Ranking Statistical Experiments via the Linear Convex Order and the Lorenz Zonoid: Economic Applications

arXiv:2502.06530 · econ.TH, math.ST, stat.TH · Submitted 2025-02-10 · 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: "Ranking Statistical Experiments via the Linear Convex Order and the Lorenz Zonoid".

Paper summary: Ruth: So, this paper introduces something they call the linear-Blackwell or LB order for ranking statistical experiments. Basically, it claims this new ranking has three equivalent ways to be defined: first, it relates to how dispersed the induced posterior and likelihood ratios are in a specific sense called the linear convex order; second, it relates to the size of something called the Lorenz zonoid which is essentially a set of statewise expectation profiles; and third, it relates to how variable the posterior mean is.

Warren: It’s interesting because they apply this LB order to compare experiments in binary-action decision problems and also in decision problems where you have payoffs that are quasi-concave. They also use it for moral hazard problems, building on earlier work from Holmström and Kim, and screening problems where you have signals after the fact.

Iris: That makes sense, but what's the main point they're pushing? They seem to be trying to find a way to rank experiments that is less demanding than the Blackwell order.

Ruth: Exactly. The Blackwell order is celebrated for ranking experiments based on informativeness, but this paper suggests it imposes requirements that are often too strict, especially in principal-agent frameworks where there's more structure on how information can be used.

Warren: They introduce this LB order because it seems more tractable when you have multiple states, specifically when the number of states is two or more. The way they characterize it involves comparisons only with univariate convex functions, which makes it less computationally demanding than the Blackwell order.

Iris: I'm looking at the characterization where F dominates G if the Lorenz zonoid of F contains the Lorenz zonoid of G. That seems like a very concrete geometric way to measure dominance.

Ruth: Right, and they show that this geometric condition, that Z(F) contains Z(G), is equivalent to both the dispersion characterization and the posterior expectation variability characterization. That means you can use whichever definition is easiest for your specific problem structure.

Warren: They also connect this LB order to another concept called the monotone-posterior-expectation or MPE order, which they say coincides with the Lehmann order under maximum likelihood ratio procedures, but works even when maximum likelihood doesn't apply.

Iris: It’s helpful to see how these different ways of measuring dominance relate. If F dominates G in LB order, it means the Lorenz zonoid of F is larger than or equal to the Lorenz zonoid of G, which intuitively suggests that F provides more flexibility in statewise expectations than G does.

Ruth: And this flexibility translates into real economic utility across different domains. In decision problems, if you use this LB order, it captures relative preferability in decision problems where the payoff is quasi-concave in a given belief.

Warren: For moral hazard problems, dominance under the LB order corresponds to both a wider range of implementable target actions and a smaller minimal disutility that the principal has to incur when implementing any target action.

Iris: That implies greater flexibility in control and also greater effectiveness because there's less cost for the principal. But what are the actual numbers on this?

Ruth: They establish bounds on penalties and rewards, showing that if F dominates G in LB order and you have a uniformly bounded scheme under G that implements a target action, then there exists a scheme under F that implements that same target action at lower disutility for the principal.

Warren: And in screening problems with ex post signals, the LB order means the principal's expected utility from her optimal mechanism is higher under F than under G for every screening environment.

Iris: It sounds like this ordering gives us a way to compare different experiments that respects the structure of how information is actually used by an agent. But what about its limitations?

Ruth: The paper flags that the Blackwell order can be overly demanding in the principal-agent framework because it's designed to compare experiments uniformly across all decision problems, whereas the agency model imposes a specific structure on how information can be used.

Warren: And they also point out that verifying the Blackwell order of two experiments is often computationally infeasible, especially when the signal space is infinite. The LB order seems to bypass some of those computational hurdles by only requiring comparisons with univariate convex functions.

Iris: So, while the LB order offers a more practical way to rank experiments, we still have this caveat about its applicability depending on the specific structure of the economic problem and whether we stick to finite or continuous state spaces.

Ruth: That’s right. For continuous state spaces, they simplify things further; for every prior and every function you define, the dominance condition boils down to that Lorenz zonoid inclusion we talked about earlier.

Warren: They conclude by showing that for finite signal spaces, LB dominance is equivalent to the Blackwell order, but in continuous state spaces, it remains a distinct measure of informativeness.

Iris: The main implication for us is that we have a more nuanced tool than just the Blackwell order when we're looking at decision-making or risk management problems where structure matters.

Ruth: It means if you're comparing two ways to gather information, the LB order tells you which one actually gives you a better guarantee of outcomes for your expected payoff under the specific constraints of your problem.

Conclusion: Ruth: So, we've been looking at how this linear-Blackwell order ranks experiments based on dispersion and geometric shapes like Lorenz zonoids, and now we need to wrap up with what this paper is actually saying about it.

