📈 economics

Blackwell-Monotone Updating Rules

This paper characterizes Blackwell-monotone updating rules, demonstrating that within a broad class of signal-independent distortions, Bayes' law is the unique strictly Blackwell-monotone rule, while non-paternalistic evaluation restricts such rules to affine distortions of Bayesian posteriors.

Original authors: Mark Whitmeyer

Published 2026-04-17

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Mark Whitmeyer

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a chef trying to cook the perfect meal. You have a recipe (your prior belief) and you get to taste the ingredients as you go (the information).

In the world of economics, there is a famous rule called Bayes' Law. It's like the "Gold Standard" of cooking: it tells you exactly how to adjust your recipe based on the taste of the ingredients. If you follow Bayes' Law perfectly, you are guaranteed that getting more information (tasting more ingredients) will never hurt your final dish. In fact, it will always help you, or at the very least, it won't make things worse. This is what the paper calls Blackwell Monotonicity: "More info is always better."

But what if you aren't a perfect Bayesian chef? What if you have bad habits? Maybe you overreact to a pinch of salt (thinking the whole dish is too salty), or maybe you underreact (ignoring the salt entirely).

This paper asks a simple but profound question: If you have any bad habit in how you process information, is there always a situation where getting more information actually makes you worse off?

The author, Mark Whitmeyer, says: Yes.

Here is the breakdown of his findings using some everyday analogies.

1. The "Strict" Rule: The Only Perfect Chef is the Bayesian

The paper defines a "Strictly Blackwell Monotone" rule as one where:

More information is always better than less.
There is always at least one situation where more information gives you a strictly better result.

The Big Discovery: If you want to be a chef who never regrets getting more information, you must be a Bayesian. You cannot have a "systematic distortion."

The Analogy: Imagine you have a rule where you always add 10% extra salt to whatever the recipe says.
- If the recipe says "add a pinch," you add a teaspoon.
- If the recipe says "add a cup," you add a bucket.
- The paper proves that with this kind of rigid, systematic error, there will always be a specific dish where adding that extra salt ruins it, whereas a perfect chef would have been fine.
- Conclusion: If you want to be safe in every scenario, you must follow the math perfectly (Bayes' Law). There is no "almost right" way to update your beliefs that works 100% of the time.

2. The "Paternalistic" vs. "Non-Paternalistic" View

The paper looks at this problem from two different perspectives, like looking at a student's test score.

Perspective A: The Strict Teacher (Paternalistic)

The View: "I am grading you based on the correct answer key (the true reality), not what you thought the answer was."
The Result: If you distort your beliefs (even slightly), you will eventually make a mistake that costs you.
- Example: You think a stock is safe because you ignored some bad news (underreaction). The teacher says, "You lost money because you didn't see the crash coming."
- The Verdict: Under this strict view, only perfect Bayesian updating is safe. Any distortion leads to a scenario where you are worse off than if you had known nothing at all.

Perspective B: The Self-Confident Student (Non-Paternalistic)

The View: "I am grading you based on what you believed at the time. If you made a choice that was logical given your beliefs, you get a good score."
The Result: Here, you can be wrong about reality but still "win" in your own mind.
- The Catch: Even here, you can't just distort your beliefs randomly. To be "Strictly Blackwell Monotone" (always happy to get more info), your distortion must be Affine.
- What is an Affine Distortion? Imagine you have a ruler, but your ruler is stretched or shifted.
  - If the ruler says "5 inches," you read it as "6 inches."
  - If it says "10 inches," you read it as "11 inches."
  - The gap between numbers stays the same; you just shifted the whole scale.
- The Verdict: If your brain just "shifts" or "scales" reality by a constant amount (like a consistent bias), you will still always prefer more information from your own perspective. But if your distortion is messy or changes depending on the situation, you will eventually regret getting more info.

3. The "Underreaction" Trap

The paper spends a lot of time on Underreaction (ignoring new info).

The Common Myth: "Underreacting is safe! If I ignore bad news, I won't panic and sell my stocks. I'm safer than the guy who overreacts."
The Paper's Twist: While underreacting might save you from panic, it violates the "Strict" rule.
- The Metaphor: Imagine you are driving and you see a sign saying "Road Closed."
  - Overreacting: You swerve into a tree. (Disaster).
  - Underreacting: You keep driving, thinking the sign is a joke. You don't crash immediately, but you miss the turn you should have taken. You arrive at the destination slower than you could have.
- The Point: Blackwell Monotonicity isn't just about avoiding disaster; it's about maximizing potential. If you underreact, you fail to take advantage of the extra information. You leave "free money" on the table. Therefore, you are not "Strictly" monotone.

