Introduction
We develop a common planning framework for episodic bandits over unknown recurring priors, incorporating delayed observations, hard per-arm budgets and per-episode caps, and risk-sensitive objectives beyond expected reward. In an episodic bandit or a meta-bandit problem there are n n n unknown arms and a learner uses these arms over N N N episodes, each with at most T T T decision steps. Each episode draws fresh arm qualities from the recurring, initially unknown prior distribution. A single episode is a finite-horizon multi-armed bandit problem. At each step, the learner plays an arm, earns reward and generates an observation. The observation provides information about the current episode’s arm quality and, across episodes, its recurring prior. Therefore the typical multi-armed bandit tradeoff of explore/use now expands to explore, use, and generalize . Meta-bandits formalize this transfer/generalization by learning shared priors [7 , 19 , 25 ] using variants of Thompson sampling. Other examples include cold-start dynamic pricing [6 ] , cloud management [31 ] , database tuning across hardware environments [43 ] , program synthesis [13 ] , and LLM routing [41 ] . The analyzed MetaTS policy supplements Thompson sampling with a forced play of every arm in each task [25 Section 4.1] . HierTS avoids forced exploration in its sequential Gaussian analysis, but its concurrent guarantee again prescribes initial exploratory actions [19 Sections 6.1 and 6.3] . These methods share information through a hierarchical posterior, while their guarantees establish the learning produced by specified sampling and exploration rules. They do not explicitly optimize which observations are worth acquiring through their contribution to later feasible decisions. Even with independent arms, optimizing each episode separately can lose linearly in N N N by never collecting observations valuable to later episodes (Examples A.1 and A.2).
We formulate a common planning objective that accounts for current reward, resource consumption and the continuation of learning. Its execution guarantee then accounts for the information and reward actually realized under the hard constraints.
Budgets are natural in distributed applications of episodic bandits. Budgeted learning is NP-hard even for a single episode with independent arms [16 ] . An unexpectedly long episode, or a favorable draw of an arm’s episode quality, can make immediate use attractive enough to exhaust the arm’s remaining budget. Those plays then become unavailable for learning and reward in later episodes. Per-episode caps limit this drawdown. Routing, coding, tuning and pricing also motivate risk-sensitive objectives beyond expected reward. Risk-sensitive objectives introduce a further structural difficulty. Mean–variance criteria need not satisfy Bellman’s principle of optimality, and maximizing expected return subject to a variance bound is NP-hard in finite-horizon Markov decision processes [27 ] . The objective depends on the return distribution across possible histories. Evaluating risk conditional on a sampled episode quality can discard variation that determines which action is preferable (Example A.6). Planning over outcome distributions retains this variation: fixed outcome scores enter as linear coefficients, while risk functionals act on the resulting law (Subsection 3.5).
Delays couple decisions within and across episodes. Concurrent Bayesian meta-bandits [19 ] analyze feedback withheld within a batch of decisions, and distributed optimization also motivates delayed feedback [2 ] . Delayed observation in learning dates back to Anderson [3 ] . Surprisingly, it was shown by Eick [12 ] that in the presence of delays an optimal policy can use a known arm and later return to an unknown arm. Thus the optimal policy of a delayed finite horizon multi-armed bandit problem is not a familiar stopping time problem [14 ] , creating a foundational divergence from standard multi-armed bandits. A late-arriving observation changes an arm’s information even when the arm is not being played. This violates the standard rested-bandit assumption, and such arms are restless [40 ] . Caro and Yoo [9 ] study no-crossover delays. General restless bandits have no known tractable algorithms, and their decision versions are PSPACE-hard [30 ] . Here the delayed observations result from our own earlier plays and arrive on a calendar fixed in advance. We call this planned restlessness .
Budgets constrain resource use on every realized path; delays constrain the information available at each decision; risk criteria specify how the reward distribution is evaluated. Under this common model, we distinguish the losses due to unavailable information from the loss incurred by the algorithm. Let V e V_e V e be a policy’s reward in episode e e e and V A = E ∑ e V e V_A = \mathbb{E}\sum_e V_e V A = E ∑ e V e its expected total. Write V s t a t e V_{\mathrm{state}} V state for the optimum when the recurring prior is revealed and each episode’s qualities are revealed at its start, V p r i o r V_{\mathrm{prior}} V prior for the optimum when only the recurring prior is revealed, and V l e a r n V_{\mathrm{learn}} V learn for the Bayes-optimal learner with our information. Then the overall loss decomposes as:
V s t a t e − V A = ( V s t a t e − V p r i o r ) + ( V p r i o r − V l e a r n ) + ( V l e a r n − V A ) . (1) V_{\mathrm{state}} - V_A = (V_{\mathrm{state}} - V_{\mathrm{prior}}) + (V_{\mathrm{prior}} - V_{\mathrm{learn}}) + (V_{\mathrm{learn}} - V_A). \tag*{(1)} V state − V A = ( V state − V prior ) + ( V prior − V learn ) + ( V learn − V A ) . (1)
The gap V s t a t e − V p r i o r V_{\mathrm{state}} - V_{\mathrm{prior}} V state − V prior is the cost of uncertainty about each episode’s fresh qualities. It can remain linear in N N N even when the recurring law is known (Example A.7). The gap V p r i o r − V l e a r n V_{\mathrm{prior}} - V_{\mathrm{learn}} V prior − V learn is the price of learning that law, and the largest fee worth paying to have it revealed before learning begins (Section 5.2). Under stationarity, independent arms, finite reward alphabets, bounded report lag and nonbinding budgets, this price is O ( N ) O(\sqrt{N}) O ( N ) , and hence vanishes per episode (Proposition 5.4). The gap V l e a r n − V A V_{\mathrm{learn}} - V_A V learn − V A is the algorithm’s loss under the same information and constraints. It measures how well the algorithm uses its information; Section 3 controls this loss for the learning-LP learner. Information is valuable only while the decisions it can improve remain available, and it is necessary that all three benchmarks face the same budgets, caps and report calendar.
We formulate a linear programming (LP) relaxation over learning trajectories. We assume independence across arms in this paper which permits separate arm programs coupled by expected episode capacities. The LP is written ex ante , over all possible observation records. Given the incoming belief, exchangeability and the fixed report calendar determine how each stopping record leads to the arm’s next learning state: its belief about the recurring prior, remaining budget and pending reports. Histories with the same sufficient state share these transitions and can be represented by one node. The transitions are specified before any particular outcome is observed; the policy determines the probability mass flowing through them. Current reward, resource use and continuation can therefore be accounted for linearly. Pricing the shared episode capacities separates the arms into local programs on their state DAGs (Figure 1 ). Their values are computed by backward induction, but their policies must be executed forward, echoing Kierkegaard’s distinction between understanding backward and living forward.2
Figure 1. One arm across an episode boundary. Shown are selected stopping nodes u u u of the current local DAG and starting nodes v v v of the next. The handoff sends stopping mass z ( u ) z(u) z ( u ) through the fixed probabilities H e ( S , u , v ) H_e(S,u,v) H e ( S , u , v ) , giving arrival masses q ( v ) q(v) q ( v ) . The continuation is ∑ v q ( v ) F ( v ) \sum_v q(v)F(v) ∑ v q ( v ) F ( v ) . Each next starting node carries its posterior, remaining budget and pending reports.
Execution proceeds ex post . The learner is an observer within the system it controls: its belief about the arms is itself part of the state being planned. Each play earns reward, spends budget and creates a report on the known calendar. The policy determines which evidence is acquired, when it becomes available, and which decisions remain feasible when it arrives. The belief from which the next episode is planned is therefore an output of the current policy. Budgets, delays and risk criteria constrain or evaluate this same process. The optimization trajectory is also a learning trajectory.
This coupling already drives the feedback restless bandits of Guha et al. [18 ] : the recurrent state process has a known transition law and evolves independently of the actions, while plays determine observations and future beliefs. Their approximation guarantees exploit further structure in the belief transitions. Liu and Zhao [26 ] establish indexability and derive closed-form indices for a related multichannel-access model. Meta-bandits extend this information problem across episodes: a play informs both the current quality and the recurring prior, and its value depends on the future decisions that its report can improve. Here planned restlessness supplies the structure for planning this evolution: reports from chosen plays change an arm’s learning state while other arms are played, on a known calendar. We plan the evolution of beliefs, budgets and pending reports jointly; execution must preserve the promised learning transitions or account for their loss. This action–information coupling is also central to dual control [24 ] .
An observed outcome selects a branch already represented in the plan. If a local experiment is completed under the specified report calendar, it produces the promised law of next learning states. The next local program is valued under that premise. But the LP fits the arms’ workloads into an episode only in expectation, and the hard horizon may prevent completion. An unfinished experiment leaves the learner without some of the evidence used to value future decisions. It changes the decision problem from which the learner must continue. Execution must therefore preserve the learning transitions on which continuation was valued, or charge for their loss and continue from the information and resources actually available.3 Planning and execution thus have a predictor–corrector structure. The LP supplies the value and transition certificate; the executor enforces pathwise feasibility. In the classic Mizuno–Todd–Ye interior-point method, the predictor reduces the duality gap and the corrector restores centrality while retaining that reduction [29 ] . Here the quantity to protect is the value of future learning. Preserving continuation while approximating current reward lets the episode guarantees compose without multiplying their reward losses across episodes (Equation (21) ).
The challenge of a proxy. The Bayes-optimal learner solves a joint program over the product of the arms’ state spaces. An algorithm therefore targets V p r i o r V_{\mathrm{prior}} V prior through a proxy V p r o x y V_{\mathrm{proxy}} V proxy , the value of the decision problem it actually solves:
V s t a t e − V A = ( V s t a t e − V p r i o r ) + ( V p r i o r − V p r o x y ) + ( V p r o x y − V A ) . (2) V_{\mathrm{state}} - V_A = (V_{\mathrm{state}} - V_{\mathrm{prior}}) + (V_{\mathrm{prior}} - V_{\mathrm{proxy}}) + (V_{\mathrm{proxy}} - V_A). \tag*{(2)} V state − V A = ( V state − V prior ) + ( V prior − V proxy ) + ( V proxy − V A ) . (2)
The second term is the choice of proxy and the third is how much of it the algorithm realizes; V p r o x y = V l e a r n V_{\mathrm{proxy}} = V_{\mathrm{learn}} V proxy = V learn recovers (1). The learning LP has value V L P ≥ V l e a r n V_{\mathrm{LP}} \ge V_{\mathrm{learn}} V LP ≥ V learn : it keeps the learner’s information but relaxes shared capacity and fast feedback. Its second term, ( V p r i o r − V l e a r n ) − ( V L P − V l e a r n ) (V_{\mathrm{prior}} - V_{\mathrm{learn}}) - (V_{\mathrm{LP}} - V_{\mathrm{learn}}) ( V prior − V learn ) − ( V LP − V learn ) , is signed, so the LP can value a learning program before we prove how much of it can be realized. Its third term is the leap from the LP certificate to a feasible learning policy. § 4 bounds this term; a guarantee V A ≥ α V L P − ϵ V_A \ge\alpha V_{\mathrm{LP}} - \epsilon V A ≥ α V LP − ϵ also gives V A ≥ α V l e a r n − ϵ V_A \ge\alpha V_{\mathrm{learn}} - \epsilon V A ≥ α V learn − ϵ , with no ordering between V L P V_{\mathrm{LP}} V LP and V p r i o r V_{\mathrm{prior}} V prior .
For posterior sampling, consider instead V p r o x y = V T S V_{\mathrm{proxy}} = V_{\mathrm{TS}} V proxy = V TS , the reward of hierarchical Thompson sampling (HierTS) told the recurring prior. Meta-regret against a prior-informed Thompson-sampling rule measures V T S − V A V_{\mathrm{TS}} - V_A V TS − V A [6 ] ; the allocation loss V p r i o r − V T S V_{\mathrm{prior}} - V_{\mathrm{TS}} V prior − V TS remains. HierTS shares information across episodes [19 ] . Its sampled-mean rule chooses the feasible arm with largest sampled episode mean, using the reports received. Even with the recurring priors known, this rule loses at least N T / 32 NT/32 N T /32 under delays and N T / 12 NT/12 N T /12 under budgets and caps (Examples A.3 and A.5). Thus V p r i o r = V l e a r n V_{\mathrm{prior}} = V_{\mathrm{learn}} V prior = V learn can coexist with a linear allocation loss. Evaluating risk conditional on a sampled episode quality also fails for mean–variance, CVaR and downside objectives (Example A.6). Appendix § 7 gives these comparisons, their calculations and the corresponding planning solutions. Bayes-regret analyses bound the complete loss against the best action for each realized task in their settings [7 , 19 , 25 ] . The LP’s prices and continuation values give a common certificate: every admissible policy’s shortfall splits into unused capacity and Bellman slacks. § 6.3 uses it to compare planning and execution, including HierTS.
Contributions
The Learning LP and risk generalizations. For general bounded reward distributions, we formulate a relaxation over learning trajectories. It decomposes into a backward cascade of the single-arm LPs of Guha and Munagala [17 ] (GM13 hereafter), linked by continuation values (Theorem 2.9). For explicit finite state graphs, its size is polynomial in their total size. Episode weights set the tradeoff between current reward and the later decisions improved by learning (§ 3.6 ). Bounded outcome scores enter linearly. For risk functionals, we identify the conditions for arm decomposition and execution.
Feasible execution. Full completion uses a time reserve whose leading lower bound matches as the caps grow; protecting selected reports avoids that reserve (§ 4 ). For a single undelayed episode, caps at most T / X T/X T / X improve GM13’s factor 2 to 1 + ( 2 π X ) − 1 / 2 + O ( X − 1 ) 1 + (2\pi X)^{-1/2} + O(X^{-1}) 1 + ( 2 π X ) − 1/2 + O ( X − 1 ) , with a matching leading lower bound. Fixed-order execution also permits up to K K K distinct arms per step (Theorem 3.13). With arbitrary caps and delays at most δ \delta δ , the factor tends to 2 whenever δ log δ = o ( T ) \delta\log\delta= o(T) δ log δ = o ( T ) .
Learning-value stability and verified execution. The episode dual bounds planning-value error by the weighted total-variation error in observable experiment laws, with sharp constant 1. Protected reports transfer this comparison through execution, paying the single-episode factor once. Short probes give an executable O ( N 2 / 3 ) O(N^{2/3}) O ( N 2/3 ) stationary comparison; the same argument permits drift (§ 5 , Appendix § 9 ).
Information prices and predictive checks. Capacity prices bound a report’s value in the learning LP; repricing gives an exact nonnegative correction to its fixed-price valuation. Purchased reports and prior buyout enter the local programs. The price V p r i o r − V l e a r n V_{\mathrm{prior}}-V_{\mathrm{learn}} V prior − V learn is O ( N ) O(\sqrt{N}) O ( N ) under Proposition 5.4’s assumptions, with other parameters fixed. A common dual certificate evaluates every admissible policy, including HierTS. Its reported innovations separate allocation loss from predictive error and support sequential model checks (§ 6 ).
Section 2 gives the model and learning LP; Section 3 develops execution and its extension to outcome scores. Section 4 proves stability and learning comparisons. Section 5 prices information and develops the policy certificates and predictive checks. Appendix A collects the illustrative examples and their planning solutions. The remaining appendices give execution proofs, statistical calculations and information-price bounds.
Other related work
Learning priors. Simchowitz et al. [38 ] bound the sensitivity of Bayesian decision rules to prior misspecification and derive meta-learning guarantees from prior estimation. Azizi et al. [5 ] study meta-learning for best-arm identification, using simple regret as the task objective.
Restless bandits and relaxations. Lagrangian relaxation separates weakly coupled dynamic programs by pricing their shared resources [1 , 40 ] . Liu and Zhao [26 ] establish indexability for two-state Markov channels, and Caro and Yoo [9 ] study no-crossover response delays. Jung and Tewari [22 ] analyze posterior sampling for unknown restless dynamics in resetting episodes, relative to a chosen model-to-policy map. Jiang et al. [20 ] study unknown transition and reward laws with unobserved states, using a planning oracle. Paid state probes also appear in Guha et al. [18 Section 7] .
Delayed and partial feedback. Joulani et al. [21 ] give reductions from delayed-feedback learning to algorithms for immediate feedback. Kalkanli and Özgür [23 ] analyze Thompson sampling with updates at batch boundaries. Impatient Bandits combines partial observations and delayed rewards through a predictive model [28 , 42 ] .
Model and the learning LP
Model
There are n n n arms, indexed by i ∈ { 1 , … , n } i \in\{1,\ldots,n\} i ∈ { 1 , … , n } , and N N N episodes, indexed by e ∈ { 1 , … , N } e \in\{1,\ldots,N\} e ∈ { 1 , … , N } . The planning horizon N N N and positive integer episode lengths T 1 , … , T N T_1,\ldots,T_N T 1 , … , T N are fixed and known before planning. Write T = ( T 1 , … , T N ) \mathbf{T}=(T_1,\ldots,T_N) T = ( T 1 , … , T N ) and T = max e T e T=\max_e T_e T = max e T e . At each step the learner plays one arm or waits. Waiting earns zero, uses no arm budget and allows scheduled reports to arrive. The LP below allows unequal known lengths. The execution and rate statements use the uniform case T e = T T_e=T T e = T unless stated otherwise.
Arm i i i has an episode quality θ i ( e ) \theta_i^{(e)} θ i ( e ) in a standard Borel parameter space Θ i \Theta_i Θ i . Given that quality, its successive rewards have a known law K i ( θ i ( e ) , d y ) K_i(\theta_i^{(e)},\mathop{}\!dy) K i ( θ i ( e ) , d y ) on [ 0 , 1 ] [0,1] [ 0 , 1 ] , with mean μ i ( θ ) = ∫ y K i ( θ , d y ) \mu_i(\theta)=\int y K_i(\theta,dy) μ i ( θ ) = ∫ y K i ( θ , d y ) . The quality specifies a reward distribution: taking Θ i \Theta_i Θ i to be the probability measures on [ 0 , 1 ] [0,1] [ 0 , 1 ] and K i ( θ , ⋅ ) = θ K_i(\theta,\cdot)=\theta K i ( θ , ⋅ ) = θ includes every bounded reward law. Successive plays consume an outcome tape ( Y i , j ( e ) ) j ≥ 1 (Y_{i,j}^{(e)})_{j\ge1} ( Y i , j ( e ) ) j ≥ 1 , which may be sampled before execution. A play earns its outcome immediately, but the learner observes it only when its report arrives. The policy’s random coins are independent of the arms.
Assumption 2.1 (Independent prior sequences across arms). Arm i i i has a random sequence of priors P i = ( P i ( e ) ) e ≥ 1 \mathcal{P}_i=(P_i^{(e)})_{e\ge1} P i = ( P i ( e ) ) e ≥ 1 on Θ i \Theta_i Θ i , with known law Q 1 , i \mathcal{Q}_{1,i} Q 1 , i . These sequences are independent across arms. Conditional
on them, the qualities are independent with θ i ( e ) ∼ P i ( e ) \theta_i^{(e)} \sim P_i^{(e)} θ i ( e ) ∼ P i ( e ) . Conditional on the qualities, tape entries are independent with law K i ( θ i ( e ) , ⋅ ) K_i(\theta_i^{(e)}, \cdot) K i ( θ i ( e ) , ⋅ ) .
The learner knows the meta-prior Q 1 = ⨂ i Q 1 , i Q_1 = \bigotimes_i Q_{1,i} Q 1 = ⨂ i Q 1 , i , but not the realized priors. In the stationary case P i ( e ) = P i P_i^{(e)} = P_i P i ( e ) = P i for every episode. There are then two levels of exchangeability: episodes are i.i.d. given the recurring prior, and plays within an episode are i.i.d. given its quality. Conversely, invariance of each arm’s infinite outcome array under permutations of episodes and of plays within each episode gives this hierarchy by the hierarchical de Finetti representation [4 ] . The model requires no parametric or conjugate family.
Arm i i i has a total play budget B i B_i B i and a per-episode cap L i L_i L i , both nonnegative integers. Write B = ( B i ) i B = (B_i)_i B = ( B i ) i and L ˉ = max i L i ≤ T \bar L = \max_i L_i \le T L ˉ = max i L i ≤ T . If its remaining budget is b b b , it can make at most min { L i , T e , b } \min\{L_i, T_e, b\} min { L i , T e , b } plays in episode e e e . Both limits hold on every realized path.
Arm i i i has known delay δ i ≥ 1 \delta_i \ge1 δ i ≥ 1 . A play at step t t t reports before the decision at t + δ i t+\delta_i t + δ i if that step lies in the same episode; δ i = 1 \delta_i = 1 δ i = 1 is undelayed feedback. In episode e e e , the arm is fast if δ i < T e \delta_i < T_e δ i < T e and slow otherwise. A fast arm’s remaining outcomes report at the next episode start. A slow arm’s episode-e e e batch reports at a known boundary rep i ( e ) ∈ { e + 1 , … , e + g } \operatorname{rep}_i(e) \in\{e+1,\ldots,e+g\} rep i ( e ) ∈ { e + 1 , … , e + g } , where the integer g ≥ 1 g \ge1 g ≥ 1 bounds the report lag in episodes. Reports from earlier episodes arrive only at boundaries. The calendar is fixed in advance and independent of outcomes and within-episode scheduling. For g = 1 g=1 g = 1 , all reports arrive before the next episode; otherwise queues can remain pending. Time between episodes is outside the reward horizon. Write δ = max i δ i \delta= \max_i \delta_i δ = max i δ i when a common delay bound is used.
Policies and states
A policy is delay-respecting if it uses only reported outcomes and independent randomness. It is admissible if it also plays at most one arm per step and respects all caps and budgets. Let H e − 1 H_{e-1} H e − 1 be the history available at episode e e e ’s start, including reports delivered there. Write V e V_e V e for its reward and OPT N ( B ) \operatorname{OPT}_N(B) OPT N ( B ) for the optimal expected reward over the N N N episodes.
A local policy π \pi π specifies one arm’s plays within one episode; we also call it an experiment. An arm program σ i \sigma_i σ i chooses its local policies over all episodes from that arm’s own reports and independent coins. At the start of episode e e e , write Q e , i = L ( P i ∣ H e − 1 ) Q_{e,i} = \mathcal{L}(P_i \mid H_{e-1}) Q e , i = L ( P i ∣ H e − 1 ) for arm i i i ’s posterior over its prior sequence, b e , i b_{e,i} b e , i for its remaining budget, and D e , i D_{e,i} D e , i for its pending reports with their episode labels and delivery dates. The queue records the plays awaiting reports, not their unobserved outcomes. These objects determine the conditional joint law of the arm’s future outcomes and reports. The arm’s boundary state is S e , i = ( Q e , i , b e , i , D e , i ) S_{e,i} = (Q_{e,i}, b_{e,i}, D_{e,i}) S e , i = ( Q e , i , b e , i , D e , i ) . Let S e , i \mathcal{S}_{e,i} S e , i be the possible boundary states of this arm in episode e e e , so S e , i ∈ S e , i S_{e,i} \in\mathcal{S}_{e,i} S e , i ∈ S e , i ; initially Y 1 , i = ( Q 1 , i , B i , ∅ ) Y_{1,i} = (Q_{1,i}, B_i, \varnothing) Y 1 , i = ( Q 1 , i , B i , ∅ ) . In a calculation for a fixed arm and episode, we write S = ( Q , b , D ) S = (Q,b,D) S = ( Q , b , D ) for a possible state and suppress the arm index on its rewards and transition kernels.
Definition 2.2 (Planned restlessness). An arm is planned-restless , if, while other arms are played, its information changes only through reports of its own earlier plays, by a delivery rule fixed in advance. A report arrives at a known offset after its play if that step lies in the same episode. Otherwise it arrives at a boundary that depends only on the play’s episode. Neither the rule nor the boundary depends on outcomes or on how arms are interleaved.
Every arm in this model is planned-restless. Its conditional reward law stays fixed within an episode. Completion preserves the law of its next boundary state under interleaving (Lemma 3.5).
For planning, fast arms get immediate access to their current outcomes; this is the delay relaxation . Slow arms use nonadaptive local policies. Each arm program still respects its cap, budget and boundary report calendar. Execution enforces the physical delays.
Single-arm policies
Use the posterior states of Guha and Munagala [17 Sections 2.1 and 3.1] . At a state u u u , let r ( u ) r(u) r ( u ) be the expected reward of a play and p ( u , v ) p(u,v) p ( u , v ) its transition probabilities. The state records the plays made, remaining budget and reports. It may be a complete observable history or a sufficient statistic preserving conditional future laws and feasible actions. The play count makes the local graph acyclic. The root has no predecessors. At each playable node u u u , p ( u , v ) ≥ 0 p(u,v) \ge0 p ( u , v ) ≥ 0 and ∑ v p ( u , v ) = 1 \sum_v p(u,v) = 1 ∑ v p ( u , v ) = 1 . For a slow arm it has no outcome-dependent branches before handoff. Bayesian updating gives r ( u ) = ∑ v p ( u , v ) r ( v ) r(u) = \sum_v p(u,v)r(v) r ( u ) = ∑ v p ( u , v ) r ( v ) , including the predictive mean at terminal states.
For a finite graph rooted at u 0 u_0 u 0 , let f ( u ) f(u) f ( u ) , w ( u ) w(u) w ( u ) and z ( u ) z(u) z ( u ) be the probability assigned to reaching, playing at and stopping at u u u . Then
f ( u ) ≤ 1 { u = u 0 } + ∑ v p ( v , u ) w ( v ) , f ( u ) ≤ 1 , w ( u ) + z ( u ) = f ( u ) , f , w , z ≥ 0. (3) \begin{aligned}
f(u) &\le\mathbf{1}_{\{u=u_0\}} + \sum_v p(v,u)w(v), \qquad f(u) \le1, \qquad w(u) + z(u) = f(u), \qquad f,w,z \ge0. \tag*{(3)}
\end{aligned} f ( u ) ≤ 1 { u = u 0 } + v ∑ p ( v , u ) w ( v ) , f ( u ) ≤ 1 , w ( u ) + z ( u ) = f ( u ) , f , w , z ≥ 0. (3)
The reach inequality permits unused probability. In particular, 0 ≤ w ( u ) , z ( u ) ≤ f ( u ) ≤ 1 0 \le w(u), z(u) \le f(u) \le1 0 ≤ w ( u ) , z ( u ) ≤ f ( u ) ≤ 1 . Set w ( u ) = 0 w(u) = 0 w ( u ) = 0 where the cap or remaining budget forbids a play.
Proposition 2.3 (Local policy representation). Every allowed local policy gives a feasible solution of (3). Conversely, every feasible solution has a completion ( f ˉ , w , z ˉ ) (\bar f,w,\bar z) ( f ˉ , w , z ˉ ) that preserves the play probabilities and satisfies equality in the reach constraints. Playing with probability w ( u ) / f ˉ ( u ) w(u)/\bar f(u) w ( u ) / f ˉ ( u ) realizes this completion. Its expected reward, workload and stopping probabilities are
R ( π ) = ∑ u r ( u ) w ( u ) , t ( π ) = ∑ u w ( u ) , Pr { stop at u } = z ˉ ( u ) . (4) R(\pi) = \sum_u r(u)w(u), \qquad\mathsf{t}(\pi) = \sum_u w(u), \qquad\Pr\{\text{stop at }u\} = \bar z(u). \tag*{(4)} R ( π ) = u ∑ r ( u ) w ( u ) , t ( π ) = u ∑ w ( u ) , Pr { stop at u } = z ˉ ( u ) . (4)
Proof. A policy’s probabilities satisfy the reach constraints with equality. For a feasible LP solution, put
f ˉ ( u ) = 1 { u = u 0 } + ∑ v p ( v , u ) w ( v ) , z ˉ ( u ) = f ˉ ( u ) − w ( u ) . \bar f(u) = \mathbf{1}_{\{u=u_0\}} + \sum_v p(v,u)w(v), \qquad\bar z(u) = \bar f(u) - w(u). f ˉ ( u ) = 1 { u = u 0 } + v ∑ p ( v , u ) w ( v ) , z ˉ ( u ) = f ˉ ( u ) − w ( u ) .
Then f ˉ ≥ f \bar f \ge f f ˉ ≥ f and z ˉ ≥ z \bar z \ge z z ˉ ≥ z . Induction over the DAG shows that playing with probability w / f ˉ w/\bar f w / f ˉ gives arrival probabilities f ˉ \bar f f ˉ , play probabilities w w w and stopping probabilities z ˉ \bar z z ˉ . Thus f ˉ ≤ 1 \bar f \le1 f ˉ ≤ 1 and ∑ u z ˉ ( u ) = 1 \sum_u \bar z(u) = 1 ∑ u z ˉ ( u ) = 1 . Summing play contributions proves (4). Choices at zero-probability states are arbitrary.
Policy and handoff calculations below use completed solutions; we write their probabilities as f , w , z f,w,z f , w , z .
For a finite reward alphabet Y i \mathcal{Y}_i Y i , the likelihood of an ordered outcome history h = ( y 1 , … , y j ) h=(y_1,\ldots,y_j) h = ( y 1 , … , y j ) under an episode prior P P P is
ℓ i ( P ; h ) = ∫ Θ i ∏ t = 1 j K i ( θ , { y t } ) P ( d θ ) . (5) \ell_i(P;h) = \int_{\Theta_i} \prod_{t=1}^{j} K_i(\theta,\{y_t\})\,P(\mathrm{d}\theta). \tag*{(5)} ℓ i ( P ; h ) = ∫ Θ i t = 1 ∏ j K i ( θ , { y t }) P ( d θ ) . (5)
It is linear in P P P and satisfies ℓ i ( P ; h ) = ∑ y ℓ i ( P ; h y ) \ell_i(P;h) = \sum_y \ell_i(P;hy) ℓ i ( P ; h ) = ∑ y ℓ i ( P ; h y ) . Suppress the arm index. Let A ( h ) A(h) A ( h ) be the probability, over a local policy’s coins, of making the plays in h h h when supplied those outcomes, and B ( h ) B(h) B ( h ) the probability of also making the next play. Then A ( ∅ ) = 1 A(\varnothing) = 1 A ( ∅ ) = 1 , A ( h y ) = B ( h ) A(hy) = B(h) A ( h y ) = B ( h ) and 0 ≤ B ( h ) ≤ A ( h ) 0 \le B(h) \le A(h) 0 ≤ B ( h ) ≤ A ( h ) . The reach, play and terminal probabilities are respectively
A ( h ) ℓ ( P ; h ) , B ( h ) ℓ ( P ; h ) , ( A ( h ) − B ( h ) ) ℓ ( P ; h ) . (6) A(h)\ell(P;h), \qquad B(h)\ell(P;h), \qquad(A(h)-B(h))\ell(P;h). \tag*{(6)} A ( h ) ℓ ( P ; h ) , B ( h ) ℓ ( P ; h ) , ( A ( h ) − B ( h )) ℓ ( P ; h ) . (6)
These identities hold for delayed policies because each decision uses only reported outcomes and independent coins. In particular,
R P ( π ) = ∑ h B ( h ) ∑ y ∈ Y y ℓ ( P ; h y ) , t P ( π ) = ∑ h B ( h ) ℓ ( P ; h ) . R^{P}(\pi) = \sum_h B(h)\sum_{y\in\mathcal{Y}} y\ell(P;hy), \qquad t^{P}(\pi) = \sum_h B(h)\ell(P;h). R P ( π ) = h ∑ B ( h ) y ∈ Y ∑ y ℓ ( P ; h y ) , t P ( π ) = h ∑ B ( h ) ℓ ( P ; h ) .
