COLLEGIUMCOLLEGIUM
Hall of Shoulders

Decision Science & OR

Leonard Savage

Leonard Savage is known for subjective expected utility (SEU), the foundations of statistics, the sure-thing principle, personal probability, the minimax-regret criterion.. **Hall of Shoulders | COLLEGIUM** Thinker ID: `savage` | Domain: decision theory / operations research / foundations of statistics

Built

Sources

49

Primary + secondary

Citations

0

ARGOS-tracked

FTS5 Chunks

49

Retrieval index

Councils

0

Memberships

Review Lens

Adversarial questions for candidates

The falsifiable questions this brain puts to a dissertation candidate. They seed the pre-Conclave initial review whenever a candidate's topic matches the Decision Science & OR lens.

  1. 1

    State the loss function. "You propose a collision-avoidance / debris-mitigation / licensing decision rule. Write down explicitly the loss (cost) of a false alarm and the loss of a missed detection, and show that your chosen threshold or rule is the one that minimizes expected loss against your stated prior. If you cannot produce the loss function, your threshold is arbitrary." (Falsifiable: either the candidate exhibits a coherent loss-minimizing derivation or the rule is shown to be unjustified.)

  2. 2

    Own the probability. "Whose degree of belief is your 'probability of collision' (or reentry-casualty probability, or hostile-intent probability)? Specify the information set it is conditioned on and demonstrate it is coherent (no Dutch book). Show how it is updated by Bayes' rule when the next observation arrives." (Falsifiable against incoherence or a frequency/logical probability masquerading as a personal one.)

  3. 3

    Defend or disavow additivity. "If your model uses Dempster-Shafer belief functions, interval probabilities, or any ambiguity-sensitive (maxmin/Choquet) criterion, you have rejected my sure-thing principle. Justify that rejection on the specific space problem, or revert to additive subjective probability." (Falsifiable: the candidate must either produce a substantive ambiguity argument or concede additivity applies.)

  4. 4

    Prove it is a small world. "Have you enumerated the state space and consequences completely enough that subjective expected utility is legitimate, or is this a grand-world problem (unforeseen constellations, non-stationary debris growth, model misspecification) where SEU is, in my own words, ridiculous? If grand-world, show why you are using a fixed-prior optimization instead of minimax-regret or adaptive robust planning." (Falsifiable: the candidate must classify the problem and match the method to that classification.)

  5. 5

    Expose the regret. "Compute the maximum regret of your recommended action across the plausible states you cannot assign a defensible prior to. If that maximum regret is catastrophic (an avoidable collision, an irreversible debris cascade), explain why an expected-value rule rather than a minimax-regret rule is the responsible choice." (Falsifiable: a candidate whose expected-value optimum carries unbounded worst-case regret must justify it or change rules.)

Core Concepts & Space Translation

Subjective Expected Utility (SEU) and the Savage axioms

*The Foundations of Statistics* (Savage 1954) shows that an agent whose preferences over "acts" (functions from states of the world to consequences) satisfy a small set of axioms (P1-P7, including completeness, transitivity, and the sure-thing principle) behaves *as if* it maximizes the expectation of a utility function with respect to a unique personal (subjective) probability. This is the foundational result that probability and utility can be *derived jointly* from coherent choice behavior, not assumed separately. It is the benchmark every "risk threshold" in space operations is implicitly measured against.

Space translation

See Space Applications below for how this framework translates to contemporary space governance, drawn directly from the dossier's applied-literature review.

Personal / subjective probability

Savage rejected the view that probability must be a frequency or a logical relation. Probability is a *degree of belief* of a specific decision-maker, measurable through their willingness to bet, and disciplined only by coherence (the Dutch-book/no-sure-loss requirement). This reframes every "probability of collision," "probability of reentry casualty," or "probability of a hostile maneuver" as a belief that must be owned by an identifiable agent, conditioned on an explicit information set, and updated by Bayes' rule.

Space translation

See Space Applications below for how this framework translates to contemporary space governance, drawn directly from the dossier's applied-literature review.

The sure-thing principle (P2)

If an act is preferred to another whenever event E obtains, and also preferred whenever E does not obtain, it must be preferred unconditionally. This is the separability axiom that makes expected-utility additivity possible. It is also the axiom most often violated in practice (Allais paradox, ambiguity aversion), so it is the precise place to interrogate whether a space decision model is coherent or is quietly using a non-additive / ambiguity-sensitive rule.

Space translation

See Space Applications below for how this framework translates to contemporary space governance, drawn directly from the dossier's applied-literature review.

The minimax-regret criterion

In *The Theory of Statistical Decision* (Savage 1951) and Chapter 9 of the *Foundations*, Savage advanced regret (the difference between the payoff of the act chosen and the best act available in hindsight for the realized state) and the minimax-regret rule as a defensible criterion when a prior cannot honestly be formed. This is his bridge between strict Bayesianism and the Wald minimax program: when you genuinely cannot quantify a prior, minimize the maximum regret rather than pretend to a probability.

Space translation

See Space Applications below for how this framework translates to contemporary space governance, drawn directly from the dossier's applied-literature review.

Statistical decision theory and the loss function (the Wald-Savage program)

Savage placed Wald's statistical decision theory on Bayesian foundations: a statistical procedure is judged by its *expected loss* (risk) under an explicit loss function, and the Bayes procedure minimizes expected loss against the prior. Sequential procedures (Wald's sequential probability ratio test) fall out as the cost-optimal way to trade observation cost against decision error. This is the most operational of Savage's ideas for space: it converts "what threshold do we maneuver at?" into "what is the loss function, and what decision rule minimizes expected loss?"

Space translation

See Space Applications below for how this framework translates to contemporary space governance, drawn directly from the dossier's applied-literature review.

Small worlds vs. grand worlds (the limits of SEU)

Savage himself flagged that SEU is compelling only in a "small world" where the state space and consequences can be fully enumerated; extending it to a "grand world" (unforeseen contingencies, model misspecification) he called "utterly ridiculous." This self-imposed boundary is itself a framework: it licenses the reviewer to ask whether a candidate has confused a tractable small-world model with the genuinely open, deeply uncertain grand world of long-run orbital sustainability.

Space translation

See Space Applications below for how this framework translates to contemporary space governance, drawn directly from the dossier's applied-literature review.

Key works:

Savage, *The Foundations of Statistics* (Wiley, 1954; 2nd ed. Dover, 1972); Savage, "The Theory of Statistical Decision," *J. Amer. Statist. Assoc.* 46 (1951): 55-67; Savage, "Elicitation of Personal Probabilities and Expectations," *JASA* 66 (1971): 783-801.

Space translation

See Space Applications below for how this framework translates to contemporary space governance, drawn directly from the dossier's applied-literature review.