Daily Summary for 2026-11-12

daily

Transcript

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

Ruth: It's the twelfth of November, twenty twenty-six, and this is the day's research.

Warren: 17 new papers came out today.

Ruth: I'm Ruth.

Warren: And I'm Warren, here with Iris.

Ruth: We'll take the day in one pass, then pull out the papers we're staying with.

Warren: Start us off with Bundling Complements.

The day's research: Ruth: Wang et al show long-lived senders advise myopic receivers using recommendation records.

Warren: That mechanism suggests nonbinding self-restraint causes selective advice by using unused capacity later.

Iris: How does that relate to the payoff cap when obedience costs something to restrain?

Ruth: The sender's payoff is capped by the best signal minus the expected cost of restraint regardless of patience.

Warren: Then we look at how combinatorial preferences dictate optimal bundling for sellers.

Iris: Does that work when a monopolist's type distribution determines virtual value but preference dictates actual transport?

Ruth: The actual required directions and covered bundles are determined by preference above a certain threshold.

Warren: In double auctions, licenses forbid recombination; what happens to welfare cost as the market grows?

Iris: The welfare cost of forbidding assembly vanishes if expected demand matches whole bundle supply at prevailing prices.

Ruth: We also examine designing markets for screening through agent-oriented optima using rather than suppressing information rents.

Warren: That leads us to identification in auctions when transaction prices are truncated, which matters greatly.

Iris: How does that clarification help estimate the screening threshold across different bidder scenarios?

Ruth: Identification depends on recovering this participation threshold across various truncation types and bidder scenarios.

Warren: This builds upon dynamic regression discontinuity designs for identifying treatment effects intertemporally.

Iris: Does reformulating assumptions from difference-in-differences literature allow for economically interpretable intertemporal effects?

Ruth: It proposes a novel argument to find such effects by reformulating assumptions in the literature.

Warren: Volatility estimation is getting more robust with debiased kernel estimation even with infinite variation jumps.

Iris: Using more general kernels achieves smaller asymptotic variances when estimating spot volatility.

Ruth: Information aggregation predicts AI chatbots can extract an advantage by aggregating news from analysts and queries.

Warren: This contrasts with the invisible hand as an emergent property showing firms reallocate under profit-seeking behavior.

Iris: Does that structural analysis show long-run efficiency is stable regardless of decentralized adjustment speed?

Ruth: Yes, it suggests long-run efficiency is globally stable regardless of decentralized adjustment speed.

Warren: Finally, coordination mechanisms examine how mediators coordinate outcomes with only partial statistics disclosed.

Iris: Implementability depends entirely on the loss function chosen when players minimize loss over consistent distributions.

Lucky paper: 2607.07982: Ruth: Okay, we're looking at Bundling Complements now, which is paper number one of five in this set.

Warren: The paper introduces a method for screening using combinatorial preferences through what they call "one duality with two reductions."

Iris: What is the first reduction that requires tangency on the valuation map?

Ruth: That first reduction reduces the flow to transport between types, which is shown in Corollary one.

Warren: And the second reduction requires linearity of the valuation map for preference independence.

Iris: Under separation, bundle cones geometrically characterize complementarity by showing how each bundle extracts surplus along specific directions.

Ruth: The comparative statics reveal two thresholds describing optimality: a lower threshold where offering a grand bundle is required and an upper threshold above which an inclusive core-peripheral menu is optimal.

Warren: That means there are three regimes based on these thresholds: below the lower, pure sales are optimal; above it but below the upper, pure bundling works; and above the upper, a core-peripheral menu is optimal.

Iris: The premium family takes that core-peripheral form because it has a finite upper threshold.

Ruth: For one-dimensional types, pure bundling can only be optimal if the valuation curve is tangent at its critical point.

Warren: And for linear valuation maps, pure bundling is also only optimal if the valuation curve is tangent at its critical point.

Iris: A curve that is tangent nowhere makes pure bundling suboptimal at every premium and for every distribution.

Ruth: The paper establishes that a common premium of at least one on the bundles an optimal menu sells keeps it optimal when prices are scaled, as shown in Corollary two.

