EconBase
← Back to paper

Minimax and Bayes Optimal Best-Arm Identification

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.

75,548 characters · 32 sections · 36 citation commands

Rendered from LaTeX for readability, not typeset faithfully. Citation keys are highlighted; maths is left as source; figures, tables and equation environments are summarised rather than reproduced; unrecognised commands are greyed out so nothing is silently dropped. Email addresses are removed.

Minimax and Bayes Optimal Best-Arm Identification

abstractThis study investigates minimax and Bayes optimal strategies for fixed-budget best-arm identification. We consider an adaptive procedure consisting of a sampling phase followed by a recommendation phase, and we design an adaptive experiment within this framework to efficiently identify the best arm, defined as the one with the highest expected outcome. In our proposed strategy, the sampling phase consists of two stages. The first stage is a pilot phase, in which we allocate samples uniformly across arms to eliminate clearly suboptimal arms and to estimate outcome variances. Before entering the second stage, we solve a Gaussian minimax game, which yields a sampling ratio and a decision rule. In the second stage, samples are allocated according to this sampling ratio. After the sampling phase, the procedure enters the recommendation phase, where we select an arm using the decision rule. We prove that this single strategy is simultaneously asymptotically minimax and Bayes optimal for the simple regret, and we establish upper bounds that coincide exactly with our lower bounds, including the constant terms.

Introduction

We investigate the problem of fixed-budget best-arm identification Audibert2010bestarm, an instance of adaptive experimental design for identifying the arm with the highest expected outcome. This problem is also known by various names across disciplines, including ordinal optimization Chen2000simulationbudget and adaptive experimental design for policy choice Kasy2021adaptivetreatment.

An adaptive experimental procedure in BAI usually consists of two phases Kaufmann2016complexity, the sampling phase and the recommendation phase. Given a total of $T$ rounds, the sampling phase allocates samples to arms at each round based on the observations obtained up to that point. After the final round, the procedure enters the recommendation phase, where an arm is selected based on the collected data.

In this setting, we propose a strategy and establish its minimax and Bayes optimality in terms of the simple regret, defined as the difference between the expected outcome of the best arm and that of the recommended arm. We first describe our strategy, Two-Stage Saddle-Point Allocation with Screening (TS-SPAS). We then derive minimax and Bayes lower bounds and show that our worst-case and average-case upper bounds coincide exactly with these lower bounds, including the constant terms.

Setup

We formulate the problem as follows. There are $K$ arms, and each arm $a \in [K] \coloneqq \{1,2,\dots, K\}$ has a potential outcome $Y_a \in {\mathcal{Y}}$, where ${\mathcal{Y}} \subseteq {\mathbb{R}}$ denotes the outcome space. Each potential outcome $Y_a$ follows a (marginal) distribution $P_{a, \mu_a}$ parameterized by $\mu_a \in {\mathcal{M}}$, where ${\mathcal{M}} \subset {\mathbb{R}}$ is a parameter space. For the parameter vector ${\bm{\mu}} \coloneqq (\mu_1, \mu_2, \dots, \mu_K) \in {\mathcal{M}}^K$, let ${\bm{P}}_{{\bm{\mu}}} \coloneqq (P_{1, \mu_1}, P_{2, \mu_2}, \dots, P_{K, \mu_K})$ be a collection of parametric distributions. The parameter $\mu_a \coloneqq {\mathbb{E}}_{P_{a,\mu_a}}[Y_a]$ is the mean of arm $a$.

Under a distribution ${\bm{P}}_{{\bm{\mu}}}$, our objective is to identify the best arm \[ a^*_{{\bm{\mu}}} = \operatorname*{arg\,max}_{a \in [K]} \mu_a, \] through an adaptive experiment in which data are sampled from ${\bm{P}}_{{\bm{\mu}}}$ according to our strategy.

\paragraph{Adaptive experiment.} Let $T$ denote the total sample size, also referred to as the budget. Assume there exist i.i.d. random vectors $(Y_{1,t},\ldots,Y_{K,t})_{t\in[T]}$ such that for each $a\in[K]$ and each $t\in[T]$, $Y_{a,t}\sim P_{a,\mu_a}$. Then, we consider an adaptive experimental procedure consisting of two phases:

enumerate• Sampling phase: For each $t \in [T] \coloneqq \{1, 2, \dots, T\}$: \begin{itemize} • An arm $A_t \in [K]$ is selected based on the past observations $\{(A_s, Y_s)\}_{s=1}^{t-1}$. • The corresponding outcome $Y_t$ is observed, where $ Y_t \coloneqq \sum_{a \in [K]} \mathbbm{1}[A_t = a]Y_{a, t}$. \end{itemize} • Recommendation phase: At the end of the experiment ($t = T$), based on the observed outcomes $\{(A_t, Y_t)\}_{t=1}^{T}$, we choose an arm $\widehat{a}_T \in [K]$ as an estimate of the best arm $a^*_{{\bm{\mu}}}$.

Our task is to design a strategy $\delta$ that determines how arms are selected during the sampling phase and how the best arm is recommended at the end of the experiment. A strategy $\delta$ is formally defined as a pair $\left(\left(A_t^{\delta}\right)_{t \in [T]}, \widehat{a}_T^{\delta}\right){\delta}}$, where $\left(A_t^{\delta}\right)_{t \in [T]} \in [K]^T$ are the arm indices (actions) selected in the sampling phase, and $\widehat{a}_T^{\delta}$ is the estimator of the best arm $a^*_{{\bm{\mu}}}$ in the recommendation phase. For simplicity, we omit the dependence on $\delta$ when it is clear from the context.

\paragraph{Regret.} The performance of a strategy $\delta$ is measured by the expected simple regret, defined as:

align*[align* omitted — 311 chars of source]

This measures the welfare loss from deploying the recommended treatment in a future population. In other words, the goal is to design a strategy $\delta$ that minimizes the simple regret ${\mathrm{Regret}}^\delta_T({\bm{P}}_{{\bm{\mu}}})$. For simplicity, we refer to the expected simple regret as the simple regret in this study, although the simple regret originally refers to the random variable $Y_{a^*_{{\bm{\mu}}}} - Y_{\widehat{a}_T^{\delta}}$ without expectation.

\paragraph{Notation.} Let ${\mathbb{P}}_{{\bm{P}}_{{\bm{\mu}}}}$ denote the probability law under ${\bm{P}}_{{\bm{\mu}}}$, and let ${\mathbb{E}}_{{\bm{P}}_{{\bm{\mu}}}}$ represent the corresponding expectation operator. For notational simplicity, depending on the context, we abbreviate ${\mathbb{P}}_{{\bm{P}}_{{\bm{\mu}}}}[\cdot]$, ${\mathbb{E}}_{{\bm{P}}_{{\bm{\mu}}}}[\cdot]$, and ${\mathrm{Regret}}^\delta_T({\bm{P}}_{{\bm{\mu}}})$ as ${\mathbb{P}}_{{\bm{\mu}}}[\cdot]$, ${\mathbb{E}}_{{\bm{\mu}}}[\cdot]$, and ${\mathrm{Regret}}^\delta_T({\bm{\mu}})$, respectively. For each $a \in [K]$, let $P_{a, \mu_a}$ denote the marginal distribution of $Y_a$ under ${\bm{P}}_{{\bm{\mu}}}$. Denote the variance of $Y_a$ under the data-generating process by $\sigma_a^2(\mu_a)$, and write $\sigma_a(\mu_a)\coloneqq \sqrt{\sigma_a^2(\mu_a)}$. Let ${\mathcal{F}}_t = \sigma(A_1, Y_1, \ldots, A_t, Y_t)$ be the sigma-algebra. For each arm $a$ and mean $\mu\in{\mathcal{M}}$, let ${\mathbb{P}}_{a,\mu}$ and ${\mathbb{E}}_{a,\mu}$ denote probability and expectation under $P_{a,\mu}$. When the arm index is clear from context, we also write ${\mathbb{E}}_{\mu}$ for ${\mathbb{E}}_{P_{\mu}}$.

We denote the gap between the expected outcomes for the best arm $a^*_{{\bm{\mu}}}$ and an arm $a\in[K]$ by $\Delta_{{\bm{\mu}}}(a) \coloneqq \max_{b\in[K]}\mu_{b} - \mu_a = \mu_{a^*_{{\bm{\mu}}}} - \mu_a$. In the bandit literature, this gap plays an important role in theoretical evaluations.

To make all argmax-based rules measurable without appealing to “ties occur with probability zero”, we fix the following deterministic tie-break: for any finite set $S\subset{\mathbb{N}}$ and any real vector $(x_a)_{a\in S}$, we define the argmax as $\operatorname*{arg\,max}_{a\in S} x_a \coloneqq \min\Big\{a\in S\colon x_a=\max_{b\in S}x_b\Big\}$. For any $m$-dimensional vector ${\bm{v}} = (v_i)_{i\in{\mathcal{V}}}\in {\mathbb{R}}^m$, where ${\mathcal{V}}$ is its index set, let $v_{(i)}$ be its $i$-th largest value; that is, $v_{(1)} = \max_{i\in{\mathcal{V}}} v_i$ and $v_{(2)} = \operatorname*{arg\,max}_{i\in{\mathcal{V}}\setminus \{\operatorname*{arg\,max}_{k\in{\mathcal{V}}} v_k\}} v_i$.

Let $N_{a,t}\coloneqq \sum_{s=1}^t \mathbbm{1}[A_s=a]$. Let $\widehat{\mu}_{a, t}$ be the sample mean of the observed outcomes for arm $a$, that is, $\widehat{\mu}_{a,t}\coloneqq \frac{1}{N_{a,t}}\sum_{s=1}^t \mathbbm{1}[A_s=a]Y_s$, with the convention $\widehat{\mu}_{a,t}=0$ if $N_{a,t}=0$. Define ${\bm{\sigma}}^2 \coloneqq (\sigma_a^2)_{a\in[K]}$ and ${\bm{\sigma}} \coloneqq (\sigma_a)_{a\in[K]}$, where $\sigma_a\coloneqq \sqrt{\sigma_a^2}$. Similarly, define $\widehat{{\bm{\sigma}}}_{t}^2\coloneqq(\widehat{\sigma}_{a,t}^2)_{a\in[K]}$ and $\widehat{{\bm{\sigma}}}_{t}\coloneqq(\widehat{\sigma}_{a,t})_{a\in[K]}$ with $\widehat{\sigma}_{a,t}\coloneqq \sqrt{\widehat{\sigma}_{a,t}^2}$.

