Statistical Decision Problems as Zero-Sum Games

From Statistical Risk to a Game

A statistical rule is chosen before the data are observed, but its performance depends on an unknown parameter. The same rule can have small risk at one parameter value and large risk at another. A game formulation makes this tension explicit: the decision maker chooses a rule, an adversary chooses a parameter value, and the payoff to the adversary is the rule’s risk at that parameter.

This is a zero-sum representation because the two players have opposite objectives. The decision maker seeks a rule with low payoff; the adversary seeks a parameter that makes that rule perform poorly. The game is a way to organize the statistical problem, not a claim that a real opponent selects the data after seeing them. The parameter determines the sampling distribution, and the rule is applied to data generated from that distribution.

Definition: Let $X$ have distribution $P_\theta$ on an observation space $\mathcal X$, where $\theta$ belongs to a parameter space $\Theta$. Let $\mathcal A$ be an action space and let $\lambda(\theta,a)$ be a nonnegative measurable loss. A statistical decision rule $\delta$ maps observations to actions. Its risk at $\theta$ is $$ R(\theta,\delta)=E_\theta[\lambda(\theta,\delta(X))]. $$ The associated zero-sum game has the decision maker’s rules as one player’s choices, the parameter values $\theta\in\Theta$ as the adversary’s choices, and payoff $R(\theta,\delta)$ to the adversary.

For a deterministic rule, the action is $\delta(X)$. A randomized rule can instead be described by a probability kernel $Q(da\mid x)$, which specifies the distribution of the action conditional on the observation. In that case,

R(\theta,Q) = \int_{\mathcal X}\int_{\mathcal A} \lambda(\theta,a)\,Q(da\mid x)\,P_\theta(dx).

This formula includes both sources of randomness: the observation generated under $P_\theta$ and any randomization used by the rule. Assume the spaces and kernel are measurable so the integrals are defined. Nonnegative loss ensures the iterated integral is valid, including when the risk is infinite. If the loss is not nonnegative, integrability is needed instead.

The adversary’s choice is a parameter, not an action selected after observing $X$. Once $\theta$ is selected, it determines the distribution of $X$. The decision maker’s choice is a whole rule, not a single action: the rule specifies how to respond to every possible observation. This difference is central to interpreting the game correctly.

Worked Examples: Reading the Payoff

Worked Example: A Bernoulli Observation and Two Hypotheses

Suppose $X\sim\operatorname{Bernoulli}(\theta)$, where $\Theta=\{0.2,0.8\}$. The action is either 0 or 1, and the loss is zero for a correct choice and one for an incorrect choice. Consider the rule $\delta(X)=X$.

At $\theta=0.2$, the correct action is 0. The rule makes an error exactly when $X=1$, which has probability $0.2$. Thus $R(0.2,\delta)=0.2$. At $\theta=0.8$, the correct action is 1. The error occurs when $X=0$, with probability $1-0.8=0.2$. Thus $R(0.8,\delta)=0.2$.

The game payoff is therefore $0.2$ whichever of these two parameter values the adversary chooses. For comparison, the constant rule $\delta(X)=0$ has risk 0 at $\theta=0.2$ and risk 1 at $\theta=0.8$. The game representation records both risks: the rule is the decision maker’s choice, while each parameter value gives a possible payoff.

Worked Example: Shrinkage Rules in a Normal Model

Let $X\sim N(\theta,1)$, with $\theta\in[-1,1]$, action space $\mathbb R$, and squared-error loss $\lambda(\theta,a)=(a-\theta)^2$. Consider the family of rules $\delta_c(X)=cX$, where $c$ is fixed.

Since $E_\theta[X]=\theta$ and $\operatorname{Var}_\theta(X)=1$, the rule has mean $c\theta$ and variance $c^2$. Decomposing mean squared error into variance plus squared bias gives $$ R(\theta,\delta_c) =c^2+(c\theta-\theta)^2 =c^2+(1-c)^2\theta^2. $$

For $c=1/2$, the risk is $1/4+\theta^2/4$. At $\theta=0$, it is $1/4$; at either endpoint, $\theta=-1$ or $\theta=1$, it is $1/2$. The adversary’s choice changes the payoff even though the rule stays fixed. If the decision maker considers only this family, the largest risk over $[-1,1]$ is $c^2+(1-c)^2$. Completing the square, $$ c^2+(1-c)^2=2(c-1/2)^2+1/2, $$ so within this family the largest risk is minimized at $c=1/2$. This calculation illustrates the game’s choices and payoffs; it does not assert that this family contains every possible rule.

Worked Example: A Threshold Rule for a Poisson Observation

Suppose $X\sim\operatorname{Poisson}(\theta)$, where $\theta\in\{1,4\}$. The action is to treat (action 1) or not treat (action 0). Treating when treatment is unnecessary costs 2, while not treating when treatment is needed costs 5. Treat when $\theta=4$ is the correct action; treat when $\theta=1$ is incorrect. Consider the rule that treats exactly when $X\geq2$.

At $\theta=1$, treatment is an error, so the risk is twice the probability of $X\geq2$: $$ R(1,\delta)=2P_1(X\geq2) =2\{1-P_1(X=0)-P_1(X=1)\} =2(1-2e^{-1}). $$ Numerically this is approximately $0.5285$.