A depth-k k k commit policy makes at most k k k adaptive plays and then fixes its remaining play count.
Corollary 2.4 (Information in an experiment). (i) The stopped outcome law, reward and workload use only likelihoods at histories reached with positive weight. (ii) Given the policy weights, the stopped law determines these likelihoods. (iii) Fixed recurring Bernoulli priors with the same first m m m moments are indistinguishable by local policies capped at m m m plays per episode. (iv) A depth-k k k commit policy with cap L L L has reward and workload determined by the law of min { L , k + 1 } \min\{L,k+1\} min { L , k + 1 } observations; its continuation can require all L L L observations.
Proof. Equation (6) gives (i). The stopped law gives each prefix probability; division by A ( h ) > 0 A(h)>0 A ( h ) > 0 gives (ii). For Bernoulli rewards a length-j j j likelihood is ∫ θ s ( 1 − θ ) j − s P ( d θ ) \int\theta^{s}(1-\theta)^{j-s}\,P(\mathrm{d}\theta) ∫ θ s ( 1 − θ ) j − s P ( d θ ) , a polynomial moment of degree j j j , proving (iii). For (iv), commitment fixes the tail length from at most k k k outcomes. Conditional on that prefix and the quality, its expected reward is unchanged by replacing the tail by its length times one further outcome. Its full outcome law still uses all L L L observations.
Proposition 2.5 (History likelihood). In a global observable history h h h , let h i , u h_{i,u} h i , u list the outcomes reported from arm i i i and episode u u u . For finite alphabets,
Pr π ( H = h ) = w π ( h ) ∏ i ℓ i ( h ) , ℓ i ( h ) = E Q 1 , i ∏ u ℓ i ( P i ( u ) ; h i , u ) , (7) \Pr_{\pi}(H=h)=w_{\pi}(h)\prod_i \ell_i(h),\qquad\ell_i(h)=\mathbb{E}_{Q_{1,i}}\prod_u \ell_i\left(P_i^{(u)};h_{i,u}\right), \tag*{(7)} π Pr ( H = h ) = w π ( h ) i ∏ ℓ i ( h ) , ℓ i ( h ) = E Q 1 , i u ∏ ℓ i ( P i ( u ) ; h i , u ) , (7) where w π ( h ) w_{\pi}(h) w π ( h ) depends on the policy and reported history but is independent of the latent prior sequences. Unreported outcomes are marginalized.
Proof. Fix the reported history. The policy’s choices contribute a factor obtained by averaging over its independent coins. Conditional on the prior sequences, integrating each episode quality gives its likelihood in (5) . Independence across arms gives the product in (7) . The fixed report calendar contributes no outcome-dependent factor.
Lemma 2.6 (The conditional episode). Given H e − 1 H_{e-1} H e − 1 , the posterior factorizes as Q e = ⨂ i Q e , i Q_e=\bigotimes_i Q_{e,i} Q e = ⨂ i Q e , i , and the episode has independent predictive priors P ˉ e , i = E [ P i ( e ) ∣ H e − 1 ] \bar{P}_{e,i}=\mathbb{E}[P_i^{(e)}\mid H_{e-1}] P ˉ e , i = E [ P i ( e ) ∣ H e − 1 ] . Writing P ˉ e = ( P ˉ e , i ) i \bar{P}_e=(\bar{P}_{e,i})_i P ˉ e = ( P ˉ e , i ) i , every episode policy satisfies E [ V e ∣ H e − 1 ] = V P ˉ e ( π e ) \mathbb{E}[V_e\mid H_{e-1}]=V^{\bar{P}_e}(\pi_e) E [ V e ∣ H e − 1 ] = V P ˉ e ( π e ) .
Proof. The policy factor in (7) cancels in Bayes’ rule, leaving independent updates of the arm meta-priors. The new tape law is linear in P i ( e ) P_i^{(e)} P i ( e ) , so averaging gives P ˉ e , i \bar{P}_{e,i} P ˉ e , i . Reports from earlier episodes arrive only at boundaries.
The predictive priors suffice for current-episode reward. Continuation requires the full beliefs Q e , i Q_{e,i} Q e , i , which determine how new reports change later decisions (Example A.2). For general observation spaces these statements use conditional kernels and integrals. Finite LPs require a supplied finite state representation.
The continuation
Fix arm i i i , an episode e e e and its incoming boundary state S S S . We suppress the arm index throughout this calculation. Each arm has N N N episode layers. In its DAG, the episode-e e e local policy moves probability mass to possible starting nodes of episode e + 1 e+1 e + 1 . Different reports can give different posterior beliefs about the recurring prior; the plays also change the remaining budget and the queue. These are the states from which the arm’s program continues.
Definition 2.7 (Planned handoff). For an allowed local policy π \pi π , the planned handoff T e π ( S , ⋅ ) T_e^{\pi}(S,\cdot) T e π ( S , ⋅ ) is the law of the arm’s state at the start of episode e + 1 e+1 e + 1 when π \pi π runs alone and to completion (undelayed if fast, nonadaptively if slow), and the reports due at that boundary are delivered.
On a finite state graph, write H e ( S , u , v ) H_e(S,u,v) H e ( S , u , v ) for the probability that stopping at local node u u u leads to next-episode starting node v v v . This transition deducts the plays, updates the queue and delivers the reports due at the boundary. The incoming state and local node determine these probabilities; the policy determines its stopping probabilities z ( u ) z(u) z ( u ) . Thus the mass arriving at v v v is
q ( v ) = ∑ u z ( u ) H e ( S , u , v ) = T e π ( S , v ) . (8) q(v)=\sum_u z(u)H_e(S,u,v)=\mathcal{T}^{\pi}_e(S,v). \tag*{(8)} q ( v ) = u ∑ z ( u ) H e ( S , u , v ) = T e π ( S , v ) . (8)
This is the boundary flow joining successive episode DAGs.
Assign a future value F ( v ) F(v) F ( v ) to each next-episode starting node. The current policy’s continuation contribution is the expected value of the state it produces:
C F ( π ) = ∑ v q ( v ) F ( v ) = ∑ u z ( u ) c F ( u ) , c F ( u ) = ∑ v H e ( S , u , v ) F ( v ) . (9) C_F(\pi)=\sum_v q(v)F(v)=\sum_u z(u)c_F(u),\qquad c_F(u)=\sum_v H_e(S,u,v)F(v). \tag*{(9)} C F ( π ) = v ∑ q ( v ) F ( v ) = u ∑ z ( u ) c F ( u ) , c F ( u ) = v ∑ H e ( S , u , v ) F ( v ) . (9)
At a stopping history, the policy weight cancels in Bayes’ rule. Its posterior, remaining budget, queue and conditional law of reports due at the boundary therefore fix H e H_e H e independently of the policy. Thus both the handoff and the continuation in (9) are linear in terminal probabilities, even when F F F is nonlinear in the posterior. Figure 1 shows these two operations. For general observation spaces the sums are integrals against the handoff kernels.
For a fast arm with a finite reward alphabet and no earlier report pending, the handoff from a terminal history h h h is deterministic:
S h + = ( Q h + , b − ∣ h ∣ , ∅ ) , Q h + ( d Π i ) ∝ ℓ i ( P i ( e ) ; h ) Q ( d Π i ) . (10) S_h^+=(Q_h^+,b-\lvert h\rvert,\varnothing),\qquad Q_h^+(d\Pi_i)\propto\ell_i(P_i^{(e)};h)\,Q(d\Pi_i). \tag*{(10)} S h + = ( Q h + , b − ∣ h ∣ , ∅ ) , Q h + ( d Π i ) ∝ ℓ i ( P i ( e ) ; h ) Q ( d Π i ) . (10)
The policy’s action probabilities cancel from Bayes’ rule, so the state attached to h h h is fixed independently of its flow weight. Pending reports can make the handoff random, as represented by the kernel in (8). The next subsection derives the arm’s future values by pricing the shared episode capacities.
The learning LP and its prices
The learning LP has N N N episode layers for each arm. A layer contains a local program for every possible incoming boundary state. Stopping in layer e e e sends probability mass to starting states in layer $e+1. At fixed capacity prices, their values are computed backward from episode N N N . The prices are chosen jointly across episodes. Thus the cascade has depth N N N ; its size also counts the states and transitions in each layer. Each local block is the single-arm LP of GM13, with its terminal values supplied by the next episode.
Let R e ( σ i ) R_e(\sigma_i) R e ( σ i ) and t e ( σ i ) t_e(\sigma_i) t e ( σ i ) be arm i i i ’s expected reward and number of plays in episode e e e . Each arm program respects its cap and total budget on every path. For known capacities C = ( C 1 , … , C N ) \mathbf{C}=(C_1,\ldots,C_N) C = ( C 1 , … , C N ) with 0 < C e ≤ T e 0<C_e\le T_e 0 < C e ≤ T e , define
LP N ( C , B ) = sup ( σ i ) ∑ e = 1 N ∑ i R e ( σ i ) subject to ∑ i t e ( σ i ) ≤ C e ( e = 1 , … , N ) . (11) \operatorname{LP}_N(\mathbf{C},B)=\sup_{(\sigma_i)}\sum_{e=1}^{N}\sum_i R_e(\sigma_i)\quad\text{subject to}\quad\sum_i t_e(\sigma_i)\le C_e\quad(e=1,\ldots,N). \tag*{(11)} LP N ( C , B ) = ( σ i ) sup e = 1 ∑ N i ∑ R e ( σ i ) subject to i ∑ t e ( σ i ) ≤ C e ( e = 1 , … , N ) . (11)
Only these shared capacity constraints hold in expectation. For uniform lengths and capacities, write LP N ( C , B ) \operatorname{LP}_N(C,B) LP N ( C , B ) for C = ( C , … , C ) \mathbf{C}=(C,\ldots,C) C = ( C , … , C ) . The subscript N N N records the number of episode layers; LP ( P ) \operatorname{LP}(P) LP ( P ) below denotes a single-episode relaxation.
Proposition 2.8 (Relaxation across episodes). (i) OPT N ( B ) ≤ LP N ( T , B ) \operatorname{OPT}_N(B)\le\operatorname{LP}_N(\mathbf{T},B) OPT N ( B ) ≤ LP N ( T , B ) . (ii) For 0 < s ≤ 1 0<s\le1 0 < s ≤ 1 , LP N ( s C , B ) ≥ s LP N ( C , B ) \operatorname{LP}_N(s\mathbf{C},B)\ge s\operatorname{LP}_N(\mathbf{C},B) LP N ( s C , B ) ≥ s LP N ( C , B ) ; in particular, LP N ( C , B ) ≥ ( C / T ) LP N ( T , B ) \operatorname{LP}_N(\mathbf{C},B)\ge(C/T)\operatorname{LP}_N(\mathbf{T},B) LP N ( C , B ) ≥ ( C / T ) LP N ( T , B ) in the uniform case. (iii) Pricing the episode capacities separates the arms; fixing the future values makes each local step an LP.
Proof. (i) Project an admissible policy onto each arm. The other arms’ latent variables and tapes, together with the policy coins, supply randomness independent of that arm. Its projected program preserves reward, workload, caps, budget and boundary reports under relaxed fast feedback. The expected workloads sum to at most T e T_e T e in episode e e e . (ii) Retain each entire arm program independently with probability s s s and otherwise make no plays of that arm. Reward and every episode’s expected workload are multiplied by s s s . (iii) Price the capacity constraints as below.
The conditional reward law is fixed within an episode. The relaxation therefore uses one capacity constraint per episode, without a within-episode time index [17 Claim 1] .
The LP on states. Join each arm’s local graphs through their handoff kernels. Let V i \mathcal{V}_i V i be its acyclic graph over episodes 1 , … , N 1,\ldots,N 1 , … , N , o i o_i o i its initial node, and e ( u ) e(u) e ( u ) the episode of node u u u . Write p i ( u , v ) p_i(u,v) p i ( u , v ) for play transitions and h i ( u , v ) h_i(u,v) h i ( u , v ) for handoffs to the next episode. Each allowed play and each handoff has total outgoing probability one. Handoffs after episode N N N enter a terminal state of value zero. Its incoming mass need not be included as a variable. The finite LP is
LP N ( C , B ) = max ∑ i ∑ u ∈ V i r i ( u ) w i ( u ) s.t. f i ( u ) ≤ 1 { u = o i } + ∑ v [ p i ( v , u ) w i ( v ) + h i ( v , u ) z i ( v ) ] , f i ( u ) ≤ 1 , w i ( u ) + z i ( u ) = f i ( u ) , f i , w i , z i ≥ 0 , ∑ i ∑ u : e ( u ) = e w i ( u ) ≤ C e ( e = 1 , … , N ) . (12) \begin{aligned}
\operatorname{LP}_N(\mathbf{C},B)&=\max\sum_i\sum_{u\in\mathcal{V}_i}r_i(u)w_i(u)\\
\text{s.t.}\quad f_i(u)&\le\mathbf{1}_{\{u=o_i\}}+\sum_v\left[p_i(v,u)w_i(v)+h_i(v,u)z_i(v)\right],\\
f_i(u)&\le1,\\
w_i(u)+z_i(u)&=f_i(u),\qquad f_i,w_i,z_i\ge0,\\
\sum_i\sum_{u:e(u)=e}w_i(u)&\le C_e\qquad(e=1,\ldots,N). \tag*{(12)}
\end{aligned} LP N ( C , B ) s.t. f i ( u ) f i ( u ) w i ( u ) + z i ( u ) i ∑ u : e ( u ) = e ∑ w i ( u ) = max i ∑ u ∈ V i ∑ r i ( u ) w i ( u ) ≤ 1 { u = o i } + v ∑ [ p i ( v , u ) w i ( v ) + h i ( v , u ) z i ( v ) ] , ≤ 1 , = f i ( u ) , f i , w i , z i ≥ 0 , ≤ C e ( e = 1 , … , N ) . (12)
There is one unit of initial mass per arm; later episode roots receive handoff mass from the preceding layer. Complete a feasible solution in topological order: keep w i w_i w i , set f ˉ i \bar f_i f ˉ i to the full initial and incoming probability, using the completed handoffs, and set z ˉ i = f ˉ i − w i \bar z_i=\bar f_i-w_i z ˉ i = f ˉ i − w i . Induction gives f ˉ i ≥ f i \bar f_i\ge f_i f ˉ i ≥ f i and z ˉ i ≥ z i \bar z_i\ge z_i z ˉ i ≥ z i . The completed arm program has the same play probabilities, reward and workload. All its boundary-state distributions have total mass one. Forbidden plays have w i ( u ) = 0 w_i(u)=0 w i ( u ) = 0 . The state records the remaining budget and current play count, so these restrictions enforce both pathwise limits. Only the capacity constraints couple the arms.
Let λ e ≥ 0 \lambda_e \ge0 λ e ≥ 0 be the dual multiplier of episode e e e ’s capacity constraint, and write λ = ( λ 1 , … , λ N ) \lambda= (\lambda_1,\ldots,\lambda_N) λ = ( λ 1 , … , λ N ) . For fixed prices, U e , i λ ( S ) U^\lambda_{e,i}(S) U e , i λ ( S ) is the maximum expected remaining reward of arm i i i from state S S S , less λ j \lambda_j λ j for each play in episode j ≥ e j \ge e j ≥ e . Put U N + 1 , i λ = 0 U^\lambda_{N+1,i}=0 U N + 1 , i λ = 0 .
Theorem 2.9 (Dual and episode decomposition). For finite state graphs, the learning LP (12) has dual
LP N ( C , B ) = min λ ≥ 0 { ∑ e = 1 N λ e C e + ∑ i U 1 , i λ ( Q 1 , i , B i , ∅ ) } . (13) \operatorname{LP}_N(\mathbf{C},\mathbf{B})=\min_{\lambda\ge0}\left\{\sum_{e=1}^{N}\lambda_e C_e+\sum_i U^\lambda_{1,i}(Q_{1,i},B_i,\varnothing)\right\}. \tag*{(13)} LP N ( C , B ) = λ ≥ 0 min { e = 1 ∑ N λ e C e + i ∑ U 1 , i λ ( Q 1 , i , B i , ∅ ) } . (13) At fixed prices, the nonnegative node values v i ( u ) v_i(u) v i ( u ) satisfy
v i ( u ) ≥ r i ( u ) − λ e ( u ) + ∑ v p i ( u , v ) v i ( v ) , v i ( u ) ≥ ∑ v h i ( u , v ) v i ( v ) , (14) v_i(u)\ge r_i(u)-\lambda_e(u)+\sum_v p_i(u,v)v_i(v),\qquad v_i(u)\ge\sum_v h_i(u,v)v_i(v), \tag*{(14)} v i ( u ) ≥ r i ( u ) − λ e ( u ) + v ∑ p i ( u , v ) v i ( v ) , v i ( u ) ≥ v ∑ h i ( u , v ) v i ( v ) , (14) with the play inequality omitted where playing is forbidden. Their smallest feasible values are computed by a backward cascade of N N N single-episode LP layers, one local LP per incoming state.
Fix arm i i i and λ \lambda λ , suppress these indices in U e U_e U e , and put U N + 1 = 0 U_{N+1}=0 U N + 1 = 0 . Root the episode-e e e local graph at S S S . For f , w , z f,w,z f , w , z satisfying (3) and next-state probabilities q q q from (8),
U e ( S ) = max f , w , z { ∑ u ( r ( u ) − λ e ) w ( u ) + ∑ S ′ q ( S ′ ) U e + 1 ( S ′ ) } . (15) U_e(S)=\max_{f,w,z}\left\{\sum_u(r(u)-\lambda_e)w(u)+\sum_{S'}q(S')U_{e+1}(S')\right\}. \tag*{(15)} U e ( S ) = f , w , z max { u ∑ ( r ( u ) − λ e ) w ( u ) + S ′ ∑ q ( S ′ ) U e + 1 ( S ′ ) } . (15)
The first sum is priced current reward; the second is continuation. With future values fixed, put
c e ( S , u ) = ∑ S ′ H e ( S , u , S ′ ) U e + 1 ( S ′ ) . c_e(S,u)=\sum_{S'}H_e(S,u,S')U_{e+1}(S'). c e ( S , u ) = S ′ ∑ H e ( S , u , S ′ ) U e + 1 ( S ′ ) .
This is GM13’s priced single-arm LP [17 Section 3.4] , with stopping value c e ( S , u ) c_e(S,u) c e ( S , u ) and recursion
v ( u ) = max { c e ( S , u ) , r ( u ) − λ e + ∑ u ′ p ( u , u ′ ) v ( u ′ ) } . (16) v(u)=\max\left\{c_e(S,u),\ r(u)-\lambda_e+\sum_{u'}p(u,u')v(u')\right\}. \tag*{(16)} v ( u ) = max { c e ( S , u ) , r ( u ) − λ e + u ′ ∑ p ( u , u ′ ) v ( u ′ ) } . (16)
Only stopping is allowed at forbidden plays, and U e ( S ) = v ( u 0 ) U_e(S)=v(u_0) U e ( S ) = v ( u 0 ) . A slow arm chooses a nonadaptive play count. Each local maximizer can be completed without loss, so q q q is a probability distribution.
Proof. The acyclic graph and unit initial mass imply f i ( u ) ≤ 1 f_i(u)\le1 f i ( u ) ≤ 1 , so these bounds are redundant. Eliminate f i = w i + z i f_i=w_i+z_i f i = w i + z i from the reach inequalities. Their nonnegative multipliers v i ( u ) v_i(u) v i ( u ) and the capacity multipliers λ e \lambda_e λ e give (14) and objective ∑ e λ e C e + ∑ i v i ( o i ) \sum_e\lambda_e C_e+\sum_i v_i(o_i) ∑ e λ e C e + ∑ i v i ( o i ) .
At fixed future values, (4) and (9) give (15). Stopping at u u u gives c e ( S , u ) c_e(S,u) c e ( S , u ) ; playing gives the other term in (16). Backward induction on each local DAG and then on episodes gives the smallest node values and identifies v i ( o i ) v_i(o_i) v i ( o i ) with U 1 , i λ U^\lambda_{1,i} U 1 , i λ at its initial state. Since c e ( S , u ) ≥ 0 c_e(S,u)\ge0 c e ( S , u ) ≥ 0 , stopping dominates discarding probability, which proves completion. Finite LP duality gives (13); complementary slackness supplies mixtures of maximizing arm programs that satisfy the capacities and attain the primal value.
The prices are optimized jointly across episodes. At fixed prices, backward induction assigns values to next-episode starting states; the local LP chooses their probability masses. A zero-play policy still receives reports due at the boundary and retains its future options. Every nonnegative price vector gives an upper bound through (13).
Proposition 2.10 (Size). Let V ∗ = ∑ i ∣ V i ∣ V_*=\sum_i|V_i| V ∗ = ∑ i ∣ V i ∣ and let E ∗ E_* E ∗ count the nonzero play and handoff transitions. Given their probabilities and reward coefficients, the LP has O ( V ∗ + N ) O(V_*+N) O ( V ∗ + N ) variables and constraints and O ( V ∗ + E ∗ + N ) O(V_*+E_*+N) O ( V ∗ + E ∗ + N ) nonzero coefficients. Fixed-price backward induction takes O ( V ∗ + E ∗ ) O(V_*+E_*) O ( V ∗ + E ∗ ) operations. For stationary priors, fixed finite reward alphabets, fixed caps and fixed report lag, a sufficient-state graph has size polynomial in N N N and the budgets.
Proof. There are three variables and a constant number of constraints per vertex, one capacity row per episode and one coefficient per transition. Backward induction visits each edge once. For an alphabet of size d i d_i d i , there are M i = ( L ˉ + d i d i ) M_i = \binom{\bar{L}+d_i}{d_i} M i = ( d i L ˉ + d i ) outcome histograms of at most L ˉ \bar{L} L ˉ plays. Write ℓ i ( P i ; s ) \ell_i(P_i; s) ℓ i ( P i ; s ) for the likelihood of any ordering of histogram s s s . If n s n_s n s completed episodes have histogram s s s , their likelihood is ∏ s ℓ i ( P i ; s ) n s \prod_s \ell_i(P_i; s)^{n_s} ∏ s ℓ i ( P i ; s ) n s . Thus at most ( N + M i M i ) \binom{N+M_i}{M_i} ( M i N + M i ) count vectors determine the stationary posterior. Record also the budget and the counts and delivery dates of at most g − 1 g-1 g − 1 pending batches. These contribute polynomially many states at fixed lag.
The bound concerns the supplied state graphs and their coefficients. Continuous observations may require an infinite graph or a specified finite approximation.
With one arm and C e ≥ min { L 1 , T e } C_e \ge\min\{L_1,T_e\} C e ≥ min { L 1 , T e } in every episode, the capacity constraints are redundant and λ e = 0 \lambda_e=0 λ e = 0 is optimal. The program gives the Bayes-optimal learner over the allowed local policies. Slow-arm policies are physically admissible; fast-arm policies use the delay relaxation unless feedback is undelayed.
Weighting use and generalization
The planner can value current use and later learning differently. Fix episode weights η e ≥ 0 \eta_e \ge0 η e ≥ 0 and bounded terminal values Ψ i ≥ 0 \Psi_i \ge0 Ψ i ≥ 0 on arm i i i ’s boundary states after episode N N N . The terminal value may represent deployment after learning. The weighted LP L P N η ( C , B ) \mathrm{LP}^{\eta}_N(C,B) LP N η ( C , B ) retains the final states t t t and maximizes, under the same constraints,
∑ i [ ∑ u ∈ V i η e ( u ) r i ( u ) w i ( u ) + ∑ t Ψ i ( t ) f i ( t ) ] (17) \sum_i \left[\sum_{u\in V_i}\eta_e(u)r_i(u)w_i(u)+\sum_t\Psi_i(t)f_i(t)\right] \tag*{(17)} i ∑ [ u ∈ V i ∑ η e ( u ) r i ( u ) w i ( u ) + t ∑ Ψ i ( t ) f i ( t ) ] (17)
Unit weights and zero terminal values recover (12) .
Proposition 2.11 (Weighted learning LP). For finite state graphs the dual has the form (13) , with reward coefficient η e r ( u ) \eta_e r(u) η e r ( u ) in episode e e e and terminal value U N + 1 = Ψ i U_{N+1}=\Psi_i U N + 1 = Ψ i for arm i i i .
Proof. Apply Theorem 2.9 with these coefficients and terminal values.
To separate early use from later learning, fix a boundary e 0 < N e_0<N e 0 < N . Let V ≤ e 0 V_{\le e_0} V ≤ e 0 and V > e 0 V_{>e_0} V > e 0 be the early and late values, including the terminal payoff in the latter. Write L P N η \mathrm{LP}^{\eta}_N LP N η for the scalar-weight objective V ≤ e 0 + η V > e 0 V_{\le e_0}+\eta V_{>e_0} V ≤ e 0 + η V > e 0 . Its value is convex and piecewise linear in η \eta η ; between breakpoints its slope is the selected program’s late value. Equivalently, η \eta η prices a feasible lower target v l a t e v_{\mathrm{late}} v late :
max { V ≤ e 0 : V > e 0 ≥ v l a t e } = min η ≥ 0 { L P N η ( C , B ) − η v l a t e } . (18) \max\{V_{\le e_0}:V_{>e_0}\ge v_{\mathrm{late}}\}=\min_{\eta\ge0}\{\mathrm{LP}^{\eta}_N(C,B)-\eta v_{\mathrm{late}}\}. \tag*{(18)} max { V ≤ e 0 : V > e 0 ≥ v late } = η ≥ 0 min { LP N η ( C , B ) − η v late } . (18)
Both sides use the original capacity and arm constraints.
For η > 0 \eta>0 η > 0 , the episode-e 0 e_0 e 0 local objective is R π − λ e 0 W π + η C F ( π ) R_\pi-\lambda_{e_0}W_\pi+\eta C_F(\pi) R π − λ e 0 W π + η C F ( π ) , with the late program valued at unit weight and prices λ e / η \lambda_e/\eta λ e / η for e > e 0 e>e_0 e > e 0 . Thus one experiment improves on another exactly when Δ R + η Δ C > λ e 0 Δ W \Delta R+\eta\Delta C>\lambda_{e_0}\Delta W Δ R + η Δ C > λ e 0 Δ W . The weight values the later decisions improved by the experiment; the capacity price charges for the plays it requires. At η = 0 \eta=0 η = 0 and e 0 = 1 e_0=1 e 0 = 1 , the objective values only the current episode.
A specified safe alternative. Some examples add an always-available action with deterministic reward a ∈ ( 0 , 1 ] a\in(0,1] a ∈ ( 0 , 1 ] per step and no binding play budget. It competes with reward plays; ordinary waiting still earns zero. In that extension, eliminating the safe action adds a ∑ e T e \sum_e T_e ∑ e T e to the objective and replaces each unknown-arm reward coefficient r i ( u ) r_i(u) r i ( u ) by r i ( u ) − a r_i(u)-a r i ( u ) − a . The total play price is then $a plus the nonnegative capacity multiplier, or a η e a\eta_e a η e plus that multiplier for weighted rewards. We use L P N \mathrm{LP}_N LP N and O P T N \mathrm{OPT}_N OPT N for the specified instance, including this action when present. The general execution and program statements below use zero-reward waiting; examples specifying a > 0 a > 0 a > 0 include this safe alternative. Thinning the arm programs leaves the safe action available, so Proposition 2.8(ii) still holds.
Example A.1 makes the distinction exact: the learning LP pays for a test that an episode planner never makes, avoiding a Θ ( N ) \Theta(N) Θ ( N ) loss. The information price is constant in that instance. Episode weights also price multi-play experiments and purchased reports through their later value (Examples A.2 and A.7 ).
Execution of weighted plans. For weights in [ 0 , 1 ] [0,1] [ 0 , 1 ] , Theorems 3.7 and 3.15 use E ∑ e η e V e \mathbb{E}\sum_e \eta_e V_e E ∑ e η e V e , with the planned value weighted alike. A terminal value depending on the retained state has its planned expectation whenever execution preserves that state’s law. If the sum of terminal values is bounded by Ψ ˉ \bar{\Psi} Ψ ˉ , a completion failure adds at most Ψ ˉ q f a i l \bar{\Psi}q_{\mathrm{fail}} Ψ ˉ q fail to the failure charge. The policy certificate of Theorem 5.6 uses the same weighted coefficients. The rate statements use unit weights.
The relaxation fits workloads only in expectation. The next section enforces the uniform hard horizon while preserving the learning transitions used to value future decisions.
From the LP to a policy
Planned restlessness lets us interleave completed arm programs without changing the learning states they pass to the next episode. Execution must fit those plays into the hard horizon. We separate current reward, continuation and resource use, then give two ways to realize the planned learning trajectory. One completes all prescribed plays using a reserve and charges failed completion. The other protects only the reports used by later programs and truncates reward-only tails. The single-episode bound quantifies that reward loss. All input arm programs are extracted from completed LP solutions as in Proposition 2.3 . Waiting earns zero throughout the execution statements.