Main Results

This study develops an asymptotically minimax and Bayes optimal strategy for fixed-budget BAI. Our proposed TS-SPAS strategy uses two-stage sampling in the sampling phase, where we sample arms according to a sampling ratio given as a solution (saddle point) of a Gaussian minimax game. We then recommend an arm according to a decision rule, which is also given as a solution of a Gaussian minimax game.

In the theoretical analysis, we focus on minimax and Bayes regret as criteria for evaluating the optimality of the proposed strategy. In the minimax analysis, we evaluate the worst-case regret of our proposed strategy over a class of distributions, and in the Bayesian analysis, we evaluate the expected regret under a prior distribution on the parameters. By showing that the upper and lower bounds coincide exactly, including the constant terms, we establish exact asymptotic minimax and Bayes optimality of our strategy.

\paragraph{Summary of main theoretical results.} To briefly illustrate our contributions, we assume in this section that $Y_a$ follows a Gaussian distribution with mean $\mu_a$ and variance $\sigma^2_a$. Let ${\mathcal{B}}_{\bm{\sigma}^2}$ denote the set of distributions ${\bm{P}}_{{\bm{\mu}}} = (P_{a, \mu_a})_{a\in[K]}$ where the means vary while the variances are fixed at $(\sigma^2_a)_{a\in[K]}$. We refer to such a set of distributions as a bandit model. In Section (ref), we define more general bandit models, which include other distributions such as Bernoulli distributions.

We define the minimax and Bayes regret as follows:

itemize• Minimax regret: $\sup_{{\bm{\mu}} \in {\mathcal{M}}^K} {\mathrm{Regret}}^\delta_T({\bm{\mu}})$. • Bayes regret: $\int_{{\bm{\mu}} \in {\mathcal{M}}^K} {\mathrm{Regret}}^\delta_T({\bm{\mu}}){\mathrm{d}} H({\bm{\mu}})$, where $H({\bm{\mu}})$ denotes a prior distribution over ${\bm{\mu}}$.

Remarkably, we show that our proposed strategy is asymptotically optimal in both minimax and Bayes senses.

\paragraph{Asymptotic Minimax optimality.} We show that our proposed algorithm attains asymptotic minimax optimality as follows:

align*[align* omitted — 335 chars of source]

where $\mathfrak C^{\mathrm{mm}}$ is called the sharp minimax constant defined in Section (ref).

\paragraph{Asymptotic Bayes optimality.} We also show that our proposed algorithm attains asymptotic Bayes optimality as follows:

align*[align* omitted — 373 chars of source]

where $C^{\mathrm{Bayes}}$ is called the sharp Bayes constant defined in Section (ref).

\paragraph{Main message.} Two phenomena determine the sharp constants: (i) minimax regret is governed by near ties where the top means differ by $O(T^{-1/2})$, and (ii) Bayes regret under regular priors is governed by pairwise near ties between the best and second-best arms. These facts imply that a single two-stage design, which combines variance estimation, screening, and a saddle-point allocation, is optimal under both criteria with exact leading constants.

\paragraph{Contents of this paper.} The remainder of this paper is organized as follows. In Section (ref), we review related work. Section (ref) defines our strategy. Section (ref) introduces a class of distributions that we consider. Sections (ref) and (ref) present the lower and upper bounds, respectively, for minimax and Bayes regret. In Section (ref), we discuss related problems.

Literature Review

This section provides a brief literature review. For a more detailed review, see Appendix (ref). The earliest BAI formulation appeared under the name ordinal optimization Chen2000simulationbudget,Glynn2004largedeviations, focusing on non-adaptive optimal designs via large-deviation principles. That literature often assumes that an experimenter knows how to sample arms to attain optimality, which requires knowledge of the distributional information of the arms' outcomes. Beginning in the 2010s, BAI was formulated by explicitly addressing the estimation of the optimal sampling rule Audibert2010bestarm,Bubeck2011pureexploration.

BAI is typically studied under two settings: the fixed-confidence setting and the fixed-budget setting Bubeck2011pureexploration,Audibert2010bestarm,Kaufmann2016complexity. While optimal strategies have been studied well in the fixed-confidence setting, they are not fully explored in the fixed-budget setting due to various technical difficulties reported in existing studies Kaufmann2020contributions,Ariu2021policychoice,Degenne2023existence.

We show the optimality of strategies by establishing matching lower and upper bounds. In fixed-budget BAI, Garivier2016optimalbest conjectures an information-theoretic lower bound for the probability of misidentification ${\mathbb{P}}_{{\bm{\mu}}}\left(\widehat{a}^\delta_T \neq a^*_{{\bm{\mu}}}\right)$, but there is no strategy that matches the lower bound. In fact, Ariu2021policychoice and Degenne2023existence establish theorems showing the impossibility and the difficulty. Note that this difficulty is closely related to the information-theoretic or distribution-dependent analysis, where we evaluate the performance measure under each fixed distribution ${\mathbb{P}}_{{\bm{\mu}}}$. See also the reverse Kullback--Leibler divergence problem in Kaufmann2020contributions. Note that the evaluation of the probability of misidentification ${\mathbb{P}}_{{\bm{\mu}}}\left(\widehat{a}^\delta_T \neq a^*_{{\bm{\mu}}}\right)$ and the (simple) regret ${\mathrm{Regret}}^\delta_T({\bm{\mu}})$ are asymptotically equivalent under exponential convergence to zero. This is because the regret is given as ${\mathrm{Regret}}^\delta_T({\bm{\mu}}) = \sum_{a \in [K]} \Delta_{{\bm{\mu}}}(a) {\mathbb{P}}_{{\bm{\mu}}}\left( \widehat{a}^\delta_T = a \right)$, where the constant $\Delta_{{\bm{\mu}}}(a)$ is negligible compared to the exponentially converging term ${\mathbb{P}}_{{\bm{\mu}}}\left( \widehat{a}^\delta_T = a \right)$.

Subsequent works consider the possibility of avoiding this difficulty by considering minimax and Bayes analyses, where we evaluate the performance measure under the worst case of ${\mathbb{P}}_{{\bm{\mu}}}\left( \widehat{a}^\delta_T = a \right)$ or the average of ${\mathbb{P}}_{{\bm{\mu}}}\left( \widehat{a}^\delta_T = a \right)$ weighted by a prior Komiyama2022minimaxoptimal,Komiyama2023rateoptimal,Atsidakou2023bayesianfixed,Kato2024generalizedneyman,Nguyen2024priordependent. In minimax and Bayes analyses, the evaluation of the probability of misidentification and the regret differ.

This study develops BAI strategies with exactly matching lower and upper bounds. Our lower bounds rely on change-of-measure identities and information inequalities for events measurable with respect to the bandit filtration, as well as existing bandit studies with other formulations such as cumulative reward maximization Lai1985asymptoticallyefficient and fixed-confidence BAI Garivier2016optimalbest. We then establish tight upper bounds for our strategy. To the best of our knowledge, this is the first work to show optimal strategies in minimax and Bayes analyses with exactly matching lower and upper bounds.

Closer to our fixed-budget simple-regret objective, Komiyama2023rateoptimal studied Bayesian simple regret for Bernoulli bandits under regular priors and obtained an explicit $1/T$ upper bound with a conjectured sharp leading constant. In the Bernoulli case, our sharp Bayes constant coincides with the constant in Komiyama2023rateoptimal, and our matching lower bound verifies its optimality; more generally, we allow heteroscedastic outcomes in mean-parameterized exponential families.

We use change-of-measure arguments with local asymptotic normality (LAN) and Le Cam's third lemma LeCam1972theoryofstatisics,LeCam1986asymptoticmethods. There are existing studies applying LAN to bandit experiments Hirano2025asymptoticrepresentations,Adusumilli2022neymanallocation,Adusumilli2023risk,Adusumilli2025samplestopsampling,Armstrong2022asymptoticefficiency. However, those works first restrict all parameters to local alternatives and discuss lower and upper bounds simultaneously. These approaches do not yield globally optimal strategies and also do not effectively eliminate suboptimal arms. To eliminate suboptimal arms effectively, as this study shows, we need to use large deviations for non-local arms to eliminate them in early rounds and then apply LAN to the final remaining arms. In fact, those existing studies do not provide optimal strategies in the fixed-budget BAI setting, as they only provide lower bounds or show optimal strategies only in the two-armed case.

table[table omitted — 931 chars of source]

Bandit Models

From this section onward, we present the theoretical analysis of our TS-SPAS strategy, aiming to establish its minimax and Bayes optimality. We first define a class of distributions ${\mathcal{P}}$ for outcomes $Y$, over which we discuss the worst-case and average-case performance of strategies.

We consider canonical exponential families for this class, which are typically defined as follows Garivier2016optimalbest: \[ \left\{ (P_\theta)_{\theta \in \Theta} \colon \frac{dP_\theta}{d\xi}(y) = \exp\big( y \theta - b\big(\theta\big) \big)) }\right\}, \] where $P_\theta$ is a distribution parameterized by a natural parameter $\theta$, rather than $P_{\mu}$ used elsewhere, $\Theta \subset {\mathbb{R}}$ is the space of natural parameters $\theta$, $\xi$ is a reference measure on ${\mathcal{Y}}$, and $b \colon \Theta \to {\mathbb{R}}$ is a convex and twice differentiable function.

