Why the Action Space Matters
A decision rule can be judged only against procedures it is permitted to compete with. If a scientific protocol requires estimates to be nonnegative, for instance, rules that report negative estimates are not available competitors in that constrained problem. This changes the class of procedures, and therefore can change which rules are admissible, even if the data model, parameter space, and loss function remain exactly the same.
Keep two restrictions separate. The parameter space says which values of the unknown parameter are considered; the action space says which decisions a procedure may report. Changing either can affect risk comparisons, but in different ways. In particular, restricting actions does not itself change the risk of a rule that remains available: it changes which other rules may be used to challenge that rule.
For the comparisons below, assume the risks being compared are finite. If a rule belongs to both procedure classes, its risk is calculated from the same data distribution and loss in either class, so its risk function is unchanged. The restriction affects its available competitors, not this calculation.
What Nested Action Spaces Imply
If \(A_0\subseteq A\), every procedure taking values in \(A_0\) is also a procedure taking values in \(A\). Thus a restricted class has fewer potential dominators. The following implication makes the direction of the effect precise.
Proof. If \(\eta\) were inadmissible in \(\mathcal D_{A_0}\), some \(\delta\in\mathcal D_{A_0}\) would have risk no greater than \(\eta\)'s risk at every \(\theta\), and strictly smaller at some \(\theta\). Since \(\delta\in\mathcal D_A\) as well, the same risk inequalities would make it a dominator in \(\mathcal D_A\). This contradicts admissibility of \(\eta\) there. Thus \(\eta\) must also be admissible in the restricted class. The argument does not establish the converse: a dominator in the larger class might use an action excluded from \(A_0\). \(\square\)
Worked Example: A Removed Action Changes Admissibility
There is no informative observation, the parameter space is \(\Theta=\{0\}\), and the loss is squared error, \(L(\theta,a)=(a-\theta)^2\). Compare the unrestricted action space \(A=\mathbb R\) with the restricted space \(A_0=\{1,2\}\). A procedure is simply a constant action.
Consider the rule \(\eta=1\). Its risk at the only parameter value is \(R(0,\eta)=(1-0)^2=1\). In the unrestricted class, the action \(0\) has risk \((0-0)^2=0\), so it strictly dominates \(\eta\). In the restricted class, the only other action is \(2\), with risk \((2-0)^2=4\). It does not dominate \(\eta\), and \(\eta\) is the only restricted action with risk no greater than 1. Hence \(\eta\) is admissible in \(\mathcal D_{A_0}\), but not in \(\mathcal D_A\).
This simple example isolates the action-space effect: \(\Theta\) and the loss did not change. The rule became admissible because its better competitor was removed. The example does not claim that every restriction creates a newly admissible rule; it shows why the reverse implication in the theorem can fail.
Projection Under Squared-Error Loss
A common estimation constraint is that reported values must lie in a specified interval. For squared-error loss, projecting an estimate onto an interval containing the parameter space cannot increase its error at any data value. This gives a direct risk comparison while holding the parameter space fixed.
Proof. Fix \(\theta\in\Theta\), so \(l\leq\theta\leq u\). If \(X\in[l,u]\), projection does not change \(X\), and the squared errors are equal. If \(X<l\), then \(X\leq l\leq\theta\), so \(0\leq\theta-l\leq\theta-X\); squaring gives \((\delta_C(X)-\theta)^2\leq(X-\theta)^2\). If \(X>u\), then \(\theta\leq u<X\), so \(0\leq u-\theta\leq X-\theta\), and again \((\delta_C(X)-\theta)^2\leq(X-\theta)^2\). Thus the inequality holds for every observed value. Taking expectations proves the risk inequality. On \(X<l\) or \(X>u\), the corresponding inequality is strict, since \(\theta\in[l,u]\). Positive probability of either event therefore makes the expected inequality strict. \(\square\)
Worked Example: Clipping a Normal-Mean Estimate to an Interval
Suppose \(X\sim N(\theta,1)\), the parameter space is \(\Theta=[0,1]\), and loss is squared error. Compare the unrestricted rule \(\delta(X)=X\) with the rule constrained to report values in \([0,1]\), \(\delta_C(X)=\min\{1,\max\{0,X\}\}\). The parameter space is the same in both comparisons.
At the possible observation \(X=1.3\), projection reports \(1\). If \(\theta=0.4\), the squared errors are \((1-0.4)^2=0.36\) and \((1.3-0.4)^2=0.81\), respectively. At \(X=-0.2\), projection reports \(0\), giving squared error \((0-0.4)^2=0.16\), compared with \((-0.2-0.4)^2=0.36\). For observations inside \([0,1]\), the two rules agree.
