Common Failure Modes in Blyth's Method

Why Blyth's Method Needs Careful Diagnostics

Blyth's method turns a possible improvement in risk into a lower bound on an integrated risk difference. The key idea is simple: if a rule is uniformly dominated and strictly worse somewhere, then suitable prior measures should detect that local improvement. The proof can fail, however, if the measures do not put enough mass where the improvement occurs, if a posterior used to define a rule is not proper, or if continuity and tail limits are asserted without justification.

The goal here is to separate the valid inference from the assumptions that support it. In particular, weak convergence of prior measures does not by itself control integrals of unbounded risk functions, and positive mass on a large neighborhood does not guarantee positive mass on the smaller region where a risk gap is bounded away from zero.

The Risk-Deficit Bookkeeping

Definition: Let \(R(\theta,\delta)\) be the risk of a decision rule \(\delta\) at parameter \(\theta\). For a finite, nonzero measure \(\mu\) on the parameter space, define its integrated risk by \[ R_\mu(\delta)=\int R(\theta,\delta)\,\mu(d\theta). \] When the infimum is over the class of allowed measurable rules, define the Bayes risk \(B(\mu)=\inf_\delta R_\mu(\delta)\). For a specified rule \(\delta_0\) with finite integrated risk, its Bayes-risk deficit is \[ D(\mu;\delta_0)=R_\mu(\delta_0)-B(\mu). \]

These definitions use an unnormalized finite measure; multiplying \(\mu\) by a positive constant multiplies both integrated risks and the deficit by that constant. The deficit is nonnegative because \(B(\mu)\) is an infimum. It measures how much integrated risk \(\delta_0\) has in excess of the best integrated risk available under \(\mu\); it is not, by itself, a pointwise risk difference.

Suppose a rule \(\delta\) dominates \(\delta_0\), meaning \(R(\theta,\delta)\leq R(\theta,\delta_0)\) for every \(\theta\). Define the pointwise gap \(d(\theta)=R(\theta,\delta_0)-R(\theta,\delta)\), so \(d(\theta)\geq 0\). If both integrated risks are finite, the definition of the infimum gives \(B(\mu)\leq R_\mu(\delta)\). Therefore

D(\mu;\delta_0) =R_\mu(\delta_0)-B(\mu) \geq R_\mu(\delta_0)-R_\mu(\delta) =\int d(\theta)\,\mu(d\theta).

This inequality is the basic bookkeeping step. It says that a sequence with \(D(\mu_n;\delta_0)\to 0\) must also have \(\int d\,d\mu_n\to 0\) for every fixed finite-risk dominating rule. An admissibility argument then asks whether such vanishing integrals are compatible with strict improvement at any parameter value.

Worked Example: A Nonvanishing Local Risk Gap

Suppose \(d\) is continuous at \(\theta_0\) and \(d(\theta_0)>0\). Continuity implies that there are a neighborhood \(V\) of \(\theta_0\) and a number \(\varepsilon>0\) such that \(d(\theta)\geq\varepsilon\) on \(V\). Assume, in addition, that a sequence of finite prior measures satisfies \(\mu_n(V)\geq\kappa\) for every \(n\), where \(\kappa>0\). Then

D(\mu_n;\delta_0) \geq \int d(\theta)\,\mu_n(d\theta) \geq \int_V d(\theta)\,\mu_n(d\theta) \geq \varepsilon\mu_n(V) \geq \varepsilon\kappa.

Thus the deficits cannot tend to zero. The mass condition must hold on the neighborhood \(V\) where the gap is bounded below. A lower bound on mass in some larger neighborhood \(U\) is not enough: all the mass in \(U\) could lie outside \(V\).

