Feasible Search Behavior
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: "Feasible Search Behavior".
Title and authors: Ruth: So this paper we're looking at is called "Feasible Search Behavior" and it deals with how people search for the best option when they have a bunch of different choices. The authors are Hiroto Sato and Ryo Shirakawa, and they're setting up a framework to look at this whole process.
Warren: Right, so it seems like the core idea here is that the way you search for things—the order you check options—isn't random; it’s shaped by how much information you get along the way. This paper tries to capture that relationship between how information is structured and the actual path a decision maker takes to find a choice, which can be quite complex because of that sequential nature.
Iris: What's interesting about this framework, as described in page one of "Feasible Search Behavior," is how the information structure isn't just about what you know at the end; it shapes your learning path before you even make a final choice. It links the ex-ante and ex-post value of investigating each option together, which really defines how the whole search path looks.
Ruth: Exactly, and they show that this whole set of possible search behaviors can be described as a polytope. They argue that there is just one specific information structure that can rationalize all these different search behaviors, and this specific structure is the one that minimizes the agent’s welfare among all possible information structures available.
Warren: That sounds like a strong claim, because minimizing welfare across all possible structures suggests finding a kind of universal optimal setup for the information designer. It’s not just about finding *a* good structure; it's about finding the one that performs best, no matter what the agent does to search.
Iris: And they characterize these vertices of this polytope, which are specific types of search behaviors where a vector has non-increasing components, and those are defined by just picking some index i and making all subsequent components equal to that value, which is what's on page three of "Feasible Search Behavior".
Ruth: That means the set of feasible search behaviors is actually quite constrained; it’s not just a random cloud of possibilities, but a specific geometric shape with defined corners. It makes sense that they focus on these vertices because they are the most extreme or defining cases in this model.
Title and authors: Warren: If we can pin down those vertices, it gives us something concrete to work with when we try to identify what information structure is actually in play in a real-world scenario. It moves the discussion from vague possibilities to specific points where things might be happening.
Iris: And they mention that a single distribution profile, which is basically the way information is distributed, can actually induce all these feasible search behaviors, and they derive the convexity of this set from that result rather than proving it directly. That's a nice structural find for the model.
Ruth: It really simplifies things by suggesting we don't need to test every possible way information could be structured; one structure should cover everything, and that structure is the welfare-minimizing one. This seems like a very elegant simplification of an otherwise messy problem in information economics.
Warren: From a trading perspective, if there's one optimal structure that minimizes agent welfare across all search patterns, it suggests there’s a stable equilibrium point for how information should be designed in this market. It reduces the search space for what kind of incentives we need to consider when setting up an online platform or any similar system.
Iris: Now they extend this into two different areas, looking at pairs of search and choice behaviors, with and without option sellers’ price competition, and they prove that the results stay the same even when you add that price competition element. That's a significant robustness result for their characterization of feasible pairs of search and choice behaviors.
Ruth: That robustness is important because it means the fundamental structure of how search interacts with choice decisions doesn't fundamentally change just because we introduce a price competition dynamic between the sellers, which is something we see all the time in markets.
Warren: So, when you look at those identified sets for a given observation v, which they characterize using inequality constraints derived from theorem one, you find that the set of equilibrium search-choice pairs doesn't shift whether sellers are competing on price at all; that’s a very stable finding for dynamic pricing environments.
Iris: And they also provide this identification result where the identified set I v is characterized by finitely many inequality constraints, which is useful because it gives us a concrete way to narrow down the possible information structures we need to consider when we are trying to identify them from actual data.
Title and authors: Ruth: If you can map observed search behavior data onto this polytope, as suggested by their work on robust identification, then you can actually estimate model parameters from that data without having to test every single structure imaginable. That’s a huge practical implication for empirical research in this area.
Warren: And if the AI can do that mapping, it means we could move toward designing information structures that minimize agent welfare across all feasible search behaviors, even when the objective function isn't simple or arbitrary; they find an optimal structure by minimizing agent welfare among all possible ones.
Iris: That leads into the idea of optimal stopping and choice modeling in equilibrium price competition settings where this model applies, because they can determine optimal stopping rules based on price-adjusted effective values to predict when an agent should stop searching or which option to pick after stopping.
Ruth: So, to wrap up on "Feasible Search Behavior," the main takeaway is that there's a single welfare-minimizing information structure that dictates the entire learning path and minimizes agent welfare across all possible search orders. This gives us a tractable tool for analyzing how information shapes decisions in ordered search environments.
Warren: And they also show that this framework is robust to price competition, meaning the set of equilibrium search-choice pairs doesn't change whether sellers compete on price, which is a solid result for dynamic market analysis.
Iris: To add one final point, the paper provides a way to characterize the set of all feasible pairs of search and choice behaviors as a polytope, which lets us verify if an observed pair is rationalizable under some information structure. This structural characterization is key for checking real-world data against the model.
Ruth: So, looking at this "Feasible Search Behavior" paper, it seems like the main implication is providing a very simple and tractable geometric description of how information structures govern sequential search paths and agent welfare in these settings.
Warren: And they've given us tools to actually identify those structures from data by mapping the observed behavior onto that polytope, which is a big step for empirical work.
