betaFeedback

WorksMathematics

Model selection for stochastic dynamics: a parsimonious and principled approach

Abstract

This thesis focuses on the discovery of stochastic differential equations (SDEs) and stochastic partial differential equations (SPDEs) from noisy and discrete time series. A major challenge is selecting the simplest possible correct model from vast libraries of candidate models, where standard information criteria (AIC, BIC) are often limited. We introduce PASTIS (Parsimonious Stochastic Inference), a new information criterion derived from extreme value theory. Its penalty term, nBln⁡(n0/p)n_\mathcal{B} \ln(n_0/p), explicitly incorporates the size of the initial library of candidate parameters (n0n_0), the number of parameters in the considered model (nBn_\mathcal{B}), and a significance threshold (pp). This significance threshold represents the probability of selecting a model containing more parameters than necessary when comparing many models. Benchmarks on various systems (Lorenz, Ornstein-Uhlenbeck, Lotka-Volterra for SDEs; Gray-Scott for SPDEs) demonstrate that PASTIS outperforms AIC, BIC, cross-validation (CV), and SINDy (a competing method) in terms of exact model identification and predictive capability. Furthermore, real-world data can be subject to large sampling intervals (ΔtΔt) or measurement noise (σσ), which can impair model learning and selection capabilities. To address this, we have developed robust variants of PASTIS, PASTIS-ΔtΔt and PASTIS-σσ, thus extending the applicability of the approach to imperfect experimental data. PASTIS thus provides a statistically grounded, validated, and practical methodological framework for discovering simple models for processes with stochastic dynamics.

TECHNICAL REPORT

Software AMU ... University

by

Andreas Ge... Engineering ...

Page de titre

Model selection for stochastic dynamics: A parsimonious and principled approach

Andonis Gerardos

Discipline Physique et Sciences de la Matière

Spécialité Biophysique

École doctorale ED 352 PHYSIQUE ET SCIENCES DE LA MATIERE

Laboratoire/Partenaires de recherche CENTURI; CINaM; Amidex

Composition du jury

Jean Barbier — Rapporteur Associate Professor, ICTP, SISSA

Christian L. Müller — Rapporteur Professeur, Helmholtz, Université de Munich

Thierry Mora — Examinateur Directeur de Recherche, CNRS, LPENS

Simon Gsell — Examinateur Maitre de Conférences, IR-PHE, amU

Claude Loverdo — Présidente du jury Directrice de Recherche, CNRS, LJP

Pierre Ronceray — Directeur de thèse Chargé de Recherche, CNRS, CINaM

Affidavit

Je soussigné, Andonis Gerardos, déclare par la présente que le travail présenté dans ce manuscrit est mon propre travail, réalisé sous la direction scientifique de Pierre Ronceray, dans le respect des principes d’honnêteté, d’intégrité et de responsabilité inhérents à la mission de recherche. Les travaux de recherche et la rédaction de ce manuscrit ont été réalisés dans le respect à la fois de la charte nationale de déontologie des métiers de la recherche et de la charte AMU relative à la lutte contre le plagiat.

Ce travail n’a pas été précédemment soumis en France ou à l’étranger dans une version identique ou similaire à un organisme examinateur.

Fait à Marseille le 05/05/2025

Cette œuvre est mise à disposition selon les termes de la Licence Creative Commons Attribution - Pas d’Utilisation Commerciale - Pas de Modification 4.0 International.

Publications et Communications Scientifiques

Publications issues des travaux de thèse

  1. [Statut : Publié] S. Amiri, Y. Zhang, A. Gerardos, C. Sykes et P. Ronceray. Inferring geometrical dynamics of cell nucleus translocation. Physical Review Research, 6, 043030 (2024). DOI: 10.1103/PhysRevResearch.6.043030

  1. [Statut : Preprint / Soumis à Physical Review Letter - Jan. 2025] A. Gerardos et P. Ronceray. Principled model selection for stochastic dynamics. https://arxiv.org/abs/2501.10339

Communications et participation aux conférences et écoles

  1. iPoLS Annual Meeting, Montpellier, France, Mai 2022.

