Extracted main text — title through conclusion, appendix excluded. This is what our citation measures are computed over, published so the extraction can be checked by eye.
90,391 characters · 31 sections · 34 citation commands
Continuous time asymptotic representations for adaptive experiments
Adaptive experiments are experiments where data is collected and analyzed continuously, allowing for adjustments or decisions to be made on an ongoing basis. Originating from early work by wald1947sequential, arrow1949bayes, and others, adaptive experimentation has evolved to encompass a wide range of strategies, including bandit experiments, A/B testing, costly sampling, and best-arm identification. These strategies are now widely used across various fields. For instance, tech companies frequently employ bandit algorithms and A/B testing for applications such as web interface optimization, dynamic pricing, and targeted ad placement (see, bouneffouf2019survey for a survey of applications). In clinical trials, multi-stage or group sequential designs (wassmer2016group) have become standard, allowing early termination of experiments when strong evidence of positive or negative effects emerges. Economics has also seen a growing adoption of adaptive experimentation. For example, kasy2019adaptive and finan2022reinforcing develop new sequential experimentation strategies for use in development contexts. More recently, chapman2024dynamically introduced an optimal dynamic strategy for eliciting economic preferences.
Despite these growing number of applications, determining optimal decision rules in adaptive experiments remains challenging due to the interactive nature of the data generating process. For instance, there is currently no well-established notion of optimal point estimation following an adaptive experiment, nor of inference procedures that remain valid at any point during the experiment. In this article, we address these challenges by deriving a continuous-time asymptotic representation for adaptive experiments. Specifically, we model the sequential data collection process using an empirical allocation process, which, at any given time $t$ specifies the fraction of time allocated to observing outcomes from a particular treatment. We then show that this empirical allocation process weakly converges to an allocation process in a limiting experiment, where signals consist of multiple Gaussian processes with unknown drifts, each corresponding to a different treatment arm. Then, by characterizing optimal decisions in the limit experiment, we can construct asymptotically optimal decision rules in the original experiment.
The limit experiment greatly simplifies the characterization of optimal decisions due to two key properties. First, its state space is considerably smaller than that of the original experiment. In the limit experiment, the sufficient statistics are the past sample paths of the signal processes. However, in many applications, we show that these statistics can be further reduced to just the current values of the signal and allocation processes, resulting in a state-space dimension of $2K$, where $K$ is the number of treatments. In contrast, the state space of the actual experiment encompasses all collected observations, making it substantially larger and more complex. Second, the limit experiment is far more tractable to analyze, as the properties of Gaussian diffusions in continuous time are well understood, allowing us to easily characterize optimal decisions in that setting.
To illustrate the broad applicability of our framework, we use our representation theorem to derive optimal decision rules for several fundamental aspects of adaptive experimentation. In particular, we address: (1) the construction of optimal estimators following adaptive experiments, (2) the analysis of in-sample regret, and (3) the development of e-processes for anytime-valid inference. An e-process is a nonnegative supermartingale (under the null hypothesis) that continuously tracks statistical evidence against the null over the course of an experiment. We introduce the first definition of an e-process for multi-treatment adaptive experiments. This enables the design of algorithm-free anytime-valid tests, i.e., tests that maintain correct size even when the sampling strategy used in the adaptive experiment is unknown.
In each of the above applications, our asymptotic framework significantly simplifies the decision problem, making it much more tractable. For example, in the estimation problem, we find that all optimal Bayes estimators share a common form that is independent of how the experiment was conducted. Furthermore, we derive explicit expressions for these estimators under Gaussian priors.
In an important prior work, hirano2023asymptotic develop an asymptotic representation theorem for batched adaptive experiments, where sampling strategies are updated only a finite number of times over the course of the experiment. In this setting, they show that the policy rule and any statistic from the finite-sample experiment can be matched to a corresponding rule and statistic in a limit experiment involving Gaussian signals from each batch.
Our asymptotic representation theorem is very different in terms of both the scope and the formulation. It applies to fully adaptive experiments and is expressed in terms of allocation processes rather than sequences of policy rules. This shift is essential, for in fully adaptive experiments, sequences of policy rules generally fail to admit weak limits. At the same time, our theory is more specialized in certain respects. It does not directly provide asymptotic representations for arbitrary statistics; instead, it characterizes the joint evolution of the score and the empirical allocation processes, which together fully determine the limit experiment. Obtaining representations for arbitrary statistics is more difficult in our setting because there is no straightforward coupling method for continuous-time processes that respects the required informational constraints (for example, ensuring that an anytime-valid test depends only on observed data).
However, for many applications such comprehensive representation theorems are not always necessary. In practice, lower bounds on losses or risk can be established using our representation theorem and standard change-of-measure arguments. For instance, in the case of point estimation, we show that the frequentist risk of any sequence of estimators is asymptotically bounded below by the risk of an estimator in the limit experiment that depends only on the terminal values of the signal and allocation processes. While this result does not establish a one-to-one mapping from finite-sample estimators to their counterparts in the limit experiment, such stronger representation is not essential for deriving lower bounds or constructing asymptotically optimal estimators.
Beyond these technical contributions, the broader value of our approach is conceptual. This article provides the first general definition of adaptive experiments in continuous time through the lens of allocation processes. This formulation not only simplifies the analysis of adaptive experiments---continuous time being more tractable than discrete time---but also yields a sharp characterization of sufficient statistics. As discussed earlier, in many applications, including point estimation and anytime-valid inference, the sufficient statistics reduce to the current values of the signal and allocation processes, resulting in a finite-dimensional state-space. More generally, allocation processes offer a natural and flexible framework for representing adaptive experiments, and our formulation of e-processes is indeed most naturally articulated in terms of these processes.
In terms of the style of asymptotic approximations, this article is most closely related to Le Cam's LeCam1979 work on stopping time representations. We extend his framework and build on his proof techniques to handle the additional complexities introduced by adaptive sampling.
The systematic study of Gaussian process approximations for adaptive experiments was initiated by fan2021diffusion and wager2021diffusion, who introduced diffusion asymptotics to analyze the behavior of adaptive algorithms such as Thompson Sampling. kalvit2021closer extended this approach to cover UCB algorithms, and adusumilli2025optimal further generalized the framework to characterize optimal bandit algorithms under both Bayesian and minimax regret criteria. adusumilli2025optimal also established that likelihood ratio processes from finite-sample adaptive experiments converge uniformly to their diffusion counterparts. While this convergence is algorithm-agnostic and well-suited for analyzing Bayesian decision criteria, it is insufficient for tasks such as anytime-valid inference, as it lacks a representation explicitly connecting the finite-sample algorithms to a suitable counterpart in the diffusion limit. This article fills that gap by providing precisely such a representation.
Finally, this article also builds on earlier work by this author adusumilli2023optimal on optimal testing following adaptive experiments. The analysis of applications such as point estimation and anytime-valid inference draws on the strategies and proof techniques developed in that work.
We begin with a simple illustration involving two-armed bandits to motivate our theoretical analysis.
Consider a scenario in which the goal is to identify the better-performing version of a website, denoted by variants $a=0,1$, each with an unknown average click-through rate $\theta^{(a)}$. The observed outcome from each variant is a binary draw from $\text{Bernoulli}(\theta^{(a)})$. To determine the optimal variant, we run a bandit algorithm for $n$ rounds, sequentially assigning users to one of the two alternatives.
Two of the most widely used bandit algorithms in practice are Thompson Sampling (TS) and the Upper Confidence Bound (UCB) algorithm. Both use accumulated data to guide user allocation but differ in their approach.
The TS algorithm begins with a prior distribution over each $\theta^{(a)}$. Given the Bernoulli outcome model, a Beta prior is standard; for this illustration, we will take it to be $\textrm{Beta}(1,1)$. At each round, the algorithm samples a draw of $\{\theta^{(a)}\}_{a}$ from its posterior distribution given the past observations, and then allocates the user to the variant with the highest draw.
In contrast, the UCB algorithm computes an index \[ \widehat{\text{UCB}}_{j}^{(a)}=\hat{\theta}_{j}^{(a)}+\sqrt{\frac{2\ln(j/n)}{N_{j}^{(a)}}}, \] where $\hat{\theta}_{j}^{(a)}$ is the sample mean of outcomes for variant $a$, and $N_{j}^{(a)}$ is the number of users allocated to variant $a$ prior to round $j$. The user is assigned to the variant with the higher index.
Define time $t:=j/n\in[0,1]$ as the fraction of the experiment completed. For either algorithm, let \[ q_{n,1}(t):=\frac{1}{n}\sum_{j=1}^{\lfloor nt\rfloor}\mathbb{I}\{A_{j}=1\} \] denote the number of assignments to variant 1 up to time $t$, normalized by $n$.
Suppose the two variants are identical, i.e., $\theta^{(0)}=\theta^{(1)}$. Figure (ref) plots the sampling distribution of $q_{n,1}(t)$ at three time points---$t=0.25,0.5,0.75$---under TS, for various values of $n$. As $n$ increases, the distributions converge, illustrating a form of asymptotic stability. The same convergence also occurs under UCB, as shown in Figure (ref), although the sampling distributions differ quite substantially. The figures are plotted for $\theta^{(0)}=\theta^{(1)}=0.1$, but changing these values would not make much of a difference to the plots (as long as $\theta^{(0)}=\theta^{(1)}$).
This convergence behavior of the empirical allocation process $q_{n,1}(\cdot)$ is not specific to these algorithms. A central result of this article is that under mild conditions, such asymptotic convergence holds for almost every adaptive allocation rule. To this end, we start by describing a general setup for adaptive experiments.
