Herding with Multiple States: When the Prior Matters

arXiv:2611.13632 · econ.TH · Submitted 2026-11-10 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.

Ruth: We're looking at the paper: "Herding with Multiple States: When the Prior Matters".

Warren: Here's the gist of the paper. In sequential observational learning with more than two states, it is possible for identical information structures and preferences to lead to either learning or herding depending on the initial full-support prior.

Ruth: The paper refutes a conjecture by Kartik, Lee, Liu, and Rappoport (2024a) by presenting a counterexample where the cascade region has a non-empty interior, yet learning occurs for generic full-support priors outside this region.

Warren: The authors establish two necessary conditions for learning: escapability and pairwise escapability.

Ruth: Escapability requires that signals can provide arbitrarily strong evidence that any given action is ex post suboptimal; pairwise escapability is a variant of this condition.

Paper summary: Ruth: So this paper, "Herding with Multiple States: When the Prior Matters," it looks at sequential observational learning where there are more than two states, and it says that even if everyone has the same information structure and preferences, whether people learn or just herd can depend on what initial full-support prior you start with.

Warren: What's interesting is that this work refutes a conjecture by Kartik, Lee, Liu, and Rappoport from 2024a because they show a counterexample where the cascade region has an interior but learning still happens for generic full-support priors outside that region <ref:2611.13632#pg1,Kartik, Lee, Liu, and Rappoport>.

Iris: That suggests that we need to look at two necessary conditions for learning: escapability and pairwise escapability. Escapability means signals can give arbitrarily strong evidence that any action is suboptimal after the fact, and pairwise escapability is a variation of that idea too.

Ruth: If excludability fails, she says herding happens for some initial prior; but if pairwise escapability fails, then herding happens for every full-support prior, according to Theorem one <ref:2611.13632#pg3>.

Warren: For environments with interval optimality and monotone likelihood ratios, these conditions completely describe learning and herding without any dependence on the initial full-support prior.

Iris: When you introduce location-shift signals into that setup, this characterization simplifies down to pairwise distinguishability, which extends unbounded beliefs over to the multiple state setting.

Ruth: The authors demonstrate that learning can actually depend on the initial prior even when we fix both preferences and information; specifically, learning happens for generic full-support priors outside a cascade belief set in Example two while herding occurs for priors that meet specific conditions on belief odds.

Warren: They characterize this dependence by showing that the failure of excludability implies herding for some prior with binary actions and finite states, but not necessarily for more than two actions.

Iris: Under the assumptions of interval optimality and monotone likelihood ratios, pairwise escapability is equivalent to pairwise distinguishability, which is Corollary one <ref:2611.13632#pg3>.

Ruth: Furthermore, in settings with location-shift information, pairwise escapability becomes equivalent to pairwise distinguishability as Proposition five shows that excludability, or its variant pairwise escapability, and pairwise distinguishability are all equivalent under these conditions.

Warren: The paper also shows that if interval optimality fails when you have monotone likelihood ratios information, then herding occurs for some initial prior according to Proposition six.

Iris: The final result states that with interval optimality and location-shift information satisfying monotone likelihood ratios, adequate learning exists if and only if both interval optimality and pairwise distinguishability hold according to Proposition six.

Ruth: The analysis uses the public belief process and its martingale property to characterize learning through limiting beliefs, which is a key part of how they set up their argument.

Warren: The main contribution here is providing sufficient conditions for herding or necessary conditions for learning that help differentiate between herding happening for some initial prior versus occurring for every full-support prior.

Iris: This characterization holds true independently of the full-support initial prior in a natural class of environments defined by interval optimality and monotone likelihood ratios.

Ruth: The paper concludes that in specific instances, learning can depend on the initial prior even when preferences and information are fixed, which highlights that identifying cascade beliefs and identifying herding are actually two different problems.

Warren: Ultimately, under Theorem two with interval optimality and monotone likelihood ratios, preferences and information determine whether learning occurs regardless of the full-support initial prior, with pairwise distinguishability serving as the characterization for location-shift information.

Conclusion: Ruth: So we're wrapping up on "Herding with Multiple States: When the Prior Matters." Basically, this paper shows that in learning with more than two states, whether people learn or herd really depends on what initial belief you start with.

Warren: Right, so it’s not just about the information itself. It means the whole setup matters for how people act. The authors are pointing out that they can have the same rules and preferences, but different starting points lead to totally different outcomes, learning versus just everyone following the crowd.

Iris: I see that dependence is being characterized by these conditions—escapability and pairwise escapability. It’s not a simple yes or no anymore; it depends on those specific conditions holding true. If those fail in a certain way, you get herding for every possible starting belief.

Ruth: Exactly, and they found that under certain technical setups, like interval optimality and monotone likelihood ratios, you can actually have learning happening for some starting beliefs while herding happens for others. It’s a distinction between who gets to learn versus who just follows the trend.

Warren: That makes sense from a strategy standpoint. If everyone starts with the same initial thought, and it still turns out they herd, that means the system itself is pushing them toward that collective action regardless of their starting point. It's about where the instability lives in the process.

Iris: And for anyone building a model, this tells you to look beyond just your own beliefs. You have to consider how the initial prior sets up those critical regions, like that cascade belief set they mentioned in Example two.