  1. École d’été de Physique de Beg Rohu, Saint-Pierre-Quiberon, France, Juin 2022.

  1. École d’été de Physique Théorique des Houches (Session Biophysique), Les Houches, France, Juillet 2023. (Poster)

  1. CENTURI Conference, Cargèse, France, Octobre 2023. (Poster)

  1. DPG Spring Meeting (DPG Frühjahrstagung - Section Biological Physics), Berlin, Allemagne, Mars 2024. (Communication orale) https://www.dpg-verhandlungen.de/year/2024/conference/berlin/part/bp/session/23/contribution/5

  1. BioInference 2024, Warwick, Royaume-Uni, Juin 2024. (Communication orale) https://bioinference.github.io/2024/

  1. École d’été Systèmes Complexes et Vitreux, Cargèse, France, Juillet 2024. (Poster)

  1. Physical Biology Circle Meeting, Lisbonne, Portugal, Janvier 2025. (Communication orale)

  1. Stochastics processes: Inferences in complex systems, Lausanne, Suisse, Mai 2025. (Poster) https://www.cecam.org/workshop-details/stochastics-processes-infere

  1. Statphys 2025 (Statphys 29), Florence, Italie, Juillet 2025. (Communication orale)

Résumé et mots clés

Abstract and keywords

Remerciements

Le 1er juillet 2025 marque la fin des montagnes russes que fut ma thèse. Ce manège, avec ses vertiges et ses accélérations, aurait sans doute déraillé sans les qualités humaines et scientifiques exceptionnelles de son incroyable co-pilote : mon directeur de thèse, Pierre Ronceray. Un immense merci à toi, Pierre, pour m’avoir maintenu sur les rails tout en me laissant les tracer.

Mes remerciements vont également à l’ensemble des membres de mon jury pour l’intérêt et le temps précieux qu’ils ont accordé à mon travail. Merci à Jean Barbier, Christian L. Müller, Thierry Mora, Simon Gsell et Claude Loverdo. Je suis particulièrement reconnaissant envers les membres de mon comité de suivi, Thierry Mora et Anna Frishman, pour leurs conseils avisés tout au long de ce parcours. Les discussions scientifiques stimulantes ont été un moteur, et je remercie chaleureusement Nicolas Levernier, Florence Bansept, Simon Gsell et Philippe Roudot. Une pensée spéciale pour João Pedro Valeriano Miranda, avec qui j’ai eu des échanges scientifiques profonds.

L’ambiance d’un laboratoire est essentielle, et je voudrais remercier tous ses membres pour leur accueil. Plus particulièrement, merci à Kheya Sengupta et Pierre Müller pour leur constante bienveillance dans mes péripéties. Merci également à Didier Tonneau, Étienne Loiseau, Carole Fauquet, Emmanuèl Helfer, Dominique Chatain et Annie Viallat. La science moderne ne serait rien sans une infrastructure robuste ; merci donc à Andres Saul et Baptiste Desmoulin pour l’entretien méticuleux du cluster de calcul du CINaM.

Cette thèse n’aurait pu voir le jour sans le soutien financier d’institutions françaises et Marseillaise. Je remercie vivement Centuri (Turing Centre for Living Systems) pour le financement de mes trois années de thèse, ainsi que le projet “Anita” d’AMIDEX pour les neuf derniers mois. L’aventure Centuri, d’ailleurs, ne serait pas aussi riche sans son incroyable équipe administrative qui orchestre des événements mémorables. Un grand merci à Marlène Salom, Melina De Oliveira, Pauline Becherel, Rémi Wojciechowski, Simon Legendre et Alizée Guarino. Dans ce quotidien, la gentillesse et le sourire des personnels du Crous ont été de vrais rayons de soleil.