Adaptive experiments involve multiple treatments, where the policy rule---i.e., the probability of allocation to each treatment---can continuously adjust over the course of the experiment.
Let $K$ denote the number of treatments under consideration. To simplify notation and proofs, we focus on the case $K=2$, though our results extend to any fixed $K$. The outcome $Y^{(a)}$ under treatment $a\in\{0,1\}$ follows a parametric model $\{P_{\theta^{(a)}}^{(a)}\}$, where $\theta^{(a)}\in\mathbb{R}^{d}$ is an unknown parameter. For simplicity, we assume $\theta^{(0)}$ and $\theta^{(1)}$ have the same dimension, though this is not required for our results.
Since only ever observe one potential outcome $Y^{(a)}$ per unit, we can assume $\{P_{\theta^{(a)}}^{(a)}\}_{a}$ are independent across $a$, conditional on $(\theta^{(1)},\theta^{(0)})$. Let $j=1,\dots,n$ index the experimental periods. We define time $t$ as the scaled number of periods, $t=j/n$, representing the fraction of the experiment completed.
The decision-maker (DM) employs a policy rule $\{\pi_{n,j}\}_{j}\equiv\{\pi_{n,\lfloor nt\rfloor}\}_{t}$, which prescribes the probability of assigning observation $j$ to treatment 1 based on past information. The treatment assignment follows $A_{j}\sim\text{Bernoulli}(\pi_{n,j})$. For the outcomes, it is useful to conceptualize a stack of observations $\ensuremath{\bm{y}_{a}:=\{Y_{i}^{(a)}\}_{i=1}^{n}}$ for each treatment, generated at the outset as i.i.d draws from $\{P_{\theta^{(a)}}^{(a)}\}$, but unobserved initially by the DM. This is the so-called stack-of-rewards model (lattimore2020bandit). Each time a treatment is sampled, it can be imagined that the DM observes the top element of the corresponding treatment stack; this element is then taken out of consideration.
As in the illustrative example, let \[ q_{n,a}(t):=\frac{1}{n}\sum_{j=1}^{\lfloor nt\rfloor}\mathbb{I}\{A_{j}=a\} \] denote the fraction of (total) observations assigned to treatment $a$ up to time $t$. We term $\{q_{n,a}(\cdot)\}_{a}$ the empirical allocation process. One can interpret the policy rule $\{\pi_{n,j}\}_{j}$ as a function mapping the stacks $\ensuremath{(\bm{y}_{1},\bm{y}_{0})}$ and an exogenous random variable $U\sim\text{Uniform}[0,1]$ to the observed trajectory of $\{q_{n,a}(\cdot)\}_{a}$. The exogenous randomness accounts for probabilistic policy rules; specifically, we subsume the policy randomizations at all $n$ stages of the adaptive experiment into a single $U$.\footnote{Indeed, a single uniform random variable can be mapped into countably infinite independent uniform random variables.} The key informational constraint is that the event $\{q_{n,1}(t)\leq\gamma_{1},q_{n,0}(t)\leq\gamma_{0}\}$ depends only on the first $\ensuremath{\lfloor n\gamma_{1}\rfloor},\ensuremath{\lfloor n\gamma_{0}\rfloor}$ observations from $(\bm{y}_{1},\bm{y}_{0})$ and the exogenous randomization $U$. Formally, \[ \{q_{n,1}(t)\leq\gamma_{1},q_{n,0}(t)\leq\gamma_{0}\}\text{ is }\mathcal{G}_{n,\gamma_{0},\gamma_{1}}:=\mathcal{F}_{n,\gamma_{1}}^{(1)}\lor\mathcal{F}_{n,\gamma_{0}}^{(0)}\lor\sigma(U)\text{-measurable} \] for each $t,\gamma_{1},\gamma_{0}\in[0,1]$, where \[ \mathcal{F}_{n,\gamma}^{(a)}:=\sigma\left(Y_{1}^{(a)},\dots,Y_{\lfloor n\gamma\rfloor}^{(a)}\right) \] is the filtration (i.e., information set) generated by the first $\lfloor n\gamma\rfloor$ observations from stack $\bm{y}_{a}$. Thus, each policy rule $\{\pi_{n,j}\}_{j}$ can be associated with monotonic empirical allocation processes $\{q_{n,a}(\cdot)\}_{a}$ satisfying the above informational constraint.
We are interested in the behavior of various statistics under local perturbations of the form ${\theta_{0}^{(a)}+h^{(a)}/\sqrt{n}:h^{(a)}\in\mathbb{R}^{d}}$, where $\theta_{0}^{(a)}$ denotes a reference parameter. This focus is motivated by the fact that many decisions in adaptive experiments hinge on distinguishing between parameter values that are close to one another. For example, in the website optimization setting discussed earlier, bandit algorithms are employed precisely because the differences between website variants tend to be small. deng2013improving survey industry practices and report that typical differences in click-through rates are often around 1% or less. In such cases, bandit algorithms are employed for detecting subtle effects using as little data as possible. This naturally places us in the domain of local asymptotics.
In this article, we analyze the behavior of adaptive algorithms under local asymptotics for parametric classes of families $\{P_{\theta}^{(a)}\}_{\theta}$. Let $\nu$ denote a dominating measure for $\{P_{\theta}^{(a)}:\theta\in\mathbb{R}^{d},a\in\{0,1\}\}$, and set $p_{\theta}^{(a)}:=dP_{\theta}^{(a)}/d\nu$. We require $\{P_{\theta}^{(a)}\}_{\theta}$ to be quadratic mean differentiable (qmd):
\begin{asm1} The class $\{P_{\theta}^{(a)}:\theta\in\mathbb{R}^{d}\}$ is qmd around $\theta_{0}^{(a)}$ for each $a\in\{0,1\}$, i.e., there exists a score function $\psi_{a}(\cdot)$ such that for each $h^{(a)}\in\mathbb{R}^{d},$ \[ \int\left[\sqrt{p_{\theta_{0}^{(a)}+h^{(a)}}^{(a)}}-\sqrt{p_{\theta_{0}^{(a)}}^{(a)}}-\frac{1}{2}h^{(a)\intercal}\psi_{a}\sqrt{p_{\theta_{0}^{(a)}}}\right]^{2}d\nu=o(\vert h^{(a)}\vert^{2}). \] Furthermore, the information matrix $I_{a}:=\mathbb{E}_{0}[\psi_{a}\psi_{a}^{\intercal}]$ is invertible for $a\in\{0,1\}$. \end{asm1}
In the illustrative example, the outcomes are Bernoulli, so Assumption 1 holds with $\psi_{a}(y)=\left(\theta_{0}^{(a)}(1-\theta_{0}^{(a)})\right)^{-1}(y-\theta_{0}^{(a)})$. More broadly, this assumption is rather mild and satisfied for almost all commonly used distributions, including the Normal, Cauchy, Exponential, and Poisson distributions.
For each $q\in[0,1]$, define $z_{n,a}(q)$ as the partial sum process \[ z_{n,a}(q):=\frac{I_{a}^{-1/2}}{\sqrt{n}}\sum_{i=1}^{\left\lfloor nq\right\rfloor }\psi_{a}(Y_{i,j}^{(a)}). \] Knowledge of the process, $z_{n,a}(\cdot)$, on the domain $[0,q]$ is equivalent to knowledge of the scores from the first $\left\lfloor nq\right\rfloor $ observations of the stack ${\bf y}_{a}$. We then define the score process for treatment $a$ as \[ x_{n,a}(t):=z_{n,a}(q_{n,a}(t));\ a\in\{0,1\}. \] As we will show, the sample paths of this process serve as an asymptotically sufficient statistic for the adaptive experiment.
The primary result of this article establishes that any adaptive experiment is asymptotically Blackwell equivalent to a limit experiment characterized by Gaussian diffusions.
In this limit experiment, the decision maker observes a Gaussian process signal $z_{a}(\cdot)$ associated with each treatment $a$, given by
where $\{\ensuremath{W_{a}(\cdot)\}_{a}}$ are independent $d$-dimensional Wiener processes, and the drifts $\{h^{(a)}\}_{a}$ are unknown. Intuitively, $z_{a}(\cdot)$ serves as the limiting counterpart of $z_{n,a}(\cdot)$ in the original experiment, while the index $q$ represents the amount of `attention' devoted to that particular treatment.
