1}\leq(1+C_{5}\delta t)a_{n}+\sum_{k=1}^{n}\frac{\chi C_{11}\delta t^{\frac{3}{2}}}{\sqrt{n-k+1}}a_{k}+\delta t\left(\Lambda+\frac{C_{8}}{\sqrt{R}}\right), (72)

where we define Λ:=χ(C4H2+C101P+C9δt)+C6δt\Lambda:=\chi\left(\frac{C_{4}}{H^{2}}+C_{10}\sqrt{\frac{1}{P}}+C_{9}\delta t\right)+C_{6}\sqrt{\delta t}.

To bound the deterministic accumulation of the error, we introduce a majorizing sequence DnD_{n} that isolates the deterministic components, defined by D0=0D_{0}=0 and

Dj+1=(1+C5δt)Dj+χC11k=1jδt2tj+1tkDk+δtΛ.\displaystyle D_{j+1}=\left(1+C_{5}\delta t\right)D_{j}+\chi C_{11}\sum_{k=1}^{j}\frac{\delta t^{2}}{\sqrt{t_{j+1}-t_{k}}}D_{k}+\delta t\Lambda. (73)

Let Un:=max0jnDjU_{n}:=\max_{0\leq j\leq n}D_{j} be the running maximum of this deterministic sequence. The recursion implies:

Uj+1\displaystyle U_{j+1}\leq (1+C5δt+χC11k=1jδt2tj+1tk)Uj+δtΛ\displaystyle\left(1+C_{5}\delta t+\chi C_{11}\sum_{k=1}^{j}\frac{\delta t^{2}}{\sqrt{t_{j+1}-t_{k}}}\right)U_{j}+\delta t\Lambda
\displaystyle\leq (1+C5δt+χC11δt0tj+11tj+1s𝑑s)Uj+δtΛ\displaystyle\left(1+C_{5}\delta t+\chi C_{11}\delta t\int_{0}^{t_{j+1}}\frac{1}{\sqrt{t_{j+1}-s}}\,ds\right)U_{j}+\delta t\Lambda
\displaystyle\leq (1+(C5+2χC11T)δt)Uj+δtΛ\displaystyle\left(1+(C_{5}+2\chi C_{11}\sqrt{T})\delta t\right)U_{j}+\delta t\Lambda (74)

By the discrete Gronwall inequality, we obtain the explicit bound for the deterministic accumulation:

Unexp(C12T)C12(χC4H2+χC101P+χC9δt+C6δt),\displaystyle U_{n}\leq\frac{\exp(C_{12}T)}{C_{12}}\left(\frac{\chi C_{4}}{H^{2}}+\chi C_{10}\sqrt{\frac{1}{P}}+\chi C_{9}\delta t+C_{6}\sqrt{\delta t}\right), (75)

where C12:=χK+2χM0TϵC_{12}:=\chi K+\frac{2\chi M_{0}\sqrt{T}}{\sqrt{\epsilon}} is a constant.

As established in Lemma 8, the RBM gradient error satisfies 𝔼[ζn,p]=0\mathbb{E}[\zeta_{n,p}]=0. This zero-mean property implies that the single-step position noise χδtζk,p\chi\delta t\zeta_{k,p} forms a martingale difference sequence with respect to the natural filtration of the particle system. Consequently, the noise terms are mutually uncorrelated, and thus orthogonal in the L2L^{2} space. Therefore, the global accumulation of the RBM noise is bounded by the square root of the sum of its variances:

k=1nχδtζk,p(k=1n(C8δtR)2)1/2C8TδtR,\displaystyle\left\|\sum_{k=1}^{n}\chi\delta t\zeta_{k,p}\right\|\leq\left(\sum_{k=1}^{n}\left(\frac{C_{8}\delta t}{\sqrt{R}}\right)^{2}\right)^{1/2}\leq C_{8}\sqrt{T}\frac{\sqrt{\delta t}}{\sqrt{R}}, (76)

where nT/δtn\leq\lfloor T/\delta t\rfloor is used. Combining this stochastic bound with the deterministic bound UnU_{n} yields the refined optimal error bound that holds with high probability:

anN0H2+N1δt+N2P+N3δtR,\displaystyle a_{n}\leq\frac{N_{0}}{H^{2}}+N_{1}\sqrt{\delta t}+\frac{N_{2}}{\sqrt{P}}+N_{3}\frac{\sqrt{\delta t}}{\sqrt{R}}, (77)

for all n0n\geq 0, where higher-order terms are omitted, and N0=χC4exp(C12T)C12,N1=C6exp(C12T)C12,N2=χC10exp(C12T)C12N_{0}=\frac{\chi C_{4}\exp(C_{12}T)}{C_{12}},N_{1}=\frac{C_{6}\exp(C_{12}T)}{C_{12}},N_{2}=\frac{\chi C_{10}\exp(C_{12}T)}{C_{12}}, and N3=C8TN_{3}=C_{8}\sqrt{T}. Moreover, we point out that C4C_{4} is a constant defined in Lemma 7, C10,C11C_{10},C_{11} are constants defined in Lemma 10, C6,C7,C8C_{6},C_{7},C_{8} are constants defined in Lemma 9.

According to the discrete and continuous dynamics defined in Eqs. (16)-(17), the 11-Wasserstein distance between the approximate and exact distributions at time tn+1t_{n+1} is given by:

𝒲1(ρ~tn+1,ρtn+1)\displaystyle\mathcal{W}_{1}(\widetilde{\rho}_{t_{n+1}},\rho_{t_{n+1}}) =infγΠ(ρ~tn+1,ρtn+1)(3×3𝐱𝐲L1𝑑γ(𝐱,𝐲)),\displaystyle=\inf_{\gamma\in\Pi(\widetilde{\rho}_{t_{n+1}},\rho_{t_{n+1}})}\left(\int_{\mathbb{R}^{3}\times\mathbb{R}^{3}}\|\mathbf{x}-\mathbf{y}\|_{L^{1}}\,d\gamma(\mathbf{x},\mathbf{y})\right), (78)

where the infimum is taken over all possible couplings of the two distributions.

Under the natural coupling induced by shared initial conditions and Brownian motion paths (i.e., X~tn\widetilde{X}_{t_{n}} and XtnX_{t_{n}} evolve via the same Wiener process WsW_{s}), we explicitly construct a joint distribution γn=Law(X~tn,Xtn)\gamma_{n}=\text{Law}(\widetilde{X}_{t_{n}},X_{t_{n}}). This coupling allows us to bound the Wasserstein distance as:

𝒲1(ρ~tn,ρtn)\displaystyle\mathcal{W}_{1}(\widetilde{\rho}_{t_{n}},\rho_{t_{n}})\!\leq S0H2+S1δt+S2P+S3δtR,\displaystyle\frac{S_{0}}{H^{2}}+S_{1}\sqrt{\delta t}+\frac{S_{2}}{\sqrt{P}}+\frac{S_{3}\sqrt{\delta t}}{\sqrt{R}}, (79)

where the constants S0,S1,S2,S3S_{0},S_{1},S_{2},S_{3} are given explicitly by S0=3C7exp(C12T)C12S_{0}=\sqrt{3}\frac{C_{7}\exp(C_{12}T)}{C_{12}}, S1=3C6exp(C12T)C12S_{1}=\sqrt{3}\frac{C_{6}\exp(C_{12}T)}{C_{12}}, S2=3χC10exp(C12T)C12S_{2}=\sqrt{3}\frac{\chi C_{10}\exp(C_{12}T)}{C_{12}}, and S3=3TC8S_{3}=\sqrt{3T}C_{8}, with C6C_{6}, C7,C8,C10C_{7},C_{8},C_{10} defined in Lemmas 8-9. Specifically, tracing back to these lemmas shows that C7L2C_{7}\propto L^{2}, while C6,C8,C9,C10,C12C_{6},C_{8},C_{9},C_{10},C_{12} are independent of LL.

The inequality follows from the fact that the Wasserstein distance is defined as the infimum over all possible couplings, and our construction provides one such coupling. This step follows from the elementary norm inequality 𝐱L13𝐱L2\|\mathbf{x}\|_{L^{1}}\leq\sqrt{3}\|\mathbf{x}\|_{L^{2}} for vectors in 3\mathbb{R}^{3}, which is a direct consequence of the Cauchy-Schwarz inequality.