Warren: The title is "Ranking Statistical Experiments via the Linear Convex Order and the Lorenz Zonoid: Economic Applications," so it’s clearly connecting some very abstract math to how things work in real-world decision problems.

Iris: It seems the authors are trying to show that this new ranking system, the LB order, isn't just a mathematical curiosity; it has direct relevance when we look at how people make choices or manage risk under uncertainty.

Ruth: Exactly. They're taking concepts from probability and statistics and applying them to things like decision problems where you have binary choices or moral hazard scenarios.

Warren: The main takeaway is that this LB order gives us a way to compare different ways of collecting data—different experiments—and tells us which one is genuinely more informative in a way that matters for the actual outcome, not just some abstract measure.

Iris: They've shown that this geometric comparison with the Lorenz zonoid actually works across different types of situations, whether you're dealing with continuous state spaces or finite ones.

Ruth: And they’ve linked it back to things like the Blackwell order, showing that under certain conditions, these two ways of ranking experiments end up being equivalent.

Warren: So, what this means for us is that we have a more practical tool to choose which experiment to run when we're trying to make decisions or set up insurance schemes.

Iris: It gives us a concrete way to assess the trade-off between the complexity of an experiment and the quality of information it provides.

Ruth: And this leads into how these results translate into real-world applications in areas like decision making and screening problems.

Aalto University

econ.TH, math.ST, stat.TH

Submitted: 2025-02-10

Updated: 2026-11-16

Comments: The main text ends on page 45, and the supplementary material follows thereafter. This paper was previously circulated under the title "Experiments in the Linear Convex Order''

License: http://creativecommons.org/licenses/by-nc-nd/4.0/

Importance score: 87/100

The gist: This paper introduces a novel ranking of statistical experiments known as the linear-Blackwell (LB) order, which is characterized by three equivalent properties…

Key concepts

Linear Convex Order
This mathematical tool measures how dispersed a random vector is by comparing it to other vectors using only univariate convex functions. If one vector has greater dispersion in this sense, it is considered 'more dispersed' than another, forming the basis for ranking experiments.
Lorenz Zonoid
The Lorenz zonoid is a geometric shape representing the set of all possible statewise expectation profiles for an experiment. Dominance in this order means one experiment's zonoid completely contains another's, indicating superior performance or informativeness.
Posterior Expectation Variability
This concept compares how much the expected posterior mean changes when moving from one experiment to another. An LB-dominant experiment ensures that the resulting posterior expectation is a 'mean-preserving spread,' meaning it provides more stable and informative results.
Blackwell Order
The Blackwell order is an older ranking of experiments. The paper shows that the new LB order is more powerful, especially in cases with multiple states, because it only requires comparisons involving univariate convex functions, making the LB order easier to work with.

Terminology

Summary

This paper introduces a novel ranking of statistical experiments known as the linear-Blackwell (LB) order, which is characterized by three equivalent properties: (i) the dispersion of induced posterior and likelihood ratios in the sense of the linear convex order, (ii) the size of the Lorenz zonoid (the set of statewise expectation profiles), or (iii) the variability of the posterior mean. p. 1

Where the paper says it

Finally, verifying the Blackwell order of two experiments is often computationally infeasible, especially when the experiments have infinite signal spaces. This paper introduces a novel ranking of experiments, the linear-Blackwell (LB) order, which has three equivalent characterizations: (i) in terms of the dispersion of the induced posterior and likelihood ratios, in the sense of the linear convex order, (ii) in terms of the size of the Lorenz zonoid (the set of statewise expectation profiles), and (iii) in terms of the variability of the posterior mean.

The LB order is formally defined such that an experiment F dominates an experiment G in LB order if and only if p F(q) lcx p G(q) for every prior q in delta n, where p F(q) is the posterior vector induced by experiment F and prior q. p. 3

Where the paper says it

The experiment F and the prior jointly generate the unconditional distribution of the random variable (signal) X and thus a random posterior vector pF (X). The characterization (i) states that F dominates G in the LB order if and only if the random vector pF (X) is more dispersed than the random vector2 pG(Y ) in the sense of the linear convex order.

The characterizations are detailed as follows:

  1. Dispersion Characterization: Experiment F dominates G if and only if the random vector p F(X) is more dispersed than the random vector p G(Y) in the sense of the linear convex order, meaning E[C[P n i=1 bip F,i(X)]] at least E[C[P n i=1 bip G,i(Y)]] for every univariate convex function C.

  2. Geometric Characterization: F dominates G if and only if the Lorenz zonoid of F, denoted as Z(F), contains the Lorenz zonoid of G, i.e., Z(F) Z(G). This geometric condition is equivalent to expectation matching: for every [0, 1] -valued function of G 's signal, there exists a [0, 1] -valued function of F 's signal with the same conditional expectation in each state.