Define the natural filtration generated by the Gaussian process $z_{a}(\cdot)$ up to a given attention-value $\gamma$ as $\mathcal{F}_{\gamma}^{(a)}:=\sigma\{z_{a}(s):s\le\gamma\}$. We can and do take $\mathcal{F}_{\gamma}^{(a)}$ to be right-continuous, i.e., $\mathcal{F}_{\gamma}^{(a)}\equiv\bigcap_{\epsilon\downarrow0}\mathcal{F}_{\gamma+\epsilon}^{(a)}.$ Similar to the stack of rewards in the original experiment, the entire process $z_{a}(\cdot)$ is not immediately observed. Instead, at time $t$, the DM observes the sample paths of $z_{1}(\cdot)$ and $z_{0}(\cdot)$ over the intervals $[0,q_{1}(t)]$ and $[0,q_{0}(t)]$, respectively, where $q_{a}(t)$ represents the amount of attention devoted to treatment $a$ up to time $t$. The quantities $\{q_{a}(t)\}_{a}$, termed allocation processes, are continuous-time analogues of $\{q_{n,a}(t)\}_{a}$, and formally defined as follows:
Definition (ref) addresses the technical challenge of defining adaptive experiments in continuous time without recourse to policy rules, which are typically non-measurable (see Section (ref) below). Instead, it defines allocation processes abstractly through their functional properties. The first two conditions on $q_{a}(\cdot)$ are straightforward. The third condition ensures that, aside from the exogenous randomization $U$, whether or not $q_{1}(t)\le\gamma_{1}$ and $q_{0}(t)\le\gamma_{0}$ hold can be determined based only on the sample paths, $\{z_{a}(s):s\le\gamma_{a}\}$, of the signal processes.\footnote{Although the informational constraint is only explicitly required for $\gamma_{1}+\gamma_{0}\ge t$, the requirement $q_{1}(t)+q_{0}(t)=t$ almost surely implies that the event $\{q_{1}(t)\le\gamma_{1},q_{0}(t)\le\gamma_{0}\}$ is a measure-zero set when $\gamma_{1}+\gamma_{0}<t$. Consequently, because $\mathcal{G}_{\gamma_{1},\gamma_{0}}$ is an augmented filtration, this event is measurable with respect to $\mathcal{G}_{\gamma_{1},\gamma_{0}}$ for any $\gamma_{1}+\gamma_{0}<t$ as well.} It is the continuous time counterpart of the information constraint on empirical allocation processes, described in Section (ref). The condition is analogous to the usual definition of a stopping time, but extended to a multi-dimensional setting.
The quantities $\{z_{a}(\cdot),q_{a}(\cdot),U\}_{a}$ characterize the limit adaptive experiment. Importantly, the inputs to the processes $z_{a}(\cdot),q_{a}(\cdot)$ are different: $\{q_{a}(\cdot)\}_{a}$ are indexed by time, while $\{z_{a}(\cdot)\}_{a}$ are indexed by the attention devoted to each treatment.
For each $a\in\{0,1\}$, define $x_{a}(t):=z_{a}(q_{a}(t))$ as the limit counterpart of $x_{n,a}(t)$. It is straightforward to verify that the sample paths of $\{x_{a}(\cdot)\}_{a}$ constitute sufficient statistics for the limit experiment up to time $t$.
Let $\mathcal{I}_{t}:=\mathcal{G}_{q_{1}(t),q_{0}(t)}$ denote the information accrued from the experiment until time $t$.\footnote{Formally, $\mathcal{G}_{q_{1}(t),q_{0}(t)}\equiv\left[A\in\mathcal{F}:A\cap\left\{ q_{1}(t)\le\gamma_{1},q_{0}(t)\le\gamma_{0}\right\} \in\mathcal{G}_{\gamma_{1},\gamma_{0}}\ \forall\ t,\gamma_{1},\gamma_{0}\in[0,1]\right]$. } An important property of $x_{1}(\cdot),x_{0}(\cdot)$ is that they are $\mathcal{I}_{t}$-martingales when $\bm{h}:=(h^{(1)},h^{(0)})=(0,0)$. Furthermore, $q_{a}(t)$ is the quadratic variation of $x_{a}(t)$, in that it captures the accumulated variability of $x_{a}(t)$.
Theorem (ref) in Section (ref) states that we can take $q_{a}(\cdot)$ to be almost surely Lipschitz continuous with a Lipschitz constant of $1$. By the fundamental theorem of Lebesgue integral calculus, almost every sample path of $q_{a}(\cdot)$ is then differentiable almost everywhere, with a Lebesgue integrable derivative $\pi_{a}(t):=\frac{dq_{a}(t)}{dt}.$ Consequently, $q_{a}(t)=\int_{0}^{t}\pi_{a}(s)ds$. Moreover, Definition (ref) and the right continuity of $\mathcal{G}_{q_{1},q_{0}}$ implies that $\pi_{a}(t)$ is $\mathcal{I}_{t}$-measurable coordinate-wise, for each $t$.
However, $\pi(\cdot)\equiv\pi_{1}(\cdot)$ cannot, in general, be interpreted as a valid policy rule, since it need not be measurable as a random process indexed by $t$. Indeed, in many continuous-time optimal control problems, the optimal policy does not belong to a separable space, and is therefore not measurable.\footnote{The measurability of sample paths---often referred to as strong or Bochner measurability---differs from the weaker notion of coordinate-wise measurability. The Pettis measurability theorem states that a stochastic process is Bochner measurable if and only if its sample paths lie in a separable subspace with probability one. See adusumilli2023optimal for further discussion.} For this reason, it is also generally impossible to define weak convergence for a sequence of policies $\pi_{n,\lfloor nt\rfloor}$, as such sequences are typically not asymptotically equicontinuous.
These observations indicate that the policy rule is not a particularly suitable object for characterizing sequential strategies in the limit experiment. In contrast, as will be established in Theorem (ref) below, any sequence of empirical allocation processes converge weakly to an allocation process in the limit experiment. In this sense, the allocation process serves as the more fundamental representation of sequential decision-making.
We now establish the asymptotic equivalence between the original sequence of adaptive experiments and the limit experiment. This equivalence follows from two key results. The first, previously demonstrated in adusumilli2025optimal, states that the likelihood ratio processes in the original experiment converge uniformly, at all possible `attention' values, to their counterparts in the limit experiment. We restate this result here for completeness. The second result, which is novel to this article, asserts that any sequence of score and allocation processes, $\{x_{n,a}(\cdot),q_{n,a}(\cdot)\}_{a}$, admits a corresponding representation in the limit experiment.
Let $\mathbb{P}_{n,\bm{h}}$ denote the induced probability over the stacked rewards and the exogenous randomization, i.e., over $({\bf y}^{(1)},{\bf y}^{(0)},U)$ when $Y^{(a)}\sim P_{\theta_{0}+h^{(a)}/\sqrt{n}}^{(a)}$. For each $a$, denote ${\bf y}_{\left\lfloor nq\right\rfloor }^{(a)}:=\{Y_{i}^{(a)}\}_{i=1}^{\left\lfloor nq\right\rfloor }$. Suppose that we observe $U$ and $\left\lfloor n\gamma_{1}\right\rfloor ,\left\lfloor n\gamma_{0}\right\rfloor $ units from each treatment, i.e., we observe $U$ and ${\bf y}_{\gamma_{1},\gamma_{0}}:=\left({\bf y}_{\left\lfloor n\gamma_{1}\right\rfloor }^{(1)},{\bf y}_{\left\lfloor n\gamma_{0}\right\rfloor }^{(0)}\right)$. Given this set of observations, the log-likelihood ratio process under the local alternative $\bm{h}:=(h^{(1)},h^{(0)})$ is:
where, for any $a\in\{0,1\}$ and $\gamma\in[0,1]$, \[ \ln\frac{dP_{\theta_{0}^{(a)}+h^{(a)}/\sqrt{n}}^{(a)}}{dP_{\theta_{0}^{(a)}}^{(a)}}\left({\bf y}_{\left\lfloor n\gamma\right\rfloor }^{(a)}\right):=\sum_{i=1}^{\left\lfloor n\gamma\right\rfloor }\ln\frac{dP_{\theta_{0}^{(a)}+h^{(a)}/\sqrt{n}}^{(a)}}{dP_{\theta_{0}^{(a)}}^{(a)}}\left(Y_{i}^{(a)}\right). \] In adusumilli2025optimal, this author showed that under Assumption 1,
Analogously, in the limit experiment, the relevant probability measure is $\mathbb{P}_{\bm{h}}:=\mathbb{P}_{h^{(1)}}^{(1)}\otimes\mathbb{P}_{h^{(0)}}^{(0)}\otimes P_{U}$, where $\mathbb{P}_{h^{(a)}}^{(a)}$ is the induced probability over the sample paths of $\{z_{a}(s);0\le s\le1\}$ when the local parameter is $h^{(a)}$, and $P_{U}$ is the probability measure induced by $U\sim\textrm{Uniform}[0,1]$. Also, for some fixed $\{\gamma_{a}\}_{a}$, let \[ \varphi^{(a)}(h^{(a)};\gamma_{a})=\mathbb{E}_{\mathbb{P}_{0}^{(a)}}\left[\left.\ln\frac{d\mathbb{P}_{h^{(a)}}^{(a)}}{d\mathbb{P}_{0}^{(a)}}\right|\bar{\mathcal{F}}_{\gamma_{a}}^{(a)}\lor\sigma(U)\right] \] denote the log-likelihood ratio under the local alternative $h^{(a)}$ given the sample path $\{z_{a}(s);s\le\gamma_{a}\}$ and $U$. Similarly, $\varphi(\bm{h};\gamma_{1},\gamma_{0})$ denotes the likelihood ratio under $\bm{h}=(h^{(1)},h^{(0)})$ given the sample paths $\{z_{a}(s);s\le\gamma_{a}\}_{a}$ and $U$. Since $z_{1}(\cdot),z_{0}(\cdot)$ are Wiener processes under $\mathbb{P}_{\bm{0}}$, the Girsanov theorem implies
Furthermore, as the Wiener processes are independent, $\varphi(\bm{h};\gamma_{1},\gamma_{0})=\varphi^{(1)}(h^{(1)};\gamma_{1})+\varphi^{(0)}(h^{(0)};\gamma_{0}).$
Equations ((ref)), ((ref)) imply $\hat{\varphi}^{(a)}(h^{(a)};\cdot)\xrightarrow[\mathbb{P}_{n,0}]{d}\varphi^{(a)}(h^{(a)};\cdot)$ for each $a,h^{(a)}$ and therefore,
By itself, ((ref)) does not reference a specific allocation process. While it describes convergence of likelihood ratio processes across the space of all possible `attention' values $(\gamma_{1},\gamma_{0})$, it does not immediately guarantee convergence at all possible time points $t$. Indeed, the relationship between the observed attention values $(\gamma_{1},\gamma_{0})$ and time $t$ is determined by the specific allocation process being employed.