Ces dernières années n’auraient pas eu la même saveur sans les discussions animées, souvent à propos d’urbanisme et du fameux Boulevard Urbain Sud, autour d’une andouillette grillée par le Crous. Merci à Sham Tlili, Marc-Eric Perrin, Martin Lardy, Jules Vanaret et Alice Gros pour ces moments précieux. Après ces festins, j’ai eu la chance de retrouver une ambiance tout aussi chaleureuse au laboratoire grâce à Alice Briole, Pauline Lahure, Marion Marchand, Sacha Soulerin et Zakaria Marmri.

Mon parcours n’aurait sans doute jamais commencé sans le goût pour la science que mes professeurs de classes préparatoires ont su m’insuffler. Merci à Monsieur Minetti, Monsieur Turner, Monsieur Petrov et Monsieur Boulesteix. Ce chemin a ensuite été jalonné de rencontres précieuses durant mes études et expériences en recherche ; merci à Yohann Geli, Ayoub Hihi, Luca Castello, Jérôme Garnier-Brun, Nicolas Dietler et Loïc Marrec pour leur soutien indéfectible. Enfin, il y a les rencontres qui changent une trajectoire. C’est pendant mon Master que j’ai eu la chance de croiser la route de ma très chère Joséphine Baillet. Son soutien a été si déterminant que, sans elle, je n’écrirais pas ces lignes aujourd’hui : j’aurais tout simplement raté la date limite de candidature à mon Master 2. Pour cela et pour tout le reste, merci ma cocotte.

Enfin, et surtout, cette thèse n’aurait tout simplement pas existé sans ma mère, qui m’a accordé une confiance absolue et a sans cesse cultivé ma curiosité par ses questions. Elle n’aurait pas existé non plus sans mon défunt grand-père, qui me répétait au potager qu’il valait mieux travailler avec son cerveau qu’avec ses mains. Je remercie aussi ma grand-mère, pour sa douceur et son incroyable talent à nous unir autour d’un bon repas.

À ma maman, mon papou et ma mamie.

Code Availability

The source code developed for the research presented in this manuscript is openly and publicly available.

The primary development repository is hosted on GitHub and can be accessed at: https://github.com/odannis/PASTIS_PhD

For long-term preservation and to ensure the specific version of the code used to generate the results in this thesis is permanently accessible, the repository has been archived on Zenodo. The archived version, which corresponds to the analyses and simulations described herein (Version [X.X.X] or commit hash [specific-commit-hash]), is available via the following Digital Object Identifier (DOI): https://doi.org/10.5281/zenodo.15814187

Use of AI tools

Generative artificial intelligence (AI) tools were employed as an aid during the research and writing process for this thesis. Among the specific uses were checking the written explanation for spelling and grammar mistakes, making it clear and to the point, creating partial code for plotting figures, debugging software implementations, finding relevant literature, and explaining math demonstrations. The core research ideas, methodologies (including the development of the PASTIS criterion), results, and interpretations presented are the original work of the author.

General Introduction: Discovering Governing Equations from Stochastic Dynamics

The mathematical modeling of complex systems across diverse scientific fields, from physics and biology to finance and climate science, increasingly relies on the framework of stochastic differential equations (SDEs) and stochastic partial differential equations (SPDEs) [1, 2]. These tools provide a powerful language for describing systems whose dynamics are influenced not only by deterministic rules but also by inherent randomness or fluctuating environments [3–45]. While the foundational theory of SDEs, originating from the study of Brownian motion [46–48], provides a basis for understanding these processes, the discovery of governing equations directly from time-series data remains a critical challenge.

Often, while we might hypothesize a range of potential mechanisms or interactions that could govern a system’s evolution, the true underlying dynamics are expected to be parsimonious – involving only a subset of these possibilities. This leads to a fundamental problem in data-driven modeling: how can we reliably identify the correct, minimal set of terms describing the system’s drift and diffusion from noisy, finite data, especially when faced with a large library of candidate functions representing potential physical laws or interactions? This task falls within the rapidly growing field of data-driven discovery of dynamical systems [49, 50].

