On the Monotonicity of Information Measures

summary

The gist

On the Monotonicity of Information Costs studies the monotonicity of information costs by providing necessary and sufficient conditions grounded in Blackwell and Lehmann information orders to…

In short

The study investigates when information costs are monotonic with respect to two key orders: Blackwell and Lehmann. The research establishes that this monotonicity can be characterized by simple, local, first-order conditions. It shows that a measure is Blackwell monotone if it decreases when replacing signals, and Lehmann monotone if it decreases under specific reverse signal replacement rules.

Key concepts

Blackwell Order
This order compares experiments based on whether one experiment dominates another across all possible decision problems. It represents a strong criterion for determining which experiment provides more informative results overall.
Lehmann Order
This order is used to compare experiments in contexts like auctions and screening problems where the comparison depends on specific monotone decision problems. It focuses on a broader class of comparisons than the Blackwell order.
Decreasing in Signal Replacement
For Blackwell monotonicity, this means that if you replace one signal with another, the associated cost must decrease. This condition ensures that more informative experiments are not always more costly.
Decreasing in Reverse Signal Replacement
For Lehmann monotonicity, this involves a more complex condition where the cost decreases when replacing signals based on how those signals relate to high or low states. It requires noise to be structured in a specific way.

Terminology used across episodes

This episode discusses

The paper

On the Monotonicity of Information Measures · Read on arXiv

Xiaoyu Cheng, Yonggyun Kim

Florida State University

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: "On the Monotonicity of Information Measures".

Warren: Here's the gist of the paper. On the Monotonicity of Information Costs studies the monotonicity of information costs by providing necessary and sufficient conditions grounded in Blackwell and Lehmann information orders to characterize when more informative experiments must entail higher costs The primary contribution is a unified framework for characterizing the monotonicity of information costs with respect to both the Blackwell and Lehmann orders using first-order conditions Foundational Concepts The paper introduces two classical information orders, the Blackwell order and the Lehmann order, as criteria for comparing experiment informativeness The Blackwell order requires one experiment to dominate another across all decision problems, while the Lehmann order focuses on a broader class of monotone decision problems such as auctions and screening problems The authors develop a unified framework that builds on the garbling representations of these two orders Characterization via Garbling The characterization for Blackwell monotonicity is based on the concept of "decreasing in signal replacement," which requires the cost function to decrease when one signal is replaced by another with some probability For Lehmann monotonicity, a similar condition is established through "decreasing in reverse signal replacement," which involves noise that is more likely to generate higher signals in lower states and lower signals in higher states The paper shows that these local conditions are sufficient for global monotonicity under additional natural conditions regulating the cost of equally informative experiments Binary Experiment Results For binary experiments, the Blackwell order is characterized by permutation invariance and decreasing in signal replacement The Lehmann order on MLRP experiments is characterized by a set of inequalities involving adjacent signals, which are equivalent to the Blackwell order in this specific case Finite Experiment Generalization The characterizations are extended to finite experiments with any number of signals, where Blackwell monotonicity requires permutation invariance, split invariance, and decreasing in signal replacement For Lehmann monotonicity over MLRP experiments, the condition is split invariance and decreasing in reverse signal replacement The proof sketch for Lehmann monotonicity involves constructing a continuous path of experiments that preserves the MLRP while decreasing the cost Application to Cost Functions The framework is applied to several cost functions, including likelihood-separable and posterior-separable costs It is shown that for likelihood separable costs, Blackwell monotonicity holds if and only if the function defining the cost is sublinear For posterior separable costs, a sufficient condition for Lehmann monotonicity involves conditions on the derivatives of the uncertainty measure H The entropy cost is shown to be Lehmann monotone However, Bregman information costs can violate both Blackwell and Lehmann monotonicity under local perturbations The paper concludes that the principle of monotonicity requires simple, local conditions for cost functions rather than complex functional forms This provides a more transparent and workable foundation for analyzing information costs in diverse economic environments The study strengthens the foundations for robust applications of costly information acquisition in economics Appendix Details The appendix provides proofs for the absolute continuity assumption and details the construction of continuous paths using directional derivatives The proof sketch for Theorem 3 relies on embedding experiments into a higher-dimensional space to construct a cost-decreasing path Lemma B.2 provides the geometric characterization of the Lehmann order in terms of cumulative distribution functions The proof for Theorem 4 demonstrates how to construct a sequence of MLRP experiments to show that cost decreases along the path The discussion on Blackwell monotonicity under quasiconvexity shows that permutation invariance and decreasing in signal replacement are necessary and sufficient conditions when quasiconvexity is imposed The examples illustrate that not all cost functions satisfy the required monotonicity properties The paper covers various related literature on information costs, including likelihood separable and posterior separable costs The discussion on Bregman information costs highlights their potential to violate monotonicity under local perturbations The final section summarizes the findings by showing that the necessary conditions derived from the order are sufficient for cost monotonicity in various settings The paper provides a flexible toolkit for researchers to build, verify, and apply cost functions in diverse economic environments The study strengthens the foundations for robust applications of costly information acquisition in economics.

