Invariant decision problem

An invariant decision problem is a statistical decision problem whose probabilistic model, action space, and loss structure are preserved by a specified group of transformations. Invariance formalizes the requirement that transformations representing equivalent descriptions of an experiment must induce corresponding transformations of admissible decisions. The resulting theory connects group actions, equivariant estimation, hypothesis testing, and minimax decision rules.

The central objects are invariant experiments and equivariant decision rules. An invariant experiment remains statistically unchanged after the simultaneous transformation of observations and parameters, whereas an equivariant rule transforms its action in the manner prescribed by the group. Under appropriate regularity conditions, the risk of an equivariant rule is constant on each orbit of the parameter space. This orbit structure permits a decision problem to be reduced from individual parameter values to equivalence classes generated by the group.

Mathematical formulation

Let (\mathcal X) be a sample space, let (\Theta) be a parameter space, and let (\mathcal A) be an action space. A statistical experiment consists of a family of probability measures

[ \mathcal P={P_\theta:\theta\in\Theta} ]

on (\mathcal X), together with a loss function

[ L:\Theta\times\mathcal A\longrightarrow[0,\infty]. ]

A group (G) acts on the sample, parameter, and action spaces through transformations denoted by (g_{\mathcal X}), (g_{\Theta}), and (g_{\mathcal A}). The experiment is invariant under (G) when

[ P_{g_{\Theta}\theta}(g_{\mathcal X}B)=P_\theta(B) ]

for every measurable (B\subseteq\mathcal X), and when the loss satisfies

[ L(g_{\Theta}\theta,g_{\mathcal A}a)=L(\theta,a). ]

These identities state that transformation of the parameter produces the corresponding transformation of the observation distribution, while simultaneous transformation of the parameter and action leaves the loss unchanged.

A randomized decision rule is a Markov kernel (\delta(\mathrm da\mid x)) from (\mathcal X) to (\mathcal A). Its risk is

[ R(\theta,\delta) =\int_{\mathcal X}\int_{\mathcal A} L(\theta,a),\delta(\mathrm da\mid x),P_\theta(\mathrm dx). ]

The group acts on decision rules by transforming both their observations and their actions. A rule is equivariant when it is fixed by this induced action, equivalently when

[ \delta(g_{\mathcal A}C\mid g_{\mathcal X}x) =\delta(C\mid x) ]

for every measurable (C\subseteq\mathcal A). Invariance of the decision problem then gives

[ R(g_{\Theta}\theta,\delta)=R(\theta,\delta) ]

for every equivariant rule. Consequently, risk is constant along each orbit (G\theta). If the action of (G) on (\Theta) is transitive, every parameter belongs to the same orbit, so an equivariant rule has constant risk throughout the parameter space.

Symmetrization

A general decision rule need not respect the transformation structure of the experiment. For a finite group, its transformed versions can be averaged to form a randomized equivariant rule. If (\delta_g) denotes the transformation of (\delta) by (g), the symmetrized rule has the form

[ \bar\delta=\frac{1}{|G|}\sum_{g\in G}\delta_g. ]

The invariance identities imply that the maximal risk of (\bar\delta) does not exceed the maximal risk of the original rule. For a compact group, normalized Haar measure replaces the finite average. The resulting integral,

[ \bar\delta=\int_G\delta_g,\nu(\mathrm dg), ]

is invariant under further transformations because the Haar probability measure (\nu) is translation invariant.

You Watanabe’s 1954 analysis of finite transformation groups expressed this averaging argument directly in terms of randomized tests. The analysis identified the invariant rule as the group average of a test and showed that its power function was constant on parameter orbits related by the group. This formulation became part of the finite-group presentation of invariant testing, in which symmetrization requires no limiting construction or improper measure.

For noncompact groups, a normalized invariant probability measure on the entire group does not exist. Direct averaging is therefore replaced by limiting averages, invariant means, or generalized Bayes constructions. The validity of these replacements depends on the structure of the group and on analytical properties of the decision space.

