Particle Systems and McKean–Vlasov Dynamics with Singular Interaction through Local Times

2025

Read Particle Systems and McKean–Vlasov Dynamics with Singular Interaction through Local Times (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-08837.txt.

Introduction

Imagine many particles sliding on a half-line, each pushed toward the origin by a force that grows with how much time all of them have spent there. Graeme Baker and co-authors study that system in this 2025 paper, where the particles bounce like reflected Brownian motions.

If the pull is too strong, the math breaks down in finite time, and the authors map exactly when that happens using a shifting graph structure. In the averaged limit, the system settles into steady or self-similar profiles, which they prove it approaches. The work is motivated by market liquidity models, a freezing problem, and cell shape models.

Catalog record

YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE134 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-08837

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Particle Systems and McKean–Vlasov Dynamics with Singular Interaction through Local Times Abstract We study a system of reflecting Brownian motions on the positive half-line in which each particle has a drift toward the origin determined by the local times at the origin of all the particles. We show that if this local time drift is too strong, such systems can exhibit a breakdown in their solutions in that there is a time beyond which the system cannot be extended. In the finite particle case we give a complete characterization of the finite time breakdown, relying on a novel dynamic graph structure. We consider the mean-field limit of the system in the symmetric case and show that there is a McKean–Vlasov representation. If the drift is too strong, the solution to the corresponding Fokker–Planck equation has a blow up in its solution. We also establish the existence of stationary and self-similar solutions to the McKean–Vlasov equation in the case where there is no breakdown of the system. This work is motivated by models for liquidity in financial markets, the supercooled Stefan problem, and a toy model for cell polarization. Graeme Baker 1 1 1 Department of Statistics, Columbia University, NY, USA [email protected] . , Ben Hambly 2 2 2 Mathematical Institute, University of Oxford, UK [email protected] . & Philipp Jettkant 3 3 3 Department of Mathematics, Imperial College London, UK, [email protected] . 1 Introduction In this article, we consider a dynamic model of Brownian particles that diffuse in the half-line [0,\infty) and are reflected at the origin. The particles interact through the reflection terms, pushing them closer to the reflecting boundary, leading to a positive feedback loop. If the interaction is too strong, this loop can lead to a breakdown of the system in finite time. We begin by studying networks with finitely many particles and, subsequently, consider an associated mean-field limit. 1.1 Finite Particle System The N particles are labelled i=1 , …, N and the state X^{i}_{t} of particle i satisfies the reflected SDE X^{i}_{t}=\xi_{i}+W^{i}_{t}-\sum_{j=1}^{N}q_{ij}L^{j}_{t}+L^{i}_{t} (1.1) for nonnegative weights q_{ij}\geq 0 . The initial conditions \xi_{1} , …, \xi_{N} are nonnegative random variables, (W^{1},\dots,W^{N}) is an N -dimensional Brownian motion with covariance matrix A=(a_{ij})_{ij}\in\mathbb{R}^{N\times N} and independent of (\xi_{1},\dots,\xi_{N}) , and the processes L^{1} , …, L^{N} are reflection terms, which ensure that each X^{i}_{t} stays nonnegative. We define the filtration \mathbb{F}^{N}=(\mathcal{F}^{N}_{t})_{t\geq 0} by \mathcal{F}^{N}_{t}=\sigma\Bigl(\xi_{i},W^{i}_{s}\mathpunct{:}s\in[0,t],\,i\in\{1,\dots,N\}\Bigr). Also, we set \mathbf{X}=(X^{1},\dots,X^{N}) and \mathbf{L}=(L^{1},\dots,L^{N}) , and define the adjacency matrix Q=(q_{ij})_{ij} , which allows us to rewrite SDE ( 1.1 ) in the form \mathbf{X}_{t}=\mathbf{X}_{0}+\mathbf{W}_{t}+(\mathbb{I}-Q)\mathbf{L}_{t}, (1.2) where \mathbb{I}\in\mathbb{R}^{N\times N} denotes the identity matrix in N dimensions The process \mathbf{X} is a semimartingale reflected Brownian motion (SRBM), i.e. an obliquely reflected semimartingale process in the first orthant [0,\infty)^{N} . Harrison & Reiman [ 18 ] established global well-posedness for SRBMs for adjacency matrices Q whose largest real left-eigenvalue \rho(Q) is strictly less than one. The objective of this work is to analyse the system when \rho(Q)\geq 1 . In this so far unexplored case, the system’s behaviour changes radically as it will typically no longer be globally well-posed. Let us illustrate it with a simple example: let N=2 , q_{ij}=1-\delta_{ij} for i , j\in\{1,2\} , and A=\mathbf{I} , so that W^{1} and W^{2} are independent standard Brownian motions. Suppose that SDE ( 1.1 ) has a global solution (X^{1}_{t},X^{2}_{t}) . Then summing X^{1}_{t} and X^{2}_{t} , we obtain X^{1}_{t}+X^{2}_{t}=\xi_{1}+W^{1}_{t}+L^{1}_{t}-L^{2}_{t}+\xi_{2}+W^{2}_{t}+L^{2}_{t}-L^{1}_{t}=(\xi_{1}+\xi_{2})+(W^{1}_{t}+W^{2}_{t}). (1.3) The left-hand side has to be nonnegative for t\geq 0 due to the reflection at the origin. However, the right-hand side becomes negative for some sufficiently large t>0 , resulting in a contradiction. Consequently, a breakdown has to occur in finite time, almost surely. The impossibility of solving SDE ( 1.1 ) globally was already pointed out by Harrison & Reiman at the end of Section 3 in [ 18 ] . However, they do not discuss whether SDE ( 1.1 ) may still be locally solvable when \rho(Q)\geq 1 and, if so, what causes the breakdown. We show in this work that local solutions for SDE ( 1.1 ) indeed exist for \rho(Q)\geq 1 and provide a precise characterisation of the breakdown time in terms of the number of particles located at the origin. If this number is too large, the feedback \sum_{j=1}^{N}q_{ij}L^{j}_{t} from the system is equally as strong as or stronger than the individual reflection term L^{i}_{t} . Thus, the reflection term can no longer compensate the feedback, and the system breaks down. Accordingly, SDE ( 1.1 ) will only be understood to hold on some random interval [0,\tau) for a stopping time \tau with values in [0,\infty] . Before we state a precise notion of solution for SDE ( 1.1 ), we set up some notation and introduce a dynamic graph structure determined by the interplay of Q and A , essential to the understanding of the breakdown phenomenon. For a matrix B=(b_{ij})_{ij}\in\mathbb{R}^{d\times d} with d\geq 1 , we let its spectral radius \rho(B) denote the largest real left-eigenvalue of B . If B has no real left-eigenvalues, we set \rho(B)=-\infty . Typically, the matrices under consideration will have nonnegative entries, in which case the Perron–Frobenius theorem implies that \rho(B)\geq 0 and that there exists a left-eigenvector corresponding to \rho(B) with nonnegative entries. Next, for a nonempty subset I\subset\mathbf{N}:=\{1,\dots,N\} and B=(b_{ij})_{ij}\in\mathbb{R}^{N\times N} , we define the reduced matrix or minor B[I]=(b_{ij})_{i,j\in I}\in\mathbb{R}^{I\times I} . If I=\emptyset , then B[I] is the empty matrix whose spectral radius we set to -\infty by convention. Usually, the set I will consist of the particles that are currently located at zero and are therefore, in principle, the only ones whose reflection term may increase. Note, however, that due to possible degeneracies of the covariance matrix A , even a particle located at the origin need not see any activity in its reflection term. Thus, within a given set I , we must further distinguish between active nodes whose reflection term increases and inactive nodes for which this is not the case. This distinction rests on the interaction between the weight matrix Q and the covariance matrix A , which determines a graph structure that we elaborate next. As before, let I be a nonempty subset of \mathbf{N} . Then the minor Q[I] induces a directed graph (I,E_{I}) , whose set of edges is given by E_{I}=\{(i,j)\in I\times I\mathpunct{:}q_{ji}>0\} . Note the inversion of indices in the definition of E_{I} , corresponding to the fact that we study left-eigenvectors of Q . We define the set of active nodes I^{a} as those i\in I such that there exists a j\in I with a_{jj}>0 and a path (i_{0},i_{1},\dots,i_{n}) in E_{I} from j to i , i.e. (i) i_{k}\in I for k=1 , …, n-1 ; (ii) i_{0}=j and i_{n}=i ; (iii) (i_{k},i_{k+1})\in E_{I} for k=0 , …, n-1 . That is, a node i\in I is active if it is reachable in (I,E_{I}) from a node j\in I whose Brownian motion W^{j} is nondegenerate. If I consists of the particles stuck at zero, then, as we alluded to above, only the active nodes will see their reflection terms increase, since they are the only ones exposed to the fluctuations of a nondegenerate Brownian motion. To

Read the complete text