Reward, continuation and resource use. Fix an episode e e e , the incoming states of all arms, a nonnegative future value function F = ∑ i F i F = \sum_i F_i F = ∑ i F i , where F i F_i F i values arm i i i ’s next state, and the planned local policies π = ( π e , i ) i \pi= (\pi_{e,i})_i π = ( π e , i ) i . Write
R π = E [ current reward ] , W π = E [ plays ] , C F ( π ) = E [ F ( S π ) ] R_\pi= \mathbb{E}[\text{current reward}], \qquad W_\pi= \mathbb{E}[\text{plays}], \qquad C_F(\pi) = \mathbb{E}[F(S^\pi)] R π = E [ current reward ] , W π = E [ plays ] , C F ( π ) = E [ F ( S π )]
for the episode as a whole: the reward of all arms, their plays, and the value of their next states. Let π ^ \hat{\pi} π ^ be the feasible joint policy produced by executing the planned local policies. Its expected reward and workload are defined as above. Its continuation is evaluated at the state retained by each arm’s program: the data it conditions on, its budget ledger and its report queue. Physical plays and budgets enter only the hard limits. An ( α e , β e ) (\alpha_e,\beta_e) ( α e , β e ) realization preserves reward and continuation if
R π ^ ≥ α e R π , C F ( π ^ ) ≥ β e C F ( π ) , 0 ≤ α e , β e ≤ 1. (19) R_{\hat{\pi}} \ge\alpha_e R_\pi,\qquad C_F(\hat{\pi}) \ge\beta_e C_F(\pi),\qquad0 \le\alpha_e,\beta_e \le1. \tag*{(19)} R π ^ ≥ α e R π , C F ( π ^ ) ≥ β e C F ( π ) , 0 ≤ α e , β e ≤ 1. (19)
If W π ^ ≤ ρ W π W_{\hat{\pi}} \le\rho W_\pi W π ^ ≤ ρ W π with ρ ≥ 0 \rho\ge0 ρ ≥ 0 , the coordinate bounds give, for $\eta,\lambda \ge 0,
R π ^ + η C F ( π ^ ) − λ W π ^ ≥ α e R π + η β e C F ( π ) − ρ λ W π . (20) R_{\hat{\pi}} + \eta C_F(\hat{\pi}) - \lambda W_{\hat{\pi}} \ge\alpha_e R_\pi+ \eta\beta_e C_F(\pi) - \rho\lambda W_\pi. \tag*{(20)} R π ^ + η C F ( π ^ ) − λ W π ^ ≥ α e R π + η β e C F ( π ) − ρ λ W π . (20)
For α e > 0 \alpha_e > 0 α e > 0 , factoring out α e \alpha_e α e reprices continuation at η β e / α e \eta\beta_e/\alpha_e η β e / α e and plays at ρ λ / α e \rho\lambda/\alpha_e ρ λ / α e . Thus reward and workload retention compare priced values at different prices. GM13’s exchangeability argument gives the following reward–workload bound before subtracting any price.
Lemma 3.1 (Reward truncation under any true prior). Let π \pi π be a nonanticipating single-arm policy, possibly designed using a prior P ^ \widehat{P} P . Couple it to a nonanticipating truncation π − \pi^{-} π − on the same episode quality, outcome tape and random coins. Write t P ( π ) = E P N π t^{P}(\pi)=\mathbb{E}_{P}N_{\pi} t P ( π ) = E P N π for expected play consumption. Suppose that, on every path, b N π ≤ N π − ≤ N π bN_{\pi}\le N_{\pi^{-}}\le N_{\pi} b N π ≤ N π − ≤ N π for a fixed b ∈ [ 0 , 1 ] b\in[0,1] b ∈ [ 0 , 1 ] . Then every true prior P P P on the quality space Θ i \Theta_i Θ i satisfies
R P ( π − ) ≥ b R P ( π ) , t P ( π − ) ≤ t P ( π ) . R^{P}(\pi^{-})\ge bR^{P}(\pi),\qquad t^{P}(\pi^{-})\le t^{P}(\pi). R P ( π − ) ≥ b R P ( π ) , t P ( π − ) ≤ t P ( π ) . The conclusion also holds conditional on external randomness independent of the arm, with b b b fixed by that randomness.
Proof. Fixing the quality, outcome tape and policy coins fixes both play counts. The hypothesis says that the truncated policy retains at least a fraction b b b of the original policy’s plays on every such run. For either policy σ \sigma σ , the decision to make play j j j uses only earlier outcomes and independent coins. Conditional on θ \theta θ , its reward Y j Y_j Y j therefore has mean μ i ( θ ) \mu_i(\theta) μ i ( θ ) independently of that decision. Summing over the possible plays gives
R P ( σ ) = ∑ j = 1 L i E P [ 1 { N σ ≥ j } Y j ] = E P [ μ i ( θ ) N σ ] . R^{P}(\sigma)=\sum_{j=1}^{L_i}\mathbb{E}_{P}\left[1\{N_{\sigma}\ge j\}Y_j\right]=\mathbb{E}_{P}\left[\mu_i(\theta)N_{\sigma}\right]. R P ( σ ) = j = 1 ∑ L i E P [ 1 { N σ ≥ j } Y j ] = E P [ μ i ( θ ) N σ ] .
Now multiply N π − ≥ b N π N_{\pi^{-}}\ge bN_{\pi} N π − ≥ b N π by μ i ( θ ) ≥ 0 \mu_i(\theta)\ge0 μ i ( θ ) ≥ 0 and take expectations. This proves the reward bound; N π − ≤ N π N_{\pi^{-}}\le N_{\pi} N π − ≤ N π gives the workload bound. The argument also holds conditional on independent external randomness and does not depend on the design prior.
The next-state law determines how episode reward guarantees compose. Let R e R_e R e and R ^ e \widehat{R}_e R e be the conditional expected rewards of the planned and realized episodes, and put J p l a n = ∑ e E R e J_{\mathrm{plan}}=\sum_e\mathbb{E}R_e J plan = ∑ e E R e . If execution reproduces the planned next-state law at every boundary, reward guarantees with fixed factors α e \alpha_e α e add:
E ∑ e R ^ e ≥ ∑ e α e E R e ≥ ( min e α e ) J p l a n . (21) \mathbb{E}\sum_e\widehat{R}_e\ge\sum_e\alpha_e\mathbb{E}R_e\ge\left(\min_e\alpha_e\right)J_{\mathrm{plan}}. \tag*{(21)} E e ∑ R e ≥ e ∑ α e E R e ≥ ( e min α e ) J plan . (21)
Matching the incoming-state laws lets us sum the episode expectations. The reward factor is therefore paid once. The constructions below preserve the planned states, or charge the event on which completion fails.
Delayed execution and the next learning state
Definition 3.2 (Executor). A map π ↦ π X \pi\mapsto\pi^{X} π ↦ π X is an executor with parameters ( ϵ , w ) (\epsilon,w) ( ϵ , w ) if, for every allowed local policy π \pi π making m m m plays:
It supplies π \pi π its outcomes in tape order, preserving the law of its decisions and planned data.
Its number X X X of plays satisfies m ≤ X ≤ min { cap , ( 1 + ϵ ) m } m\le X\le\min\{\operatorname{cap},(1+\epsilon)m\} m ≤ X ≤ min { cap , ( 1 + ϵ ) m } on every path.
After at most w w w of its own steps, each remaining own step is a play.
Its actions are determined by the arm’s tape, its coins and π \pi π . Interleaving changes only when those actions occur.
Direct execution is an executor with ( ϵ , w ) = ( 0 , 0 ) (\epsilon,w)=(0,0) ( ϵ , w ) = ( 0 , 0 ) for undelayed arms and for slow arms’ nonadaptive local policies. For delayed fast arms, buffered execution makes nonadaptive blocks, waits for their reports, and supplies outcomes to the planned policy in tape order. Extra outcomes are excluded from its planned data.
Definition 3.3 (Buffered execution). Let π \pi π be an undelayed policy for arm i i i and ϵ ∈ ( 0 , 1 ] \epsilon\in(0,1] ϵ ∈ ( 0 , 1 ] . The buffered executor A ϵ ( π ) A_{\epsilon}(\pi) A ϵ ( π ) makes actual plays and supplies their reported outcomes to π \pi π , at most once each and in play order.
Suppose π \pi π has consumed all q q q outcomes generated so far and requests another play. Make a nonadaptive block of
b = min { max { 1 , ⌈ ϵ q ⌉ } , L i − q } b = \min\{\max\{1,\lceil\epsilon q\rceil\},L_i-q\} b = min { max { 1 , ⌈ ϵ q ⌉} , L i − q } plays, one per step. Wait for their reports. Supply the outcomes to π \pi π one at a time until it stops, commits, or requests another block. If it commits inside a block, credit unused block plays toward its committed tail. Make any remaining tail plays without waiting for further reports. Extra plays beyond the original policy are excluded from its planned data.
For a general policy, switch modes once q ≥ ⌈ δ i / ϵ ⌉ q \ge\lceil\delta_i/\epsilon\rceil q ≥ ⌈ δ i / ϵ ⌉ . Make up to δ i \delta_i δ i further plays and enter the delay-free mode of Guha and Munagala [17 Lemma 19] . At each subsequent request, supply the oldest unused outcome. Also make a new play on that step if the cap has not been reached. Stop making new plays when π \pi π stops or the cap is exhausted.
Lemma 3.4 (Buffered executor, prefix bounds proved in Appendix § 8 .§ 1 ). A ϵ ( π ) A_{\epsilon}(\pi) A ϵ ( π ) reproduces the law of π \pi π ’s decision path and satisfies R ( A ϵ ( π ) ) ≥ R ( π ) R(A_{\epsilon}(\pi)) \ge R(\pi) R ( A ϵ ( π )) ≥ R ( π ) . On every path it makes at most min { L i , ( 1 + ϵ ) m } \min\{L_i,(1+\epsilon)m\} min { L i , ( 1 + ϵ ) m } plays, where m m m is the number of plays made by π \pi π .
For a commit policy, the prefix consists of its own play and wait steps before the committed plays. For a general policy, it also includes the plays that start delay-free mode. Its length is at most
W i = { δ i B ϵ ( k ) + ( 1 + ϵ ) k + 1 , depth- k commit , 5 ( δ i + 1 ) ( 1 + ln ( δ i + 1 ) ) / ϵ , general , W_i =
\begin{cases}
\delta_i B_{\epsilon}(k) + (1+\epsilon)k+1, & \text{depth-}k\text{ commit},\\
5(\delta_i+1)(1+\ln(\delta_i+1))/\epsilon, & \text{general},
\end{cases} W i = { δ i B ϵ ( k ) + ( 1 + ϵ ) k + 1 , 5 ( δ i + 1 ) ( 1 + ln ( δ i + 1 )) / ϵ , depth- k commit , general , where B ϵ ( k ) = min { k , ( 3 + 1.45 ln ( 1 + ϵ k ) ) / ϵ } B_{\epsilon}(k)=\min\{k,(3+1.45\ln(1+\epsilon k))/\epsilon\} B ϵ ( k ) = min { k , ( 3 + 1.45 ln ( 1 + ϵ k )) / ϵ } . After the prefix, each remaining own step is a play. Hence A ϵ A_{\epsilon} A ϵ is an executor with parameters ( ϵ , max i W i ) (\epsilon,\max_i W_i) ( ϵ , max i W i ) .
Proof. Couple the two policies on the same quality, tape and coins. Supplying outcomes in tape order reproduces π \pi π ’s decisions. Every supplied outcome comes from a physical play already made, so the executor makes at least the m m m plays requested by π \pi π . If π \pi π stops in a block started after q q q outcomes, it has requested at least q + 1 q+1 q + 1 plays and leaves at most b − 1 < ϵ q b-1 < \epsilon q b − 1 < ϵ q unused plays, unless b = 1 b=1 b = 1 , when there is no excess. At the switch to delay-free mode, the unused buffer is at most δ i ≤ ϵ q \delta_i \le\epsilon q δ i ≤ ϵ q . Unused block plays count toward a committed tail, and every block respects the cap. Thus m ≤ X ≤ min { L i , ( 1 + ϵ ) m } m \le X \le\min\{L_i,(1+\epsilon)m\} m ≤ X ≤ min { L i , ( 1 + ϵ ) m } . The latent-mean identity in Lemma 3.1 gives R ( A ϵ ( π ) ) ≥ R ( π ) R(A_{\epsilon}(\pi)) \ge R(\pi) R ( A ϵ ( π )) ≥ R ( π ) . The actions use only this arm’s tape, coins and own progress.
After the singleton blocks, each completed block grows the number of available outcomes by a factor of at least 1 + ϵ 1+\epsilon 1 + ϵ . Counting these blocks and their waits gives the stated prefix bounds; the calculation is in Appendix § 8 .§ 1 . After commitment or the switch to delay-free mode, each remaining own step is a play. □
Lemma 3.5 (Placement invariance). Fix arm i i i ’s incoming state in episode e e e . Couple its standalone and interleaved executions on the same qualities, tapes and coins, using the same executor. If the interleaved execution completes within the episode, both executions earn the same reward and retain the same state at the start of episode e + 1 e+1 e + 1 .
Proof. Condition iv preserves the executor’s ordered plays and supplied observations under interleaving. The same tape therefore gives the same rewards and retained data. Under the report calendar of Definition 2.2 , reports from earlier episodes arrive only at boundaries, and any reports still pending at the next boundary have delivery dates independent of placement. The retained belief, remaining budget and report queue consequently agree. The fixed within-episode reward law is essential (Example A.8 ). □
Fix an arm order for the episode. At each step, the priority scheduler plays the first executor requesting a play, and waits if none does. Report delays elapse while other arms are played.
Lemma 3.6 (Handoff under completion). Run arm programs ( σ i ) i (\sigma_i)_i ( σ i ) i through an executor with parameters ( ϵ , w ) (\epsilon,w) ( ϵ , w ) under the priority scheduler , in any fixed order. Couple this run with the n n n independent standalone runs on the same qualities, tapes and coins. Let X e , i X_{e,i} X e , i be arm i i i ’s standalone plays in episode e e e and S e = ∑ i X e , i S_e = \sum_i X_{e,i} S e = ∑ i X e , i . On the event
E ≤ e = { S u ≤ T − w for all u ≤ e } , \mathcal{E}_{\le e} = \{S_u \le T-w\text{ for all }u\le e\}, E ≤ e = { S u ≤ T − w for all u ≤ e } , the two runs make the same plays of unknown arms, earn the same rewards from them, and reach the same planned data, budgets and queues at the start of episode e + 1 e+1 e + 1 .
Proof. Induct on e e e . Suppose the runs agree at the start of episode e e e . Then they choose the same local policies. Index the arms in priority order and let M e , i M_{e,i} M e , i be the plays the scheduler makes of arm i i i . Plays of higher-priority arms do not depend on arm i i i , and there are at most ∑ j < i X e , j \sum_{j<i}X_{e,j} ∑ j < i X e , j of them. Arm i i i ’s waiting steps advance in calendar time, and a higher-priority play blocks it only when it requests a play. By condition iii, after at most w w w own steps, every unblocked step is a play until the policy ends. Hence
M e , i ≥ min { X e , i , ( T − w − ∑ j < i X e , j ) + } . (22) M_{e,i} \ge\min\left\{X_{e,i},\left(T-w-\sum_{j<i}X_{e,j}\right)^{+}\right\}. \tag*{(22)} M e , i ≥ min { X e , i , ( T − w − j < i ∑ X e , j ) + } . (22)
On S e ≤ T − w S_e \le T-w S e ≤ T − w the right side equals X e , i X_{e,i} X e , i for every i i i , whatever the order, so every executed policy completes. Lemma 3.5 then gives the same rewards and the same boundary states.
Completion preserves the standalone arm programs’ data, budgets and queues. In particular, extra executor plays earn reward, but their outcomes are ignored for learning: using them would change the transition law valued by the LP. The boundary-aligned calendar is essential to this handoff: a report from an earlier episode arriving partway through a local policy would insert a posterior update into its decision graph, which Lemma 3.5 does not cover.
The learner and its completion guarantee
Algorithm 1 (The learning-LP learner). Choose an executor for each arm: buffered execution A ϵ A_{\epsilon} A ϵ for delayed fast arms, and direct execution otherwise. Let w < T w<T w < T bound every prefix ( max i W i ≤ w ) (\max_i W_i\le w) ( max i W i ≤ w ) , with ϵ = w = 0 \epsilon=w=0 ϵ = w = 0 when no fast arm is delayed. Choose a reserve 0 ≤ s < T − w 0\le s<T-w 0 ≤ s < T − w and plan with
C = T − w − s 1 + ϵ , B i ′ = { B i , B i ≤ 1 / ϵ , ⌊ B i / ( 1 + ϵ ) ⌋ , B i > 1 / ϵ , C=\frac{T-w-s}{1+\epsilon},\qquad
B'_i=
\begin{cases}
B_i, & B_i\le1/\epsilon,\\
\lfloor B_i/(1+\epsilon)\rfloor, & B_i>1/\epsilon,
\end{cases} C = 1 + ϵ T − w − s , B i ′ = { B i , ⌊ B i / ( 1 + ϵ )⌋ , B i ≤ 1/ ϵ , B i > 1/ ϵ ,
with B ′ = B B'=B B ′ = B when ϵ = 0 \epsilon=0 ϵ = 0 .
Complete a feasible solution of L P N ( C , B ′ ) \mathrm{LP}_N(C,B') LP N ( C , B ′ ) and extract its arm programs ( σ i ) i (\sigma_i)_i ( σ i ) i . Their objective F F F is the value certificate .
In episode e e e , let π e , i \pi_{e,i} π e , i be the local policy of σ i \sigma_i σ i given arm i i i ’s planned history. Execute it with the chosen executor and cap { min [ L i , remaining budget ] } \{\min[L_i,\text{remaining budget}]\} { min [ L i , remaining budget ]} . Interleave all arms with the priority scheduler in any fixed order. No admission coins are needed: the reduced capacity already reserves time.
At each boundary, update σ i \sigma_i σ i using only the reported outcomes of π e , i \pi_{e,i} π e , i ’s own plays. Pending outcomes stay in the queue. Extra executor plays earn reward, but their outcomes are excluded from the planned data.
If some local policy does not complete, the prediction fails. Switch to any admissible recovery for later episodes, built from the actual reported data, budgets and queues. Idling is admissible; so is re-solving the learning LP from the actual state.
Theorem 3.7 (Self-verification). Under Assumption 2.1 , Algorithm 1 respects every cap and budget on every path. Let its arm programs have objective F F F . Run them standalone through the executor, let X e , i X_{e,i} X e , i be arm i i i ’s workload in episode e e e , and let E = E ≤ N \mathcal{E} = \mathcal{E}_{\leq N} E = E ≤ N be the completion event of Lemma 3.6 . Any certified bound q f a i l ≥ Pr ( E c ) q_{\mathrm{fail}} \geq\Pr(\mathcal{E}^{c}) q fail ≥ Pr ( E c ) on a failed prediction gives
E ∑ e V e ≥ F − n L ‾ N q f a i l . \mathbb{E}\sum_{e} V_e \geq F - n\overline{L}Nq_{\mathrm{fail}}. E e ∑ V e ≥ F − n L N q fail . This holds for every admissible recovery after a failed prediction. If no arm has positive cap, the learner waits and F = 0 F=0 F = 0 .
For a uniform completion certificate, put Λ e x e c = log ( n L ‾ N 2 ) \Lambda_{\mathrm{exec}} = \log(n\overline{L}N^2) Λ exec = log ( n L N 2 ) for n L ‾ > 0 n\overline{L}>0 n L > 0 and choose
s = 2 L ‾ T Λ e x e c + 2 3 L ‾ Λ e x e c < T − w . s = \sqrt{2\overline{L}T\Lambda_{\mathrm{exec}}}+\frac{2}{3}\overline{L}\Lambda_{\mathrm{exec}} < T-w. s = 2 L T Λ exec + 3 2 L Λ exec < T − w . Then the failure charge is at most 1, and
E ∑ e V e ≥ F − 1. \mathbb{E}\sum_{e} V_e \geq F-1. E e ∑ V e ≥ F − 1. For an optimal solution, put ω = 1 − C / T = ( ( w + s ) / T + ϵ ) / ( 1 + ϵ ) \omega= 1-C/T = ((w+s)/T+\epsilon)/(1+\epsilon) ω = 1 − C / T = (( w + s ) / T + ϵ ) / ( 1 + ϵ ) . Then
E ∑ e V e ≥ ( 1 − ω ) OPT N ( B ′ ) − 1. \mathbb{E}\sum_{e} V_e \geq(1-\omega)\operatorname{OPT}_{N}(B')-1. E e ∑ V e ≥ ( 1 − ω ) OPT N ( B ′ ) − 1.
Proof. Direct execution uses the planned plays. Buffered execution inflates their count by at most 1 + ϵ 1+\epsilon 1 + ϵ . For B i > 1 / ϵ B_i>1/\epsilon B i > 1/ ϵ , ( 1 + ϵ ) B i ′ ≤ B i (1+\epsilon)B'_i\leq B_i ( 1 + ϵ ) B i ′ ≤ B i ; for B i ≤ 1 / ϵ B_i\leq1/\epsilon B i ≤ 1/ ϵ , it uses singleton blocks and makes exactly the planned plays. These budget bounds hold on every path. The executor respects each local cap, and recovery respects the actual remaining resources. Run each arm program standalone on its own qualities, tape and coins. These runs are independent across arms and retain the planned data law. Write V e , i o V^o_{e,i} V e , i o for their rewards and S e = ∑ i X e , i S_e=\sum_i X_{e,i} S e = ∑ i X e , i for their physical plays. Then
E ∑ i V e , i o ≥ ∑ i R e ( σ i ) , E S e ≤ ( 1 + ϵ ) ∑ i t e ( σ i ) ≤ ( 1 + ϵ ) C . \mathbb{E}\sum_i V^o_{e,i}\geq\sum_i R_e(\sigma_i),\qquad\mathbb{E}S_e\leq(1+\epsilon)\sum_i t_e(\sigma_i)\leq(1+\epsilon)C. E i ∑ V e , i o ≥ i ∑ R e ( σ i ) , E S e ≤ ( 1 + ϵ ) i ∑ t e ( σ i ) ≤ ( 1 + ϵ ) C .
Lemma 3.6 couples the scheduled run to these runs on E \mathcal{E} E . Put V e o = ∑ i V e , i o V^o_e=\sum_i V^o_{e,i} V e o = ∑ i V e , i o . Its expectation sums to at least F F F , and V e o ≤ S e ≤ n L ‾ V^o_e\leq S_e\leq n\overline{L} V e o ≤ S e ≤ n L . The scheduled reward agrees with it on E \mathcal{E} E and is nonnegative off it. Hence
E ∑ e V e ≥ F − n L ‾ N Pr ( E c ) . \mathbb{E}\sum_e V_e\geq F-n\overline{L}N\Pr(\mathcal{E}^{c}). E e ∑ V e ≥ F − n L N Pr ( E c ) .
This proves the certificate bound for every admissible recovery.
For each episode E S e ≤ ( 1 + ϵ ) C = T − w − s \mathbb{E}S_e\leq(1+\epsilon)C=T-w-s E S e ≤ ( 1 + ϵ ) C = T − w − s . Put μ e = E S e ≤ T − w − s \mu_e=\mathbb{E}S_e\leq T-w-s μ e = E S e ≤ T − w − s . The standalone workloads are independent across arms, with ∑ i Var ( X e , i ) ≤ L ‾ μ e \sum_i\operatorname{Var}(X_{e,i})\leq\overline{L}\mu_e ∑ i Var ( X e , i ) ≤ L μ e . Bernstein’s inequality gives
Pr ( S e > T − w ) ≤ exp { − s 2 2 L ‾ μ e + ( 2 / 3 ) L ‾ s } ≤ e − Λ e x e c . \Pr(S_e>T-w)\leq\exp\left\{-\frac{s^2}{2\overline{L}\mu_e+(2/3)\overline{L}s}\right\}\leq e^{-\Lambda_{\mathrm{exec}}}. Pr ( S e > T − w ) ≤ exp { − 2 L μ e + ( 2/3 ) L s s 2 } ≤ e − Λ exec .
A union bound yields Pr ( E c ) ≤ N e − Λ e x e c \Pr(\mathcal{E}^{c})\leq Ne^{-\Lambda_{\mathrm{exec}}} Pr ( E c ) ≤ N e − Λ exec and failure charge at most 1. Finally, F = L P N ( C , B ′ ) ≥ ( C / T ) OPT N ( B ′ ) F=LP_N(C,B')\geq(C/T)\operatorname{OPT}_N(B') F = L P N ( C , B ′ ) ≥ ( C / T ) OPT N ( B ′ ) by Proposition 2.8, C / T = 1 − ω ≥ 1 − ( w + s ) / T − ϵ C/T=1-\omega\geq1-(w+s)/T-\epsilon C / T = 1 − ω ≥ 1 − ( w + s ) / T − ϵ . These give the optimal-solution bound. □ \square □
For L ‾ = T / X \overline{L}=T/X L = T / X , the capacity loss tends to zero when Λ e x e c = o ( X ) \Lambda_{\mathrm{exec}}=o(X) Λ exec = o ( X ) , w = o ( T ) w=o(T) w = o ( T ) and ϵ = o ( 1 ) \epsilon=o(1) ϵ = o ( 1 ) . The following lower bound shows that the leading reserve is unavoidable in the worst case when every prescribed experiment must complete.
A matching lower bound for full completion.
Theorem 3.8 (Necessary reserve for full completion, proved in Appendix B.2). Fix integers m ≥ 2 m \ge2 m ≥ 2 , n ≥ m n \ge m n ≥ m and N ≥ 1 N \ge1 N ≥ 1 . There are n n n independent undelayed one-shot arms with cap L ˉ = L = m 2 \bar{L}=L=m^2 L ˉ = L = m 2 , known stationary priors and nonbinding budgets B i = N L B_i=NL B i = N L . For planning capacity C = n m C=nm C = nm and episode horizon T = C + s T=C+s T = C + s , where s ≥ 0 s \ge0 s ≥ 0 is an integer, every optimal solution of L P N ( C , B ) \mathrm{LPN}(C,B) LPN ( C , B ) prescribes, almost surely, one probe of each arm and continuation to its cap after success, stopping after failure.
Let s min s_{\min} s m i n be the smallest integer reserve for which all these policies can complete in all N N N episodes with probability at least 1 − δ 1-\delta 1 − δ , where 0 < δ ≤ 1 / 2 0<\delta\le1/2 0 < δ ≤ 1/2 . Put
h = log ( N / δ ) , v = n m ( m + 1 ) 2 . h=\log(N/\delta), \qquad v=\frac{nm}{(m+1)^2}. h = log ( N / δ ) , v = ( m + 1 ) 2 nm . Along any sequence with h → ∞ h\to\infty h → ∞ and h = o ( v ) h=o(v) h = o ( v ) ,
s min = ( 1 + o ( 1 ) ) ( m − 1 ) 2 n m log ( N / δ ) . (23) s_{\min}=(1+o(1))(m-1)\sqrt{2nm\log(N/\delta)}. \tag*{(23)} s m i n = ( 1 + o ( 1 )) ( m − 1 ) 2 nm log ( N / δ ) . (23) At T = C + s min T=C+s_{\min} T = C + s m i n this is
s min = ( 1 + o ( 1 ) ) ( 1 − 1 L ) 2 L T log ( N / δ ) . s_{\min}=(1+o(1))\left(1-\frac{1}{\sqrt{L}}\right)\sqrt{2LT\log(N/\delta)}. s m i n = ( 1 + o ( 1 )) ( 1 − L 1 ) 2 L T log ( N / δ ) . In particular, if also m → ∞ m\to\infty m → ∞ , the leading constant is 2 \sqrt{2} 2 , matching the uniform reserve of Theorem 3.7.
For the failure budget δ = 1 / ( n L ˉ N ) \delta=1/(n\bar{L}N) δ = 1/ ( n L ˉ N ) used above, log ( N / δ ) = Λ e x e c \log(N/\delta)=\Lambda_{\mathrm{exec}} log ( N / δ ) = Λ exec . Along this construction, when Λ e x e c = o ( T / L ˉ ) \Lambda_{\mathrm{exec}}=o(T/\bar{L}) Λ exec = o ( T / L ˉ ) , the lower bound is 1 − 1 / L ˉ 1-1/\sqrt{\bar{L}} 1 − 1/ L ˉ times the leading uniform reserve, up to a 1 + o ( 1 ) 1+o(1) 1 + o ( 1 ) factor. The leading constants match when also L ˉ → ∞ \bar{L}\to\infty L ˉ → ∞ .
The lower-bound construction also admits a feasible policy that collects every informative observation and approaches the LP value without a reserve. Each arm’s episode mean is 1 1 1 with probability p = 1 / ( m + 1 ) p=1/(m+1) p = 1/ ( m + 1 ) and 0 0 0 otherwise. The first play reveals it; later plays supply no further information. If Z ∼ Binomial ( n , p ) Z\sim\operatorname{Binomial}(n,p) Z ∼ Binomial ( n , p ) , the prescribed work is S = n + ( L − 1 ) Z S=n+(L-1)Z S = n + ( L − 1 ) Z , with E S = C = n m \mathbb{E}S=C=nm E S = C = nm , and the LP’s per-episode value is exactly n p L npL n p L , at both capacities C C C and T = C + s T=C+s T = C + s . Probe all n n n arms first, then spend the remaining steps on successful arms. This feasible policy earns
V p r o b e = n p L − E ( S − T ) + . V_{\mathrm{probe}}=npL-\mathbb{E}(S-T)^+. V probe = n p L − E ( S − T ) + .
Even with no reserve, Var ( S ) = ( m − 1 ) 2 n m \operatorname{Var}(S)=(m-1)^2nm Var ( S ) = ( m − 1 ) 2 nm gives
V p r o b e n p L ≥ 1 − 1 − m − 2 2 m n . \frac{V_{\mathrm{probe}}}{npL}\ge1-\frac{1-m^{-2}}{2}\sqrt{\frac{m}{n}}. n p L V probe ≥ 1 − 2 1 − m − 2 n m .
because E ( S − C ) + = 1 2 E ∣ S − C ∣ ≤ 1 2 Var ( S ) \mathbb{E}(S-C)^+=\frac{1}{2}\mathbb{E}|S-C|\le\frac{1}{2}\sqrt{\operatorname{Var}(S)} E ( S − C ) + = 2 1 E ∣ S − C ∣ ≤ 2 1 Var ( S ) . The lower-bound regime forces n / m → ∞ n/m\to\infty n / m → ∞ , so this ratio tends to 1 1 1 . Protecting the first reports preserves all the information while allowing reward-only tails to be truncated. We next bound that reward loss, then extend this separation to learning across episodes.