Standard statistical approaches to model selection, such as those based on the Akaike Information Criterion (AIC) [51] or Bayesian Information Criterion (BIC) [52], often struggle in this sparse discovery context. While valuable for comparing a small number of pre-defined models, these criteria can be prone to overfitting, selecting overly complex models laden with superfluous terms when searching through large model spaces. This "multiple comparisons" problem fundamentally limits their ability to reliably uncover the true sparse structure underlying stochastic dynamics.

This thesis addresses this critical gap by developing and validating a novel, statistically principled framework for Parsimonious Stochastic Inference, termed PASTIS. Rooted in likelihood-based inference [53] and leveraging insights from extreme value theory [54], PASTIS introduces an information criterion specifically designed to penalize model complexity in a way that explicitly accounts for the size of the candidate function library being searched. By establishing a direct link between the penalty term and a user-defined statistical significance level, PASTIS aims to control the rate of false discoveries, enabling more robust identification of the truly relevant terms governing the system’s dynamics.

The work presented herein details the theoretical derivation of the PASTIS criterion, explores its properties, and demonstrates its efficacy through comprehensive benchmarking against established methods (including AIC, BIC, and sparse regression techniques like SINDy [55, 56]) across a variety of SDE systems. Furthermore, the framework is extended to the significantly more challenging domain of identifying sparse SPDEs from spatio-temporal data. Finally, recognizing that real-world data are rarely ideal, we investigate and propose adaptations to the PASTIS methodology to enhance its robustness against common data imperfections, namely finite sampling intervals and the presence of measurement noise.

This thesis is structured as follows:

  • Part I lays the necessary groundwork. Chapters 1 and 2 introduce the Langevin equation, SDEs, and the essential tools of stochastic calculus.

  • Part II establishes the techniques for inferring parameters for a given SDE or SPDE model, discussing likelihood-based approaches and estimators robust to data imperfections.

  • Part III focuses on the core contribution: sparse model selection. Chapter 8 introduces likelihood-based inference in the context of model selection and highlights the limitations of standard criteria like AIC and BIC. Chapter § 14.4 derives the PASTIS information criterion using extreme value statistics. Chapter § 15 presents extensive validation of PASTIS on benchmark SDE systems. Chapter § 15.2 extends the PASTIS framework to SPDE discovery. Chapter § 15.3 details the adaptation of PASTIS for robustness against non-ideal data. Finally, Chapter § 15.4 provides a concluding discussion for this part, summarizing the advantages and limitations of the PASTIS framework.

Through this work, we aim to provide a robust and practical tool for uncovering the fundamental equations governing complex stochastic phenomena directly from data.

Visual abstract of the PASTIS method, showing noisy and sampled data being processed by PASTIS to recover a minimal model

Figure 1. Visual Abstract of the PASTIS method.

Introduction to the Langevin equation

An Introduction to the Langevin Equation and Stochastic Calculus

“Life is far too important a thing ever to talk seriously about.”

Oscar Wilde, Lady Windermere’s Fan

The Birth of the Langevin Equation

Emergence of a New Mathematical Object from Observations of Brownian Motion

The study of microscopic dynamics often begins with Brownian motion: the incessant, seemingly random jiggling of microscopic particles suspended in a fluid (liquid or gas). While the particles themselves are typically too small to see with the naked eye, their erratic dance becomes strikingly apparent under a microscope. Figure 1.1a captures this, showing trajectories recorded by Jean Perrin in 1909. This microscopic chaos has macroscopic consequences; the gradual spreading of a drop of ink in water, depicted in Figure 1.1b, is a visible manifestation driven by the underlying Brownian motion of countless ink particles (alongside convection).

The phenomenon owes its name to the Scottish botanist Robert Brown, who, in 1827, provided the first detailed scientific account while observing pollen grains suspended in water. He meticulously documented the motion but could not determine its origin.