In this study, however, we consider worst-case and average-case performance over the mean parameter and characterize the lower and upper bounds in terms of variances, where the mean corresponds to $\dot{b}(\theta)$ and the variance corresponds to $\ddot{b}(\theta)$. Therefore, it is more convenient to define a class of distributions in terms of mean and variance parameters. This section provides such a definition and introduces a bandit model as a collection of $K$ distribution families.

Mean-Parameterized Canonical Exponential Families

We define the following mean-parameterized exponential families. Note that this class is essentially the same as the standard canonical exponential family, but we introduce it for the following purposes: (i) to define a distribution class parameterized by the mean, (ii) to ensure the correspondence between the inverse Fisher information and the variance, and (iii) to guarantee finite third moments, which are required for our analysis. Common distributions such as Gaussian and Bernoulli are included in this class.

definition[Mean-parameterized canonical exponential family] Fix a measurable outcome space $({\mathcal{Y}},\mathscr{A})$ and a compact interval ${\mathcal{M}}\subset\mathbb R$. For each arm $a\in[K]$, let $\xi_a$ be a $\sigma$-finite measure on $({\mathcal{Y}},\mathscr A)$ and let $\sigma_a^2\colon{\mathcal{M}}\to(0,\infty)$ be twice continuously differentiable. We say that $\{P_{a,\mu}\colon\mu\in{\mathcal{M}}\}$ belongs to ${\mathcal{P}}(\sigma_a^2,{\mathcal{M}},{\mathcal{Y}})$ if there exist an open interval $\Theta_a\subset\mathbb R$, a strictly convex $C^3$ log-partition function $b_a\colon\Theta_a\to\mathbb R$, and a $C^1$ map $\theta_a\colon{\mathcal{M}}\to\Theta_a$ such that for all $\mu\in{\mathcal{M}}$: \begin{enumerate} • $P_{a,\mu}\ll \xi_a$ with density $\frac{dP_{a,\mu}}{d\xi_a}(y)=\exp\{y\theta_a(\mu)-b_a(\theta_a(\mu))\}$; • $\dot b_a(\theta_a(\mu))=\mu$; • $\ddot b_a(\theta_a(\mu))=\sigma_a^2(\mu)$; • (Sub-Gaussian outcome) There exist constants $\lambda_{0,a}>0$ and $C_a<\infty$ such that \[ \sup_{\mu\in{\mathcal{M}}}{\mathbb{E}}_{a,\mu}\big[\exp\left(\lambda (Y-\mu)\right)\big]\le \exp(C_a\lambda^2) \qquad\forall |\lambda|\le \lambda_{0,a}. \] \end{enumerate}
remarkThe outcome space ${\mathcal{Y}}$ and the parameter space ${\mathcal{M}}$ should be chosen to satisfy the conditions in Definition (ref). For example, if the outcome $Y_a$ follows a Bernoulli distribution with support ${\mathcal{Y}} = \{0, 1\}$, we can choose ${\mathcal{M}}=[\underline\mu,\overline\mu]\subset(0,1)$ with $0<\underline\mu<\overline\mu<1$. If we instead choose ${\mathcal{M}} = [0, 1]$, the variance equals $0$ at $\mu = 0$ and $\mu = 1$, which violates our definition of the class. In this case, the Fisher information does not exist at $\mu = 0$ and $\mu = 1$, since the Fisher information is given by $I(\mu) = \frac{1}{\mu(1 - \mu)}$.

Bandit Model

For each arm $a\in[K]$, let $\sigma_a^2\colon{\mathcal{M}}\to(0,\infty)$ be continuous and bounded away from $0$ and $\infty$. Define the $K$-arm outcome model \[ {\mathcal{B}}_{\bm\sigma^2} \coloneqq \Big\{ (P_{a,\mu_a})_{a\in[K]}\colon \forall a, \{P_{a,\mu}\}_{\mu\in{\mathcal{M}}}\in{\mathcal{P}}(\sigma_a^2,{\mathcal{M}},{\mathcal{Y}})\Big\}, \qquad \bm\sigma^2=(\sigma_a^2)_{a\in[K]}. \]

example[Bandit instances] Our bandit class ${\mathcal{B}}_{\bm{\sigma}^2}$ allows heterogeneity across arms. For example, we can consider mixed families with $K=2$, e.g., arm $1$ is Bernoulli$(\mu_1)$ so $\sigma_1^2(\mu)=\mu(1-\mu)$, and arm $2$ is Gaussian with (possibly constant) variance $\sigma_2^2(\mu)\equiv \sigma_2^2$.

Two-Stage Saddle-Point Allocation with Screening

We propose an adaptive design that attains both the minimax and Bayes sharp constants. Our strategy consists of two phases, the sampling phase and the recommendation phase. In the sampling phase, we adopt a two-stage sampling rule. Before the experiment begins, we divide the total number of rounds $T$ into two stages, the first stage consists of $T_0$ rounds and the second stage consists of $T - T_0$ rounds, where $T_0$ is defined in Section (ref). In the first stage, we sample arms uniformly across all arms, assigning $\lfloor T_0/K\rfloor$ rounds to each arm (In the following arguments, for notational simplicity, we assume that $T_0/K$ is an integer). At the end of this stage, we apply screening and eliminate clearly suboptimal arms using a concentration inequality, resulting in a candidate set of potentially optimal arms. In the second stage, we sample the remaining arms according to a sampling ratio that is given as a solution of the Gaussian minimax game defined in Section (ref). Once all $T$ sampling rounds have been completed, we proceed to the recommendation phase. In the recommendation phase, we select an arm using a decision rule obtained from the Gaussian minimax game. Since the solution of the Gaussian minimax game is given as a saddle point, we refer to our strategy as the Two-Stage Saddle-Point Allocation with Screening (TS-SPAS) strategy and denote it by $\delta^{\text{TS}\mathchar`-\text{SPAS}}$. Algorithm (ref) presents the pseudocode.

Gaussian Game

Before describing our strategy in detail, we define the Gaussian game, which plays an important role in our strategy and the subsequent optimality arguments. Fix an integer $m\ge 2$. Let ${\mathcal{S}} \subseteq [K]$ be a nonempty subset of $[K]$. For the case where $|{\mathcal{S}}| = m$, let us define a vector of standard deviations as ${\bm{\sigma}} =(\sigma_a)_{a\in{\mathcal{S}}}\in(0,\infty)^m$. Let us define \[ {\mathcal{W}}^{{\mathcal{S}}} \coloneqq \Bigg\{{\bm{w}} \coloneqq (w_a)_{a\in{\mathcal{S}}}\in[0,1]^m\colon \sum_{a\in{\mathcal{S}}} w_a=1\Bigg\}, \qquad {\mathcal{W}}^{{\mathcal{S}}\circ} \coloneqq \Bigg\{{\bm{w}}\in{\mathcal{W}}^{{\mathcal{S}}}\colon \min_a w_a>0\Bigg\}. \] Given ${\bm{w}}\in{\mathcal{W}}^{{\mathcal{S}}\circ}$ and ${\mathcal{S}} \subseteq [K]$, define the diagonal covariance

align[align omitted — 198 chars of source]

In the Gaussian proxy experiment, Nature chooses an unknown mean vector ${\bm{h}} = (h_a)_{a\in{\mathcal{S}}}\in{\mathbb{R}}^m$ and we observe \[ {\bm{X}} = {\bm{h}}+{\bm{Z}},\qquad {\bm{Z}}\sim{\mathcal{N}}\left(0,\Sigma^{{\mathcal{S}}}\left({\bm{w}}, {\bm{\sigma}}\right)\right)a\right)}, \] which can also be written as $(X_a)_{a\in{\mathcal{S}}} = (h_a)_{a\in{\mathcal{S}}} + (Z_a)_{a\in{\mathcal{S}}}$, that is, indices in ${\bm{X}}$ and ${\bm{Z}}$ also depend on ${\mathcal{S}}$. Let $\zeta^{{\mathcal{S}}}({\bm{h}}, {\bm{w}}, {\bm{\sigma}})$ be the distribution of ${\bm{X}}$.

A (possibly randomized) decision rule is a Markov kernel $g(\cdot\mid {\bm{x}})$ on ${\mathcal{S}}$ given ${\bm{X}} = {\bm{x}} = (x_a)_{a\in{\mathcal{S}}}\in{\mathbb{R}}^m$, and the decision-maker outputs $A\sim g(\cdot\mid {\bm{X}})$. We evaluate decisions under the simple regret

align[align omitted — 116 chars of source]

We may denote $A$ by $A(g)$ to emphasize the dependency of $A$ on $g$.

definition[Gaussian minimax value] Let ${\mathcal{S}}$ be a subset of $[K]$ with $|{\mathcal{S}}| = m \geq 2$. For ${\bm{w}}\in{\mathcal{W}}^{{\mathcal{S}} \circ}$, we define the minimax risk of the Gaussian proxy experiment \[ V^{{\mathcal{S}}}\left({\bm{w}}, {\bm{\sigma}}\right)\coloneqq \inf_{g} \sup_{{\bm{h}}\in{\mathbb{R}}^m} {\mathbb{E}}_{\zeta^{{\mathcal{S}}}({\bm{h}}, {\bm{w}}, {\bm{\sigma}})}\left[\Delta^{{\mathcal{S}}}_{{\bm{h}}}(A(g))\right], \] where the infimum is over all Markov kernels $g$ from ${\mathbb{R}}^m$ to ${\mathcal{S}}$. The Gaussian minimax value is \begin{align} \mathfrak C^{{\mathcal{S}}}\left({\bm{\sigma}}\right)\coloneqq \inf_{{\bm{w}}\in{\mathcal{W}}^{{\mathcal{S}} \circ}} V^{{\mathcal{S}}}\left({\bm{w}}, {\bm{\sigma}}\right). \end{align} Equivalently, we may take the infimum over the closed simplex ${\bm{w}}\in{\mathcal{W}}^{{\mathcal{S}}}$ by adopting the convention $V^{{\mathcal{S}}}\left({\bm{w}}, {\bm{\sigma}}\right)=+\infty$ whenever $\min_a w_a=0$.