To derive the bound for the Fourier coefficients in Eq. (25), we explicitly incorporate the Fourier mode HH into the error analysis. Unlike the strong trajectory error ana_{n}, the Fourier coefficients represent macroscopic weak functionals (expectations of the test functions eiωjxe^{-i\omega_{j}\cdot x}). By the convergence theory of SDEs, the stochastic strong fluctuations of order 𝒪(δt)\mathcal{O}(\sqrt{\delta t}) cancel out in expectation. Instead, the Brownian increments and the zero-mean RBM noise ζk,p\zeta_{k,p} contribute to the macroscopic bias only through second-order Taylor expansion terms, yielding a single-step bias of 𝒪(δt2)\mathcal{O}(\delta t^{2}) from temporal discretization and 𝒪(δt2/R)\mathcal{O}(\delta t^{2}/R) from the RBM approximation.

Accordingly, we refine the accumulation of the source term error over nT/δtn\leq\lfloor T/\delta t\rfloor steps (i.e., up to the final time TT) by replacing the stochastic part of aka_{k} with its weak bias, while retaining the deterministic spatial part N0/H2N_{0}/H^{2}. Applying the standard accumulation estimates along with the frequency bound ω𝐣3πH/L\|\omega_{\mathbf{j}}\|\leq\sqrt{3}\pi H/L, we arrive at the error bound:

max𝐣|α~tn;𝐣αtn;𝐣|S4H+S5Hδt+S6HP+S7HδtR,\displaystyle\max_{\mathbf{j}\in\mathcal{H}}|\widetilde{\alpha}_{t_{n};\mathbf{j}}-\alpha_{t_{n};\mathbf{j}}|\leq\frac{S_{4}}{H}+S_{5}H\delta t+\frac{S_{6}H}{\sqrt{P}}+S_{7}\frac{H\delta t}{\sqrt{R}}, (80)

where we define S4=3πTC3N0LS_{4}=\sqrt{3}\pi TC_{3}N_{0}L, S5=3πTC3ϵLS_{5}=\frac{\sqrt{3}\pi TC_{3}}{\epsilon L}, S6=3πTC2ϵLS_{6}=\frac{\sqrt{3}\pi TC_{2}}{\epsilon L}, S7=3πTC3M0M4ϵLS_{7}=\frac{\sqrt{3}\pi TC_{3}M_{0}M_{4}}{\epsilon L}. Here, C1,C2C_{1},C_{2}, and C3C_{3} are constants defined in Lemma 6. We also omit higher-order terms in the leading-order bound. This completes the proof of Theorem 4.

3.4 Numerical Unconditional Stability of the SIPF-rr Method

The SIPF-rr method exhibits inherent numerical stability regarding the chemical concentration field. This property stems directly from the spectral discretization and the implicit time-stepping scheme, a property that is independent of the particle distribution. We establish that both the chemical concentration c~\widetilde{c} and its gradient c~\nabla\widetilde{c} remain uniformly bounded for any fixed Fourier mode HH, thereby justifying the boundedness assumptions employed in the convergence analysis.

Lemma 11 (Unconditional stability of c~\widetilde{c} and c~\nabla\widetilde{c}).

Let HH be the finite number of Fourier modes and Ω\Omega be the spatial domain. For any time step n0n\geq 0 and the fixed Fourier mode HH, the reconstructed concentration field c~n\widetilde{c}_{n} and its gradient satisfy the uniform bounds:

c~nL(Ω)CstabH+C0,\displaystyle\|\widetilde{c}_{n}\|_{L^{\infty}(\Omega)}\leq C_{\mathrm{stab}}H+C_{0}, (81)

and

c~nL(Ω):=sup𝐱Ωc~n(𝐱)L2Creg(H),\displaystyle\|\nabla\widetilde{c}_{n}\|_{L^{\infty}(\Omega)}:=\operatorname{\,sup}_{\mathbf{x}\in\Omega}\|\nabla\widetilde{c}_{n}(\mathbf{x})\|_{L^{2}}\leq C_{\mathrm{reg}}(H), (82)

where Cstab,C0>0C_{\mathrm{stab}},C_{0}>0 are constants depending only on the initial data, M0M_{0}, λ\lambda, and the domain size LL, and Creg(H)C_{\mathrm{reg}}(H) is a constant that depends on HH, LL, M0M_{0}, and λ\lambda, but is independent of the time step size δt\delta t and the particle positions {X~np}p=1P\{\widetilde{X}_{n}^{p}\}_{p=1}^{P}.

Proof.

The update formula for the Fourier coefficients α~n;𝐣\widetilde{\alpha}_{n;\mathbf{j}} in the SIPF-rr algorithm is derived from the implicit Euler discretization. By rearranging the terms in Eq. (8) and applying the Fourier transform, the update relates α~n;𝐣\widetilde{\alpha}_{n;\mathbf{j}} to the previous state α~n1;𝐣\widetilde{\alpha}_{n-1;\mathbf{j}} and the current empirical density ρ~^n;𝐣\widehat{\widetilde{\rho}}_{n;\mathbf{j}} via the amplification factor

α~n;𝐣=11+Z𝐣α~n1;𝐣+δt/ϵ1+Z𝐣ρ~^n;𝐣,with Z𝐣=δtϵ(ω𝐣2+λ2).\widetilde{\alpha}_{n;\mathbf{j}}=\frac{1}{1+Z_{\mathbf{j}}}\widetilde{\alpha}_{n-1;\mathbf{j}}+\frac{\delta t/\epsilon}{1+Z_{\mathbf{j}}}\widehat{\widetilde{\rho}}_{n;\mathbf{j}},\quad\text{with }Z_{\mathbf{j}}=\frac{\delta t}{\epsilon}(\|\omega_{\mathbf{j}}\|^{2}+\lambda^{2}). (83)

The particle source term satisfies

|ρ~^n;𝐣|=|M0Pp=1Peiω𝐣X~np|M0.|\widehat{\widetilde{\rho}}_{n;\mathbf{j}}|=|\frac{M_{0}}{P}\sum_{p=1}^{P}e^{-i\omega_{\mathbf{j}}\cdot\widetilde{X}_{n}^{p}}|\leq M_{0}. (84)

Applying the recursive relation and the triangle inequality yields a bound via geometric series for the magnitude of the coefficients:

|α~n;𝐣|\displaystyle|\widetilde{\alpha}_{n;\mathbf{j}}| 11+Z𝐣|α~n1;𝐣|+M0δt/ϵ1+Z𝐣\displaystyle\leq\frac{1}{1+Z_{\mathbf{j}}}|\widetilde{\alpha}_{n-1;\mathbf{j}}|+\frac{M_{0}\delta t/\epsilon}{1+Z_{\mathbf{j}}}
(11+Z𝐣)n|α~0;𝐣|+M0δtϵ(1+Z𝐣)k=0n1(11+Z𝐣)k\displaystyle\leq\left(\frac{1}{1+Z_{\mathbf{j}}}\right)^{n}|\widetilde{\alpha}_{0;\mathbf{j}}|+\frac{M_{0}\delta t}{\epsilon(1+Z_{\mathbf{j}})}\sum_{k=0}^{n-1}\left(\frac{1}{1+Z_{\mathbf{j}}}\right)^{k}
|α~0;𝐣|+M0ω𝐣2+λ2.\displaystyle\leq|\widetilde{\alpha}_{0;\mathbf{j}}|+\frac{M_{0}}{\|\omega_{\mathbf{j}}\|^{2}+\lambda^{2}}. (85)

