Blackwell Informativeness Order¶

Compare experiments by whether one can be garbled into another, equivalently by universal Bayes decision value under standard finite assumptions.

Version: v1 · 2026-10-03 · History
Domain-specific #: 13018
Domain group: Formal Sciences
Origin domain: Mathematics
Subdomain: Statistical Decision Theory → Mathematics

Core Idea¶

The Blackwell informativeness order compares two statistical experiments over the same hidden state space by what their signals can enable a decision maker to do. In a standard finite formulation, experiment $P$ is at least as informative as experiment $Q$ if a state-independent stochastic post-processing of $P$'s signal produces $Q$'s state-conditional signal law: $Q=PM$. Blackwell's equivalence theorem says this is equivalent to $P$ providing at least as high an optimized ex ante expected payoff as $Q$ for every matched Bayesian decision problem in the theorem's class. The name in the frozen Wikipedia request is Blackwell's informativeness theorem; this entry admits the reusable comparison relation that the theorem characterizes, not merely a historical theorem recital.^[1]^[2]

An experiment is a channel from states to signals, not a single realized observation. In matrix notation rows index hidden states and columns index possible signals; $M$ acts after the $P$ signal and cannot inspect the hidden state afresh. That explains the easy direction: a decision maker seeing $P$ can simulate $Q$ and follow any rule available after $Q$, or ignore extra detail. The converse—that universal decision-value dominance supplies a garbling witness—is the substantive theorem, not a claim licensed by observing success in one task.^[1]

The relation is reflexive and transitive on raw experiment representations but need not be antisymmetric: relabeling signals, or splitting a signal into state-independent duplicate labels, gives different channels that can garble each other. It is a preorder there; after quotienting by mutual Blackwell equivalence, it becomes a partial order. Live prime Order's current definition requires antisymmetry, so this draft proposes Relation, not Order, as strict DAG parent.^[1]^[3]^[4]

Structural Signature¶

Sig role-phrases: common hidden states — pair of state-to-signal experiments — state-independent garbling — universal Bayes-value comparison — posterior contraction — equivalence-class qualification.

Common hidden states. $P$ and $Q$ must concern the same state space (or an explicitly specified translation). An apparently better signal about a different unknown does not enter this comparison.^[1]
Compared experiments. For each state, each experiment specifies a probability distribution over its own signal alphabet. The alphabets may differ; that is why a stochastic map between signals matters.^[1]^[2]
State-independent garbling. A stochastic matrix $M$ maps a $P$ signal to a $Q$ signal with probabilities that depend on the observed signal, not the unobserved state. $Q=PM$ is a structural witness for $P\succeq_B Q$ in the finite formulation.^[1]
Universal value test. Compare optimized expected utilities using common states, priors, feasible actions and payoffs while varying over the full stated decision-problem class. One problem's result is weaker than the universal quantifier.^[1]^[2]
Posterior contraction. Under a fixed common prior, each posterior induced by a garbled signal is a probability-weighted average of posteriors induced by original signals. The worse law of posteriors is consequently a mean-preserving contraction, not an arbitrary identical posterior law.^[1]^[5]
Equivalence-class qualification. Mutual garbling identifies experiments that may look different at the raw signal-label level; quotienting those representations is needed before calling the induced order antisymmetric.^[1]

These roles identify a comparison method, not one experiment. They remain available whenever new state-dependent measurement or information designs are to be compared under the theorem's assumptions.^[1]

What It Is Not¶

Not “more accurate” for one chosen question. One classifier may have higher accuracy at one threshold, or higher payoff for one action menu, without dominating across every allowed decision problem. Blackwell dominance is deliberately demanding and many pairs are incomparable.^[1]^[2]

Not any post-processing whatever. If a supposed garbling reads the hidden state after the first signal, it may create information unavailable from the signal and does not witness the relation. The stochastic kernel must act on the observed signal alone.^[1]