When convenient, for ${\mathcal{S}}\subseteq [K]$ such that $|{\mathcal{S}}| = m \geq 2$, we identify deterministic decision rules with measurable maps $g:{\mathbb{R}}^m\to{\mathcal{S}}$, and write $A=A(g)=g({\bm{X}})$.

remark[Translation and scale invariance] The loss ((ref)) is translation-invariant and positively homogeneous: for any $c\in{\mathbb{R}}$ and $\alpha>0$, \[ \Delta^{{\mathcal{S}}}_{{\bm{h}} + c\bm{1}}(A) = \Delta^{{\mathcal{S}}}_{{\bm{h}}}(A), \qquad \Delta^{{\mathcal{S}}}_{\alpha{\bm{h}}}(A)=\alpha \Delta^{{\mathcal{S}}}_{{\bm{h}}}(A). \] Consequently, $V^{{\mathcal{S}}}\left({\bm{w}}, {\bm{\sigma}}\right)$ and $\mathfrak C^{{\mathcal{S}}}\left({\bm{\sigma}}\right)$ inherit the same positive homogeneity in ${\bm{\sigma}}$: $\mathfrak C^{{\mathcal{S}}}\left(\alpha {\bm{\sigma}}\right)=\alpha\mathfrak C^{{\mathcal{S}}}\left({\bm{\sigma}}\right)$ (Lemma (ref) below).

Sampling Phase: The Two-Stage Rule

The sampling phase consists of two stages. For simplicity, we assume that $\lfloor T_0/K\rfloor \geq 2$. In the first stage, each arm is sampled an equal number of times, that is, $\lfloor T_0/K\rfloor$ rounds per arm. Based on the outcomes, we identify the empirical best arm and select candidate arms that are competitive with it. When the number of remaining arms is one, we immediately return that arm as the estimated best arm. If the number is greater than or equal to two, then in the second stage, we compute the sampling ratio by solving the Gaussian minimax game. Note that when the number of remaining arms is two, there exists a closed-form solution to the Gaussian minimax game such that we sample each arm proportionally to its standard deviation. Such a sampling rule is called Neyman allocation Neyman1934ontwo. We describe the strategy in detail below.

\paragraph{Sample size in the first stage.} Fix $\epsilon\in(0,1/3)$. We define the first-stage budget as $T_0 \coloneqq \lceil T^{1-\epsilon}\rceil$, where $T_0\to\infty$ and $T_0 / T\to0$ as $T\to\infty$.

First stage. We sample each arm $\lfloor T_0/K\rfloor$ times.

Suboptimal arm elimination after the first stage. For each arm $a \in [K]$, we construct the following lower and upper confidence bounds:

align*[align* omitted — 159 chars of source]

where \[ \overline C\coloneqq \max_{a\in[K]} C_a, \qquad C_{\mathrm{scr}}>6, \qquad \tau_{T_0} \coloneqq \sqrt{\frac{C_{\mathrm{scr}}\overline C K\log T}{T_0}}. \] Here we use the unbiased sample variance (division by $\lfloor T_0/K\rfloor-1$), but one can also use the biased variant dividing by $\lfloor T_0/K\rfloor$.

Using these bounds, we construct the set of candidate arms: \[ \widehat{{\mathcal{S}}}_{T_0} \coloneqq \left\{a \in [K] \colon \widehat{u}_{a, T_0} \geq \widehat{l}_{\widehat{a}_{T_0}, T_0}\right\}, \] where $\widehat{a}_{T_0} \coloneqq \operatorname*{arg\,max}_{a\in[K]} \widehat{\mu}_{a, T_0}$.

This stage serves two purposes. First, it gathers enough data to estimate the variances used in the second stage. Second, it eliminates clearly suboptimal arms early, allowing greater focus on distinguishing between the top-performing arms.

Second stage. We obtain an estimate of the sampling ratio. The sampling in the second stage depends on the cardinality of $\widehat{{\mathcal{S}}}_{T_0}$. Let $\widehat{\sigma}^2_{b, T_0}$ is the empirical variance estimator defined as

align*[align* omitted — 175 chars of source]

We also let $\widehat{\sigma}_{b,T_0}\coloneqq \sqrt{\widehat{\sigma}^2_{b,T_0}}$.

\paragraph{Case 1: $\left|\widehat{{\mathcal{S}}}_{T_0}\right| = 1$.} If $|\widehat{{\mathcal{S}}}_{T_0}|=1$, we set the recommendation to be the unique arm in $\widehat{{\mathcal{S}}}_{T_0}$. The remaining budget can be spent arbitrarily (e.g., repeatedly pulling this arm), without affecting the recommendation.

\paragraph{Case 2: $\left|\widehat{{\mathcal{S}}}_{T_0}\right| = 2$.} We sample arms in the second stage following the Neyman allocation, that is, in proportion to the estimated standard deviations $\widehat{\sigma}_{a,T_0}$ for $a \in \widehat{{\mathcal{S}}}_{T_0}$. Equivalently, we set $\widehat{w}_{a,T_0} \coloneqq \frac{\widehat{\sigma}_{a,T_0}}{\sum_{c \in \widehat{{\mathcal{S}}}_{T_0}} \widehat{\sigma}_{c,T_0}}$. Since we have already sampled a certain number of samples in the first stage, we take these samples into account and sample arms according to $(\widehat{\pi}_{a, T_0})_{a \in \widehat{{\mathcal{S}}}_{T_0}}$, which is defined as $\widehat{\pi}_{a, T_0} \coloneqq \frac{\widetilde{\pi}_{a, T_0}}{\sum_{a\in\widehat{{\mathcal{S}}}_{T_0}}\widetilde{\pi}_{a, T_0}}$, where $\widetilde{\pi}_{a, T_0} \coloneqq \max\left\{\frac{\widehat{w}_{a, T_0} - T_0/(KT)}{1-T_0/T},0\right\}$.

\paragraph{Case 3: $\left|\widehat{{\mathcal{S}}}_{T_0}\right| \geq 3$.} If $\left|\widehat{{\mathcal{S}}}_{T_0}\right| \geq 3$, we sample arms so that the empirical sampling ratio $\sum_{t=1}^T \mathbbm{1}[A_t = a] / T$ converges to the sampling ratio $w_a$ defined as

align*[align* omitted — 236 chars of source]

where ${\bm{\sigma}}_{\widehat{{\mathcal{S}}}_{T_0}}\coloneqq\bigl(\sigma_a(\mu_a)\bigr)_{a\in\widehat{{\mathcal{S}}}_{T_0}}$, and $\sigma_a^2(\cdot)$ denotes the variance function in the model, and the true variance equals $\sigma_a^2(\mu_a)$ under $P_{a,\mu_a}$. If $Y_a$ is generated from a distribution $P_{a, \mu_a}$ with parameter $\mu_a$, then $\sigma^2_a = \sigma^2_a(\mu_a)$. We refer to this sampling ratio as an ideal sampling ratio given $\widehat{{\mathcal{S}}}_{T_0}$. In many studies such as Garivier2016optimalbest, such an ideal sampling ratio is used as a target in the sampling phase of adaptive experimental design. One common ideal sampling ratio is the Neyman allocation, and our ideal sampling ratio can be regarded as a variant of it.

Since the variances are unknown, the ideal sampling ratio is also unknown. Therefore, by estimating the variances, we estimate the ideal sampling ratio $(\widehat{w}_{a, T_0})_{a \in \widehat{{\mathcal{S}}}_{T_0}}$ as

align*[align* omitted — 241 chars of source]

We then sample arms in the second stage by sampling from a multinomial distribution with probabilities $(\widehat{\pi}_{a, T_0})_{a \in \widehat{{\mathcal{S}}}_{T_0}}$, which is defined as $\widehat{\pi}_{a, T_0} \coloneqq \frac{\widetilde{\pi}_{a, T_0}}{\sum_{a\in\widehat{{\mathcal{S}}}_{T_0}}\widetilde{\pi}_{a, T_0}}$, where $\widetilde{\pi}_{a, T_0} \coloneqq \max\left\{\frac{\widehat{w}_{a, T_0} - T_0/(KT)}{1-T_0/T},0\right\}$.

remarkOur choice $T_0=\lceil T^{1-\epsilon}\rceil$ ensures $T_0\to\infty$ and $T_0/T\to0$, so Stage 1 is large enough for consistent variance estimation and screening, while its contribution to the $\sqrt{T}$- and $T$-scaled asymptotic regrets is negligible.
remarkSince $T_0/T \to 0$, we can ignore the first-stage samples $T_0$ asymptotically. Therefore, it is acceptable to sample directly according to $\widehat{w}_{a, T_0}$. However, we use $\widehat{\pi}$ to stabilize finite-sample performance by accounting for the first-stage samples.
remarkIn the estimation of the sampling ratio, we compute an (approximate) saddle point $(\widehat{{\bm{w}}}_{T_0},\widehat{g})$ for the Gaussian minimax value $\mathfrak C^{\widehat{{\mathcal{S}}}_{T_0}}\left(\widehat{{\bm{\sigma}}}_{T_0}\right)$ as in Definition (ref) (Section (ref)), e.g. with tolerance $\varepsilon_T=o(1)$: \[ \sup_{{\bm{h}}\in{\mathbb{R}}^m} {\mathbb{E}}_{\zeta^{\widehat{{\mathcal{S}}}_{T_0}}({\bm{h}}, \widehat{{\bm{w}}}_{T_0}, \widehat{{\bm{\sigma}}}_{T_0})} \left[\Delta^{\widehat{{\mathcal{S}}}_{T_0}}_{{\bm{h}}}\big(A(\widehat g)\big)\right] \le\mathfrak C^{\widehat{{\mathcal{S}}}_{T_0}}(\widehat{{\bm{\sigma}}}_{T_0})+\varepsilon_T, \] where ${\bm{X}}={\bm{h}}+{\bm{Z}}$ and ${\bm{Z}}\sim{\mathcal{N}}\left(0,\Sigma^{\widehat{{\mathcal{S}}}_{T_0}}(\widehat{{\bm{w}}}_{T_0},\widehat{{\bm{\sigma}}}_{T_0})\right)$.