Single-episode execution and the effect of caps
The remaining ingredient bounds reward lost at the horizon when continuation has already been protected. For known, independent recurring priors P = ( P 1 , … , P n ) P=(P_1,\ldots,P_n) P = ( P 1 , … , P n ) , write O P T δ ( P ) \mathrm{OPT}_{\delta}(P) OPT δ ( P ) and O P T ( P ) \mathrm{OPT}(P) OPT ( P ) for the delayed and undelayed single-episode optima.
For an undelayed single-arm policy π i \pi_i π i of cap L i L_i L i , write R ( π i ) R(\pi_i) R ( π i ) and t ( π i ) t(\pi_i) t ( π i ) for its expected reward and plays. GM13’s relaxation is
L P ( P ) = max ( π i ) ∑ i R ( π i ) subject to ∑ i t ( π i ) ≤ T . \mathrm{LP}(P)=\max_{(\pi_i)}\sum_i R(\pi_i)
\qquad\text{subject to}\qquad\sum_i t(\pi_i)\le T. LP ( P ) = ( π i ) max i ∑ R ( π i ) subject to i ∑ t ( π i ) ≤ T .
Every local cap holds on every path. Shared time holds only in expectation. The local policies may require more than T T T plays on a realized joint path. Write LP ( k ) ( P ) \operatorname{LP}^{(k)}(P) LP ( k ) ( P ) for this relaxation restricted to depth-k k k commit policies.
Projecting a feasible policy onto its arms, as in Proposition 2.8(i), and relaxing all within-episode delays gives
OPT δ ( P ) ≤ LP ( P ) , OPT ( P ) ≤ LP ( P ) . (24) \operatorname{OPT}_{\delta}(P) \le\operatorname{LP}(P), \qquad\operatorname{OPT}(P) \le\operatorname{LP}(P). \tag*{(24)} OPT δ ( P ) ≤ LP ( P ) , OPT ( P ) ≤ LP ( P ) . (24)
The projection preserves reward, workload and caps and ignores outcomes until the original policy would receive them [17 Claim 1] .
Theorem 3.9 (Cap-sensitive undelayed execution, proved in Appendix § 8 .§ 4 ). Fix local policies with independent coins, caps L i ≤ T / X L_i \le T/X L i ≤ T / X for X ≥ 1 X \ge1 X ≥ 1 , and total expected workload at most T T T . Let their value be J = ∑ i R ( π i ) J = \sum_i R(\pi_i) J = ∑ i R ( π i ) . Execute them in decreasing order of expected reward per expected play, and idle at unused steps. Truncate at the horizon. For integer X X X , let Z X ∼ Poisson ( X ) Z_X \sim\operatorname{Poisson}(X) Z X ∼ Poisson ( X ) . This undelayed policy earns at least c X J c_XJ c X J , where
c X = max { 1 2 , 1 − Pr ( Z X = X ) − Pr ( Z X ≥ X ) 2 X } , c_X = \max\left\{\frac{1}{2}, 1-\Pr(Z_X=X)-\frac{\Pr(Z_X\ge X)}{2X}\right\}, c X = max { 2 1 , 1 − Pr ( Z X = X ) − 2 X Pr ( Z X ≥ X ) } , and
1 / c X = 1 + ( 2 π X ) − 1 / 2 + O ( X − 1 ) . 1/c_X = 1+(2\pi X)^{-1/2}+O(X^{-1}). 1/ c X = 1 + ( 2 π X ) − 1/2 + O ( X − 1 ) . With the safe action of § 3.6 , fill unused steps with it: its expected reward is at least a ( T − ∑ i t ( π i ) ) a(T-\sum_i t(\pi_i)) a ( T − ∑ i t ( π i )) , so the total is at least c X ∑ i R ( π i ) + a ( T − ∑ i t ( π i ) ) ≥ c X [ a T + ∑ i ( R ( π i ) − a t ( π i ) ) ] c_X\sum_i R(\pi_i)+a(T-\sum_i t(\pi_i)) \ge c_X[aT+\sum_i(R(\pi_i)-at(\pi_i))] c X ∑ i R ( π i ) + a ( T − ∑ i t ( π i )) ≥ c X [ a T + ∑ i ( R ( π i ) − a t ( π i ))] , since c X ≤ 1 c_X \le1 c X ≤ 1 .
The order is fixed before execution. For any arm, the time consumed by higher-priority arms is independent of its tape. The scheduling bound (22) and Lemma 3.1 turn the available time into a reward guarantee. The total workload is a sum of independent variables bounded by their caps. A Poisson comparison bounds its excess and gives c X c_X c X ; the calculation is in Appendix § 8 .§ 4 . Thus smaller caps improve retention of reward that may be cut at the horizon.
Corollary 3.10 (Delayed execution, proved in Appendix § 8 .§ 4 ). Take the feasible local policies of Theorem 3.9 . Choose ϵ ∈ ( 0 , 1 ] \epsilon\in(0,1] ϵ ∈ ( 0 , 1 ] and γ ∈ ( 0 , 1 ) \gamma\in(0,1) γ ∈ ( 0 , 1 ) with max i W i ≤ γ T \max_i W_i \le\gamma T max i W i ≤ γ T . Admit each transformed policy independently with probability ( 1 − γ ) / ( 1 + ϵ ) (1-\gamma)/(1+\epsilon) ( 1 − γ ) / ( 1 + ϵ ) . Order admitted policies by expected reward per expected play. At each step, run the highest-priority policy requesting a play, otherwise idle. The policy earns at least κ J \kappa J κ J , where
κ = 1 − γ 1 + ϵ c X , γ . \kappa=\frac{1-\gamma}{1+\epsilon}c_{X,\gamma}. κ = 1 + ϵ 1 − γ c X , γ . The finite-cap factor c X , γ c_{X,\gamma} c X , γ is given in (49) in Appendix § 8 .§ 4 . It is at least ( 1 − γ ) / 2 (1-\gamma)/2 ( 1 − γ ) /2 . For fixed integer X X X , c X , γ → c X c_{X,\gamma}\to c_X c X , γ → c X as γ → 0 \gamma\to0 γ → 0 .
When no report arrives within an episode, local policies are nonadaptive. Example A.4 gives the exact planning solution and a T / 16 T/16 T /16 loss for sampled-mean Thompson sampling even with the prior known.
GM13 retains half the LP value for general caps. Small caps concentrate the independent workloads and improve the factor. The bound applies to every feasible solution. For noninteger X X X , use c X , 0 c_{X,0} c X , 0 from (49). The leading square-root loss is necessary:
Proposition 3.11 (The square-root loss is necessary, proved in Appendix § 8 .§ 4 ). For every integer X ≥ 1 X \ge1 X ≥ 1 , there are undelayed one-shot instances with L i = T / X L_i=T/X L i = T / X such that
OPT ( P ) LP ( P ) ≤ 1 − Pr { Poisson ( X ) = X } + o ( 1 ) \frac{\operatorname{OPT}(P)}{\operatorname{LP}(P)} \le1-\Pr\{\operatorname{Poisson}(X)=X\}+o(1) LP ( P ) OPT ( P ) ≤ 1 − Pr { Poisson ( X ) = X } + o ( 1 ) as T → ∞ T\to\infty T → ∞ through multiples of X X X .
Proposition 3.12 (The factor 2 is tight, proved in Appendix § 8 .§ 4 ). For every T ≥ 2 T \ge2 T ≥ 2 , there is an undelayed one-shot instance with L i = T L_i=T L i = T such that OPT ( P ) / LP ( P ) ≤ 1 / 2 + 1 / T \operatorname{OPT}(P)/\operatorname{LP}(P) \le1/2+1/T OPT ( P ) / LP ( P ) ≤ 1/2 + 1/ T .
At X = 1 X=1 X = 1 the factor 1 / 2 1/2 1/2 is therefore exact for this relaxation. In Proposition 3.11 the loss comes from successes competing for time. Here it comes from the time spent finding one, which the single time constraint does not see.
If every unknown arm is fast with delay at most δ \delta δ , put D δ = ( δ + 1 ) ( 1 + ln ( δ + 1 ) ) D_\delta=(\delta+1)(1+\ln(\delta+1)) D δ = ( δ + 1 ) ( 1 + ln ( δ + 1 )) . The additional delay loss is O ( D δ / T ) O(\sqrt{D_\delta/T}) O ( D δ / T ) at fixed X X X , and the factor for arbitrary caps tends to 2 2 2 when δ log δ = o ( T ) \delta\log\delta=o(T) δ log δ = o ( T ) (Corollary B.5 ). Appendix § 8 .§ 4 gives the parameter calculation and the direct comparison with GM13. These bounds price reward truncation. Full completion in Theorem 3.7 instead pays the reserve and executor overhead.
Several plays per step. With up to K ≥ 1 K \ge1 K ≥ 1 distinct arms played per step, each arm still makes at most one play per step. Write LP K ( P ) \operatorname{LP}_K(P) LP K ( P ) for the same relaxation with total expected workload at most K T KT K T ; projecting a feasible policy gives OPT K ( P ) ≤ LP K ( P ) \operatorname{OPT}_K(P) \le\operatorname{LP}_K(P) OPT K ( P ) ≤ LP K ( P ) .
Theorem 3.13 (Fixed-order execution with K K K plays per step, proved in Appendix § 8 .§ 5 ). Fix undelayed local policies with independent coins, caps L i ≤ L ≤ T L_i \le L \le T L i ≤ L ≤ T , value J = ∑ i R ( π i ) J=\sum_i R(\pi_i) J = ∑ i R ( π i ) and expected workload W = ∑ i t ( π i ) ≤ K T W=\sum_i t(\pi_i)\le KT W = ∑ i t ( π i ) ≤ K T . Order them by decreasing expected reward per expected play. At each step play the at most K K K highest-priority policies that have not stopped, and truncate at T T T . With ν = T / L \nu=T/L ν = T / L , the expected reward is at least
max { 1 − W 2 K T , ρ P , K ( ν ) } J . \max\left\{1-\frac{W}{2KT},\rho_{P,K}(\nu)\right\}J. max { 1 − 2 K T W , ρ P , K ( ν ) } J . where, for Z ∼ Poisson ( K ν ) Z\sim\operatorname{Poisson}(K\nu) Z ∼ Poisson ( K ν ) ,
ρ P , K ( ν ) = 1 − E H ν ( Z / K ) ν , H ν ( x ) = ∫ 0 x min { 1 , ( y − ν + 1 ) + } d y . (25) \rho_{P,K}(\nu)=1-\frac{\mathbb{E}H_\nu(Z/K)}{\nu},\qquad H_\nu(x)=\int_0^x\min\{1,(y-\nu+1)^+\}\,\mathrm{d}y. \tag*{(25)} ρ P , K ( ν ) = 1 − ν E H ν ( Z / K ) , H ν ( x ) = ∫ 0 x min { 1 , ( y − ν + 1 ) + } d y . (25) For K = 1 K=1 K = 1 and integer ν = X \nu=X ν = X , the uniform factor is c X c_X c X .
The uniform factor max { 1 / 2 , ρ P , K ( ν ) } \max\{1/2,\rho_{P,K}(\nu)\} max { 1/2 , ρ P , K ( ν )} is nondecreasing in K K K and tends to 1 − L / ( 2 T ) 1-L/(2T) 1 − L / ( 2 T ) . Example B.6 shows that this limit is tight relative to LP K \operatorname{LP}_K LP K ; at L = T L=T L = T , the worst-case factor is 1 / 2 1/2 1/2 for every K K K . These are undelayed single-episode bounds.
Predicting less
We now preserve the planned learning transitions on every path. Each arm program learns from a designated prefix of its episode’s plays. Execution completes these plays in a fixed window and may truncate the remaining reward plays. The program ignores their outcomes and charges their planned count to its budget ledger, so truncation does not change the state from which it continues.
Definition 3.14 (Prediction-limited local policy). Fix an integer r ≥ 0 r \ge0 r ≥ 0 . A local policy π \pi π is prediction-limited at depth r r r if it commits to its remaining play count after at most r r r adaptive plays; for a slow arm, the count is fixed at the episode’s start. If N π N_\pi N π is its total play count, the first min { N π , r } \min\{N_\pi,r\} min { N π , r } plays are predicted and the rest are unpredicted . Its next boundary state uses predicted reports only, and its budget ledger deducts every planned play.
Fix depths r e , i r_{e,i} r e , i in advance. A prediction-limited arm program chooses its local policies from this class using its own retained reports. Choose executors with parameters ( ϵ , w ) (\epsilon,w) ( ϵ , w ) and comparison budgets B ′ B' B ′ as in Algorithm 1. Choose a window T p < T T_p<T T p < T satisfying w + ( 1 + ϵ ) ∑ i r e , i ≤ T p w+(1+\epsilon)\sum_i r_{e,i}\le T_p w + ( 1 + ϵ ) ∑ i r e , i ≤ T p in every episode, and put C 2 = T − T p C_2=T-T_p C 2 = T − T p . The restricted learning LP is
LP N [ r ] ( C 2 , B ) = sup ( σ i ) e , i ∑ e , i R e ( σ i ) s.t. ∑ i t i u ( σ i ) ≤ C 2 ( e = 1 , … , N ) . \operatorname{LP}^{[r]}_N(C_2,B)=\sup_{(\sigma_i)_{e,i}}\sum_{e,i}R_e(\sigma_i)\quad\text{s.t.}\quad\sum_i t_i^u(\sigma_i)\le C_2\quad(e=1,\ldots,N). LP N [ r ] ( C 2 , B ) = ( σ i ) e , i sup e , i ∑ R e ( σ i ) s.t. i ∑ t i u ( σ i ) ≤ C 2 ( e = 1 , … , N ) .
over these arm programs with budgets B B B . Here t e u ( σ i ) t^{\mathrm{u}}_e(\sigma_i) t e u ( σ i ) is arm i i i ’s expected number of unpredicted plays in episode e e e . The window accommodates the predicted plays and executor overhead on every path; the constraint limits expected unpredicted work to the remaining C 2 C_2 C 2 steps.
Each commitment specifies an expected reward, a workload and a distribution of retained next states. Its probability weights these quantities linearly. Pricing the workload constraint at λ e \lambda_e λ e separates the arms: each local program maximizes current reward plus continuation, less λ e \lambda_e λ e times its unpredicted workload.
Algorithm 2 (Predict less). Use the depths, window, executors and comparison budgets chosen above.
Complete a feasible solution of L P N [ r ] ( C 2 , B ′ ) \mathrm{LP}^{[\mathrm{r}]}_N(C_2,B') LP N [ r ] ( C 2 , B ′ ) and extract its arm programs ( σ i ) i (\sigma_i)_i ( σ i ) i . Let F F F be their objective.
Phase 1 , steps 1 to T p T_{\mathrm{p}} T p . Run the predicted plays through their executors under the priority scheduler. Wait when no executor requests a play.
Phase 2 , steps T p + 1 T_{\mathrm{p}}+1 T p + 1 to T T T . Run the remaining unpredicted plays arm by arm, in an order fixed before the run by decreasing expected reward per expected play of their standalone blocks. Wait when all blocks have ended, and truncate at T T T .
At each boundary, update each arm program from predicted reports, retain those still pending, and deduct every planned play from its budget ledger.
Buffered execution may make unpredicted plays during phase 1; phase 2 plays the rest.
Theorem 3.15 (Predict less). Under Assumption 2.1, Algorithm 2 respects every cap and budget on every path and preserves the planned law of the retained learning states. Every predicted play is made in phase 1. Put X 2 = C 2 / L ‾ ≥ 1 X_2=C_2/\overline{L}\ge1 X 2 = C 2 / L ≥ 1 and c = c X 2 , 0 c=c_{X_2,0} c = c X 2 , 0 . Then
E ∑ e V e ≥ F − ( 1 − c ) ∑ e , i R e u ( σ i ) ≥ c F , \mathbb{E}\sum_e V_e \ge F-(1-c)\sum_{e,i}R^{\mathrm{u}}_e(\sigma_i)\ge cF, E e ∑ V e ≥ F − ( 1 − c ) e , i ∑ R e u ( σ i ) ≥ c F , where R e u ( σ i ) R^{\mathrm{u}}_e(\sigma_i) R e u ( σ i ) is the expected reward of arm i i i ’s unpredicted plays in episode e e e .
Proof. Let X e , i p X^{\mathrm{p}}_{e,i} X e , i p be arm i i i ’s standalone physical plays in phase 1. A block is requested only before r e , i r_{e,i} r e , i outcomes have been supplied, so X e , i p ≤ ( 1 + ϵ ) r e , i X^{\mathrm{p}}_{e,i}\le(1+\epsilon)r_{e,i} X e , i p ≤ ( 1 + ϵ ) r e , i . The window condition and (22) therefore make every predicted prefix complete by T p T_{\mathrm{p}} T p on every path. Couple to the standalone prefixes on the same qualities, tapes and coins. Lemma 3.5 gives the same predicted reports, committed tail counts and pending queue. The ledger deducts every planned play, including truncated ones, so the retained states have the planned law in every episode. The physical budget suffices by the executor allowance in B ′ B' B ′ , and every executor respects its cap.
The remaining tails are nonadaptive capped blocks, independent across arms in their standalone runs, with total expected work at most C 2 C_2 C 2 . The single-episode execution bound retains at least c c c of their expected reward. Unpredicted plays made early by the executor are all kept. Write J e = ∑ i ( R e , i p + R e , i u ) J_e=\sum_i(R^{\mathrm{p}}_{e,i}+R^{\mathrm{u}}_{e,i}) J e = ∑ i ( R e , i p + R e , i u ) for the planned episode reward. Predicted plays retain their reward, hence
E V e ≥ J e − ( 1 − c ) ∑ i R e , i u . \mathbb{E}V_e\ge J_e-(1-c)\sum_i R^{\mathrm{u}}_{e,i}. E V e ≥ J e − ( 1 − c ) i ∑ R e , i u . The retained states have the planned law, so ∑ e J e = F \sum_e J_e=F ∑ e J e = F . Summing and using ∑ e , i R e , i u ≤ F \sum_{e,i}R^{\mathrm{u}}_{e,i}\le F ∑ e , i R e , i u ≤ F proves the claim. Continuation is preserved throughout; only unpredicted reward pays the single-episode factor. □ \square □
For one-shot arms, the first report contains all the information available from that arm in the episode.
Corollary 3.16 (One-shot arms). Suppose every episode mean lies in { 0 , 1 } \{0,1\} { 0 , 1 } . Take r e , i = 1 r_{e,i}=1 r e , i = 1 and T p = w + ⌈ ( 1 + ϵ ) n ⌉ T_{\mathrm p}=w+\lceil(1+\epsilon)n\rceil T p = w + ⌈( 1 + ϵ ) n ⌉ with ( T − T p ) / L ˉ ≥ 1 (T-T_{\mathrm p})/\bar L\ge1 ( T − T p ) / L ˉ ≥ 1 , and use an optimal restricted LP solution. Then
E ∑ e V e ≥ c ( 1 − T p T ) OPT N ( B ′ ) . \mathbb{E}\sum_{e} V_e \ge c\left(1-\frac{T_{\mathrm p}}{T}\right)\operatorname{OPT}_{N}(B'). E e ∑ V e ≥ c ( 1 − T T p ) OPT N ( B ′ ) . Without delayed fast arms, T p = n T_{\mathrm p}=n T p = n , ϵ = w = 0 \epsilon=w=0 ϵ = w = 0 and B ′ = B B'=B B ′ = B .
Proof. After its first report, a fast arm’s outcomes are determined. Replacing its local policy by depth-1 commit form preserves reward, workload and the next learning state; slow arms are already nonadaptive. In particular, LP ( 1 ) ( P ) = LP ( P ) \operatorname{LP}^{(1)}(P)=\operatorname{LP}(P) LP ( 1 ) ( P ) = LP ( P ) for a single episode. Retain each arm program independently with probability 1 − T p / T 1-T_{\mathrm p}/T 1 − T p / T . The total expected unpredicted workload is then at most T − T p T-T_{\mathrm p} T − T p , so
LP N [ 1 ] ( T − T p , B ′ ) ≥ ( 1 − T p T ) LP N ( T , B ′ ) ≥ ( 1 − T p T ) OPT N ( B ′ ) . \operatorname{LP}^{[1]}_{N}(T-T_{\mathrm p},B')\ge\left(1-\frac{T_{\mathrm p}}{T}\right)\operatorname{LP}_{N}(T,B')\ge\left(1-\frac{T_{\mathrm p}}{T}\right)\operatorname{OPT}_{N}(B'). LP N [ 1 ] ( T − T p , B ′ ) ≥ ( 1 − T T p ) LP N ( T , B ′ ) ≥ ( 1 − T T p ) OPT N ( B ′ ) .
Apply Theorem 3.15 .
Allocating the remaining plays by conditional mean gives a sharper factor (Proposition B.3 ); it is asymptotically sharp for one-shot arms whose probes take o ( T ) o(T) o ( T ) steps (Corollary B.4 ).
Objectives beyond expected reward
A fixed bounded score ϕ \phi ϕ of a local policy’s plays and outcomes enters as a payoff coefficient. Use equality in the reach constraints of (3) and (12) , so every outcome receives its score; the node multipliers are then unrestricted in sign. With the history weights of (6) ,
E P ϕ = ∑ h ( A ( h ) − B ( h ) ) ℓ ( P ; h ) ϕ ( h ) . (26) \mathbb{E}^{P}\phi=\sum_{h}(A(h)-B(h))\ell(P;h)\phi(h). \tag*{(26)} E P ϕ = h ∑ ( A ( h ) − B ( h )) ℓ ( P ; h ) ϕ ( h ) . (26)
For general observation spaces, integrate against the terminal-flow measure. Downside penalties, failure indicators and squared loss against a fixed target are examples. Scores additive across arms preserve the dual decomposition. A score of their joint return generally requires joint histories. The state must retain any running reward total, or its conditional law while reports remain pending, needed to evaluate the score.
A risk functional acts on the resulting distribution. Mean–variance utility is convex in trajectory probabilities, so its maximization is not an LP and is hard in general [27 ] . CVaR has a threshold representation whose fixed-threshold payoff is linear [33 ] ; the threshold is an outer optimization parameter. Risk of joint return can still couple the arms.
Direct completion preserves the scored trajectory law, with failures charged by a bound on the score difference. Reward truncation also applies to fixed nonnegative per-play scores; a general terminal score or risk functional needs a corresponding execution bound. The reward factors above therefore do not automatically extend to every distributional objective. Example A.6 shows why conditioning on a sampled quality can discard variation decisive for mean–variance, CVaR and downside decisions. Related mean–variance and CVaR Thompson-sampling formulations already exist [8 , 44 ] ; the example concerns conditioning on a sampled episode quality.
Learning-value stability and verified execution
An error in the experiment law changes both reward and expected workload. The episode dual bounds these changes together. We then transfer the comparison through the reports preserved by execution.
Stability of the learning value
The quantitative comparisons in this subsection assume finite reward alphabets Y i \mathcal{Y}_i Y i , of sizes d i d_i d i . The model and execution results allow general bounded reward laws. Fix adaptation depths k i k_i k i , with k i = 0 k_i=0 k i = 0 for slow arms, and write L P ( k ) \mathrm{LP}^{(k)} LP ( k ) for these arm-specific classes, with k = max i k i k=\max_i k_i k = max i k i . Put r i = min { L i , k i + 1 } r_i=\min\{L_i,k_i+1\} r i = min { L i , k i + 1 } . We need the law of an r i r_i r i -play experiment to compare its reward and workload.
For r r r nonadaptive plays of arm i i i , let s = ( s y ) y ∈ Y i \boldsymbol{s}=(s_y)_{y\in\mathcal{Y}_i} s = ( s y ) y ∈ Y i be the outcome histogram, with ∣ s ∣ = r \lvert\boldsymbol{s}\rvert=r ∣ s ∣ = r . Its law and distance are
q r ( P ) = ( ( r s ) ∫ ∏ y ∈ Y i K i ( θ , { y } ) s y P ( d θ ) ) ∣ s ∣ = r , d r ( P , P ′ ) = 1 2 ∥ q r ( P ) − q r ( P ′ ) ∥ 1 . (27) \mathbf{q}_r(P)=\left(\binom{r}{\boldsymbol{s}}\int\prod_{y\in\mathcal{Y}_i}K_i(\theta,\{y\})^{s_y}P(\mathrm{d}\theta)\right)_{\lvert\boldsymbol{s}\rvert=r},\qquad d_r(P,P')=\frac{1}{2}\left\lVert\mathbf{q}_r(P)-\mathbf{q}_r(P')\right\rVert_1. \tag*{(27)} q r ( P ) = ( s r ) ∫ y ∈ Y i ∏ K i ( θ , { y } ) s y P ( d θ ) ∣ s ∣ = r , d r ( P , P ′ ) = 2 1 ∥ q r ( P ) − q r ( P ′ ) ∥ 1 . (27)
There are D i = ( r i + d i − 1 d i − 1 ) D_i=\binom{r_i+d_i-1}{d_i-1} D i = ( d i − 1 r i + d i − 1 ) possible histograms. Conditional on a histogram, exchangeability makes the ordered outcomes a uniform permutation under every prior. The ordered tape has the same total variation distance. Write ∥ x − x ′ ∥ T V = 1 2 ∥ x − x ′ ∥ 1 \lVert x-x'\rVert_{\mathrm{TV}}=\frac{1}{2}\lVert x-x'\rVert_1 ∥ x − x ′ ∥ TV = 2 1 ∥ x − x ′ ∥ 1 also for arbitrary vectors.
Lemma 4.1 (Reward and workload stability). For a depth-k i k_i k i commit policy of cap L i L_i L i , its reward and workload each change by at most L i d r i ( P , P ′ ) L_i d_{r_i}(P,P') L i d r i ( P , P ′ ) when P P P is replaced by P ′ P' P ′ .
Proof. Pre-sample r i r_i r i outcomes and the policy’s coins. These determine its adaptive prefix, commitment and workload. If it commits to b b b further plays at depth j j j , represent their reward by b b b times the next unused outcome. Conditional on the quality and prefix, this preserves expected reward. The surrogate reward and workload both lie in [ 0 , L i ] [0,L_i] [ 0 , L i ] and use at most r i r_i r i outcomes. Average over coins and the uniform ordering given the histogram, and apply the range-times-total-variation bound.
Theorem 4.2 (Stability between fixed experiment laws). For two product priors P , P ′ P,P' P , P ′ and the same caps and adaptation depths, put D e x p = ∑ i L i d r i ( P i , P i ′ ) D_{\mathrm{exp}}=\sum_i L_i d_{r_i}(P_i,P_i') D exp = ∑ i L i d r i ( P i , P i ′ ) . Then
∣ L P ( k ) ( P ) − L P ( k ) ( P ′ ) ∣ ≤ D e x p . \left|\mathrm{LP}^{(k)}(P)-\mathrm{LP}^{(k)}(P')\right|\leq D_{\mathrm{exp}}. LP ( k ) ( P ) − LP ( k ) ( P ′ ) ≤ D exp . The constant 1 is sharp.
Proof. For a fixed local policy, the surrogate reward in Lemma 4.1 satisfies 0 ≤ R ~ π ≤ W π ≤ L i 0\leq\widetilde{R}_{\pi}\leq W_{\pi}\leq L_i 0 ≤ R π ≤ W π ≤ L i on each sampled tape and coin sequence. For 0 ≤ λ ≤ 1 0\leq\lambda\leq1 0 ≤ λ ≤ 1 ,
− λ L i ≤ R ~ π − λ W π ≤ ( 1 − λ ) L i . -\lambda L_i\leq\widetilde{R}_{\pi}-\lambda W_{\pi}\leq(1-\lambda)L_i. − λ L i ≤ R π − λ W π ≤ ( 1 − λ ) L i .
The interval has length L i L_i L i . The range-times-total-variation bound therefore controls the change in the policy’s priced value by L i d r i ( P i , P i ′ ) L_i d_{r_i}(P_i,P_i') L i d r i ( P i , P i ′ ) . Maximizing over policies preserves this bound. The episode dual is
L P ( k ) ( P ) = min 0 ≤ λ ≤ 1 { λ T + ∑ i sup π i E P i [ R ~ π i − λ W π i ] } . \mathrm{LP}^{(k)}(P)=\min_{0\leq\lambda\leq1}\left\{\lambda T+\sum_i\sup_{\pi_i}\mathbb{E}_{P_i}\left[\widetilde{R}_{\pi_i}-\lambda W_{\pi_i}\right]\right\}. LP ( k ) ( P ) = 0 ≤ λ ≤ 1 min { λ T + i ∑ π i sup E P i [ R π i − λ W π i ] } .
Prices above 1 are unnecessary: all priced rewards are then nonpositive, and each arm can make no plays. Summing the arm bounds and minimizing over the common price proves the claim. One arm with cap T T T , giving identically 0 under one prior and identically 1 under the other, attains equality.
