Smoothness and Boundary Behavior of Target Maps

Why Boundary Behavior Matters

A target map converts response parameters into desired treatment allocation shares. In practice, those parameters are estimated from accumulating data, so the map is evaluated at estimates that may lie on, or close to, the edge of their possible range. A target that behaves well for interior parameter values can be undefined or unstable at a boundary. That matters because the resulting target shares are inputs to the allocation rule; a formula that is undefined cannot provide a target, and a sharply changing target can make allocations sensitive to small changes in the estimates.

The question here is not whether the target is scientifically desirable. It is whether its mathematical definition continues to make sense across the parameter values the estimator can produce. The introductory material on plugging response estimates into a prespecified target map establishes the plug-in operation; this tutorial examines the conditions that make the map itself well-behaved.

Target Maps and Their Domain

Let there be \(K\) treatment arms, indexed by \(k=1,\ldots,K\). Let \(\theta=(\theta_1,\ldots,\theta_K)\) denote the vector of response parameters, and let \(\Theta\) be the set of parameter values allowed by the model. A target map assigns a vector of shares \(\rho(\theta)=(\rho_1(\theta),\ldots,\rho_K(\theta))\). With estimated parameters, the plug-in target is \(\rho(\hat{\theta})\), provided \(\hat{\theta}\) is in the map's domain and the map is defined there.

Definition: A valid target share vector has \(\rho_k(\theta)\geq 0\) for every arm and \(\sum_{k=1}^K\rho_k(\theta)=1\). The target map is well-defined on a set if it assigns exactly one such vector to every parameter value in that set. It is continuous at a parameter value if nearby parameter values produce target vectors arbitrarily close to the target at that value. Differentiability is a stronger local property: small input changes are represented, to first order, by a finite linear change in the output.

The distinction between the interior and the boundary is important. For example, Bernoulli response parameters lie in \([0,1]\): values strictly between 0 and 1 are interior, while 0 and 1 are boundary values. An estimator may attain a boundary value when all observed responses in an arm are identical. A formula derived only for \(0<\theta_k<1\) therefore does not automatically define a usable plug-in target for every possible estimate.

A common construction starts with nonnegative scores \(g_k(\theta)\) and normalizes them:

\rho_k(\theta)=\frac{g_k(\theta)}{S(\theta)},\qquad S(\theta)=\sum_{j=1}^K g_j(\theta).

This is a valid target wherever each score is defined, each score is nonnegative, and \(S(\theta)>0\). If all the scores vanish, normalization divides by zero and does not define a target. If some scores vanish but their sum is positive, the formula is still defined, but the corresponding target shares are zero. That is mathematically different from an undefined target.

What Smoothness Does and Does Not Guarantee

Suppose the scores \(g_k\) are continuous and \(S(\theta)>0\) on a region. Their normalized shares are continuous there: sums of continuous functions are continuous, and division by a continuous denominator that is nonzero preserves continuity. If the scores are differentiable and \(S\) remains nonzero, the shares are differentiable as well. For a component \(\theta_\ell\), the quotient rule gives

\frac{\partial \rho_k}{\partial \theta_\ell} = \frac{ \left(\frac{\partial g_k}{\partial \theta_\ell}\right)S - g_k\left(\frac{\partial S}{\partial \theta_\ell}\right) }{S^2}, \qquad \frac{\partial S}{\partial \theta_\ell} = \sum_{j=1}^K\frac{\partial g_j}{\partial \theta_\ell}.

This calculation also shows why a small normalizing sum is a warning sign: the denominator \(S^2\) can make derivatives large. Continuity alone does not guarantee a finite derivative at a boundary. Nor does smoothness on the interior guarantee a continuous extension to a boundary point that was excluded from the formula's domain.

