Similarity of Information in Games
Deepal Basak Indiana University Joyee Deb New York University Aditya Kuvalekar University of Essex · 2026
Read Similarity of Information in Games by Deepal Basak Indiana University Joyee Deb New York University Aditya Kuvalekar University of Essex (2026) 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-2603-28190.txt.
Introduction
Basak, Deb, and Kuvalekar argue that common similarity measures miss what matters when people act strategically on shared information. Targeted feeds make information more alike across agents, as in the Silicon Valley Bank run. Existing measures ignore how agents update beliefs about each other. The authors define a family of orders, Concentration Along the Diagonal, built on those beliefs. In coordination games, more similarity in this sense exactly matches agents choosing the same strategy.
Catalog record
| CREATOR | Deepal Basak Indiana University Joyee Deb New York University Aditya Kuvalekar University of Essex |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 135109 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.28190
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Similarity of Information in Games This research was supported in part by National Science Foundation award 2417694. Kuvalekar thanks the Indian School of Business for their hospitality. We are grateful to Nageeb Ali, Dirk Bergemann, Geoffroy De Clippel, Duarte Goncalves, Nima Haghpanah, Johannes Hörner, Elliot Lipnowski, Michael Mandler, Meg Meyer, Dilip Mookherjee, Stephen Morris, Alessandro Pavan, Ludvig Sinander, Bruno Strulovici, and conference audiences and seminar participants at Bonn University, Brown University, Carnegie Mellon University, Cornell University, Indian Statistical Institute Delhi, National University of Singapore, Ohio State University, Oxford University, Paris School of Economics, Royal Holloway University, Toulouse School of Economics, University of Rochester, University of Surrey, Duke & UNC, and Penn State for their helpful comments and suggestions. Basak: Indiana University, Kelley School of Business, email: [email protected]; Deb: New York University, email: [email protected], Kuvalekar: University of Essex, email: [email protected]. Deepal Basak Indiana University Joyee Deb New York University Aditya Kuvalekar University of Essex Abstract Algorithmic content targeting homogenizes information, with implications for strategic interactions. For example, this increased homogenization was arguably responsible for the run on the Silicon Valley Bank. We argue that existing measures of similarity are inappropriate for studying games—especially coordination games—because they do not discipline agents’ conditional beliefs. We propose a class of stochastic orders, Concentration Along the Diagonal (CAD), built on agents’ conditional beliefs. In canonical binary-action coordination games, greater CAD-similarity is both necessary and sufficient for strategic similarity—agents adopt the same strategy. We further demonstrate CAD’s applicability in congestion games, collective action, and second-price auctions. Keywords: coordination games, information similarity. JEL Codes: D82, D83 1 Introduction In recent times, algorithmic targeting of content has resulted in increased homogenization of information among like-minded individuals: People with similar demographic traits or preferences are exposed to the same content. 1 1 1 A large literature documents the decrease in diversity of content consumption across users as a result of recommendation algorithms. See, for example, 43 ; 2 ; 44 ; 9 ; 3 for evidence of homogenization in various contexts from online purchases to news consumption. This homogenization can affect agent behavior in strategic situations. For instance, both empirical research and news media have highlighted how access to the same information on Twitter encouraged many people to withdraw deposits, precipitating a run on Silicon Valley Bank in 2023. 2 2 2 See 1 , 2 , 3 , or 14 ; 21 on bank runs. The narrative is that people accessed the same information about the health of the bank and reacted to this information in the same way, causing a run which may not have been possible without the homogenization of information. Motivated by this discourse, we seek to systematically investigate how increased information similarity across agents affects their strategic interactions. We present a simple bank-run example that motivates the central question of our paper. This example demonstrates that increasing information similarity across agents in some natural ways—like increasing correlation—does not increase the expected number of people that would run on the bank! Motivated by this observation, we ask: What type of information homogenization encourages agents to follow the same strategy as others in the face of incomplete information in