Not an unconditional claim that more information improves every strategic or costly outcome. Blackwell's Bayes decision maker can use, ignore or garble a free signal with an unchanged action menu. Signal costs, privacy constraints, incentives, forced actions or other strategic responses require additional assumptions. Cabrales et al.'s restricted investment entropy order ranks some cases differently and is compatible with—not identical to—the Blackwell benchmark.^[1]^[2]

Not literally a partial order on every raw channel representation. Two signal alphabets can encode the same informational experiment. Mutual garbling creates equivalence even when their matrices or labels differ; antisymmetry belongs to the quotient.^[1]

Scope of Application¶

In statistical decision theory, compare two measurement channels before choosing an action to predict or respond to an unknown state. The all-decision test says a better channel cannot reduce the best available Bayes payoff when the decision maker may disregard it. Garbling gives a constructive comparison certificate and explains exactly how a weaker measurement can be simulated from a stronger one.^[1]

In investment information economics, Cabrales, Gossner and Serrano specify finite payoff-relevant states, signal structures, a prior and feasible assets selected after information arrives. They describe Blackwell's garbling benchmark, then study a different complete entropy-based order for restricted ruin-averse, no-arbitrage investment decisions. That contrast is instructive: Blackwell dominance, if present, still licenses the full Bayes benchmark, but winning under their restricted investor class alone does not prove it.^[2]

In posterior-belief geometry, Karni–Safra's equation shows that after a garbled signal the posterior is a conditional average of better-signal posteriors. Yang and Yang describe the related order on fixed-prior posterior distributions using convex indirect value functions and martingale transitions. These are alternative representations under stated setups, not a license to equate all possible posterior distributions across distinct priors.^[1]^[5]

Clarity¶

Set the matrix orientation before using $Q=PM$: $P_{\theta s}=\Pr(s\mid\theta)$, $M_{st}=\Pr(t\mid s)$ and $Q_{\theta t}=\sum_s P_{\theta s}M_{st}$. A row of $M$ sums to one and has no $\theta$ index. This equation is the decisive structural test. A shared prior is needed for posterior and Bayes-value calculations; a fixed-prior posterior comparison should not be silently promoted to a claim over all priors.^[1]

State whether “order” refers to raw kernels or equivalence classes. For raw kernels, the natural relation is a preorder. One can split each original signal into two labels by an independent coin, then merge labels to recover the original; the split and unsplit channels are mutually garbling even though their alphabets differ. Quotienting this mutual dominance yields the antisymmetric object.^[1]

Manages Complexity¶

Universal Bayes-value dominance looks like an infinite family of payoff comparisons. In the finite theorem setting, a single stochastic garbling witness verifies all of them at once; conversely, failure to find such a witness reflects the theorem's exact, strong order. This compresses a family of decision problems into a relation between channels while retaining the universal scope rather than approximating it with one score.^[1]

The simplification is partial, not total. Two experiments may be incomparable, and a restricted setting can still have a meaningful preference between them. Cabrales et al. obtain a complete order by narrowing preference and investment classes; the resulting ranking is a different abstraction and cannot be substituted for the original universal test.^[2]

Abstract Reasoning¶

The garbling direction can be reasoned through as simulation. Suppose $P\succeq_B Q$ via $M$. After observing $s$ from $P$, draw $t$ from $M(\cdot\mid s)$ and take exactly the action a $Q$-observer would take after $t$. Because the induced law of $t$ conditional on each state equals $Q$, this policy duplicates the $Q$ payoff distribution. Optimizing after $P$ can do at least as well. The reverse direction is a theorem under its assumptions; it cannot be inferred from this simulation argument alone.^[1]

At fixed prior, the posterior after $t$ averages the posteriors after contributing $s$ values, weighted by their probabilities of producing $t$. Thus garbling contracts posterior dispersion in convex order. It does not mean the weaker experiment can reproduce every posterior distribution attainable under the stronger, or vice versa without specifying the stochastic coupling.^[1]^[5]

