Choosing Weights to Detect Risk Differences

Why the Choice of Weights Matters

Suppose two procedures have risk functions $R_0(\theta)$ and $R_1(\theta)$, and $R_1$ has strictly smaller risk on some region of the parameter space. A prior weight concentrated on that region can turn the local advantage into a strict advantage in integrated risk. The construction needs care: if the risks are unbounded, or if they can be signed, a weight that detects the difference need not give finite weighted risks.

The aim here is a precise selection principle. Starting from a measurable region of positive size where one risk is better, we will choose nonnegative weights supported on a subset where the improvement is uniformly positive and both risks are bounded in absolute value. Restricting to a finite-measure subset makes it possible to normalize the weights into a proper prior.

Definition: Let $(\Theta,\mathcal{A},\mu)$ be a parameter space with a sigma-finite measure $\mu$. A weight is a measurable function $w:\Theta\to[0,\infty)$. Its weighted risk for a risk function $R$ is $\int_\Theta R(\theta)w(\theta)\,d\mu(\theta)$, provided this integral is defined. We say the weighted risk is finite in the signed setting when $\int_\Theta |R(\theta)|w(\theta)\,d\mu(\theta)<\infty$. If $\int_\Theta w\,d\mu=1$, the weight defines a proper prior probability measure relative to $\mu$.

Sigma-finiteness means that $\Theta$ can be covered by countably many measurable sets of finite $\mu$-measure. The reference measure might be counting measure on a discrete space or Lebesgue measure on a Euclidean parameter space. We assume the risk functions are measurable and finite $\mu$-almost everywhere on the region under discussion. Write

D(\theta)=R_0(\theta)-R_1(\theta).

Thus $D(\theta)>0$ means procedure 1 has smaller risk at $\theta$. In standard decision problems with nonnegative loss, risks are nonnegative. The construction below is deliberately valid in the more general signed case too; that requires controlling each risk in absolute value.

A Weight-Selection Result

The useful set is not necessarily the whole region where $D>0$. On that region the improvement may approach zero, the risks may become arbitrarily large, or the reference measure of the region may be infinite. Instead, separate the requirements into three countable restrictions: keep only points where the improvement exceeds a positive threshold, where the risks are bounded in absolute value, and where the reference measure is finite.

Theorem: Suppose $(\Theta,\mathcal{A},\mu)$ is sigma-finite, $R_0$ and $R_1$ are measurable and finite $\mu$-almost everywhere, and $$ \mu\{\theta:R_0(\theta)>R_1(\theta)\}>0. $$ Then there is a measurable weight $w\geq 0$ with $\int w\,d\mu=1$ such that both weighted risks are finite and $$ \int_\Theta R_0(\theta)w(\theta)\,d\mu(\theta)> \int_\Theta R_1(\theta)w(\theta)\,d\mu(\theta). $$

Proof. Because the risks are finite almost everywhere, the set $S=\{\theta:D(\theta)>0,\ |R_0(\theta)|+|R_1(\theta)|<\infty\}$ has positive measure. Sigma-finiteness supplies a sequence of measurable sets $H_j$ of finite measure whose union is $\Theta$. For positive integers $j,k,l$, define

E_{jkl}=H_j\cap\{\theta:D(\theta)\geq 1/k\} \cap\{\theta:|R_0(\theta)|+|R_1(\theta)|\leq l\}.

The union of these countably many sets covers $S$: at any point of $S$, the positive number $D(\theta)$ is at least $1/k$ for some $k$, the finite quantity $|R_0(\theta)|+|R_1(\theta)|$ is at most some integer $l$, and the point belongs to some $H_j$. If every $E_{jkl}$ had measure zero, their countable union would have measure zero, contradicting $\mu(S)>0$. Hence one such set, say $E=E_{jkl}$, has positive measure. It also has finite measure because $E\subseteq H_j$.

Set $w(\theta)=\mathbf{1}_E(\theta)/\mu(E)$. This weight is nonnegative and integrates to one. On $E$, both absolute risks are bounded by $l$, so

\int_\Theta |R_i(\theta)|w(\theta)\,d\mu(\theta) \leq \frac{l\mu(E)}{\mu(E)}=l,\qquad i=0,1.

Thus both weighted risks are absolutely finite. Also $D\geq 1/k$ on $E$, and therefore

\int_\Theta R_0w\,d\mu-\int_\Theta R_1w\,d\mu =\int_\Theta Dw\,d\mu \geq \frac{\mu(E)/k}{\mu(E)} =\frac{1}{k}>0.

The equality is valid because both risks are absolutely integrable under $w$, so their integrals can be subtracted. This proves the claim. $\square$

The theorem is an existence result, not a recipe for finding the largest possible integrated improvement. Its point is that a positive-measure improvement region always contains a subset where improvement is bounded below and both risks are controlled. The normalized indicator weight on that subset is enough to certify a strict weighted comparison.

Worked Examples

Worked Example: A Signed Risk Comparison on an Interval

Let $\Theta=[0,1]$ with Lebesgue measure, and consider the measurable risk-comparison functions $R_0(\theta)=\theta$ and $R_1(\theta)=\theta-1$. Their difference is $D(\theta)=1$ throughout the interval. Use the uniform prior density $w(\theta)=1$ on $[0,1]$.

The weighted risks are

\int_0^1 R_0(\theta)w(\theta)\,d\theta =\int_0^1\theta\,d\theta=\frac12, \qquad \int_0^1 R_1(\theta)w(\theta)\,d\theta =\int_0^1(\theta-1)\,d\theta=-\frac12.

Both are finite, and the first exceeds the second by $1$. In particular, $\int_0^1|R_1(\theta)|\,d\theta=\int_0^1(1-\theta)\,d\theta=1/2$, so finiteness does not rely on cancellation between positive and negative parts. These functions illustrate the signed-integrability issue; they are not intended as risks from a nonnegative loss.

Worked Example: Unbounded Risks with a Finite Weighted Comparison

Let $\Theta=(0,\infty)$, with $R_0(\theta)=\theta^2+2$ and $R_1(\theta)=\theta^2+2-\theta$. Both functions are unbounded on $\Theta$, but $D(\theta)=\theta>0$. Choose $E=[1,2]$ and the prior density $w(\theta)=1$ on $E$, zero elsewhere. Since $E$ has length one, this is already normalized.

For $R_0$, direct integration gives

\int_1^2(\theta^2+2)\,d\theta =\left[\frac{\theta^3}{3}+2\theta\right]_1^2 =\frac{7}{3}+2 =\frac{13}{3}.

For $R_1$, subtract the integral of $\theta$:

\int_1^2(\theta^2+2-\theta)\,d\theta =\frac{13}{3}-\left[\frac{\theta^2}{2}\right]_1^2 =\frac{13}{3}-\frac32 =\frac{17}{6}.

The integrated improvement is $13/3-17/6=3/2$, which also equals $\int_1^2\theta\,d\theta$. On the selected interval the risks are bounded, so both weighted risks are finite despite their being unbounded over the full parameter space.

Worked Example: Geometric Weights on a Countable Parameter Space

Let $\Theta=\{1,2,\ldots\}$ with counting measure, and set $R_0(n)=(-1)^n n^2$ and $R_1(n)=(-1)^n n^2-1$. These signed functions satisfy $D(n)=1$ for every $n$. Choose prior weights $w(n)=2^{-n}$; they are nonnegative and sum to one.

The identity $\sum_{n=1}^\infty n^2x^n=x(1+x)/(1-x)^3$, valid for $|x|<1$, gives

\sum_{n=1}^\infty n^2 2^{-n}=6, \qquad \sum_{n=1}^\infty (-1)^n n^2 2^{-n}=-\frac{2}{27}.

Consequently the weighted risks are

\sum_{n=1}^\infty R_0(n)w(n)=-\frac{2}{27}, \qquad \sum_{n=1}^\infty R_1(n)w(n)=-\frac{2}{27}-1=-\frac{29}{27}.

To check finiteness without relying on the signed sums, use $|R_0(n)|=n^2$ and $|R_1(n)|\leq n^2+1$. Thus $\sum_n|R_0(n)|w(n)=6$, and $\sum_n|R_1(n)|w(n)\leq 6+1=7$. The weighted difference is $1$, as it must be since $D(n)=1$ at every point and the weights sum to one.

Why Absolute Risk Bounds Are Essential

In the signed setting, bounding $R_0+R_1$ does not bound either risk separately. For example, on $(0,\infty)$, take $R_0(\theta)=\theta$ and $R_1(\theta)=-\theta$. Then $R_0+R_1=0$ everywhere, but integrating either risk against the unnormalized indicator of $(0,\infty)$ does not give a finite integral. Cancellation in a sum cannot establish finiteness of its separate terms.

The construction in the theorem instead selects a set where $|R_0|+|R_1|\leq l$. This guarantees the bound on each absolute risk and hence on each weighted integral. If it is already known that both risks are nonnegative, a suitable bound on their sum does bound each risk; without that nonnegativity assumption, use absolute values.

Key takeaway: To convert strict risk improvement on a positive-measure region into a finite, proper weighted comparison, restrict to a positive-measure subset where the improvement is bounded below and the absolute risks are bounded above. Normalize the indicator of that subset to obtain prior weights.

Common Mistake

Wrong reasoning: “On the selected set, $R_0+R_1\leq K$, so each weighted risk is finite.” This does not follow for signed risks: one risk can be arbitrarily large and positive while the other is equally large and negative.
Correction: Either establish that both risks are nonnegative, or impose a bound such as $|R_0|+|R_1|\leq K$ on the selected set. Combine that bound with finite measure of the set and normalized nonnegative weights. Then each absolute weighted risk is finite.
Another caution: A strict improvement at a single parameter value need not have positive mass under a continuous reference measure. The selection argument requires a positive-measure improvement region (or a separate reason that the chosen prior assigns positive mass to the improved points).

Check Your Understanding

Use the definitions and selection argument above to answer these questions.

Why does sigma-finiteness help produce a positive-measure subset with finite measure?
In the theorem’s proof, what role does the condition $D\geq 1/k$ play after the weight is normalized?
Why does $|R_0|+|R_1|\leq l$ guarantee that both weighted risks are finite?
For $\Theta=(0,\infty)$, suppose $D(\theta)>0$ only on a set of Lebesgue measure zero. Does the theorem’s positive-measure hypothesis hold? Explain.
In the countable example, verify directly from $\sum_n w(n)=1$ and $D(n)=1$ that the weighted risk difference equals one.