Corollary 4.3 (Learning-value stability). Fix an episode and available history H H H , with remaining budgets at least the supplied caps. Let P = ( P i ) i P=(P_i)_i P = ( P i ) i be the current recurring priors and P ‾ \overline{P} P their conditional predictive priors. Then
L P ( k ) ( P ‾ ) ≥ E [ L P ( k ) ( P ) ∣ H ] − ∑ i L i E [ d r i ( P i , P ‾ i ) ∣ H ] . (28) \mathrm{LP}^{(k)}(\overline{P})\geq\mathbb{E}\left[\mathrm{LP}^{(k)}(P)\mid H\right]-\sum_i L_i\mathbb{E}\left[d_{r_i}(P_i,\overline{P}_i)\mid H\right]. \tag*{(28)} LP ( k ) ( P ) ≥ E [ LP ( k ) ( P ) ∣ H ] − i ∑ L i E [ d r i ( P i , P i ) ∣ H ] . (28)
For any H H H -measurable vectors q ^ i ∈ R D i \hat{\mathbf{q}}_i \in\mathbb{R}^{D_i} q ^ i ∈ R D i ,
LP ( k ) ( P ˉ ) ≥ E [ LP ( k ) ( P ) ∣ H ] − 2 ∑ i L i E [ ∥ q r i ( P i ) − q ^ i ∥ T V ∣ H ] . (29) \operatorname{LP}^{(k)}(\bar{P}) \ge\mathbb{E}[\operatorname{LP}^{(k)}(P) \mid H] - 2 \sum_i L_i \mathbb{E}[\lVert\mathbf{q}_{r_i}(P_i) - \hat{\mathbf{q}}_i\rVert_{\mathrm{TV}} \mid H]. \tag*{(29)} LP ( k ) ( P ˉ ) ≥ E [ LP ( k ) ( P ) ∣ H ] − 2 i ∑ L i E [∥ q r i ( P i ) − q ^ i ∥ TV ∣ H ] . (29)
The vectors need not be realizable experiment laws.
Proof. Apply Theorem 4.2 with P ′ = P ˉ P'=\bar{P} P ′ = P ˉ and average over P P P given H H H . Since q r i ( P ˉ i ) = E [ q r i ( P i ) ∣ H ] \mathbf{q}_{r_i}(\bar{P}_i)=\mathbb{E}[\mathbf{q}_{r_i}(P_i)\mid H] q r i ( P ˉ i ) = E [ q r i ( P i ) ∣ H ] , the triangle inequality and Jensen’s inequality bound the posterior spread by twice the error about any H H H -measurable center.
Probes and verified learning
A current-episode solution gives an immediate reward floor.
Theorem 4.4 (Per-episode floor). Fix episode e e e , assume every remaining budget is at least its cap, and take a feasible solution of LP ( k ) ( P ˉ e ) \operatorname{LP}^{(k)}(\bar{P}_e) LP ( k ) ( P ˉ e ) of value at least ( 1 − ϵ o p t ) LP ( k ) ( P ˉ e ) (1-\epsilon_{\mathrm{opt}})\operatorname{LP}^{(k)}(\bar{P}_e) ( 1 − ϵ opt ) LP ( k ) ( P ˉ e ) , where 0 ≤ ϵ o p t < 1 0 \le\epsilon_{\mathrm{opt}} < 1 0 ≤ ϵ opt < 1 . Under the conditions of Theorem 3.9 or Corollary 3.10, with execution factor κ \kappa κ ,
E [ V e ∣ H e − 1 ] ≥ κ ( 1 − ϵ o p t ) LP ( k ) ( P ˉ e ) . \mathbb{E}[V_e \mid H_{e-1}] \ge\kappa(1-\epsilon_{\mathrm{opt}})\operatorname{LP}^{(k)}(\bar{P}_e). E [ V e ∣ H e − 1 ] ≥ κ ( 1 − ϵ opt ) LP ( k ) ( P ˉ e ) . Proof. Conditional on H e − 1 H_{e-1} H e − 1 , Lemma 2.6 gives the independent predictive priors P ˉ e \bar{P}_e P ˉ e . Apply the single-episode execution guarantee to the supplied solution.
The following policy acquires observations through ordinary plays and plans one episode at a time. It supplies both an executable learner and a reward comparison for the cumulative learning LP.
Algorithm 3 (Probing with episode LPs). Partition the arms into J g r p J_{\mathrm{grp}} J grp probe groups with ∑ i ∈ g r o u p r i ≤ T \sum_{i\in\mathrm{group}} r_i \le T ∑ i ∈ group r i ≤ T in each group. One can take J g r p ≤ 1 + 2 ∑ i r i / T J_{\mathrm{grp}} \le1+2\sum_i r_i/T J grp ≤ 1 + 2 ∑ i r i / T . If all probes fit, use one group. Fix ξ ∈ ( 0 , 1 ) \xi\in(0,1) ξ ∈ ( 0 , 1 ) . In each episode:
With probability 1 − ξ 1-\xi 1 − ξ , execute an optimal LP ( k ) ( P ˉ e ) \operatorname{LP}^{(k)}(\bar{P}_e) LP ( k ) ( P ˉ e ) solution with the known-prior guarantee of § 4.4 .
Otherwise, choose a group uniformly. Make r i r_i r i nonadaptive plays of each arm in that group, and idle at the remaining steps.
Update the posterior when reports arrive. Use reported outcomes from both branches.
Policy weights cancel from the posterior (Proposition 2.5), so the update uses reports from both branches. Each arm is probed with probability p = ξ / J g r p p=\xi/J_{\mathrm{grp}} p = ξ / J grp per episode, independently of its qualities. With the histogram dimensions D i D_i D i from (27) , put
L Σ = ∑ i L i , C e s t = J g r p ∑ i L i D i . (30) L_{\Sigma}=\sum_i L_i,\qquad C_{\mathrm{est}}=\sqrt{J_{\mathrm{grp}}}\sum_i L_i\sqrt{D_i}. \tag*{(30)} L Σ = i ∑ L i , C est = J grp i ∑ L i D i . (30)
The comparison need retain only the reports protected by execution. Its estimation error then transfers to the cumulative learning LP.
Theorem 4.5 (Learning comparison through protected reports). Assume finite reward alphabets, fixed caps and depths, and nonbinding comparison budgets B i ′ ≥ N L i B'_i \ge NL_i B i ′ ≥ N L i . Choose prediction depths r e , i = r i r_{e,i}=r_i r e , i = r i and a window satisfying w + ( 1 + ϵ ) ∑ i r i ≤ T p < T w+(1+\epsilon)\sum_i r_i \le T_p<T w + ( 1 + ϵ ) ∑ i r i ≤ T p < T , with X 2 = ( T − T p ) / L ˉ ≥ 1 X_2=(T-T_p)/\bar{L}\ge1 X 2 = ( T − T p ) / L ˉ ≥ 1 . Put
c = c X 2 , 0 , κ p l = c ( 1 − T p / T ) . c=c_{X_2,0},\qquad\kappa_{\mathrm{pl}}=c(1-T_p/T). c = c X 2 , 0 , κ pl = c ( 1 − T p / T ) .
Fix ξ ∈ ( 0 , 1 ) \xi\in(0,1) ξ ∈ ( 0 , 1 ) . In the planned comparison, every arm is probed for r i r_i r i plays with probability ξ \xi ξ ; otherwise the arms follow an optimal current-episode LP at capacity T − T p T-T_p T − T p . Let H e − 1 H_{e-1} H e − 1 contain its reported predicted outcomes, and take any H e − 1 H_{e-1} H e − 1 -measurable estimates q ^ e , i \widehat{q}_{e,i} q e , i . Define
E N ( ξ ) = ∑ e = 1 N ∑ i L i E ∥ q r i ( P i ( e ) ) − q ^ e , i ∥ T V . \mathcal{E}_N(\xi)=\sum_{e=1}^{N}\sum_i L_i\mathbb{E}\lVert q_{r_i}(P_i^{(e)})-\widehat{q}_{e,i}\rVert_{\mathrm{TV}}. E N ( ξ ) = e = 1 ∑ N i ∑ L i E ∥ q r i ( P i ( e ) ) − q e , i ∥ TV .
If Algorithm 2 uses a feasible restricted LP solution within ζ ≥ 0 \zeta\ge0 ζ ≥ 0 of its optimum, then
V A ≥ κ p l ( 1 − ξ ) { ∑ e ELP ( k ) ( P ( e ) ) − 2 E N ( ξ ) } − c ζ . (31) V_A \ge\kappa_{\mathrm{pl}}(1-\xi)\left\{\sum_e \operatorname{ELP}^{(k)}(P^{(e)})-2\mathcal{E}_N(\xi)\right\}-c\zeta. \tag*{(31)} V A ≥ κ pl ( 1 − ξ ) { e ∑ ELP ( k ) ( P ( e ) ) − 2 E N ( ξ ) } − c ζ . (31)
Proof. The comparison learns only from the first r i r_i r i outcomes of each arm in an episode. Its depth-k i k_i k i policies therefore form prediction-limited arm programs. Project them onto the arms as in Proposition 2.8(i). Their unpredicted workload fits T − T p T-T_p T − T p in expectation, so they are feasible for LP N [ r ] ( T − T p , B ′ ) \operatorname{LP}^{[r]}_N(T-T_p,B') LP N [ r ] ( T − T p , B ′ ) .
Thinning a current-episode solution from capacity T T T to T − T p T-T_p T − T p loses at most the factor 1 − T p / T 1-T_p/T 1 − T p / T . Probe rewards are nonnegative. If F ∗ F^* F ∗ is the restricted learning-LP optimum, then
F ∗ ≥ ( 1 − ξ ) ( 1 − T p / T ) ∑ e ELP ( k ) ( P ‾ e ) , F^* \ge(1-\xi)(1-T_p/T)\sum_e \operatorname{ELP}^{(k)}(\overline{P}_e), F ∗ ≥ ( 1 − ξ ) ( 1 − T p / T ) e ∑ ELP ( k ) ( P e ) ,
where P ‾ e \overline{P}_e P e conditions on the retained predicted reports. Apply Corollary 4.3 to this history and sum. Theorem 3.15 gives V A ≥ c F V_A \ge cF V A ≥ c F for the chosen value F F F . Finally F ≥ F ∗ − ζ F \ge F^*-\zeta F ≥ F ∗ − ζ , which proves the result. □ \square □
Corollary 4.6 (Stationary learning with episode LPs, proved in Appendix § 9 .§ 1 ). Assume stationary independent arms, finite reward alphabets, report lag at most g g g , fixed caps and depths, and nonbinding budgets B i ≥ N L i B_i \ge NL_i B i ≥ N L i . Let κ \kappa κ be the execution factor in Theorem 4.4 . Algorithm 3, with
ξ = min { 1 2 , ( C e s t 2 T 2 N ) 1 / 3 } , \xi=\min\left\{\frac{1}{2},\left(\frac{C_{\mathrm{est}}^2}{T^2N}\right)^{1/3}\right\}, ξ = min { 2 1 , ( T 2 N C est 2 ) 1/3 } , satisfies
E ∑ e = 1 N [ κ LP ( k ) ( P ) − V e p r o b e ] ≤ 5 κ ( T C e s t 2 ) 1 / 3 N 2 / 3 + 2 κ ( g − 1 ) L Σ . (32) \mathbb{E}\sum_{e=1}^{N}\left[\kappa\operatorname{LP}^{(k)}(P)-V_e^{\mathrm{probe}}\right]\le5\kappa(TC_{\mathrm{est}}^2)^{1/3}N^{2/3}+2\kappa(g-1)L_{\Sigma}. \tag*{(32)} E e = 1 ∑ N [ κ LP ( k ) ( P ) − V e probe ] ≤ 5 κ ( T C est 2 ) 1/3 N 2/3 + 2 κ ( g − 1 ) L Σ . (32) Here P P P is the stationary recurring-prior vector, and expectation averages under Q Q Q and the policy’s coins. Each planning episode optimizes a single-episode LP over arm programs. Under the conditions of Theorem 4.5 , the same bound holds for Algorithm 2, with κ \kappa κ replaced by κ p l \kappa_{\mathrm{pl}} κ pl . A planning gap ζ \zeta ζ adds c ζ c\zeta c ζ .
The first bound uses short probes and episode LPs; it does not require a joint Bellman planner. The protected learner transfers the same reported evidence through its retained states. Appendix § 9 gives the estimator comparison (Theorem C.1 ), its specializations to slow drift and known affine evolution (Theorems C.3 and C.4 ), and its transfer through completion and protected reports (Corollaries C.5 and C.6). Posterior sampling gives a separate square-root episode-LP comparison (Theorem D.2 in Appendix § 10 ); its information-price consequence is stated in § 6.2 .
Pricing and corrections
An observation is worth the decisions it improves, including those made after further episodes of learning. We first price that benefit in the learning LP, then use execution certificates to bound its feasible value. The same certificates distinguish allocation loss from discrepancies in the model’s reported predictions.
Definition 5.1 (Information benchmarks). V l e a r n = OPT N ( B ) V_{\mathrm{learn}}=\operatorname{OPT}_{N}(B) V learn = OPT N ( B ) is the optimum with the learner’s information. V p r i o r V_{\mathrm{prior}} V prior is the Q \mathcal{Q} Q -average optimum when the recurring priors are revealed before episode 1. V s t a t e V_{\mathrm{state}} V state also reveals each episode’s qualities at its start; future qualities remain hidden. All three use the same budgets, caps and report calendar. If purchased reports are available, they use the same menu and subtract fees from reward.
Extra information can be ignored, so V s t a t e ≥ V p r i o r ≥ V l e a r n V_{\mathrm{state}} \ge V_{\mathrm{prior}} \ge V_{\mathrm{learn}} V state ≥ V prior ≥ V learn . Thus
V s t a t e − V l e a r n = ( V s t a t e − V p r i o r ) + ( V p r i o r − V l e a r n ) (33) V_{\mathrm{state}}-V_{\mathrm{learn}}=(V_{\mathrm{state}}-V_{\mathrm{prior}})+(V_{\mathrm{prior}}-V_{\mathrm{learn}}) \tag*{(33)} V state − V learn = ( V state − V prior ) + ( V prior − V learn ) (33) separates the fresh-draw uncertainty from the price of learning the recurring priors.
Information at capacity prices
Fix arm i i i at the start of episode e e e , its remaining budget and report queue, and the capacity prices λ \lambda λ . Let χ \chi χ be its complete hidden record of recurring priors, qualities and potential outcomes, and let Q Q Q be the conditional law of that record. At the model’s incoming state, Q Q Q is induced by Q e , i Q_{e,i} Q e , i and the outcome kernels. For a history-based program σ \sigma σ , let g σ λ ( χ ) g^\lambda_\sigma(\chi) g σ λ ( χ ) be its remaining reward less the capacity prices of its plays, averaged over its independent coins. Then
U λ ( Q ) = sup σ ⟨ Q , g σ λ ⟩ . (34) U^\lambda(Q)=\sup_{\sigma}\langle Q,g^\lambda_\sigma\rangle. \tag*{(34)} U λ ( Q ) = σ sup ⟨ Q , g σ λ ⟩ . (34)
We suppress the fixed physical state and the indices e , i e,i e , i . The programs use the learning LP’s information: immediate feedback for fast arms, nonadaptive play for slow arms, and the specified boundary calendar. Programs subject to actual within-episode delays form a smaller class. A program that acquires information over N ′ N' N ′ episodes contributes its entire reward and workload vectors; these costs are already included in (34).
The value is convex in Q Q Q , as a supremum of linear functions [39 ] . Let g ∗ g_* g ∗ attain the supremum at Q Q Q and put D U λ ( Q ′ , Q ; g ∗ ) = U λ ( Q ′ ) − U λ ( Q ) − ⟨ Q ′ − Q , g ∗ ⟩ ≥ 0 D_{U}^{\lambda}(Q',Q;g_*)=U^\lambda(Q')-U^\lambda(Q)-\langle Q'-Q,g_*\rangle\ge0 D U λ ( Q ′ , Q ; g ∗ ) = U λ ( Q ′ ) − U λ ( Q ) − ⟨ Q ′ − Q , g ∗ ⟩ ≥ 0 . For an instantaneous report experiment Y Y Y with posterior Q Y Q_Y Q Y ,
p i λ ( Y ) = E U λ ( Q Y ) − U λ ( Q ) = E D U λ ( Q Y , Q ; g ∗ ) . (35) p_i^\lambda(Y)=\mathbb{E}U^\lambda(Q_Y)-U^\lambda(Q)=\mathbb{E}D_U^\lambda(Q_Y,Q;g_*). \tag*{(35)} p i λ ( Y ) = E U λ ( Q Y ) − U λ ( Q ) = E D U λ ( Q Y , Q ; g ∗ ) . (35)
The supporting inequality follows by retaining the old program, and the equality follows from E Q Y = Q \mathbb{E}Q_Y=Q E Q Y = Q . This is the classical value of information at fixed prices. Its link to the coupled plan requires reoptimizing those prices.
Theorem 5.2 (Information value and repricing). Take a finite learning LP at a fixed incoming boundary state. Let L P 0 \mathrm{LP}_0 LP 0 be its value without an additional report and L P Y \mathrm{LP}_Y LP Y its value when arm i i i receives the report experiment Y Y Y before further decisions. The report concerns only arm i i i , uses no resources, and changes no dates. Both LPs impose the same expected episode capacities. Let D 0 ( λ ) D_0(\lambda) D 0 ( λ ) and D Y ( λ ) D_Y(\lambda) D Y ( λ ) be their dual objectives, minimized at λ ∗ \lambda^* λ ∗ and λ Y ∗ \lambda_Y^* λ Y ∗ , respectively. Then
p i λ Y ∗ ( Y ) ≤ L P Y − L P 0 ≤ p i λ ∗ ( Y ) , (36) p_i^{\lambda_Y^*}(Y)\le\mathrm{LP}_Y-\mathrm{LP}_0\le p_i^{\lambda^*}(Y), \tag*{(36)} p i λ Y ∗ ( Y ) ≤ LP Y − LP 0 ≤ p i λ ∗ ( Y ) , (36) and
L P Y − L P 0 = p i λ ∗ ( Y ) − [ D Y ( λ ∗ ) − D Y ( λ Y ∗ ) ] . (37) \mathrm{LP}_Y-\mathrm{LP}_0=p_i^{\lambda^*}(Y)-\left[D_Y(\lambda^*)-D_Y(\lambda_Y^*)\right]. \tag*{(37)} LP Y − LP 0 = p i λ ∗ ( Y ) − [ D Y ( λ ∗ ) − D Y ( λ Y ∗ ) ] . (37) The subtracted correction is nonnegative.
Proof. Independence leaves the other arms’ priced values unchanged. At every fixed λ \lambda λ , D Y ( λ ) = D 0 ( λ ) + p i λ ( Y ) D_Y(\lambda)=D_0(\lambda)+p_i^{\lambda}(Y) D Y ( λ ) = D 0 ( λ ) + p i λ ( Y ) . Evaluating the informed dual at λ ∗ \lambda^* λ ∗ gives the upper bound. At λ ∗ \lambda^* λ ∗ , use D 0 ( λ ∗ ) ≥ L P 0 D_0(\lambda^*)\geq\mathrm{LP}_0 D 0 ( λ ∗ ) ≥ LP 0 for the lower bound. Finally substitute D Y ( λ ∗ ) = L P 0 + p i λ ∗ ( Y ) D_Y(\lambda^*)=\mathrm{LP}_0+p_i^{\lambda^*}(Y) D Y ( λ ∗ ) = LP 0 + p i λ ∗ ( Y ) and D Y ( λ Y ∗ ) = L P Y D_Y(\lambda_Y^*)=\mathrm{LP}_Y D Y ( λ Y ∗ ) = LP Y .
Both price vectors are chosen for the ex-ante LPs. In particular, λ Y ∗ \lambda_Y^* λ Y ∗ is one vector for the report experiment, rather than a different vector for each realized report. The fixed-price value predicts the gain; repricing corrects it by the improvement in the informed dual. The comparison includes future play-based learning in both branches.
Purchased reports and prior buyout
Allow an arm’s finite DAG to include report requests with specified fees, resource charges, signal laws and delivery dates. Their laws are arm-local, conditional on the arm’s hidden record. Add any observation-resource ledger to its state. Write R e Q ( π ) R_e^Q(\pi) R e Q ( π ) , W e Q ( π ) W_e^Q(\pi) W e Q ( π ) and C e Q ( π ) C_e^Q(\pi) C e Q ( π ) for reward, active-slot use and fees.
Lemma 5.3 (Priced observations on a finite DAG). At fixed capacity prices the recursion is
U e f e e ( S ) = max π { R e Q ( π ) − λ e W e Q ( π ) − C e Q ( π ) + ∑ S ′ T e ( S , S ′ ) U e + 1 f e e ( S ′ ) } , (38) U_e^{\mathrm{fee}}(S)=\max_{\pi}\left\{R_e^Q(\pi)-\lambda_e W_e^Q(\pi)-C_e^Q(\pi)+\sum_{S'}T^e(S,S')U_{e+1}^{\mathrm{fee}}(S')\right\}, \tag*{(38)} U e fee ( S ) = π max { R e Q ( π ) − λ e W e Q ( π ) − C e Q ( π ) + S ′ ∑ T e ( S , S ′ ) U e + 1 fee ( S ′ ) } , (38) with U N + 1 f e e = 0 U_{N+1}^{\mathrm{fee}}=0 U N + 1 fee = 0 . Each local step is an LP with fixed continuation values.
Proof. The supplied signal likelihoods fix the posterior at each history. Reward, resource use, fees and boundary transitions are linear in the history probabilities, as in Proposition 2.5.
Reports that consume time are priced through their resulting states and remaining opportunities. For a feasible purchase decision, let J 0 , J 1 J_0,J_1 J 0 , J 1 be the optimal remaining rewards when the report is declined or requested, before subtracting its quoted fee. Let L b ≤ J b ≤ U b L_b\leq J_b\leq U_b L b ≤ J b ≤ U b be execution and LP certificates for each full branch, b = 0 , 1 b=0,1 b = 0 , 1 . Then
L 1 − U 0 ≤ J 1 − J 0 ≤ U 1 − L 0 . (39) L_1-U_0\leq J_1-J_0\leq U_1-L_0. \tag*{(39)} L 1 − U 0 ≤ J 1 − J 0 ≤ U 1 − L 0 . (39)
A fee below the lower endpoint favors the certified purchase policy over every decline policy; a fee above the upper endpoint favors the certified decline policy over every purchase policy. The correction to the LP’s predicted price is ( U 0 − J 0 ) − ( U 1 − J 1 ) (U_0-J_0)-(U_1-J_1) ( U 0 − J 0 ) − ( U 1 − J 1 ) . Each branch includes subsequent learning, and its certificates must include any fees paid on unfinished paths. The execution theorems for the original play/report model do not by themselves bound these additional fees.
The price of the recurring prior. Revealing the recurring priors before episode 1 is worth p ∗ = V p r i o r − V l e a r n p^*=V_{\mathrm{prior}}-V_{\mathrm{learn}} p ∗ = V prior − V learn . This price can be linear when reports arrive too late to improve decisions (Example A.7 ). Under timely stationary learning it has the following bound. An exact joint episode planner serves here as a witness for information value.
Proposition 5.4 (Price of the prior by posterior sampling). *Assume stationary independent arms, finite reward alphabets of sizes d i d_i d i , caps L i L_i L i , report lag at most g g g , and budgets B i ≥ N L i B_i\geq NL_i B i ≥ N L i . Put D i t a p e = ( L i + d i − 1 ) d i − 1 D_i^{\mathrm{tape}}=(L_i+d_i-1)^{d_i-1} D i tape = ( L i + d i − 1 ) d i − 1 , D t a p e = ∏ i D i t a p e D_{\mathrm{tape}}=\prod_i D_i^{\mathrm{tape}} D tape = ∏ i D i tape and L Σ = ∑ i L i L_{\Sigma}=\sum_i L_i L Σ = ∑ i L i . Let π ∗ ( P ) \pi^*(P) π ∗ ( P ) be an optimal delay-respecting episode policy for P P P , with fixed tie-breaking, in the finite class Π \Pi Π of deterministic feasible episode policies. Then
p ∗ ≤ min { T , L Σ } 1 2 g D t a p e N H ( π ∗ ) ≤ min { T , L Σ } 1 2 g D t a p e N log ∣ Π ∣ . (40) p^*\leq\min\{T,L_{\Sigma}\}\sqrt{\frac{1}{2}gD_{\mathrm{tape}}NH(\pi^*)}\leq\min\{T,L_{\Sigma}\}\sqrt{\frac{1}{2}gD_{\mathrm{tape}}N\log|\Pi|}. \tag*{(40)} p ∗ ≤ min { T , L Σ } 2 1 g D tape N H ( π ∗ ) ≤ min { T , L Σ } 2 1 g D tape N log ∣Π∣ . (40) For fixed arms, horizon, caps and lag, the price is O ( N ) O(\sqrt{N}) O ( N ) .
Proof. Sample the recurring priors and execute an optimal joint episode policy for the sample, using q q q inference threads. Lemma D.1 bounds its loss against the revealed-prior optimum. The Bayes-optimal learner does at least as well, which bounds p ∗ p^* p ∗ .
Corollary 5.5 (The learner against the revealed prior). Under Proposition 5.4’s conditions, suppose B i ′ ≥ N L i B'_i \ge NL_i B i ′ ≥ N L i and run Algorithm 1 with an optimal solution and the uniform completion certificate of Theorem 3.7 . With p ˉ \bar p p ˉ the bound in (40) and 1 − ω = C / T 1-\omega=C/T 1 − ω = C / T as in that theorem,
V A ≥ ( 1 − ω ) ( V p r i o r − p ˉ ) − 1. V_A \ge(1-\omega)(V_{\mathrm{prior}}-\bar p)-1. V A ≥ ( 1 − ω ) ( V prior − p ˉ ) − 1.
Proof. The comparison budgets do not bind, so OPT N ( B ′ ) ≥ V p r i o r − p ˉ \operatorname{OPT}_{N}(B') \ge V_{\mathrm{prior}}-\bar p OPT N ( B ′ ) ≥ V prior − p ˉ . Apply Theorem 3.7 .
The execution factor is charged separately from the information price. Example A.7 also distinguishes buying the recurring law from buying the current draw; both options enter (38) .
Certificate corrections and predictive checks
Use the full learning relaxation with zero-reward waiting. For dual-feasible prices and node values in (14) , put
D ( λ , v ) = T ∑ e λ e + ∑ i v i ( o i ) . (41) D(\lambda,v)=T\sum_e\lambda_e+\sum_i v_i(o_i). \tag*{(41)} D ( λ , v ) = T e ∑ λ e + i ∑ v i ( o i ) . (41)
Its nonnegative play and handoff slacks are
d i p l a y ( u ) = v i ( u ) − r i ( u ) + λ e ( u ) − ∑ v p i ( u , v ) v i ( v ) , d i s t o p ( u ) = v i ( u ) − ∑ v h i ( u , v ) v i ( v ) . (42) \begin{aligned}
d_i^{\mathrm{play}}(u)&=v_i(u)-r_i(u)+\lambda_{e(u)}-\sum_v p_i(u,v)v_i(v),\\
d_i^{\mathrm{stop}}(u)&=v_i(u)-\sum_v h_i(u,v)v_i(v). \tag*{(42)}
\end{aligned} d i play ( u ) d i stop ( u ) = v i ( u ) − r i ( u ) + λ e ( u ) − v ∑ p i ( u , v ) v i ( v ) , = v i ( u ) − v ∑ h i ( u , v ) v i ( v ) . (42)
Set the play slack to zero where playing is forbidden.
Theorem 5.6 (A common policy certificate). Every admissible policy A A A has projected play and handoff probabilities w i A w_i^A w i A , z i A z_i^A z i A feasible for the full learning LP. Let t e ( A ) = ∑ i ∑ u : e ( u ) = e w i A ( u ) t_e(A)=\sum_i\sum_{u:e(u)=e}w_i^A(u) t e ( A ) = ∑ i ∑ u : e ( u ) = e w i A ( u ) . Then
D ( λ , v ) − V A = ∑ e λ e ( T − t e ( A ) ) + ∑ i , u [ w i A ( u ) d i p l a y ( u ) + z i A ( u ) d i s t o p ( u ) ] . (43) D(\lambda,v)-V_A=\sum_e\lambda_e(T-t_e(A))+\sum_{i,u}\left[w_i^A(u)d_i^{\mathrm{play}}(u)+z_i^A(u)d_i^{\mathrm{stop}}(u)\right]. \tag*{(43)} D ( λ , v ) − V A = e ∑ λ e ( T − t e ( A )) + i , u ∑ [ w i A ( u ) d i play ( u ) + z i A ( u ) d i stop ( u ) ] . (43) Every term is nonnegative.
Proof. Project the policy as in Proposition 2.8(i). Actual policy probabilities satisfy flow conservation with equality. Multiply by v i ( u ) v_i(u) v i ( u ) and sum; internal transitions cancel, leaving the initial values. Substitute (42) and add the unused-capacity charges.
This identity evaluates the learning-LP executor and HierTS with the same certificate. Decisions of slack at least d > 0 d>0 d > 0 contribute at least d d d times their expected number of uses. For the optional safe action with reward a a a , use r i ( u ) − a r_i(u)-a r i ( u ) − a , add a T N aTN a T N to D D D , and add a a a times the expected number of zero-reward waiting steps to the right side. The base model has a = 0 a=0 a = 0 .
Repricing continuation. Let B e \mathcal{B}_e B e be the integrable remaining-value certificate chosen from history H e − 1 H_{e-1} H e − 1 , and let B ^ e + 1 \widehat{\mathcal{B}}_{e+1} B e + 1 be its continuation at the state actually reached. After correction, let B e + 1 \mathcal{B}_{e+1} B e + 1 be the new certificate at that state. Define
s e = B e − E [ V e + B ^ e + 1 ∣ H e − 1 ] , Δ e = B e + 1 − B ^ e + 1 . s_e=\mathcal{B}_e-\mathbb{E}[V_e+\widehat{\mathcal{B}}_{e+1}\mid H_{e-1}],\qquad\Delta_e=\mathcal{B}_{e+1}-\widehat{\mathcal{B}}_{e+1}. s e = B e − E [ V e + B e + 1 ∣ H e − 1 ] , Δ e = B e + 1 − B e + 1 .
With terminal certificates zero, cancellation gives
E B 1 − V A = E ∑ e s e − E ∑ e < N Δ e . (44) \mathbb{E}\mathcal{B}_1-V_A=\mathbb{E}\sum_e s_e-\mathbb{E}\sum_{e<N}\Delta_e. \tag*{(44)} E B 1 − V A = E e ∑ s e − E e < N ∑ Δ e . (44)
The correction is charged with its sign. In particular, if continuation is preserved and E [ V e + α e F e + 1 ∣ H e − 1 ] ≥ α e F e \mathbb{E}[V_e+\alpha_eF_{e+1}\mid H_{e-1}]\ge\alpha_eF_e E [ V e + α e F e + 1 ∣ H e − 1 ] ≥ α e F e , where 0 ≤ α e ≤ 1 0\le\alpha_e\le1 0 ≤ α e ≤ 1 is chosen from the incoming history, then
V A ≥ E [ α 1 F 1 ] + E ∑ e = 2 N ( α e − α e − 1 ) F e . (45) V_A\ge\mathbb{E}[\alpha_1F_1]+\mathbb{E}\sum_{e=2}^{N}(\alpha_e-\alpha_{e-1})F_e. \tag*{(45)} V A ≥ E [ α 1 F 1 ] + E e = 2 ∑ N ( α e − α e − 1 ) F e . (45)
A fixed execution factor is paid once; changing it reprices the remaining continuation. The completion and protected-report executors supply the required transition guarantees. Re-solving from an actual state must also account for Δ e \Delta_e Δ e .