Iris: We get concrete characterizations for identification problems and robust results concerning price competition, all stemming from this framework of characterizing the set of all search behaviors as a polytope.
The paper's summary: Ruth: So, to recap, this paper is essentially giving us a mathematical map for how information structures actually dictate the way people search for their best option in a sequence of choices.
Warren: Right, so it's not just about what you know at the end; it’s about how you learn along the way because that structure shapes your entire search path before you even pick something.
Iris: The authors show that all these possible ways people could search—all feasible behaviors—can be described as a specific geometric shape, which they call a polytope.
Ruth: And they found one single information structure that actually makes sense for all those different search paths, and this structure is the one that minimizes how much the agent suffers overall across every possible information setup.
Warren: That means if you're designing an incentive system or some platform, you don't need to test a million structures; there’s this one optimal configuration that handles everyone’s behavior.
Iris: They characterize the corners of this shape, those vertices, and they say those vertices correspond to very specific search orders where the vector components are in non-increasing order.
Ruth: So, we can figure out exactly what kinds of search strategies are actually possible under this model by looking at these defined points rather than just guessing.
Warren: And they also showed that if you look at price competition between sellers, the set of stable search and choice pairs doesn't actually change, which is a pretty robust finding for dynamic pricing scenarios.
Iris: They’ve also managed to define exactly what the identified set looks like by listing a finite number of inequality constraints, which is useful when we are trying to pull real data into the model.
Ruth: If an analyst observes some search behavior and maps it onto this polytope, they can actually estimate the model parameters because they have these concrete constraints to work with.
Warren: This moves us from just theoretical modeling to actually using observed data to pin down what’s going on in the market.
Iris: It suggests that we can design information structures that minimize agent welfare across all those feasible search patterns, even if our objective function isn't totally simple or arbitrary.
Ruth: So, the big picture here is that we have a tool to figure out how information shapes decisions sequentially and find the setup that minimizes harm for everyone involved.
Warren: And this framework opens up ways to actually test if real-world search patterns are rational under some specific information structure by checking them against these geometric constraints.
Iris: The next thing we look at is how this connects to optimal stopping rules when there's price competition involved in the choice part of the process.
The paper's improvements: Ruth: So, we've seen how the paper maps search behavior onto a polytope, but now let's look at what they suggest we can actually *do* with this characterization of feasible search and choice pairs.
Warren: Right, so they aren't just describing the problem; they’re giving us ways to use this geometric description for identification and optimization.
Iris: They propose using these constraints to identify the specific information structures that are generating the observed data, which is a big step toward empirical work because it gives us something concrete to test against.
Ruth: And they also suggest designing those information structures in a way that minimizes agent welfare across all those possible search orders, even when the objective function isn't straightforward.
Warren: That sounds like they are suggesting how an AI could actually build a system that sets incentives optimally by minimizing the worst-case outcome for the decision maker.
Iris: They also address what happens when you introduce price competition between sellers; they prove that this dynamic doesn't change the fundamental set of stable search and choice pairs we identified earlier.
Ruth: That robustness is key because it means our results hold up even in competitive pricing environments, which is a major plus for any real-world application.
Warren: They also showed that the identified set for a given observation, that set I v, can be characterized by a finite number of inequality constraints, which makes the identification problem much more tractable.
Iris: So if we see some search data, we can use those constraints to narrow down the possibilities for what information structure could be in play.
Ruth: And they point toward using these results to determine optimal stopping rules when price competition is present, telling the agent exactly when to quit searching or which option to pick.
Warren: That's a practical application; knowing the best time to stop searching based on price-adjusted values sounds like something a trader could actually use.
Iris: The limitation they state is that this framework focuses heavily on characterizing search behavior and choice pairs, but it doesn't necessarily cover every possible type of interaction between information and decision making.
Ruth: That makes sense; it’s a focused tool for ordered search paths, not a universal model for every single market interaction out there.
Warren: So we've seen the tools they offer for identifying structures and optimizing incentives, but the next thing we need to look at is how these ideas scale up when we move beyond simple sequential search.
Conclusion: Ruth: So we've covered how the paper "Feasible Search Behavior" uses polytopes to map out all possible search paths and find the welfare-minimizing information structure that rationalizes them.
Warren: It really boils down to a geometric description of how information structures force agents into certain search patterns, and they give us tools to identify those structures from observed data.
Iris: And we also saw how this framework handles price competition robustly, meaning the stable outcomes don't shift even when sellers start competing on price.
Ruth: For someone just listening, it means we have a mathematical way to understand why people search the way they do based on what information they get along the line.
Warren: It’s about having a specific model for how an AI system should design incentives if it wants to minimize losses for everyone involved in that search process.
Iris: The caveat is that this framework is focused on these ordered sequences of search and choice, so it doesn't necessarily cover every kind of complex decision-making interaction out there.
Ruth: Right, a tool for a specific area, not the whole world economy but very powerful for understanding sequential learning.
Warren: Exactly; it gives us a solid foundation for testing those hypotheses about how information structures influence behavior in real markets.
