Uniform Asymptotics Over Trial Parameters

Why Uniformity Matters

An asymptotic approximation can be accurate at each fixed response parameter and still be poor for parameter values that change with the trial size. This distinction matters when a design or analysis is intended to work across a range of plausible treatment responses, rather than only at one unknown point. Uniform asymptotics make that ambition precise: they require the approximation error to vanish even at the most difficult parameter value in the specified set.

The focus here is a compact set of response parameters in an adaptive $K$-arm trial. We consider arm-specific sample means and a fixed contrast of those means. The central question is whether a normal approximation holds uniformly over the parameter set, and what conditions are needed to support that claim.

Definition: Let $\Theta$ be a parameter set and let $T_n(\theta)$ be a statistic under response parameter $\theta$. The sequence $T_n(\theta)$ converges to $t(\theta)$ uniformly in probability over $\Theta$ if, for every $\epsilon>0$, $$ \sup_{\theta\in\Theta} P_\theta\bigl(|T_n(\theta)-t(\theta)|>\epsilon\bigr)\longrightarrow 0. $$ For a scalar statistic with parameter-dependent normal approximation $G_\theta$, a uniform central limit approximation means $$ \sup_{\theta\in\Theta}\sup_{x\in\mathbb R} \left|P_\theta(T_n(\theta)\leq x)-G_\theta(x)\right|\longrightarrow 0. $$

The second definition compares each trial's distribution with the appropriate normal distribution at that trial's parameter value. It does not require the normal distributions to be identical across $\Theta$. The two suprema are important: approximation error must be small for every threshold $x$ and uniformly over all response parameters in the set.

Adaptive Trial Setup and Conditions

Suppose $n$ participants are assigned sequentially to $K$ arms. Write $A_i$ for participant $i$'s assigned arm and $Y_i$ for the response observed after assignment. The assignment probabilities may depend on earlier assignments and outcomes, but are determined before participant $i$'s response is observed. Let $N_k(n)=\sum_{i=1}^n I(A_i=k)$, and let $\bar Y_k$ be the mean of the observed responses in arm $k$, when $N_k(n)>0$. Under parameter $\theta$, arm $k$'s response has mean $\mu_k(\theta)$ and variance $\sigma_k^2(\theta)$.

A convenient set of sufficient conditions for a uniform normal approximation is as follows. These conditions deliberately rule out parameter values at which an arm becomes asymptotically unobserved.

Compact parameter set and regular response moments. The set $\Theta$ is compact. Responses have the stated arm-specific means and variances, and for some $\delta>0$, $\sup_{\theta\in\Theta} E_\theta[|Y_i-\mu_k(\theta)|^{2+\delta}\mid A_i=k]$ is finite for every arm. The variances are bounded away from zero and infinity uniformly over $\Theta$.
Stable, positive allocation limits. For continuous functions $\rho_k(\theta)$, the realized allocation fractions satisfy $N_k(n)/n\to\rho_k(\theta)$ uniformly in probability over $\Theta$. There is an $\eta>0$ such that $\rho_k(\theta)\geq\eta$ for every arm and every $\theta\in\Theta$.
Predictable allocation and a uniform central limit condition. Assignments use only information available before the current response. The centered arm response contributions form martingale differences, their conditional variance sums converge uniformly to the stated limits, and their conditional Lindeberg terms vanish uniformly. A uniform $2+\delta$ moment bound together with the allocation lower bound is a standard sufficient way to control the Lindeberg terms.

Here, the conditional variance sum is the accumulated variance of the centered response contributions, conditional on the information available before each response. For arm $k$, its leading term, divided by $n$, is $\sigma_k^2(\theta)$ times the average predictable probability of assigning arm $k$. The condition requires this average to approach $\rho_k(\theta)$ uniformly. Predictability prevents the current response from influencing its own assignment; it does not prohibit assignments from adapting to earlier responses.

Theorem: Under the conditions above, for a fixed vector of contrast coefficients $c=(c_1,\ldots,c_K)$, define $$ T_n(\theta)=\sqrt n\sum_{k=1}^K c_k\{\bar Y_k-\mu_k(\theta)\}, \qquad V(\theta)=\sum_{k=1}^K\frac{c_k^2\sigma_k^2(\theta)}{\rho_k(\theta)}. $$ If $V(\theta)$ is bounded away from zero on $\Theta$, then $$ \sup_{\theta\in\Theta}\sup_{x\in\mathbb R} \left| P_\theta\left(\frac{T_n(\theta)}{\sqrt{V(\theta)}}\leq x\right)-\Phi(x) \right|\longrightarrow 0, $$ where $\Phi$ is the standard normal cumulative distribution function.