At $\theta=4$, failing to treat is the error. The rule fails to treat exactly when $X\leq1$, so $$ R(4,\delta)=5P_4(X\leq1) =5\{e^{-4}+4e^{-4}\} =25e^{-4}, $$ which is approximately $0.4579$. These two values are the adversary’s possible payoffs against this particular rule. The loss costs matter: the payoff is not merely an error probability.

What the Game Formulation Does—and Does Not—Say

The risk function is the payoff function. For each fixed rule, the adversary can compare $R(\theta,\delta)$ across parameter values; for each fixed parameter, the decision maker can compare rules. When the objective is protection against the least favorable parameter, the relevant criterion for a rule is its largest risk,

\sup_{\theta\in\Theta}R(\theta,\delta).

The corresponding minimax objective is to choose a rule that makes this largest risk as small as possible. This description identifies the optimization problem without implying that the order of optimization can be reversed or that an optimal rule or parameter must exist. Such claims require additional conditions.

The representation also keeps statistical randomization distinct from the adversary’s choice. A randomized rule uses randomness in selecting an action, conditional on the data, and its risk averages over that randomness. The adversary’s choice in this formulation remains a particular value $\theta$. No claim about an adversary randomizing over parameter values is needed to define the payoff.

Result: If the loss is nonnegative and measurable and $Q$ is a measurable randomized rule, its game payoff at $\theta$ is exactly its risk: $$ u(Q,\theta) = \int_{\mathcal X}\int_{\mathcal A} \lambda(\theta,a)\,Q(da\mid x)\,P_\theta(dx) =R(\theta,Q). $$

Proof. Under parameter $\theta$, the observation has distribution $P_\theta$. Conditional on observing $x$, the rule selects an action with distribution $Q(da\mid x)$. Therefore the joint distribution of observation and action is $P_\theta(dx)Q(da\mid x)$. The expected loss under this joint distribution is the displayed iterated integral. Because the loss is nonnegative and measurable, Tonelli’s theorem justifies evaluating that expectation as an iterated integral. By definition, the expected loss is the risk, and the game payoff was defined to be this risk. Hence $u(Q,\theta)=R(\theta,Q)$. $\square$

Comparing Rules Through Their Payoffs

A useful consequence of the game view is that a risk comparison is a payoff comparison, parameter by parameter. In particular, if one rule has no greater expected loss at any parameter value, then the adversary cannot obtain a larger payoff against that rule at any of its choices.

Result: Let $Q_1$ and $Q_2$ be two rules. Suppose, for every $\theta$ and $P_\theta$-almost every $x$, their conditional expected losses satisfy $$ \int_{\mathcal A}\lambda(\theta,a)\,Q_1(da\mid x) \leq \int_{\mathcal A}\lambda(\theta,a)\,Q_2(da\mid x). $$ Then $R(\theta,Q_1)\leq R(\theta,Q_2)$ for every $\theta$.

Proof. Fix any $\theta$. Integrate the assumed conditional inequality with respect to $P_\theta(dx)$. Integration preserves the inequality, giving $$ \int_{\mathcal X}\int_{\mathcal A}\lambda(\theta,a)\,Q_1(da\mid x)\,P_\theta(dx) \leq \int_{\mathcal X}\int_{\mathcal A}\lambda(\theta,a)\,Q_2(da\mid x)\,P_\theta(dx). $$ Both sides are the corresponding risks by the risk formula, so $R(\theta,Q_1)\leq R(\theta,Q_2)$. Since $\theta$ was arbitrary, the inequality holds at every parameter value. $\square$

If both rules have finite risks, this pointwise comparison also implies that the largest risk of $Q_1$ is no greater than the largest risk of $Q_2$. Indeed, every value $R(\theta,Q_1)$ is bounded above by $R(\theta,Q_2)$, and taking suprema preserves that order. This is why dominance and risk calculations can be read directly as statements about the adversary’s payoff.

Common Mistake: Letting the Adversary See the Data

A common misinterpretation is to imagine that the adversary observes the realized $X$ and then chooses a parameter value to maximize the loss for that particular observation. That would replace the fixed-parameter risk with a different, data-dependent construction. In the statistical game here, $\theta$ determines the distribution from which $X$ is generated; the risk averages loss over that distribution. The adversary chooses among parameter values, not among data-dependent responses to individual observations.

A second mistake is to treat the payoff as the loss at one observation. For example, a rule may make an error for some possible samples and succeed for others. Its payoff at $\theta$ is the expected loss across those samples, including any randomization by the rule—not the loss from one realized sample.

Key takeaway: In the statistical zero-sum game, the decision maker chooses a complete rule, the adversary chooses a parameter value, and the payoff is $R(\theta,\delta)$. The sampling model and loss determine that payoff; neither player’s role should be confused with choosing a realized observation.

Check Your Understanding

Use the definitions and examples above to answer these questions.

In the game representation, what does the decision maker choose, and what does the adversary choose?
Why is the payoff at a fixed parameter the expected loss rather than the loss at one observed data value?
For $X\sim N(\theta,1)$ and $\delta_c(X)=cX$, derive the risk $c^2+(1-c)^2\theta^2$ from the estimator’s mean and variance.
In the Poisson treatment example, why is the risk at $\theta=4$ five times $P_4(X\leq1)$?
What additional interpretation would be introduced by letting the adversary choose a parameter after seeing the data, and why is that not the game defined here?