Paper summary: Ruth: So this paper we're looking at is called "On the Monotonicity of Information Measures". It's about information costs, which are basically how much it costs to get a certain piece of data or an experiment. The main idea they’re pushing here is that more informative experiments have to cost more.

Warren: Right, so we're talking about this fundamental requirement that if you learn something better, you should pay a higher price for it. The paper claims they give us simple necessary and sufficient conditions for when a cost function is monotone with respect to both the Blackwell and Lehmann orders.

Iris: That's the core contribution mentioned in the abstract: providing these first-order conditions that ground it in those two classical information orders, which are Blackwell and Lehmann. They aim to unify how we check this monotonicity using a common analytical structure, even though they build on different garbling representations of the two orders.

Ruth: Exactly. The paper suggests that we can characterize this behavior by looking at local, first-order conditions instead of trying to find some complicated global condition for every single cost function out there. It makes verification much more transparent for researchers in the field.

Warren: And they give specific ways to look at it depending on whether you care about Blackwell monotonicity or Lehmann monotonicity. For example, Blackwell monotonicity is characterized by something called "decreasing in signal replacement," which means the cost has to go down when you swap one signal for another with some probability.

Iris: That's a key mechanism they use for the Blackwell side. But then there's the Lehmann order, and that condition is phrased as "decreasing in reverse signal replacement." That involves noise that is more likely to generate higher signals when you are already in a lower state, or lower signals when you are in a high state.

Ruth: It sounds complicated, but the paper applies these ideas to specific types of measures, like likelihood-separable and posterior-separable costs. They show that for likelihood separable costs, Blackwell monotonicity holds if and only if the function defining that cost is sublinear.

Warren: And for posterior separable costs, they find a sufficient condition for Lehmann monotonicity involves looking at the derivatives of an uncertainty measure called H. They specifically show that the entropy cost is Lehmann monotone, which is pretty interesting because it's a very natural measure to start with.

Iris: That means we have a concrete example where we know it works, but then they point out that other types of costs, like Bregman information costs, can violate both Blackwell and Lehmann monotonicity under local perturbations. That’s an important caveat for anyone trying to use these conditions universally.

Ruth: It’s a useful warning because it shows that just because a cost function looks okay locally doesn't guarantee it respects these broader information orders globally. The paper emphasizes that the principle of monotonicity really requires these simple, local conditions rather than complex functional forms.

Warren: So to wrap up the summary: this paper provides a unified framework using first-order conditions derived from Blackwell and Lehmann orders to characterize when more informative experiments must entail higher costs, and it shows how these conditions apply to several important cost functions like likelihood separable ones.

Iris: And finally, they identify a specific subclass of posterior-separable measures called posterior-mean-ratio measures that characterizes the Lehmann order specifically. That gives us a clear identifier for that particular class of information costs within the broader set.

Ruth: So what this means for the world is that it gives researchers a more transparent toolkit to build, verify, and apply these cost functions in economic environments where agents are making costly information choices. It strengthens the foundations for analyzing costly information acquisition in economics generally.

Conclusion: Ruth: So we've been looking at how information costs behave when we add more data, and this paper is called "On the Monotonicity of Information Measures".

Warren: Right, they're figuring out the conditions for when those costs actually go up as you get better information. It boils down to these Blackwell and Lehmann orders.

Iris: They're using some first-order conditions to characterize that monotonicity. It’s not just a general idea; they’ve given us specific mathematical rules based on how experiments are replaced with other experiments.

Ruth: Exactly, so instead of looking for some huge global rule, they suggest we check these local conditions. It makes the whole analysis way more manageable for anyone trying to build or test these cost functions in economics.

Warren: And they apply this to things like likelihood separable and posterior separable costs, showing that entropy cost is one of those monotone ones. But there are still some tricky cases, like Bregman costs where things can break down under small changes.

Iris: That's the caveat they flag—that not every cost function will follow these rules everywhere. They also identify a specific subclass of posterior separable measures called posterior mean ratio measures that precisely defines the Lehmann order.

Ruth: It’s interesting because it gives us a clearer way to identify which costs belong to which order, moving beyond just naming them. This whole paper is about establishing robust rules for when more information demands more cost.

Warren: So if you're building a model where agents are making decisions based on imperfect knowledge, you need these monotonicity checks to make sure your cost structure makes economic sense.

Iris: It sets up the framework, but the next step is seeing exactly how these conditions play out when we actually run the experiments.

More episodes

← Home