We next bound c~nL\|\widetilde{c}_{n}\|_{L^{\infty}} by approximating the partial Fourier sum. Using ω𝐣=2πL𝐣\|\omega_{\mathbf{j}}\|=\frac{2\pi}{L}\|\mathbf{j}\| and treating the sum over {𝟎}\mathcal{H}\setminus\{\mathbf{0}\} as a Riemann sum, which is approximated by an integral in spherical coordinates:

c~nL\displaystyle\|\widetilde{c}_{n}\|_{L^{\infty}} 1L3𝐣|α~n;𝐣|\displaystyle\leq\frac{1}{L^{3}}\sum_{\mathbf{j}\in\mathcal{H}}|\widetilde{\alpha}_{n;\mathbf{j}}|
1L3(𝐣|α~0;𝐣|+M0λ2+𝐣{𝟎}M0L24π2𝐣2)\displaystyle\leq\frac{1}{L^{3}}\left(\sum_{\mathbf{j}\in\mathcal{H}}|\widetilde{\alpha}_{0;\mathbf{j}}|+\frac{M_{0}}{\lambda^{2}}+\sum_{\mathbf{j}\in\mathcal{H}\setminus\{\mathbf{0}\}}\frac{M_{0}L^{2}}{4\pi^{2}\|\mathbf{j}\|^{2}}\right)
1L3(𝐣|α~0;𝐣|+M0λ2+M0L24π2132H1r2 4πr2𝑑r)\displaystyle\leq\frac{1}{L^{3}}\left(\sum_{\mathbf{j}\in\mathcal{H}}|\widetilde{\alpha}_{0;\mathbf{j}}|+\frac{M_{0}}{\lambda^{2}}+\frac{M_{0}L^{2}}{4\pi^{2}}\int_{1}^{\frac{\sqrt{3}}{2}H}\frac{1}{r^{2}}\,4\pi r^{2}\,dr\right)
CstabH+C0,\displaystyle\leq C_{\mathrm{stab}}H+C_{0}, (86)

where Cstab=3M02πLC_{\mathrm{stab}}=\frac{\sqrt{3}M_{0}}{2\pi L} and C0=1L3(𝐣|α~0;𝐣|+M0λ2)C_{0}=\frac{1}{L^{3}}\left(\sum_{\mathbf{j}\in\mathcal{H}}|\widetilde{\alpha}_{0;\mathbf{j}}|+\frac{M_{0}}{\lambda^{2}}\right) are constants.

Regarding the gradient c~\nabla\widetilde{c}, the SIPF-rr method employs a specific discretization to handle the singularity of the Green’s function 𝒦ϵ,δt\mathcal{K}_{\epsilon,\delta t}. According to Alg. 2, the gradient at a particle position X~np\widetilde{X}^{p}_{n} is computed as:

c~(X~np,tn)\displaystyle\nabla\widetilde{c}(\widetilde{X}^{p}_{n},t_{n}) =ϵδtL3H3𝐤𝐱𝒦ϵ,δt(X~np+X¯np𝐱𝐤)c~n1(𝐱𝐤X¯np)\displaystyle=-\frac{\epsilon}{\delta t}\frac{L^{3}}{H^{3}}\sum_{\mathbf{k}\in\mathcal{H}}\nabla_{\mathbf{x}}\mathcal{K}_{\epsilon,\delta t}(\widetilde{X}^{p}_{n}+\bar{X}^{p}_{n}-\mathbf{x}_{\mathbf{k}})\widetilde{c}_{n-1}(\mathbf{x}_{\mathbf{k}}-\bar{X}^{p}_{n})
M0RsCp,sp𝐱𝒦ϵ,δt(X~npX~ns),\displaystyle\quad-\frac{M_{0}}{R}\sum_{s\in C_{p},s\neq p}\nabla_{\mathbf{x}}\mathcal{K}_{\epsilon,\delta t}(\widetilde{X}^{p}_{n}-\widetilde{X}^{s}_{n}), (87)

where the spatial shift X¯np=L2H+X~npL/HLHX~np\bar{X}^{p}_{n}=\frac{L}{2H}+\lfloor\frac{\widetilde{X}^{p}_{n}}{L/H}\rfloor\frac{L}{H}-\widetilde{X}_{n}^{p} ensures that the evaluation point is bounded away from the singularity of Green’s function at grid points 𝐱𝐤\mathbf{x}_{\mathbf{k}}. Specifically, let r:=X~np+X¯np𝐱𝐤L2Hr:=\big\|\widetilde{X}^{p}_{n}+\bar{X}^{p}_{n}-\mathbf{x}_{\mathbf{k}}\big\|\geq\frac{L}{2H} and β=λ2+ϵ/δt\beta=\sqrt{\lambda^{2}+\epsilon/\delta t}. The gradient of the Green’s function satisfies

ϵδt𝒦ϵ,δt(X~np+X¯np𝐱𝐤)β2eβr4π(1r2+βr).\left\|\frac{\epsilon}{\delta t}\nabla\mathcal{K}_{\epsilon,\delta t}(\widetilde{X}^{p}_{n}+\bar{X}^{p}_{n}-\mathbf{x}_{\mathbf{k}})\right\|\leq\beta^{2}\frac{e^{-\beta r}}{4\pi}\left(\frac{1}{r^{2}}+\frac{\beta}{r}\right). (88)

Letting y=βry=\beta r, and using the inequality supy0ykey=(k/e)k\sup_{y\geq 0}y^{k}e^{-y}=(k/e)^{k}, we can bound the term to be independent of β\beta (and thus independent of δt\delta t):

supβ>0(β2r2eβr+β3reβr)\displaystyle\sup_{\beta>0}\left(\frac{\beta^{2}}{r^{2}}e^{-\beta r}+\frac{\beta^{3}}{r}e^{-\beta r}\right) =1r4supy(y2ey+y3ey)\displaystyle=\frac{1}{r^{4}}\sup_{y}(y^{2}e^{-y}+y^{3}e^{-y})
1(L/2H)4(4e2+27e3).\displaystyle\leq\frac{1}{(L/2H)^{4}}\left(\frac{4}{e^{2}}+\frac{27}{e^{3}}\right). (89)

Since c~n1L\|\widetilde{c}_{n-1}\|_{L^{\infty}} is bounded (as shown above) and the sum over 𝐤\mathbf{k}\in\mathcal{H} is finite for a fixed HH, the first term in Eq. (3.4) is uniformly bounded by a constant depending on HH but independent of δt\delta t.

For the second term in Eq. (3.4), the RBM excludes self-interaction (sps\neq p). For any fixed grid resolution HH, the corresponding effective kernel corresponds to a spectrally truncated approximation that is smooth. Thus, for any finite number of particles, this sum is finite.

Combining these results, c~nL(Ω)Creg(H)\|\nabla\widetilde{c}_{n}\|_{L^{\infty}(\Omega)}\leq C_{\mathrm{reg}}(H) is guaranteed by the algorithm’s design, where the LL^{\infty}-norm for the vector-valued gradient is defined by c~nL(Ω):=sup𝐱Ωc~n(𝐱)L2\|\nabla\widetilde{c}_{n}\|_{L^{\infty}(\Omega)}:=\operatorname{\,sup}_{\mathbf{x}\in\Omega}\|\nabla\widetilde{c}_{n}(\mathbf{x})\|_{L^{2}}, and Creg(H)C_{\mathrm{reg}}(H) is a constant depending on HH, LL, M0M_{0}, and λ\lambda.

This lemma confirms that the boundedness condition in Assumption 2 is not merely an external hypothesis but a property guaranteed by the SIPF-rr method itself. This demonstrates that the SIPF-rr method effectively regularizes the singular Keller-Segel kernel. While the exact solution may exhibit finite-time blow-up (where c\|\nabla c\|\to\infty), the numerical field remains finite for any fixed Fourier mode HH. This unconditional numerical stability ensures that the algorithm is robust: it enables the simulation of blow-up phenomena by capturing the solution’s growth trend as HH increases, while avoiding numerical breakdown at fixed resolutions.

4 Numerical Experiments

The numerical experiments are organized into three main parts to evaluate the SIPF-rr method. In Subsection 4.1, we validate the SIPF-rr method by quantifying its accuracy against a high-resolution radial finite difference benchmark and verifying its convergence rates with respect to the time step and batch size. Subsection 4.2 investigates the method’s capability to detect finite-time blow-up phenomena and critical mass thresholds under various conditions. Finally, Subsection 4.3 provides empirical verification of the key theoretical assumptions used in our analysis, specifically the spatial Lipschitz continuity of the concentration gradient.

