The statistic is not yet a decision

Classical statistical problems were often presented under separate headings: test a hypothesis, estimate a parameter, construct an interval. Wald’s decisive reorganization was to place these procedures inside one architecture of action. There is an unknown distribution or parameter, a possible experiment, an observation, a set of terminal actions and a numerical account of what each action costs when each state is true. A decision function connects observations to actions. Once that connection is explicit, tests and estimators become special cases rather than separate species.

This shift changes the professional question. The analyst no longer asks only whether an estimator is unbiased or a test has a certain significance level. The analyst asks what whole rule will be followed for every observation that could arrive, what losses that rule creates under every state that might govern the data, and whether another rule performs no worse everywhere. The realized action is one leaf of a policy. Evaluating only that leaf after the fact confuses an outcome with the rule that generated it.

Wald’s 1950 formulation, which expands the 1945 one-stage paper, makes the ingredients unusually explicit. The class of possible distributions describes what is unknown. The terminal-decision space describes what the experimenter may do. A weight or loss function records the damage of pairing a state with an action. A cost function records the burden of experimentation. The decision function specifies both how information is gathered and how a terminal decision follows. The 1945 theory kept experimentation to one stage; the later book generalized the architecture to sequential designs.

Loss is not an ornamental layer pasted onto probability. It determines what counts as a good inference procedure. If false acceptance is catastrophic and false rejection is mild, the same evidence should produce a different action than it would under symmetric loss. If another observation is costly, delay and sampling expense belong in the evaluation. A method cannot be optimal in the abstract; it is optimal relative to a state space, action space, experiment and loss specification.

That statement is both powerful and dangerous. It frees statistics from pretending that scientific action is value-free, but it also exposes the dependence of recommendations on modelling choices outside the likelihood. A loss table may encode money, safety, delay, missed discoveries or political priorities. Quantifying these dimensions does not make them objective. It makes them inspectable. The honest decision analysis reports which quantities came from data, which came from value judgements, and how the result moves when either changes.

A rule has a risk profile, not a single grade

For a fixed state, risk is the expected loss produced by a decision function. Change the state and the data distribution changes; the same rule may then make different errors at different rates. The full object is therefore a function R(θ,δ). It can be plotted as a curve over a continuous parameter or displayed as a vector over a finite set of states. The question “which rule has lower risk?” is incomplete until one asks where.

This statewise view makes dominance precise. Rule δ₁ is uniformly better than rule δ₂ if its risk is no larger for every possible state and smaller for at least one. A dominated rule has no compensating region: another available rule improves it somewhere without making it worse anywhere. Excluding such rules is a comparatively weak rationality demand. It does not select one winner when sensible risk curves cross.

Admissibility is the corresponding negative criterion: a rule is admissible when no other rule uniformly improves it. A class is complete when every rule outside it is dominated by some rule inside. Wald’s programme used Bayes rules and complete-class results to organize the otherwise immense space of possible procedures. But admissible does not mean best for a particular institution. Several admissible rules can distribute risk very differently, and loss misspecification can make an apparently admissible comparison irrelevant to the real decision.

Randomization can look like indecision if it is interpreted psychologically: why should a serious analyst flip a coin? In decision theory it has a geometric function. If two deterministic rules have risk vectors r₁ and r₂, choosing the first with probability p and the second with probability 1−p produces the componentwise average pr₁+(1−p)r₂. All such mixtures fill the line segment between the two risk points. The enlarged set may contain a profile whose highest component is below the highest component of either endpoint.

Operationally, randomization may also arise without an analyst literally tossing a coin. A population can be assigned policies probabilistically; an inspection system can sample cases; an experiment can allocate treatments; or a policy can randomize at a boundary to hit an exact error constraint. What matters is that the chance mechanism is specified in advance and its consequences are included in risk. Post-hoc arbitrariness is not the same thing.

The smallest possible maximum

When a credible prior distribution over states exists, average risk provides a direct criterion. But Wald focuses attention on the harder case in which a prior cannot be assumed or is unknown. The minimax rule minimizes the supremum of statewise risk. First, Nature is allowed to expose each rule’s worst state. Then the statistician selects the rule whose exposed worst value is lowest. The order of operations matters: minimize after taking the maximum.

