betaFeedback

WorksMathematics

The microstructure of information through mean field learning interaction

Abstract

We study how knowledge about partially observable model parameters is being created and propagated in large scale systems. Information is generated from observational signals whose dynamics depend endogenously on the learning history, the learning duration, the decisions and the related costs encountered by a continuum of decision makers. Such characteristics give rise to a new class of sequential testing problems in which there is continuous endogenous feedback. Adopting the classical approach of Shiryaev, we relate the problem to an optimal stopping one for the representative belief process, whose generator now depends on the evolving interactions. Considering a fixed point for the latter, we introduce a notion of mean field game and prove the existence of an equilibrium. Our arguments, which are based on a combination of pde and probabilistic techniques, involve a detailed analysis of the underlying variational inequality and the geometry, stability and continuity of the free boundaries.

Introduction

This paper studies how interactive and collective information acquisition about unknown model parameters generates and shapes aggregate learning about them in large-scale systems. Information is dynamically created in a Bayesian framework by costly signals which generate belief processes (posterior, conditional on the signals, distributions). The key feature in our work is that the stochastic dynamics of the signals are directly affected by the collective information-seeking behavior, like duration of learning and assessment about the unknown parameter and decisions when learning terminates. Acquiring information induces a cost consisting of a running component while signals are being used and a final aggregate cost caused by these possibly erroneous assessments and decisions at exit. The focus is on how to (i) minimize the representative information acquisition cost, (ii) investigate through its solution how representative knowledge is being created and shaped by a continuous interaction of signals, (iii) examine its evolution and (iv) build an equilibrium model for its validity. Our results spread across the areas of hypothesis testing, Bayesian learning, filtering, optimal stopping, and mean field games with partial observation.

The motivation to study large-scale models of interactive learning comes from numerous applications in marketing, advertising, trading, autonomous driving systems, and supply-chain systems, to name a few. Indeed, in e-commerce platforms, knowledge about the quality of a new product is generated from reviews and usage data induced by earlier buyers; see Kakhbod, Lanzani and Xing [22]. Because some consumers provide detailed feedback while others leave only short comments, the informativeness of the public, collective signals depends considerably on how many and who has chosen to adopt the new product so far. Indeed, if early buyers provide meaningful, detailed feedback, then the signal about the quality of the product is strong. On the other hand, if many early buyers provide only sparse or uninformative feedback, the collective signal is inferior. As a result, the quality of acquired information is being shaped directly by the previous adoption decisions, the population size of this group and the duration of their learning.

In financial markets with liquidity effects, if asset liquidation becomes more intense, the overall market depth deteriorates, making price movements more volatile and less informative about the underlying price fundamentals. As a consequence, the ability to extract information from observed prices depends critically on how many positions are already closed and by when this actually occurred; see, for example, Casgrain and Jaimungal [11]. The resulting knowledge acquisition problem is thus affected critically by collective learning interaction, as the value of waiting to gather further information and the quality of information itself are determined by the evolving distribution of decisions and learning duration.

The case of a single decision maker, who learns about a binary random variable Θ\Theta with possible values 0 and 1, has been extensively analyzed in the literature, starting with the work of Wald in [33] on costly sequential learning in Bayesian statistics. Since then, it has a long history of development and constitutes a central area in statistical learning and hypothesis testing, with hypotheses being H0:Θ=0H_0 : \Theta= 0 and H1:Θ=1H_1 : \Theta= 1, for this binary case. The underlying problem is to minimize, up to the time that learning stops, the total expected cost consisting of a running part due to signal usage and a terminal component. The latter is generated by erroneous assessments about the values of Θ\Theta at exit time and related decisions. Thus, the controls are formulated in terms of the pair of stopping times and decisions at them, (τ(ω);di(τ(ω)))(\tau(\omega); d_i(\tau(\omega))) for i=0,1i = 0, 1.