To demonstrate that a risk function in the finite-sample experiment admits a corresponding representation or lower bound in the limit experiment, we need to go further and match the joint distribution of the asymptotically sufficient statistics $\{x_{n,a}(\cdot),q_{n,a}(\cdot)\}_{a}$ as well. Combining this joint convergence with ((ref)) would yield uniform convergence of likelihood ratios across all time points, and thereby imply asymptotic equivalence between the actual and limit experiments in the sense of le1986asymptotic. The following key result establishes this joint convergence.
Theorem (ref) establishes that the distribution of $\left\{ x_{n_{k},a}(\cdot),q_{n_{k},a}(\cdot)\right\} _{a}$ in the original experiment can be matched with that of $\left\{ x_{a}(\cdot),q_{a}(\cdot)\right\} _{a}$ in the limit experiment, where $\{q_{a}(\cdot)\}_{a}$ is a suitably defined allocation process. Although these statistics are path-valued processes, the convergence is uniform over time. In concert with ((ref)), Theorem (ref) enables us to derive lower bounds on losses or statistical risk in various applications by employing change of measure arguments.
While the proof of Theorem (ref) is somewhat involved, the underlying intuition is relatively straightforward. The processes $\{z_{n,a}(\cdot)\}_{a}$ are asymptotically tight (being standard partial sum processes), and likewise, the processes $\{q_{n,a}(\cdot)\}_{a}$ are also tight since, by definition, \[ \sup_{t}\vert q_{n,a}(t+\delta)-q_{n,a}(t)\vert\le\delta+n^{-1},\quad\forall\ \delta>0. \] This ensures that $\{z_{n,a}(\cdot),q_{n,a}(\cdot)\}_{a}$ converges to some weak limit $\{z_{a}(\cdot),q_{a}(\cdot)\}_{a}$. Moreover, the measurability of the events $\{q_{n,1}(t)\le\gamma_{1},q_{n,0}(t)\le\gamma_{0}\}$ with respect to $\mathcal{G}_{n,\gamma_{0},\gamma_{1}}:=\mathcal{F}_{n,\gamma_{1}}^{(1)}\lor\mathcal{F}_{n,\gamma_{0}}^{(0)}\lor\sigma(U)$ suggests that $q_{a}(\cdot)$ can be constructed to inherit the appropriate adaptedness properties required by Definition (ref). The construction requires some care and utilizes some results from the theory of stable convergence hausler2015stable; it is perhaps the most intricate part of the proof. Setting $x_{n,a}(\cdot):=z_{n,a}(q_{n,a}(\cdot))$ and $x_{a}(\cdot):=z_{a}(q_{a}(\cdot))$ then gives the desired result.
Although Theorem (ref) is stated for sub-sequences, most applications require weak convergence of the full sequence $\{x_{n,a}(\cdot),q_{n,a}(\cdot)\}_{a}$. This motivates:
\begin{asm2} The sequence of policy rules $\{\pi_{n,j}\}_{j}$ is such that $\{x_{n,a}(\cdot),q_{n,a}(\cdot)\}_{a}$ has a weak limit under $\mathbb{P}_{n,0}$.\end{asm2}
Theorem (ref) already ensures $\{x_{n,a}(\cdot),q_{n,a}(\cdot)\}_{a}$ is tight under $\mathbb{P}_{n,0}$. Assumption 2 strengthens this to weak convergence. The assumption is needed to rule out pathological sequences of policy rules, e.g., sequences where the policy rules differ for even and odd $n$. It is therefore rather mild: if it does not hold, one should extract convergent subsequences and treat each as arising from a distinct protocol.
Theorem (ref) describes the behavior of $\left\{ x_{n_{k},a}(\cdot),q_{n_{k},a}(\cdot)\right\} _{a}$ under the reference distribution $\mathbb{P}_{n,0}$. Under local alternatives of the form $\mathbb{P}_{n,\bm{h}}$, the partial sum processes, $z_{n,a}(\cdot)$, acquire an asymptotic drift, converging weakly to $z_{a}(\cdot)\sim I_{a}^{1/2}h_{a}\cdot+W_{a}(\cdot)$.
Given that $\left\{ x_{a}(\cdot),q_{a}(\cdot)\right\} _{a}$ is adapted to the filtrations generated by $\{z_{a}(\cdot)\}_{a}$ and the exogenous randomization $U\sim\textrm{Uniform}[0,1]$, it follows that $\left\{ x_{n,a}(\cdot),q_{n,a}(\cdot)\right\} _{a}$ should converge weakly to $\{x_{a}(\cdot),q_{a}(\cdot)\}_{a}$, where the only difference is that the underlying processes $\{z_{a}(\cdot)\}_{a}$ now exhibit a linear drift. This is formalized in the following corollary.
Corollary (ref) is established in Appendix A using Theorem (ref) and standard change-of-measure arguments analogous to Le Cam's third lemma.
Theorem (ref) and Corollary (ref) are existence results: they establish that $\{q_{n,a}(\cdot)\}_{a}$ converge weakly to an allocation process $\{q_{a}(\cdot)\}_{a}$ in the limit experiment. While these results do not characterize the explicit form of $\{q_{a}(\cdot)\}_{a}$, this is generally not a limitation in practice. As the applications below illustrate, it is often possible to determine the form of optimal decisions in the limit experiment without knowing the specific structure of $\{q_{a}(\cdot)\}_{a}$. Theorem (ref), combined with change-of-measure arguments, then allows us to transfer these decisions back to the finite-sample setting and show that they are asymptotically optimal.
In this section, we illustrate how Theorem (ref) can be used to analyze estimation problems following an adaptive experiment.
Suppose that, upon completion of the experiment, we aim to estimate the unknown parameter vector $\bm{\theta}:=(\theta^{(1)},\theta^{(0)})$. Let $T_{n}$ denote a proposed estimator based on the entire data collected before the terminal time $t=1$. By construction, $T_{n}$ must be $\mathcal{I}_{n,1}:=\mathcal{G}_{n,q_{n,1}(1),q_{n,0}(1)}$ measurable.
Let $l(\cdot)$ be a non-negative convex loss function. Following the setup introduced earlier in this paper, we fix a reference parameter $\bm{\theta}_{0}$ and evaluate estimator loss under local alternatives of the form $\bm{\theta}_{0}+\bm{h}/\sqrt{n}$, where $\bm{h}\in\mathbb{R}^{d}$. Unlike classical settings, however, the choice of $\bm{\theta}_{0}$ is subject to important restrictions. In many adaptive experiments, only certain reference points yield non-degenerate diffusion asymptotics. Let $\bm{\Theta}_{0}$ denote the equivalence class of such admissible reference parameters. For example, in two-armed bandit experiments, this class consists of all parameter vectors satisfying $\theta^{(1)}=\theta^{(0)}$; otherwise, the resulting allocation processes become asymptotically degenerate, collapsing to either 0 or 1. More generally, we define $\bm{\Theta}_{0}$ to be the set of all parameter values $\bm{\theta}$ for which the weak limit of $q_{n,a}(\cdot)$ is not trivial (i.e., not identically 0) for any arm $a$. The reference parameter $\bm{\theta}_{0}$ must therefore lie in this set; otherwise some components of $\bm{\theta}$ will not be consistently estimable.\footnote{The set $\bm{\Theta}_{0}$ can be enlarged, however, if we are only interested in estimating some sub-components of $\bm{\theta}$. }
Given such a choice of $\bm{\theta}_{0}$, the frequentist risk of $T_{n}$, evaluated at the local parameter $\bm{h}$, is defined as \[ R_{n}(T_{n},\bm{h})=\mathbb{E}_{n,\bm{h}}\left[l\left(\sqrt{n}(T_{n}-\bm{\theta}(\bm{h}))\right)\right], \] where $\mathbb{E}_{n,\bm{h}}[\cdot]$ is the expectation under $\mathbb{P}_{n,\hm{h}}$, and $\bm{\theta}(\bm{h}):=\bm{\theta}_{0}+\bm{h}/\sqrt{n}$.
The estimation problem in the limit experiment is defined analogously. Given access to the sample paths of $\left\{ x_{a}(\cdot),q_{a}(\cdot)\right\} _{a}$ over $t\in[0,1]$, we seek an estimate of the local parameter $\bm{h}$. Let $T$ denote a candidate estimator, required to be $\mathcal{I}_{1}\equiv\mathcal{G}_{q_{1}(1),q_{0}(1)}$ measurable. The frequentist risk of this estimator is \[ R(T,\bm{h}):=\mathbb{E}_{\bm{h}}\left[l(T-\bm{h})\right]. \]
We term a sequence of estimators, $\{T_{n}\}_{n}$, tight at $\bm{\theta}_{0}$ if $\sqrt{n}(T_{n}-\bm{\theta}_{0})$ is asymptotically tight, i.e., bounded in probability, under $\mathbb{P}_{n,0}$. Tightness at $\bm{\theta}_{0}$ is a substantial relaxation of the usual notion of regularity---which requires the limit distribution of $\sqrt{n}(T_{n}-\bm{\theta}(\bm{h}))$ under $\mathbb{P}_{n,\bm{h}}$ to be the same for all $\bm{h}$.