The criterion does not claim that the worst state is likely. It protects a guarantee. If the state space and loss function are accepted, the minimax value is a ceiling on expected loss: no included state produces more. This is attractive when prior probabilities would be invented, when a decision must withstand adversarial review, or when a catastrophic region deserves explicit protection. It is also conservative in a literal mathematical sense: ordinary performance can be sacrificed to reduce an extreme component.

The finite geometry is simple. Plot each decision rule by its risk vector. For two states, a contour of maximum risk is a square anchored at the origin: every point inside has both coordinates below the same ceiling. Expand the square until it first touches the attainable risk set. That contact point is minimax. If the attainable frontier crosses the equal-risk diagonal, the contact often equalizes the active worst states. If one state remains strictly below the ceiling, it is not binding.

Equalization is a diagnostic, not a universal law. A dominated endpoint may be minimax only at a boundary; three or more states can leave some non-binding; constraints can prevent exact equalization; and infinite spaces require existence and compactness conditions. Wald’s later book devotes substantial work to these technical questions. A friendly diagram must not be mistaken for a general proof.

Minimax also depends entirely on which states are allowed. Add an implausible state with enormous loss and the rule can pivot toward it. Remove a difficult state and the guarantee improves. A robust-sounding answer can therefore conceal a fragile modelling boundary. The state space is part of the decision, not a neutral container. Good practice publishes the included states, the excluded states and the reason the supremum is being taken over exactly that set.

A prior can play Nature’s mixed strategy

Wald’s connection to game theory is not a loose metaphor. Put possible distributions on one side and decision functions on the other. Nature chooses a state; the experimenter chooses a rule; the numerical outcome is risk, which Nature is modelled as maximizing and the experimenter as minimizing. A prior distribution over states becomes a mixed strategy for Nature. A randomized decision function supplies the experimenter’s corresponding convexity.

Wald adds an essential caution. Nature need not have a mind or a desire to injure the statistician. The experimenter’s ignorance makes it reasonable, in some settings, to behave as if Nature maximized risk. The game formalism is useful even without literal hostility because it exposes the max–min structure and supports results about admissible rules and complete classes. Anthropomorphic language should never substitute for the mathematical correspondence.

For any prior π, the Bayes rule minimizes average risk Σπ(θ)R(θ,δ). Now let Nature choose the prior that makes this minimized Bayes risk as large as possible. Such a distribution is least favourable. It concentrates weight where the experimenter is hardest to protect. Under appropriate conditions, a Bayes rule for a least-favourable prior is minimax. The two optimization directions meet at a saddle point.

In the two-rule example, let q be the prior probability of state θ₁. Rule A’s Bayes risk is q×1+(1−q)×7 = 7−6q. Rule B’s is q×6+(1−q)×2 = 2+4q. Nature wants the lower of these two lines: the statistician’s best response: to be as high as possible. The lines meet at q=0.5 and value 4. At that prior the statistician is indifferent between A and B, so any mixture is Bayes; the particular mixture p=0.4 also equalizes the state risks and is minimax.

The equalities reveal complementary indifference. At the minimax mixture, Nature is indifferent between the active states because each gives risk 4. At the least-favourable prior, the statistician is indifferent between the active pure rules because each has average risk 4. Move away from either mixture and the opposing side has a profitable response. That mutual stability is the finite saddle-point picture.

What minimax protects: and what it cannot decide

Minimax offers a clean guarantee only after the problem has been built. It cannot choose the loss function, decide which parameters are possible, determine whether data collection is ethical, or tell an organization how much ordinary performance to trade for extreme protection. These are not minor implementation details. They shape the risk surface whose maximum is being minimized.

The criterion can also be insensitive to improvements away from the worst state. If two rules have the same maximum risk but one is much better everywhere else, a bare minimax comparison may tie them. Lexicographic refinements, Bayes criteria, minimax regret, constraints or domain-specific preferences may break the tie. Each refinement changes the decision principle and should be named.