4.1 Validation of the SIPF-rr Method

4.1.1 Comparison with FDM

We first demonstrate the accuracy of the SIPF-rr method. In the radially symmetric case, the fully parabolic KS system (1) in 3D can be expressed as ρ(x,y,z,t)=ρ(r,t)\rho(x,y,z,t)=\rho(r,t) and c(x,y,z,t)=c(r,t)c(x,y,z,t)=c(r,t), where r=x2+y2+z2r=\sqrt{x^{2}+y^{2}+z^{2}}. The system is then rewritten as follows in 1D:

{ρt=μ(2ρr2+2rρr)χ(ρrcr+ρ(2cr2+2rcr)),ϵct=(2cr2+2rcr)λ2c+ρ.\left\{\begin{aligned} \rho_{t}&=\mu\,\left(\frac{\partial^{2}\rho}{\partial r^{2}}+\frac{2}{r}\frac{\partial\rho}{\partial r}\right)-\chi\,\left(\frac{\partial\rho}{\partial r}\frac{\partial c}{\partial r}+\rho\cdot(\frac{\partial^{2}c}{\partial r^{2}}+\frac{2}{r}\frac{\partial c}{\partial r})\right),\\ \epsilon c_{t}&=\,\left(\frac{\partial^{2}c}{\partial r^{2}}+\frac{2}{r}\frac{\partial c}{\partial r}\right)-\lambda^{2}c+\rho.\end{aligned}\right. (90)

The radial representation in 1D allows us to compute a reference solution using a very fine mesh by the finite difference method (FDM). It will serve as a benchmark to quantify the accuracy of the SIPF-rr method in 3D. We define the relative error between the cumulative distribution functions (CDFs) obtained from the FDM and the SIPF-rr method as

Relative Error=1Ni=1N{0,if FFDM(si)=0,|FSIPF-r(si)FFDM(si)|FFDM(si),otherwise,\text{Relative Error}=\frac{1}{N}\sum_{i=1}^{N}\begin{cases}0,&\text{if }F_{\text{FDM}}(s_{i})=0,\\ \frac{|F_{\text{SIPF-$r$}}(s_{i})-F_{\text{FDM}}(s_{i})|}{F_{\text{FDM}}(s_{i})},&\text{otherwise,}\end{cases} (91)

where FSIPF-r(si)F_{\text{SIPF-$r$}}(s_{i}) and FFDM(si)F_{\text{FDM}}(s_{i}) represent the CDFs of ρ\rho computed via the SIPF-rr and FDM methods, respectively, and sis_{i} denotes the ii-th radial mesh point in the FDM, which are the discrete points along the radial direction starting from the origin. To ensure the relative error is well-defined, we set it to zero wherever FFDM(si)=0F_{\text{FDM}}(s_{i})=0.

Here, the initial distribution ρ0\rho_{0} is assumed to be a uniform distribution over a ball centered at (0,0,0)T(0,0,0)^{T} with radius 1. The model parameters are chosen as follows:

μ=χ=1,ϵ=104,λ=101.\displaystyle\mu=\chi=1,\quad\epsilon=10^{-4},\quad\lambda=10^{-1}. (92)

For the numerical computation, we use H=24H=24 Fourier basis functions in each spatial dimension to discretize the chemical concentration cc and use P=10000P=10000 particles to represent the approximated distribution ρ\rho, where the batch size in Algorithm 2 is R=P=100R=\lfloor\sqrt{P}\rfloor=100. The computational domain is Ω=[L/2,L/2]3\Omega=[-L/2,L/2]^{3}, where L=8L=8, and the total mass is chosen to be M0=20M_{0}=20. The evolution of cc and ρ\rho is computed using Algorithm 3 with a time step size δt=104\delta t=10^{-4}, up to the final simulation time T=0.1T=0.1.

In Fig. 1, we present the evolution of particles over time, showing the dynamic behavior of ρ\rho. Additionally, in Fig. 2, we compare the cumulative probability curves of ρ\rho obtained from the radial FDM and the SIPF-rr method at T=0.1T=0.1, with a mean relative error of 0.05512 as defined in Eq. (91). This comparison demonstrates that the SIPF-rr algorithm achieves high accuracy in approximating the true solution. These results validate the effectiveness of the SIPF-rr algorithm in capturing the behavior of the particle distribution.

Refer to caption
(a) t=0
Refer to caption
(b) t=0.025
Refer to caption
(c) t=0.05
Refer to caption
(d) t=0.1
Figure 1: Scatter plot of particles with M0M_{0} = 20.
Refer to caption
Figure 2: Cumulative distribution of ρ\rho computed by the SIPF-rr method and radial FDM.

4.1.2 Convergence of the SIPF-rr Method

In this subsection, we validate the convergence of the SIPF-rr method numerically. Based on Eq. (80), the error between c~\widetilde{c} and cc can be quantified by the L2L^{2} error between their Fourier coefficients α~\widetilde{\alpha} and α\alpha. We adopt the same initial conditions as in Subsection 4.1.1. To eliminate the uncertainty introduced by the RBM, the reference solution is computed using the original SIPF method [51] with parameters δt=106\delta t=10^{-6}, H=24H=24, and P=10000P=10000. Additionally, we set M0=20,T=0.01M_{0}=20,T=0.01 to ensure that the system remains free of singularities, as verified in Fig. 3 of [51].

To investigate the convergence with respect to the time step δt\delta t, we vary δt\delta t from 28T2^{-8}T to 24T2^{-4}T. Since Theorem 4 holds with high probability, we perform 100 independent experiments for each δt\delta t to empirically validate the algorithm’s accuracy. The mean L2L^{2} error of the Fourier coefficients is computed over these 100 trials. As shown in Fig. 3(a), the slope of the mean L2L^{2} error versus δt\delta t on a logarithmic scale indicates an approximate first-order convergence rate, with e(δt)=𝒪(δt1.023)e(\delta t)=\mathcal{O}(\delta t^{1.023}). This result aligns with the theoretical bound given in Eq. (25) of Theorem 4.

Furthermore, we examine the mean L2L^{2} error of c~(,T)\widetilde{c}(\cdot,T) for varying batch sizes R=100,200,400,800,1600R=100,200,400,800,1600, while keeping P=10000P=10000. From Eq. (25), with other parameters fixed, the theoretical L2L^{2} error of c~\widetilde{c} with respect to the batch size RR should scale as 𝒪(R12)\mathcal{O}(R^{-\frac{1}{2}}). This is empirically verified in Fig. 3(b), where the fitted convergence rate is e(R)=𝒪(R0.495)e(R)=\mathcal{O}(R^{-0.495}), closely matching the theoretical prediction.

Refer to caption
(a) vs. time step δt\delta t (log-scale)
Refer to caption
(b) vs. batch size RR (log-scale)
Figure 3: L2L^{2} error of c~\widetilde{c} in the SIPF-rr method.

4.2 Numerical Detection of Finite-time Blow-up

While the error analysis in Section 3 assumes regular solutions, our algorithm can effectively detect finite-time blow-up phenomena under practical conditions. The parameters are the same as in Eq. (92), with the radial system (90) serving as a reference benchmark. We consider a more concentrated initial condition than in the preceding subsections: a uniform distribution within a ball of radius r=1.5r=1.5.

Fig. 4 illustrates the effect of varying the initial mass M0M_{0} on solution behavior at T=2T=2. We note that unlike the 2D case, the 3D KS system has no universal critical mass, and the blow-up formation depends on both the total mass and the initial spatial distribution. For our prescribed initial configuration, defined as a uniform density supported over a spherical ball of fixed radius, we numerically observed a sharp blow-up transition. Figs. 4(a) and 4(b) show particle distributions for M0=72M_{0}=72 and M0=73M_{0}=73, respectively, with the latter exhibiting clear concentration and potential blow-up. In Fig. 4(c), we present the maximum value of cc over space at T=2T=2 for different initial masses, computed via the radial FDM as a benchmark. We denote

c,FDM=supr|c(r,T)|.\|c\|_{\infty,\mathrm{FDM}}=\sup_{r}|c(r,T)|. (93)

The FDM exhibits numerical instabilities (marked in red) for initial masses between 7272 and 7373, which identifies the threshold for blow-up specific to this concentrated initial configuration. This agrees with the blow-up transition observed in our proposed method in Fig. 4(b). Remarkably, the concentration and potential blow-up are captured with only P=104P=10^{4} particles and a moderate value of HH, demonstrating the algorithm’s ability to detect singularities without the strict constraints imposed in the convergence analysis.

To further investigate blow-up detection with our method, we examine the maximum chemical concentration cc as a function of time TT for different discretization levels HH and masses M0M_{0}. As shown in Fig. 5, when M0=40M_{0}=40 (Fig. 5(a)), the maximum cc remains stable and shows minimal variation across HH values, indicating a non-blow-up regime. In contrast, for M0=100M_{0}=100 (Fig. 5(b)), the maximum cc exhibits strong dependence on HH, with curves diverging significantly. Notably, this blow-up signature is visible even with coarse discretization (H=8H=8 vs H=12H=12), demonstrating that singularity detection does not require high-resolution computations. The divergence of solutions at different HH levels provides a practical diagnostic for intense concentration and blow-up phenomena without demanding highly refined discretizations.