  3. Posterior Expectation Variability: F dominates G if and only if the posterior expectation E[phi(theta tilde) X] under F is a mean-preserving spread of that under G for every function phi: theta to R, specifically E[C[p F(X)]] at least E[C[p G(Y)]] for every convex function C: R n to R.

The paper establishes that the LB order is more tractable than the Blackwell order, particularly when n at least 2, because it only requires comparisons involving univariate convex functions. p. 3

Where the paper says it

However, when n ≥ 2, the Blackwell order is far more complex, as noted by Blackwell and Girshick (1954): in general, it is computationally infeasible to check whether one experiment dominates another, and many pairs of experiments are not even comparable. This is because the set of multivariate convex functions has no tractable characterization (see, e.

Furthermore, the geometric characterization (ii) is shown to be equivalent to both (i) and (iii). p. 1
Where the paper says it

We take (i) as the definition of the LB order; (ii) and (iii) are equivalent reformulations. Let us start by discussing the geometric characterization (ii).

The paper also relates the LB order to other concepts, proving that F LB G if and only if F MPE G for every ordinal ranking of the state space. p. 7
Where the paper says it

Furthermore, MLRP, the Lehmann order, and the MPE order are all defined with respect to an ordinal ranking of the state space. We show that F dominates G in the LB order if and only if F dominates G in the MPE order for every ordinal ranking of the state space.

For finite signal spaces, irredundancy implies equivalence with the Blackwell order (F LB G F B G). p. 6
Where the paper says it

, Blackwell, 1951, Blackwell, 1953, Moscarini and Smith, 2002, Ganuza and Penalva, 2010, Cabrales, Gossner, and Serrano, 2013, Mu, Pomatto, Strack, and Tamuz, 2021, Brooks, Frankel, and Kamenica, 2024, Frick, Iijima, and Ishii, 2024, and Whitmeyer, 2025). The most closely related work is that of Bertschinger and Rauh (2014), who compare two experiments with finite signal spaces by examining the zonotopes generated by their likelihood vectors.

Crucially, in continuous state spaces, the LB dominance condition simplifies to requiring that for every prior q in P(theta) and every function phi: theta to R, E[C[phi(theta tilde) X]] at least E[C[phi(theta tilde) Y]], which is equivalent to the condition Z(F) Z(G). p. 2
Where the paper says it

Note that the geometric characterization of the LB order requires expectation matching: F dominates G if, for every [0, 1]-valued function of G’s signal, there exists a [0, 1]-valued function of F’s signal with the same conditional expectation in each state. In contrast, the garbling condition for the Blackwell order requires more stringent distribution matching: F dominates G if, for every [0, 1]-valued random variable derived from G’s signal, there exists a [0, 1]-valued random variable derived from F’s signal with the same conditional distribution in each state.

The paper concludes that F about LB G if and only if F about B G. p. 0
Where the paper says it

RANKING STATISTICAL EXPERIMENTS VIA THE LINEAR CONVEX ORDER AND THE LORENZ ZONOID: ECONOMIC APPLICATIONS KAILIN CHEN February 1, 2026 This paper introduces a novel ranking of statistical experiments, the linearBlackwell (LB) order, which can equivalently be characterized by (i) the dispersion of the induced posterior and likelihood ratios in the sense of the linear convex order, (ii) the size of the Lorenz zonoid (the set of statewise expectation profiles), or (iii) the variability of the posterior mean. We apply the LB order to compare experiments in binary-action decision problems and in decision problems with quasi-concave payoffs, as analyzed by Kolotilin, Corrao, and Wolitzky (2025).

The LB order is noted as a broader measure of informativeness than the Blackwell order, allowing for experiments that are LB-dominant but not Blackwell-dominant under small noise levels. p. 3
Where the paper says it

Furthermore, when n = 1 the Blackwell order is relatively easy to verify, because the set of univariate convex functions admits a simple characterization. However, when n ≥ 2, the Blackwell order is far more complex, as noted by Blackwell and Girshick (1954): in general, it is computationally infeasible to check whether one experiment dominates another, and many pairs of experiments are not even comparable.

The economic applications demonstrate the utility of this ordering across several domains:

  • Decision Problems: The LB order captures relative preferability in decision problems. For binary-action problems, F LB G if and only if the decision-maker’s ex ante expected payoff is higher under F than under G in every such problem. p. 4

    Where the paper says it

    We show that F dominates G in the LB order if and only if, in every decision problem with |A| = 2 (i. , with a binary action space), the DM’s ex ante expected payoff is higher under F than under G.

    This extends to all decision problems where the payoff is quasi-concave in a given belief. p. 4
    Where the paper says it

    , with a binary action space), the DM’s ex ante expected payoff is higher under F than under G. This result extends to all decision problems such that (i) A ⊆ R, and (ii) the DM’s expected payoff is quasi-concave in a under each belief.