Recommendation Phase

After the sampling phase, we recommend an arm using a decision rule $\widehat{g}$ obtained as a solution (saddle point) of the Gaussian minimax game.

When $|\widehat{{\mathcal{S}}}_{T_0}| = 2$, there exists a closed form solution for the decision rule, which is given as the empirical-best arm (EBA); that is, $\widetilde{g}(x)\in\operatorname*{arg\,max}_{a\in\{a,b\}} x_a$ for every ${\bm{w}}$. When $|\widehat{{\mathcal{S}}}_{T_0}|\ge 3$, there is no closed form except for some special cases such as equal variances. In this case, we recommend an arm following the decision rule $\widehat{g}$, a saddle point of the Gaussian minimax game, that is, \[ \widehat{a}^{\delta^{\text{TS}\mathchar`-\text{SPAS}}}_T \coloneqq \widehat{g}\left(\widehat {\bm{X}}_T\right), \] where $\widehat {\bm{X}}_T \coloneqq \sqrt{T}\left(\widehat\mu_{a, T}\right)_{a\in\widehat{{\mathcal{S}}}_{T_0}}\in{\mathbb{R}}^m$.

algorithm[algorithm omitted — 3,401 chars of source]
remark[Implementation of $\widehat{g}$] In theory, $\widehat{g}$ is the minimax (or Bayes) decision rule for the Gaussian proxy game. Section (ref) shows that minimax rules may be taken deterministic and as Bayes rules for least-favorable priors. In particular, if one computes a (possibly approximate) least-favorable prior $\widehat{\pi}^{\mathrm{LF}}$ (e.g. a finitely-supported prior) then $\widehat{g}$ can be implemented as a posterior-mean argmax: $\widehat{g}(x)\in\operatorname*{arg\,max}_{b}{\mathbb{E}}_{\widehat{\pi}^{\mathrm{LF}}}[h_b\mid X=x]$ (Proposition (ref)).
remark[Centering and translation equivariance] The loss in the Gaussian proxy game depends only on differences of coordinates. Lemma (ref) implies that minimax (and Bayes) rules may be taken translation-equivariant, so $\widehat{g}(x)=\widehat{g}(x+c\bm 1)$ for all $c\in{\mathbb{R}}$. Hence the input $\sqrt{T} \widehat\mu_{\widehat{{\mathcal{S}}}_{T_0},T}$ in the recommendation step may be replaced by any centered version, e.g. $\sqrt{T}\big(\widehat\mu_{\widehat{{\mathcal{S}}}_{T_0},T}-\overline{\widehat\mu}_{\widehat{{\mathcal{S}}}_{T_0},T}\bm 1\big)$.
remark[Tie-breaking and discrete outcomes] When outcomes are continuous, sample means are almost surely distinct and the tie-breaking rule is immaterial. With discrete outcomes, ties may occur but can be broken deterministically (e.g. by the smallest index); this does not affect any of our asymptotic constants because ties occur with probability $o(T^{-1/2})$ under our moment assumptions.

Asymptotic Minimax Optimality

In this section, we establish the minimax lower bound and the worst-case upper bound for the simple regret for the TS-SPAS strategy. The performance upper bound depends on the parameters of the distributions. By taking the worst case for the parameters, we develop the lower and upper bounds.

Sharp Minimax Constant

The sharp minimax constant is given as the Gaussian minimax value defined in Section (ref) with ${\mathcal{S}} = [K]$. We define the mean-dependent standard deviation as ${\bm{\sigma}}(\mu)=(\sigma_1(\mu),\dots,\sigma_K(\mu))\in(0,\infty)^K$. Recall that we have defined the open simplex as ${\mathcal{W}}_{[K]}^{\circ}\coloneqq\{{\bm{w}}\in(0,1)^K\colon\sum_{a=1}^K w_a=1\}$. Given ${\bm{w}}\in{\mathcal{W}}_{[K]}^{\circ}$ and ${\bm{\sigma}}(\mu)$, let us define a covariance matrix as \[ \Sigma^{[K]}\left({\bm{w}},{\bm{\sigma}}(\mu)\right)\coloneqq \text{diag}\left(\frac{\sigma_1(\mu)^2}{w_1},\dots,\frac{\sigma_K(\mu)^2}{w_K}\right). \] Let ${\bm{h}} = (h_a)_{a\in{\mathcal{S}}}\in{\mathbb{R}}^K$ and \[ {\bm{X}} = {\bm{h}}+{\bm{Z}},\qquad {\bm{Z}}\sim{\mathcal{N}}\left(0,\Sigma^{[K]}\left({\bm{w}}, {\bm{\sigma}}(\mu)\right)\right))\right)}, \] whose distribution is denoted by $\zeta^{[K]}({\bm{h}}, {\bm{w}}, {\bm{\sigma}}(\mu))$. Given ${\bm{X}}$, we sample an arm $A\sim g(\cdot\mid {\bm{X}})$, where $g(\cdot\mid {\bm{x}})$ is a Markov kernel on $[K]$. We define the sharp minimax constant as

align[align omitted — 139 chars of source]

where $\mathfrak C^{[K]}\left({\bm{\sigma}}\right)$ is the Gaussian minimax value from Definition (ref) (Section (ref)) evaluated at $m=K$ and ${\bm{\sigma}}={\bm{\sigma}}(\mu)$: \[ \mathfrak C^{[K]}\bigl({\bm{\sigma}}(\mu)\bigr) \coloneqq \inf_{{\bm{w}}\in{\mathcal{W}}^{[K]\circ}}\inf_{g}\sup_{{\bm{h}}\in{\mathbb{R}}^K}{\mathbb{E}}_{\zeta^{[K]}({\bm{h}}, {\bm{w}}, {\bm{\sigma}})}\left[\Delta^{[K]}_{{\bm{h}}}(A)\right], \] where $\Delta^{{\mathcal{S}}}_{{\bm{h}}}(A)\coloneqq \max_{a\in{\mathcal{S}}} h_a - h_{A}$, and $A\sim g(\cdot\mid X)$. In the minimax regime, least favorable instances are near-ties: $\mu_a-\mu_b=O(T^{-1/2})$ among competing treatments. Thus, competing treatments share a common limit $\mu$ as $T\to\infty$, and their variances evaluate at the same $\mu$. This is why the sharp constant depends on ${\bm{\sigma}}(\mu)$.

remark[Binary case and Neyman allocation] When $K=2$, the auxiliary value admits a closed form (see Appendix (ref), Corollary (ref)), implying \[ \mathfrak C^{\mathrm{mm}} = \sup_{\mu\in{\mathcal{M}}} \big(\sigma_1(\mu)+\sigma_2(\mu)\big) c_{\mathrm{mm}}, \qquad c_{\mathrm{mm}} \coloneqq \sup_{x\ge0} x\Phi(-x)\approx 0.1699712. \] Moreover, the minimax sampling proportion at tie value $\mu$ reduces to Neyman allocation $w_1^*(\mu)=\sigma_1(\mu)/(\sigma_1(\mu)+\sigma_2(\mu))$.

Asymptotic Minimax Lower Bound

We now present the minimax lower bound, which characterizes the best possible performance in the worst-case distribution.

theorem[Sharp minimax lower bound] Fix $K\ge2$, outcome space ${\mathcal{Y}}$, parameter space ${\mathcal{M}}$, and variance functions $\bm\sigma^2=(\sigma_a^2)_{a\in[K]}$. Assume $(P_{a,\mu})_{a\in[K]}\in{\mathcal{B}}_{\bm\sigma^2}$. Then, for the class ${\mathcal{A}}$ of all adaptive experiments, \[ \inf_{\delta\in{\mathcal{A}}} \liminf_{T\to\infty} \sqrt{T} \sup_{{\bm{\mu}}\in{\mathcal{M}}^K} {\mathrm{Regret}}_T^\delta({\bm{\mu}}) \ge \mathfrak C^{\mathrm{mm}}, \] where $\mathfrak C^{\mathrm{mm}}$ is defined in ((ref)).
theorem[Asymptotic worst-case upper bound for TS-SPAS] Under the assumptions of Theorem (ref), let $\delta^{\mathrm{TS}\mathchar`-\text{SPAS}}$ denote Algorithm (ref) with $(T_0,\tau_{T_0})$ defined in Section (ref). Then \[ \limsup_{T\to\infty} \sqrt{T} \sup_{{\bm{\mu}}\in{\mathcal{M}}^K} {\mathrm{Regret}}_T^{\delta^{\mathrm{TS}\mathchar`-\text{SPAS}}}({\bm{\mu}}) \le\ \mathfrak C^{\mathrm{mm}}. \]

Thus, we upper bounded the simple regret of the proposed strategy in Theorem (ref).

The results in the minimax lower bound (Theorem (ref)) and the worst-case upper bound (Theorem (ref)) imply the asymptotic minimax optimality.

Asymptotic Worst-Case Upper Bound

Then, we derive the following worst-case upper bound for the simple regret under the TS-SPAS strategy. The proof is shown in Appendix (ref).

corollary[Asymptotic minimax optimality] Under the conditions of Theorems (ref) and (ref), \[ \lim_{T\to\infty} \sqrt{T} \sup_{{\bm{\mu}}\in{\mathcal{M}}^K} {\mathrm{Regret}}_T^{\delta^{\mathrm{TS}\mathchar`-\text{SPAS}}}({\bm{\mu}}) = \mathfrak C^{\mathrm{mm}}. \]

Asymptotic Bayes Optimality

In this section, we establish the Bayes lower bound and the average upper bound for the simple regret for the TS-SPAS strategy. By taking the average of the regret bound weighted by the prior distribution, we develop the lower and upper bounds.