Knowledge Transfer¶

The roles transfer from a simple hidden-state prediction channel to financial signals because both have common states, state-conditional signal kernels, possible signal-only processing and downstream Bayes actions. What does not transfer automatically is a specific policy conclusion about information acquisition, regulation or incentives. Once the objective class changes, the order being tested may change.^[1]^[2]

Live prime Relation is a proposed strict parent: the pair of experiments is either in or out of the garbling/Bayes-dominance relation. Live prime Order is a related quotient-level pattern but is not a strict parent of the raw preorder under its current antisymmetry signature. Information and Comparison are conceptual neighbors, not exact typed coverage.^[3]^[4]^[1]

Examples¶

Binary-state observation. This is an explicit constructed illustration, not an empirical study. Let hidden state be $0$ or $1$. Experiment $P$ reports it perfectly. Experiment $Q$ reports it correctly with probability $0.75$ and flips it otherwise. A binary symmetric channel $M$ applied to $P$'s report yields $Q$ without state access. The common state space is $\{0,1\}$; the two state-to-signal kernels are $P,Q$; the universal value test is passed because any $Q$ action can be simulated after $P$; under a fixed prior, each $Q$ posterior averages the compatible perfect-signal posteriors. The experiments are not mutually equivalent when noise is nondegenerate. Mapped back: all six blueprint roles are filled, and the numeric value illustrates the theorem rather than adding a new hypothesis.^[1]

Investment information. Cabrales et al. specify finite market-payoff states, signal structures and an investor selecting an asset from a feasible set after seeing a signal. If one structure admits a signal-only garbling to another, the first can reproduce the second's signal-and-action policy. Those are the same state, experiment, garbling and Bayes-value roles as above; with common prior, the poorer posteriors are conditional averages, and mutually garbling structures belong to one information class. Their paper's separate restricted entropy order may rank more investor-specific pairs but is not itself a Blackwell witness. Mapped back: the Blackwell relation remains the universal comparison benchmark within this financial setting, not a claim that every investor preference or contract obeys it.^[2]^[1]

Structural Tensions¶

Preference-robust guarantee versus comparability. Requiring no loss for every matched Bayes decision problem yields a robust comparison but leaves many experiment pairs incomparable. Restricting to ruin-averse no-arbitrage investment decisions can rank more pairs, but sacrifices the universal Blackwell guarantee. Diagnostic: Is the desired judgment about every Bayes decision problem or only a declared class of investor objectives?^[1]^[2]

Retaining signal detail versus discarding it. A richer signal preserves distinctions that may matter to some future action; garbling simplifies or compresses the signal at the cost of those distinctions, though it cannot improve universal cost-free Bayes value. Practical storage, attention or privacy benefits of discarding detail belong to an expanded objective and cannot be smuggled into this theorem. Diagnostic: Is the comparison about information alone with unchanged action opportunity, or do resource and constraint costs need to be modeled separately?^[1]^[2]

Structural–Framed Character¶

Evaluative weight. “More informative” sounds evaluative, but the formal relation is a mathematically checkable weak dominance across a specified value class. Selecting a class of values is a modeling decision, not an unconditional claim that collecting data is good.^[1]^[2]

Human-practice dependence. Decision makers and statisticians choose states, signals, actions and priors; once those inputs are fixed, the stochastic matrix witness is not a survey of their opinions. The practical usefulness of one signal in a particular institution remains a separate question.^[1]

Institutional origin. Blackwell's named theorem arose in statistical decision theory rather than a policy or laboratory standard. The original 1953 full paper could not be inspected here, so historical detail is bounded to the publisher record and later author research.^[1]

Vocabulary travel. “Signal,” “garbling” and “value” appear in measurement and investment work, but the exact comparison travels only with the state-to-signal and universal-decision roles intact. Colloquial “better information” does not imply Blackwell dominance.^[1]^[2]

Import versus recognition. Applying the theorem to a new setting requires specifying comparable kernels and proving a state-independent garbling or universal value claim. It is an imported formal test, not something established by thematic resemblance to ordinary information loss.^[1]