Checking the reported predictions. A large Bellman slack can arise under a correct model. For finite report alphabets, index arriving batches by j j j and let F j − \mathcal{F}_{j-} F j − be the actual information just before batch Y j Y_j Y j , including the plays made and the pending-report queue. A working model P ^ \widehat{\mathbb{P}} P predicts p ^ j ( y ) = P ^ ( Y j = y ∣ F j − ) \widehat{p}_j(y)=\widehat{\mathbb{P}}(Y_j=y\mid\mathcal{F}_{j-}) p j ( y ) = P ( Y j = y ∣ F j − ) . Choose a bounded score ψ j \psi_j ψ j before seeing Y j Y_j Y j and form
Z j = ψ j ( Y j ) − ∑ y p ^ j ( y ) ψ j ( y ) . (46) Z_j=\psi_j(Y_j)-\sum_y\widehat{p}_j(y)\psi_j(y). \tag*{(46)} Z j = ψ j ( Y j ) − y ∑ p j ( y ) ψ j ( y ) . (46)
For an undelayed play, suppose u u u is sufficient for this history and the tested law supplies the LP transitions. With score ψ j = R + v i ( u + ) \psi_j=R+v_i(u^+) ψ j = R + v i ( u + ) ,
v i ( u ) − R + λ e ( u ) − v i ( u + ) = d i p l a y ( u ) − Z j . v_i(u)-R+\lambda_e(u)-v_i(u^+)=d_i^{\mathrm{play}}(u)-Z_j. v i ( u ) − R + λ e ( u ) − v i ( u + ) = d i play ( u ) − Z j .
The realized shortfall separates the predicted allocation loss from the prediction error. With delays, use the conditional law of the reports actually arriving, including conditional expectations of hidden outcomes in any continuation score.
Proposition 5.7 (Predictive checking under adaptive play). For any admissible policy, under the tested model E ^ [ Z j ∣ F j − ] = 0 \widehat{\mathbb{E}}[Z_j\mid\mathcal{F}_{j-}]=0 E [ Z j ∣ F j − ] = 0 , and
E m t e s t = ∏ j = 1 m exp { ψ j ( Y j ) } ∑ y p ^ j ( y ) exp { ψ j ( y ) } . (47) E_m^{\mathrm{test}}=\prod_{j=1}^{m}\frac{\exp\{\psi_j(Y_j)\}}{\sum_y\widehat{p}_j(y)\exp\{\psi_j(y)\}}. \tag*{(47)} E m test = j = 1 ∏ m ∑ y p j ( y ) exp { ψ j ( y )} exp { ψ j ( Y j )} . (47) is a nonnegative martingale starting at one. For 0 < q < 1 0<q<1 0 < q < 1 , P ^ { sup m E m t e s t ≥ 1 / q } ≤ q \widehat{\mathbb{P}}\{\sup_m E_m^{\mathrm{test}}\ge1/q\}\le q P { sup m E m test ≥ 1/ q } ≤ q . If the policy changes only upon rejection, followed by admissible recovery, the expected reward loss under P ^ \widehat{\mathbb{P}} P is at most q N T qNT q N T in the original bounded-reward model.
Proof. Each multiplier has conditional mean one. Ville’s inequality gives the crossing bound [36 ] . Couple the checked and unchecked policies on the same tapes and coins: they agree until rejection and their rewards differ by at most N T NT N T thereafter.
A report assigned zero probability is an immediate rejection. For HierTS, the tested law is the hierarchical posterior predictive mixture for the selected arm, averaged over its latent parameters. Under an actual conditional law p j ⋆ p_j^{\star} p j ⋆ , the innovation has drift E ⋆ [ Z j ∣ F j − ] = ∑ y ( p j ⋆ ( y ) − p ^ j ( y ) ) ψ j ( y ) \mathbb{E}^{\star}[Z_j \mid\mathcal{F}_{j-}] = \sum_y (p_j^{\star}(y) - \hat{p}_j(y))\psi_j(y) E ⋆ [ Z j ∣ F j − ] = ∑ y ( p j ⋆ ( y ) − p ^ j ( y )) ψ j ( y ) . Certificate scores therefore test discrepancies relevant to the remaining program. A fixed mixture of the tests for ψ j \psi_j ψ j and − ψ j -\psi_j − ψ j checks both directions with the same false-alarm bound. Detection requires distinguishing observations; the cap-based indistinguishability of Corollary 2.4 applies here too.
The false-alarm guarantee is under the specified predictive law, averaged over a Bayesian meta-prior. Repeated episodes reveal one recurring law, not independent draws from that meta-prior. Rejection concerns the predictive specification; attributing it to the prior requires holding the kernels, stationarity and calendar fixed. Correction retains the actual clock, budget and queue and charges the change of certificate in (44) . A regret bound under misspecification would further require control of detection delay and replacement-model errors.
Illustrative Examples
The planning examples use the model of § 3 . In examples specifying a > 0 a>0 a > 0 , an idle step means taking the safe alternative of § 3.6 . Example A.8 changes the within-episode reward law to show why the fixed-law assumption is needed. The numerical comparisons take K i ( θ , ⋅ ) = Bernoulli ( θ ) K_i(\theta,\cdot)=\operatorname{Bernoulli}(\theta) K i ( θ , ⋅ ) = Bernoulli ( θ ) . Available-report HierTS draws the current episode qualities from their posterior given received reports, then chooses the feasible arm with largest sampled mean, including the idle option. Feasibility checks caps and remaining budgets. Sampling the recurring law and then the episode quality gives the same posterior draw.
Example A.1 (Pricing one test). Let idle steps earn a a a , and let one arm have cap 1, immediate feedback and a nonbinding budget. Its quality is 1 in every episode with probability q < a q<a q < a and 0 in every episode otherwise, so its first play reveals which law recurs. Assume q ( N − 2 ) ( 1 − a ) > a − q q(N-2)(1-a)>a-q q ( N − 2 ) ( 1 − a ) > a − q . No single episode gains from the arm: its one play earns q < a q<a q < a in expectation, and what it reveals cannot be used before the episode ends. The learning LP tests in episode 1 and, after a success, plays once in every later episode, gaining G N = q ( N − 1 ) ( 1 − a ) − ( a − q ) G_N=q(N-1)(1-a)-(a-q) G N = q ( N − 1 ) ( 1 − a ) − ( a − q ) over idling.
Take e 0 = 1 e_0=1 e 0 = 1 . Without a test in episode 1, the late phase tests in episode 2. Testing in episode 1 instead changes the early value by Δ R = − ( a − q ) \Delta R=-(a-q) Δ R = − ( a − q ) and the late value by Δ C = ( a − q ) + q ( 1 − a ) = a ( 1 − q ) \Delta C=(a-q)+q(1-a)=a(1-q) Δ C = ( a − q ) + q ( 1 − a ) = a ( 1 − q ) : the late phase saves its own test and gains one informed episode. Hence the test is made in episode 1 when η > η 1 = ( a − q ) / { a ( 1 − q ) } \eta>\eta_1=(a-q)/\{a(1-q)\} η > η 1 = ( a − q ) / { a ( 1 − q )} , and in episode 2 when 0 < η < η 1 0<\eta<\eta_1 0 < η < η 1 , whatever N N N . For late targets between these two plans, η 1 \eta_1 η 1 is the shadow price in (18) . At η = 0 \eta=0 η = 0 episode 1 does not test, and re-solving with η = 0 \eta=0 η = 0 at every episode never tests.
Posterior sampling can escape this episode-planning lock-in: HierTS samples quality 1 with probability q q q at each step until the arm is tested, giving it a positive probability of testing in each untested episode.
Planning solution. With one arm and cap 1, the capacity rows at C = T C=T C = T are redundant. The cascade of Theorem 2.9 is the exact Bayesian program, with value
LP N ( T , B ) = V l e a r n = a T N + G N , V p r i o r − V l e a r n = a ( 1 − q ) . \operatorname{LP}_N(T,B)=V_{\mathrm{learn}}=aTN+G_N,\qquad V_{\mathrm{prior}}-V_{\mathrm{learn}}=a(1-q). LP N ( T , B ) = V learn = a T N + G N , V prior − V learn = a ( 1 − q ) .
The test and subsequent plays fit on every path, so execution has no loss. Revealing the law gives V p r i o r = a T N + q N ( 1 − a ) V_{\mathrm{prior}}=aTN+qN(1-a) V prior = a T N + q N ( 1 − a ) . The episode planner always uses the safe action and loses G N = Θ ( N ) G_N=\Theta(N) G N = Θ ( N ) for fixed q , a q,a q , a .
Example A.2 (Cumulative versus episode planning on a fast arm). Let idle steps earn a = 1 / 2 a=1/2 a = 1/2 . Two priors recur, both with mean μ = 1 / 4 \mu=1/4 μ = 1/4 . Under P f l a t P^{\mathrm{flat}} P flat , the mean is θ = 1 / 4 \theta=1/4 θ = 1/4 in every episode. Under P s p r e a d P^{\mathrm{spread}} P spread , it is θ = 1 \theta=1 θ = 1 with probability 1 / 4 1/4 1/4 and θ = 0 \theta=0 θ = 0 otherwise.
Here the episode planner and every one-step rule stop learning, while the cumulative planner buys tests. Give the arm cap 4 ≤ L ≤ T 4\le L\le T 4 ≤ L ≤ T , a nonbinding total budget and immediate feedback (δ = 1 \delta=1 δ = 1 ). Under P f l a t P^{\mathrm{flat}} P flat the best episode policy never plays the arm. Under P s p r e a d P^{\mathrm{spread}} P spread it plays once and, after a success, makes L − 1 L-1 L − 1 further plays, gaining
Δ s p r e a d = μ ( 1 − a ) ( L − 1 ) − ( a − μ ) = L − 1 8 − 1 4 \Delta_{\mathrm{spread}}=\mu(1-a)(L-1)-(a-\mu)=\frac{L-1}{8}-\frac{1}{4} Δ spread = μ ( 1 − a ) ( L − 1 ) − ( a − μ ) = 8 L − 1 − 4 1
over a T aT a T .
One play cannot tell. Under both priors a single play is Bernoulli ( 1 / 4 ) \operatorname{Bernoulli}(1/4) Bernoulli ( 1/4 ) (Corollary 2.4 (iii)). A second play after a success can tell: under P s p r e a d P^{\mathrm{spread}} P spread it is a sure success, so a failure rules P s p r e a d P^{\mathrm{spread}} P spread out.
Certainty equivalence locks in. Let q q q be the posterior probability of P s p r e a d P^{\mathrm{spread}} P spread . If
q < q 0 = a − μ Δ s p r e a d + a − μ , q<q_0=\frac{a-\mu}{\Delta_{\mathrm{spread}}+a-\mu}, q < q 0 = Δ spread + a − μ a − μ ,
the episode policy that is optimal under P ˉ e \bar{P}_{e} P ˉ e never plays the arm. Then no data arrive and q q q never changes, although under P s p r e a d P^{\mathrm{spread}} P spread every episode could gain Δ s p r e a d \Delta_{\mathrm{spread}} Δ spread .
To verify the threshold, let ϱ \varrho ϱ be the probability of making any play. It is the same under both priors because the first play precedes any observation. Under P f l a t P^{\mathrm{flat}} P flat the loss is at least ϱ ( a − μ ) \varrho(a-\mu) ϱ ( a − μ ) . Under P s p r e a d P^{\mathrm{spread}} P spread the gain is at most ϱ Δ s p r e a d \varrho\Delta_{\mathrm{spread}} ϱ Δ spread : a bad quality loses at least a a a , and a good quality gains at most L ( 1 − a ) L(1-a) L ( 1 − a ) . The mixture’s gain is at most ϱ { q Δ s p r e a d − ( 1 − q ) ( a − μ ) } \varrho\{q\Delta_{\mathrm{spread}}-(1-q)(a-\mu)\} ϱ { q Δ spread − ( 1 − q ) ( a − μ )} , which is negative whenever ϱ > 0 \varrho>0 ϱ > 0 .
The test. Play once, and after a success play once more. Under P f l a t P^{\mathrm{flat}} P flat , this costs ( a − μ ) ( 1 + μ ) = 5 / 16 (a-\mu)(1+\mu)=5/16 ( a − μ ) ( 1 + μ ) = 5/16 in expectation. It rules out P s p r e a d P^{\mathrm{spread}} P spread with probability μ ( 1 − μ ) = 3 / 16 \mu(1-\mu)=3/16 μ ( 1 − μ ) = 3/16 per episode. Under P s p r e a d P^{\mathrm{spread}} P spread , a success followed by a success multiplies the odds of P s p r e a d P^{\mathrm{spread}} P spread by 4. Both plays count toward reward, and their outcomes are reported before the next episode.
What buys the test. After one test, q q q becomes 0, stays at q q q , or becomes 4 q / ( 1 + 3 q ) 4q/(1+3q) 4 q / ( 1 + 3 q ) . While 4 q / ( 1 + 3 q ) < q 0 4q/(1+3q)<q_0 4 q / ( 1 + 3 q ) < q 0 , no outcome of one test changes the next episode policy. The one-step value of information of the test is then zero, so a rule based on it never buys the test [32 ] . The obstruction is not specific to this test. The data of one episode multiply the odds of P s p r e a d P^{\mathrm{spread}} P spread by at most 4 L − 1 4^{L-1} 4 L − 1 , which happens after L L L successes. Under the condition 4 L − 1 q / ( 1 − q + 4 L − 1 q ) < q 0 4^{L-1}q/(1-q+4^{L-1}q)<q_0 4 L − 1 q / ( 1 − q + 4 L − 1 q ) < q 0 , no local policy within one episode can change the next episode policy. Every one-step rule therefore stops exploring.
The full belief about the recurring law is needed here. Knowing P ˉ = q P s p r e a d + ( 1 − q ) P f l a t \bar{P}=qP^{\mathrm{spread}}+(1-q)P^{\mathrm{flat}} P ˉ = q P spread + ( 1 − q ) P flat and being uncertain between P s p r e a d P^{\mathrm{spread}} P spread and P f l a t P^{\mathrm{flat}} P flat give the same current predictive prior. Only the second belief can learn which law recurs and change its future episode policy.
Planning solution. The single-arm learning LP is exact here: its capacity rows are redundant, and every local policy executes within T T T . The revealed-prior value is a T N + q N Δ s p r e a d aTN+qN\Delta_{\mathrm{spread}} a T N + q N Δ spread . For fixed L L L and q > 0 q>0 q > 0 , Proposition 5.4 gives
L P N ( T , B ) = V l e a r n ≥ a T N + q N Δ s p r e a d − O ( N ) . \mathrm{LP}_{N}(T,B)=V_{\mathrm{learn}}\ge aTN+qN\Delta_{\mathrm{spread}}-O(\sqrt{N}). LP N ( T , B ) = V learn ≥ a T N + q N Δ spread − O ( N ) .
The episode planner remains at a T N aTN a T N under the stated condition on q q q , so the advantage is linear in N N N . Direct execution realizes the LP program without loss and preserves every learning transition.
Example A.3 (Probe, wait, and return). Let T = 2 d T=2d T = 2 d , where d ≥ 3 d\ge3 d ≥ 3 is an integer. There is one unknown arm with known recurring prior
Pr ( θ ( e ) = 1 ) = 1 4 , Pr ( θ ( e ) = 0 ) = 3 4 , \Pr(\theta^{(e)}=1)=\frac{1}{4},\qquad\Pr(\theta^{(e)}=0)=\frac{3}{4}, Pr ( θ ( e ) = 1 ) = 4 1 , Pr ( θ ( e ) = 0 ) = 4 3 , and idle steps earn a = 1 / 2 a=1/2 a = 1/2 . A play at step t t t is reported before the decision at step t + d t+d t + d . Caps and total budgets do not bind. Then the Bayes-optimal reward and the available-report HierTS reward in each episode are
V ∗ = 9 d 8 − 1 4 , V T S = 17 d 16 − 9 16 + 9 16 ( 3 4 ) d . V_*=\frac{9d}{8}-\frac{1}{4},\qquad V_{\mathrm{TS}}=\frac{17d}{16}-\frac{9}{16}+\frac{9}{16}\left(\frac{3}{4}\right)^d. V ∗ = 8 9 d − 4 1 , V TS = 16 17 d − 16 9 + 16 9 ( 4 3 ) d . Consequently,
V ∗ − V T S = d 16 + 5 16 − 9 16 ( 3 4 ) d ≥ T 32 . V_*-V_{\mathrm{TS}}=\frac{d}{16}+\frac{5}{16}-\frac{9}{16}\left(\frac{3}{4}\right)^d\ge\frac{T}{32}. V ∗ − V TS = 16 d + 16 5 − 16 9 ( 4 3 ) d ≥ 32 T . Repeating the instance for N N N episodes gives an expected-reward gap of at least N T / 32 NT/32 N T /32 . The optimal policy probes at step 1, idles until the report and returns to the arm only on success. Its reward is
1 4 + d − 1 2 + d ( 1 4 + 3 4 ⋅ 1 2 ) = 9 d 8 − 1 4 . \frac{1}{4}+\frac{d-1}{2}+d\left(\frac{1}{4}+\frac{3}{4}\cdot\frac{1}{2}\right)=\frac{9d}{8}-\frac{1}{4}. 4 1 + 2 d − 1 + d ( 4 1 + 4 3 ⋅ 2 1 ) = 8 9 d − 4 1 . Extra probes before the first report have mean 1 / 4 < 1 / 2 1/4<1/2 1/4 < 1/2 and reveal nothing sooner. A first probe at step τ \tau τ gains − 1 / 4 + max { T − τ − d + 1 , 0 } / 8 -1/4+\max\{T-\tau-d+1,0\}/8 − 1/4 + max { T − τ − d + 1 , 0 } /8 over idling, maximized at τ = 1 \tau=1 τ = 1 .
Before a report, HierTS plays the arm with probability 1 / 4 1/4 1/4 and earns 7 / 16 7/16 7/16 per step in expectation; afterward it earns 5 / 8 5/8 5/8 . At step t t t , no report has arrived with probability ( 3 / 4 ) max { t − d , 0 } (3/4)^{\max\{t-d,0\}} ( 3/4 ) m a x { t − d , 0 } , independently of the quality. Hence
V T S = 5 8 ( 2 d ) − 3 16 [ d + ∑ k = 1 d ( 3 / 4 ) k ] . V_{\mathrm{TS}}=\frac{5}{8}(2d)-\frac{3}{16}\left[d+\sum_{k=1}^{d}(3/4)^k\right]. V TS = 8 5 ( 2 d ) − 16 3 [ d + k = 1 ∑ d ( 3/4 ) k ] .
This gives the displayed value and gap.
Unknown recurring prior. Let the success probability p p p be 1 / 8 1/8 1/8 or 3 / 8 3/8 3/8 , equally likely, and T > 32 T>32 T > 32 be even. At every episode start, q e = E [ p ∣ H e − 1 ] ∈ [ 1 / 8 , 3 / 8 ] q_e=\mathbb{E}[p\mid H_{e-1}]\in[1/8,3/8] q e = E [ p ∣ H e − 1 ] ∈ [ 1/8 , 3/8 ] . In the first d = T / 2 d=T/2 d = T /2 steps HierTS loses q e ( 1 / 2 − q e ) ≥ 3 / 64 q_e(1/2-q_e)\ge3/64 q e ( 1/2 − q e ) ≥ 3/64 per step relative to idling. Even granting it the quality for the remaining steps, its expected reward is at most T / 2 − 3 d / 64 + ( T − d ) / 8 T/2-3d/64+(T-d)/8 T /2 − 3 d /64 + ( T − d ) /8 , since E q e = 1 / 4 \mathbb{E}q_e=1/4 E q e = 1/4 . Probe–wait–return needs no knowledge of p p p and earns T / 2 − 1 / 4 + ( T − d ) / 8 T/2-1/4+(T-d)/8 T /2 − 1/4 + ( T − d ) /8 . Its advantage is at least 3 T / 128 − 1 / 4 ≥ T / 64 3T/128-1/4\ge T/64 3 T /128 − 1/4 ≥ T /64 per episode.
Planning solution. In the known-prior instance, the undelayed local LP probes once and, after success, continues to the cap. Its relaxed value per episode is 5 T / 8 − 3 / 8 5T/8-3/8 5 T /8 − 3/8 . To execute this plan, protect the first probe and its report, then truncate the reward tail to the d d d remaining steps. This is precisely probe–wait–return, with
V e x e c = 9 T 16 − 1 4 = V ∗ , L P N ( T , B ) N − V ∗ = T − 2 16 . V_{\mathrm{exec}}=\frac{9T}{16}-\frac{1}{4}=V_*,
\qquad
\frac{\mathrm{LP}_N(T,B)}{N}-V_*=\frac{T-2}{16}. V exec = 16 9 T − 4 1 = V ∗ , N LP N ( T , B ) − V ∗ = 16 T − 2 .
The first report reveals the episode quality completely; further plays supply no additional information, including in the unknown-prior extension. With nonbinding budgets, truncating those plays leaves the value of later decisions unchanged. Thus this execution is optimal even at delay T / 2 T/2 T /2 , while the immediate-feedback relaxation remains an upper bound.
Example A.4 (A completely slow arm with known prior). Let the safe action earn a = 1 / 2 a=1/2 a = 1/2 . One unknown arm has cap T T T , a nonbinding total budget and known recurring prior Pr ( θ ( e ) = 1 ) = 1 / 4 \Pr(\theta^{(e)}=1)=1/4 Pr ( θ ( e ) = 1 ) = 1/4 , Pr ( θ ( e ) = 0 ) = 3 / 4 \Pr(\theta^{(e)}=0)=3/4 Pr ( θ ( e ) = 0 ) = 3/4 . No report arrives within its episode. Available-report Thompson sampling therefore plays the unknown arm with probability 1 / 4 1/4 1/4 at every step. Each such play has expected reward 1 / 4 1/4 1/4 , so its expected episode reward is T [ ( 1 / 4 ) ( 1 / 4 ) + ( 3 / 4 ) ( 1 / 2 ) ] = 7 T / 16 T[(1/4)(1/4)+(3/4)(1/2)]=7T/16 T [( 1/4 ) ( 1/4 ) + ( 3/4 ) ( 1/2 )] = 7 T /16 .
Planning solution. Without within-episode reports, the number of unknown-arm plays is independent of its fresh quality. Every play replaces safe reward 1 / 2 1/2 1/2 by expected reward 1 / 4 1/4 1/4 . The slow-arm LP therefore makes no unknown-arm plays and is exactly executable: L P N ( T , B ) = V p r i o r = V l e a r n = N T / 2 \mathrm{LP}_N(T,B)=V_{\mathrm{prior}}=V_{\mathrm{learn}}=NT/2 LP N ( T , B ) = V prior = V learn = N T /2 . Thompson sampling loses T / 16 T/16 T /16 per episode, or N T / 16 NT/16 N T /16 in total, although the recurring prior is known.
Example A.5 (Spending a scarce arm too early). Let N N N be even and T = 3 m T=3m T = 3 m , with integer m ≥ 1 m\ge1 m ≥ 1 . Idle steps earn 0 0 0 . Two unknown arms have known recurring priors concentrated at means 1 1 1 and 1 / 2 1/2 1/2 , respectively. Both arms have per-episode cap 2 m 2m 2 m . Their total budgets are B 1 = N m B_1=Nm B 1 = N m and B 2 = 2 N m B_2=2Nm B 2 = 2 N m . Feedback is immediate. HierTS restricted to feasible arms earns 7 N m / 4 7Nm/4 7 N m /4 , whereas the optimal feasible reward is 2 N m 2Nm 2 N m . The expected-reward gap is N T / 12 NT/12 N T /12 .
HierTS always prefers arm 1. It plays it 2 m 2m 2 m times in each of the first N / 2 N/2 N /2 episodes, exhausting its budget. Arm 2 is played m m m times in each of those episodes and 2 m 2m 2 m times in every later one; the last m m m steps of those later episodes are idle. Total reward is N m + ( 1 / 2 ) ( 3 N m / 2 ) = 7 N m / 4 Nm+(1/2)(3Nm/2)=7Nm/4 N m + ( 1/2 ) ( 3 N m /2 ) = 7 N m /4 . Playing arm 1 m m m times and arm 2 2 m 2m 2 m times in every episode spends both budgets and earns 2 N m 2Nm 2 N m . This is optimal because it equals the sum of the budgets times their means.
Unknown recurring priors. Independent meta-priors supported on arm 1 laws in [ 3 / 4 , 1 ] [3/4,1] [ 3/4 , 1 ] and arm 2 laws in [ 1 / 4 , 1 / 2 ] [1/4,1/2] [ 1/4 , 1/2 ] preserve the same sample ordering. If μ i \mu_i μ i is arm i i i ’s unconditional mean, the two allocations earn ( μ 1 + 3 μ 2 / 2 ) N m (\mu_1+3\mu_2/2)Nm ( μ 1 + 3 μ 2 /2 ) N m and ( μ 1 + 2 μ 2 ) N m (\mu_1+2\mu_2)Nm ( μ 1 + 2 μ 2 ) N m , respectively. The gap is μ 2 N m / 2 ≥ N T / 24 \mu_2Nm/2\ge NT/24 μ 2 N m /2 ≥ N T /24 .
Planning solution. In the known-prior instance, the deterministic allocation ( m , 2 m ) (m,2m) ( m , 2 m ) in every episode is an optimal learning-LP solution. Its value 2 N m = B 1 + B 2 / 2 2Nm=B_1+B_2/2 2 N m = B 1 + B 2 /2 attains the budget upper bound, and its workload is exactly T T T on every path. Hence LP N ( T , B ) = OPT N ( B ) = 2 N m \operatorname{LP}_N(T,B)=\operatorname{OPT}_N(B)=2Nm LP N ( T , B ) = OPT N ( B ) = 2 N m and execution has no loss. The same fixed allocation realizes the stated improvement in the unknown-prior extension.
Example A.6 (A one-step risk separation). For lower-tail CVaR with tail mass β \beta β , write
C β ( V ) = sup z { z − 1 β E ( z − V ) + } . C_\beta(V)=\sup_z\left\{z-\frac{1}{\beta}\mathbb{E}(z-V)^+\right\}. C β ( V ) = z sup { z − β 1 E ( z − V ) + } . Let T = 1 T=1 T = 1 , with immediate feedback and nonbinding budgets. Idling earns a = 1 / 2 a=1/2 a = 1/2 . The unknown arm has known recurring prior Pr ( θ ( e ) = 1 ) = 3 / 4 \Pr(\theta^{(e)}=1)=3/4 Pr ( θ ( e ) = 1 ) = 3/4 and Pr ( θ ( e ) = 0 ) = 1 / 4 \Pr(\theta^{(e)}=0)=1/4 Pr ( θ ( e ) = 0 ) = 1/4 . Let V V V be the episode’s realized reward, including the policy’s randomization. Idling is optimal for each of the following three episode criteria, with value 1 / 2 1/2 1/2 :
E V − 2 Var ( V ) , C 1 / 4 ( V ) , E [ V − 3 ( 1 / 2 − V ) + ] . \mathbb{E}V-2\operatorname{Var}(V),\qquad C_{1/4}(V),\qquad\mathbb{E}[V-3(1/2-V)^+]. E V − 2 Var ( V ) , C 1/4 ( V ) , E [ V − 3 ( 1/2 − V ) + ] . HierTS attains, respectively,
49 128 , 1 8 , 13 32 . \frac{49}{128},\qquad\frac{1}{8},\qquad\frac{13}{32}. 128 49 , 8 1 , 32 13 . The same failures hold for a rule that samples the episode quality and then maximizes any of these three criteria conditional on that sample. If a policy chooses the arm with probability r r r , its reward probabilities at 0 0 0 , 1 / 2 1/2 1/2 and 1 1 1 are r / 4 r/4 r /4 , 1 − r 1-r 1 − r and 3 r / 4 3r/4 3 r /4 . Consequently,
E V − 2 Var ( V ) = 1 2 − r 4 + r 2 8 , C 1 / 4 ( V ) = 1 − r 2 , E [ V − 3 ( 1 / 2 − V ) + ] = 1 2 − r 8 . \mathbb{E}V-2\operatorname{Var}(V)=\frac{1}{2}-\frac{r}{4}+\frac{r^2}{8},\qquad C_{1/4}(V)=\frac{1-r}{2},\qquad\mathbb{E}[V-3(1/2-V)^+]=\frac{1}{2}-\frac{r}{8}. E V − 2 Var ( V ) = 2 1 − 4 r + 8 r 2 , C 1/4 ( V ) = 2 1 − r , E [ V − 3 ( 1/2 − V ) + ] = 2 1 − 8 r . Each is maximized at r = 0 r=0 r = 0 . HierTS takes r = 3 / 4 r=3/4 r = 3/4 , giving the stated values. Conditional on a sampled quality, rewards are deterministic; all three criteria therefore make the same sampled-mean choice.
Unknown recurring prior. Let p p p be 7 / 10 7/10 7/10 or 4 / 5 4/5 4/5 , equally likely, with θ ( e ) ∣ p ∼ Bernoulli ( p ) \theta^{(e)}\mid p\sim\operatorname{Bernoulli}(p) θ ( e ) ∣ p ∼ Bernoulli ( p ) . At any available history, q = E [ p ∣ H ] ∈ [ 7 / 10 , 4 / 5 ] q=\mathbb{E}[p\mid H]\in[7/10,4/5] q = E [ p ∣ H ] ∈ [ 7/10 , 4/5 ] . HierTS’s reward masses at 0 0 0 , 1 / 2 1/2 1/2 and 1 1 1 are q ( 1 − q ) q(1-q) q ( 1 − q ) , 1 − q 1-q 1 − q and q 2 q^2 q 2 . Its conditional lower-quarter CVaR is 1 / 2 − 2 q ( 1 − q ) ≤ 9 / 50 1/2-2q(1-q)\le9/50 1/2 − 2 q ( 1 − q ) ≤ 9/50 , whereas idling gives 1 / 2 1/2 1/2 . Thus its loss is at least 8 / 25 8/25 8/25 in every episode.