Iris: It’s useful because it provides the constraints needed to actually pull empirical data into the model and see if our assumptions hold up against what we observe.
Ruth: So that's what they delivered with "Feasible Search Behavior," a structured way to analyze how information shapes sequential search learning paths.
Warren: It’s a clean characterization, even if it doesn't solve every single market mystery at once.
Iris: Next week, we look at how this kind of structural analysis connects to dynamic pricing and optimal stopping rules when competition is in play.
Nagoya University · Massachusetts Institute of Technology
econ.TH
Submitted: 2026-09-14
Updated: 2026-11-18
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 88/100
The gist: Consider a situation wherein a decision maker sequentially searches for the best alternative among heterogeneous options with an arbitrary search order The information structure jointly determines…
Key concepts
- Search Behavior Polytope (V)
- This is the set of all feasible search strategies an agent might employ. The paper shows this set forms a polytope, which means it has a well-defined geometric shape defined by linear inequalities. Vertices of this shape correspond to specific, extreme search orders.
- Information Structure
- This refers to the way information is structured and revealed during the sequential search process. The paper investigates how different information structures jointly determine both the expected value and the actual value an agent places on investigating each option.
- Welfare Minimization
- The goal is to find an information structure that minimizes the agent's overall welfare across all possible feasible search behaviors. This means designing a system where, regardless of how an agent searches, their expected payoff is as low as possible under the best possible information design.
- Equilibrium Search-Choice Pairs
- These are pairs of search and choice decisions that are mutually consistent given the underlying information structure and agent preferences. The paper proves that these equilibrium pairs remain stable even when option sellers compete on price.
Terminology
Summary
Consider a situation wherein a decision maker sequentially searches for the best alternative among heterogeneous options with an arbitrary search order The information structure jointly determines the ex-ante and ex-post value of investigating each option, thereby shaping the entire learning path. p. 0
Core Framework and Characterization
The paper provides a simple framework for analyzing how information structures shape ordered learning paths by characterizing the set of all search behaviors compatible with some information structure, which forms a polytope A single information structure rationalizes all these search behaviors, which minimizes the agent’s welfare among all information structures. This result suggests that our model not only captures the interaction between information structures and the ordered learning path in perhaps the simplest possible way, but also serves as a tractable tool for analyzing it. p. 2
Feasible Search Behaviors and Vertices
Theorem 1 summarizes our main findings, stating that the set of all feasible search behaviors is a polytope Any vector v with v1 ≥ v2 ≥ · · · ≥ vn is a vertex of V if and only if v = (v1,..., vi, vi+1,..., vn) for some i. Furthermore, a single distribution profile F induces all feasible search behaviors. p. 7 The convexity of V is obtained as a corollary of this result, rather than being proved directly.
Information Design Implications
In applications, information designers may influence agents’ behavior not only through information but also through other instruments Theorem 2 shows that the welfare implications change when such a platform also earns sales commissions. The second statement is about the case of linear objective functions, where the information designer can robustly implement the optimal search behavior. p. 2
Identification and Price Competition
The closed-form characterization in Theorem 1 can be directly mapped to the closed-form characterization for this partial identification problem The identified set I(v; Γ) is characterized by finitely many inequality constraints. Theorem 3 proves that the set of equilibrium search-choice pairs does not change at all whether sellers compete on price
Extensions to Search and Choice Behaviors
Theorem 2 extends the characterization to feasible pairs of search and choice behaviors, with and without price competition The set S∗ is a polytope Lemma 5 shows that F minimizes Pandora’s expected payoff among all feasible distributions, which also equals z.
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
-
Robust Information Structure Identification: The AI can now derive a tractable set of model parameters based on observed search behavior data by mapping it to a polytope defined in Theorem 1, allowing for identification under minimal informational assumptions.
-
Welfare-Minimizing Information Design: The system can design information structures that minimize agent welfare across all feasible search behaviors, even when the objective function is arbitrary, by finding an optimal information structure that
minimizes the agent’s welfare among all information structures.
-
Feasibility Characterization of Search and Choice Pairs: The AI can characterize the set of all feasible pairs of search and choice behaviors as a polytope, enabling it to verify if a specific observed pair is rationalizable under some information structure.
-
Price Competition Robustness: The system can analyze the effect of option sellers' price competition by proving that
the set of equilibrium search-choice pairs does not change at all whether sellers compete on price,
providing stable results in dynamic pricing environments. -
Identification from Search Data: Analysts can estimate model parameters from observed search behavior data by computing the
identified set for a given observation v, i.e., the set of model parameters that rationalize the data v with some information structure.
-
Optimal Stopping and Choice Modeling: In equilibrium price competition settings, the AI can determine optimal stopping rules based on price-adjusted effective values, allowing it to predict when an agent should stop searching or which option to choose upon termination.
Sources
- Limits of Disclosure in Search Markets
- A Note on Equivalent Conditions for Majorization
- Rationalizing dynamic choices
- Revealed Information
- Competitive Information Disclosure with Heterogeneous Consumer Search
- Attention Capture
- How to Segment a Search Market: Information Design and Directed Search
- Competition, Persuasion, and Search
- Persuasion Produces the (Diamond) Paradox