We then demonstrate that our method can also capture concentration and potential blow-up for non-radial initial data. We consider two non-overlapping spheres of radius r=0.5r=0.5 with centers at (0.6,0,0)(0.6,0,0) and (0.6,0,0)(-0.6,0,0), each containing equal mass M0/2M_{0}/2 (total mass M0=100M_{0}=100). Fig. 6 shows the time evolution for two different mass regimes: the top row (M0=10M_{0}=10) remains stable at all times, while the bottom row (M0=100M_{0}=100) exhibits clear concentration towards blow-up formation at T=0.2T=0.2. This demonstrates that our algorithm effectively detects blow-up even for non-symmetric initial configurations, highlighting its robustness beyond the radially symmetric case.

Refer to caption
(a) Scatter plot of particles at T=2T=2 with M0=72M_{0}=72.
Refer to caption
(b) Scatter plot of particles at T=2T=2 with M0=73M_{0}=73.
Refer to caption
(c) c,FDM\|c\|_{\infty,\text{FDM}} at T=2T=2 vs M0M_{0} (the red star shape marker denotes the numerical instabilities).
Figure 4: Effects of initial mass M0M_{0} on concentration behavior (finite time blow-up).
Refer to caption
(a) M0=40M_{0}=40
Refer to caption
(b) M0=100M_{0}=100
Figure 5: Maximum chemical concentration cc vs computation time TT for different discretization parameters HH and total masses M0M_{0}
Refer to caption
(a) M0=10M_{0}=10, T=0T=0
Refer to caption
(b) M0=10M_{0}=10, T=0.1T=0.1
Refer to caption
(c) M0=10M_{0}=10, T=0.2T=0.2
Refer to caption
(d) M0=100M_{0}=100, T=0T=0
Refer to caption
(e) M0=100M_{0}=100, T=0.1T=0.1
Refer to caption
(f) M0=100M_{0}=100, T=0.2T=0.2
Figure 6: Finite-time blow-up detection for non-radial initial data (two spheres). Top row: M0=10M_{0}=10, stable evolution at times T=0,0.1,0.2T=0,0.1,0.2. Bottom row: M0=100M_{0}=100, evolution showing blow-up formation at T=0.2T=0.2. Both simulations use P=104P=10^{4} particles, H=24H=24, and initial spheres of radius r=0.5r=0.5 centered at (±0.6,0,0)(\pm 0.6,0,0).

4.3 Validation of Theoretical Assumptions

To verify the spatial Lipschitz continuity in Assumption 2(a), we adjust the spatial discretization by varying HH from 6 to 24. At the final time T=0.1T=0.1, we randomly select 1000 pairs of particle points from a total of 10,000 particles in each calculation. The spatial Lipschitz constant L(H)L(H) for c~\nabla\widetilde{c} is defined as the maximum ratio of the gradient difference to the spatial distance over all pairs of particle points {𝐱,𝐲}\{\mathbf{x},\mathbf{y}\}:

L(H):=max{𝐱,𝐲}c~(𝐱,T)c~(𝐲,T)𝐱𝐲.L(H):=\max_{\{\mathbf{x},\mathbf{y}\}}\frac{\|\nabla\widetilde{c}(\mathbf{x},T)-\nabla\widetilde{c}(\mathbf{y},T)\|}{\|\mathbf{x}-\mathbf{y}\|}. (94)

The results, shown in Table 1, list the computed Lipschitz constant L(H)L(H) for each value of HH. The variation in these values is relatively small, confirming that the spatial Lipschitz continuity holds for c~\nabla\widetilde{c} computed by the SIPF-rr method.

Fourier modes (HH) Spatial Lipschitz constant (L(H)L(H))
6 0.002085
12 0.002106
18 0.002036
24 0.001957
Table 1: Spatial Lipschitz constant of c~\nabla\widetilde{c} vs. HH.
5 Discussion

In previous sections, we developed and analyzed the SIPF-rr method for the classical 3D fully parabolic KS system (1). The underlying SIPF-rr framework, however, is flexible and not limited to this classical model. Since the evolution of the chemical field and particle dynamics are updated separately, we can modify the SIPF-rr method to incorporate additional biological mechanisms and simulate more complicated mathematical biology models [21, 22]. In what follows, we illustrate this flexibility through three representative extensions: multi-species chemotaxis systems with volume-exclusion effects, nonlinear diffusion, and anisotropic cell motility, and we leave their detailed implementation, numerical validation, and convergence analysis for future work.

Multi-species chemotaxis system with volume-exclusion effects. The SIPF-rr framework is readily extendable to multi-species chemotaxis with volume-exclusion (often referred to as volume-filling) effects, which are essential for modeling realistic biological scenarios where cells of different species compete for finite physical space [19, 37, 3]. A representative two-species chemotaxis model with cross-volume exclusion is

tρk\displaystyle\partial_{t}\rho_{k} =(Dk(ρtot)ρkχkρkΦ(ρtot)ck),\displaystyle=\nabla\cdot\left(D_{k}(\rho_{\mathrm{tot}})\nabla\rho_{k}-\chi_{k}\rho_{k}\Phi(\rho_{\mathrm{tot}})\nabla c_{k}\right), (95)
ϵtck\displaystyle\epsilon\partial_{t}c_{k} =Δckλk2ck+l=12βklρl,k=1,2,\displaystyle=\Delta c_{k}-\lambda_{k}^{2}c_{k}+\sum_{l=1}^{2}\beta_{kl}\rho_{l},\qquad k=1,2,

where ρtot=ρ1+ρ2\rho_{\mathrm{tot}}=\rho_{1}+\rho_{2} denotes the total cell density, ckc_{k} is the chemical concentration of species kk, ϵ\epsilon, χk\chi_{k}, and λk\lambda_{k} are positive constants, βkl\beta_{kl} are the production rates of chemical kk by species ll, Dk(ρtot)D_{k}(\rho_{\mathrm{tot}}) is the density-dependent diffusivity, and Φ(ρtot)=(1ρtot/ρmax)+\Phi(\rho_{\mathrm{tot}})=(1-\rho_{\mathrm{tot}}/\rho_{\max})_{+} is the crowding factor (derived from the volume-filling probability in [37]) that modulates chemotactic mobility. This formulation naturally extends to mm species by redefining ρtot=l=1mρl\rho_{\mathrm{tot}}=\sum_{l=1}^{m}\rho_{l}. Crucially, the system  (95) reduces to the classical multi-species KS model when DkμkD_{k}\equiv\mu_{k} and Φ1\Phi\equiv 1.

Since the equations governing the chemical field ckc_{k} remain linear, we apply Algorithm 1 to update the Fourier coefficients of the chemical field ckc_{k} as follows:

α~n;𝐣k=ϵδtα~n1;𝐣k+l=1mβklρ~^l,n;𝐣ϵ/δt+ω𝐣2+λk2,k=1,,m,\widetilde{\alpha}_{n;\mathbf{j}}^{k}=\frac{\frac{\epsilon}{\delta t}\widetilde{\alpha}_{n-1;\mathbf{j}}^{k}+\sum_{l=1}^{m}\beta_{kl}\widehat{\widetilde{\rho}}_{l,n;\mathbf{j}}}{\epsilon/\delta t+\|\omega_{\mathbf{j}}\|^{2}+\lambda_{k}^{2}},\qquad k=1,\ldots,m, (96)

where α~n;𝐣k\widetilde{\alpha}_{n;\mathbf{j}}^{k} denotes the Fourier coefficient of the chemical field ckc_{k}, δt\delta t is the time step, ω𝐣=2πL𝐣\omega_{\mathbf{j}}=\frac{2\pi}{L}\mathbf{j} is the frequency vector associated with the multi-index 𝐣=(j1,j2,j3)\mathbf{j}=(j_{1},j_{2},j_{3}) (with LL denoting the domain size), 𝐣\mathbf{j}\in\mathcal{H} (the index set defined in (4)), and ρ~^l,n;𝐣\widehat{\widetilde{\rho}}_{l,n;\mathbf{j}} is the Fourier coefficient of the empirical density of species ll with 1lm1\leq l\leq m.

