On the Monotonicity of Information Costs

Imagine you are a detective trying to solve a mystery. You have a choice: you can buy a cheap, blurry snapshot of the crime scene, or you can pay a fortune for a high-definition, 3D video with audio.

In economics, we assume that better information should cost more. This seems obvious, right? But in the world of complex math and decision theory, "better" is a tricky word. Does a blurry photo count as "better" if it helps you catch a thief in a specific situation but fails in another?

This paper, titled "On the Monotonicity of Information Costs," by Xiaoyu Cheng and Yonggyun Kim, tackles a fundamental question: How do we mathematically guarantee that a cost function (a price tag for information) actually charges more for better info?

Here is the breakdown using simple analogies.

1. The Two Rules of "Better"

The authors look at two different ways to define what "better information" means. Think of these as two different rulebooks for a game.

The Blackwell Rulebook (The Universal Judge):
Imagine a judge who says, "Experiment A is better than Experiment B only if A helps you win every single possible game you could play."
- The Problem: This rule is incredibly strict. It's like saying a Ferrari is only "better" than a bicycle if it's faster at driving on the highway, and faster at climbing a mountain, and better at swimming, and better at cooking dinner. Because no car can do everything better than a bike in every scenario, the judge often says, "I can't compare them." They are "incomparable."
- The Cost: If you follow this rulebook, your price tag must go up whenever an experiment helps in any possible scenario.
The Lehmann Rulebook (The Specialized Judge):
This judge is more practical. They say, "Experiment A is better than B if it helps you win in specific, logical situations where things move in one direction."
- The Scenario: Think of an auction. If you know the item is worth more, you bid higher. If you know it's worth less, you bid lower. This is a "monotone" situation (things move in one direction).
- The Advantage: The Lehmann judge can compare experiments that the Blackwell judge couldn't. It's like saying, "For the specific game of 'Auction,' this high-def video is definitely better than the blurry photo, even if the photo is better for 'Swimming.'"

2. The Big Discovery: The "Price Tag" Test

The authors wanted to know: What rules must a price tag follow to be fair under these two rulebooks?

They discovered that you don't need to check every single possible experiment to see if the price is fair. You only need to check tiny, local changes. They call this "Decreasing in Signal Replacement."

The Analogy: The "Signal Swap"
Imagine your information is a set of clues.

Blackwell Test: If you take a clue that is "High Quality" and swap it for a "Low Quality" clue (making the whole set worse), the price must go down. If the price stays the same or goes up, your pricing model is broken.
Lehmann Test: This is trickier. It's not just about swapping any clue. It's about swapping clues in a specific order.
- Imagine you have a list of clues ranked from "Low State" (bad news) to "High State" (good news).
- The Lehmann rule says: If you take a "High State" clue and accidentally swap it with a "Low State" clue only when the news is bad, the price must drop.
- Why? Because in monotone problems (like auctions), mixing up the order of clues confuses the decision-maker specifically in the way that matters most.

3. The "Path" Problem (The Hard Part)

The authors had to prove that if your price tag passes these tiny, local tests, it will pass the big, global test.

The Blackwell Path: Imagine you have a very informative map. You want to turn it into a less informative map by slowly adding fog. The authors proved you can do this by adding fog in small, step-by-step chunks. If your price drops every time you add a little fog, the total price will be correct.
The Lehmann Path (The Mountain Climb): This was the hard part. The set of "good" experiments (those that follow the Lehmann rules) is not a smooth hill; it's a jagged, non-convex mountain. You can't just walk in a straight line from a good map to a bad map without falling off the mountain (violating the rules).
- The Solution: The authors built a special "rope bridge" (a mathematical path) that stays strictly on the mountain ridge. They showed that you can transform a complex experiment into a simpler one by making tiny, specific swaps that keep the "logic" of the experiment intact while lowering the quality. If the price drops on every step of this bridge, the pricing model is valid.

4. Testing Famous Price Tags

Finally, the authors took famous "price tags" used by economists and tested them against their new rules.