Warren: The second contribution to comparative statics shows that for linear valuation maps, pure bundling can only be optimal if the valuation curve is tangent at its critical point again.

Iris: Does this mean non-tangency always leads to suboptimality regardless of the distribution?

Ruth: Yes, a curve tangent nowhere makes pure bundling suboptimal at every premium and for every distribution in that context.

Warren: So, the method itself is defined by that initial duality with two reductions where maximal revenue equals the minimum cost of a flow of differential virtual value across valuation vectors.

Iris: That first reduction represents the differential virtual value on types, which exists if and only if tangency holds according to Proposition one.

Ruth: On a convex valuation manifold, that flow reduces to transport between types as shown in Corollary one.

Warren: The second reduction requires linearity of the valuation map for independence from preference across the setup.

Iris: When we are under separation, bundle cones geometrically characterize complementarity and show how each bundle extracts surplus along specific directions.

Ruth: The paper concludes that for linear valuation maps, a menu is optimal when the bundles each buyer chooses cover the required directions through that buyer as per Proposition three.

Warren: And to keep that menu optimal, you need a common premium of at least one on the bundles sold, with prices scaled according to Corollary two.

Iris: So it's not just about geometry; it’s about those specific price scaling rules keeping the menu optimal when you move past the pure bundling regime.

Lucky paper: 2608.06134: Ruth: Okay, let's talk about this paper, Large-Market Discipline in Combinatorial Double Auctions: No Recombination, Bundle Selection, and Complementarities.

Warren: It focuses on double auctions where platforms can't legally or technically combine components from different sellers into a single bundle.

Iris: So the authors formulate this as a combinatorial buyer’s-bid double auction under that constraint.

Ruth: Under explicit stability and price-influence conditions, each bundle submarket inherits the large-market discipline of single-good double auctions.

Warren: That means bid shading vanishes, and clearing prices concentrate on competitive levels tracking the common value, which is price discovery.

Iris: The key incentive step—that bidding on a bundle creates no first-order strategic distortion beyond single-good logic—is proven for two goods.

Ruth: And that condition holds even for larger item sets, which is interesting given the complexity of those markets.

Warren: Multi-agent reinforcement learning simulations decompose the welfare loss, showing that No Assembly, not strategic shading, is the binding finite-market friction.

Iris: The simulation evidence suggests structural costs are larger than allocational losses in moderately thick markets.

Ruth: And both those costs diminish as market growth and complementarity between goods increase.

Warren: The model assumes correlated private values where every trader knows their own value exactly, linked only through this common component.

Iris: The central analytical device is submarket reduction, conditional on market sorting traders into thick regular bundle submarkets.

Ruth: That intuition for reduced strategic behavior is that as a submarket thickens, any single trader’s chance of setting the price shrinks.

Warren: They use XOR bidding language to be fully expressive over bundles, and the allocation rule selects the one with the highest total clearing prices among welfare-maximizing allocations.

Iris: The structural loss from No Assembly is defined as ELstructural one minus E of W Truthful over E of W Optimal.

Ruth: While the allocational efficiency loss, or strategic loss, is defined as ELallocational one minus E of W Realized over E of W Truthful.

Warren: Simulations show both losses are small in moderately thick markets, which supports the idea that a decentralized bundle auction can achieve efficient allocation.

Iris: The theory concludes that once the market is moderately thick and goods are genuine complements, it requires no central authority for assembly.

Ruth: Their key results include a conditional large-market rate result for the two-item combinatorial BBDA where lambda eta buyer of S is big O one over eta.

Warren: And lambda eta seller of S is big O one over eta as eta goes to infinity.

Iris: The structural loss bound is conjectured to be at most C times eta, and numerical verification supports that conjecture.

Ruth: They also conjectured that the No-Assembly price gap satisfies delta p with respect to eta at most C times one over e of c times eta plus C two over eta.

Warren: The empirical results confirm that structural loss from No Assembly exceeds allocational loss in every cell, and complementarity helps reduce both components as synergy increases.