Why the Variance Has This Form

The arm-specific centered response sum is $S_{k,n}=\sum_{i=1}^n I(A_i=k)\{Y_i-\mu_k(\theta)\}$. The sample-mean error can be written exactly as $\bar Y_k-\mu_k(\theta)=S_{k,n}/N_k(n)$. Thus $$ \sqrt n\{\bar Y_k-\mu_k(\theta)\} = \frac{n}{N_k(n)}\frac{S_{k,n}}{\sqrt n}. $$ The uniform allocation condition makes $n/N_k(n)$ close in probability to $1/\rho_k(\theta)$, uniformly over $\Theta$. The conditional variance of $S_{k,n}/\sqrt n$ approaches $\rho_k(\theta)\sigma_k^2(\theta)$. Consequently, the limiting variance for the arm's scaled mean is $\sigma_k^2(\theta)/\rho_k(\theta)$.

At a given participant, only one arm indicator can be nonzero. Therefore the conditional cross-products between centered response contributions for distinct arms are zero. The limiting covariance matrix of the vector of scaled arm means is diagonal, with entries $\sigma_k^2(\theta)/\rho_k(\theta)$. Taking the contrast with coefficients $c_k$ gives $V(\theta)$ in the theorem.

For justification of the uniform conclusion, apply the martingale central limit result under the stated uniform conditional-variance and Lindeberg conditions. Those conditions give a normal approximation for the vector of response sums, uniformly over $\theta$. The arm fractions can then replace their limits in the denominator: their uniform convergence in probability and the bound $\rho_k(\theta)\geq\eta$ keep the reciprocals stable. The resulting variance is $V(\theta)$; its positive lower bound permits standardization. The normal distribution has a continuous distribution function, so this uniform standardization yields the stated uniform cdf approximation.

Worked Example: A Uniform Approximation for Bernoulli Responses

Consider a hypothetical two-arm trial with binary responses. Let $\theta=(p_1,p_2)$ range over $\Theta=[0.2,0.8]^2$, where $p_k$ is the success probability in arm $k$. Suppose the design's limiting allocation fraction for arm 1 is $\rho_1(\theta)=0.5+(p_1-p_2)/4$, with $\rho_2(\theta)=1-\rho_1(\theta)$. Consider the contrast $\bar Y_1-\bar Y_2$, so $c_1=1$ and $c_2=-1$.

For every $\theta\in\Theta$, $-0.6\leq p_1-p_2\leq0.6$. Therefore $0.35\leq\rho_1(\theta)\leq0.65$, and the same bounds hold for $\rho_2(\theta)$. This verifies a uniform allocation lower bound with $\eta=0.35$. Bernoulli responses have variance $p_k(1-p_k)$, which lies between $0.16$ and $0.25$ on $[0.2,0.8]$. They also have uniformly bounded moments of every positive order. If the stated predictable-allocation and uniform conditional-limit conditions hold, the theorem applies.

At the hypothetical parameter value $p_1=0.4,\ p_2=0.6$, the limiting fractions are $\rho_1=0.5+(-0.2)/4=0.45$ and $\rho_2=0.55$. The contrast variance on the $\sqrt n$ scale is $$ V(\theta)=\frac{p_1(1-p_1)}{\rho_1}+ \frac{p_2(1-p_2)}{\rho_2} =\frac{0.4(0.6)}{0.45}+\frac{0.6(0.4)}{0.55} =0.5333+0.4364=0.9697. $$ For $n=1000$, the corresponding approximate standard error for the unscaled difference of means is $\sqrt{V(\theta)/n}=\sqrt{0.9697/1000}=0.03114$. This calculation is an asymptotic standard error at the specified parameter value; the theorem's uniform claim concerns approximation quality across the entire square $\Theta$, not just at this point.

Checking Uniformity in Practice

A pointwise argument fixes one $\theta$ and lets $n$ grow. A uniform argument must also control parameter sequences $\theta_n$ that vary with $n$. This is why checking only a few interior parameter values is insufficient. The most informative checks target the weakest uniform bounds: the smallest allocation fraction, the smallest response variance if standardization requires it, and the largest moment or conditional-variance errors.

For a design being assessed analytically or by simulation, useful diagnostics include:

Find a lower bound for every limiting allocation fraction across the full parameter set; a lower bound that deteriorates toward zero does not establish the theorem's condition.
Check whether response moment bounds hold uniformly, rather than separately at each parameter value with constants that may grow without limit.
Evaluate conditional variance and tail contributions near parameter-set boundaries, where a pointwise approximation may be least reliable.
In simulation, compare normal-approximation errors across a grid that includes boundary and near-boundary values. A grid is a diagnostic, not a proof of a supremum bound.