The following theorem states that the asymptotic performance of any tight sequence of estimators $\{T_{n}\}_{n}$ is lower bounded, along subsequences, by the performance of some estimator $T$ in the limit experiment, and that this limit estimator depends only on the terminal values $\{x_{a}(1),q_{a}(1)\}_{a}$.
The surprising implication of Theorem (ref) is that knowledge of the just the terminal values of $\{x_{a}(\cdot),q_{a}(\cdot)\}_{a}$ is sufficient to characterize optimal estimators. The paths taken by these processes are not directly informative for estimation.
Let $m_{\bm{\theta}}(\cdot)$ denote a given prior over $\bm{\theta}$. This induces a local prior, $m(\cdot)$, over $\bm{h}$ through the transformation $\bm{h}=\sqrt{n}(\bm{\theta}-\bm{\theta}_{0})$. We consider an asymptotic regime wherein $m(\cdot)$ is assumed to be independent of $n$. The influence of the prior thus remains asymptotically non-negligible and the Bernstein-von Mises theorem does not apply. As discussed in adusumilli2025optimal, local priors offer a better approximation to finite-sample behavior because their influence does not diminish with sample size.
Theorem (ref) implies a lower bound on the Bayes risk corresponding to $m(\cdot)$:
Let $T^{*}$ denote the optimal Bayes estimator in the limit experiment. Then, Corollary (ref) implies that $R^{*}(m):=R(T^{*},m)$ is an asymptotic lower bound on the Bayes risk of any tight sequence of estimators $T_{n}$, i.e., \[ \liminf_{n\to\infty}R_{n}(T_{n},m)\ge R^{*}(m)\ \forall\ T_{n}. \] This lower bound does not require the use of subsequences.
By the likelihood principle, the optimal Bayes estimator in the limit experiment is algorithm independent and depends only on $\{x_{a}(1),q_{a}(1)\}_{a}$. For example, consider a prior $m_{0}(\cdot)$ on $\bm{h}\equiv(h^{(1)},h^{(0)})\in\mathbb{R}^{2}$ with independent Gaussian components: $\mathcal{N}(\mu_{0}^{(1)},\nu_{(1)}^{2})\times\mathcal{N}(\mu_{0}^{(0)},\nu_{(0)}^{2})$. Then, by standard results in stochastic filtering, the posterior distribution of $h^{(a)}$ at the end of the experiment is \[ h^{(a)}\vert q_{a}(1),x_{a}(1)\sim\mathcal{N}\left(\frac{I_{a}^{1/2}x_{a}(1)+\nu_{(a)}^{-2}\mu_{0}^{(a)}}{I_{a}q_{a}(1)+\nu_{(a)}^{-2}},\frac{1}{I_{a}q_{a}(1)+\nu_{(a)}^{-2}}\right), \] for each $a$. The optimal Bayes estimator of $h^{(a)}$ in the limit experiment, under the squared error loss, $l(\delta)=\delta^{2}$, is therefore \[ T^{*}=\frac{I_{a}^{1/2}x_{a}(1)+\nu_{(a)}^{-2}\mu_{0}^{(a)}}{I_{a}q_{a}(1)+\nu_{(a)}^{-2}}. \] Notably, this estimator is invariant to the choice of the sampling algorithm. As $\nu^{(a)}\to\infty$, we get the MLE estimator $T_{\textrm{mle}}=I_{a}^{-1/2}x_{a}(1)/q_{a}(1)$.
Corollary (ref) implies that the set of estimators depending only on $\{x_{a}(1),q_{a}(1)\}_{a}$ constitute a complete class in the limit experiment. Furthermore, we can lower bound the Bayes risk of any tight sequence of estimators using the Bayes risk of estimators in the limit experiment. These results are useful because determining the optimal estimator is a lot easier in the limit experiment. As seen above, Gaussian priors are particularly straightforward to analyze due to conjugacy. For general priors, computing the posterior is more involved, but one can employ approximate methods such as MCMC.
Given an optimal Bayes estimator, $T^{*}\left(\{x_{a}(1),q_{a}(1)\}_{a}\right)$, in the limit experiment, we can construct a finite sample version, \[ T_{n}^{*}:=\bm{\theta}_{0}+n^{-1/2}T^{*}\left(\{x_{n,a}(1),q_{n,a}(1)\}_{a}\right), \] by replacing $x_{a}(\cdot),q_{a}(\cdot)$ with the sample counterparts $x_{n,a}(\cdot),q_{n,a}(\cdot)$. Since $T^{*}$ is an estimator of $\bm{h}$, the transformation above converts it into an estimator of $\bm{\theta}$.
In practice, because $x_{n,a}(\cdot)$ depends on the information matrix, $I_{a}$, one would need to replace it with a consistent estimate. This can be supplied by the standard variance estimator, which remains consistent under general conditions, even if only at slower-than $n^{-1/2}$ rates. The construction also requires knowledge of the reference parameter $\bm{\theta}_{0}$. We suggest choosing this as the element from the equivalence class, $\bm{\Theta}_{0}$, that is closest to the prior median under $m_{\bm{\theta}}(\cdot)$.
Suppose that $T^{*}(\cdot)$ satisfies the conditions for a continuous mapping theorem. Together with ((ref)) and Theorem (ref), this implies \[ \left(
\right)\xrightarrow[P_{nT,0}]{d}\left(
\right), \] for any $\bm{h}$. Then, a similar argument as in the proof of Theorem (ref) shows that the frequentist risk of $T_{n}^{*}$ converges to that of $T^{*}$ in the limit experiment, as long as the loss function is bounded. But $T^{*}$ is the optimal Bayes estimator in the limit experiment, so the above implies that $T_{n}^{*}$ is asymptotically Bayes optimal as well, in the sense that its Bayes risk is arbitrarily close to $R^{*}(m)$ as $n\to\infty$.
Minimax risk is defined as $\inf_{T_{n}}\sup_{m(\cdot)}\int R_{n}(T_{n},\bm{h})dm(\bm{h})$, where $T_{n}$ ranges over all tight sequences of estimators, and $m(\cdot)$ ranges over all possible priors. The estimator $T_{n}^{*}$ that solves this problem is referred to as the minimax estimator. This estimator can also be interpreted as the equilibrium outcome of a zero-sum game between a decision-maker and nature: nature selects a prior $m(\cdot)$ to maximize the Bayes risk, while the decision-maker selects an estimator to minimize it.
In contrast to classical experiments, the minimax risk in adaptive experiments is often infinite. To see why, consider the two-armed bandit experiment from the illustrative example, and suppose the objective is to estimate $h^{(1)}$ in the limit experiment. Nature can make the problem arbitrarily hard by choosing a flat prior over $h^{(1)}$ and taking $h^{(0)}\to\infty$. In this case, because $h^{(0)}/h^{(1)}\to\infty$ with probability one, the bandit algorithm devotes negligible attention to arm 1, effectively yielding no data from which to estimate $h^{(1)}$. This leads to an infinite risk.
Continuing with the illustrative example from Section (ref), suppose we aim to estimate the parameter $\theta^{(1)}$ after conducting the experiment using a bandit algorithm (we show results under both UCB and Thompson Sampling). We assume independent and identical Gaussian priors over $\theta^{(0)}$ and $\theta^{(1)}$, given by $\mathcal{N}(\bar{\theta}_{},\bar{\sigma}^{2})$.
To apply our asymptotic framework, we reformulate the problem as a local estimation problem. Specifically, we treat $\bm{\theta}_{0}=(\bar{\theta},\bar{\theta})$---the vector of prior medians---as the reference value and define the local parameter $h^{(a)}:=\sqrt{n}(\theta^{(a)}-\bar{\theta}_{})$. This transformation induces a prior over $h^{(a)}$ of the form $\mathcal{N}(0,\nu^{2})$, where $\nu^{2}:=n\bar{\sigma}^{2}$. As detailed in Section (ref), $\nu^{2}$ is held fixed in our asymptotic regime even as $n$ increases, implying that the prior over $\theta^{(a)}$ increasingly concentrates around $\bar{\theta}$.
In the Bernoulli case, the score function is $\psi(y)=\left[\bar{\theta}_{}(1-\bar{\theta}_{})\right]^{-1}(y-\bar{\theta}_{})$, and the Fisher information is $I=\left[\bar{\theta}_{}(1-\bar{\theta}_{})\right]^{-1}$. Then, \[ x_{n,a}(t):=\frac{1}{\sqrt{n\bar{\theta}_{}(1-\bar{\theta}_{})}}\sum_{j=1}^{\lfloor nq_{n,a}(t)\rfloor}(Y_{j}^{(a)}-\bar{\theta}_{}). \] As shown in Section (ref), the asymptotically optimal Bayes estimator of $h^{(1)}$ is \[ \hat{h}^{(1)}=\left(Iq_{n,1}(1)+\nu^{-2}\right)^{-1}I^{1/2}x_{n,1}(1). \] This leads to the corresponding estimator for $\theta^{(1)}$: \[ \hat{\theta}^{(1)}=\frac{I^{-1}\nu^{-2}}{q_{n,1}(1)+I^{-1}\nu^{-2}}\bar{\theta}_{}+\frac{1}{q_{n,1}(1)+I^{-1}\nu^{-2}}\left(\frac{1}{n}\sum_{j=1}^{\lfloor nq_{n,1}(1)\rfloor}Y_{j}^{(1)}\right). \] Clearly, $\hat{\theta}^{(1)}$ has the same form as the usual shrinkage estimator of the population mean under a Gaussian prior. As $\nu\to\infty$, $\hat{\theta}^{(1)}$ becomes the MLE estimator $\frac{1}{nq_{n,1}(1)}\sum_{j=1}^{\lfloor nq_{n,1}(1)\rfloor}Y_{j}^{(1)}.$ Both estimators are independent of the adaptive sampling algorithm used.