Iris: So the binding friction is finite-market dependent, shrinking at a rate close to one over eta as market size grows.

Lucky paper: 2604.09340: Ruth: Today we’re looking at Designing Markets for Screening Optimal Composition and the Geometry of Surplus.

Warren: This paper explores market compositions that generate Pareto efficient outcomes within a screening model featuring type-dependent costs.

Iris: When you maximize principal or total surplus, the market concentrates at the highest type, extracting all surplus.

Ruth: But agent-oriented optima serve all agents without bunching and efficiently serve a positive share of agents with the highest type.

Warren: Separation is valuable even under agent surplus maximization because it propagates information rents.

Iris: Under mild cost-curvature conditions, every Pareto efficient outcome corresponds to a unique market structure.

Ruth: If you have stronger conditions, the surplus set is convex when values are private or when types without gains from trade can be included in the market.

Warren: Otherwise, it can become nonconvex due to type-dependent costs.

Iris: Market segmentation generates missing outcomes but does not increase maximum agent surplus at any level of principal surplus.

Ruth: The paper details Theorem one which says if beta is at least one the unique optimizer is a degenerate market structure delta theta bar.

Warren: If beta is in the range of zero to one, the optimizer has connected support and no realized types are assigned zero quality.

Iris: There's also a positive atom at the highest type served at first-best quality under Assumption three.

Ruth: Under Assumption three uniqueness is guaranteed with a support defined by beta, theta, and theta bar where phi of zero is greater than zero.

Warren: Furthermore, under Assumption three the weighted problems trace the entire Pareto frontier for beta in the range of zero to one.

Iris: That means every Pareto payoff is realized by a unique market structure identified in parts one and two of that paper.

Ruth: Theorem two describes the geometry of the surplus set showing how adding mass to the highest first-best no-trade type scales payoffs.

Warren: Under Assumption four V zero is compact and convex, which ensures every payoff has a single market implementation.

Iris: Private values without a payoff-neutral type and under strong curvature conditions maintain a compact and convex set even without such types.

Ruth: Conversely, type-dependent costs without a payoff-neutral type can lead to nonconvex single market payoff sets.

Warren: Comparative statics show that greater consumer orientation raises expected consumer surplus while lowering rent for types above a minimum quality.

Lucky paper: 2412.09321: Ruth: So we're looking at Learning to Choose with Categories, which uses Coarse Q-learning to handle stochastically varying menus.

Warren: The paper describes organizing alternatives into similarity classes based on an exogenously defined equivalence relation and learning at that class level using a multinomial logit rule.

Iris: That coarse aggregation compresses the exponential complexity of alternative-menu pairs into linear growth in the number of classes, which suggests it’s a form of stateless Q-learning for tractability.

Ruth: Exactly, and they derive continuous-time mean-field dynamics using stochastic approximation, yielding that ordinary differential equation v dot s equals g s(v) minus v s.

Iris: The paper establishes that the set of steady states under Coarse Q-Learning is non-empty and generically finite with odd parity.

Ruth: They show that as the payoff sensitivity parameter approaches infinity, the accumulation points of these steady states select a subset of the Valuation Equilibria from Jehiel and Samet two thousand seven.

Warren: That sounds like a big result because it links this coarse learning to existing valuation equilibria concepts.

Iris: One key finding is that in settings with two similarity classes, both multiple steady states and a unique steady state where class valuations are equalized in the high-sensitivity limit are demonstrated.

Ruth: Then they use three similarity classes to show that the learning dynamics can exhibit instability, cycling perpetually through all possible rankings of similarity classes.

Iris: The long-run dynamics can exhibit indeterminacy, indifference, or instability depending on whether the expected payoffs in singleton menus are sufficiently large or low.

Warren: When expected payoffs in singleton menus are low, the feedback logic supports multiple self-confirming rankings of similarity classes instead of just pushing valuations toward indifference.

Iris: The paper connects coarse categorization to limited salience and cognitive complexity, interpreting Coarse Q-learning as a deliberate form of stateless Q-learning for tractability.