The "Entropy" Cost (The Standard): This is the most common way to price information (based on uncertainty). They confirmed it works for both the Blackwell and Lehmann judges. It's a fair price tag.
The "Bregman" Cost (The New Kid): This is a newer, more flexible pricing model used in machine learning and advanced economics.
- The Verdict: The authors found that this model fails the test. It sometimes charges more for information that is actually worse (or charges the same for different levels of quality).
- Why? It's like a store that charges you more for a slightly damaged apple than a perfect one because of how they calculated the price. The authors showed exactly why it fails: it doesn't respect the "Signal Swap" rule.

Summary: Why Does This Matter?

In the real world, economists and AI researchers build models where agents (people or computers) "buy" information to make decisions.

If the "price tag" they use is flawed—meaning it doesn't strictly charge more for better info—the whole model breaks. The agents might make irrational choices, or the math might predict things that never happen in reality.

The Takeaway:
This paper gives economists a simple checklist (the "Signal Swap" test) to verify if their information pricing models are fair.

If you want a universal price tag: Make sure the price drops whenever you swap a good clue for a bad one.
If you want a specialized price tag (for auctions, markets, etc.): Make sure the price drops when you swap clues in a specific, ordered way.

By providing these simple "local" tests, the authors made it much easier to build robust economic models that accurately reflect how we value information.

Here is a detailed technical summary of the paper "On the Monotonicity of Information Costs" by Xiaoyu Cheng and Yonggyun Kim.

1. Problem Statement

The paper addresses a fundamental requirement in models of costly information acquisition (Rational Inattention): Monotonicity. The principle states that more informative experiments must be more costly. However, "informativeness" is not a scalar quantity but is defined via partial orders.

The authors investigate the conditions under which a cost function $C$ is monotone with respect to two standard information orders:

Blackwell Order ( $\succeq_B$ ): An experiment is more informative if it yields higher expected payoffs in all decision problems.
Lehmann Order ( $\succeq_L$ ): An experiment is more informative if it yields higher expected payoffs in monotone decision problems (e.g., auctions, screening). This order is a refinement of the Blackwell order, applicable when experiments satisfy the Monotone Likelihood Ratio Property (MLRP).

The Gap: While Blackwell monotonicity has been studied extensively, a general characterization of Lehmann monotonicity has been missing. Furthermore, existing characterizations of Blackwell monotonicity often rely on global properties or specific functional forms. The authors aim to provide a unified, local, first-order characterization for both orders.

2. Methodology

The authors employ a geometric and path-construction approach grounded in the garbling representations of information orders.

Garbling Characterization:
- Blackwell: $f \succeq_B g$ if $g$ is a garbling (noisy version) of $f$ .
- Lehmann: For MLRP experiments, $f \succeq_L g$ if $g$ is a garbling of $f$ via "reverse monotone" noise (Kim, 2023).
Local First-Order Conditions: Instead of checking global inequalities, the authors derive necessary and sufficient conditions based on directional derivatives (gradients) of the cost function. They analyze how the cost changes under infinitesimal perturbations that reduce informativeness.
Path Construction: To prove sufficiency, they construct continuous paths between two comparable experiments ( $f \succeq g$ $f ⪰ g$ ) such that the cost decreases monotonically along the path.
- Challenge: The set of MLRP experiments is non-convex. Standard convex combination paths often leave the MLRP set.
- Solution: They utilize specific geometric representations (Zonotopes for Blackwell; Probability-Probability (PP) plots for Lehmann) to construct paths that remain strictly within the MLRP set while reducing informativeness.

3. Key Contributions

A. Characterization of Blackwell Monotonicity

For a cost function $C$ defined on experiments, the authors establish that $C$ is Blackwell monotone if and only if:

Permutation Invariance: $C(f) = C(fP)$ for any signal permutation $P$ . (Relabeling signals does not change information).
Split Invariance: $C(f) = C(f')$ if a signal is split into two copies. (Splitting does not change information).
Decreasing in Signal Replacement: The directional derivative of $C$ $C$ in the direction of replacing signal $j$ $j$ with signal $k$ $k$ is non-positive.
- Intuition: If you replace a signal with another (making the experiment noisier), the cost must not increase.

