Contraction and concentration of measures with applications to theoretical neuroscience
Simone Betteti, Francesco Bullo · 2025
Read Contraction and concentration of measures with applications to theoretical neuroscience by Simone Betteti, Francesco Bullo (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-2504-05666.txt.
Introduction
Simone Betteti and Francesco Bullo studied in twenty twenty-five how random systems settle into steady patterns over time. They look at stochastic systems whose drift pulls states together, a property called contraction. Classic results assume steady noise and drifts from a potential, but many real systems break both assumptions.
The authors allow noise that varies across space and drifts that no potential describes. Under global contraction, they prove a unique steady distribution exists and the system reaches it, as long as contraction beats the noise's unevenness. When contraction holds only far out, mass piles up near the low points of a linked potential. They show what this means for Hopfield networks, a classic model of memory, where noise can blur or protect stored patterns.
Catalog record
| CREATOR | Simone Betteti, Francesco Bullo |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 41 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2504-05666
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Contraction and concentration of measures with applications to theoretical neuroscience Simone Betteti, Francesco Bullo Abstract We investigate the asymptotic behavior of probability measures associated with stochastic dynamical systems featuring either globally contracting or B r B_{r} -contracting drift terms. While classical results often assume constant diffusion and gradient-based drifts, we extend the analysis to spatially inhomogeneous diffusion and non-integrable vector fields. We establish sufficient conditions for the existence and uniqueness of stationary measures under global contraction, showing that convergence is preserved when the contraction rate dominates diffusion inhomogeneity. For systems contracting only outside of a compact set and with constant diffusion, we demonstrate mass concentration near the minima of an associated non-convex potential, like in multistable regimes. The theoretical findings are illustrated through Hopfield networks, highlighting implications for memory retrieval dynamics in noisy environments. 1 Introduction The effective functioning of dynamical systems in real world context relies on convergence to desirable states. From cellular reactions to ecosystems, from small circuitry to entire plants, systems evolve dynamically to attain robust and stable configurations, ensuring operational efficiency. Contraction theory ( Lohmiller and Slotine, 1998 ; Simpson-Porco and Bullo, 2014 ; Davydov et al., 2022 ) has recently emerged as a powerful and insightful framework to study the convergence properties of dynamical systems. In essence, contraction theory establishes the asymptotic regularity of a system, whether through convergence to stable equilibria, limit cycles, or trajectories entrainment. Canonically, much of the literature on contraction theory has focused on deterministic system. However, in many practical scenarios, dynamical systems are inherently stochastic, either due to unknown environmental noise or because of the collective behavior of numerous interacting components. Therefore, it is critical to develop formal tools enabling the study of the asymptotic properties of stochastic dynamical systems. The study of stochastic system asymptotics has a rich and evolving history, and has primarily leveraged the associated probability measure as investigative tool. The canonical approach assumes that the drift field follows the gradient of a potential function ( Jordan et al., 1997 ) . More general dynamics consider drift vector fields that are sum of a gradient term and a rotational (divergence-free) term. Under the hypothesis of a constant diffusion term, convergence in Wasserstein distance to a unique stationary probability measure has been established in Bolley et al. (2012) . The case of globally contracting but not necessarily integrable drift fields has been investigated under similar diffusion assumptions ( Natile et al., 2011 ) . A first thrust to expand beyond the case of constant diffusivity has been investigated in Pham and Slotine (2013) ; Bouvrie and Slotine (2019) , with convergence bounds depending both on the contraction and diffusion constants. Finally, in Luo and Wang (2016) ; Monmarché (2023) it has been shown that when drift term is contracting only outside a compact set there exists a unique stationary probability measures and initial distributions converge exponentially fast towards it. Given the pervasiveness of stochastic processes in application, we believe in the need for additional investigation of the associated measures for spatially inhomogeneous diffusivity. Additionally, in our recent study ( Betteti et al., 2025 ) we investigated a stochastic system characterized by a drift term contracting only outside of a compact set and with constant diffusion. Aside from the question on the existence and uniqueness of an associated stationary measure, we were mostly interested in a rigorous understanding of the regions where most of its mass concentrates. Motivated by these open problems, the paper is structured as follows. Section 2 provides the necessary