  • Moral Hazard Problems: Dominance in LB order corresponds to both a wider range of implementable target actions and a smaller minimal disutility incurred by the principal when implementing any target action, implying greater flexibility (breadth of control) and greater effectiveness (depth of control). Bounds on penalties and rewards are established, showing that if F LB G and a uniformly bounded scheme under G implements delta*, there exists a scheme under F implementing delta* at lower disutility for the principal. p. 7

    Where the paper says it

    , Holmström, 1979, Gjesdal, 1982, Kim, 1995, Jewitt, 1997, Jewitt, 2007, Dewatripont, Jewitt, and Tirole, 1999, Fagart and Sinclair-Desgagné, 2007, Xie, 2017, and Frick, Iijima, and Ishii, 2023). We establish a new informativeness criterion based on the set of implementable actions and show that it is equivalent to the established criterion based on the principal’s minimal disutility from implementing her target action.

  • Screening Problems: In screening problems with ex post signals, F LB G if and only if the principal’s expected utility from her optimal mechanism is higher under F than under G for every screening environment E, indicating that greater informativeness in the LB sense leads to a higher payoff guarantee for the investor.

The analysis extends these results to continuous state spaces, confirming that binary-action and ex ante payoff comparisons hold via Lorenz-zonoid inclusion, and that principal–agent results in moral hazard and screening problems extend analogously by using separating hyperplane arguments generalized to positive Borel measures. The paper also introduces the monotone-posterior-expectation (MPE) order, which coincides with the Lehmann order under MLRP but remains applicable when MLRP fails, and establishes its equivalence to LB dominance for state space relabelings. p. 7

Where the paper says it

In the supplementary material for this paper, we introduce an extension of the Lehmann order called the monotone-posteriorexpectation (MPE) order. The MPE order coincides with the Lehmann order for experiments satisfying MLRP but also applies when MLRP does not hold.

Numerical methods are provided for verification, reducing the check to verifying zonotope inclusion. p. 4
Where the paper says it

5 contrast, the LB order is straightforward to work with for all n, since it relies on the linear convex order, which involves only univariate convex functions. In particular, the LB order of two experiments is fairly easy to check either analytically or through well-established numerical methods.

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. Enhanced experiment ranking for decision problems by using linear-Blackwell order (LB) instead of Blackwell order. This allows for verification via a simpler characterization involving only univariate convex functions, making it computationally feasible for multidimensional state spaces (n > 1).

  2. Improved moral hazard problem analysis through the LB order by capturing both the breadth and depth of principal control. This means dominance in the LB order corresponds to both a wider range of implementable target actions and a smaller minimal disutility to the principal from implementing any target action.

  3. Application of decision theory under uncertainty to improve investment and procurement systems. The paper establishes that F dominates G in the LB order if and only if V F (q) ≥ V G(q) for every QCC problem and every q in Delta b n, providing a robust ranking criterion across various economic scenarios.

  4. Development of screening mechanisms with ex post signals by using the LB order to compare experiment informativeness. Theorem 3 shows that F dominates G in the LB order if and only if WF (E) ≥ WG(E) for every screening environment E in E, which directly relates experiment quality to the principal's expected utility from her optimal mechanism.

  5. Generalization of information orderings to continuous state spaces using the linear-Blackwell order definition via posterior expectation characterization. This extends the ranking criterion to situations where "F dominates G if and only if for every prior q in P(Theta) and every phi in Bb(Theta), Z Theta phi Piq F (dtheta X) >= Z Theta phi Piq G (dtheta Y)."

  6. Robust verification of the LB order using geometric characterization via the Lorenz zonoid. The condition F dominates G if and only if Z(F) contains Z(G) provides a tractable, geometric tool for comparing experiments, which can be verified numerically by checking the inequality in Appendix E.

  7. Creation of product and mixture experiment comparison rules for the LB order. Proposition 6 establishes that For any experiments F1, F2, G1, G2 with F1 ⪰LB G1 and F2 ⪰LB G2, we have F1 oproducts LBLG1 ⊗ LBLG2, simplifying the comparison of complex information structures.

  8. Establishing equivalence between LB order and Blackwell order under irredundancy for finite signal spaces. Proposition 8 states that If F is irredundant and F ⪰LB G, then F ⪰B G, meaning for these specific experiments, the two orders yield the same ranking result.

Abstract

This paper introduces a novel ranking of statistical experiments, the linear-Blackwell (LB) order, which can equivalently be characterized by (i) the dispersion of the induced posterior and likelihood ratios in the sense of the linear convex order, (ii) the size of the Lorenz zonoid (the set of statewise expectation profiles), or (iii) the variability of the posterior mean. We show that the LB order provides a necessary and sufficient criterion for comparing experiments uniformly across decision problems with quasi-concave payoffs and principal-agent problems with separable payoffs.

Related papers