Sharp Bayes Constant

We now show the asymptotic Bayes optimality. Let $H$ be a prior distribution on ${\mathcal{M}}^K$. We assume the following regularity conditions for the prior distribution:

assumption[Regular prior: uniform continuity and boundedness of conditional densities] For each $a\in[K]$, let $h_a(\cdot\mid {\bm{\mu}}_{\setminus\{a\}})$ be a (regular) conditional density of $\mu_a$ given ${\bm{\mu}}_{\setminus\{a\}}$ w.r.t. Lebesgue measure on ${\mathcal{M}}$. For each distinct $a\neq b$, let $h_{ab}(\cdot,\cdot\mid {\bm{\mu}}_{\setminus\{a,b\}})$ be a conditional density of $(\mu_a,\mu_b)$ given ${\bm{\mu}}_{\setminus\{a,b\}}$ w.r.t. Lebesgue measure on ${\mathcal{M}}^2$. Assume the following uniform continuity: for every $\varepsilon>0$, there exists $\delta(\varepsilon)>0$ such that \begin{itemize} • for every $a\in[K]$, every ${\bm{\mu}}_{\setminus\{a\}}\in{\mathcal{M}}^{K-1}$ and every $\mu,\lambda\in{\mathcal{M}}$ with $|\mu-\lambda|\le\delta(\varepsilon)$, \[ \big|h_a(\mu\mid{\bm{\mu}}_{\setminus\{a\}})-h_a(\lambda\mid{\bm{\mu}}_{\setminus\{a\}})\big|\le\varepsilon; \] • for every $a\neq b$, every ${\bm{\mu}}_{\setminus\{a,b\}}\in{\mathcal{M}}^{K-2}$ and every $(\mu_a,\mu_b),(\lambda_a,\lambda_b)\in{\mathcal{M}}^2$ with $|\mu_a-\lambda_a|\le\delta(\varepsilon)$ and $|\mu_b-\lambda_b|\le\delta(\varepsilon)$, \[ \big|h_{ab}(\mu_a,\mu_b\mid{\bm{\mu}}_{\setminus\{a,b\}})-h_{ab}(\lambda_a,\lambda_b\mid{\bm{\mu}}_{\setminus\{a,b\}})\big|\le\varepsilon. \] \end{itemize} Assume additionally that these conditional densities are uniformly bounded: \[ \sup_{a, {\bm{\mu}}_{\setminus\{a\}}, \mu\in{\mathcal{M}}} h_a(\mu\mid{\bm{\mu}}_{\setminus\{a\}})<\infty, \qquad \sup_{a\neq b, {\bm{\mu}}_{\setminus\{a,b\}}, (\mu_a,\mu_b)\in{\mathcal{M}}^2} h_{ab}(\mu_a,\mu_b\mid{\bm{\mu}}_{\setminus\{a,b\}})<\infty. \]

This assumption follows from those in Theorem 1 of Lai1987adaptivetreatment and Assumption 1 of Komiyama2023rateoptimal.

For each $a\in[K]$, let ${\bm{\mu}}_{\setminus\{a\}} = (\mu_b)_{b\in[K]\setminus\{a\}} \in {\mathcal{M}}^{K-1}$. For each $a\in[K]$, define the (tie-broken) best-of-others index and its value \[ b^*_{\setminus \{a\}}\coloneqq \operatorname*{arg\,max}_{b\in[K]\setminus\{a\}} \mu_b, \qquad m_a({\bm{\mu}}_{\setminus \{a\}})\coloneqq \mu_{b^*_{\setminus \{a\}}}. \] Let $H^{\setminus \{a\}}$ denote the marginal of $H$ on ${\bm{\mu}}_{\setminus \{a\}}$. Let $h_a(\cdot\mid{\bm{\mu}}_{\setminus\{a\}})$ be a version of the conditional density from Assumption (ref), and let $H_a(\mu\mid{\bm{\mu}}_{\setminus\{a\}})\coloneqq {\mathbb{P}}_H(\mu_a\le \mu\mid{\bm{\mu}}_{\setminus\{a\}})$ denote the corresponding conditional cumulative density function.

We define the sharp Bayes constant as

align[align omitted — 399 chars of source]

Asymptotic Bayes Lower Bound

We now show a Bayes lower bound. Let $H$ be a prior distribution on ${\mathcal{M}}^K$.

theorem[Asymptotic Bayes lower bound] Assume the same outcome model as in Theorem (ref), and let the prior $H$ satisfy Assumption (ref). In particular, $H$ is absolutely continuous w.r.t. Lebesgue measure on ${\mathcal{M}}^K$ (so ${\mathbb{P}}_H(\mu_i=\mu_j)=0$ for all $i\neq j$). Then \[ \inf_{\delta\in{\mathcal{A}}} \liminf_{T\to\infty} T \int_{{\bm{\mu}} \in {\mathcal{M}}^K} {\mathrm{Regret}}^\delta_T({\bm{\mu}}){\mathrm{d}} H({\bm{\mu}})\ge C^{\mathrm{Bayes}}, \] where $C^{\mathrm{Bayes}}$ is given in ((ref)).

Asymptotic Average Upper Bound

We derive the following average upper bound under the TS-SPAS strategy. The proof is shown in Appendix (ref).

theorem[Asymptotic Average upper bound for TS-SPAS] Under the assumptions of Theorem (ref), let $\delta^{\mathrm{TS}\mathchar`-\text{SPAS}}$ denote Algorithm (ref) with $(T_0,\tau_{T_0})$ defined in Section (ref). Then \[ \limsup_{T\to\infty} T\int_{{\bm{\mu}} \in {\mathcal{M}}^K} {\mathrm{Regret}}^{\delta^{\mathrm{TS}\mathchar`-\text{SPAS}}}_T({\bm{\mu}}){\mathrm{d}} H({\bm{\mu}}) \le C^{\mathrm{Bayes}}. \]

The results in the Bayes lower bound (Theorem (ref)) and the average upper bound (Theorem (ref)) imply the asymptotic Bayes optimality.

corollary[Asymptotic Bayes optimality] Under the conditions of Theorems (ref) and (ref), \[ \lim_{T\to\infty}T\int_{{\bm{\mu}} \in {\mathcal{M}}^K} {\mathrm{Regret}}^{\delta^{\mathrm{TS}\mathchar`-\text{SPAS}}}_T({\bm{\mu}}){\mathrm{d}} H({\bm{\mu}}) = C^{\mathrm{Bayes}}. \]

Remarks on the Gaussian Minimax Game

This section records decision-theoretic facts about the Gaussian proxy game introduced in Section (ref). The proxy game determines the sharp constants in Theorems (ref)--(ref), and it also provides an implementable characterization of the Stage 2 objects $(\widehat{{\bm{w}}}_{T_0},\widehat g)$ used by TS-SPAS (Algorithm (ref)).

Throughout, we fix a nonempty ${\mathcal{S}}\subseteq[K]$ with $m\coloneqq|{\mathcal{S}}|\ge 2$. Recall from Section (ref) that, for ${\bm{w}}\in{\mathcal{W}}^{{\mathcal{S}}\circ}$ and ${\bm{\sigma}}\in(0,\infty)^m$, the Gaussian proxy experiment observes ${\bm{X}}={\bm{h}}+{\bm{Z}}$ with ${\bm{Z}}\sim{\mathcal{N}}(0,\Sigma^{{\mathcal{S}}}({\bm{w}},{\bm{\sigma}}))$ and loss $\Delta^{{\mathcal{S}}}_{{\bm{h}}}(A)=\max_{a\in{\mathcal{S}}}h_a-h_A$. The associated minimax risk is $V^{{\mathcal{S}}}({\bm{w}},{\bm{\sigma}})$ and the Gaussian minimax value is $\mathfrak C^{{\mathcal{S}}}({\bm{\sigma}})$ (Definition (ref)).

The main uses of this section are:

itemize[leftmargin=*] • Existence of Stage 2 saddle-point objects. Lemma (ref)(ii) guarantees existence of a minimizer $w^*$ of $w\mapsto V^{{\mathcal{S}}}({\bm{w}},{\bm{\sigma}})$ and an associated minimax rule $g^*$. This justifies defining Stage 2 by (approximately) computing a saddle point $(\widehat{{\bm{w}}}_{T_0},\widehat g)$ for $\mathfrak C^{\widehat{{\mathcal{S}}}_{T_0}}(\widehat{{\bm{\sigma}}}_{T_0})$ in Algorithm (ref) and in Section (ref). • How to compute $\widehat g$ in practice. Proposition (ref) and Lemmas (ref)--(ref) imply that a minimax rule can be implemented as a posterior-mean argmax under an (approximately) least-favorable prior. This is the formal content behind Remark (ref). • Centering of the recommendation statistic. Lemma (ref) implies that minimax (and Bayes) rules may be taken translation-equivariant, which justifies centering $\widehat {\bm{X}}_T$ in Remark (ref) and in the local Gaussian reductions (e.g. Appendix (ref), Step 6). • Plug-in of variance estimates. Lemma (ref)(iv) (continuity of ${\bm{\sigma}}\mapsto\mathfrak C^{{\mathcal{S}}}({\bm{\sigma}})$) is used when replacing the unknown ${\bm{\sigma}}$ by the Stage 1 estimator $\widehat{{\bm{\sigma}}}_{T_0}$ in Stage 2, as in the proofs of Theorems (ref) and (ref). • Binary special case. Lemma (ref) explains why, when $|{\mathcal{S}}|=2$, Stage 2 reduces to Neyman allocation and the recommendation reduces to the empirical-best arm, as used explicitly in Section (ref) (Case 2) and in the recommendation phase of TS-SPAS. • \textbf{Numerical approximation of the saddle point for $m\ge 3$.} Subsection (ref) and Algorithm (ref) provide a concrete numerical routine for approximating a saddle point $(w^*,g^*)$ when no closed form is available. The procedure combines (i) a finite-support approximation of least-favorable priors (Lemma (ref)), (ii) implementation of the resulting Bayes rule via posterior-mean argmax (Proposition (ref)), and (iii) an outer search over allocations ${\bm{w}}$. This is the practical mechanism used to compute $(\widehat{{\bm{w}}}_{T_0},\widehat g)$ in Stage 2 of TS-SPAS when $|\widehat{{\mathcal{S}}}_{T_0}|\ge 3$.