4. The "Systematic" vs. "Random" Distortion

The paper also looks at whether your bad habits depend on the type of information you get.

Systematic Distortion: You always distort beliefs the same way, no matter the experiment. (e.g., "I always think probabilities are 10% higher than they are.")
Non-Systematic: Your distortion changes depending on the experiment.
The Finding: Even if you drop the "systematic" rule and allow for complex, changing distortions, the paper finds that if you want to be "Strictly Blackwell Monotone," you still have to be a Bayesian—unless you have very specific, weird properties (like being "Convex" or "Focused").
- The "Focused" Rule: If you treat identical-looking signals the same way (e.g., if you see a red light in Experiment A and a red light in Experiment B, you react the same), you are "Focused."
- The Result: Even if you are "Focused" and "Grounded" (you start with the right prior), the only way to guarantee you always love more information is to be a perfect Bayesian.

Summary: The Takeaway for Everyday Life

The "Perfect" Standard: If you want to be 100% sure that gathering more data will never hurt you (and will always help you find the best solution), you must update your beliefs exactly according to the math (Bayes' Law).
The Danger of Habits: If you have a consistent bias (like always being too optimistic or too pessimistic), there is always a specific situation where that bias will cause you to make a worse decision than if you had just stayed ignorant.
The "Safe" Bias: The only time you can have a bias and still be happy to get more info is if your bias is a simple, consistent shift (like a broken scale that always adds 5 pounds). But even then, you are only happy from your perspective, not necessarily the "real" perspective.
Underreaction is a Trap: Just because you aren't panicking (overreacting) doesn't mean you are doing well. If you are too slow to update your beliefs, you are missing out on opportunities.

In short: The universe of "good" information processing is very small. If you want to be strictly better off with more information, you have to be a perfect Bayesian. Any other way of thinking leaves you vulnerable to a scenario where "less is more."

1. Problem Statement

The paper investigates the relationship between Blackwell monotonicity and Bayesian updating.

Background: In standard decision theory, Blackwell's theorem establishes that for a Bayesian expected-utility maximizer, more informative experiments (in the Blackwell order) always yield higher welfare, regardless of the decision problem. This property is termed Blackwell monotonicity.
The Converse Question: The paper asks the converse: If an agent's updating rule satisfies Blackwell monotonicity, must the agent be Bayesian?
The Ambiguity: For non-Bayesian agents, the definition of "welfare" is ambiguous. Should welfare be evaluated:
1. Paternalistically: Using the true (Bayesian) posterior to calculate expected utility, even if the agent acts on a distorted belief?
2. Non-paternalistically: Using the agent's own (distorted) posterior to calculate expected utility?
Core Objective: To characterize the class of updating rules that satisfy Blackwell monotonicity under both evaluation criteria and to determine if these rules uniquely identify Bayes' law.

2. Methodology and Framework

The author employs a formal decision-theoretic framework involving:

Setup: A finite set of states $\Theta$ , a prior $\mu_0$ , and a set of statistical experiments $\Pi$ .
Updating Rules: Defined as a map $U$ $U$ from an experiment to a set of posteriors. The paper focuses on rules where the agent's posterior $\hat{\mu}$ $\overset{μ}{^}$ is a distortion $\phi$ $ϕ$ of the Bayesian posterior $\mu$ $μ$ .
- Systematic Distortions: $\hat{\mu} = \phi(\mu)$ , where $\phi$ is independent of the specific experiment (depends only on the Bayesian posterior).
- General Distortions: $\hat{\mu}_s = \phi_\pi(\mu_s)$ , where the distortion may depend on the experiment $\pi$ .
Definitions of Monotonicity:
- Blackwell Monotone: More informative experiments yield weakly higher welfare.
- Strictly Blackwell Monotone: More informative experiments yield strictly higher welfare in at least one decision problem.
- Paternalistic vs. Non-paternalistic: Determined by whether the welfare function $W$ uses the true belief $\mu$ or the distorted belief $\phi(\mu)$ .
Key Concepts Introduced:
- Grounded: The updating rule preserves the prior when no information is received ( $\phi(\mu_0) = \mu_0$ ).
- Convex: An agent prefers randomizing between two experiments to observing their convex combination (pooling).
- Focused: Identical signal realizations (in terms of likelihood ratios) induce identical beliefs, regardless of the experiment context.
- Extreme-Belief Aversion: A specific distortion pattern (relevant for binary states) where extreme beliefs are "clamped" to intermediate values.