To evaluate the finite-sample performance of $\hat{\theta}^{(1)}$, we conduct simulations based on the illustrative example from Section (ref), fixing $\theta^{(0)}=\bar{\theta}_{}$ and setting $\theta^{(1)}=\bar{\theta}+h^{(1)}/\sqrt{n}$ for values of $h^{(1)}$ ranging from $-0.5$ to $0.5$. We also take $\bar{\theta}_{}=0.1$ and set the prior standard deviation over $h^{(1)}$ to be $\nu=0.2$. This implies that the 95% prior credible interval for $\theta^{(1)}$ is approximately $[0.04,0.14]$ when $n=100$.
Panel A of Figure (ref) displays the corresponding frequentist risk profiles of the estimator for different values of $n$, when the data is obtained through UCB. Notably, the risk profiles are nearly identical across sample sizes, even when $n$ is as small as 100, highlighting the robustness of the estimator's performance in small samples. Panel B of the same figure plots the resulting Bayes risk under the local prior $(h^{(1)},h^{(0)})\sim\textrm{i.i.d}\ \mathcal{N}(0,\nu^{2})$. The distributions of risk are again almost identical across $n$.
Figure (ref) displays equivalent results when the data is obtained through Thompson Sampling. Surprisingly, the estimator attains very similar values of Bayes risk under both algorithms.
In-sample regret measures the difference between the (welfare) performance of the best possible action---which is unknown---and a chosen policy, evaluated on the same dataset used to select the policy. As we show below, it follows from Theorem (ref) that the in-sample regret from any sequence of policy rules $\{\pi_{n,j}\}_{j}$ can be asymptotically matched by that in the Gaussian diffusion limit experiment.\footnote{See Appendix (ref) for extensions to out-of-sample regret, also known as simple regret.}
Let $\mu_{n,a}(\bm{h}):=\mathbb{E}_{n,\bm{h}}[Y_{i}^{(a)}]$ denote the average reward corresponding to treatment $a$ when the local parameter is $\bm{h}$. Following hirano2023asymptotic and adusumilli2025optimal, the reference parameter is chosen such that $\mu_{n,a}(\bm{0})=0$. The frequentist regret of any $\pi_{n}\equiv\{\pi_{n,j}\}_{j}$ is given by \[ W_{n}(\bm{h})=\sqrt{n}\left\{ \max_{a}\mu_{n,a}(\bm{h})-\sum_{a}\mu_{n,a}(\bm{h})\mathbb{E}_{n,\bm{h}}[q_{n,a}(1)]\right\} . \] Analogously, in the limit experiment, the frequentist regret is given by \[ W(\bm{h})=\max_{a}\dot{\mu}_{a}^{\intercal}\bm{h}-\sum_{a}\dot{\mu}_{a}^{\intercal}\bm{h}\mathbb{E}_{\bm{h}}[q_{a}(1)], \] where $\dot{\mu}_{a}(\cdot)$ is defined in the assumption below:
\begin{asm3} There exists $\dot{\mu}_{a}\in\mathbb{R}^{d}$ and $\delta_{n}\to0$ such that $\sqrt{n}\mu_{n,a}(\bm{h})=\dot{\mu}_{a}^{\intercal}\bm{h}+\delta_{n}\vert\bm{h}\vert^{2}\ \forall\ \bm{h}$. \end{asm3}
Recall from the measurability requirement on $q_{a}(\cdot)$ that the allocation process at any time $t$ needs to be adapted to the sample paths of $\{x_{a}(s);s\le t\}_{a}$. Theorem (ref) thus implies that the regret profile of any policy can be asymptotically matched by one that depends only on the sample paths of $\{x_{a}(s);s\le t\}_{a}$.
Previous work by this author (adusumilli2025optimal) characterized the optimal Bayes and minimax risks in this setting, and showed that they can be attained by a sequence of policy rules that depend on just $\left\{ x_{a}(t),q_{a}(t)\right\} _{a}$, i.e., the past values of these variables are not relevant. Theorem (ref) is more general, in that it applies to arbitrary sequences of policy rules, but it makes use of a larger information set that includes the entire sample paths of $\{x_{a}(s);s\le t\}_{a}$ until time $t$.
Anytime-valid inference refers to statistical procedures that maintain valid error control (e.g., Type I error or confidence coverage) uniformly over time, without compromising inference guarantees. A central tool in this framework is the e-process, a nonnegative stochastic process that starts at 1 and is a super-martingale (i.e., its expectations is always less than or equal to 1) under the null hypothesis. Much of the existing literature on anytime-valid inference has focused on sequential experiments with fixed sampling strategies, where the only adaptive element is the stopping rule. Here, we extend the notion of an e-process to the more general setting of multi-treatment adaptive experiments.
Formally, we analyze e-processes and anytime-valid inference for tests of the form $H_{0}:\bm{\theta}\in\Theta_{0}$ vs $H_{1}:\bm{\theta}\in\Theta_{1}$. We assume that the reference parameter $\bm{\theta}_{0}:=(\theta_{0}^{(1)},\theta_{0}^{(0)})$ from Section (ref) always lies in the null region $\Theta_{0}$. As in Section (ref), we index each $\bm{\theta}$ by $\bm{h}\in\mathbb{R}^{d}\times\mathbb{R}^{d}$ such that $\bm{\theta}=\bm{\theta}_{0}+\bm{h}/\sqrt{n}$. Denote the set of all $\bm{h}$ representing $\bm{\theta}\in\Theta_{0}$ by $\mathcal{H}_{0}$ and those representing $\bm{\theta}\in\Theta_{1}$ by $\mathcal{H}_{1}$. We restrict attention to `asymptotically stable' hypothesis testing problems, wherein the regions $\mathcal{H}_{0},\mathcal{H}_{1}$ do not change with $n$.
Let $\mathcal{G}_{n,\gamma_{1},\gamma_{0}}$ denote the filtration introduced in Section (ref). The e-process for multi-treatment adaptive experiments is formally defined as follows:
Definition (ref) generalizes the usual notion of an e-process to a multi-indexed super-martingale, where the indices are the treatment allocation proportions. In this framework, the value of $\varepsilon_{n}(\cdot)$ depends on the trajectory of the empirical allocation process, and the super-martingale property holds across all possible empirical allocation processes, i.e., all possible adaptive experiments. A central feature of the e-process is that it is algorithm-free: it remains a valid supermartingale at any time point of any adaptive experiment.
Notably, Definition (ref) does not require $t$ to be a stopping time. Optimal stopping can be incorporated by introducing a designated “default” treatment, such that assigning units to this treatment is equivalent to halting the experiment. For notational simplicity, we focus on the two-treatment case, though the extension to more arms is conceptually straightforward. If there were optimal stopping, the interpretation of $n$ would, however, change; it then no longer denotes a fixed sample size, but instead serves as a normalization constant. In particular, it associates the time index $t$ to the closeness of alternatives being considered: if the aim is to analyze performance against local alternatives of the form $\bm{\theta}_{0}+\bm{h}/\sqrt{n}$, then $t$ should represent the number of observations collected in units of $n$.
E-processes serve as dynamic measures of evidence against the null hypothesis $H_{0}$. At any point $(q_{0},q_{1})$, the value $\varepsilon_{n}(q_{0},q_{1})$ can be interpreted as the current payoff from wagering one unit against $H_{0}$. This interpretation holds uniformly over time and across different adaptive experiments. Moreover, taking the inverse defines a p-process $p_{n}(q_{1},q_{0}):=1/\varepsilon_{n}(q_{1},q_{0})$, which yields an anytime-valid p-value. Specifically, for any $\bm{h}\in\mathcal{H}_{0}$, empirical allocation $\{q_{n,a}(\cdot)\}_{a}$, and time $t$,
Thus, $p_{n}\left(q_{n,1}(t),q_{n,0}(t)\right)$ is a valid classical $p$-value at any time-point of any adaptive experiment.
Following ramdas2023game and grunwald2024safe, a common approach to evaluating e-processes in stopping-time experiments is through the Growth Rate Optimality (GRO) criterion. GRO assesses the quality of an e-process $\varepsilon_{n}(\cdot)$ based on its expected log-growth under an alternative hypothesis $\bm{h}\in\mathcal{H}_{1}$. We extend this criterion to the multi-treatment adaptive setting by defining the GRO score
where $\{q_{n,a}(t)\}_{a}$ denotes the empirical allocation process at time $t$. The GRO score thus depends jointly on the e-process, the alternative $\bm{h}$, and the adaptive experiment employed (as indexed by the empirical allocation process).