theoretical background, introducing the concept of c-strong contractivity, unique strong solutions to stochastic differential equations and the associated probability measures. Section 3 investigates stochastic systems with globally contracting drift fields and spatially inhomogeneous diffusion terms, providing sharper bounds on the convergence to a unique stationary measure than in Bouvrie and Slotine (2019) . Section 4 explores mass concentration properties for stochastic systems with drift field contracting outside of a ball centered at the origin. In particular, we prove that the mass concentrates around stable equilibria provided that the basins of attraction and local contraction rates are large enough. In addition, if the non-integrable drift field is associated to a non-convex potential, we prove that the mass tends to concentrates around the potential deepest minima. Finally, section 5 frames the theoretical results in the context of the Hopfield network, a dynamical system describing memory retrieval in the brain known for its multistable nature. Notation. We identify with B_{r}(x)\subset\real^{d} the ball of radius r>0 and centered at x\in\real^{d} . For a smooth manifold \mathcal{M}\subseteq{\real{}^{d}} we identify with T_{x}\mathcal{M} the tangent space of \mathcal{M} at x\in\mathcal{M} , and with \partial\mathcal{M} its boundary. We identify with C^{k}(\mathcal{X};\mathcal{Y}) the class of k -differentiable functions from \mathcal{X} into \mathcal{Y} . Let f\in C^{1}(\real^{d};\real) and denote with \nabla f(x)\in\real^{d} its gradient. We identify the partial derivative of f with respect to x_{i} as \partial_{x_{i}}f(x) . Let g\in C^{1}(\real^{d};\real^{d}) and denote with J_{x}g(x)\in\real^{d\times d} its Jacobian, and with \nabla\cdot g(x)\in\real its divergence. We identify the identity matrix of dimension d\in\mathbb{N} as \mathcal{I}_{d}\in\real^{d\times d} . We refer to a positive definite matrix A\in\real^{d\times d} as A\succ 0 . Let B\in\real^{d\times d} and we denote its trace operator as \mathrm{Tr}(B) . We denote with (\ ,\ )_{2}:\real^{d}\times\real^{d}\to\real the standard Euclidean inner product. We denote the weighted inner product as (\ ,\ )_{A}:\real^{d}\times\real^{d}\to\real , for A\succ 0 . We identify with \mathcal{P}_{2}(\mathcal{X}) the space of probability measures over \mathcal{X} having finite second moment. Stochastic processes are identified by boldface upper roman letters, i.e. \mathbf{X}_{t} . 2 Theoretical Background Consider the Ito stochastic differential equation (S.D.E.) \begin{cases}d\mathbf{X}_{t}=f(t,\mathbf{X}_{t})\>dt+G(t,\mathbf{X}_{t})\>d\mathbf{B}_{t}\\ \mathbf{X}(0)=x_{0}\in\real^{d}\end{cases} (1) where \mathbf{B}_{t} is the standard d-dimensional Brownian motion. Assumption 1 (Lipschitzianity and sublinearity) . Let f:\real\times\real^{d}\rightarrow\real^{d} be the drift term, globally Lipschitz with constant L_{f}>0 and sublinear with constant s_{f}>0 . \displaystyle(f(t,x)-f(t,y),x-y)_{2} \displaystyle\leq L_{f}\|x-y\|^{2}\qquad\forall x,y\in\real^{d} (2) \displaystyle\|f(t,x)\|^{2} \displaystyle\leq s_{f}(1+\|x\|^{2}) (3) The diffusion term G:\real\times\real^{d}\rightarrow\real^{d\times d} obeys equivalent constraints. Under these standard assumptions ( Karatzas and Shreve, 2014 ) we can guarantee the existence and uniqueness of a strong solution to eq. ( 1 ). Definition 2 (Infinitesimal generator) . Let h:\real\times\real^{d}\rightarrow\real be a Lebesgue-measurable function. The infinitesimal generator of h along the process ( 1 ) is defined as \displaystyle\mathcal{A}h(t,x_{t})= \displaystyle\lim_{\updelta t\rightarrow 0}\frac{\operatorname{\mathbb{E}}_{\mathbf{B_{t}}}[h(t+\updelta t,\mathbf{X}_{t+\updelta t})]-h(t,x_{t})}{\updelta t} \displaystyle= \displaystyle\partial_{t}h(t,x_{t})+\nabla h(t,x_{t})^{\top}f(x_{t})\ \displaystyle+\frac{1}{2}\mathrm{Tr}\{G(t,x_{t})\nabla^{2}h(t,x_{t})G(t,x_{t})^{\top}\}\ . (4) The probability measure \upmu\in\mathcal{P}_{2}(\real_{\geq 0}\times\real^{d}) is related ( Pavliotis, 2014 ) to the S.D.E. ( 1 ) via the Kolmogorov forward equation (or Fokker-Planck equation) \displaystyle\partial_{t}\upmu(t,x)= \displaystyle\nabla\cdot\{-f(t,x)\upmu(t,x)+\frac{1}{2}\nabla\cdot[D(t,x)\upmu(t,x)]\} (5) \displaystyle\upmu(0,x)= \displaystyle\upmu_{0}(x)\qquad\forall x\in\real^{d} (6) where D(t,x)=G(t,x)G(t,x)^{\top} and \upmu_{0}\in\mathcal{P}_{2}(\real^{d}) is the initial distribution of the states, i.e. the distribution of the initial conditions for eq. ( 1 ). Additionally, we introduce the notion of distance in \mathcal{P}_{2}(\real^{d}) between two probability measures ( Givens and Shortt, 1984 ) . Definition 3 (Wasserstein metric) . Let \upmu_{1},\upmu_{2}\in\mathcal{P}_{2}(\real^{d}) be two probability measures having finite second moment. Define \Gamma(\upmu_{1},\upmu_{2}) the set of the joint probability measures \upgamma having \upmu_{1} and \upmu_{2} as marginals. The p -Wasserstein metric is defined as \displaystyle W_{p}(\upmu_{1},\upmu_{2}) \displaystyle=\left(\inf_{\upgamma\in\Gamma(\upmu_{1},\upmu_{2})}\operatorname{\mathbb{E}}_{\upgamma}[\|x-y\|^{p}]\right)^{\frac{1}{p}}. (7) For the sake of brevity, from now on we will abbreviate \Gamma(\upmu_{1},\upmu_{2}) as simply \Gamma , and \upmu(t,x)=\upmu_{t}(x) . 3 Contracting drifts and convergence to stationary measures We begin by introducing the definition of global contractivity, and henceforth consider only