Ruth: But they also relate it to Bayesian learning under model misspecification, which is an interesting connection for the learning mechanism itself.

Warren: The authors confirm structural properties of the solution space by establishing that for every finite number of classes, the correspondence V beta is non-empty and compact on

zero infinity). [Iris: Theorem one states that for generic decision trees and almost every beta, every steady state is regular, meaning the set V beta is finite for almost all beta.

Ruth: They also describe the convergence of relative valuation x and mean valuation y under an ordinary differential equation system.

Warren: For sufficiently large beta, Theorem six proves that every Valuation Equilibrium must be fully mixed, meaning total indifference among all classes.

Iris: That's a strong claim because it suggests a certain level of complete mixing in the long run under specific assumptions.

Ruth: The paper also shows the existence of a unique steady state for large enough z and beta, which is globally asymptotically stable.

Warren: They prove this stability by showing that the eigenvalues of the reduced Jacobian coincide with those of the full system except for a strictly negative eigenvalue.

Iris: Theorem three proves there exists a strict pure LSVE under certain conditions where v infinity is a steady state satisfying v infinity equals g(v infinity).

Ruth: And they also show that for every order pi, there exists a beta hat pi of z less than infinity such that for all beta at least beta hat pi of z, the unique steady state in a neighborhood is locally asymptotically stable.

Warren: So it connects the coarse perception directly to limited salience and cognitive complexity, which is what they are interpreting Coarse Q-learning as.

Lucky paper: 2512.20851: Ruth: Alright, we're moving on to paper number five of five this day, and it’s Testing Exclusion and Shape Restrictions in Potential Outcomes Models.

Warren: This paper deals with how exclusion and shape restrictions define causal effects in potential outcomes models using a graph-based representation.

Iris: So, what is the core idea here for someone listening who isn't deep into causal inference?

Ruth: They are developing a general framework to characterize sharp testable implications of general support restrictions on potential response functions.

Warren: The model involves an exogenous input Z, like an instrument or treatment, and an endogenous response R, which can be treatments or outcomes themselves.

Iris: When you have these support restrictions defined by a set R* in RX almost surely, what does the graph actually encode?

Ruth: A graph-based representation encodes the compatibility structure of the model. If two vertices are disconnected in this graph G, then certain events cannot occur jointly under those modeling assumptions.

Warren: Theorem one states that if Assumptions one and two hold, for any maximal independent set I of the potential response graph G, the inequality X v r z in I P(R = r Z = z) less than or equal to one is necessary for a joint distribution Q to exist.

Iris: That sounds like you are saying if that inequality fails, there's no way for the required distribution Q to exist under those conditions.

Ruth: Exactly, and Theorem two generalizes that by stating that for any subset V of the vertices of G, the inequality X v r z in V P(R = r Z = z) less than or equal to G(V) is necessary and sufficient for Q's existence.

Warren: Assumption three requires a regular support restriction, which means the graph G must be perfect and every maximal clique has to correspond to a support point r* in R*.

Iris: If Assumption three holds, Theorem three establishes that those inequalities from Theorem one become sharp. What does that sharpness mean for our understanding?

Ruth: It means we have found the tightest possible bounds for the necessary conditions when the support restrictions are well-behaved.

Warren: The analysis extends to continuous covariates through a sequence of potential response graphs G n and a generalized theorem based on partitions A n, k.

Iris: And computationally, what are the complexities for constructing this graph G?

Ruth: There are two methods: checking pairwise incompatibility which has a complexity of O(Y squared D squared Z squared), or enumerating all r* in R* which is O(Y D D Z).

Warren: Inference procedures use moment inequalities derived from the half-space representation of conditional distributions, plus a test based on vertex representation involving solving a quadratic program to define critical values.

Iris: That sounds quite involved; how do you handle continuous covariates when testing these implications?

Ruth: You accommodate them by discretizing them into bins when testing such implications.

Warren: Testing exclusion and shape restrictions in potential outcomes models is what this paper is about, we're looking at the paper Testing Exclusion and Shape Restrictions in Potential Outcomes Models.

More episodes

← Home