We say that an e-process $\varepsilon_{n}(\cdot)$ uniformly dominates another process $\varepsilon_{n}'(\cdot)$ in terms of GRO if \[ R_{n}\left(\varepsilon_{n};\bm{h},\{q_{n,a}(t)\}_{a}\right)\ge R_{n}\left(\varepsilon_{n}';\bm{h},\{q_{n,a}(t)\}_{a}\right) \] for all $\bm{h}\in\mathcal{H}_{1}$, at all time points $t$, and for all possible experiments, i.e., all possible empirical allocation processes $\{q_{n,a}(t)\}_{a}$. This notion of uniform GRO dominance is quite strong and, in general, no single e-process may achieve it. One way to relax this requirement is to instead use the mixture-GRO (mGRO) criterion, which averages the GRO score using a prior, or weight function, $w(\cdot)$ over $\mathcal{H}_{1}$: \[ R_{n}\left(\varepsilon_{n};w(\cdot),\{q_{n,a}(t)\}_{a}\right)=\int\mathbb{E}_{n,\bm{h}}\left[\ln\varepsilon_{n}(q_{n,1}(t),q_{n,0}(t))\right]dw(\bm{h}). \] The mGRO criterion ranks e-processes in terms of their average performance over plausible alternatives, rather than requiring dominance for every possible alternative.
An alternative criterion, following grunwald2024safe, is the REGROW (RElative GRowth Optimality in Worst case) score:
REGROW measures the (negative of the) worst-case GRO-regret, where GRO-regret is defined as the difference between the GRO value of $\varepsilon_{n}(\cdot)$ and the GRO value of the log-likelihood ratio process corresponding to a specific alternative $\bm{h}_{1}\in\mathcal{H}_{1}$. The latter is the ideal, i.e., uniformly GRO optimal, e-process for testing $H_{0}:\bm{h}=0$ versus $H_{1}:\bm{h}=\bm{h}_{1}$. REGROW thus benchmarks the performance of $\varepsilon_{n}(\cdot)$ against the optimal e-process that would be achievable if $\bm{h}_{1}$ were known in advance. A higher REGROW score indicates more robust performance across alternatives.
GRO and its variants, mGRO and REGROW, offer an alternative paradigm to the classical power criterion for hypothesis testing. However, determining and computing optimal e-processes becomes considerably more challenging when $\mathcal{H}_{0}$ is composite or when the REGROW criterion is used. Closed-form expressions are only known for certain parametric families $\mathbb{P}_{n,h}$.
In order to circumvent this complexity in the fixed $n$ setting, we propose employing a local-asymptotic criterion. Accordingly, we relax the definition of e-processes given previously, and call a sequence of non-negative, $\mathcal{G}_{n,q_{1},q_{0}}$-adapted stochastic processes, $\varepsilon_{n}(\cdot)$, asymptotic e-processes if \[ \limsup_{n}\mathbb{E}_{n,\bm{h}}\left[\varepsilon_{n}\left(q_{n,1}(t),q_{n,0}(t)\right)\right]\le1\ \forall\ \bm{h}\in\mathcal{H}_{0},\ \forall\ t\in[0,1], \] and for all possible empirical allocation processes $\{q_{n,a}(\cdot)\}_{a}$.
We then define an equivalent notion of an e-process in the limit experiment. Recall that the limit experiment is characterized by Gaussian process signals $z_{a}(q):=I_{a}^{1/2}h^{(a)}q+W_{a}(q)$. The e-process in the limit experiment is a $\mathcal{G}_{q_{1},q_{0}}$-adapted stochastic process designed for testing $H_{0}:\bm{h}\in\mathcal{H}_{0}$ versus $H_{1}:\bm{h}\in\mathcal{H}_{1}$. The formal definition is given below. First, a bit of terminology: let $\bm{\mathcal{Q}}$ denote the collection of all allocation processes in the limit experiment that are the weak limits of some possible sequence of empirical allocation processes $\{q_{n,a}(\cdot)\}_{a}$.
Relative to Definition (ref), the above definition restricts the set of allocation processes to $\bm{\mathcal{Q}}$. The rationale behind this is mainly technical: it allows us to avoid dealing with $\{q_{a}(\cdot)\}_{a}$ that cannot be generated by any sequence of experimental protocols in the actual experiment. Whether $\bm{\mathcal{Q}}$, in fact, includes all possible allocation processes is currently unknown (to this author).
The GRO, mGRO and REGROW criteria in the limit experiment retain the same form as ((ref)), except that $\ensuremath{\mathbb{E}_{n,\bm{h}}[\cdot]}$ is replaced by $\mathbb{E}_{\bm{h}}[\cdot]$, e.g., the GRO criterion in the limit experiment becomes \[ R\left(\varepsilon_{};\bm{h},\{q_{a}(t)\}_{a}\right)=\mathbb{E}_{\bm{h}}\left[\ln\varepsilon_{}(q_{1}(t),q_{0}(t))\right];\ \bm{h}\in\mathcal{H}_{1}. \]
We derive an asymptotic representation theorem for e-processes, which establishes that for any sequence of asymptotic e-processes, there exists a dominating e-process in the limit experiment with respect to the GRO criterion and its variants. This is based on the following regularity conditions:
\begin{asm4} As functions of $\{q_{a}\}_{a}$, the sequence $\{z_{n,a}(\cdot),\varepsilon_{n}(\cdot,\cdot)\}_{a}$ converges weakly under $\mathbb{P}_{n,\bm{0}}$. Furthermore, for each $t$ and each possible weakly convergent sequence of empirical processes $\{q_{n,a}(\cdot)\}_{a}$, the sequence $\left\{ \ln\varepsilon_{n}\left(q_{n,1}(t),q_{n,0}(t)\right)\right\} _{n}$ is uniformly integrable with respect to each element in $\{\mathbb{P}_{n,\bm{h}}\}_{\bm{h}\in\mathcal{H}_{1}}$.\end{asm4}
The first part of Assumption 4 is an analogue of Assumption 2. It restricts the class of asymptotic e-processes by requiring them to be asymptotically equicontinuous (so that they have a weak limit). The second part of Assumption 4 is an additional regularity condition ensuring that the GRO scores are asymptotically convergent under the alternative hypotheses. Both properties will need to be verified on a case-by-case basis. For an example, see Appendix (ref).
An important implication of Theorem (ref) is that any asymptotically optimal e-process need depend only on $\{z_{a}(q_{a})\}_{a}$, in addition to its index $\{q_{a}\}_{a}$. Thus, at any time-point $t$, the set of sufficient statistics for anytime-valid inference is again $\{x_{a}(t),q_{a}(t)\}_{a}$, and the past values of these processes are not relevant for inference.
The extension to the mGRO criterion is a straightforward consequence of Theorem (ref) and the monotone convergence theorem.
In most cases, the function $g(\cdot)$ in the statement of Corollary (ref) can be set to 0; see Appendix (ref), for instance.
Analysis of the REGROW criterion requires an additional assumption:
\begin{asm5} For any time $t$, $\bm{h}_{1}\in\mathcal{H}_{1}$, and any sequence of empirical allocation processes $\{q_{n,a}(\cdot)\}_{a}$ weakly converging to some $\{q_{n,a}(\cdot)\}_{a}\in\bm{\mathcal{Q}}$, we have that $\mathbb{E}_{n,\bm{h}_{1}}\left[\ln\frac{d\mathbb{P}_{n,\hm{h}_{1}}}{d\mathbb{P}_{n,\bm{0}}}(q_{n,1}(t),q_{n,0}(t))\right]$ converges to $\mathbb{E}_{\bm{h}_{1}}\left[\ln\frac{d\mathbb{P}_{\hm{h}_{1}}}{d\mathbb{P}_{\bm{0}}}(q_{1}(t),q_{0}(t))\right]$.\end{asm5}
The assumption states that the KL-divergences, $\textrm{KL}\left(\mathbb{P}_{n,\bm{h}_{1}}\mid\mid\mathbb{P}_{n,\bm{0}}\right)$, in the original experiment converge asymptotically to KL-divergences, $\textrm{KL}\left(\mathbb{P}_{\bm{h}_{1}}\mid\mid\mathbb{P}_{\bm{0}}\right)$, in the limit experiment. This involves restrictions on the parametric models allowed.
In what follows, we simplify matters by assuming that the null hypothesis $\mathcal{H}_{0}$ is a singleton, consisting solely of the reference parameter $\bm{\theta}_{0}$.\footnote{When there is only adaptive stopping, compound nulls are typically addressed in anytime-valid inference using the method of reverse information projection (see ramdas2023game, for a survey). Extending this approach to the adaptive sampling setting is more involved and left for future work. See, however, Appendix (ref) for the simpler case of testing parameters corresponding to a single treatment arm.} Theorem (ref) and Corollaries (ref), (ref) establish asymptotic upper bounds on the GRO, mGRO and REGROW criteria. These bounds are obtained by optimizing the respective criteria within the limit experiment.
A natural question is whether the restriction $\{q_{a}(\cdot)\}_{a}\in\bm{\mathcal{Q}}$, imposed in the definition of the limiting e-process (Definition (ref)), affects the outcome of the optimization. In practice, it does not. The general approach to optimizing these criteria involves constructing, for each fixed allocation process $\{q_{a}(\cdot)\}_{a}$ and time $t$, an e-value: an $\mathcal{I}_{t}$-measurable random variable $\varepsilon$ satisfying \[ \mathbb{E}_{\bm{h}}[\varepsilon]\leq1\quad\text{for all }\bm{h}\in\mathcal{H}_{0}. \] We then identify the point-wise optimal e-value, $\varepsilon_{q_{1}(t),q_{0}(t)}^{*}$, that maximizes the desired criterion (GRO, mGRO, or REGROW) at the given allocation point $\{q_{a}(t)\}_{a}$. The GRO, mGRO or REGROW value of $\varepsilon_{q_{1}(t),q_{0}(t)}^{*}$ furnishes a sharp upper bound---uniformly over all possible e-processes---for the corresponding criterion evaluated at that allocation.