Minimax risk is especially easy to confuse with minimax regret. Risk compares absolute expected loss. Regret compares a rule’s loss with the best loss that could have been achieved if the state were known. A rule can have high absolute loss but low regret in an intrinsically difficult state, or low absolute loss but high regret where an alternative would have been excellent. Leonard Savage later developed minimax regret as a distinct response to uncertainty. Importing it into Wald’s 1945 title erases a real conceptual difference.

Robust deployment therefore needs two audits. The internal audit verifies the optimization: risk arithmetic, dominance, envelope minimization, randomization and any saddle-point certificate. The external audit examines whether states, losses and distributions represent the real environment. A laboratory can prove that p=0.4 is the exact minimax mixture for a two-by-two table; it cannot prove that the table describes a clinic, market or intelligence system.

These audits fail in different ways. An internal failure is mathematical or computational: the loss was averaged under the wrong sampling distribution, a maximum was taken before rather than after mixing, a local search missed a better rule, or a dual value does not match the proposed solution. An external failure can survive every unit test: a security model omits an attacker’s adaptation, a medical model compresses quality of life into mortality, or an intelligence estimate treats a politically convenient scenario as impossible. Passing the first audit is necessary because arithmetic errors are avoidable. Passing it never licenses silence about the second.

The framework is also valuable when exact minimax optimization is not the final policy. Risk profiles expose the trade before leaders choose how to resolve it. One team may accept a higher ceiling for much lower risk in common states; another may impose a safety constraint and optimize average performance inside it; a third may buy information that shrinks the state set. Those departures are intelligible because Wald’s architecture has separated observation, action, consequence and criterion. Without that separation, “better model” can hide several incompatible claims.

Wald’s framework also clarifies why a model score alone does not own the decision. A predictive model proposes a distribution for observations or outcomes. The decision function includes thresholds, abstention, additional tests and action randomization. The loss function expresses consequences. Governance belongs around the assembled policy. Teams that monitor only model accuracy can miss a threshold change, a cost shift or a newly relevant state that invalidates the original guarantee.

A practical protocol for worst-case decision design

First, enumerate states at the level needed to change action. Avoid a catalogue so broad that every imaginable disaster dominates, but do not remove inconvenient states merely to improve the guarantee. Record the evidence for inclusion. Second, enumerate terminal actions, including abstention, escalation and information gathering. A narrow action set can make every statistical method look poor.

Third, specify loss in units that stakeholders can interrogate. Separate false-positive harm, false-negative harm, delay, experimentation, resource use and irreversible consequences before combining them. Publish sensitivity ranges when conversion to one scale is controversial. Fourth, define the decision function as a full mapping, including thresholds, ties, missing data and any randomization. “Use model M” is not a complete policy.

Fifth, calculate statewise risk and remove dominated rules. Sixth, identify whether the decision principle is empirical Bayes, subjective Bayes, minimax risk, minimax regret or a constraint formulation. If minimax is chosen, show the worst-state envelope and the states that bind. Seventh, search for a least-favourable prior or another dual certificate when the conditions and scale permit it; a matching lower bound is stronger evidence than a numerical optimizer alone.

Eighth, stress the boundary. Perturb losses, state definitions and sampling distributions. Test whether a modest change swaps the selected rule. Ninth, deploy with monitoring of both predictive behaviour and decision consequences. Tenth, reopen the policy when action costs, prevalence, institutional capacity or strategic behaviour changes. A minimax guarantee is durable only relative to the problem it actually solved.

Keep the final record reproducible. Store the risk table or model, the code that computes it, the chosen criterion, the optimizer’s tolerance, the dual or enumeration check, and the approved interpretation. Record who owns the loss judgements and when they expire. This turns a worst-case number from a rhetorical shield into an inspectable decision artefact. It also lets a future reviewer distinguish a changed world from a broken calculation: the former calls for a new problem specification, while the latter calls for correction of the existing one.

Executable minimax laboratory

The laboratory below solves a finite two-state decision problem exactly. Select a risk profile and, for the crossing cases, choose any proposed probability of using Rule A. The engine computes the statewise mixture risks, the current worst case, the analytic minimax mixture, the least-favourable prior and three self-tests. The chart shows the two state risks and their upper envelope across all mixtures.

Memory studio