Its character: structural but decision-theory-framed, with evaluative wording disciplined by a universal mathematical criterion and little institutional dependence; its exact stochastic/decision equivalence is domain-specific, whereas the broad pairwise-relation skeleton is already prime.^[3]^[1]

Structural Core vs. Domain Accent¶

The core is the equivalence between a channel comparison by signal-only garbling and a universal Bayes-value comparison under specified conditions. The state space, experiment kernels, downstream actions and universal quantifier are indispensable. The posterior averaging statement is a derived fixed-prior interpretation, not a replacement definition. A single empirical accuracy score or one investor's advantage does not instantiate the core.^[1]^[2]^[5]

The portable outer skeleton—a relation deciding whether one pair dominates another—is live prime Relation. The Blackwell-specific garbling kernel, Bayesian decision values and quotient-by-mutual-equivalence qualification supply the domain-specific core, not detachable jargon. The possibly broader idea “information never hurts if it can be ignored” is not promoted as a new prime by this draft; it is only one direction of the sourced theorem in its particular setting.^[3]^[1]

This entry is a kind of Relation.

Relation is the broader abstraction this entry instantiates. Comparison, Information and Decision are related prime vocabularies, but none alone supplies the state-to-signal/garbling equivalence. Order becomes appropriate after quotienting raw experiments by mutual Blackwell equivalence; the live Order prime currently requires antisymmetry, so it is not recorded as the raw relation's strict parent.^[3]^[4]^[1]

Relationships to Other Abstractions¶

Current abstraction Blackwell Informativeness Order Domain-specific

Parents (1) — more general patterns this builds on

Blackwell Informativeness Order is a kind of Relation Prime

Blackwell dominance is a specialized relation between two compatible statistical experiments.

Hierarchy path (1) — routes to 1 parentless root

Blackwell Informativeness Order → Relation

Neighborhood in Abstraction Space¶

Blackwell Informativeness Order sits in a sparse region of the domain-specific corpus (80^th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Unclustered & Miscellaneous (2551 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-10-08

Not to Be Confused With¶

Blackwell's 1953 theorem is the equivalence result characterizing this reusable order; the frozen requested name remains provenance for the reframe. Shannon mutual information or an entropy reduction gives numerical summaries, not an identical universal comparison; Cabrales et al.'s restricted investor entropy order is explicitly a different completion. Posterior convex order is a fixed-prior representation with a particular coupling, not permission to equate arbitrary posterior laws. Signal accuracy for one classification task, statistical sufficiency under another definition, and strategic information value with changed incentives require their own criteria.^[1]^[2]^[5]

References¶

[1] E. Karni and Z. Safra, “Hybrid Decision Model and the Ranking of Experiments”, author-hosted original research (2021), §§2.1–2.2 Definitions 1–2 and displayed theorem, §3.1 posterior derivation; inspected 2026-10-01. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l ↩m ↩n ↩o ↩p ↩q ↩r ↩s ↩t ↩u ↩v ↩w ↩x ↩y ↩z ↩27 ↩28 ↩29 ↩30 ↩31 ↩32 ↩33 ↩34 ↩35 ↩36 ↩37

[2] A. Cabrales, O. Gossner and R. Serrano, “Entropy and the Value of Information for Investors”, author-hosted original research (2011), Introduction and §2.3; inspected 2026-10-01. The article's own investor entropy order is not equated with Blackwell dominance. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l ↩m ↩n ↩o

[3] Encyclopedia of Abstractions, live prime Relation, Core Idea and Structural Signature, inspected 2026-10-01. registry ↩a ↩b ↩c ↩d ↩e

[4] Encyclopedia of Abstractions, live prime Order, Core Idea and Structural Signature, inspected 2026-10-01. registry ↩a ↩b ↩c

[5] F. Yang and K. H. Yang, “Stochastic Optimization and Coupling”, author-hosted original research (2026), “Consistent Comparisons of Experiments” subsection, PDF pp.4–5; inspected 2026-10-01. registry ↩a ↩b ↩c ↩d ↩e