Putty grain trajectories

(a) (a) Putty grain trajectories

Ink diffusion in water

(b) (b) Ink diffusion in water

Figure 1.1. (a) From Jean Perrin’s 1909 work [57, p. 81]: Three tracings show the motion of a putty grain (radius 0,53 µm) observed under a microscope. Successive positions recorded every 30 seconds are connected by straight lines (grid mesh size is 3,2 µm). (b) Diffusion of an ink drop in a beaker of water, a macroscopic consequence of Brownian motion and convection (Credit: McGraw-Hill/Charles D. Winters).

Explaining the physical mechanism behind this ceaseless motion became a major scientific puzzle. A pivotal breakthrough arrived in 1905 when Albert Einstein, building upon the kinetic theory of heat, developed a purely theoretical description [46]. He derived a quantitative relationship linking the particle’s mean squared displacement (MSD) to observable macroscopic quantities:

MSD=E[(x(t)−x(0))2]=2Dt(1)\mathrm{MSD} = \mathbb{E}\left[(x(t)-x(0))^2\right] = 2Dt \tag*{(1)}

Here, E[⋅]\mathbb{E}[\cdot] denotes the statistical average, x(t)x(t) is the particle’s position at time tt, and DD is the diffusion coefficient. Einstein further related DD to the thermal energy and fluid properties via the Stokes-Einstein relation: D=kBTμD = k_B T\mu, where kBk_B is Boltzmann’s constant, TT is the temperature, and μ\mu is the particle’s mobility (the inverse of the friction coefficient, γ=μ−1\gamma= \mu^{-1}). For a spherical particle of radius aa in a fluid with viscosity η\eta, Stokes’ law gives the friction coefficient as γ=6πηa\gamma= 6\pi\eta a, so μ=(6πηa)−1\mu= (6\pi\eta a)^{-1}. Einstein’s theory provided a powerful connection between the microscopic random walk and macroscopic thermodynamics, but precise experimental verification was initially lacking.

Independently, in 1906, Marian Smoluchowski approached the problem from a different statistical perspective, also deriving results consistent with a diffusive process, although some initial calculations yielded slightly different numerical factors compared to Einstein’s specific predictions [47]. These theoretical advances spurred the need for careful experimental tests.

The decisive experimental confirmation came from the meticulous work of Jean Perrin around 1909 [57]. His experiments validated Einstein’s MSD relation Eq. (1) with remarkable accuracy. Furthermore, by measuring the MSD and knowing the other parameters, Perrin was able to determine Avogadro’s number NAN_A (kB=R/NAk_B = R/N_A, where RR is the gas constant), providing compelling quantitative evidence for the existence of atoms and molecules. This groundbreaking work earned Jean Perrin the Nobel Prize in Physics in 1926.

Amidst these developments, in 1908, Paul Langevin proposed a conceptually different, strikingly direct approach to derive Einstein’s result [48]. Instead of focusing solely on statistical descriptions, Langevin applied Newton’s second law (F=maF = ma) directly to the Brownian particle. His crucial insight was to model the total force on the particle as the sum of two components: a systematic frictional drag force opposing the particle’s velocity, and a rapidly fluctuating random force representing the incessant, unbalanced impacts from the surrounding fluid molecules:

md2xdt2=−γdxdt⏟Frictional force+Frandom(t)⏟Random force(2)m\frac{\mathrm{d}^{2}x}{\mathrm{d}t^{2}} = \underbrace{-\gamma\frac{\mathrm{d}x}{\mathrm{d}t}}_{\text{Frictional force}} + \underbrace{F_{\mathrm{random}}(t)}_{\text{Random force}} \tag*{(2)}

Here, mm is the particle’s mass, and γ=μ−1\gamma= \mu^{-1} is the friction coefficient. Langevin characterized Frandom(t)F_{\mathrm{random}}(t) as being "sometimes positive, sometimes negative" and having a magnitude sufficient to counteract the damping effect of friction, thus maintaining the particle’s thermal agitation [48]. He argued that because the molecular impacts are extremely rapid and random compared to the particle’s velocity changes, this random force should be statistically uncorrelated with the particle’s position x(t)x(t). Hence, his method relied on the assumption that the expectation of the product of the particle’s velocity and this random force is zero. By multiplying Eq. (2) by xx, using integration by parts, averaging, and invoking the equipartition theorem of statistical mechanics (E[12mv2]=12kBT\mathbb{E}\left[\frac{1}{2}mv^{2}\right] = \frac{1}{2}k_B T for thermal equilibrium in one dimension), Langevin elegantly recovered Einstein’s relation Eq. (1).