A Local-Mass Criterion, with Its Exact Hypothesis

Theorem: Suppose \(R_\mu(\delta_0)\) and \(R_\mu(\delta)\) are finite for each measure in a sequence \((\mu_n)\), and suppose \(B(\mu_n)\) is finite. Let \(\delta\) dominate \(\delta_0\), with finite risks, and let \(d(\theta)=R(\theta,\delta_0)-R(\theta,\delta)\). If \(d\) is continuous at \(\theta_0\), \(d(\theta_0)>0\), and there are a neighborhood \(V\) of \(\theta_0\) and constants \(\varepsilon,\kappa>0\) such that \(d(\theta)\geq\varepsilon\) on \(V\) and \(\mu_n(V)\geq\kappa\) for every \(n\), then \(D(\mu_n;\delta_0)\geq\varepsilon\kappa\) for every \(n\). In particular, these deficits do not converge to zero.

Proof. Since \(\delta\) is an allowed rule, \(B(\mu_n)\leq R_{\mu_n}(\delta)\). Subtracting this inequality from \(R_{\mu_n}(\delta_0)\) and using the finiteness assumptions gives

D(\mu_n;\delta_0) \geq R_{\mu_n}(\delta_0)-R_{\mu_n}(\delta) =\int d(\theta)\,\mu_n(d\theta) \geq \int_V d(\theta)\,\mu_n(d\theta) \geq \varepsilon\mu_n(V) \geq \varepsilon\kappa.

The second inequality uses \(d\geq 0\) on the whole parameter space, which follows from domination; the next uses the lower bound on \(V\). The resulting positive constant is independent of \(n\), proving the claim. \(\square\)

A useful sufficient condition for an admissibility conclusion is that every nonempty open neighborhood \(V\) of every parameter value receives mass bounded away from zero along the sequence, and that the relevant risks are continuous. If a rule dominated \(\delta_0\) strictly at \(\theta_0\), continuity would give a neighborhood with a positive gap, and the theorem would contradict \(D(\mu_n;\delta_0)\to 0\). This is a conditional argument: the prior-mass property and the risk continuity must actually hold for the sequence and model being used.

Worked Example: Why Mass on a Larger Set Does Not Suffice

Let \(\Theta=\mathbb{R}\), and suppose the observation is constant. There are two actions, \(a_0\) and \(a_1\), with losses \(L(\theta,a_0)=d(\theta)\) and \(L(\theta,a_1)=0\), where \(d(\theta)=\max(1-4|\theta|,0)\). Let \(\delta_0\) always choose \(a_0\), and let \(\delta\) always choose \(a_1\). Then \(\delta\) dominates \(\delta_0\), \(d\) is continuous, and \(d(0)=1\).

Take \(U=(-1,1)\) and \(\mu_n=\delta_{1/2}\) for every \(n\), where \(\delta_{1/2}\) denotes unit mass at \(1/2\). Then \(\mu_n(U)=1\), but \(d(1/2)=\max(1-2,0)=0\). Hence \(R_{\mu_n}(\delta_0)=R_{\mu_n}(\delta)=0\), \(B(\mu_n)=0\), and \(D(\mu_n;\delta_0)=0\). Although continuity gives \(d(\theta)\geq 1/2\) when \(|\theta|<1/8\), these measures put no mass there. The needed hypothesis is a lower bound on mass in that smaller neighborhood, not merely in \(U\).

Failure Mode: Treating an Improper Prior as a Proper Posterior

An improper prior is a measure with infinite total mass, often used to define a generalized Bayes rule. It does not automatically produce a probability distribution after observing data. If the likelihood at observed value \(x\) is \(f_\theta(x)\), the posterior is proper only when its normalizing constant satisfies \(0<\int f_\theta(x)\,\pi(d\theta)<\infty\). Even when this holds, a posterior-based decision also requires the relevant posterior expected loss to be defined and a measurable rule to be chosen.

Worked Example: An Improper Prior with No Proper Posterior

Let \(X\mid\theta\sim N(\theta,1)\), with parameter space \(\mathbb{R}\), and take the improper prior with density \(\pi(\theta)=e^{\theta^2}\) relative to Lebesgue measure. For an observed \(x\), the unnormalized posterior density is proportional to

f_\theta(x)\pi(\theta) \propto \exp\left\{-\frac{(x-\theta)^2}{2}+\theta^2\right\} =\exp\left\{\frac{\theta^2}{2}+x\theta-\frac{x^2}{2}\right\}.