In a DBCD, the target is not itself necessarily the probability of assigning the next patient to each arm. For example, with positive current allocation proportions \(\pi_k\), a common multi-arm rule uses weights proportional to \(\rho_k(\rho_k/\pi_k)^\gamma\), where \(\gamma\geq 0\). Thus a target map must first provide meaningful shares; the allocation rule then uses them to form assignment probabilities. If a target share is exactly zero, its weight in this expression is zero when the expression is defined. The target's boundary behavior therefore has direct implications for which arms the rule can assign.

Worked Example: A Linear Map That Extends to the Boundary

Suppose a two-arm target is defined from a Bernoulli response parameter \(\theta\in[0,1]\) by \(\rho_1(\theta)=\theta\) and \(\rho_2(\theta)=1-\theta\). This is a hypothetical target map; the parameter space includes both boundary values.

At the hypothetical estimate \(\hat{\theta}=0.30\), the plug-in shares are \(\rho_1(0.30)=0.30\) and \(\rho_2(0.30)=1-0.30=0.70\). They are nonnegative and sum to \(1.00\).

At the boundary estimate \(\hat{\theta}=0\), the shares are \((0,1)\); at \(\hat{\theta}=1\), they are \((1,0)\). The formula is defined at both endpoints and is continuous there. Its derivatives are \(d\rho_1/d\theta=1\) and \(d\rho_2/d\theta=-1\), including the corresponding one-sided behavior at the endpoints. Thus this map is well-defined and has finite, constant sensitivity throughout its allowed parameter range, although a boundary target assigns zero share to one arm.

Worked Example: A Continuous Target with an Unbounded Boundary Derivative

Consider a hypothetical two-arm target based on nonnegative variance parameters \(v_1,v_2\): \(\rho_1=\sqrt{v_1}/(\sqrt{v_1}+\sqrt{v_2})\) and \(\rho_2=1-\rho_1\). The formula is defined whenever \(v_1,v_2\geq0\) and at least one is positive.

For hypothetical estimates \(\hat v_1=0.04\) and \(\hat v_2=0.09\), the square roots are \(0.20\) and \(0.30\). Therefore \(\hat\rho_1=0.20/(0.20+0.30)=0.400\) and \(\hat\rho_2=0.30/0.50=0.600\). The shares are valid and sum to \(1.000\).

Fix \(v_2=c>0\). Writing \(x=v_1\), the first share is \(f(x)=\sqrt{x}/(\sqrt{x}+\sqrt{c})\). For \(x>0\), differentiation gives \[ f'(x)=\frac{\sqrt{c}}{2\sqrt{x}(\sqrt{x}+\sqrt{c})^2}. \] As \(x\) decreases to 0, \(f(x)\) tends to 0, so the target is continuous at \(x=0\). But \(f'(x)\) grows without bound because of the factor \(\sqrt{x}\) in the denominator. The boundary value is well-defined and continuous, yet the map is not differentiable there with a finite right derivative. Consequently, describing the map as simply “smooth” without specifying where would be misleading.

Worked Example: A Ratio That Has No Boundary Value at the Origin

Consider the hypothetical two-arm map \(\rho_1(\theta_1,\theta_2)=\theta_1/(\theta_1+\theta_2)\) and \(\rho_2=1-\rho_1\), for nonnegative parameters with \(\theta_1+\theta_2>0\). At estimates \((\hat{\theta}_1,\hat{\theta}_2)=(0.02,0.08)\), the target is \(\hat\rho_1=0.02/(0.02+0.08)=0.200\) and \(\hat\rho_2=0.800\).

The origin \((0,0)\) is excluded because the denominator is zero there. More importantly, no single continuous value can be assigned there to extend the map. Along the path \((\theta_1,\theta_2)=(t,t)\), for \(t>0\), the first share is \(t/(t+t)=0.500\). Along \((t,2t)\), it is \(t/(t+2t)=0.333\). Both paths approach \((0,0)\), but their target limits differ. The map is well-defined on its stated domain, but has no path-independent boundary limit at the origin.