For the density update, species ll with total mass MlM_{l} is represented at time tnt_{n} using PlP_{l} particles {X~l,np}p=1Pl\{\widetilde{X}_{l,n}^{p}\}_{p=1}^{P_{l}}, with the corresponding empirical measure

ρ~l,n(𝐱)=MlPlp=1Plδ(𝐱X~l,np).\widetilde{\rho}_{l,n}(\mathbf{x})=\frac{M_{l}}{P_{l}}\sum_{p=1}^{P_{l}}\delta(\mathbf{x}-\widetilde{X}_{l,n}^{p}). (97)

The Fourier coefficients of the density source in Eq. (96) can be computed directly from this empirical measure as

ρ~^l,n;𝐣=𝐣[ρ~l,n]=MlPlp=1Pleiω𝐣X~l,np.\widehat{\widetilde{\rho}}_{l,n;\mathbf{j}}=\mathcal{F}_{\mathbf{j}}[\widetilde{\rho}_{l,n}]=\frac{M_{l}}{P_{l}}\sum_{p=1}^{P_{l}}e^{-i\omega_{\mathbf{j}}\cdot\widetilde{X}_{l,n}^{p}}. (98)

The density of species ll is then reconstructed at the Fourier grid points via the inverse Fourier transform, and the total density is obtained by summing these reconstructed densities over all species:

ρ~l,nH(𝐱)=1L3𝐣ρ~^l,n;𝐣eiω𝐣𝐱,ρ~tot,nH(𝐱)=l=1mρ~l,nH(𝐱).\widetilde{\rho}_{l,n}^{H}(\mathbf{x})=\frac{1}{L^{3}}\sum_{\mathbf{j}\in\mathcal{H}}\widehat{\widetilde{\rho}}_{l,n;\mathbf{j}}e^{i\omega_{\mathbf{j}}\cdot\mathbf{x}},\qquad\widetilde{\rho}_{\mathrm{tot},n}^{H}(\mathbf{x})=\sum_{l=1}^{m}\widetilde{\rho}_{l,n}^{H}(\mathbf{x}). (99)

Using the reconstructed total density, the functions Dk,n(𝐱)=Dk(ρ~tot,nH(𝐱))D_{k,n}(\mathbf{x})=D_{k}(\widetilde{\rho}_{\mathrm{tot},n}^{H}(\mathbf{x})) and Φn(𝐱)=Φ(ρ~tot,nH(𝐱))\Phi_{n}(\mathbf{x})=\Phi(\widetilde{\rho}_{\mathrm{tot},n}^{H}(\mathbf{x})) are first evaluated at the Fourier grid points x𝐪=L𝐪/Hx_{\mathbf{q}}=L\mathbf{q}/H, 𝐪\mathbf{q}\in\mathcal{H}. The unnormalized Fourier coefficients of these functions are approximated as

D^k,n;𝐣\displaystyle\widehat{D}_{k,n;\mathbf{j}} =L3H3𝐪Dk,n(x𝐪)eiω𝐣x𝐪,\displaystyle=\frac{L^{3}}{H^{3}}\sum_{\mathbf{q}\in\mathcal{H}}D_{k,n}(x_{\mathbf{q}})e^{-i\omega_{\mathbf{j}}\cdot x_{\mathbf{q}}}, (100)
Φ^n;𝐣\displaystyle\widehat{\Phi}_{n;\mathbf{j}} =L3H3𝐪Φn(x𝐪)eiω𝐣x𝐪,𝐣.\displaystyle=\frac{L^{3}}{H^{3}}\sum_{\mathbf{q}\in\mathcal{H}}\Phi_{n}(x_{\mathbf{q}})e^{-i\omega_{\mathbf{j}}\cdot x_{\mathbf{q}}},\qquad\mathbf{j}\in\mathcal{H}.

This coefficient computation extends Step 5 of Algorithm 1 to handle the density-dependent fields Dk,nD_{k,n} and Φn\Phi_{n}. The values required for the particle update are obtained via Fourier interpolation:

Dk,n(X~k,np)\displaystyle D_{k,n}(\widetilde{X}_{k,n}^{p}) =1L3𝐣D^k,n;𝐣eiω𝐣X~k,np,\displaystyle=\frac{1}{L^{3}}\sum_{\mathbf{j}\in\mathcal{H}}\widehat{D}_{k,n;\mathbf{j}}e^{i\omega_{\mathbf{j}}\cdot\widetilde{X}_{k,n}^{p}}, (101)
Dk,n(X~k,np)\displaystyle\nabla D_{k,n}(\widetilde{X}_{k,n}^{p}) =1L3𝐣(iω𝐣)D^k,n;𝐣eiω𝐣X~k,np,\displaystyle=\frac{1}{L^{3}}\sum_{\mathbf{j}\in\mathcal{H}}(i\omega_{\mathbf{j}})\widehat{D}_{k,n;\mathbf{j}}e^{i\omega_{\mathbf{j}}\cdot\widetilde{X}_{k,n}^{p}},
Φn(X~k,np)\displaystyle\Phi_{n}(\widetilde{X}_{k,n}^{p}) =1L3𝐣Φ^n;𝐣eiω𝐣X~k,np.\displaystyle=\frac{1}{L^{3}}\sum_{\mathbf{j}\in\mathcal{H}}\widehat{\Phi}_{n;\mathbf{j}}e^{i\omega_{\mathbf{j}}\cdot\widetilde{X}_{k,n}^{p}}.

These quantities, determined by the density at time tnt_{n}, are kept fixed during the update from tnt_{n} to tn+1t_{n+1}. Assuming that the frozen diffusivity Dk,nD_{k,n} is sufficiently smooth and nonnegative (after regularization if needed), the Fokker–Planck correspondence leads to the following update for the pp-th particle of species kk:

X~k,n+1p=\displaystyle\widetilde{X}_{k,n+1}^{p}={} X~k,np+Dk,n(X~k,np)δt+2Dk,n(X~k,np)δtNk,npχkΦn(X~k,np)[\displaystyle\widetilde{X}_{k,n}^{p}+\nabla D_{k,n}(\widetilde{X}_{k,n}^{p})\delta t+\sqrt{2D_{k,n}(\widetilde{X}_{k,n}^{p})\delta t}\,N_{k,n}^{p}-\chi_{k}\Phi_{n}(\widetilde{X}_{k,n}^{p})\Biggl[ (102)
𝒦ϵ,δtk(ϵδtc~k,n1)(X~k,np)+l=1mβklMlRlsCp,l𝒦ϵ,δtk(X~k,npX~l,ns)]δt,\displaystyle\nabla\mathcal{K}_{\epsilon,\delta t}^{k}\ast\left(\frac{\epsilon}{\delta t}\widetilde{c}_{k,n-1}\right)(\widetilde{X}_{k,n}^{p})+\sum_{l=1}^{m}\frac{\beta_{kl}M_{l}}{R_{l}}\sum_{s\in C_{p,l}}\nabla\mathcal{K}_{\epsilon,\delta t}^{k}(\widetilde{X}_{k,n}^{p}-\widetilde{X}_{l,n}^{s})\Biggr]\delta t,

where Nk,np𝒩(𝟎,I3)N_{k,n}^{p}\sim\mathcal{N}(\mathbf{0},I_{3}) denote independent standard Gaussian vectors, and 𝒦ϵ,δtk\mathcal{K}_{\epsilon,\delta t}^{k} is the Green’s function for the operator Δλk2ϵ/δt\Delta-\lambda_{k}^{2}-\epsilon/\delta t. For each ll, the batch Cp,lC_{p,l} consists of RlR_{l} indices sampled with replacement from {1,,Pl}\{1,\ldots,P_{l}\}; when l=kl=k, any occurrence with s=ps=p is omitted from the sum. The bracketed expression approximates c~k,n-\nabla\widetilde{c}_{k,n}, with its density contribution evaluated by random batching. Using Eq. (97), the updated positions yield the empirical densities and hence the source coefficients for the subsequent chemical field update. If DkD_{k} is constant and Φ1\Phi\equiv 1, Eq. (102) reduces to the standard SIPF-rr update.

Chemotaxis systems with nonlinear diffusion. Density-dependent diffusion is widely used in chemotaxis models to describe changes in cellular dispersal induced by the local population density, particularly when crowding leads to nonlinear motility effects [28, 8, 18, 29, 20]. For the porous-medium diffusion term Δ(ργ)=(γργ1ρ)\Delta(\rho^{\gamma})=\nabla\cdot(\gamma\rho^{\gamma-1}\nabla\rho) with γ>1\gamma>1, the diffusivity at time tnt_{n} can be approximated by Dn=γρ~nγ1D_{n}=\gamma\widetilde{\rho}_{n}^{\gamma-1}. The values of DnD_{n} and Dn\nabla D_{n} evaluated at particle positions are then computed from the density reconstruction in Eq. (99) and the Fourier interpolation in Eq. (101). After freezing the regularized DnD_{n} over one substep, we formally represent the diffusion equation using the associated Itô SDE, discretize it via the Euler–Maruyama scheme, and then apply the original chemotactic update.