3. Key Contributions and Results

A. Systematic Distortions (The Main Class)

The paper first restricts attention to rules where distortions are systematic (experiment-independent).

1. Paternalistic Evaluation (Theorem 3.1)

Result: Within the class of systematically distorting rules, Bayes' law is the unique strictly Blackwell-monotone updating rule.
Implication: Any systematic deviation from Bayes' law (e.g., overreaction, underreaction, or conservative updating) leads to a violation of Blackwell monotonicity. There will always exist a decision problem where the agent is strictly worse off with more information than with less.
Nuance: If the "strict" requirement is dropped, non-Bayesian rules can be Blackwell monotone (e.g., extreme-belief aversion in binary state spaces), but they fail the strict condition because they cannot exploit all incremental information.

2. Non-Paternalistic Evaluation (Theorem 3.8)

Result: An updating rule is strictly non-paternalistically Blackwell monotone if and only if the distortion is affine ( $\hat{\mu} = A\mu + b$ ).
Implication: Unlike the paternalistic case, non-Bayesian agents can satisfy non-paternalistic monotonicity. However, their distortions must be linear transformations (scaling and translation) of the Bayesian posterior. This ensures that the agent's perceived value of information remains positive and convex.

B. Beyond Systematic Distortions

The paper relaxes the assumption of systematic distortions to explore broader classes of rules.

1. Convex and Grounded Rules (Proposition 4.3)

Result: If an updating rule is both grounded (correctly updates on no info) and convex (prefers randomization over pooling), then it is strictly Blackwell monotone if and only if it is Bayes' law.
Significance: This shows that even without assuming systematic distortions, the combination of groundedness and convexity forces the agent to be Bayesian to maintain strict monotonicity.

2. Focused and Grounded Rules (Proposition 4.4)

Result: If an updating rule is grounded and focused (consistent interpretation of identical signals), it is strictly Blackwell monotone if and only if it is Bayes' law.
Significance: This extends the uniqueness of Bayes' law to a broader class of cognitive models that do not require the agent to understand all probabilistic ratios (systematic distortions), only that identical signals yield identical beliefs.

3. Existence of Non-Bayesian Strictly Monotone Rules (Section 4.1)

Counter-example: The paper constructs a specific updating rule that is grounded and strictly Blackwell monotone but is not Bayes' law.
Mechanism: This rule uses an experiment-dependent weight to clamp extreme beliefs (a form of "experiment-dependent extreme-belief aversion").
Takeaway: This demonstrates that groundedness alone is insufficient to pin down Bayes' law. The additional structural assumptions (convexity, focus, or systematic distortion) are necessary to recover the uniqueness of Bayesian updating.

4. Technical Intuition and Mechanisms

Overreaction vs. Underreaction:
- Overreaction: Leads to "Type I errors" (taking extreme actions based on noise). This can make information strictly worse than no information, violating monotonicity.
- Underreaction: Leads to "Type II errors" (failing to act on good news). While underreaction never makes an agent worse off than having no information (paternalistically), it violates strict monotonicity because the agent fails to fully exploit the value of additional information.
The Role of Strictness: The "strict" modifier is crucial. Without it, rules like extreme-belief aversion (which clamp beliefs) can be monotone. With it, any failure to perfectly track the Bayesian posterior (or an affine transformation thereof) creates a gap where more information yields no additional value, violating the strict condition.

5. Significance and Implications

Characterization of Rationality: The paper provides a powerful characterization result: Strict Blackwell monotonicity is essentially equivalent to Bayesianism (or affine distortions in the non-paternalistic case). This suggests that if an agent consistently values more information and can exploit it, they must be updating in a Bayesian (or linear) manner.
Evaluation of Biases: The results imply that common non-Bayesian biases (e.g., confirmation bias, conservative updating, overconfidence) are inherently "suboptimal" in the Blackwell sense. They will inevitably lead to scenarios where the agent would have been better off with less information (paternalistically) or failed to gain the full value of information.
Policy and Design: For mechanism designers or principals evaluating agents, the paper suggests that checking for Blackwell monotonicity is a robust test for Bayesian rationality. If an agent's updating rule is not Bayesian (and not affine), there exists a decision problem where their information processing is detrimental.
Welfare Analysis: The distinction between paternalistic and non-paternalistic evaluation highlights that non-Bayesian agents may feel they benefit from information (non-paternalistic view) even when they are objectively making suboptimal decisions (paternalistic view).

In summary, Whitmeyer demonstrates that Bayes' law is the unique updating rule that guarantees strict Blackwell monotonicity across a wide range of structural assumptions, reinforcing the normative power of Bayesian updating in decision theory.