Shiryaev developed in [31] a connection to an optimal stopping problem for the belief process. In particular, it was shown in [31] that the underlying free boundary problem has two continuous and not intersecting boundaries which completely characterize the optimal learning. Specifically, the boundaries divide the space in an information seeking region and two idle regions to its left and right sides. If the belief process starts, at t=0t = 0 in the left (resp. right) region, no learning occurs and hypothesis H0:Θ=0H_0 : \Theta= 0 (resp. H1:Θ=1H_1 : \Theta= 1) is instantly accepted. The cost then comes from the instant decision d0(0)d_0(0) (resp. d1(0)d_1(0)). If, on the other hand, the belief process starts in the middle region, learning takes place till the first hitting time to either boundary. If the left boundary (resp. right) is reached first at τ∗(ω)\tau^*(\omega), hypothesis H0:Θ=0H_0 : \Theta= 0 (resp. H1:Θ=1H_1 : \Theta= 1) is accepted and decision d0(τ∗(ω))d_0(\tau^*(\omega)) (resp. d1(τ∗(ω))d_1(\tau^*(\omega)) is made. The total cost is the one incurred by the latter and the running learning cost in [0,τ∗(ω)][0,\tau^*(\omega)].

The results of [31] initiated a very rich and active line of research, and providing a complete bibliography is beyond the scope herein (see, among others, Moscarini and Smith [28], Miao and Xing [27], Ke, Tang, Villas-Boas and Zhang [24]). We also mention the works of Ekström, Karatzas and Vaicenavicius [17], Ekström and Vaicenavicius [18], Ekström and Wang [19] and Campbell, Gaitsgori, Groenewald and Karatzas [7].

Herein, we build a mean field game of interactive and collective information acquisition, which we study using the technical approach of [31]. We assume that information is being acquired and propagated through a continuous interaction within a population continuum which uses signals with dynamics directly affected by the aggregate learning, decisions and costs. This continuum starts with a distribution of prior probabilities for the values of Θ\Theta and interacts through time till learning stops. The key quantity is the temporal aggregate of the fraction of the population that has completed learning, selected a decision, encountered the relevant costs and exited. This temporal aggregate alters the learning mechanism of the remaining population through the drift of its, continuously updated, signal. In turn, it affects the representative value function which models the minimal total expected informational cost and constitutes the best response to the effects of realized learning and costs. We are then motivated to introduce a fixed-point notion in terms of this aggregate feedback which leads to a mean field game.

We stress that contrary to the standard mean field games where players are entirely identified by their controlled state processes, herein the notion of a player is identified with the individual prior probability and subsequent belief process. In other words, there is no representative controlled diffusion as such but rather a representative belief process whose dynamics are affected by the learning and decisions of the peer continuum. To our knowledge, the only work close to ours is the preprint of Campbell and Zhang [8], but both their technical assumptions and modeling elements are different than ours. We provide further details at the end of the introduction.

The model and the results. In what follows we explain our main results and refer for the precise definitions and assumptions later in the paper.

Let (Ω,F,P)(\Omega,\mathcal{F},\mathbb{P}) be a probability space that supports

(i) a binary random variable Θ∈{0,1}\Theta\in\{0,1\},

(ii) a random variable P0∈[0,1]P_0 \in[0,1] with law η\eta having a continuous density representing the initial belief (prior) that P(Θ=1∣P0)=P0\mathbb{P}(\Theta=1 \mid P_0)=P_0,

(iii) a standard Brownian motion WW.

(i) a binary random variable Θ∈{0,1},(ii) a random variable P0∈[0,1] with law η having a continuous densityrepresenting the initial belief (prior) that P(Θ=1∣P0)=P0,(iii) a standard Brownian motion W.(1)\begin{aligned} &\text{(i) a binary random variable }\Theta\in\{0,1\},\\ &\text{(ii) a random variable }P_0 \in[0,1]\text{ with law }\eta\text{ having a continuous density}\\ &\qquad\text{representing the initial belief (prior) that }\mathbb{P}(\Theta=1\mid P_0)=P_0,\\ &\text{(iii) a standard Brownian motion }W. \tag*{(1)} \end{aligned}

It is assumed that

W is independent of (Θ,P0).(2)W\text{ is independent of }(\Theta,P_0). \tag*{(2)}

We study information acquisition about the unknown binary state Θ\Theta from continuous time signals in a large-scale system. Knowledge is being created from a representative observational signal YμY^\mu whose dynamics depend on aggregate learning interaction in the continuum. Specifically, YμY^\mu is taken to solve

dYtμ=b(t,Θ;F(t;μ≤t)) dt+dWtfor t∈[0,T] and Y0μ=0.(3)\mathrm{d}Y_t^\mu=b(t,\Theta;F(t;\mu_{\leq t}))\,\mathrm{d}t+\mathrm{d}W_t\quad\text{for }t\in[0,T]\text{ and }Y_0^\mu=0. \tag*{(3)}