coordination games such as bank runs? Our main contribution is to answer this question by proposing a new class of stochastic orders—Concentration Along the Diagonal (CAD)—to compare similarity of information. Correlation and Coordination: A puzzle Two agents each decide whether to stay (S) or run (R) on a bank. The payoff from running is 0 . The payoff from staying is \theta if the other player stays, and \theta-1 if the other runs, where \theta is an underlying state, say the health of the bank, which is unknown: \theta\in\left\{\frac{3}{2},\frac{1}{2},-\frac{1}{2}\right\} . Stay Run Stay \theta,\theta \theta-1,0 Run 0,\theta-1 0,0 Notice that staying is dominant if \theta=\frac{3}{2} and running is dominant if \theta=-\frac{1}{2} . The state \theta is drawn from the prior P(\theta=\frac{3}{2})=P(\theta=-\frac{1}{2})=\epsilon and P(\theta=\frac{1}{2})=1-2\epsilon . Each agent i sees a private signal s_{i} about the realized \theta . The signals (s_{1},s_{2}) in state \theta are drawn from some joint distribution \mathcal{G}^{\theta} over \left\{-\frac{1}{2},\frac{1}{2},\frac{3}{2}\right\}\times\left\{-\frac{1}{2},\frac{1}{2},\frac{3}{2}\right\} . Assume that signals are sufficiently accurate so that R is strictly dominated for an agent with signal s_{i}=\frac{3}{2} and S is strictly dominated for an agent with signal s_{i}=-\frac{1}{2} . 3 3 3 A sufficient condition is P(s_{i}=\theta|\theta)=p and P(s_{i}=\theta^{\prime}|\theta)=\frac{1-p}{2} for \theta^{\prime}\neq\theta with p>\frac{3-3\epsilon}{3-\epsilon} . Consider the largest run on the bank possible in any symmetric pure strategy equilibrium. This would involve each agent i staying if and only if she gets a signal s_{i}=\frac{3}{2} . Essentially, the existence of such an equilibrium depends only on whether a player with a signal \frac{1}{2} wishes to run or not. This, in turn, solely depends on how likely it is, according to this player, that the other player receives a signal \frac{1}{2} or \frac{-1}{2} . Suppose we started with an information structure where such an equilibrium exists and make players’ signals more correlated. Intuitively, we should expect more correlated information to cause more people to run on the bank, because, conditional on their own private signal, each agent now thinks that it is more likely that others have also seen a similar signal and will run on the bank. For instance, at the extreme, if all the agents see exactly the same information, then for any observed signal, if all others are willing to run on the bank, and running undominated, then any individual will also run on the bank. Surprisingly, it is not true that more correlated information makes larger bank runs possible. The figure below shows two joint distributions \mathcal{G}^{\theta=\frac{1}{2}} and \mathcal{F}^{\theta=\frac{1}{2}} of signals in state \theta=\frac{1}{2} , where \mathcal{F}^{\theta=\frac{1}{2}} has a higher correlation. 4 4 4 For simplicity, assume that the signals are conditionally independent when \theta\in\{-\frac{1}{2},\frac{3}{2}\} . For example, the -\alpha and +\alpha in the top row mean that the probability mass on (-\frac{1}{2},\frac{3}{2}) under \mathcal{F}^{\theta=\frac{1}{2}} is \alpha less than under \mathcal{G}^{\theta=\frac{1}{2}} , and the probability mass on (\frac{1}{2},\frac{3}{2}) under \mathcal{F}^{\theta=\frac{1}{2}} is \alpha more than under \mathcal{G}^{\theta=\frac{1}{2}} . It is easy to verify that \mathcal{F} not only has higher correlation than \mathcal{G} , but also is greater than \mathcal{G} in the supermodular order, positive quadrant dependence order or concordance orders, for any positive and feasible \alpha . -\frac{1}{2} \frac{1}{2} \frac{3}{2} -\frac{1}{2} \frac{1}{2} \frac{3}{2} X_{1} X_{2} Distribution \mathcal{G}^{\theta=\frac{1}{2}} -\frac{1}{2} \frac{1}{2} \frac{3}{2} -\frac{1}{2} \frac{1}{2} \frac{3}{2} X_{1} X_{2} -\alpha -2\alpha +\alpha +\alpha +\alpha +\alpha -\alpha Distribution \mathcal{F}^{\theta=\frac{1}{2}} Figure 1 : Increasing interdependence according to existing orders ( \alpha>0 ). Notice that when the signals become more correlated in this way, the probability that a player with a signal \frac{1}{2} assigns to the other player receiving a signal \frac{-1}{2} or \frac{1}{2} —and therefore the other player running—decreases. Thus, we may no longer be able to support an equilibrium where players run on \frac{-1}{2} and \frac{1}{2} . In a nutshell, in this simple example, increasing correlation of information decreases the maximal (across all equilibria) expected number of people that would run on the bank! Why do we get this intuitively implausible result? The reason is that in a game of incomplete information, players’ incentive constraints—and hence their choice of actions—are affected by their conditional beliefs , and an increase in correlation (or other commonly used stochastic orders