Uniform asymptotic results are not finite-sample guarantees. Even with all assumptions satisfied, an approximation can be poor for a particular modest $n$. Conversely, an apparently successful simulation at selected parameter values cannot establish uniform convergence over a continuum.

Worked Example: A Parameter Boundary That Breaks the Conditions

Suppose a two-arm design has a parameter $\theta\in[0,1]$ and limiting fraction $\rho_1(\theta)=\theta$, with $\rho_2(\theta)=1-\theta$. At every fixed $\theta>0$, arm 1 has a positive limiting fraction. But over the full compact set $[0,1]$, $\inf_{\theta\in[0,1]}\rho_1(\theta)=0$. The uniform allocation lower bound therefore fails.

The problem is not merely formal. At $\theta=1/n$, the target arm-1 count is $n\rho_1(\theta)=1$, rather than a count that grows proportionally to $n$. Thus one cannot use the theorem's stable denominator approximation uniformly over the stated parameter set. Restricting the parameter set to $[\epsilon,1]$ for a fixed $\epsilon>0$ would restore a lower bound for arm 1, though all the other assumptions would still need checking. This example shows why a result proved separately for every fixed positive $\theta$ does not automatically extend to the boundary.

Pointwise Results, Uniform Results, and Limitations

Approach	What it establishes	Main limitation
Pointwise asymptotics	Approximation at each fixed response parameter	Error may not be controlled when parameters vary with $n$
Uniform asymptotics on compact $\Theta$	Approximation error vanishes uniformly over $\Theta$	Requires common bounds, especially positive allocation fractions and uniform tail control
Finite-sample simulation or bounds	Performance at specified sample sizes and parameter values, or a finite-sample guarantee if a valid bound is proved	A finite grid does not by itself establish a uniform asymptotic result

The compactness of $\Theta$ is useful because continuous parameter-dependent quantities can attain their extrema on it. Compactness alone, however, does not guarantee positive allocations or bounded response moments; those are separate assumptions. Nor does a uniform normal approximation automatically justify every downstream procedure. For example, a studentized statistic or a data-dependent contrast requires conditions ensuring its estimated scaling is itself stable uniformly.

Common mistake: “The central limit theorem holds for every fixed $\theta$, so it must hold uniformly over all $\theta\in\Theta$.” This reasoning confuses separate pointwise limits with a limit of the worst-case error. The correction is to establish bounds that hold simultaneously over $\Theta$, such as a common moment bound, a uniform Lindeberg condition, and an allocation fraction bounded below by the same positive constant for every parameter value.

Key takeaway: Uniform asymptotic normality is a guarantee across a parameter set, not a stronger-sounding synonym for pointwise normality. Its central protections are uniform control of response tails, stable allocation limits, and allocation fractions bounded away from zero.

Check Your Understanding

Answer each question, then compare with the solutions below.

In the definition of a uniform central limit approximation, what do the two suprema range over?
For the contrast $\bar Y_1-\bar Y_2$, why does arm $k$'s contribution to the limiting variance contain $1/\rho_k(\theta)$?
In the Bernoulli example, verify the allocation fraction at $p_1=0.8,\ p_2=0.2$, and state whether the uniform lower bound is satisfied.
Why does a pointwise central limit theorem not alone rule out poor approximation at a sequence of parameter values $\theta_n$?
What assumption fails for arm 1 in the boundary example with $\rho_1(\theta)=\theta$, $\theta\in[0,1]$?

Solutions. (1) One supremum ranges over $\theta\in\Theta$, and the other ranges over thresholds $x\in\mathbb R$; the largest cdf error over both must vanish. (2) The centered arm-$k$ sum has variance of order $n\rho_k\sigma_k^2$, while its sample mean divides by $N_k$, approximately $n\rho_k$. On the $\sqrt n$ scale this gives variance $\sigma_k^2/\rho_k$. (3) $\rho_1=0.5+(0.8-0.2)/4=0.65$, and $\rho_2=0.35$; both are at least $0.35$, so the bound holds. (4) Pointwise convergence controls each fixed parameter separately, not the supremum of the error over parameters that may change with $n$. (5) There is no positive lower bound for the arm-1 allocation fraction on the full parameter set: its infimum is zero at $\theta=0$.

Uniform Asymptotics Over Trial Parameters

What You'll Learn