The term F(t;μ≤t)F(t;\mu_{\leq t}), which is defined in (16), models the collective information feedback and incorporates the learning duration, exit times, and decisions about the unknown values and related costs up to time tt. The input μ\mu is a pair (μ0,μ1)(\mu^0,\mu^1) of two nonnegative subadditive measures representing stopping times and decisions at them, given, for each Borel subset AA of [0,T][0,T] and i=0,1i=0,1, by

μi(A)=P(τ∈A,d=i),\mu^i(A)=\mathbb{P}(\tau\in A,d=i),

and such that μ0+μ1\mu^0+\mu^1 is a probability measure on [0,T][0,T]. In what follows, we denote the set of such pairs by M\mathcal{M}, that is,

M:={(μ0,μ1):μ0,μ1∈M+([0,T]), μ0+μ1∈P([0,T])},\mathcal{M}:=\left\{(\mu^0,\mu^1):\mu^0,\mu^1\in\mathcal{M}_+([0,T]),\ \mu^0+\mu^1\in\mathcal{P}([0,T])\right\},

where M+([0,T])\mathcal{M}_{+}([0,T]) stands for the set of nonnegative subadditive measures on [0,T][0,T].

For a fixed input μ\mu, the observation signal YμY^\mu generates the representative posterior belief process PP given by

Pt:=P(Θ=1∣Ftμ) for t∈[0,T] and P0=p∈[0,1],P_t := \mathbb{P}(\Theta= 1 \mid\mathcal{F}^{\mu}_t) \text{ for } t \in[0,T] \text{ and } P_0 = p \in[0,1],

where Fμ=(Ftμ)0≤t≤T\mathcal{F}^{\mu} = (\mathcal{F}^{\mu}_t)_{0\leq t\leq T} denotes the usual augmentation of the filtration (σ(P0,Ysμ):0≤s≤t))0≤t≤T(\sigma(P_0,Y^\mu_s):0\leq s\leq t))_{0\leq t\leq T}.

Classical results in filtering yield that the process PP solves the sde

dPtμ=Δbμ(t)Ptμ(1−Ptμ) dZtμfor t∈(0,T] and P0=p∈[0,1],(4)\mathrm{d}P^\mu_t = \Delta b^\mu(t)P^\mu_t(1-P^\mu_t)\,\mathrm{d}Z^\mu_t \quad\text{for } t \in(0,T] \text{ and } P_0 = p \in[0,1], \tag*{(4)}

where

Δbμ(t):=Δb(t;F(t;μ≤t))=b(t,1;F(t;μ≤t))−b(t,0;F(t;μ≤t))(5)\Delta b^\mu(t) := \Delta b(t;F(t;\mu_{\leq t})) = b(t,1;F(t;\mu_{\leq t})) - b(t,0;F(t;\mu_{\leq t})) \tag*{(5)}

is the difference between the signal drifts under the two states 0 and 1, and ZμZ^\mu is a Fμ\mathcal{F}^\mu-Brownian motion referred to as the innovation process; see, for example, Bain and Crisan [1].

We assume (see (17) and (18)) that bb and Δb\Delta b satisfy the assumptions that are needed for (3) and (4) to have unique strong solutions.

The representative information acquisition problem is an optimization one over all stopping learning times and decisions related to parameter value assessment.

Following [31], we formulate the original problem as an optimal stopping problem for the representative belief process over all stopping learning times.

Specifically, we introduce the value function

Vμ(0,p)=inf⁡τ∈T0,TμE[∫0τfμ(s,Psμ) ds+G(Pτμ):P0μ=p](6)V^\mu(0,p) = \inf_{\tau\in\mathcal{T}^\mu_{0,T}} \mathbb{E}\left[\int_0^\tau f^\mu(s,P^\mu_s)\,\mathrm{d}s + G(P^\mu_\tau):P^\mu_0=p\right] \tag*{(6)}

where, for t∈[0,T]t\in[0,T], Tt,Tμ\mathcal{T}^\mu_{t,T} denotes the set of [t,T][t,T]-valued Fμ\mathbb{F}^\mu-stopping times, fμ:[0,T]×[0,1]→Rf^\mu:[0,T]\times[0,1]\to\mathbb{R} is the running cost during learning given by

fμ(t,p):=f(t,p;F(t;μ≤t)),(7)f^\mu(t,p):=f(t,p;F(t;\mu_{\leq t})), \tag*{(7)}

with ff satisfying (20), and G:[0,1]→RG:[0,1]\to\mathbb{R} is the penalty incurred when learning stops, taken to be

G(p):=min⁡i=0,1G‾(p,i).G(p):=\min_{i=0,1}\overline{G}(p,i).

