Introduction
A risk comparison is made parameter value by parameter value. Two procedures may have exactly the same risk for some values, while one has lower risk for others. The key question is not whether they tie anywhere, but whether one procedure is no worse everywhere and strictly better somewhere.
The set of parameter values where the risks tie makes this logic visible. In particular, equality at a boundary does not automatically block domination: what matters is whether the competing procedure is no worse throughout the parameter space and achieves a strict improvement at at least one parameter value.
Equality and Strict-Improvement Sets
Let \(\delta\) and \(\eta\) be two procedures in the same decision problem, with finite risks \(R(\theta,\delta)\) and \(R(\theta,\eta)\) for every \(\theta\) in the parameter space \(\Theta\). Define the risk difference, equality set, and strict-improvement set by
A positive difference means that \(\eta\) has lower risk than \(\delta\). A zero difference means a tie. A negative difference means that \(\eta\) has higher risk.
Under weak improvement, every parameter value is either a tie or a strict improvement: \(E(\eta,\delta)\cup I(\eta,\delta)=\Theta\), and these sets are disjoint. Consequently, weak improvement with \(E(\eta,\delta)=\Theta\) is risk equality everywhere, not domination. The procedures themselves need not be identical for their risks to agree.
The Domination Test
Proof. By definition, domination means weak improvement everywhere together with strict improvement somewhere. Weak improvement is exactly the condition \(\Delta_{\eta,\delta}(\theta)\ge 0\) for every \(\theta\). Under that condition, strict improvement at \(\theta_0\) is exactly \(\Delta_{\eta,\delta}(\theta_0)>0\), which holds for some \(\theta_0\) if and only if the equality set is not all of \(\Theta\). Combining these statements proves both directions. \(\square\)
This criterion separates two issues that are easy to conflate. The equality set records where improvement is absent; it does not tell us whether there is a reversal elsewhere. Before using the equality set to conclude domination, first check the risk order at every parameter value.
Worked Example: Ties at Two Parameter Values
Consider a decision problem with no informative observation, parameter space \(\Theta=\{0,1,2\}\), and two available actions \(a\) and \(b\). Their losses are:
| Parameter \(\theta\) | Loss of \(a\) | Loss of \(b\) |
|---|---|---|
| 0 | 2 | 1 |
| 1 | 1 | 1 |
| 2 | 2 | 2 |
Let \(\delta\) always choose \(a\), and let \(\eta\) always choose \(b\). With no informative observation, each procedure's risk is its action's loss at the parameter value in question. Thus the risks of \(\delta\) are \((2,1,2)\), and those of \(\eta\) are \((1,1,2)\), in the order \(\theta=0,1,2\).
The risk differences \(R(\theta,\delta)-R(\theta,\eta)\) are \(1,0,0\). Therefore \(E(\eta,\delta)=\{1,2\}\) and \(I(\eta,\delta)=\{0\}\). The risks tie at two parameter values, but \(\eta\) is no worse anywhere and is strictly better at \(\theta=0\). It dominates \(\delta\).
Worked Example: Improvement at Some Values Is Not Enough
In another no-observation decision problem, let \(\Theta=\{0,1,2,3\}\), with losses for actions \(a\) and \(b\) as follows. Again, \(\delta\) always chooses \(a\), and \(\eta\) always chooses \(b\).
| Parameter \(\theta\) | Loss of \(a\) | Loss of \(b\) |
|---|---|---|
| 0 | 4 | 3 |
| 1 | 3 | 3 |
| 2 | 2 | 1 |
| 3 | 1 | 2 |
The risk differences are \(1,0,1,-1\). Thus \(\eta\) has strictly lower risk at \(\theta=0\) and \(\theta=2\), and ties at \(\theta=1\). But at \(\theta=3\), its risk is 2 rather than 1. Since \(\eta\) is worse at that parameter value, it does not weakly improve on \(\delta\), and therefore does not dominate it.