B. Characterization of Lehmann Monotonicity (Primary Contribution)

This is the paper's novel result. For cost functions restricted to MLRP experiments, $C$ is Lehmann monotone if and only if:

Split Invariance: (Same as above).
Decreasing in Reverse Signal Replacement: The directional derivative is non-positive when replacing a signal $j$ $j$ with its neighbor $j+1$ $j + 1$ (or $j-1$ $j - 1$ ) in a state-dependent manner.
- Specifically, replacing a low signal with a high signal is only allowed in low states, and replacing a high signal with a low signal is only allowed in high states.
- Significance: This condition is strictly stronger than Blackwell monotonicity. It requires the cost function to respect the specific structure of monotone decision problems.

C. Unified Framework

The paper unifies the analysis of both orders by showing they share a common structure: global monotonicity is equivalent to local monotonicity along specific "garbling" directions, provided the cost function satisfies invariance properties (permutation/split) and continuity.

4. Key Results and Applications

Theoretical Results

Theorem 1 & 3 (Blackwell): Proved that local "decreasing in signal replacement" combined with invariance is sufficient for global Blackwell monotonicity.
Theorem 2 & 4 (Lehmann): Proved that local "decreasing in reverse signal replacement" combined with split invariance is sufficient for global Lehmann monotonicity.
Path Construction: The authors developed a novel method to decompose arbitrary information loss into continuous paths of local replacements that preserve the MLRP, overcoming the non-convexity of the MLRP set.

Application to Existing Cost Functions

The authors apply their conditions to well-known cost classes:

Likelihood-Separable Costs: $C(f) = \sum \psi(f^j) - \psi(1)$ $C (f) = \sum ψ (f^{j}) - ψ (1)$ .
- Blackwell monotonicity holds if $\psi$ is sublinear.
- Lehmann monotonicity requires additional conditions on the derivatives of $\psi$ (Proposition 2 & 3). Not all sublinear functions satisfy Lehmann monotonicity.
Posterior-Separable Costs: $C(f) = H(\mu) - \sum \tau^j H(q^j)$ $C (f) = H (μ) - \sum τ^{j} H (q^{j})$ .
- Blackwell monotonicity holds if $H$ is concave.
- Lehmann monotonicity holds if $H$ is concave and satisfies specific derivative inequalities related to First-Order Stochastic Dominance (FOSD) (Proposition 4).
- Result: The Entropy Cost (Sims, 2003) is shown to be Lehmann monotone.
Bregman Information Costs:
- These costs (generalizing entropy) are often not permutation invariant.
- The authors show that for nested logit specifications, Bregman costs can violate the "decreasing in signal replacement" condition locally, meaning they are neither Blackwell nor Lehmann monotone. This challenges their universal applicability in monotone decision problems.
State-wise Divergence Costs:
- Costs based on statistical divergences (KL, Rényi) satisfy monotonicity if the divergence satisfies the Data-Processing Inequality.

5. Significance

Foundational Clarity: The paper isolates monotonicity as a standalone property, providing a "first-order" test for any proposed cost function without needing to derive a full functional form from axioms.
Lehmann Monotonicity: It fills a critical gap in the literature by providing the first general necessary and sufficient conditions for Lehmann monotonicity, which is crucial for applications in auctions, screening, and finance where monotone decision problems are prevalent.
Practical Utility: The conditions are computationally tractable (checking derivatives) rather than requiring global optimization over all decision problems.
Critique of Existing Models: By showing that popular Bregman costs may fail monotonicity even locally, the paper warns researchers to be cautious when applying these costs to monotone decision environments.
Geometric Insight: The use of PP-plots and zonotopes to navigate the non-convex MLRP set offers a new toolkit for analyzing statistical experiments.

In summary, Cheng and Kim provide a rigorous, unified, and geometrically intuitive framework for verifying whether information costs behave logically under standard definitions of "more information," with a specific and novel focus on the economically vital Lehmann order.