Best equivariant rules

A best equivariant rule minimizes risk within the class of equivariant rules. In a transitive problem, every member of this class has constant risk, so comparison reduces to comparison of those constants. The best equivariant rule is not automatically minimax among all rules, because restricting attention to equivariant procedures can exclude rules with lower maximal risk.

The connection with Bayes estimators arises from Haar measure on the group. When the parameter space can be represented as a homogeneous space of (G), a right or left Haar measure induces an invariant or relatively invariant measure on (\Theta). A generalized Bayes rule under this measure frequently coincides with the best equivariant rule. The measure may be improper, so the resulting rule is generalized Bayes rather than Bayes with respect to a probability distribution.

The distinction between left and right Haar measures matters for non-unimodular groups. Their relationship is governed by the modular function, which changes the form of the induced prior and can alter the associated equivariant rule. Compact groups and abelian groups are unimodular, eliminating this distinction in many standard applications.

Location estimation

A basic example is the location model

[ X_i=\theta+\varepsilon_i,\qquad i=1,\ldots,n, ]

where the distribution of the error vector does not depend on (\theta). The additive group of real numbers acts by

[ x\mapsto x+c,\qquad \theta\mapsto\theta+c,\qquad a\mapsto a+c. ]

Under a loss depending only on estimation error, such as

[ L(\theta,a)=(a-\theta)^2, ]

the decision problem is invariant. An estimator (T) is equivariant precisely when

[ T(x_1+c,\ldots,x_n+c)=T(x_1,\ldots,x_n)+c. ]

Its risk is independent of (\theta), because translating the model converts the risk at one location into the risk at every other location.

E. J. G. Pitman developed the corresponding invariant estimation theory for location families. Under squared-error loss, the resulting Pitman estimator agrees with a generalized Bayes estimator based on Lebesgue measure, which is the Haar measure of the additive group. The sample mean is obtained when the errors are normally distributed, while other error distributions produce different equivariant estimators.

This example also illustrates the difference between invariance and ancillary reduction. Translation-invariant functions of the sample, such as centered residual configurations, contain information about the shape of the observed sample but not about its absolute location. An equivariant estimator combines such invariant information with one location-carrying component of the data.

Invariant tests

In a testing problem, the action space records rejection or non-rejection, possibly with randomization. A group preserving both the null and alternative parameter sets induces a transformed test with the same size and correspondingly transformed power. An invariant test is constant on sample-space orbits, so it depends on the data only through a maximal invariant.

A maximal invariant takes the same value for two observations exactly when they lie in the same orbit. Replacing the original sample by this statistic removes variation generated solely by the group while retaining distinctions between different orbits. The induced testing problem is then formulated on the quotient space (\mathcal X/G).

The relationship between invariant tests and minimaxity was developed through the Hunt–Stein theorem, associated with George Hunt and Charles Stein. In decision-theoretic formulations satisfying the relevant compactness and group conditions, the theorem equates an unrestricted minimax value with the value obtained from invariant procedures. Later formulations connect this result to amenable groups, whose invariant means provide an analogue of averaging for many noncompact transformation groups.

Decision-theoretic significance

Abraham Wald placed invariant procedures within the general framework of decision functions, risk, and minimaxity. In that framework, invariance is a structural property of the entire decision problem rather than a property attached only to an estimator or test. This distinction prevents an arbitrary transformation of the observations from being treated as a symmetry when it fails to preserve either the statistical model or the loss.

Invariant analysis separates three logically distinct reductions. The model reduction identifies parameter values connected by the group. The sample reduction replaces observations by their maximal invariant when decisions themselves are invariant. The decision-class reduction restricts attention to equivariant rules, whose risks are constant on parameter orbits. These reductions coincide only under additional assumptions about the group action and the form of the loss.

The theory does not imply that every scientifically meaningful procedure must be invariant. A proposed group can suppress distinctions represented in the loss or in the parameter space, in which case the underlying decision problem is not invariant under that group. Within a genuinely invariant formulation, however, orbit structure determines the precise relation between transformed experiments, transformed decisions, and their risks.