Planning solution. The one-step formulations of Subsection 3.5 retain the displayed outcome probabilities. The CVaR and downside objectives are linear in r r r and attain 1 / 2 1/2 1/2 at r = 0 r=0 r = 0 . Mean–variance has the linear upper certificate
1 2 − r 4 + r 2 8 ≤ 1 2 − r 8 ≤ 1 2 ( 0 ≤ r ≤ 1 ) . \frac{1}{2}-\frac{r}{4}+\frac{r^2}{8}\le\frac{1}{2}-\frac{r}{8}\le\frac{1}{2}\qquad(0\le r\le1). 2 1 − 4 r + 8 r 2 ≤ 2 1 − 8 r ≤ 2 1 ( 0 ≤ r ≤ 1 ) . Idling attains this bound as well, so all three known-prior optima are exact. The one-step decisions require no execution correction.
Example A.7 (Learning the law versus observing the draw). Each benchmark gap can be linear in N N N , and a different paid report closes each. Consider one unknown arm with cap L ≤ T L\le T L ≤ T , total budget B ≥ N L B\ge NL B ≥ N L , and idle steps earning a ∈ ( 0 , 1 ) a\in(0,1) a ∈ ( 0 , 1 ) . The additional perfect reports below are idealized finite signals with no active-slot or play-budget charge; their fees are subtracted from reward. The three benchmarks first use only ordinary reports; V f e e V_{\mathrm{fee}} V fee denotes the optimum after adding the stated paid offer.
Learning the recurring law. Let the unknown prior be P + P_+ P + , the point mass at u u u , with probability q ∈ ( 0 , 1 ) q\in(0,1) q ∈ ( 0 , 1 ) , and P − P_- P − , the point mass at ℓ \ell ℓ , otherwise, where 0 ≤ ℓ < a < u ≤ 1 0\le\ell<a<u\le1 0 ≤ ℓ < a < u ≤ 1 . The arm is slow,
and every ordinary batch is scheduled after episode N N N (so q ≥ N q \ge N q ≥ N ). Thus ordinary plays yield no information during the horizon, and each play has predictive mean μ ˉ = q u + ( 1 − q ) ℓ \bar{\mu}=qu+(1-q)\ell μ ˉ = q u + ( 1 − q ) ℓ . Using the arm pays exactly when μ ˉ > a \bar{\mu}>a μ ˉ > a ; with the law known, it pays only under P + P_{+} P + . Hence
V l e a r n = a T N + N L ( μ ˉ − a ) + , V p r i o r = a T N + N L q ( u − a ) , V_{\mathrm{learn}}=aTN+NL(\bar{\mu}-a)^{+}, \qquad V_{\mathrm{prior}}=aTN+NLq(u-a), V learn = a T N + N L ( μ ˉ − a ) + , V prior = a T N + N L q ( u − a ) ,
D N ≔ V p r i o r − V l e a r n = N L min { q ( u − a ) , ( 1 − q ) ( a − ℓ ) } . D_N\coloneqq V_{\mathrm{prior}}-V_{\mathrm{learn}}=NL\min\{q(u-a),(1-q)(a-\ell)\}. D N : = V prior − V learn = N L min { q ( u − a ) , ( 1 − q ) ( a − ℓ )} .
Here x + = max { x , 0 } x^{+}=\max\{x,0\} x + = max { x , 0 } . If a single report offered before episode 1 reveals P P P for fee c ≥ 0 c\ge0 c ≥ 0 , buying it realizes the law-informed value less c c c , so
V f e e = V l e a r n + ( D N − c ) + , V p r i o r − V f e e = min { D N , c } . V_{\mathrm{fee}}=V_{\mathrm{learn}}+(D_N-c)^{+}, \qquad V_{\mathrm{prior}}-V_{\mathrm{fee}}=\min\{D_N,c\}. V fee = V learn + ( D N − c ) + , V prior − V fee = min { D N , c } .
The report is worth buying when c < D N c<D_N c < D N : its price is supported by all N N N episodes even though it supplies no operating reward itself. Knowing P P P also fixes every episode mean here, so V s t a t e = V p r i o r V_{\mathrm{state}}=V_{\mathrm{prior}} V state = V prior . Put D 1 = D N / N D_1=D_N/N D 1 = D N / N . With weight 1 on episode 1 and η \eta η on later episodes, the report is bought when c < D 1 { 1 + η ( N − 1 ) } c<D_1\{1+\eta(N-1)\} c < D 1 { 1 + η ( N − 1 )} . When D 1 < c < D N D_1<c<D_N D 1 < c < D N , optimizing only episode 1 rejects this one-time offer, whereas the cumulative planner buys it and gains D N − c D_N-c D N − c .
Observing the current draw. Now let P P P be known, with each episode mean independently equal to 1 with probability p > a p>a p > a and 0 otherwise. Ordinary reports arrive only after their episode, so every episode starts at success probability p p p . Without a current report the optimal episode policy uses all L L L allowed plays, each gaining p − a p-a p − a ; with the mean revealed it uses them only when the mean is 1, each gaining 1 − a 1-a 1 − a . Consequently,
V l e a r n = V p r i o r = N [ a T + L ( p − a ) ] , V s t a t e − V p r i o r = N L ( 1 − p ) a . V_{\mathrm{learn}}=V_{\mathrm{prior}}=N[aT+L(p-a)], \qquad V_{\mathrm{state}}-V_{\mathrm{prior}}=NL(1-p)a. V learn = V prior = N [ a T + L ( p − a )] , V state − V prior = N L ( 1 − p ) a .
A perfect report at each episode start is worth buying when its fee is below L ( 1 − p ) a L(1-p)a L ( 1 − p ) a . For p = 3 / 4 p=3/4 p = 3/4 and a = 1 / 2 a=1/2 a = 1/2 , the no-report gap is N L / 8 NL/8 N L /8 , despite complete knowledge of the recurring law.
Planning solution. The single-arm program is exact for each of these information menus. The cap L ≤ T L\le T L ≤ T makes shared capacity redundant, and the instantaneous reports use no active slots. Adding their request nodes as in Lemma 5.3 therefore gives exactly the purchase thresholds D N D_N D N for the law and L ( 1 − p ) a L(1-p)a L ( 1 − p ) a for the current draw, with no scheduling loss. With ordinary reports alone it attains the displayed V l e a r n V_{\mathrm{learn}} V learn .
Example A.8 (The fixed-law boundary). Movement within an episode can invalidate the relaxation. Take n n n arms with caps at least T / n T/n T / n , each of mean 1 in the same window of T / n T/n T / n steps and 0 outside it, and let 0 ≤ a < 1 0\le a<1 0 ≤ a < 1 . Each arm alone gains ( T / n ) ( 1 − a ) (T/n)(1-a) ( T / n ) ( 1 − a ) over a T aT a T , so the relaxation predicts a gain of T ( 1 − a ) T(1-a) T ( 1 − a ) and total value T T T . A global policy gains at most ( T / n ) ( 1 − a ) (T/n)(1-a) ( T / n ) ( 1 − a ) , because the good windows coincide.
Scope. This construction violates the fixed within-episode law assumed in § 3 . It establishes why that assumption is needed for the relaxation with one shared capacity row per episode.
Execution proofs
Prefix bounds for buffered execution
Prefix bounds in Lemma 3.4 . Each block supplies at least one new outcome. There are at most k k k blocks before commitment. Blocks have size one while q ≤ 1 / ϵ q\le1/\epsilon q ≤ 1/ ϵ , so this initial phase uses at most ⌊ 1 / ϵ ⌋ + 1 \lfloor1/\epsilon\rfloor+1 ⌊ 1/ ϵ ⌋ + 1 blocks. Afterwards, every block except possibly the last has b ≥ ϵ q b\ge\epsilon q b ≥ ϵ q . Completed blocks therefore increase q q q by a factor of at least 1 + ϵ 1+\epsilon 1 + ϵ . Since ln ( 1 + ϵ ) ≥ ϵ ln 2 \ln(1+\epsilon)\ge\epsilon\ln2 ln ( 1 + ϵ ) ≥ ϵ ln 2 , the number of further blocks is at most ln ( ϵ k ) / ( ϵ ln 2 ) + 1 \ln(\epsilon k)/(\epsilon\ln2)+1 ln ( ϵ k ) / ( ϵ ln 2 ) + 1 when ϵ k ≥ 1 \epsilon k\ge1 ϵ k ≥ 1 . Since 1 / ϵ + 2 ≤ 3 / ϵ 1/\epsilon+2\le3/\epsilon 1/ ϵ + 2 ≤ 3/ ϵ and 1 / ln 2 < 1.45 1/\ln2<1.45 1/ ln 2 < 1.45 , this gives the stated B ϵ ( k ) B_\epsilon(k) B ϵ ( k ) .
A block of b b b plays starting at step s s s receives its last report at step s + b − 1 + δ i s+b-1+\delta_i s + b − 1 + δ i . It occupies b + δ i − 1 b+\delta_i-1 b + δ i − 1 own steps. The plays made before commitment number at most ( 1 + ϵ ) k + 1 (1+\epsilon)k+1 ( 1 + ϵ ) k + 1 . This proves the commit-policy prefix bound.
For a general policy, replace k k k in this count by ⌈ δ i / ϵ ⌉ \lceil\delta_i/\epsilon\rceil ⌈ δ i / ϵ ⌉ . Before switching modes there are at most ( 3 + 1.45 ln ( δ i + 1 ) ) / ϵ (3+1.45\ln(\delta_i+1))/\epsilon ( 3 + 1.45 ln ( δ i + 1 )) / ϵ blocks and 2 δ i / ϵ + 3 2\delta_i/\epsilon+3 2 δ i / ϵ + 3 plays. The switch adds at most δ i \delta_i δ i plays. The total prefix is bounded by 5 ( δ i + 1 ) ( 1 + ln ( δ i + 1 ) ) / ϵ 5(\delta_i+1)(1+\ln(\delta_i+1))/\epsilon 5 ( δ i + 1 ) ( 1 + ln ( δ i + 1 )) / ϵ . In delay-free mode, the oldest unused outcome was generated at least δ i \delta_i δ i own steps earlier and is available. Each remaining play can occur at the next own step. No physical plays remain after the cap is reached.
The reserve lower bound
Proof of Theorem 3.8. Each arm’s fresh episode mean is 1 1 1 with probability p = 1 / ( m + 1 ) p=1/(m+1) p = 1/ ( m + 1 ) and 0 0 0 otherwise, independently across arms and episodes. A degenerate meta-prior at these known priors is allowed by Assumption 2.1. Let π ⋆ \pi^\star π ⋆ probe once, continue to the cap after success, and stop after failure. Its workload and reward are
X = 1 + ( L − 1 ) G , G ∼ Bernoulli ( p ) , E X = m , R ( π ⋆ ) = m 2 m + 1 . X=1+(L-1)G,\qquad G\sim\operatorname{Bernoulli}(p),\qquad\mathbb{E}X=m,\qquad R(\pi^\star)=\frac{m^2}{m+1}. X = 1 + ( L − 1 ) G , G ∼ Bernoulli ( p ) , E X = m , R ( π ⋆ ) = m + 1 m 2 .
Thus the collection uses expected capacity exactly C = n m C=nm C = nm in each episode. At price λ e = 1 / 2 \lambda_e=1/2 λ e = 1/2 , a success makes every remaining play profitable and a failure makes every remaining play unprofitable. Starting has strictly positive priced value
R ( π ⋆ ) − 1 2 E X = m ( m − 1 ) 2 ( m + 1 ) > 0. R(\pi^\star)-\frac{1}{2}\mathbb{E}X=\frac{m(m-1)}{2(m+1)}>0. R ( π ⋆ ) − 2 1 E X = 2 ( m + 1 ) m ( m − 1 ) > 0.
These choices uniquely maximize the priced arm problem on reachable histories. Priors are known and budgets do not bind, so this holds in every episode. The collection is primal feasible and its dual value is
1 2 C N + N n ( m 2 m + 1 − m 2 ) = N n m 2 m + 1 , \frac{1}{2}CN+Nn\left(\frac{m^2}{m+1}-\frac{m}{2}\right)=Nn\frac{m^2}{m+1}, 2 1 C N + N n ( m + 1 m 2 − 2 m ) = N n m + 1 m 2 ,
equal to its reward. Every optimal solution must therefore attain each priced arm maximum and prescribe the stated plays.
Let Z e Z_e Z e count the good arms in episode e e e . Then the Z e Z_e Z e are independent Binomial ( n , p ) \operatorname{Binomial}(n,p) Binomial ( n , p ) variables and total standalone work is
S e = n + ( L − 1 ) Z e , S e − C = ( L − 1 ) ( Z e − n p ) . S_e=n+(L-1)Z_e,\qquad S_e-C=(L-1)(Z_e-np). S e = n + ( L − 1 ) Z e , S e − C = ( L − 1 ) ( Z e − n p ) .
All prescribed policies can finish if and only if S e ≤ T S_e\le T S e ≤ T : serial execution suffices, while excess prescribed work cannot fit under any scheduler. With
β ( s ) = Pr ( Z e > n p + s L − 1 ) , \beta(s)=\Pr\left(Z_e>np+\frac{s}{L-1}\right), β ( s ) = Pr ( Z e > n p + L − 1 s ) ,
the largest possible full-completion probability is exactly ( 1 − β ( s ) ) N (1-\beta(s))^N ( 1 − β ( s ) ) N . Hence a probability of at least 1 − δ 1-\delta 1 − δ requires
β ( s ) ≤ 1 − ( 1 − δ ) 1 / N ≤ − log ( 1 − δ ) N ≤ 2 δ N = 2 e − h . (48) \beta(s)\le1-(1-\delta)^{1/N}\le\frac{-\log(1-\delta)}{N}\le\frac{2\delta}{N}=2e^{-h}. \tag*{(48)} β ( s ) ≤ 1 − ( 1 − δ ) 1/ N ≤ N − log ( 1 − δ ) ≤ N 2 δ = 2 e − h . (48)
Put s ⋆ = ( L − 1 ) 2 v h = ( m − 1 ) 2 n m h s_\star=(L-1)\sqrt{2vh}=(m-1)\sqrt{2nmh} s ⋆ = ( L − 1 ) 2 v h = ( m − 1 ) 2 nmh . The binomial estimate below, Lemma B.1, gives, for every fixed ϵ r e s ∈ ( 0 , 1 ) \epsilon_{\mathrm{res}}\in(0,1) ϵ res ∈ ( 0 , 1 ) ,
log β ( ( 1 ± ϵ r e s ) s ⋆ ) = − ( 1 ± ϵ r e s ) 2 h ( 1 + o ( 1 ) ) . \log\beta((1\pm\epsilon_{\mathrm{res}})s_\star)=-(1\pm\epsilon_{\mathrm{res}})^2h(1+o(1)). log β (( 1 ± ϵ res ) s ⋆ ) = − ( 1 ± ϵ res ) 2 h ( 1 + o ( 1 )) .
The assumptions h → ∞ h \to\infty h → ∞ and h = o ( v ) h=o(v) h = o ( v ) put the binomial deviations between v \sqrt{v} v and v v v ; integer rounding is negligible. For s ≤ ( 1 − ε r e s ) s ⋆ s \le(1-\varepsilon_{\mathrm{res}})s_\star s ≤ ( 1 − ε res ) s ⋆ , the lower tail estimate contradicts (48) . For s ≥ ( 1 + ε r e s ) s ⋆ s \ge(1+\varepsilon_{\mathrm{res}})s_\star s ≥ ( 1 + ε res ) s ⋆ , the upper estimate gives β ( s ) ≤ δ / N \beta(s) \le\delta/N β ( s ) ≤ δ / N eventually, so a union bound gives full completion with probability at least 1 − δ 1-\delta 1 − δ . Letting ε r e s \varepsilon_{\mathrm{res}} ε res tend to zero proves (23) .
Finally, s ⋆ 2 / C 2 = 2 ( m − 1 ) 2 h / ( n m ) = o ( 1 ) s_\star^2/C^2=2(m-1)^{2h}/(nm)=o(1) s ⋆ 2 / C 2 = 2 ( m − 1 ) 2 h / ( nm ) = o ( 1 ) , so T ∼ C T\sim C T ∼ C at the smallest reserve. Since L = m 2 L=m^2 L = m 2 , substitution gives the last two statements.
Lemma B.1 (Binomial moderate deviations). Let Z ∼ Binomial ( n , p ) Z\sim\operatorname{Binomial}(n,p) Z ∼ Binomial ( n , p ) , with p p p allowed to vary along the sequence, and put v = n p ( 1 − p ) v=np(1-p) v = n p ( 1 − p ) . If v → ∞ v\to\infty v → ∞ , t / v → ∞ t/\sqrt{v}\to\infty t / v → ∞ and t / v → 0 t/v\to0 t / v → 0 , then
log Pr ( Z ≥ n p + t ) = − ( 1 + o ( 1 ) ) t 2 2 v . \log\Pr(Z\ge np+t)=-(1+o(1))\frac{t^2}{2v}. log Pr ( Z ≥ n p + t ) = − ( 1 + o ( 1 )) 2 v t 2 . A noninteger threshold is rounded upward.
Proof. Let k = ⌈ n p + t ⌉ k=\lceil np+t\rceil k = ⌈ n p + t ⌉ and let D ( ⋅ ∥ ⋅ ) D(\cdot\mathbin{\|}\cdot) D ( ⋅ ∥ ⋅ ) be the binary relative entropy. The Chernoff bound gives the upper estimate Pr p ( Z ≥ k ) ≤ exp { − n D ( k / n ∥ p ) } \Pr_p(Z\ge k)\le\exp\{-nD(k/n\mathbin{\|}p)\} Pr p ( Z ≥ k ) ≤ exp { − n D ( k / n ∥ p )} . For the lower estimate, put r = ( k + 2 v ) / n r=(k+2\sqrt{v})/n r = ( k + 2 v ) / n ; since t = o ( v ) t=o(v) t = o ( v ) , r − p = o ( p ( 1 − p ) ) r-p=o(p(1-p)) r − p = o ( p ( 1 − p )) and n r ( 1 − r ) = ( 1 + o ( 1 ) ) v nr(1-r)=(1+o(1))v n r ( 1 − r ) = ( 1 + o ( 1 )) v . Under Binomial ( n , r ) \operatorname{Binomial}(n,r) Binomial ( n , r ) , the event A = { ∣ Z − n r ∣ ≤ 2 v } A=\{|Z-nr|\le2\sqrt{v}\} A = { ∣ Z − n r ∣ ≤ 2 v } lies above k k k and has probability at least 1 / 2 1/2 1/2 by Chebyshev. On A A A the likelihood ratio Pr p ( Z = z ) / Pr r ( Z = z ) = exp { − n D ( r ∥ p ) − ( z − n r ) ℓ } \Pr_p(Z=z)/\Pr_r(Z=z)=\exp\{-nD(r\mathbin{\|}p)-(z-nr)\ell\} Pr p ( Z = z ) / Pr r ( Z = z ) = exp { − n D ( r ∥ p ) − ( z − n r ) ℓ } , with ℓ = log r ( 1 − p ) p ( 1 − r ) \ell=\log\frac{r(1-p)}{p(1-r)} ℓ = log p ( 1 − r ) r ( 1 − p ) , is at least exp { − n D ( r ∥ p ) − 2 v ℓ } \exp\{-nD(r\mathbin{\|}p)-2\sqrt{v}\ell\} exp { − n D ( r ∥ p ) − 2 v ℓ } . Taylor expansion gives n D ( r ∥ p ) nD(r\mathbin{\|}p) n D ( r ∥ p ) and n D ( k / n ∥ p ) nD(k/n\mathbin{\|}p) n D ( k / n ∥ p ) both equal to ( 1 + o ( 1 ) ) t 2 / ( 2 v ) (1+o(1))t^2/(2v) ( 1 + o ( 1 )) t 2 / ( 2 v ) , and 2 v ℓ = O ( t / v + 1 ) = o ( t 2 / v ) 2\sqrt{v}\ell=O(t/\sqrt{v}+1)=o(t^2/v) 2 v ℓ = O ( t / v + 1 ) = o ( t 2 / v ) .
Single-episode execution and comparison with GM13
Depth-restricted programs. A depth-k k k commit policy may commit earlier and randomize its tail length. Committed outcomes do not change that length, but may be retained for later learning. Fix a prior P P P , a depth k k k and the caps L i L_i L i . For arm i i i , put k i = min { k , L i } k_i=\min\{k,L_i\} k i = min { k , L i } and truncate its posterior graph at depth k i k_i k i . At a state u u u of depth j j j , let z u z_u z u be the probability of reaching u u u before commitment, x u x_u x u that of an adaptive play there, y u y_u y u that of committing there, and c u c_u c u the expected committed plays. The constraints are
z u = 1 { u = o i } + ∑ v p i ( v , u ) x v , x u + y u ≤ z u , 0 ≤ c u ≤ ( L i − j ) y u . z_u=\mathbf{1}_{\{u=o_i\}}+\sum_v p_i(v,u)x_v,\qquad x_u+y_u\le z_u,\qquad0\le c_u\le(L_i-j)y_u. z u = 1 { u = o i } + v ∑ p i ( v , u ) x v , x u + y u ≤ z u , 0 ≤ c u ≤ ( L i − j ) y u .
with x u = 0 x_u=0 x u = 0 at depth k i k_i k i ; unassigned reach mass stops. The objective is ∑ i , u r i ( u ) ( x u + c u ) \sum_{i,u}r_i(u)(x_u+c_u) ∑ i , u r i ( u ) ( x u + c u ) subject to ∑ i , u ( x u + c u ) ≤ T \sum_{i,u}(x_u+c_u)\le T ∑ i , u ( x u + c u ) ≤ T . Playing adaptively with probability x u / z u x_u/z_u x u / z u , committing with probability y u / z u y_u/z_u y u / z u , and randomly rounding c u / y u c_u/y_u c u / y u to a tail length realizes the flows; conditionally i.i.d. rewards give each committed play mean r i ( u ) r_i(u) r i ( u ) , so the objective is exact. For continuation, assign unallocated reach probability to a zero-length commitment. Index commitment by its tail length b b b : a prior-free weight y ( h , b ) y(h,b) y ( h , b ) contributes b y ( h , b ) ∑ y ∈ Y i b y ℓ i ( P i ; h y ) by(h,b)\sum_{y\in\mathcal{Y}_i^b}y\ell_i(P_i;hy) b y ( h , b ) ∑ y ∈ Y i b y ℓ i ( P i ; h y ) to reward and b y ( h , b ) ℓ i ( P i ; h ) by(h,b)\ell_i(P_i;h) b y ( h , b ) ℓ i ( P i ; h ) to workload, and each ordered tail t ∈ Y i b t\in\mathcal{Y}_i^b t ∈ Y i b has terminal probability y ( h , b ) ℓ i ( P i ; h t ) y(h,b)\ell_i(P_i;ht) y ( h , b ) ℓ i ( P i ; h t ) . For a one-shot arm, commitment after the first outcome uses c u ≤ ( L i − 1 ) y u c_u\le(L_i-1)y_u c u ≤ ( L i − 1 ) y u . Replacing this tail cap by the post-observation horizon T − δ i T-\delta_i T − δ i would exclude feasible nonadaptive plays.
The general guarantee. Admission controls expected workload. The cap bounds each arm’s realized workload. We use both facts to bound the reward lost at the horizon.
For ν > 0 \nu>0 ν > 0 , let Z ν ∼ Poisson ( ν ) Z_\nu\sim\operatorname{Poisson}(\nu) Z ν ∼ Poisson ( ν ) and define
H ν ( x ) = ∫ 0 x min { 1 , ( y − ν + 1 ) + } d y , ρ P ( ν ) = 1 − E [ H ν ( Z ν ) ] / ν . H_\nu(x)=\int_0^x\min\{1,(y-\nu+1)^+\}\,dy,\qquad\rho_P(\nu)=1-\mathbb{E}[H_\nu(Z_\nu)]/\nu. H ν ( x ) = ∫ 0 x min { 1 , ( y − ν + 1 ) + } d y , ρ P ( ν ) = 1 − E [ H ν ( Z ν )] / ν .
The subscript P P P denotes the Poisson comparison. For X ≥ 1 X \ge1 X ≥ 1 and 0 ≤ γ < 1 0 \le\gamma< 1 0 ≤ γ < 1 , set
c X , γ = max { ( 1 − γ ) / 2 , ρ P ( ( 1 − γ ) X ) } . (49) c_{X,\gamma} = \max\{(1-\gamma)/2, \rho_P((1-\gamma)X)\}. \tag*{(49)} c X , γ = max {( 1 − γ ) /2 , ρ P (( 1 − γ ) X )} . (49)
The first term retains GM13’s triangle bound. The second gives the small-cap improvement. Related Poisson supply bounds yield the same leading constant in sequential posted pricing [10 ] .
Lemma B.2 (Single-episode execution guarantee). Fix feasible local policies π i \pi_i π i with ∑ i t ( π i ) ≤ T \sum_i t(\pi_i) \le T ∑ i t ( π i ) ≤ T and value
J = ∑ i R ( π i ) . J = \sum_i R(\pi_i). J = i ∑ R ( π i ) . Let ϵ ∈ ( 0 , 1 ] \epsilon\in(0,1] ϵ ∈ ( 0 , 1 ] and γ ∈ ( 0 , 1 ) \gamma\in(0,1) γ ∈ ( 0 , 1 ) . Suppose W i ≤ γ T W_i \le\gamma T W i ≤ γ T for every used unknown arm. Choose a common cap 0 < L ≤ T 0 < L \le T 0 < L ≤ T with L i ≤ L L_i \le L L i ≤ L , and put ν = ( 1 − γ ) T / L \nu= (1-\gamma)T/L ν = ( 1 − γ ) T / L . Admit each transformed policy independently with probability α = ( 1 − γ ) / ( 1 + ϵ ) \alpha= (1-\gamma)/(1+\epsilon) α = ( 1 − γ ) / ( 1 + ϵ ) . The priority scheduler of § 4.1 , with admission, earns
V P ( A L G ) ≥ ( 1 − γ ) ρ 1 + ϵ J , ρ = max { ( 1 − γ ) / 2 , ρ P ( ν ) } . V^{P}(\mathrm{ALG}) \ge\frac{(1-\gamma)\rho}{1+\epsilon}J,\qquad\rho= \max\{(1-\gamma)/2,\rho_P(\nu)\}. V P ( ALG ) ≥ 1 + ϵ ( 1 − γ ) ρ J , ρ = max {( 1 − γ ) /2 , ρ P ( ν )} . A ( 1 + ϵ ′ ) (1+\epsilon') ( 1 + ϵ ′ ) -approximate solution of L P ( P ) \mathrm{LP}(P) LP ( P ) therefore earns at least
( 1 − γ ) ρ ( 1 + ϵ ) ( 1 + ϵ ′ ) L P ( P ) . \frac{(1-\gamma)\rho}{(1+\epsilon)(1+\epsilon')}\mathrm{LP}(P). ( 1 + ϵ ) ( 1 + ϵ ′ ) ( 1 − γ ) ρ LP ( P ) . For a depth-restricted solution, replace L P ( P ) \mathrm{LP}(P) LP ( P ) by L P ( k ) ( P ) \mathrm{LP}^{(k)}(P) LP ( k ) ( P ) . Use the prefix bound for the supplied policy class. For one-shot arms, depth 1 1 1 gives L P ( 1 ) ( P ) = L P ( P ) ≥ O P T δ ( P ) \mathrm{LP}^{(1)}(P)=\mathrm{LP}(P)\ge\mathrm{OPT}_{\delta}(P) LP ( 1 ) ( P ) = LP ( P ) ≥ OPT δ ( P ) without a commitment loss. Direct undelayed execution uses the original policies. The same guarantee then holds with ϵ = γ = 0 \epsilon=\gamma=0 ϵ = γ = 0 and α = 1 \alpha=1 α = 1 .
Order the transformed policies π i x \pi_i^x π i x by decreasing R ( π i x ) / t ( π i x ) R(\pi_i^x)/t(\pi_i^x) R ( π i x ) / t ( π i x ) , breaking ties in favor of smaller delays, and run the priority scheduler of § 4.1 in this order. Admission replaces a local policy by a null policy with probability 1 − α 1-\alpha 1 − α .
Proof of Lemma B.2. Let π i x = A ϵ ( π i ) \pi_i^x=A_{\epsilon}(\pi_i) π i x = A ϵ ( π i ) and let π i a d m \pi_i^{\mathrm{adm}} π i adm be its admitted version. Index the arms in the scheduler’s order. Write N i N_i N i for the number of plays made by π i a d m \pi_i^{\mathrm{adm}} π i adm when run alone. Put
τ i = E N i , R i = R ( π i a d m ) = α R ( π i x ) . \tau_i=\mathbb{E}N_i,\qquad R_i=R(\pi_i^{\mathrm{adm}})=\alpha R(\pi_i^x). τ i = E N i , R i = R ( π i adm ) = α R ( π i x ) .
The arm tapes and admission coins are independent. Hence the N i N_i N i are independent. Lemma 3.4 gives
N i ≤ L , ∑ i τ i ≤ α ( 1 + ϵ ) ∑ i t ( π i ) ≤ ( 1 − γ ) T = ν L . N_i \le L,\qquad\sum_i \tau_i \le\alpha(1+\epsilon)\sum_i t(\pi_i)\le(1-\gamma)T=\nu L. N i ≤ L , i ∑ τ i ≤ α ( 1 + ϵ ) i ∑ t ( π i ) ≤ ( 1 − γ ) T = ν L .
Discard policies with τ i = 0 \tau_i=0 τ i = 0 . If none remain, idling fills the horizon and proves the claim.
Pathwise truncation. Pre-sample every arm’s episode quality, tape and coins. Let S i = ∑ j < i N j S_i=\sum_{j<i}N_j S i = ∑ j < i N j . Condition on the tapes and coins of higher-priority arms. Their actual plays do not depend on arm i i i and number at most S i S_i S i .