Following [31] and subsequent papers, we assume that, for a0,a1>0a_0,a_1>0,

G‾(p,0)=a0pandG‾(p,1)=a1(1−p).\overline{G}(p,0)=a_0p \quad\text{and} \quad\overline{G}(p,1)=a_1(1-p).

With this choice of G‾\overline{G}, the terminal cost GG can be written as

G(p)=min⁡{a0p,a1(1−p)}.(8)G(p)=\min\{a_0p,a_1(1-p)\}. \tag*{(8)}

The maximum in (8) is attained at

m=a1a0+a1.(9)m=\frac{a_1}{a_0+a_1}. \tag*{(9)}

Then, we can rewrite GG as

G(p)=G‾(p,D(p)),G(p)=\overline{G}(p,D(p)),

where D:[0,1]→{0,1}D:[0,1]\to\{0,1\} is the so-called decision rule, with the tie at mm assigned to decision 0,

D(p):={0if p≤m,1if p>m.D(p):= \begin{cases} 0 & \text{if }p\le m,\\ 1 & \text{if }p>m. \end{cases}

References

  1. [1]Alan Bain and Dan Crisan. Fundamentals of Stochastic Filtering. Springer, New York, 2009.
  2. [2]Charles Bertucci. Optimal stopping in mean field games, an obstacle problem approach. J. Math. Pures Appl., 120:165–194, 2018.arxiv.org/abs/1704.06553
  3. [3]Adrien Blanchet. On the regularity of the free boundary in the parabolic obstacle problem. application to American options. Nonlinear Analysis, 65(7):1362–1378, 2006.DOI
  4. [4]Adrien Blanchet, Jean Dolbeault, and Régis Monneau. On the continuity of the time derivative of the solution to the parabolic obstacle problem with variable coefficients. Journal de Mathématiques Pures et Appliquées, 85(3):371–414, 2006.
  5. [5]Adrien Blanchet, Jean Dolbeault, and Régis Monneau. Erratum to “on the continuity of the time derivative of the solution to the parabolic obstacle problem with variable coefficients”. Journal de Mathématiques Pures et Appliquées, 94(4):447–449, 2010.
  6. [6]Géraldine Bouveret, Roxana Dumitrescu, and Peter Tankov. Mean-field games of optimal stopping: a relaxed solution approach. SIAM J. Control Optim., 58(4):1795–1821, 2020.arxiv.org/abs/1812.06196
  7. [7]Steven Campbell, Georgy Gaitsgori, Richard Groenewald, and Ioannis Karatzas. Grab it before it’s gone: testing uncertain rewards under a stochastic deadline. Stochastic Process. Appl., 201:Paper No. 105069, 32, 2026.arxiv.org/abs/2503.06856
  8. [8]Steven Campbell and Yuchong Zhang. A mean field game of sequential testing. arXiv preprint arXiv:2403.18297, 2024.arxiv.org/abs/2403.18297
  9. [9]Pierre Cardaliaguet, Joe Jackson, and Panagiotis E Souganidis. Mean field control with stopping. arXiv preprint arXiv:2603.21204, 2026.arxiv.org/abs/2603.21204
  10. [10]René Carmona, François Delarue, and Daniel Lacker. Mean field games of timing and models for bank runs. Appl. Math. Optim., 76(1):217–260, 2017.arxiv.org/abs/1606.03709
  11. [11]Philippe Casgrain and Sebastian Jaimungal. Mean-field games with differing beliefs for algorithmic trading. Mathematical Finance, 30(3):995–1034, 2020.arxiv.org/abs/1810.06101
  12. [12]Andrea Cosso and Laura Perelli. Mean field optimal stopping with uncontrolled state. arXiv preprint arXiv:2503.04269, 2025.arxiv.org/abs/2503.04269
  13. [13]Michael G Crandall, Hitoshi Ishii, and Pierre-Louis Lions. User’s guide to viscosity solutions of second order partial differential equations. Bulletin of the American mathematical society, 27(1):1–67, 1992.
  14. [14]Tiziano De Angelis and Gabriele Stabile. On lipschitz continuous optimal stopping boundaries. SIAM Journal on Control and Optimization, 57(1):402–436, 2019.
