Countable Risk-Dense Families and Essential Completeness

Why Countable Risk Approximation Matters

A decision rule is often described by its risk function: the expected loss it incurs at each parameter value. A useful reduction is to find a countable collection of rules whose risks approximate the risks of every rule in the decision problem. Countability makes it possible to organize approximations into sequences and to use a diagonal argument when the parameter space is countable.

But pointwise approximation is a limited guarantee. At each fixed parameter value, the risk can be made close; the rule doing the approximating may depend on the parameter value. That does not automatically provide one rule that is nearly no worse at all parameter values at once. This distinction is central when approximation is used in a complete-class argument.

Risk Functions and Sequential Density

The word “sequentially” matters. Pointwise convergence of risk functions means convergence separately at each parameter. It does not assert uniform convergence over \(\Theta\). Nor, when \(\Theta\) is arbitrary, is sequential density equivalent to topological density for the topology of pointwise convergence.

A basic neighborhood of a risk function \(R\) in the pointwise topology specifies tolerances at only finitely many parameter values. Thus topological density means that every finite set of coordinates can be approximated simultaneously. In an uncountable product space, finite-coordinate approximation need not be achievable by a single convergent sequence.

Finite-Coordinate Approximation and a Diagonal Argument

Proof. Fix \(\delta\in\mathcal D\). For each positive integer \(m\), apply the hypothesis to \(K=\{\theta_1,\ldots,\theta_m\}\) and \(\varepsilon=1/m\). It gives a rule \(\gamma_m\in\mathcal C\) satisfying

Fix any index \(j\). For every \(m\geq j\), the displayed inequality applies to \(\theta_j\). Consequently, as \(m\to\infty\), the difference at \(\theta_j\) tends to zero. Since this holds for every \(j\), \(R_{\gamma_m}(\theta)\to R_\delta(\theta)\) at every \(\theta\in\Theta\). \(\square\)

This argument explains the role of a countable parameter space: it lets us satisfy successively larger finite lists of requirements, while ensuring each fixed coordinate eventually remains on the list. It is a sufficient condition, not a claim that every decision problem automatically has such a family.

A Statistical Sufficient Condition: Rational Rules

A common route to countable risk-dense families is to approximate actions by a countable grid. A uniform bound on how much loss changes with the action turns that action approximation into a risk approximation.

Proof. There are countably many rational-valued assignments on a finite set, so the proposed family is countable. Fix a rule \(\delta:\mathcal X\to[0,1]\) and a positive integer \(m\). For each \(x\), choose a rational \(q_m(x)\in[0,1]\) with \(\lvert q_m(x)-\delta(x)\rvert<1/m\). The Lipschitz condition gives, for every \(\theta,x\),

Taking expectations under the observation distribution at \(\theta\), and using \(\lvert E[Y]\rvert\leq E[\lvert Y\rvert]\) for an integrable random variable \(Y\), gives

The bound holds for every \(\theta\), so the convergence is uniform in the parameter, and in particular is pointwise. \(\square\)

Pointwise Density Is Not Approximate Domination

Sequential pointwise risk density asks for two-sided closeness at each fixed parameter, eventually. Approximate domination asks for a single rule that stays below a specified upper bound at all parameter values simultaneously. On a finite parameter space, pointwise convergence along a sequence implies approximate domination: there are only finitely many coordinates, so one sequence member can satisfy all the \(\varepsilon\)-bounds at once. For an infinite parameter space, this conclusion can fail.

Pointwise Topology: A Necessary Qualification

When \(\Theta\) is countable, pointwise topology can be described by a metric. If \(\Theta=\{\theta_j:j\geq1\}\), one such metric on finite-valued risk functions is

Convergence under this metric is equivalent to pointwise convergence: convergence in the metric forces each summand, and hence each fixed-coordinate difference, to tend to zero; conversely, pointwise convergence makes any finite initial sum tend to zero, while the remaining tail is bounded by \(\sum_{j>J}2^{-j}\), which can be made arbitrarily small. Thus for countable \(\Theta\), topological density and sequential density agree.

For uncountable \(\Theta\), do not make that identification without an additional condition. For example, in \(\{0,1\}^I\) with uncountable index set \(I\), the vectors with finite support are topologically dense: any neighborhood restricts only finitely many coordinates, and a finite-support vector can match them. But no sequence of finite-support vectors converges pointwise to the all-ones vector. The union of the supports in a sequence is countable, so some coordinate lies outside that union and remains zero throughout. This is why “dense” must be qualified as topological or sequential when the parameter space is unrestricted.

Common Mistake and Practical Use

• Mistake: “A countable family is risk-dense, so one of its rules is nearly as good as the target rule at every parameter.” Pointwise convergence only guarantees closeness eventually at each fixed parameter; the index needed can vary with the parameter.
• Correction: Check the required quantifiers. Pointwise density gives, for each \(\theta\), eventual closeness at that \(\theta\). Uniform approximate domination requires one selected rule to satisfy the upper risk bound for all \(\theta\). Exact essential completeness requires a rule with no greater risk anywhere.
• Mistake: Treating finite-coordinate topological density as sequential density on an arbitrary parameter space.
• Correction: State a countability or first-countability condition, or use the precise finite-coordinate formulation instead of claiming a sequence exists.

Countable risk-dense families are useful because they can reduce risk approximation questions to a manageable collection of rules. The diagonal theorem gives a reliable route from finite-coordinate approximation to pointwise approximation when \(\Theta\) is countable. To draw a stronger conclusion about replacement or completeness, however, one needs a separate argument establishing simultaneous upper bounds or exact domination; risk density alone supplies neither.