To achieve this consistency, the statistical properties of Frandom(t)F_{\mathrm{random}}(t) must be precisely linked to the friction γ\gamma and temperature TT. This requirement dictates the magnitude of the random force. It is common today to express the random force in terms of a standardized noise term ξt\xi_t, leading to the modern form of the Langevin equation:

md2xdt2=−γdxdt+γ2Dξt(3)m\frac{\mathrm{d}^{2}x}{\mathrm{d}t^{2}} = -\gamma\frac{\mathrm{d}x}{\mathrm{d}t} + \gamma\sqrt{2D}\xi_t \tag*{(3)}

Langevin’s work was pioneering in its explicit inclusion of a stochastic forcing term within a dynamical equation based on physical laws. Consequently, differential equations featuring such random terms and derived from physical principles are often referred to as Langevin equations in physics. From a mathematical perspective, they fall under the broader category of Stochastic Differential Equations (SDEs).

While Langevin’s physical intuition about Frandom(t)F_{\mathrm{random}}(t) was powerful, its precise mathematical nature remained undefined. What mathematical properties must this "random force" possess to be consistent with observations? This question motivates the need for a more formal mathematical description. Thus, we will first look at a mathematical definition motivated by physical arguments.

Mathematical Formalism for the Physicists’ Random Term ξt\xi_t

We now delve deeper into the properties required of the normalized random force ξt\xi_t, guided by physical considerations and the known behavior of Brownian motion, particularly: (1) its prominence for microscopic particles where fluid friction is significant, and (2) the linear-in-time growth of the MSD Eq. (1).

From (1), a crucial simplification arises, which is called the overdamped limit, relevant for many systems like colloidal particles in liquids. Here, viscous damping forces (∝γ∝a\propto\gamma\propto a) dominate inertial forces (m∝a3m \propto a^3) because the particle’s mass is very small. On the characteristic timescale of velocity relaxation (τv=m/γ\tau_v = m/\gamma), the acceleration term md2xdt2m\frac{d^2x}{dt^2} becomes negligible compared to the frictional term −γdxdt-\gamma\frac{dx}{dt}.

Setting the inertial term to zero in Eq. (3) yields the overdamped Langevin equation:

γdxdt=2γkBTξt⟹dxdt=2(γ−1)kBTξt=2μkBTξt=2Dξt(4)\gamma\frac{dx}{dt} = \sqrt{2\gamma k_{\mathrm{B}}T}\xi_t \Longrightarrow\frac{dx}{dt} = \sqrt{2\left(\gamma^{-1}\right)k_{\mathrm{B}}T}\xi_t = \sqrt{2\mu k_{\mathrm{B}}T}\xi_t = \sqrt{2D}\xi_t \tag*{(4)}

This can be written in differential form as:

dx=2Dξt dt(5)dx = \sqrt{2D}\xi_t\,dt \tag*{(5)}

In this picture, ξt\xi_t represents the effective, normalized random impulse delivered by the countless fluid molecule collisions over an infinitesimal time interval dtdt. Because this net impulse results from a vast number of quasi-independent molecular impacts, the Central Limit Theorem suggests that its integral over any finite time interval should be Gaussian distributed. Furthermore, by symmetry (no preferred direction for the net impacts), the average force must be zero:

E[ξt]=0(6)\mathbb{E}[\xi_t] = 0 \tag*{(6)}

Now consider the temporal correlations. The molecular collisions at time tt are essentially independent of those at a different time t′t', provided ∣t−t′∣\lvert t-t'\rvert is larger than the extremely short duration of molecular collision events. This strongly suggests that the random force has a very short memory, ideally zero for t≠t′t \ne t'. To quantify this, we examine what correlation structure is needed to recover Einstein’s MSD relation from the overdamped equation Eq. (5). Integrating Eq. (5) gives x(t)−x(0)=∫0t2Dξs dsx(t)-x(0)=\int_0^t\sqrt{2D}\xi_s\,ds. Squaring and taking the expectation yields:

E[(x(t)−x(0))2]=2D∫0t∫0tE[ξsξs′] ds ds′(7)\mathbb{E}\left[\left(x(t)-x(0)\right)^2\right] = 2D\int_0^t\int_0^t \mathbb{E}\left[\xi_s\xi_{s'}\right]\,ds\,ds' \tag*{(7)}

Comparing this with the target result E[(x(t)−x(0))2]=2Dt\mathbb{E}\left[\left(x(t)-x(0)\right)^2\right]=2Dt, we require the double integral to equal tt:

∫0t∫0tE[ξsξs′] ds ds′=t(8)\int_{0}^{t}\int_{0}^{t} \mathbb{E}[\xi_s \xi_{s'}]\,\mathrm{d}s\,\mathrm{d}s' = t \tag*{(8)}

The simplest mathematical object satisfying both the requirement of zero correlation for s≠s′s \ne s' and this integral identity (Eq. (8)) is the Dirac delta function:

E[ξsξs′]=δ(s−s′)(9)\mathbb{E}[\xi_s \xi_{s'}] = \delta(s-s') \tag*{(9)}

This delta correlation precisely yields ∫0t∫0tδ(s−s′) ds ds′=∫0tds=t\int_{0}^{t}\int_{0}^{t}\delta(s-s')\,\mathrm{d}s\,\mathrm{d}s' = \int_{0}^{t}\mathrm{d}s = t. Thus, the physicist’s operational picture of ξt\xi_t is that of Gaussian white noise: a stationary stochastic process characterized by a Gaussian distribution at any time tt, zero mean (E[ξt]=0\mathbb{E}[\xi_t] = 0), and a delta-function autocorrelation (E[ξsξs′]=δ(s−s′)\mathbb{E}[\xi_s \xi_{s'}] = \delta(s-s')).

While intuitive and useful, this description presents profound mathematical difficulties. A process with delta-correlated noise intrinsically has infinite variance (E[ξt2]→∞\mathbb{E}[\xi_t^2] \to\infty as suggested by δ(0)\delta(0)), which is physically unrealistic and mathematically problematic. Furthermore, the trajectory x(t)x(t) resulting from Eq. (5), while continuous, turns out to be nowhere differentiable. This lack of a well-defined derivative dx/dt\mathrm{d}x/\mathrm{d}t echoes Jean Perrin’s own remarks on the seemingly "pathological" nature of the experimentally observed Brownian paths1. Interpreting equations like Eq. (5) requires a more sophisticated mathematical framework: the Wiener process.

A Rigorous Mathematical Formulation: The Wiener Process

While the physical description of Langevin captured the essence of Brownian motion, and Perrin provided compelling experimental evidence, the mathematical nature of the "random force" or "white noise" remained problematic from a rigorous standpoint. The concept of a function that is continuous everywhere but differentiable nowhere challenged classical calculus. Furthermore, defining integrals involving the highly irregular white noise required new mathematical tools. Norbert Wiener, working in the early 1920s, sought to address these challenges. His goal was to construct a rigorous mathematical model for such paths. This endeavor was pivotal not only for understanding Brownian motion itself but also for laying the foundations of the modern theory of stochastic processes and stochastic calculus.

Central to Wiener’s framework is the Wiener process (or Brownian motion), denoted here as WtW_t, which is defined by the following properties:

  1. W0=0W_0 = 0 (starts at the origin almost surely).

  1. Independent increments: For any time sequence 0≤t0<t1<⋯<tn0 \le t_0 < t_1 < \cdots< t_n, the random variables (increments) Wt1−Wt0,Wt2−Wt1,…,Wtn−Wtn−1W_{t_1} - W_{t_0}, W_{t_2} - W_{t_1}, \ldots, W_{t_n} - W_{t_{n-1}} are mutually independent.