The theorem applies because \(0.4\in[0,1]\), and likewise applies at every parameter value in the stated parameter space. Under this normal model, observations outside the interval have positive probability, so the risk inequality is strict at every \(\theta\in[0,1]\). The comparison shows how restricting the reported action can yield a rule with smaller risk than the unrestricted estimate on a fixed parameter space.
Constraints Need Not Improve Risk Everywhere
Projection helps in the preceding result because the allowed interval contains every parameter value under consideration. If the action constraint and parameter space are misaligned, clipping may instead increase risk at some parameter values. A risk improvement on part of the parameter space is not, by itself, domination: domination requires the inequality at every parameter value being considered.
Worked Example: Nonnegative Estimates with an Unrestricted Parameter
Let \(X\sim N(\theta,1)\), now with fixed parameter space \(\Theta=\mathbb R\) and squared-error loss. The unrestricted rule is \(\delta(X)=X\). The rule \(\delta_+(X)=\max\{X,0\}\) obeys the action constraint \(A_0=[0,\infty)\). The unrestricted rule has risk \(R(\theta,\delta)=1\) at every \(\theta\), since \(X-\theta\sim N(0,1)\).
At \(\theta=0\), \(X\sim N(0,1)\). The constrained rule reports \(X\) when \(X>0\) and 0 otherwise, so \(R(0,\delta_+)=E[X^2\mathbf 1_{\{X>0\}}]=\tfrac12\). The last equality follows from symmetry of the standard normal density and \(E[X^2]=1\). Thus its risk is smaller than 1 at \(\theta=0\).
At \(\theta=-1\), write \(X=-1+Z\), where \(Z\sim N(0,1)\). The rule reports 0 when \(Z\leq1\), incurring squared error 1, and reports \(X\) when \(Z>1\), incurring squared error \(Z^2\). Consequently, \(R(-1,\delta_+)=\Phi(1)+\int_1^\infty z^2\phi(z)\,dz\), where \(\phi\) and \(\Phi\) are the standard normal density and distribution function. Since \(\phi'(z)=-z\phi(z)\), integration by parts gives \(\int_1^\infty z^2\phi(z)\,dz=\phi(1)+1-\Phi(1)\). Therefore \(R(-1,\delta_+)=1+\phi(1)\), approximately \(1.242\), which exceeds the unrestricted rule’s risk of 1.
The action-constrained rule improves risk at \(\theta=0\) but has higher risk at \(\theta=-1\). It does not dominate the unrestricted rule over \(\mathbb R\). More fundamentally, \(X\) is not an allowable competitor in the constrained class when it can be negative. A risk comparison with that rule is informative, but it is not a dominance comparison within \(\mathcal D_{A_0}\).
Common Mistakes and Practical Consequences
When analyzing a constrained problem, first specify the parameter space, action space, loss, and procedure class. Then check that each proposed competitor belongs to that class before using its risk to claim dominance. This prevents two frequent errors.
- Wrong reasoning: “The restricted rule has lower risk than the unrestricted rule at some parameter values, so it dominates.” Correction: dominance requires weak improvement everywhere and strict improvement somewhere; a single parameter value with higher risk defeats it.
- Wrong reasoning: “The constrained rule dominates the unrestricted one, so it must be inadmissible among constrained rules.” Correction: the unrestricted rule may not be an allowable constrained competitor. Even a valid comparison across classes does not establish a dominator within the restricted class.
- Wrong reasoning: “Restricting actions changes the risk of every procedure.” Correction: a rule still available in both classes has the same risk function when the model and loss are unchanged. What changes is the set of procedures available for comparison.
- Wrong reasoning: “The parameter space and action space are interchangeable restrictions.” Correction: the parameter space indexes the distributions and risk comparisons; the action space limits the decisions the procedure can report. State which one changed before drawing an admissibility conclusion.
Action constraints matter in estimation because they encode practical or scientific requirements, such as nonnegative estimates or probabilities restricted to \([0,1]\). They can remove potential dominators, and can also motivate procedures whose risks differ from unrestricted estimates. Admissibility is therefore always relative to a stated class of available procedures—not a property inferred from one risk comparison alone.
Check Your Understanding
Use the same data model and loss unless a question says otherwise.
- If \(\mathcal D_{A_0}\subseteq\mathcal D_A\), which direction of the admissibility implication is guaranteed? Explain why the reverse need not hold.
- For squared-error loss and \(\Theta\subseteq[l,u]\), why does projecting an observation below \(l\) onto \(l\) reduce its squared error?
- In the normal example with \(\Theta=\mathbb R\), which parameter value demonstrates that the nonnegative rule does not dominate \(X\)?
- Why is a risk comparison between a constrained rule and an excluded unrestricted rule not, by itself, a dominance comparison within the constrained class?
- If a shared rule belongs to both procedure classes and the model and loss are unchanged, does restricting the action space alter that rule’s risk function? Why?