Chemotaxis systems with anisotropic cell motility. Anisotropic motility arises in structured environments such as extracellular-matrix fibers and neural tracts [36, 38]. It can be modeled as

tρ=(𝔻(𝐱)ρρχ(𝐱)c),\partial_{t}\rho=\nabla\cdot\left(\mathbb{D}(\mathbf{x})\nabla\rho-\rho\chi(\mathbf{x})\nabla c\right), (103)

where 𝔻(𝐱)=(Dij(𝐱))ij\mathbb{D}(\mathbf{x})=(D_{ij}(\mathbf{x}))_{ij} denotes a sufficiently smooth, symmetric positive semidefinite motility tensor. If chemical diffusion remains isotropic, the spectral update of c~\widetilde{c} remains unchanged. The Fokker–Planck correspondence yields the formal particle update

X~n+1p=X~np+[(𝔻)(X~np)+χ(X~np)c~(X~np,tn)]δt+2𝔻(X~np)δtNnp,\widetilde{X}_{n+1}^{p}=\widetilde{X}_{n}^{p}+\left[(\nabla\cdot\mathbb{D})(\widetilde{X}_{n}^{p})+\chi(\widetilde{X}_{n}^{p})\nabla\widetilde{c}(\widetilde{X}_{n}^{p},t_{n})\right]\delta t+\sqrt{2\mathbb{D}(\widetilde{X}_{n}^{p})\delta t}\,N_{n}^{p}, (104)

where [(𝔻)(𝐱)]i=jxjDij(𝐱)[(\nabla\cdot\mathbb{D})(\mathbf{x})]_{i}=\sum_{j}\partial_{x_{j}}D_{ij}(\mathbf{x}), 2𝔻\sqrt{2\mathbb{D}} denotes the matrix square root, and Nnp𝒩(𝟎,I3)N_{n}^{p}\sim\mathcal{N}(\mathbf{0},I_{3}) denote independent standard Gaussian vectors. Thus, anisotropy modifies the particle solver, whereas the field update and random-batch evaluation of c~\nabla\widetilde{c} preserve the structure of Algorithm 2. We leave the interpolation of 𝔻\mathbb{D} and its divergence, together with the convergence analysis of the resulting scheme, for future work in a separate implementation study.

6 Conclusions

In this paper, we introduced a random batch variant [24] of the original SIPF method [51] to simulate the 3D fully parabolic KS system. This modification leverages the randomness in batch sampling to bypass the mean-field limit, which reduces computational complexity without sacrificing accuracy. We established a comprehensive convergence analysis for the fully coupled particle-field system. Specifically, we proved the convergence with high probability for both the density ρ~(𝐱,t)\widetilde{\rho}(\mathbf{x},t) and the concentration field c~(𝐱,t)\widetilde{c}(\mathbf{x},t) to their respective exact solutions ρ(𝐱,t){\rho}(\mathbf{x},t) and c(𝐱,t)c(\mathbf{x},t). The error bounds revealed a dependence on δt\delta t, HH, and PP, with the density and concentration fields exhibiting distinct but interrelated convergence behaviors.

Finally, we performed numerical experiments to verify these theoretical rates and confirm the robustness of the SIPF-rr method. Notably, the algorithm effectively captures density concentration and potential finite-time blow-up phenomena subject to the critical threshold of initial mass in three space dimensions under specific initial data configurations, even with discretization parameters milder than those required by the worst-case theoretical bounds.

Future work will focus on refining the numerical analysis of the classical model and extending the SIPF-rr method to broader biological applications. From a theoretical perspective, we will refine the error estimates, particularly to weaken the strong dependence on the Fourier mode HH, and improve the efficiency of the algorithm in high-dimensional settings. Beyond refinements to the classical model, we will build on the formulations presented in Section 5 to develop SIPF-rr schemes for multi-species chemotaxis systems with volume-exclusion effects, nonlinear diffusion, and anisotropic cell motility, and to investigate their numerical performance and convergence properties.

Acknowledgements

ZZ was partially supported by the National Natural Science Foundation of China (Projects 92470103 and 12171406), the Hong Kong RGC grant (projects 17304324 and 17300325), the Seed Funding Programme for Basic Research (HKU), and the Hong Kong RGC Research Fellow Scheme 2025. ZW was partially supported by NTU SUG-023162-00001 and MOE AcRF Tier 1 Grant RG17/24. JX was partially supported by NSF grant DMS-2309520, the Swedish Research Council grant no. 2021-06594 at the Institut Mittag-Leffler in Djursholm, Sweden, and the E. Schrödinger Institute in Vienna, Austria, during his stay in the Fall of 2025. The simulations were performed on the research computing facilities of the Information Technology Services at the University of Hong Kong, and the Greenplanet Cluster at UC Irvine.