As \(|\theta|\to\infty\), the positive quadratic term \(\theta^2/2\) dominates the linear term \(x\theta\). More explicitly, if \(|\theta|\geq 4|x|\), then \(\theta^2/2+x\theta\geq\theta^2/4\). The integral of \(\exp(\theta^2/4)\) over either sufficiently distant tail is infinite. Thus the posterior normalizing integral is infinite for every \(x\); no posterior probability distribution is defined. It is invalid to invoke a posterior mean or posterior expected loss for this prior without a different, justified construction.

The diagnostic is to compute the normalizing integral, rather than infer posterior propriety from a familiar model or from the fact that the prior has a density. In a Blyth argument based on proper priors, a separate issue is whether the measures used in the integrated-risk comparison are finite and whether their Bayes risks are well-defined. A generalized Bayes rule and a sequence of proper priors are related tools, not interchangeable assumptions.

Failure Mode: Asserting Risk Continuity Without a Limit Argument

Continuity of the likelihood at each fixed data value does not alone establish continuity of risk. Risk integrates a loss against a distribution, and the contribution from extreme observations can change even when the integrand converges at each fixed observation. A standard sufficient condition is a dominated-convergence argument.

Sufficient condition for risk continuity: Let the observation have density \(p_\theta(x)\) with respect to a common measure \(\nu\), and let \(\delta\) be a fixed measurable rule. Suppose that whenever \(\theta_k\to\theta_0\), the functions \(L(\theta_k,\delta(x))p_{\theta_k}(x)\) converge to \(L(\theta_0,\delta(x))p_{\theta_0}(x)\) for \(\nu\)-almost every \(x\). Suppose also that, for all \(\theta\) in some neighborhood of \(\theta_0\), their absolute values are bounded by a single \(\nu\)-integrable function \(g(x)\). Then \(R(\theta,\delta)\) is continuous at \(\theta_0\).

Indeed, for any sequence \(\theta_k\to\theta_0\), the stated pointwise convergence and the common integrable bound permit the dominated convergence theorem: \[ R(\theta_k,\delta)=\int L(\theta_k,\delta(x))p_{\theta_k}(x)\,\nu(dx) \longrightarrow \int L(\theta_0,\delta(x))p_{\theta_0}(x)\,\nu(dx)=R(\theta_0,\delta). \] The parameter space is understood with its usual metric topology, so sequential continuity is continuity. In particular, continuity of a difference \(d(\theta)\) follows if each of its two risk functions is continuous and finite near \(\theta_0\).

Worked Example: Pointwise Continuity but Discontinuous Risk

Let \(X\) have a fixed Pareto density \(p(x)=x^{-2}\) for \(x\geq1\), independent of \(\theta\in[0,1]\), and use a fixed decision rule. Choose a continuous function \(\phi\) that is nonnegative, zero outside \([0,1]\), and positive on some subinterval of \((0,1)\). Define the loss, for that rule, by \(L(\theta,x)=\theta x^2\phi(\theta x)\) when \(\theta>0\), and \(L(0,x)=0\).

For each fixed \(x\), \(L(\theta,x)\to0=L(0,x)\) as \(\theta\to0\); the loss is pointwise continuous at zero. But for \(\theta>0\), its risk is

R(\theta,\delta) =\int_1^\infty \theta x^2\phi(\theta x)x^{-2}\,dx =\theta\int_1^\infty\phi(\theta x)\,dx =\int_\theta^\infty\phi(u)\,du.

The change of variable is \(u=\theta x\). Since \(\phi\) vanishes outside \([0,1]\), this integral is finite for every \(\theta>0\), while \(R(0,\delta)=0\). As \(\theta\downarrow0\), the risk tends to \(\int_0^1\phi(u)\,du>0\), not to zero. There is no common integrable bound of the kind in the sufficient condition; the losses at increasingly large \(x\) retain nonnegligible expected contribution. Thus pointwise continuity of the loss is not a substitute for controlling the integral.