The equality and improvement sets are \(E(\eta,\delta)=\{1\}\) and \(I(\eta,\delta)=\{0,2\}\), but those sets do not account for the negative risk difference at \(\theta=3\). A complete comparison must check for such reversals as well as locate ties and improvements.
When the Parameter Space Has a Boundary
The definition of domination applies to every parameter value in \(\Theta\), including its boundary. A tie at an endpoint is simply one equality in the comparison. It cannot cancel a strict improvement at another parameter value, just as a tie at an interior value cannot.
Worked Example: Equality at an Endpoint
Let \(\Theta=[0,1]\) in a no-observation decision problem. Suppose two fixed actions have losses \(L(\theta,a)=1+\theta\) and \(L(\theta,b)=1\). Let \(\delta\) always choose \(a\) and \(\eta\) always choose \(b\). Their risks are therefore \(R(\theta,\delta)=1+\theta\) and \(R(\theta,\eta)=1\).
The risk difference is \(\Delta_{\eta,\delta}(\theta)=\theta\). At the boundary point \(\theta=0\), the procedures tie. For every \(\theta\in(0,1]\), the difference is positive, so \(\eta\) has strictly lower risk. Hence \(E(\eta,\delta)=\{0\}\), \(I(\eta,\delta)=(0,1]\), and \(\eta\) dominates \(\delta\).
The strict improvement does not have to occur at the boundary. It is enough that it occur at one parameter value in the space, while the risk comparison remains nonnegative everywhere else.
Continuity and the Location of Strict Improvement
When the risk difference varies continuously with the parameter, a strict improvement at one point persists nearby. At a boundary, “nearby” means nearby within the parameter space; the neighborhood need not extend outside it.
Proof. The interval \((0,\infty)\) is open in \(\mathbb{R}\). By continuity, its preimage under \(\Delta_{\eta,\delta}\) is open in \(\Theta\). That preimage is exactly \(I(\eta,\delta)\). The singleton \(\{0\}\) is closed in \(\mathbb{R}\); by continuity its preimage is closed in \(\Theta\), and that preimage is \(E(\eta,\delta)\). \(\square\)
For example, if \(\Theta=[0,1]\) and the risk difference is continuous and strictly positive at \(\theta=0\), then it remains positive on some relative neighborhood of 0, containing parameter values in \([0,1]\). Continuity does not, however, replace the global no-worse condition: a positive difference near one point does not rule out a negative difference elsewhere.
Common Mistakes
- Treating any tie as a failure of domination. A tie at one or many parameter values is compatible with domination. Check whether at least one strict improvement remains and whether there are any reversals.
- Counting strict improvements without checking the whole parameter space. A procedure can be better at several values and still fail to dominate because it is worse at another value.
- Assuming a boundary tie is special. The boundary is part of \(\Theta\), so its risk comparison must be included. But equality there is not an exception to the domination criterion.
- Confusing weak improvement with domination. If the risks are equal everywhere, neither procedure has a strict improvement over the other, even if their rules differ.
Check Your Understanding
For each question, distinguish ties, strict improvements, and any parameter values where the proposed improvement fails.
- If \(R(\theta,\eta)=R(\theta,\delta)\) for every \(\theta\), can \(\eta\) dominate \(\delta\)? Why or why not?
- On a finite parameter space, the risk differences \(R(\theta,\delta)-R(\theta,\eta)\) are \((0,2,0)\). Does \(\eta\) dominate \(\delta\)? Identify the equality and strict-improvement sets.
- Suppose the risk difference is positive at one parameter value but negative at another. What does this imply about domination by \(\eta\)?
- For \(\Theta=[0,1]\), suppose the risk difference is zero at 1 and positive for every \(\theta<1\). Does the endpoint tie prevent domination?
- If the risk difference is continuous and strictly positive at an endpoint of an interval parameter space, what can be said about strict improvement near that endpoint within the parameter space?