Integrated Risk Differences as Admissibility Certificates

Why Integrate a Risk Difference?

Admissibility is a pointwise comparison across the parameter space. A decision rule is inadmissible if another rule has risk no greater at every parameter value and strictly smaller at least somewhere. In a complicated problem, however, it can be easier to compare the rules after averaging their risks against a probability measure on the parameter space. The key is to preserve the pointwise comparison: under domination, the risk difference has one sign, so its integral cannot conceal cancellation between positive and negative parts.

An integrated comparison is therefore a useful certificate only when its hypotheses and conclusion are kept precise. A zero integral of a nonnegative function implies that the function is zero almost everywhere with respect to the integrating measure. It does not, by itself, imply equality at every parameter value.

Risk Differences and Integrated Risk

Definition: Let $\delta$ and $\eta$ be decision rules in a statistical decision problem with parameter space $\Theta$ and loss $L(\theta,a)$. Their risks are $R(\theta,\delta)=E_\theta[L(\theta,\delta(X))]$ and $R(\theta,\eta)=E_\theta[L(\theta,\eta(X))]$, whenever these expectations are defined. Define the pointwise risk difference, with $\delta$ listed first, by $$ g_{\delta,\eta}(\theta)=R(\theta,\delta)-R(\theta,\eta). $$ Rule $\eta$ dominates $\delta$ if $g_{\delta,\eta}(\theta)\geq 0$ for every $\theta\in\Theta$, and $g_{\delta,\eta}(\theta)>0$ for at least one $\theta$. Rule $\delta$ is admissible if no rule dominates it.

The sign convention matters: a positive difference means that $\eta$ has smaller risk. If $\pi$ is a probability measure on $\Theta$ and the difference is measurable and integrable, define the integrated risk difference by

D_\pi(\delta,\eta) =\int_\Theta \bigl[R(\theta,\delta)-R(\theta,\eta)\bigr]\,\pi(d\theta) =\int_\Theta g_{\delta,\eta}(\theta)\,\pi(d\theta).

When the individual risks are nonnegative and integrable under $\pi$, this is also the difference between the integrated risks. More generally, it is enough that $g_{\delta,\eta}$ itself be integrable. If $\eta$ dominates $\delta$, then $g_{\delta,\eta}\geq 0$, and therefore $D_\pi(\delta,\eta)\geq 0$. The central question is what can be concluded if this integral is zero.

The Zero-Integral Argument

Theorem: Let $(\Theta,\mathcal{F},\pi)$ be a measure space, and let $g:\Theta\to[0,\infty]$ be measurable. If $\int_\Theta g\,d\pi=0$, then $g(\theta)=0$ for $\pi$-almost every $\theta$.

This fact is the core of the certificate. The nonnegativity assumption is essential: without it, positive and negative values can cancel in the integral.

Proof. For each positive integer $k$, let $A_k=\{\theta:g(\theta)\geq 1/k\}$. Measurability of $g$ makes $A_k$ measurable. Since $g\geq 1/k$ on $A_k$,

0=\int_\Theta g\,d\pi \geq \int_{A_k}g\,d\pi \geq \frac{1}{k}\pi(A_k).

Hence $\pi(A_k)=0$ for every $k$. If $g(\theta)>0$, then $g(\theta)\geq 1/k$ for some positive integer $k$, so $\{g>0\}=\bigcup_{k=1}^{\infty}A_k$. Countable subadditivity now gives $\pi(\{g>0\})=0$, which is precisely $g=0$ $\pi$-almost everywhere. $\square$

Applied to risks, this gives an immediate consequence. Suppose $\eta$ dominates $\delta$, the risk difference is measurable, and $D_\pi(\delta,\eta)=0$. The theorem says $R(\theta,\delta)=R(\theta,\eta)$ for $\pi$-almost every $\theta$. Thus, if every possible strict improvement would have to occur on a set of positive $\pi$-measure, a dominator cannot exist.

Certificate: If every rule $\eta$ that dominates $\delta$ would satisfy $\pi(\{\theta:R(\theta,\eta)<R(\theta,\delta)\})>0$, then $D_\pi(\delta,\eta)=0$ is impossible for a dominator. In particular, $\delta$ is admissible if $D_\pi(\delta,\eta)=0$ for every candidate $\eta$ and every dominator would have a strict-improvement set of positive $\pi$-measure.

A useful special case is a finite parameter space with positive probability assigned to every parameter value. There, a strict improvement at even one state contributes a positive amount to the integral. The analogous claim for a continuous parameter space needs care: a single point can have measure zero.

Worked Examples

Worked Example: Positive Weights Detect a Strict Improvement

Consider a decision problem with two parameter values, $\theta_1$ and $\theta_2$. The observation is uninformative, and the available deterministic rules have the following risks:

Rule	$R(\theta_1,\cdot)$	$R(\theta_2,\cdot)$
$\delta$	4	2
$\eta$	3	2

Take $\pi(\{\theta_1\})=1/3$ and $\pi(\{\theta_2\})=2/3$. The pointwise risk differences are

g_{\delta,\eta}(\theta_1)=4-3=1,\qquad g_{\delta,\eta}(\theta_2)=2-2=0.

The difference is nonnegative at both states, and strictly positive at $\theta_1$, so $\eta$ dominates $\delta$. Its integrated difference is

D_\pi(\delta,\eta) =\frac{1}{3}(1)+\frac{2}{3}(0) =\frac{1}{3}>0.

The positive weight on $\theta_1$ detects the strict improvement. In this finite setting, any prior giving positive mass to each state has the same property: a nonnegative difference with a positive value at one state must have a positive integral.

Worked Example: A Zero Integral Does Not Detect Every Strict Improvement

Let $\Theta=[0,1]$, let the observation be constant, and consider two actions $a_0,a_1$ with losses

L(\theta,a_0)=\mathbf{1}\{\theta=0\},\qquad L(\theta,a_1)=0.

Let $\delta$ always choose $a_0$, and let $\eta$ always choose $a_1$. Their risks are the displayed losses because there is no random observation. Thus

g_{\delta,\eta}(\theta) =R(\theta,\delta)-R(\theta,\eta) =\mathbf{1}\{\theta=0\}.

Rule $\eta$ dominates $\delta$: the difference is nonnegative everywhere and equals 1 at $\theta=0$. Now let $\pi$ be the uniform probability measure on $[0,1]$. Since a single point has uniform measure zero,

D_\pi(\delta,\eta) =\int_{[0,1]}\mathbf{1}\{\theta=0\}\,\pi(d\theta) =\pi(\{0\})=0.

There is strict improvement, despite the zero integrated difference. The theorem is not contradicted: the difference is zero $\pi$-almost everywhere, and the strict-improvement set $\{0\}$ is a $\pi$-null set. This example shows why equality almost everywhere cannot be silently upgraded to equality everywhere.

Worked Example: Improvement on a Positive-Measure Set

Again take $\Theta=[0,1]$ and a constant observation, but now use actions $a_0,a_1$ with losses $L(\theta,a_0)=\theta$ and $L(\theta,a_1)=0$. Let $\delta$ always choose $a_0$, let $\eta$ always choose $a_1$, and let $\pi$ be uniform on $[0,1]$. The risks are

R(\theta,\delta)=\theta,\qquad R(\theta,\eta)=0,\qquad g_{\delta,\eta}(\theta)=\theta.

For every $\theta\in[0,1]$, the difference is nonnegative; it is strictly positive for every $\theta\in(0,1]$. Its integrated value is

D_\pi(\delta,\eta) =\int_0^1\theta\,d\theta =\left[\frac{\theta^2}{2}\right]_0^1 =\frac{1}{2}>0.

Here the strict-improvement set has positive measure, so its improvement is detected by the integral. If a separate argument established that a proposed rule has the smallest possible integrated risk, this calculation would rule out any dominator whose strict-improvement set has positive $\pi$-measure.

Using the Certificate Correctly

An integrated-risk argument is a comparison tool, not a replacement for the definition of admissibility. To use it, first identify the proposed competitor and write the difference in the direction $R(\theta,\delta)-R(\theta,\eta)$. Next, establish the required pointwise sign. Only then does a zero integral imply equality almost everywhere. Reversing this order can lead to an invalid conclusion: an integral of zero alone does not establish domination or equality.

Define the comparison.
For the proposed rules, set $g(\theta)=R(\theta,\delta)-R(\theta,\eta)$, so positive values favor $\eta$.

Check the hypotheses.
Verify that the risk difference is measurable and integrable, and that domination would make it nonnegative at every parameter value.

Interpret the integral.
If the integrated difference is zero, conclude $g=0$ almost everywhere with respect to the chosen measure—not automatically at every parameter value.

Connect equality to admissibility.
Show that any strict improvement would have to occur on a set of positive measure, or use positive mass at every state in a finite parameter space. Without such a condition, a null-set improvement may remain undetected.

Two technical points deserve attention. First, the risk difference must be measurable for its integral and the sets used in the proof to be well-defined. Second, the integral must be meaningful. In particular, do not subtract two infinite integrated risks and call the result zero; that expression is undefined. Establish finite integrated risks, or directly establish integrability of the difference.

Common Mistake and Key Takeaway

Wrong reasoning: “The risk difference is nonnegative and its integral is zero, so the two rules have equal risk at every parameter value; therefore there is no strict improvement.”

Correction: The conclusion is equality only almost everywhere under the integrating measure. A strict improvement can occur on a null set, as the second worked example demonstrates. To rule out strict domination, also show that any strict-improvement set must receive positive measure, or use positive weights on every state in a finite parameter space.

Key takeaway: A zero integrated risk difference rules out a nonnegative risk improvement on any set of positive measure. It rules out strict domination altogether only when the measure cannot ignore the location of a possible strict improvement.

Check Your Understanding

Use the sign convention and the almost-everywhere conclusion carefully.

If $R(\theta,\delta)-R(\theta,\eta)\geq 0$ for every $\theta$, what does an integrated difference of zero imply?
Why does assigning positive probability to every state in a finite parameter space detect a strict improvement at any one state?
Can a nonnegative risk difference have integral zero and still be positive somewhere? State the condition that permits this.
Why is an integral of zero insufficient to establish that one rule dominates another?
What measurability and integrability checks are needed before applying the zero-integral argument?

Rule	\(R(\theta_1,\cdot)\)	\(R(\theta_2,\cdot)\)
\(\delta\)	4	2
\(\eta\)	3	2