Checking a Candidate Map

A reliable check begins with the estimator's possible values, not merely the values expected under a typical sample. For each candidate target, identify the parameter domain, substitute a generic estimate from that domain, and test the defining formulas at its edges. In particular, inspect every denominator, square root, logarithm, and other operation with a restricted domain. Then verify that the resulting shares are nonnegative and sum to one.

State the domain.
List the parameter values allowed by the model and the values the estimator can actually take, including endpoints.

Check definition and normalization.
Confirm that every score is defined and that the normalizing sum is positive at every value where the target is meant to be used.

Examine boundary limits.
For an excluded boundary point, approach it along different paths. A common limit permits a continuous extension; different limits do not.

Check regularity and allocation meaning.
When local stability matters, inspect derivatives or one-sided behavior, and note whether any valid target share is zero.

For comparison, a softmax map from finite real-valued scores \(u_k(\theta)\) is \(\rho_k=\exp(u_k)/\sum_j\exp(u_j)\). Its denominator is strictly positive, each share is positive, and if the scores are smooth on the domain, the shares are smooth there. It avoids zero shares at finite scores, unlike a normalized-score map whose component can vanish. But it is not a universal fix: the scores must still be defined, and if scores diverge at a boundary, their limiting shares need separate examination. It also encodes a particular way of translating scores into shares; it is not interchangeable with every scientifically chosen target.

Common Mistakes and Limitations

Assuming interior formulas automatically cover endpoints. A ratio can be valid whenever its denominator is positive yet fail at a boundary estimate. Check the estimator's actual range.
Treating a zero share as an undefined target. If the denominator remains positive and shares sum to one, a zero component is a valid boundary share. It may have important allocation consequences, but it is not the same as division by zero.
Calling continuity differentiability. The square-root example has a continuous boundary value and an unbounded derivative nearby. State which property has been checked.
Checking only one path to a boundary. The ratio example has different limits along different paths. A single convenient sequence cannot establish a continuous extension.

Smoothness is a mathematical property of a specified map on a specified domain; it does not establish that a target is optimal or appropriate for a clinical objective. Conversely, a map need not be differentiable everywhere to be usable. What matters for a given design is whether it is defined for the estimates that can arise and whether its boundary behavior is acceptable. If the map permits a zero target share, the allocation design must account for the resulting possibility of assigning no future patients to that arm. Boundary checks diagnose that issue; they do not by themselves prescribe a remedy.

Key takeaway: For a normalized target, verify that its scores are defined and nonnegative and that their sum stays positive. Then assess continuity and, if needed, differentiability at boundary values the estimator can attain. A zero share can be valid; an undefined or path-dependent target cannot be silently treated as valid.

Check Your Understanding

Answer each question, then compare with the brief solutions below.

For nonnegative scores \(g_1,g_2,g_3\), what condition is needed to normalize them into target shares?
A two-arm target is \(\rho_1=\theta/2\), \(\rho_2=1-\theta/2\), with \(\theta\in[0,1]\). Evaluate the shares at \(\theta=0\) and check whether they form a valid target.
For \(\rho_1=x/(x+y)\), where \(x,y\geq0\) and \(x+y>0\), compare the limits approaching the origin along \(y=x\) and \(y=3x\). Is there a continuous extension at the origin?
Why does continuity of a target map not guarantee a finite derivative at a boundary?

Solutions. (1) Each score must be defined and nonnegative, and \(g_1+g_2+g_3>0\). (2) At \(\theta=0\), the shares are \((0,1)\), which are nonnegative and sum to one, so they form a valid boundary target. (3) Along \(y=x\), the first share is \(1/2\); along \(y=3x\), it is \(1/4\). The limits differ, so no continuous extension exists at the origin. (4) Continuity requires the function values to converge to the boundary value, but a finite derivative additionally requires the change in output divided by the change in input to have a finite limiting linear rate. A function can converge while that rate becomes unbounded.