Existence and Continuity of the Gaussian Minimax Value Function

The next results provide regularity properties of the Gaussian minimax value. They are used when plugging in $\widehat{{\bm{\sigma}}}_{T_0}$ in Stage 2 and when establishing existence of an interior minimizer $w^*$.

lemma[Pairwise lower bound] For any ${\bm{w}}\in{\mathcal{W}}^{{\mathcal{S}} \circ}$ and ${\bm{\sigma}}\in(0,\infty)^m$, \[ V^{{\mathcal{S}}}({\bm{w}},{\bm{\sigma}}) \ge c_{\mathrm{mm}}\max_{a\neq b}\sqrt{\frac{\sigma_a^2}{w_a}+\frac{\sigma_b^2}{w_b}}, \qquad c_{\mathrm{mm}}\coloneqq \sup_{x\ge 0} x\Phi(-x)\approx 0.1699712. \]
proofFix $a\neq b$ and consider a restricted two-point subproblem in which Nature selects two mean vectors that differ only on coordinates $a$ and $b$ (all other coordinates are made strictly dominated by shifting them far below). This reduction yields a binary Gaussian testing problem based on $X_a-X_b\sim{\mathcal{N}}(\pm\Delta, v)$ with $v=\sigma_a^2/w_a+\sigma_b^2/w_b$. The minimax risk of this binary subproblem equals $\sup_{\Delta\ge0}\Delta\Phi(-\Delta/\sqrt v)=c_{\mathrm{mm}}\sqrt v$. Taking the maximum over pairs gives the bound.
lemma[Existence, interiority, homogeneity, and continuity] Fix $m\ge 2$ and a set ${\mathcal{S}}$ with $|{\mathcal{S}}|=m$. \begin{enumerate} • (Finiteness) For every ${\bm{\sigma}}\in(0,\infty)^m$, the value $\mathfrak C^{{\mathcal{S}}}({\bm{\sigma}})$ is finite. • (Existence and interiority) For every ${\bm{\sigma}}\in(0,\infty)^m$, there exists $w^*\in{\mathcal{W}}^{{\mathcal{S}} \circ}$ and a decision rule $g^*$ such that \[ \mathfrak C^{{\mathcal{S}}}({\bm{\sigma}}) = \sup_{{\bm{h}}\in{\mathbb{R}}^m}{\mathbb{E}}_{\zeta^{{\mathcal{S}}}({\bm{h}},{\bm{w}}^*,{\bm{\sigma}})}\left[\Delta^{{\mathcal{S}}}_{{\bm{h}}}(A(g^*))\right] = V^{{\mathcal{S}}}({\bm{w}}^*,{\bm{\sigma}}), \] and ${\bm{w}}^*$ minimizes ${\bm{w}}\mapsto V^{{\mathcal{S}}}({\bm{w}},{\bm{\sigma}})$ over ${\mathcal{W}}^{{\mathcal{S}}\circ}$. • (Positive homogeneity) For every $\alpha>0$, $\mathfrak C^{{\mathcal{S}}}(\alpha{\bm{\sigma}})=\alpha \mathfrak C^{{\mathcal{S}}}({\bm{\sigma}})$. • (Continuity) The map ${\bm{\sigma}}\mapsto \mathfrak C^{{\mathcal{S}}}({\bm{\sigma}})$ is continuous on $(0,\infty)^m$. In particular, if $\widehat{{\bm{\sigma}}}_{T_0}\to {\bm{\sigma}}$, then $\mathfrak C^{{\mathcal{S}}}(\widehat{{\bm{\sigma}}}_{T_0})\to \mathfrak C^{{\mathcal{S}}}({\bm{\sigma}})$. \end{enumerate}

Bayes Rules, Least-Favorable Priors, and Deterministic Minimax Rules

\paragraph{Deterministic rules and tie-breaking.} We call a decision rule $g$ deterministic if there exists a Borel measurable map $a\colon{\mathbb{R}}^m\to{\mathcal{S}}$ such that $g(b\mid x)=\mathbbm{1}[a(x)=b]$. Since ${\bm{X}}\sim{\mathcal{N}}({\bm{h}},\Sigma^{{\mathcal{S}}}({\bm{w}},{\bm{\sigma}}))$ has a continuous density for every fixed ${\bm{h}}$, ties occur with probability zero; hence any deterministic tie-breaking convention is immaterial for the risk. (We still fix a deterministic tie-break to make argmax selectors measurable.)

definition[Bayes risk and least-favorable priors] Fix ${\bm{w}}\in{\mathcal{W}}^{{\mathcal{S}} \circ}$ and ${\bm{\sigma}}\in(0,\infty)^m$. For a prior $\varrho$ on ${\mathbb{R}}^m$, define the Bayes risk \[ {\mathcal{R}}^{{\mathcal{S}}}({\bm{w}},{\bm{\sigma}};\varrho) \coloneqq \inf_{g}\int {\mathbb{E}}_{\zeta^{{\mathcal{S}}}({\bm{h}}, {\bm{w}}, {\bm{\sigma}})}\left[\Delta^{{\mathcal{S}}}_{{\bm{h}}}(A)\right]\varrho(dh). \] A prior $\varrho^*$ is least-favorable for $({\bm{w}},{\bm{\sigma}})$ if it maximizes the Bayes risk: $\varrho^*\in\operatorname*{arg\,max}_{\varrho}{\mathcal{R}}^{{\mathcal{S}}}({\bm{w}},{\bm{\sigma}};\varrho)$.
proposition[Bayes rule is posterior-mean argmax] Fix ${\bm{w}}\in{\mathcal{W}}^{{\mathcal{S}} \circ}$, ${\bm{\sigma}}\in(0,\infty)^m$, and a prior $\varrho$ on ${\bm{h}}\in{\mathbb{R}}^m$. Let $\varrho( \cdot\mid {\bm{X}}=x)$ denote the posterior distribution of ${\bm{h}}$ given ${\bm{X}}=x$. Then any Bayes rule for the loss ((ref)) can be taken deterministic and satisfies \[ g_\varrho(x)\in\operatorname*{arg\,max}_{b\in{\mathcal{S}}} {\mathbb{E}}_\varrho\left[h_b\mid {\bm{X}}=x\right]. \] In particular, for a finitely supported prior $\varrho=\sum_{\ell=1}^L p_\ell \delta_{h^{(\ell)}}$, the posterior weights satisfy \[ \varrho\bigl(h^{(\ell)}\mid {\bm{X}}=x\bigr) \propto p_\ell\varphi_{\Sigma^{{\mathcal{S}}}({\bm{w}},{\bm{\sigma}})}\left(x-h^{(\ell)}\right), \] where $\varphi_{\Sigma}$ is the density of ${\mathcal{N}}(0,\Sigma)$, and hence ${\mathbb{E}}_\varrho[{\bm{h}}\mid {\bm{X}}=x]=\sum_{\ell=1}^L \varrho(h^{(\ell)}\mid {\bm{X}}=x)h^{(\ell)}$.
proofConditioning on ${\bm{X}}=x$, the Bayes conditional risk of a (possibly randomized) decision rule $g$ is \[ {\mathbb{E}}_\varrho\left[\Delta^{{\mathcal{S}}}_{{\bm{h}}}(A)\mid {\bm{X}}=x\right] = {\mathbb{E}}_\varrho\left[\max_{i\in{\mathcal{S}}} h_i\mid {\bm{X}}=x\right] - \sum_{b\in{\mathcal{S}}} {\mathbb{E}}_\varrho[h_b\mid {\bm{X}}=x] g(b\mid x). \] The first term does not depend on the chosen arm. Thus, minimizing the conditional risk over the simplex $\{g(\cdot\mid x)\}$ is equivalent to maximizing $\sum_{b\in{\mathcal{S}}}{\mathbb{E}}_\varrho[h_b\mid {\bm{X}}=x]g(b\mid x)$, which is achieved by a deterministic choice of an index maximizing the posterior mean.
lemma[Least-favorable prior and minimax rule for fixed $({\bm{w}},{\bm{\sigma}})$] Fix ${\bm{w}}\in{\mathcal{W}}^{{\mathcal{S}} \circ}$ and ${\bm{\sigma}}\in(0,\infty)^m$ and assume $V^{{\mathcal{S}}}({\bm{w}},{\bm{\sigma}})<\infty$. Then there exists a least-favorable prior $\varrho^*$ (Definition (ref)) such that \[ V^{{\mathcal{S}}}({\bm{w}},{\bm{\sigma}}) = \inf_{g} \sup_{{\bm{h}}\in{\mathbb{R}}^m} {\mathbb{E}}_{\zeta^{{\mathcal{S}}}({\bm{h}}, {\bm{w}}, {\bm{\sigma}})}\left[\Delta^{{\mathcal{S}}}_{{\bm{h}}}(A)\right] = \sup_{\varrho} {\mathcal{R}}^{{\mathcal{S}}}({\bm{w}},{\bm{\sigma}};\varrho) = {\mathcal{R}}^{{\mathcal{S}}}({\bm{w}},{\bm{\sigma}};\varrho^*). \] Moreover, any Bayes rule $g^*$ for $\varrho^*$ is minimax for $({\bm{w}},{\bm{\sigma}})$ and may be taken deterministic. In particular, $g^*$ can be chosen to satisfy $g^*(x)\in\operatorname*{arg\,max}_{b\in{\mathcal{S}}}{\mathbb{E}}_{\varrho^*}[h_b\mid {\bm{X}}=x]$.
proofThis is a standard consequence of minimax theorems in statistical decision theory for dominated experiments with finite action spaces after convexifying the decision-maker's strategies by allowing randomization. One reference is VanderVaart1998asymptoticstatistics; see also Lecam2000asymptoticsin. Proposition (ref) yields existence of a deterministic Bayes rule and its posterior-mean characterization.
lemma[Finite-support $\varepsilon$-least-favorable priors] Fix ${\bm{w}}\in{\mathcal{W}}^{{\mathcal{S}} \circ}$, ${\bm{\sigma}}\in(0,\infty)^m$, and $\varepsilon>0$. There exists a finitely supported prior $\varrho_\varepsilon=\sum_{\ell=1}^L p_\ell \delta_{h^{(\ell)}}$ such that \[ {\mathcal{R}}^{{\mathcal{S}}}({\bm{w}},{\bm{\sigma}};\varrho_\varepsilon)\ge V^{{\mathcal{S}}}({\bm{w}},{\bm{\sigma}})-\varepsilon. \] Consequently, one may compute an $\varepsilon$-minimax decision rule by optimizing over discrete priors and then applying the posterior-mean argmax rule of Proposition (ref).
proofLet $\varrho^*$ be a least-favorable prior from Lemma (ref). After truncating $\varrho^*$ to a large compact set and renormalizing, approximate it by a discrete prior (using density of finitely supported measures on compact metric spaces under weak convergence). For the Gaussian proxy experiment with fixed $({\bm{w}},{\bm{\sigma}})$, the Bayes risk depends continuously on $\varrho$ under weak convergence on compact supports (dominated convergence). Choosing the approximation fine enough yields the claim.
remark[Markov kernels] We allow Markov kernels in Definition (ref) because the class of kernels is convex, which streamlines minimax-interchange and saddle-point existence arguments. At the minimax solution one may work with deterministic rules (Proposition (ref) and Lemma (ref)).
remark[Closed form vs. numerical computation of $g^*$] When $m=2$, the minimax decision rule is the empirical-best arm and the optimal allocation is Neyman allocation (Lemma (ref)). For $m\ge 3$, we do not know a general closed form for the minimax decision boundary, so Stage 2 computes an approximate saddle point numerically (Algorithm (ref) and Remark (ref)).