References

  • [1] N. Bellomo, A. Bellouquid, Y. Tao, and M. Winkler (2015) Toward a mathematical theory of Keller-Segel models of pattern formation in biological tissues. Mathematical Models and Methods in Applied Sciences 25 (09), pp. 1663–1763. Cited by: §1.
  • [2] F. Bubba, C. Pouchol, N. Ferrand, G. Vidal, L. Almeida, B. Perthame, and M. Sabbah (2019) A chemotaxis-based explanation of spheroid formation in 3D cultures of breast cancer cells. Journal of Theoretical Biology 479, pp. 73–80. Cited by: §1.
  • [3] M. Burger, M. Di Francesco, and Y. Dolak-Struss (2006) The Keller–Segel model for chemotaxis with prevention of overcrowding: linear vs. nonlinear diffusion. SIAM Journal on Mathematical Analysis 38 (4), pp. 1449–1475. Cited by: §5.
  • [4] Z. Cai, J. Liu, and Y. Wang (2024) Convergence of random batch method with replacement for interacting particle systems. arXiv preprint arXiv:2407.19315. Cited by: §1, §2.
  • [5] A. Chertock, Y. Epshteyn, H. Hu, and A. Kurganov (2018) High-order positivity-preserving hybrid finite-volume-finite-difference methods for chemotaxis systems. Advances in Computational Mathematics 44, pp. 327–350. Cited by: §1.
  • [6] A. Chertock and A. Kurganov (2008) A second-order positivity preserving central-upwind scheme for chemotaxis and haptotaxis models. Numerische Mathematik 111, pp. 169–205. Cited by: §1, §1.
  • [7] K. Craig and A. Bertozzi (2016) A blob method for the aggregation equation. Mathematics of Computation 85 (300), pp. 1681–1717. Cited by: §1.
  • [8] H. J. Eberl, D. F. Parker, and M. C. Van Loosdrecht (2001) A new deterministic spatio-temporal continuum model for biofilm development. Computational and Mathematical Methods in Medicine 3 (3), pp. 161–175. Cited by: §5.
  • [9] Y. Epshteyn and A. Izmirlioglu (2009) Fully discrete analysis of a discontinuous finite element method for the Keller-Segel chemotaxis model. Journal of Scientific Computing 40, pp. 211–256. Cited by: §1.
  • [10] Y. Epshteyn (2009) Discontinuous Galerkin methods for the chemotaxis and haptotaxis models. Journal of Computational and Applied Mathematics 224 (1), pp. 168–181. Cited by: §1.
  • [11] Y. Epshteyn (2012) Upwind-difference potentials method for Patlak-Keller-Segel chemotaxis model. Journal of Scientific Computing 53, pp. 689–713. Cited by: §1.
  • [12] I. Fatkullin (2013) A study of blow-ups in the Keller–Segel model of chemotaxis. Nonlinearity 26 (1), pp. 81–94. Cited by: §1.
  • [13] F. Filbet (2006) A finite volume scheme for the Patlak-Keller-Segel chemotaxis model. Numerische Mathematik 104, pp. 457–488. Cited by: §1.
  • [14] D. Godinho and C. Quininao (2015) Propagation of chaos for a subcritical Keller-Segel model. In Annales de l’IHP Probabilités et Statistiques, Vol. 51, pp. 965–992. Cited by: §1.
  • [15] I. Goodfellow, Y. Bengio, and A. Courville (2016) Deep learning. MIT press. Cited by: §2.
  • [16] J. Haškovec and C. Schmeiser (2009) Stochastic particle approximation for measure valued solutions of the 2D Keller-Segel system. Journal of Statistical Physics 135, pp. 133–151. Cited by: §1.
  • [17] M. A. Herrero, E. Medina, and J. J. Velázquez (1998) Self-similar blow-up for a reaction-diffusion system. Journal of Computational and Applied Mathematics 97 (1-2), pp. 99–119. Cited by: §1.
  • [18] T. Hillen and K. J. Painter (2009) A user’s guide to PDE models for chemotaxis. Journal of Mathematical Biology 58 (1), pp. 183–217. Cited by: §5.
  • [19] T. Hillen and K. Painter (2001) Global existence for a parabolic chemotaxis model with prevention of overcrowding. Advances in Applied Mathematics 26 (4), pp. 280–301. Cited by: §5.
  • [20] T. Höfer, J. A. Sherratt, and P. K. Maini (1995) Dictyostelium discoideum: cellular self-organization in an excitable biological medium. Proceedings of the Royal Society of London. Series B: Biological Sciences 259 (1356), pp. 249–257. Cited by: §5.
  • [21] B. Hu, Z. Wang, J. Xin, and Z. Zhang (2024) A stochastic interacting particle-field algorithm for a haptotaxis advection-diffusion system modeling cancer cell invasion. arXiv:2407.05626. Cited by: §5.
  • [22] J. Hu, Z. Wang, J. Xin, and Z. Zhang (2026) A novel stochastic particle-field algorithm for a reaction-diffusion-advection cancer invasion model. arXiv:2605.20140. Cited by: §5.
  • [23] Z. Huang, S. Jin, and L. Li (2025) Mean field error estimate of the random batch method for large interacting particle system. ESAIM: Mathematical Modelling and Numerical Analysis 59 (1), pp. 265–289. Cited by: §1, §2.
  • [24] S. Jin, L. Li, and J. Liu (2020) Random batch methods (RBM) for interacting particle systems. Journal of Computational Physics 400, pp. 108877. Cited by: §1, §2, §3.2, §6.
  • [25] S. Jin and L. Li (2021) Random batch methods for classical and quantum interacting particle systems and statistical samplings. In Active Particles, Volume 3: Advances in Theory, Models, and Applications, pp. 153–200. Cited by: §1, §2.
  • [26] E. F. Keller and L. A. Segel (1970) Initiation of slime mold aggregation viewed as an instability. Journal of Theoretical Biology 26 (3), pp. 399–415. Cited by: §1.
  • [27] T. W. Körner (2022) Fourier analysis. Cambridge University Press. Cited by: §1.
  • [28] R. Kowalczyk (2005) Preventing blow-up in a chemotaxis model. Journal of Mathematical Analysis and Applications 305 (2), pp. 566–588. Cited by: §5.
  • [29] H. Kuiper and L. Dung (2007) Global attractors for cross diffusion systems on domains of arbitrary dimension. The Rocky Mountain Journal of Mathematics, pp. 1645–1668. Cited by: §5.
  • [30] X. H. Li, C. Shu, and Y. Yang (2017) Local discontinuous Galerkin method for the Keller-Segel chemotaxis model. Journal of Scientific Computing 73 (2), pp. 943–967. Cited by: §1, §1.
  • [31] J. Liu, L. Wang, and Z. Zhou (2018) Positivity-preserving and asymptotic preserving method for 2D Keller-Segel equations. Mathematics of Computation 87 (311), pp. 1165–1189. Cited by: §1.
  • [32] J. Liu and R. Yang (2017) A random particle blob method for the Keller-Segel equation and convergence analysis. Mathematics of Computation 86 (304), pp. 725–745. Cited by: §1, §1.
  • [33] J. Liu and R. Yang (2019) Propagation of chaos for the Keller-Segel equation with a logarithmic cut-off. Methods and Applications of Analysis 26 (4), pp. 319–348. Cited by: §1, §1.
  • [34] A. I. Markushevich (2013) Theory of functions of a complex variable. American Mathematical Society. Cited by: §1.
  • [35] T. Nagai (1995) Blow-up of radially symmetric solutions to a chemotaxis system. Advances in Mathematical Sciences and Applications 5, pp. 581. Cited by: §1.
  • [36] H. G. Othmer and T. Hillen (2002) The diffusion limit of transport equations II: chemotaxis equations. SIAM Journal on Applied Mathematics 62 (4), pp. 1222–1250. Cited by: §5.
  • [37] K. J. Painter and T. Hillen (2002) Volume-filling and quorum-sensing in models for chemosensitive movement. Canadian Applied Mathematics Quarterly 10 (4), pp. 501–543. Cited by: §5, §5.
  • [38] K. J. Painter (2009) Continuous models for cell migration in tissues and applications to cell sorting via differential chemotaxis. Bulletin of Mathematical Biology 71 (5), pp. 1117–1147. Cited by: §5.
  • [39] K. J. Painter (2019) Mathematical models for chemotaxis and their applications in self-organisation phenomena. Journal of Theoretical Biology 481, pp. 162–182. Cited by: §1.
  • [40] C. S. Patlak (1953) Random walk with persistence and external bias. The Bulletin of Mathematical Biophysics 15, pp. 311–338. Cited by: §1.
  • [41] B. Perthame (2006) Transport equations in biology. Springer Science & Business Media. Cited by: §1.
  • [42] N. Saito and T. Suzuki (2005) Notes on finite difference schemes to a parabolic-elliptic system modelling chemotaxis. Applied Mathematics and Computation 171 (1), pp. 72–90. Cited by: §1.
  • [43] N. Saito (2007) Conservative upwind finite-element method for a simplified Keller-Segel system modelling chemotaxis. IMA Journal of Numerical Analysis 27 (2), pp. 332–365. Cited by: §1.
  • [44] N. Saito (2011) Error analysis of a conservative finite-element approximation for the Keller-Segel system of chemotaxis. Communications on Pure and Applied Analysis 11 (1), pp. 339–364. Cited by: §1.
  • [45] J. Shen and J. Xu (2020) Unconditionally bound preserving and energy dissipative schemes for a class of Keller-Segel equations. SIAM Journal on Numerical Analysis 58 (3), pp. 1674–1695. Cited by: §1.
  • [46] J. A. Sherratt (1994) Chemotaxis and chemokinesis in eukaryotic cells: the Keller-Segel equations as an approximation to a detailed model. Bulletin of Mathematical Biology 56 (1), pp. 129–146. External Links: ISSN 0092-8240 Cited by: §1.
  • [47] M. Sommerfeld, J. Schrieber, Y. Zemel, and A. Munk (2019) Optimal transport: fast probabilistic approximation with exact solvers. Journal of Machine Learning Research 20 (105), pp. 1–23. Cited by: §2.
  • [48] A. Stevens (2000) The derivation of chemotaxis equations as limit dynamics of moderately interacting stochastic many-particle systems. SIAM Journal on Applied Mathematics 61 (1), pp. 183–212. Cited by: §1, §1.
  • [49] Z. Wang, J. Xin, and Z. Zhang (2022) DeepParticle: learning invariant measure by a deep neural network minimizing Wasserstein distance on data generated by an interacting particle method. Journal of Computational Physics 464, pp. 111309. Cited by: §2.
  • [50] Z. Wang, J. Xin, and Z. Zhang (2024) A DeepParticle method for learning and generating aggregation patterns in multi-dimensional Keller-Segel chemotaxis systems. Physica D: Nonlinear Phenomena 460, pp. 134082. Cited by: §2.
  • [51] Z. Wang, J. Xin, and Z. Zhang (2025) A novel stochastic interacting particle-field algorithm for 3D parabolic-parabolic Keller-Segel chemotaxis system. Journal of Scientific Computing 102 (3), pp. 1–23. Cited by: §1, §1, §2, §4.1.2, §6.
  • [52] Y. Xie, Z. Wang, and Z. Zhang (2024) Randomized methods for computing optimal transport without regularization and their convergence analysis. Journal of Scientific Computing 100 (2), pp. 37. Cited by: §2.
  • [53] T. Zhang, Z. Wang, J. Xin, and Z. Zhang (2026) A Bidirectional DeepParticle method for efficiently solving low-dimensional transport map problems. Journal of Computational Physics 561, pp. 114983. Cited by: §2.
  • [54] G. Zhou and N. Saito (2017) Finite volume methods for a Keller-Segel system: discrete energy, error estimates and numerical blow-up analysis. Numerische Mathematik 135 (1), pp. 265–311. Cited by: §1, §1.