  1. Gaussian increments: For any 0≤t0<t10 \le t_0 < t_1, the increment Wt1−Wt0W_{t_1} - W_{t_0} follows a normal (Gaussian) distribution with mean 0 and variance t1−t0t_1 - t_0, denoted as N(0,t1−t0)\mathcal{N}(0,t_1-t_0).

  1. Continuous paths: The function t↦Wtt \mapsto W_t is continuous in tt (almost surely).

We use WtW_t to represent the mathematical construct, distinguishing it from the physical phenomenon of Brownian motion itself.

Using the Wiener process, the overdamped Langevin equation Eq. (5) is rigorously rewritten in differential form as:

dxt=2D dWt(10)\mathrm{d}x_t = \sqrt{2D}\,\mathrm{d}W_t \tag*{(10)}

References

  1. [138]138M. G. Usai, M. E. Goddard, and B. J. Hayes, “LASSO with cross-validation for genomic selection”, Genetics Research 91, 427–436 (2009) (cit. on p. 124).DOI
  2. [139]139S. Bhattacharya and P. D. McNicholas, “A LASSO-penalized BIC for mixture model selection”, Advances in Data Analysis and Classification 8, 45–61 (2014) (cit. on p. 124).DOI
  3. [140]140Y. Ninomiya and S. Kawano, “AIC for the Lasso in generalized linear models”, Electronic Journal of Statistics 10, 2537–2560 (2016) (cit. on p. 124).DOI
  4. [141]141“Incomplete gamma function”, Wikipedia (2025) (cit. on p. 134).
  5. [142]142J. Chen and Z. Chen, “Extended Bayesian Information Criteria for Model Selection with Large Model Spaces”, Biometrika 95, 759–771 (2008) (cit. on p. 138).DOI
  6. [143]143C. Soussen, J. Idier, D. Brie, and J. Duan, “From Bernoulli–Gaussian Deconvolution to Sparse Signal Restoration”, IEEE Transactions on Signal Processing 59, 4572–4584 (2011) (cit. on p. 140).DOI
  7. [144]144M. Ragwitz and H. Kantz, “Indispensable Finite Time Corrections for Fokker-Planck Equations from Time Series Data”, Physical Review Letters 87, 254501 (2001) (cit. on p. 158).
  8. [145]145F. Morcos, A. Pagnani, B. Lunt, A. Bertolino, D. S. Marks, C. Sander, R. Zecchina, J. N. Onuchic, T. Hwa, and M. Weigt, “Direct-coupling analysis of residue coevolution captures native contacts across many protein families”, Proceedings of the National Academy of Sciences 108, E1293–E1301 (2011) (cit. on p. 169).DOI
  9. [146]146A.-F. Bitbol, R. S. Dwyer, L. J. Colwell, and N. S. Wingreen, “Inferring interaction partners from protein sequences”, Proceedings of the National Academy of Sciences 113, 12180–12185 (2016) (cit. on p. 169).DOI
  10. [147]147L. Rosset, R. Netti, A. P. Muntoni, M. Weigt, and F. Zamponi, adabmDCA 2.0 – a flexible but easy-to-use package for Direct Coupling Analysis, Feb. 2025 (cit. on p. 169).DOI
  11. [148]148S. S. Mukherjee, Consistent model selection in the spiked Wigner model via AIC-type criteria, Feb. 2025 (cit. on p. 169).
  12. [149]149J. Barbier, M. Dia, N. Macris, F. Krzakala, T. Lesieur, and L. Zdeborová, “Mutual information for symmetric rank-one matrix estimation: A proof of the replica formula”, in Advances in Neural Information Processing Systems, Vol. 29 (2016) (cit. on p. 169).arxiv.org/abs/1606.04142

Paper details

vlnc:W_dbrejdpg
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Generative artificial intelligence (AI) tools

View statement
Looking for related papers…
Contents