Translation Equivariance

Recall from Remark (ref) that the loss depends only on differences of coordinates. The following lemma formalizes the corresponding invariance reduction and is used to justify centering.

lemma[Translation equivariance] Fix ${\bm{w}}\in{\mathcal{W}}^{{\mathcal{S}} \circ}$ and ${\bm{\sigma}}\in(0,\infty)^m$. There exists a minimax decision rule $g^*$ attaining $V^{{\mathcal{S}}}({\bm{w}},{\bm{\sigma}})$ that is translation-equivariant: \[ g^*(x+c\bm 1)=g^*(x)\qquad\text{for all }x\in{\mathbb{R}}^m, c\in{\mathbb{R}}. \]
proofConsider the additive group $G=\{c\bm 1:c\in{\mathbb{R}}\}$ acting on both parameters and observations by translation. The Gaussian experiment is $G$-invariant and the loss ((ref)) is $G$-invariant. By standard results on invariant decision rules (e.g. Hunt--Stein type arguments for amenable groups), there exists a minimax rule that is $G$-invariant, i.e., translation-equivariant.

Closed Form in the Binary Case

For later reference (and for sanity-checking numerical solvers), the binary case admits a closed form.

lemma[Binary closed form] Let ${\mathcal{S}} = \{a, b\} \subset [K]$ such that $a\neq b$ and ${\bm{\sigma}}=(\sigma_a,\sigma_b)\in(0,\infty)^2$. Then \[ \mathfrak C^{{\mathcal{S}}}({\bm{\sigma}}) = \inf_{{\bm{w}}\in{\mathcal{W}}^{{\mathcal{S}} \circ}} \inf_{g} \sup_{{\bm{h}}\in{\mathbb{R}}^2} {\mathbb{E}}_{\zeta^{{\mathcal{S}}}({\bm{h}}, {\bm{w}}, {\bm{\sigma}})}[\Delta^{{\mathcal{S}}}_{{\bm{h}}}(A)] = \left(\sigma_a+\sigma_b\right)c_{\mathrm{mm}}, \] where $c_{\mathrm{mm}}\coloneqq \sup_{x\ge 0} x\Phi(-x)\approx 0.1699712$. Moreover, the optimizer is the Neyman allocation $w_a^*=\sigma_a/(\sigma_a+\sigma_b)$ and $w_b^*=\sigma_b/(\sigma_a+\sigma_b)$, and the minimax decision rule is the empirical-best arm.

Numerical Solution for the Gaussian Saddle Point Game

Stage 2 of TS-SPAS requires computing an (approximate) saddle point \( (\widehat{{\bm{w}}}_{T_0},\widehat g) \) for the Gaussian minimax value \( \mathfrak C^{{\mathcal{S}}}({\bm{\sigma}}) \) (Definition (ref)) on the candidate set ${\mathcal{S}}=\widehat{{\mathcal{S}}}_{T_0}$ with $m\coloneqq|{\mathcal{S}}|\ge 3$. When $m=2$ there is a closed form (Lemma (ref)), but for $m\ge 3$ we solve the saddle-point problem numerically. In Appendix (ref), we describe a practical procedure that is consistent with the decision-theoretic characterization in Section (ref), in particular Proposition (ref) and Lemma (ref) (finite-support approximation of least-favorable priors).

Discussion

Minimax Optimal Strategy

If we focus solely on minimax optimality, we do not need to eliminate suboptimal arms in the first stage. This is because the worst case is governed by instances where all arms have gaps of order $O(1/\sqrt{T})$. After the first stage, even without eliminating suboptimal arms, we can sample each arm according to the solution of the Gaussian minimax game and recommend an arm according to the same solution. Such a strategy also guarantees minimax optimality.

Bayes Optimal Strategy

If our interest lies solely in a Bayes optimal strategy, we can omit solving the Gaussian minimax game. In the second stage, if $|\widehat{{\mathcal{S}}}_{T_0}| = 2$, we apply the Neyman allocation; if $|\widehat{{\mathcal{S}}}_{T_0}| \geq 3$, we draw the other arms uniformly. Since, in the Bayes optimal case, only two arms dominate the regret, this simplification still works.

LAN Approaches in Existing Studies

If we apply LAN to all arms, we cannot obtain an optimal strategy in a global sense. For example, Adusumilli2022neymanallocation,Adusumilli2023risk adopt such an approach, but these works do not provide an optimal strategy. This is because screening suboptimal arms using large-deviation arguments remains important, even when we apply LAN.

Change of Measure in Bandit Problems and Le Cam's Third Lemma

The proofs combine two complementary viewpoints. Bandit lower bounds often start from a change-of-measure identity Lai1985asymptoticallyefficient,Kaufmann2016complexity. Asymptotic statistics instead emphasizes LAN expansions and Le Cam's third lemma. These approaches are actually based on the same change-of-measure arguments. Therefore, our arguments closely follow the existing literature on bandits and asymptotic statistics.

Bernoulli Reduction

When outcomes follow a Bernoulli distribution, our strategy can be simplified by omitting the variance estimation step. We describe this simplification in this section.

As discussed above, in both minimax and Bayesian analyses, regret is primarily influenced by instances in which the gap between the best and suboptimal arms shrinks at the rate $1/\sqrt{T}$. Formally, as $T \to \infty$, we have $\mu_{a^*_{{\bm{\mu}}}} - \mu_b \to 0$. Recall that for Bernoulli outcomes, the variance of arm $a$ is $\sigma^2_a(\mu_a) = \mu_a(1 - \mu_a)$. As the mean differences converge to zero, the variances of the best arm and its competitors also converge, $\sigma_{a^*_{{\bm{\mu}}}}(\mu_{a^*_{{\bm{\mu}}}}) - \sigma_b(\mu_b) \to 0$. When these variances become asymptotically equivalent, variance-based sampling in the second stage of the sampling phase is unnecessary. Instead, we sample arms in the candidate set $\widehat{{\mathcal{S}}}_{T_0}$ uniformly, that is, with probability $1 / |\widehat{{\mathcal{S}}}_{T_0}|$ for each arm. We then obtain the following result.

proposition[Bernoulli reduction] Suppose $P_{a,\mu}=\mathrm{Bernoulli}(\mu)$ for all $a\in[K]$ so that ${\mathcal{M}}=[\underline\mu,\overline\mu]\subset(0,1)$ with $0<\underline\mu<\overline\mu<1$, and $\sigma_a^2(\mu)=\mu(1-\mu)$. Let the prior $H$ satisfy Assumption (ref) with density $h$. Write $\mu_{*\setminus \{a\}}\coloneqq \max_{b\neq a}\mu_b$ and let $h_a(\cdot\mid{\bm{\mu}}_{\setminus \{a\}})$ be a regular conditional density of $\mu_a$ given ${\bm{\mu}}_{\setminus \{a\}}$. Then the sharp Bayes constant $C^{\mathrm{Bayes}}$ in ((ref)) equals \[ C_{\mathrm{opt}} \coloneqq \sum_{a\in[K]} \int_{[0,1]^{K-1}} \mu_{*\setminus \{a\}}\bigl(1-\mu_{*\setminus \{a\}}\bigr) h_a\bigl(\mu_{*\setminus \{a\}}\mid{\bm{\mu}}_{\setminus \{a\}}\bigr){\mathrm{d}} H^{\setminus \{a\}}({\bm{\mu}}_{\setminus \{a\}}), \] which is exactly the constant from Komiyama2023rateoptimal.

Conclusion

In this study, for fixed-budget BAI, we proposed the TS-SPAS strategy, which eliminates clearly suboptimal arms early and samples the remaining arms to distinguish the best arm from the others. In our theoretical analysis, we derived minimax and Bayes lower bounds for the simple regret, establishing fundamental performance limits for any regular strategy. We further proved that the simple regret of the proposed strategy matches these lower bounds, including the constant term, not only the rate.