Blackwell Informativeness Order¶
Compare experiments by whether one can be garbled into another, equivalently by universal Bayes decision value under standard finite assumptions.
Core Idea¶
The Blackwell informativeness order compares two experiments over the same hidden states. In a finite standard setting, experiment \(P\) is at least as informative as \(Q\) if a stochastic map \(M\) applied only to \(P\)'s observed signal yields \(Q\)'s state-conditional signal law: \(Q=PM\). Blackwell's theorem equates this signal-garbling test with \(P\) weakly improving optimized Bayes value over \(Q\) for every matched expected-utility decision problem in its stated class. The frozen requested title names the theorem; this entry reframes it as the reusable comparison relation that theorem characterizes.[ref-3eaf505d1677][ref-95e5517c8e74]
The raw relation is reflexive and transitive, but different signal representations can mutually garble one another. It is therefore a preorder on raw experiments and a partial order only after quotienting by mutual Blackwell equivalence. The proposed strict live parent is prime Relation, not the live antisymmetric Order.[ref-3eaf505d1677][ref-23865c39f73c][^ref-23865c39f73c-2]
Scope of Application¶
In a constructed binary-state prediction example, \(P\) reports the hidden state perfectly while \(Q\) flips its report with probability $0.25$. A state-independent noise channel turns \(P\) into \(Q\); an observer of \(P\) can simulate any decision rule available after \(Q\). With a fixed prior, \(Q\) posteriors are conditional averages of \(P\) posteriors. The numerical example illustrates the theorem, not a reported empirical result.[^ref-3eaf505d1677]
In investment information economics, Cabrales, Gossner and Serrano specify finite payoff-relevant states, state-to-signal structures and feasible asset choices after a signal. Blackwell garbling is their universal benchmark. Their own complete entropy-based order is for restricted ruin-averse/no-arbitrage investor choices and must not be mistaken for the same partial Blackwell order; advantage in that restricted setting does not by itself prove universal dominance.[^ref-95e5517c8e74]
Clarity¶
Write \(P_{\theta s}=\Pr(s\mid\theta)\) and \(M_{st}=\Pr(t\mid s)\), so \(Q_{\theta t}=\sum_sP_{\theta s}M_{st}\). The post-processing \(M\) cannot read \(\theta\) after the original signal. A claim that one experiment was more useful for one payoff function is weaker than Blackwell dominance, which ranges over all decision problems of the theorem's class.[^ref-3eaf505d1677]
Under a common fixed prior, a posterior after a garbled signal is an average of better-signal posteriors, and the worse distribution of posteriors is a mean-preserving contraction in the appropriate convex-order description. It is not literally the same law of posteriors and should not be compared across changed priors without redoing the setup.[ref-3eaf505d1677][ref-9d3bb6815380]
Manages Complexity¶
The theorem replaces many Bayes payoff comparisons with a structural witness: one stochastic signal-only garbling matrix. If such a matrix exists, any policy available with the worse experiment can be simulated with the better one. The reverse direction—from universal payoff comparison to a garbling witness—is the nontrivial theorem and depends on its stated conditions.[^ref-3eaf505d1677]
The price of this robustness is incomparability. Many experiment pairs admit no garbling in either direction. A restricted investor order can rank more pairs, but then makes a weaker, task-class-specific assertion. Costs of obtaining or processing information and strategic responses need separate modeling.[^ref-95e5517c8e74]
Abstract Reasoning¶
To test an instance, identify the common states, state-to-signal kernels and whether a state-independent \(M\) satisfies \(Q=PM\). If yes, the universal Bayes-value direction follows by simulating \(Q\) after seeing \(P\). For posteriors, first fix a common prior; then each garbled posterior is a weighted average of the contributing original posteriors. This is a contraction statement, not arbitrary posterior-law equality.[ref-3eaf505d1677][ref-9d3bb6815380]
Raw kernels with different alphabets can be equivalent: split an original signal into independent duplicate labels and merge them back. They mutually garble, so antisymmetry appears only on their equivalence classes. Live prime Order's current definition cannot be the raw relation's strict parent.[ref-3eaf505d1677][ref-23865c39f73c-2]
Knowledge Transfer¶
The hidden-state, experiment-pair, signal-only garbling and universal-decision roles transfer from prediction to investment signals. The conclusion travels only while the decision maker can use, ignore or simulate the free signal with an unchanged action opportunity. A one-task success, restricted entropy comparison or strategic contract effect is not automatically a Blackwell claim.[ref-3eaf505d1677][ref-95e5517c8e74]
The source hierarchy remains explicit: the full original Blackwell 1953 article was not accessible here, while later original research directly states and uses the finite theorem. The staged identity is a domain-specific formal comparison, not yet independently reviewed or canonically promoted.[^ref-3eaf505d1677]
[^ref-3eaf505d1677]: E. Karni and Z. Safra, “Hybrid Decision Model and the Ranking of Experiments”, author-hosted original research (2021), §§2.1–2.2 and §3.1, inspected 2026-10-01. [^ref-95e5517c8e74]: 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. [^ref-9d3bb6815380]: F. Yang and K. H. Yang, “Stochastic Optimization and Coupling”, author-hosted original research (2026), posterior-comparison discussion, PDF pp.4–5, inspected 2026-10-01. [^ref-23865c39f73c]: Encyclopedia of Abstractions, live prime Relation, inspected 2026-10-01. [^ref-23865c39f73c-2]: Encyclopedia of Abstractions, live prime Order, inspected 2026-10-01.
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 (80th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.
Family — Unclustered & Miscellaneous (2551 abstractions)
Nearest neighbors
- Particle Filter — 0.84
- Kushner–Stratonovich Equation — 0.83
- Matrix Analytic Method — 0.82
- Quantum State — 0.82
- Noisy Channel Model — 0.82
Computed from structural-signature embeddings · 2026-10-08