The final step is to determine whether these pointwise-optimal e-values can be coherently combined, or “strung together”, into a full e-process satisfying Definition (ref). If such a construction is not possible, it usually implies that a globally optimal e-process---one that simultaneously achieves pointwise optimality at all allocation points---does not exist.
As a first illustration of this approach, consider the mGRO criterion. For a given $\{q_{a}(t)\}_{a}$, the mGRO optimal e-value in the limit experiment is \[ \varepsilon_{q_{1}(t),q_{0}(t)}^{*}=\int\exp\sum_{a}\left\{ h^{(a)\intercal}I_{a}^{1/2}z_{a}(q_{a}(t))-\frac{q_{a}(t)}{2}h^{(a)\intercal}I_{a}h^{(a)}\right\} dw(\bm{h}). \] Importantly, the form of $\varepsilon_{q_{1}(t),q_{0}(t)}^{*}$ does not change with $\{q_{a}(t)\}_{a}$, implying that the optimal e-process $\varepsilon^{*}(\cdot,\cdot)$ can be constructed as $\varepsilon^{*}(q_{1},q_{0})=\varepsilon_{q_{1},q_{0}}^{*}$. It is straightforward to verify that the resulting e-process satisfies ((ref)) using Lemma (ref) and standard martingale arguments.
Replacing $\{z_{a}(\cdot)\}_{a}$ with $\{z_{n,a}(\cdot)\}_{a}$ yields the asymptotically mGRO-optimal e-process:
Appendix (ref) describes primitive conditions under which $\varepsilon_{n}^{*}(q_{1},q_{0})$ satisfies the requirements for Corollary (ref). Essentially, we need $w(\cdot)$ to be a sub-Gaussian distribution and $\psi_{a}(Y_{i}^{(a)})$ to have finite $2+p$ moments, for some $p>0$.
grunwald2024safe show that the REGROW-optimal e-value, $\bar{\varepsilon}_{q_{1}(t),q_{0}(t)}^{*}$, coincides with the mGRO-optimal e-value for a specific, least-favorable weighting function $w_{q_{1}(t),q_{0}(t)}^{*}(\cdot)$. The authors also show that this weighting function is obtained as the solution to
where $\Delta(\mathcal{H}_{1})$ denotes the set of all probability measures supported on $\mathcal{H}_{1}$, \[ \mathbb{P}_{w}(\cdot):=\int\mathbb{P}_{\bm{h}}(\cdot)dw(\bm{h}), \] and $\textrm{KL}_{q_{1}(t),q_{0}(t)}\left(\mathbb{P}_{\bm{h}}\mid\mid\mathbb{P}_{w}\right)$ represents the KL divergence between $\mathbb{P}_{\bm{h}},\mathbb{P}_{w}$ when these probability measures are restricted to the filtration $\mathcal{I}_{t}\equiv\mathcal{G}_{q_{1}(t),q_{0}(t)}$.\footnote{For instance, the restriction of $\mathbb{P}_{\bm{h}}$ to $\mathcal{I}_{t}$ is the probability measure induced by the sample paths of $z_{a}(s)=I_{a}^{1/2}h^{(a)}s+W_{a}(s)$ between $0$ and $q_{a}(t)$.}
grunwald2024safe further demonstrate that the optimized value of the objective in ((ref)) provides an upper bound on the REGROW criterion $\mathcal{R}\left(\varepsilon;\{q_{a}(t)\}_{a}\right)$: \[ \sup_{\varepsilon}\mathcal{R}\left(\varepsilon;\{q_{a}(t)\}_{a}\right)\le\mathbb{E}_{\bm{h}\sim w_{q_{1}(t),q_{0}(t)}^{*}}\left[\textrm{KL}_{q_{1}(t),q_{0}(t)}\left(\mathbb{P}_{\bm{h}}\mid\mid\mathbb{P}_{w}\right)\right]. \]
Replacing the $q_{1}(t),q_{0}(t)$ subscripts with $(\cdot)$ for ease of notation, note that $w_{(\cdot)}^{*}$ can be alternatively characterized as:
where $p_{\bm{h}},p_{w}$ denote the densities of $\mathbb{P}_{\bm{h}},\mathbb{P}_{w}$ with respect to some dominating measure, and $I_{(\cdot)}\left(\cdot;\cdot\right)$ represents mutual information under the restricted filtration. The optimization problem ((ref)) has a natural information-theoretic interpretation. Consider an information transmission channel with input $\hm{h}$ and output $\{z_{a}(s);s\le q_{a}(t)\}_{a}$. Then, the quantity $\sup_{w}I_{(\cdot)}\left(w;p_{w}\right)$ corresponds to the channel capacity, and $w_{(\cdot)}^{*}$ represents the optimal signal distribution.
Crucially, the least-favorable distribution $w_{(\cdot)}^{*}$ depends on both the chosen allocation process and the structure of the alternative hypothesis set $\mathcal{H}_{1}$. It is essential that $\mathcal{H}_{1}$ be compact---without it, the channel capacity (and hence the REGROW value) becomes infinite. Because of the dependence on the allocation process, a globally optimal REGROW e-process does not exist.
As an alternative, one could fix a specific allocation process $\{q_{a}(t)\}_{a}$ and an alternative region $\mathcal{H}_{1}$, from which an optimal weighting function $w_{(\cdot)}^{*}$ can be derived, e.g., by employing the Blahut-Arimoto algorithm (see, arimoto1972algorithm,blahut1972computation). The resulting e-process is thus locally REGROW optimal relative to the chosen design and alternative set (it would also be globally mGRO optimal relative to the least favorable distribution $w^{*}(\cdot)$). Appendix (ref) provides an example of such a locally REGROW optimal e-process, constructed to be optimal against fixed values of $\{q_{a}\}_{a}$.
We revisit again the illustrative example from Section (ref), now focusing on constructing an anytime-valid test of the null hypothesis $H_{0}:\theta^{(1)}=0.1$ against the two-sided alternative $H_{1}:\theta^{(1)}\neq0.1$. We take the reference parameter vector to be $\bm{\theta}_{0}=(0.1,0.1)$---as it is the only reference parameter that induces non-trivial asymptotic limits---and consider local alternatives of the form $\bm{\theta}=\bm{\theta}_{0}+\bm{h}/\sqrt{n}$.
Although the null is composite due to the unrestricted nature of $h^{(0)}$, Appendix (ref) shows that it is without loss of generality to ignore observations from arm 0. This dimensionality reduction allows us to construct optimal e-processes using only the data from arm 1, leveraging the techniques developed in Section (ref).
We employ the mGRO criterion with the weighting function $w_{1}(\cdot)\sim\mathcal{N}(0,\nu^{2})$, where $\nu^{2}=1$. This corresponds to a $\mathcal{N}(0.1,1/n)$ prior over $\theta^{(1)}$. We examine sample sizes $n\in\{500,1000,1500,2000\}$. For instance, when $n=1000$, this prior places 95% of the mass within the credible interval $[0.06,0.14]$ for $\theta^{(1)}$. As noted earlier, in real-world scenarios, the difference in click-through rates between various variants of a website is typically less than 1%. Hence, our chosen weighting aligns well with realistic alternative values of $\theta^{(1)}$.
Under this setup, the asymptotically optimal mGRO e-process (from equation (ref)) takes the explicit form:
where $I_{1}:=[\theta_{0}^{(1)}(1-\theta_{0}^{(1)})]^{-1}$ denotes the Fisher information.
Panel A of Figure (ref) reports the uniform-over-time finite-sample size of this e-process under Thompson Sampling and UCB allocation rules, calculated as\footnote{The least-favorable configuration for $h^{(0)}$ appears to be $-\infty$, corresponding to $\theta^{(0)}=0$.} \[ \sup_{h^{(0)}}P_{n,(0,h^{(0)})}\left(\max_{t}\varepsilon_{n}^{*}(q_{n,1}(t))\ge20\right), \] The critical value of $20$ implies that the p-process conversion (from the e-process) targets an anytime valid size of $5$%. However, the size under specific policies can be smaller and it indeed turns out that the test is conservative for the policies considered---a behavior that is expected, since the validity of the e-process applies to all adaptive algorithms and not just Thompson Sampling or UCB.
Panel B plots the evolution of the GRO value over time for this e-process under a local alternative $\bm{h}\equiv(h^{(1)},h^{(0)})=(1/\sqrt{n},0)$, for varying sample sizes, all under UCB (see Appendix (ref) for equivalent results under Thompson Sampling). The resulting curves exhibit remarkable stability across $n$, indicating that the asymptotic approximation is already accurate at such small sample sizes as $n=500$.
This article introduces a continuous-time formalism for analyzing fully adaptive experiments. The formalism is based on the notion of allocation processes, also introduced in this article. We show that any empirical allocation process, as induced by some policy rule, converges weakly to a corresponding allocation process in a limit experiment governed by Gaussian diffusions. The various applications, ranging from point-estimation to anytime-valid inference illustrate the utility of this framework.
Beyond these applications, the continuous-time formulation offers a powerful tool for addressing design problems in adaptive experimentation. Though not based on the results reported in this article, prior work by this author has applied the continuous-time framework to derive optimal algorithms for bandit experiments and costly sampling problems (adusumilli2025optimal,adusumilli2025sample). The current results provide an easy-to-use and generalizable template for transferring optimal designs from the limit experiment back to the finite-sample setting. Looking ahead, we expect this approach to be broadly useful in addressing a range of open questions in adaptive experimentation---e.g., in deriving optimal strategies for best-arm identification with multiple treatments, or in designing adaptive experiments in strategic environments involving interactions between an experimenter and a regulator.