  15. [15]Roxana Dumitrescu, Marcos Leutscher, and Peter Tankov. Control and optimal stopping mean field games: a linear programming approach. Electronic Journal of Probability, 26:1–49, 2021.arxiv.org/abs/2011.11533
  16. [16]Roxana Dumitrescu, Marcos Leutscher, and Peter Tankov. Linear programming fictitious play algorithm for mean field games with optimal stopping and absorption. ESAIM Math. Model. Numer. Anal., 57(2):953–990, 2023.arxiv.org/abs/2202.11428
  17. [17]Erik Ekström, Ioannis Karatzas, and Juozas Vaicenavicius. Bayesian sequential least-squares estimation for the drift of a Wiener process. Stochastic Process. Appl., 145:335–352, 2022.
  18. [18]Erik Ekström and Juozas Vaicenavicius. Bayesian sequential testing of the drift of a Brownian motion. ESAIM Probab. Stat., 19:626–648, 2015.
  19. [19]Erik Ekström and Yuqiong Wang. Stopping problems with an unknown state. J. Appl. Probab., 61(2):515–528, 2024.DOI
  20. [20]Wendell H. Fleming and H. Mete Soner. Controlled Markov Processes and Viscosity Solutions. Springer, New York, second edition, 2006.DOI
  21. [21]Xihao He, Xiaolu Tan, and Jun Zou. A mean-field version of bank-el karoui’s representation of stochastic processes. arXiv preprint arXiv:2302.03300, 2023.arxiv.org/abs/2302.03300
  22. [22]Ali Kakhbod, Giacomo Lanzani, and Hao Xing. Heterogeneous learning in product markets. 2021.
  23. [23]Ioannis Karatzas and Steven E. Shreve. Brownian Motion and Stochastic Calculus. Springer, New York, 2 edition, 1991.
  24. [24]T Tony Ke, Wenpin Tang, J Miguel Villas-Boas, and Yuming Paul Zhang. Parallel search for information in continuous time–optimal stopping and geometry of the pde. Applied Mathematics & Optimization, 85(2):3, 2022.
  25. [25]N. V. Krylov. One more square root law for Brownian motion and its application to SPDEs. Probability Theory and Related Fields, 127(4):496–512, 2003.DOI
  26. [26]Pascal Massart. The tight constant in the Dvoretzky–Kiefer–Wolfowitz inequality. The Annals of Probability, 18(3):1269–1283, 1990.DOI
  27. [27]Jianjun Miao and Hao Xing. Dynamic discrete choice under rational inattention. Economic Theory, 77(3):597–652, 2024.DOI
  28. [28]Giuseppe Moscarini and Lones Smith. The optimal level of experimentation. Econometrica, 69(6):1629–1644, 2001.
  29. [29]Marcel Nutz. A mean field game of optimal stopping. SIAM Journal on Control and Optimization, 56(2):1206–1221, 2018.arxiv.org/abs/1605.09112
  30. [30]Dylan Possamai and Mehdi Talbi. Mean-field games of optimal stopping: master equation and weak equilibria. arXiv preprint arXiv:2307.09278, 2023.arxiv.org/abs/2307.09278
  31. [31]Albert N Shiryaev. Optimal stopping rules, volume 8. Springer Science & Business Media, 2007.
  32. [32]Mehdi Talbi, Nizar Touzi, and Jianfeng Zhang. Dynamic programming equation for the mean field optimal stopping problem. SIAM Journal on Control and Optimization, 61(4):2140–2164, 2023.arxiv.org/abs/2103.05736
  33. [33]Abraham Wald. Foundations of a general theory of sequential decision functions. Econometrica, Journal of the Econometric Society, pages 279–313, 1947.DOI

Paper details

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

views · 10

downloads · 0

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

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents