Risk Continuity and Extending Almost-Everywhere Comparisons

Why Null Sets Matter in Risk Comparisons

In admissibility arguments, an integrated comparison may first yield a statement about risks that holds only almost everywhere under a prior. That is not automatically a comparison at every parameter value: a prior can assign probability zero to points, or even to entire regions. When the risks are continuous and the prior puts positive mass in every neighborhood of each parameter value, however, a strict reversal at any point would persist on a neighborhood. That neighborhood would have positive prior mass, contradicting the almost-everywhere comparison.

The issue is therefore a precise interaction between two properties: continuity of the risks and where the prior places mass. Neither property can be dropped without further conditions.

Definition: Let \(\Theta\) be a topological parameter space and \(\pi\) a probability measure on its Borel sets. The prior \(\pi\) has full topological support on \(\Theta\) if every nonempty relatively open subset of \(\Theta\) has positive \(\pi\)-probability. A property holds \(\pi\)-almost everywhere if the set where it fails has \(\pi\)-probability zero.

For a decision rule \(\delta\), its risk at parameter value \(\theta\) is \(R(\theta,\delta)=E_\theta[L(\theta,\delta(X))]\), when this expectation exists. The comparison of interest concerns two such risk functions. It is useful to write their difference as \(q(\theta)=R(\theta,\delta_1)-R(\theta,\delta_0)\); then \(q(\theta)\leq 0\) means that \(\delta_1\) has no greater risk than \(\delta_0\) at \(\theta\).

The Continuity Extension Theorem

The central result is topological: a continuous function that is nonpositive almost everywhere under a full-support measure must be nonpositive everywhere. The statement applies to risk differences because the difference of two continuous risk functions is continuous.

Theorem: Let \(\Theta\) be a topological space, let \(\pi\) have full topological support on \(\Theta\), and let \(q:\Theta\to\mathbb{R}\) be continuous. If \(q(\theta)\leq 0\) for \(\pi\)-almost every \(\theta\), then \(q(\theta)\leq 0\) for every \(\theta\in\Theta\).

Proof. Suppose, to the contrary, that there is a point \(\theta_0\in\Theta\) with \(q(\theta_0)>0\). Set \(\varepsilon=q(\theta_0)/2>0\). By continuity, the set \(U=\{\theta\in\Theta:q(\theta)>\varepsilon\}\) is an open neighborhood of \(\theta_0\). It is nonempty, so full support gives \(\pi(U)>0\). But \(q(\theta)>\varepsilon>0\) throughout \(U\), contradicting \(q\leq0\) \(\pi\)-almost everywhere. Thus no such \(\theta_0\) exists, and \(q\leq0\) everywhere. \(\square\)

Taking \(q=R(\,\cdot\,,\delta_1)-R(\,\cdot\,,\delta_0)\), the theorem gives the risk comparison pointwise, provided both risks are finite and continuous and the prior has full support. If \(\Theta\) is a subset of Euclidean space, “open” here means relatively open in \(\Theta\). For example, a density positive Lebesgue-almost everywhere on an interval gives positive mass to every nonempty relatively open subset of that interval.

Worked Example: Extending a Comparison to the Endpoints

Suppose \(X\sim N(\theta,1)\), with parameter space \(\Theta=[-1,1]\), and use squared-error loss \(L(\theta,a)=(a-\theta)^2\). Compare \(\delta_1(X)=0\) with \(\delta_0(X)=X\). Let the prior be uniform on \([-1,1]\).

For the constant rule, \(R(\theta,\delta_1)=E_\theta[(0-\theta)^2]=\theta^2\). For the rule \(\delta_0(X)=X\), \(X-\theta\sim N(0,1)\), so \(R(\theta,\delta_0)=E_\theta[(X-\theta)^2]=1\). Hence \(q(\theta)=R(\theta,\delta_1)-R(\theta,\delta_0)=\theta^2-1\).

For \(-1<\theta<1\), \(\theta^2-1<0\); at \(\theta=-1\) and \(\theta=1\), it equals zero. The prior assigns zero probability to either endpoint, so an almost-everywhere comparison alone would not state what happens there. Both risks are continuous, and the uniform prior has full support relative to \([-1,1]\). The theorem therefore gives \(R(\theta,\delta_1)\leq R(\theta,\delta_0)\) at every \(\theta\in[-1,1]\), including both endpoints.

What an Integrated Inequality Does—and Does Not—Show

An integrated risk inequality is not, by itself, a pointwise comparison. For example, \(\int q(\theta)\,\pi(d\theta)\leq0\) allows \(q\) to be positive on some regions if it is sufficiently negative elsewhere. A route from an integral to an almost-everywhere statement requires an additional sign condition.

Result: If \(h\) is measurable, \(h\geq0\) \(\pi\)-almost everywhere, and \(\int h\,d\pi=0\), then \(h=0\) \(\pi\)-almost everywhere.

Proof. For each positive integer \(m\), let \(A_m=\{\theta:h(\theta)\geq 1/m\}\). Nonnegativity gives \(0=\int h\,d\pi\geq (1/m)\pi(A_m)\), so \(\pi(A_m)=0\). Every point where \(h>0\) belongs to \(A_m\) for some positive integer \(m\). Thus \(\{h>0\}=\bigcup_{m=1}^{\infty}A_m\) has probability zero, proving the result. \(\square\)

For instance, if a known nonnegative risk deficit has prior integral zero, this result first makes the deficit zero \(\pi\)-almost everywhere. The continuity extension theorem can then carry that equality to all parameter values, if the prior has full support and the deficit is continuous. Without the nonnegativity condition, a zero integral can result from positive and negative parts cancelling; without full support, an almost-everywhere conclusion need not reach every parameter value.

Worked Example: A Prior That Does Not Reach the Whole Parameter Space

Keep \(X\sim N(\theta,1)\), squared-error loss, and the same rules \(\delta_1(X)=0\) and \(\delta_0(X)=X\), but now take \(\Theta=\mathbb{R}\) and let the prior be uniform on \([-1,1]\). The risks and their difference are still \(R(\theta,\delta_1)=\theta^2\), \(R(\theta,\delta_0)=1\), and \(q(\theta)=\theta^2-1\).

The prior-almost-everywhere comparison \(q\leq0\) holds, since \(\theta^2\leq1\) throughout the prior’s support. Its integrated difference is \(\int q\,d\pi=\frac{1}{2}\int_{-1}^{1}(\theta^2-1)\,d\theta=\frac{1}{2}(2/3-2)=-2/3\). Nevertheless, at \(\theta=2\), \(q(2)=3>0\): the comparison fails.

There is no contradiction with the theorem. The prior gives zero mass to open neighborhoods lying outside \([-1,1]\), so it does not have full topological support on \(\mathbb{R}\). Continuity can only extend the comparison through regions that the prior’s support reaches.

Checking Risk Continuity

Continuity of a risk function is an assumption to verify, not a consequence of having a statistical model. One common sufficient condition follows directly from the dominated convergence theorem. Let the distributions of \(X\) have densities or mass functions \(f_\theta\) with respect to a common measure \(\mu\). For a fixed measurable rule \(\delta\), write \(R(\theta,\delta)=\int L(\theta,\delta(x))f_\theta(x)\,\mu(dx)\).

Proposition: Suppose that for every sequence \(\theta_k\to\theta\), the integrands \(L(\theta_k,\delta(x))f_{\theta_k}(x)\) converge to \(L(\theta,\delta(x))f_\theta(x)\) for \(\mu\)-almost every \(x\). Suppose also that there is a \(\mu\)-integrable function \(g\) such that, for all sufficiently large \(k\), their absolute values are bounded by \(g(x)\). Then \(R(\theta_k,\delta)\to R(\theta,\delta)\). If these assumptions hold for every convergent parameter sequence, the risk is continuous.

Proof. For each sufficiently large \(k\), the integrand is measurable and bounded in absolute value by the same integrable function \(g\). It converges \(\mu\)-almost everywhere to the integrand defining \(R(\theta,\delta)\). The dominated convergence theorem therefore gives \(\lim_{k\to\infty}\int L(\theta_k,\delta(x))f_{\theta_k}(x)\,\mu(dx)=\int L(\theta,\delta(x))f_\theta(x)\,\mu(dx)\). The finitely many earlier terms do not affect the limit. Thus \(R(\theta_k,\delta)\to R(\theta,\delta)\). In a metric parameter space, sequential continuity is equivalent to continuity, proving the claim. \(\square\)

Worked Example: Verifying Continuity in a Normal Location Model

Let \(X\sim N(\theta,1)\), let \(\delta(X)=X\), and use squared-error loss. The risk is \(R(\theta,\delta)=\int (x-\theta)^2\phi(x-\theta)\,dx\), where \(\phi\) is the standard normal density. With the substitution \(z=x-\theta\), this is \(\int z^2\phi(z)\,dz=1\), so the risk is constant and hence continuous.

The proposition also verifies continuity directly. For \(\theta_k\to\theta\), the integrand \((x-\theta_k)^2\phi(x-\theta_k)\) converges pointwise to \((x-\theta)^2\phi(x-\theta)\). Choose \(M\) so that \(|\theta_k|\leq M\) for all sufficiently large \(k\). Then \((x-\theta_k)^2\leq 2x^2+2M^2\). Also, \(|x-\theta_k|\geq (|x|-M)_+\), where \(u_+=\max(u,0)\). Consequently, \[ (x-\theta_k)^2\phi(x-\theta_k) \leq \frac{2x^2+2M^2}{\sqrt{2\pi}}\exp\!\left(-\frac{((|x|-M)_+)^2}{2}\right). \] The right-hand side is integrable: it is bounded on bounded intervals and has a Gaussian tail multiplied by a quadratic polynomial. Dominated convergence applies, confirming continuity of the risk.

Common Mistakes and Their Corrections

Treating an integrated inequality as pointwise dominance. A negative integrated difference may hide a positive difference on part of the parameter space. First establish a pointwise almost-everywhere comparison, for example through a nonnegative deficit with integral zero.
Ignoring the prior’s support. An almost-everywhere statement under a prior says nothing directly about open regions to which that prior assigns zero probability. Check full support on the parameter space where the conclusion is wanted.
Extending across a discontinuity. A function can satisfy a comparison almost everywhere and violate it at a single point. Continuity of both risks, or equivalently of their difference, is essential to this extension argument.

Worked Example: Why Continuity Cannot Be Omitted

Let \(X=\theta\) deterministically, with \(\Theta=[-1,1]\), squared-error loss, and a uniform prior. Consider \(\delta_1(x)=\mathbf{1}\{x=0\}\) and \(\delta_0(x)=0\). Because \(X=\theta\), the risks are \(R(\theta,\delta_1)=(\mathbf{1}\{\theta=0\}-\theta)^2\) and \(R(\theta,\delta_0)=\theta^2\).

For every \(\theta\neq0\), \(\mathbf{1}\{\theta=0\}=0\), so both risks equal \(\theta^2\). At \(\theta=0\), however, \(R(0,\delta_1)=1\) while \(R(0,\delta_0)=0\). Thus \(R(\theta,\delta_1)\leq R(\theta,\delta_0)\) holds prior-almost everywhere but fails at zero. The prior has full support, but \(R(\theta,\delta_1)\) is discontinuous at zero, so the continuity theorem does not apply.

The practical lesson is to check the two conditions separately: the prior must reach every neighborhood relevant to the conclusion, and the risk difference must be continuous there. When both hold, a comparison that excludes only a prior-null set cannot conceal a strict reversal at an isolated point or elsewhere in the parameter space.

Check Your Understanding

Use the support and continuity conditions to decide what an almost-everywhere risk comparison actually implies.

State the definition of full topological support for a prior on a parameter space.
Why does continuity rule out a point where a risk difference is strictly positive if the difference is nonpositive almost everywhere under a full-support prior?
What additional sign condition lets a zero integrated risk deficit imply that the deficit is zero almost everywhere?
In the normal-model example with prior uniform on \([-1,1]\), why does the prior not support an extension of the comparison to all of \(\mathbb{R}\)?
In the deterministic-observation example, which condition for the extension theorem fails, and at which parameter value?