Failure Mode: Uncontrolled Tails in Prior-Sequence Limits

Weak convergence controls integrals of bounded continuous functions, not arbitrary unbounded risk differences. Tail control is therefore essential when a proof passes from convergence of prior measures to convergence of integrated risks. Uniform integrability or an explicit tail bound may provide the needed control, but it must apply to the actual functions being integrated.

Worked Example: Weak Convergence Does Not Control an Unbounded Gap

On \(\Theta=[0,\infty)\), define \(\mu_n=(1-1/n)\delta_0+(1/n)\delta_{n^2}\) for integers \(n\geq2\). For every bounded continuous function \(h\), \[ \int h\,d\mu_n=(1-1/n)h(0)+(1/n)h(n^2)\longrightarrow h(0), \] because the second term has absolute value at most \(\|h\|_\infty/n\). Thus \(\mu_n\) converges weakly to unit mass at zero.

Now take the nonnegative unbounded function \(d(\theta)=\theta\). Direct calculation gives \[ \int d(\theta)\,\mu_n(d\theta) =(1-1/n)0+(1/n)n^2=n, \] which diverges rather than converging to \(\int d\,d\delta_0=0\). This can be made an exact risk-gap example: with a constant observation, let the losses of two actions be \(L(\theta,a_0)=\theta\) and \(L(\theta,a_1)=0\). The rule choosing \(a_1\) dominates the rule choosing \(a_0\); the deficit for the latter is \(D(\mu_n;\delta_0)=n\). Every \(\mu_n\) is a finite probability measure and every integrated risk here is finite, but weak convergence alone says nothing about convergence of these unbounded risk integrals.

How to Audit a Blyth-Style Argument

A reliable proof makes each link in the chain explicit. First verify that the candidate rule and the prior sequence have well-defined integrated risks. Next, if a posterior rule is used, check the posterior normalizing constant and the posterior risk or loss condition needed to define the decision. Then identify the pointwise risk gap against an arbitrary putative dominator. A claim that the gap is positive nearby needs a continuity argument; a claim that prior measures detect that gap needs a mass bound on that same neighborhood. Finally, any limit involving integrated risk needs its own tail or domination justification.

Check definitions and finiteness.
Confirm that the risks, prior integrals, Bayes risks, and deficits used in the proof are defined under the stated measures.

Form the pointwise gap.
For a proposed dominator \(\delta\), write \(d(\theta)=R(\theta,\delta_0)-R(\theta,\delta)\) and verify \(d\geq0\).

Justify local positivity and mass.
Establish continuity (or another valid lower-bound argument) on a neighborhood where \(d\geq\varepsilon\), and establish \(\mu_n(V)\geq\kappa\) on that very neighborhood.

Control every limit.
When prior measures or risk integrals converge, state the theorem and the integrability, domination, or tail hypotheses that justify the passage.

The most common logical shortcut is to say that strict improvement at one point, continuity, and nonvanishing prior mass somewhere nearby force a positive integrated gap. The conclusion follows only if the mass remains in a region where the gap has a positive lower bound. Similarly, a generalized Bayes label does not certify posterior propriety, and weak convergence does not certify convergence of unbounded risk integrals. These are not technical decorations: each protects a different step of the admissibility argument.

Key takeaway: To rule out domination using vanishing Bayes-risk deficits, prove the inequality \(D(\mu_n;\delta_0)\geq\int d\,d\mu_n\), then show that the measures actually detect a positive portion of \(d\). The local mass, posterior, continuity, and tail conditions must be verified where they are used.

Check Your Understanding

For each question, identify the precise hypothesis that an admissibility proof would need.

Why does a uniform lower bound on \(\mu_n(U)\) fail to guarantee a positive integrated gap if the gap is only known to be positive on a smaller set \(V\subset U\)?
State the inequality connecting the Bayes-risk deficit \(D(\mu;\delta_0)\) to the integrated risk gap for a rule that dominates \(\delta_0\).
What must be checked before an improper prior can be used to define a posterior-based decision rule?
Give the two main ingredients in the dominated-convergence condition that can establish continuity of risk.
Why does weak convergence of \(\mu_n\) not, by itself, imply convergence of \(\int d\,d\mu_n\) when \(d\) is unbounded?