Ruth: Yeah, it underscores that identifying cascade beliefs and predicting herding are actually two different identification problems. It's not just about seeing if the signals are good; it's about understanding how the people interpret those signals based on where they started.

Warren: So, even if we nail the information structure and preferences perfectly, the initial conditions can still dictate whether you get a stable learning outcome or a runaway herding event. That’s a lot of work to untangle.

Xiaoyu Chen

University of Tokyo

econ.TH

Submitted: 2026-11-10

Updated: 2026-11-10

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 98/100

The gist: In sequential observational learning with more than two states, it is possible for identical information structures and preferences to lead to either learning or herding depending on the initial…

Key concepts

Excludability
This condition requires that signals can provide strong evidence to show any given action is suboptimal after the fact. If excludability fails, it implies that herding will occur for at least one possible initial prior.
Pairwise Escapability
A stronger version of escapability, this condition is also necessary for learning. If pairwise escapability fails, the paper proves that herding occurs for every possible full-support prior.
Interval Optimality and MLRP
These are specific environmental assumptions—interval optimality and monotone likelihood ratios—that simplify the characterization of learning and herding. Under these conditions, learning is completely determined by pairwise distinguishability, which relates to how well signals can distinguish between states.

Terminology

Summary

In sequential observational learning with more than two states, it is possible for identical information structures and preferences to lead to either learning or herding depending on the initial full-support prior. p. 5 The paper refutes a conjecture by Kartik, Lee, Liu, and Rappoport (2024a) by presenting a counterexample where the cascade region has a non-empty interior, yet learning occurs for generic full-support priors outside this region. p. 0

The authors establish two necessary conditions for learning: escapability and pairwise escapability. p. 0 Escapability requires that signals can provide arbitrarily strong evidence that any given action is ex post suboptimal; pairwise escapability is a variant of this condition. p. 0 The paper proves that if excludability fails, herding occurs for some initial prior; conversely, if pairwise escapability fails, herding occurs for every full-support prior (Theorem 1). p. 0

For environments characterized by interval optimality and monotone likelihood ratios (MLRP), these conditions completely characterize learning and herding independently of the initial full-support prior. p. 0 When location-shift signals are present, this characterization simplifies to pairwise distinguishability, which extends unbounded beliefs to the multi-state setting. p. 0 The authors demonstrate that learning can indeed depend on the initial prior even when preferences and information are fixed; specifically, learning occurs for generic full-support priors outside a cascade belief set in Example 2, whereas herding occurs for priors satisfying specific conditions on belief odds. p. 2 This dependence is characterized by showing that the failure of excludability implies herding for some prior with binary actions and finite states, but not necessarily for more than two actions. p. 5

Under the assumptions of interval optimality and MLRP, pairwise escapability is equivalent to pairwise distinguishability (Corollary 1). p. 4 Furthermore, in settings with location-shift information, pairwise escapability is equivalent to pairwise distinguishability (Proposition 5), establishing that excludability, (pairwise) escapability, and pairwise distinguishability are all equivalent under these conditions. p. 4 The paper also shows that if interval optimality fails under MLRP information, herding occurs for some initial prior (Proposition 6). p. 4 The final result states that with interval optimality and location-shift information satisfying MLRP, adequate learning exists if and only if both interval optimality and pairwise distinguishability hold (Proposition 6). p. 4

The analysis leverages the public belief process and its martingale property to characterize learning through limiting beliefs. p. 4 The main contribution is providing sufficient conditions for herding or necessary conditions for learning that differentiate between herding occurring for some initial prior versus occurring for every full-support prior. p. 5 This characterization is independent of the full-support initial prior in a natural class of environments defined by interval optimality and MLRP. p. 5 The paper concludes that in specific instances, learning can depend on the initial prior even when preferences and information are fixed, underscoring that identifying cascade beliefs and identifying herding are distinct problems. p. 2 Ultimately, under Theorem 2 (interval optimality, MLRP), preferences and information determine whether learning occurs independently of the full-support initial prior, with pairwise distinguishability serving as the characterization for location-shift information. p. 0

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improved AI systems can distinguish between learning and herding in sequential observational learning with more than two states by checking for escapability or pairwise escapability of utility functions relative to information structures. This allows the system to determine if a specific initial prior leads to learning for every full-support prior or herding for some initial prior.

  2. Improved AI systems can identify when the familiar binary state conditions extend to multi-state environments by checking for pairwise distinguishability under location-shift information and single-crossing differences of preferences. This allows the system to characterize learning and herding independently of the full-support initial prior in specific classes of environments.

  3. Improved AI systems can determine if information alone can characterize learning within a class where interval optimality holds by checking for pairwise distinguishability when the information structure is location-shift. This allows the system to verify if every pair of states is distinguishable from each other under these conditions.

  4. Improved AI systems can analyze the dynamics of public belief convergence by tracking whether it enters a cascade region or converges to a stationary belief with adequate knowledge using martingale properties. This allows the system to predict when learning occurs for generic full-support priors outside specific belief sets in multi-state environments.

  5. Improved AI systems can handle prior dependence by simulating different initial priors and observing how the resulting asymptotic learning outcome changes, especially in environments where cascade regions have nonempty interior but learning occurs for generic full-support priors outside it.

Sources

Related papers