Let M i M_i M i be the plays of arm i i i made by the scheduler. Its waiting steps advance in calendar time. A higher-priority play blocks it only when it requests a play. After at most W i ≤ γ T W_i\le\gamma T W i ≤ γ T own prefix steps, every unblocked step is a play until the policy ends. Thus
M i ≥ min { N i , ( ( 1 − γ ) T − S i ) + } ≥ ϕ i N i , ϕ i = min { 1 , ( ν L − S i ) + / L } . M_i \ge\min\{N_i,((1-\gamma)T-S_i)^+\}\ge\phi_iN_i,\qquad\phi_i=\min\{1,(\nu L-S_i)^+/L\}. M i ≥ min { N i , (( 1 − γ ) T − S i ) + } ≥ ϕ i N i , ϕ i = min { 1 , ( ν L − S i ) + / L } .
If the prefix does not fit, the first lower bound is zero. The factor ϕ i \phi_i ϕ i is fixed by the higher-priority tapes and is independent of arm i i i . Lemma 3.1 therefore gives expected reward at least R i E ϕ i R_i\mathbb{E}\phi_i R i E ϕ i .
The cap-independent bound. Since L ≤ T L \le T L ≤ T , we have ϕ i ≥ 1 − γ − S i / T \phi_i \ge1-\gamma-S_i/T ϕ i ≥ 1 − γ − S i / T . Consequently,
E ϕ i ≥ 1 − γ − ∑ j < i τ j T . \mathbb{E}\phi_i \ge1-\gamma-\frac{\sum_{j<i}\tau_j}{T}. E ϕ i ≥ 1 − γ − T ∑ j < i τ j .
The densities R i / τ i R_i/\tau_i R i / τ i are nonincreasing, and ∑ i τ i ≤ ( 1 − γ ) T \sum_i \tau_i \le(1-\gamma)T ∑ i τ i ≤ ( 1 − γ ) T . The triangle argument of Guha and Munagala [17 Lemma 3] yields
∑ i R i E ϕ i ≥ 1 − γ 2 ∑ i R i . \sum_i R_i\mathbb{E}\phi_i \ge\frac{1-\gamma}{2}\sum_i R_i. i ∑ R i E ϕ i ≥ 2 1 − γ i ∑ R i .
The Poisson bound. Define h ( s ) = min { 1 , ( s − ( ν − 1 ) L ) + / L } h(s)=\min\{1,(s-(\nu-1)L)^+/L\} h ( s ) = min { 1 , ( s − ( ν − 1 ) L ) + / L } , so that 1 − ϕ i = h ( S i ) 1-\phi_i=h(S_i) 1 − ϕ i = h ( S i ) . Put S = ∑ i N i S=\sum_i N_i S = ∑ i N i . The function h h h is nondecreasing and S i + 1 = S i + N i S_{i+1}=S_i+N_i S i + 1 = S i + N i . Hence
∑ i N i h ( S i ) ≤ ∫ 0 S h ( y ) d y = : H ( S ) . \sum_i N_i h(S_i)\le\int_0^S h(y)\,\mathrm{d}y=:H(S). i ∑ N i h ( S i ) ≤ ∫ 0 S h ( y ) d y =: H ( S ) .
The function H H H is convex, and H ( L y ) = L H ν ( y ) H(Ly)=LH_\nu(y) H ( L y ) = L H ν ( y ) .
Every convex function lies below its chord on [ 0 , L ] [0,L] [ 0 , L ] . It follows that N i N_i N i is dominated in convex order by L L L times a Bernoulli variable of mean τ i / L \tau_i/L τ i / L . A Bernoulli variable of mean q q q is dominated by a Poisson variable of mean q q q . To see this, take a convex function f f f . For every integer m ≥ 0 m\ge0 m ≥ 0 ,
f ( m ) ≥ f ( 0 ) + m ( f ( 1 ) − f ( 0 ) ) . f(m)\ge f(0)+m(f(1)-f(0)). f ( m ) ≥ f ( 0 ) + m ( f ( 1 ) − f ( 0 )) .
Averaging over the Poisson law gives at least ( 1 − q ) f ( 0 ) + q f ( 1 ) (1-q)f(0)+qf(1) ( 1 − q ) f ( 0 ) + q f ( 1 ) , the Bernoulli expectation. Convex order is preserved under independent sums [37 Chapter 3] . Equivalently, condition on the other summands and replace one at a time. Thus S S S is dominated by L Z ν 0 LZ_{\nu_0} L Z ν 0 , where ν 0 = ∑ i τ i / L ≤ ν \nu_0=\sum_i\tau_i/L\le\nu ν 0 = ∑ i τ i / L ≤ ν .
Independence of N i N_i N i and S i S_i S i now gives
∑ i τ i E h ( S i ) = E ∑ i N i h ( S i ) ≤ E H ( S ) ≤ L E H ν ( Z ν 0 ) . \sum_i\tau_i\mathbb{E}h(S_i)=\mathbb{E}\sum_iN_i h(S_i)\le\mathbb{E}H(S)\le L\mathbb{E}H_\nu(Z_{\nu_0}). i ∑ τ i E h ( S i ) = E i ∑ N i h ( S i ) ≤ E H ( S ) ≤ L E H ν ( Z ν 0 ) .
The function u ↦ E H ν ( Z u ) u\mapsto\mathbb{E}H_\nu(Z_u) u ↦ E H ν ( Z u ) vanishes at zero and is convex. Its second derivative is
E [ H ν ( Z u + 2 ) − 2 H ν ( Z u + 1 ) + H ν ( Z u ) ] ≥ 0. \mathbb{E}\left[H_\nu(Z_u+2)-2H_\nu(Z_u+1)+H_\nu(Z_u)\right]\ge0. E [ H ν ( Z u + 2 ) − 2 H ν ( Z u + 1 ) + H ν ( Z u ) ] ≥ 0.
Therefore
E H ν ( Z ν 0 ) ≤ ν 0 ν E H ν ( Z ν ) , \mathbb{E}H_\nu(Z_{\nu_0})\le\frac{\nu_0}{\nu}\mathbb{E}H_\nu(Z_\nu), E H ν ( Z ν 0 ) ≤ ν ν 0 E H ν ( Z ν ) ,
which implies
∑ i τ i E h ( S i ) ≤ ( 1 − ρ P ( ν ) ) ∑ i τ i . \sum_i\tau_i\mathbb{E}h(S_i)\le(1-\rho_P(\nu))\sum_i\tau_i. i ∑ τ i E h ( S i ) ≤ ( 1 − ρ P ( ν )) i ∑ τ i .
Both R i / τ i R_i/\tau_i R i / τ i and 1 − E h ( S i ) 1-\mathbb{E}h(S_i) 1 − E h ( S i ) are nonincreasing in i i i . Weighted Chebyshev’s inequality gives
∑ i R i ( 1 − E h ( S i ) ) ≥ ( ∑ i R i ) ∑ i τ i ( 1 − E h ( S i ) ) ∑ i τ i ≥ ρ P ( ν ) ∑ i R i . \sum_iR_i(1-\mathbb{E}h(S_i))\ge\left(\sum_iR_i\right)\frac{\sum_i\tau_i(1-\mathbb{E}h(S_i))}{\sum_i\tau_i}\ge\rho_P(\nu)\sum_iR_i. i ∑ R i ( 1 − E h ( S i )) ≥ ( i ∑ R i ) ∑ i τ i ∑ i τ i ( 1 − E h ( S i )) ≥ ρ P ( ν ) i ∑ R i .
Combining the bounds. The unknown arms earn at least ρ ∑ i R i \rho\sum_iR_i ρ ∑ i R i . Lemma 3.4 also gives
∑ i R i = α ∑ i R ( π i x ) ≥ α ∑ i R ( π i ) . \sum_iR_i=\alpha\sum_iR(\pi_i^x)\ge\alpha\sum_iR(\pi_i). i ∑ R i = α i ∑ R ( π i x ) ≥ α i ∑ R ( π i ) .
Consequently,
V P ( A L G ) ≥ ρ α ∑ i R ( π i ) = ρ α J . V^P(\mathrm{ALG})\ge\rho\alpha\sum_iR(\pi_i)=\rho\alpha J. V P ( ALG ) ≥ ρ α i ∑ R ( π i ) = ρ α J .
The approximate-solution guarantee follows by substituting J ≥ L P ( P ) / ( 1 + ϵ ′ ) J\ge\mathrm{LP}(P)/(1+\epsilon') J ≥ LP ( P ) / ( 1 + ϵ ′ ) , or the restricted LP value for a restricted policy class.
For direct undelayed execution, take π i x = π i \pi_i^x=\pi_i π i x = π i , W i = 0 W_i=0 W i = 0 and α = 1 \alpha=1 α = 1 . The proof then applies verbatim with ϵ = γ = 0 \epsilon=\gamma=0 ϵ = γ = 0 .
Proof of Theorem 3.9 and Corollary 3.10. For direct undelayed execution, apply Lemma B.2 with L = T / X L=T/X L = T / X , ϵ = γ = 0 \epsilon=\gamma=0 ϵ = γ = 0 and α = 1 \alpha=1 α = 1 . This gives c X , 0 J c_{X,0}J c X , 0 J . For delayed execution, use the same common cap and the stated admission probability. Then ν = ( 1 − γ ) X \nu=(1-\gamma)X ν = ( 1 − γ ) X and ρ = c X , γ \rho=c_{X,\gamma} ρ = c X , γ , which gives Corollary 3.10.
For integer X X X ,
E H X ( Z X ) = 1 2 Pr ( Z X ≥ X ) + E ( Z X − X ) + , E ( Z X − X ) + = X Pr ( Z X = X ) . \mathbb{E}H_X(Z_X)=\frac{1}{2}\Pr(Z_X\ge X)+\mathbb{E}(Z_X-X)^+,\qquad\mathbb{E}(Z_X-X)^+=X\Pr(Z_X=X). E H X ( Z X ) = 2 1 Pr ( Z X ≥ X ) + E ( Z X − X ) + , E ( Z X − X ) + = X Pr ( Z X = X ) .
Hence c X , 0 c_{X,0} c X , 0 equals the stated c X c_X c X . Continuity of ρ P \rho_P ρ P gives c X , γ → c X c_{X,\gamma}\to c_X c X , γ → c X at fixed X X X . As ϵ , γ → 0 \epsilon,\gamma\to0 ϵ , γ → 0 , the delayed factor therefore tends to the undelayed factor. Stirling’s formula gives
Pr ( Z X = X ) = ( 2 π X ) − 1 / 2 + O ( X − 3 / 2 ) , \Pr(Z_X=X)=(2\pi X)^{-1/2}+O(X^{-3/2}), Pr ( Z X = X ) = ( 2 π X ) − 1/2 + O ( X − 3/2 ) ,
and the remaining tail term is O ( X − 1 ) O(X^{-1}) O ( X − 1 ) . Taking the reciprocal gives the expansion of 1 / c X 1/c_X 1/ c X . For X = 4 , 8 , 16 X=4,8,16 X = 4 , 8 , 16 , the factors are below 1.363, 1.211 and 1.132.
Integrality gaps
Proof of Proposition 3.11. Let L = T / X L=T/X L = T / X , take p = T − 1 / 2 p=T^{-1/2} p = T − 1/2 , and set n = ⌊ T / ( 1 + p ( L − 1 ) ) ⌋ n=\lfloor T/(1+p(L-1))\rfloor n = ⌊ T / ( 1 + p ( L − 1 ))⌋ . Each arm has θ = 1 \theta=1 θ = 1 with probability p p p and zero otherwise. The local policy probes the arm and continues to its cap after success. The collection is feasible for the relaxation. Its expected work is n ( 1 + p ( L − 1 ) ) ≤ T n(1+p(L-1))\le T n ( 1 + p ( L − 1 )) ≤ T and its reward is n p L = ( 1 − o ( 1 ) ) T npL=(1-o(1))T n p L = ( 1 − o ( 1 )) T . Also, n p → X np\to X n p → X .
Let B g o o d ∼ Binomial ( n , p ) B_{\mathrm{good}}\sim\operatorname{Binomial}(n,p) B good ∼ Binomial ( n , p ) count the good arms.
Even a clairvoyant earns at most L min { B g o o d , X } L\min\{B_{\mathrm{good}},X\} L min { B good , X } . The bounded variable min { B g o o d , X } \min\{B_{\mathrm{good}},X\} min { B good , X } converges in mean to min { Z X , X } \min\{Z_X,X\} min { Z X , X } for Z X ∼ Poisson ( X ) Z_X\sim\operatorname{Poisson}(X) Z X ∼ Poisson ( X ) . Since
E Z X = X , E ( Z X − X ) + = X Pr ( Z X = X ) , \mathbb{E}Z_X=X,\qquad\mathbb{E}(Z_X-X)^+=X\Pr(Z_X=X), E Z X = X , E ( Z X − X ) + = X Pr ( Z X = X ) ,
division by the relaxation’s feasible value proves the claim.
Proof of Proposition 3.12. Take n = T − 1 n=T-1 n = T − 1 arms, each with θ = 1 \theta=1 θ = 1 with probability p = 1 / ( T − 1 ) 2 p=1/(T-1)^2 p = 1/ ( T − 1 ) 2 and θ = 0 \theta=0 θ = 0 otherwise. Probing every arm and continuing to the cap after success uses n ( 1 + p ( T − 1 ) ) = T n(1+p(T-1))=T n ( 1 + p ( T − 1 )) = T plays in expectation, so LP ( P ) ≥ n p T \operatorname{LP}(P)\ge npT LP ( P ) ≥ n pT . A policy earns nothing before its first play of a good arm. If that play is at step τ \tau τ , it earns at most T − τ + 1 T-\tau+1 T − τ + 1 . The arms are i.i.d. and one play reveals each, so τ \tau τ is at least the position of the first good arm in the order in which the policy probes arms, a geometric variable with parameter p p p truncated at n n n . Hence
OPT ( P ) ≤ ∑ k = 1 n p ( 1 − p ) k − 1 ( T − k + 1 ) ≤ p ∑ j = 2 T j = p ( T − 1 ) ( T + 2 ) 2 . \operatorname{OPT}(P)\le\sum_{k=1}^{n}p(1-p)^{k-1}(T-k+1)\le p\sum_{j=2}^{T}j=\frac{p(T-1)(T+2)}{2}. OPT ( P ) ≤ k = 1 ∑ n p ( 1 − p ) k − 1 ( T − k + 1 ) ≤ p j = 2 ∑ T j = 2 p ( T − 1 ) ( T + 2 ) .
and OPT ( P ) / LP ( P ) ≤ ( T + 2 ) / ( 2 T ) \operatorname{OPT}(P)/\operatorname{LP}(P)\le(T+2)/(2T) OPT ( P ) / LP ( P ) ≤ ( T + 2 ) / ( 2 T ) .
In the instances of Proposition 3.12, add the safe alternative of § 3.6 with reward a a a just below the common density of the probing policies. Then OPT ( P ) / LP ( P ) → 1 \operatorname{OPT}(P)/\operatorname{LP}(P)\to1 OPT ( P ) / LP ( P ) → 1 , but the optimal LP solution is still the search, and executing it earns about LP ( P ) / 2 \operatorname{LP}(P)/2 LP ( P ) /2 .
In the future-utilization problem of Guha and Munagala [17 Section 7] , the budget is spent before a terminal choice of L L L arms, and the factor of Guha [15 ] tends to 1 − Pr ( Z L = L ) 1-\Pr(Z_L=L) 1 − Pr ( Z L = L ) in its non-vanishing regime. With caps T / X T/X T / X , cumulative reward loses the same Poisson term Pr ( Z X = X ) \Pr(Z_X=X) Pr ( Z X = X ) , plus Pr ( Z X ≥ X ) / ( 2 X ) \Pr(Z_X\ge X)/(2X) Pr ( Z X ≥ X ) / ( 2 X ) for the local policy cut at the horizon. At general caps, exploiting a find uses the clock that finding it used, and Proposition 3.12 makes 1 / 2 1/2 1/2 tight; a terminal choice takes no time, so future utilization keeps 1 − e − 1 1-e^{-1} 1 − e − 1 .
Greedy allocation after protected reports.
Proposition B.3 (Greedy second phase). Suppose phase 2 of Algorithm 2 plays the remaining unpredicted plays in decreasing order of their conditional means given the reports available at the start of phase 2 (for a slow arm, its predictive mean at the start of the episode). Put X 2 = ( T − T p ) / L ˉ > 0 X_2=(T-T_{\mathrm{p}})/\bar{L}>0 X 2 = ( T − T p ) / L ˉ > 0 . The reward bounds of Theorem 3.15 and Corollaries 3.16 and C.6 hold with c = ρ X 2 c=\rho_{X_2} c = ρ X 2 , where
ρ X 2 = 1 − E ( Z − X 2 ) + X 2 , Z ∼ Poisson ( X 2 ) . \rho_{X_2}=1-\frac{\mathbb{E}(Z-X_2)^{+}}{X_2},\qquad Z\sim\operatorname{Poisson}(X_2). ρ X 2 = 1 − X 2 E ( Z − X 2 ) + , Z ∼ Poisson ( X 2 ) . These greedy bounds hold for every X 2 > 0 X_2>0 X 2 > 0 , and ρ X 2 = 1 − Pr ( Z = ⌊ X 2 ⌋ ) \rho_{X_2}=1-\Pr(Z=\lfloor X_2\rfloor) ρ X 2 = 1 − Pr ( Z = ⌊ X 2 ⌋) . The function ν ↦ ρ ν \nu\mapsto\rho_\nu ν ↦ ρ ν is continuous and nondecreasing on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) . For X 2 ≥ 1 X_2\geq1 X 2 ≥ 1 , ρ X 2 ≥ c X 2 , 0 \rho_{X_2}\geq c_{X_2,0} ρ X 2 ≥ c X 2 , 0 .
Proof of Proposition B.3. Fix e e e and put C 2 = T − T p C_2=T-T_{\mathrm{p}} C 2 = T − T p . Let R i \mathcal{R}_i R i be arm i i i ’s local record: its incoming retained state, its coins, and the outcomes supplied to its policy in phase 1. It determines the number M i ≤ L ˉ M_i\leq\bar{L} M i ≤ L ˉ of remaining unpredicted plays. Put p i = E [ μ i ( θ i ) ∣ R i ] p_i=\mathbb{E}[\mu_i(\theta_i)\mid\mathcal{R}_i] p i = E [ μ i ( θ i ) ∣ R i ] . By condition iv these pairs do not depend on the interleaving, so they are independent across arms, and E ∑ i M i ≤ C 2 \mathbb{E}\sum_i M_i\leq C_2 E ∑ i M i ≤ C 2 . First allocate phase 2 greedily by the p i p_i p i . Let S t = ∑ i M i 1 { p i > t } S_t=\sum_i M_i\mathbf{1}\{p_i>t\} S t = ∑ i M i 1 { p i > t } . The greedy order makes up to C 2 C_2 C 2 plays with the largest p i p_i p i , so phase 2 earns ∫ 0 1 min { C 2 , S t } d t \int_0^1\min\{C_2,S_t\}\,dt ∫ 0 1 min { C 2 , S t } d t in conditional expectation. As in the proof of Lemma B.2 , S t S_t S t is dominated in convex order by L ˉ Z ν t \bar{L}Z_{\nu_t} L ˉ Z ν t , where ν t = E S t / L ˉ ≤ X 2 \nu_t=\mathbb{E}S_t/\bar{L}\leq X_2 ν t = E S t / L ˉ ≤ X 2 . Since u ↦ E ( Z u − X 2 ) + u\mapsto\mathbb{E}(Z_u-X_2)^+ u ↦ E ( Z u − X 2 ) + is convex and vanishes at 0,
E ( S t − C 2 ) + ≤ L ˉ E ( Z ν t − X 2 ) + ≤ L ˉ ν t X 2 E ( Z X 2 − X 2 ) + = ( 1 − ρ X 2 ) E S t . \mathbb{E}(S_t-C_2)^+\leq\bar{L}\mathbb{E}(Z_{\nu_t}-X_2)^+\leq\bar{L}\frac{\nu_t}{X_2}\mathbb{E}(Z_{X_2}-X_2)^+=(1-\rho_{X_2})\mathbb{E}S_t. E ( S t − C 2 ) + ≤ L ˉ E ( Z ν t − X 2 ) + ≤ L ˉ X 2 ν t E ( Z X 2 − X 2 ) + = ( 1 − ρ X 2 ) E S t .
Hence E min { C 2 , S t } ≥ ρ X 2 E S t \mathbb{E}\min\{C_2,S_t\}\geq\rho_{X_2}\mathbb{E}S_t E min { C 2 , S t } ≥ ρ X 2 E S t . Integrating over t t t , phase 2 earns at least ρ X 2 E ∑ i M i p i \rho_{X_2}\mathbb{E}\sum_i M_ip_i ρ X 2 E ∑ i M i p i . Supplied outcomes have been reported, and the learner knows its coins and the incoming retained states, so the information available at the start of phase 2 determines every R i \mathcal{R}_i R i . Greedy allocation by the conditional means given this information maximizes the conditional expected value of phase 2 over allocations determined by it, which include the allocation above; its expected value is therefore at least as large. The steps of the phase-2 plays that are not made are also idle. The accounting in the proof of Theorem 3.15 then gives
E V e ≥ J e − ( 1 − ρ X 2 ) E ∑ i M i p i . \mathbb{E}V_e\geq J_e-(1-\rho_{X_2})\mathbb{E}\sum_i M_ip_i. E V e ≥ J e − ( 1 − ρ X 2 ) E i ∑ M i p i .
and E ∑ i M i p i ≤ ∑ i R u \mathbb{E}\sum_i M_ip_i\leq\sum_i R^u E ∑ i M i p i ≤ ∑ i R u . This is the theorem’s bound with c c c replaced by ρ X 2 \rho_{X_2} ρ X 2 ; the corollaries follow as before. The argument holds for every X 2 > 0 X_2>0 X 2 > 0 . For Z ν ∼ Poisson ( ν ) Z_\nu\sim\operatorname{Poisson}(\nu) Z ν ∼ Poisson ( ν ) and m = ⌊ ν ⌋ m=\lfloor\nu\rfloor m = ⌊ ν ⌋ ,
E ( Z ν − ν ) + = ν Pr ( Z ν = m ) , ρ ν = 1 − e − ν ν m m ! . \mathbb{E}(Z_\nu-\nu)^+=\nu\Pr(Z_\nu=m),\qquad\rho_\nu=1-e^{-\nu}\frac{\nu^m}{m!}. E ( Z ν − ν ) + = ν Pr ( Z ν = m ) , ρ ν = 1 − e − ν m ! ν m .
On ( m , m + 1 ) (m,m+1) ( m , m + 1 ) its derivative is Pr ( Z ν = m ) ( 1 − m / ν ) ≥ 0 \Pr(Z_\nu=m)(1-m/\nu)\geq0 Pr ( Z ν = m ) ( 1 − m / ν ) ≥ 0 , and the expressions agree at integers. Thus ρ ν \rho_\nu ρ ν is continuous and nondecreasing. Finally, H X ( x ) ≥ ( x − X ) + H_X(x)\geq(x-X)^+ H X ( x ) ≥ ( x − X ) + gives ρ P ( X ) ≤ ρ X \rho_P(X)\leq\rho_X ρ P ( X ) ≤ ρ X , and
E ( Z X − X ) + = E ∣ Z X − X ∣ / 2 ≤ X / 2 \mathbb{E}(Z_X-X)^+=\mathbb{E}|Z_X-X|/2\leq\sqrt{X}/2 E ( Z X − X ) + = E ∣ Z X − X ∣/2 ≤ X /2
gives ρ X ≥ 1 / 2 \rho_X\geq1/2 ρ X ≥ 1/2 for X ≥ 1 X\geq1 X ≥ 1 ; hence ρ X 2 ≥ c X 2 , 0 \rho_{X_2}\geq c_{X_2,0} ρ X 2 ≥ c X 2 , 0 for X 2 ≥ 1 X_2\geq1 X 2 ≥ 1 . □
Corollary B.4 (Sharp factor with short probes). Fix an integer X ≥ 1 X\geq1 X ≥ 1 . For undelayed one-shot single-episode instances with caps L i ≤ T / X L_i\leq T/X L i ≤ T / X , budgets B i ≥ L i B_i\geq L_i B i ≥ L i and n = o ( T ) n=o(T) n = o ( T ) , the greedy variant of Algorithm 2, with ϵ = ω = 0 \epsilon=\omega=0 ϵ = ω = 0 , T p = n T_{\mathrm{p}}=n T p = n and an optimal restricted LP solution, earns
E V ≥ ( ρ X − o ( 1 ) ) LP ( P ) . \mathbb{E}V\geq(\rho_X-o(1))\operatorname{LP}(P). E V ≥ ( ρ X − o ( 1 )) LP ( P ) . The factor is worst-case tight relative to LP ( P ) \operatorname{LP}(P) LP ( P ) .
Proof. Put X 2 = ( T − n ) / L ˉ ≥ X ( 1 − n / T ) X_2=(T-n)/\bar{L}\geq X(1-n/T) X 2 = ( T − n ) / L ˉ ≥ X ( 1 − n / T ) . Monotonicity and continuity give ρ X 2 ≥ ρ X − o ( 1 ) \rho_{X_2}\geq\rho_X-o(1) ρ X 2 ≥ ρ X − o ( 1 ) . The proof of Corollary 3.16 , using the greedy bound for every X 2 > 0 X_2>0 X 2 > 0 , gives E V ≥ ρ X 2 ( 1 − n / T ) LP ( P ) \mathbb{E}V\geq\rho_{X_2}(1-n/T)\operatorname{LP}(P) E V ≥ ρ X 2 ( 1 − n / T ) LP ( P ) , proving the guarantee. For the instances of Proposition 3.11, n = O ( X T ) = o ( T ) n=O(X\sqrt{T})=o(T) n = O ( X T ) = o ( T ) and OPT ( P ) ≤ ( ρ X + o ( 1 ) ) LP ( P ) \operatorname{OPT}(P)\leq(\rho_X+o(1))\operatorname{LP}(P) OPT ( P ) ≤ ( ρ X + o ( 1 )) LP ( P ) . Thus the worst-case ratio OPT / LP \operatorname{OPT}/\operatorname{LP} OPT / LP in this class is ρ X + o ( 1 ) \rho_X+o(1) ρ X + o ( 1 ) . □
The short-probe condition. The restriction to short probes is necessary. Let T T T be a multiple of 4, and take n = T / 4 n=T/4 n = T /4 one-shot arms with L i = L = T / 2 L_i=L=T/2 L i = L = T /2 , p = 3 / ( L − 1 ) p=3/(L-1) p = 3/ ( L − 1 ) and a = 0 a=0 a = 0 . Probing every arm and continuing to the cap after success fits in expectation, n ( 1 + p ( L − 1 ) ) = T n(1+p(L-1))=T n ( 1 + p ( L − 1 )) = T , and collects every good arm’s reward, so LP ( P ) = n p L \operatorname{LP}(P)=npL LP ( P ) = n p L . The optimum exhausts each good arm before probing further. If the second good arm is the j j j th probed, the first contributes L L L and the second L − j + 2 L-j+2 L − j + 2 , so
OPT ( P ) = L ( 1 − ( 1 − p ) n ) + ∑ j = 2 n ( L − j + 2 ) ( j − 1 ) p 2 ( 1 − p ) j − 2 , \operatorname{OPT}(P)=L(1-(1-p)^n)+\sum_{j=2}^{n}(L-j+2)(j-1)p^2(1-p)^{j-2}, OPT ( P ) = L ( 1 − ( 1 − p ) n ) + j = 2 ∑ n ( L − j + 2 ) ( j − 1 ) p 2 ( 1 − p ) j − 2 ,
OPT ( P ) LP ( P ) → 8 9 − 13 18 e − 3 / 2 ≈ 0.7277 , \frac{\operatorname{OPT}(P)}{\operatorname{LP}(P)}\to\frac{8}{9}-\frac{13}{18}e^{-3/2}\approx0.7277, LP ( P ) OPT ( P ) → 9 8 − 18 13 e − 3/2 ≈ 0.7277 ,
below ρ 2 = 1 − 2 e − 2 ≈ 0.7293 \rho_2=1-2e^{-2}\approx0.7293 ρ 2 = 1 − 2 e − 2 ≈ 0.7293 . Here probes take a quarter of the horizon: the time spent finding good arms adds to their competition for time.
Delays and comparison with GM13. Table 1 compares the two guarantees.
Table 1. Approximation factors relative to the single-episode LP; in the delayed rows, GM13’s factor is relative to its delay-respecting relaxation and ours to the undelayed LP ( P ) \operatorname{LP}(P) LP ( P ) , and y = δ / T y=\delta/\sqrt{T} y = δ / T . Proposition 3.12 shows that the first and last rows are tight, and Proposition 3.11 that the leading term in the fifth row is necessary for this relaxation.
Fix X ≥ 1 X\ge1 X ≥ 1 and use the general-policy prefix bound. Put
D δ = ( δ + 1 ) ( 1 + ln ( δ + 1 ) ) , ϵ = γ = 5 D δ / T < 1. D_\delta=(\delta+1)(1+\ln(\delta+1)),\qquad\epsilon=\gamma=\sqrt{5D_\delta/T}<1. D δ = ( δ + 1 ) ( 1 + ln ( δ + 1 )) , ϵ = γ = 5 D δ / T < 1.
Then W i ≤ 5 D δ / ϵ = γ T W_i\le5D_\delta/\epsilon=\gamma T W i ≤ 5 D δ / ϵ = γ T , so Corollary 3.10 applies. The function ρ P \rho_P ρ P is locally Lipschitz on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) : on a compact parameter interval, H ν ( z ) H_\nu(z) H ν ( z ) is Lipschitz in ν \nu ν with a summable Poisson envelope. Hence c X , γ = c X , 0 + O ( γ ) c_{X,\gamma}=c_{X,0}+O(\gamma) c X , γ = c X , 0 + O ( γ ) at fixed X X X . Since c X , 0 ≥ 1 / 2 c_{X,0}\ge1/2 c X , 0 ≥ 1/2 , the approximation factor is
1 + ϵ ( 1 − γ ) c X , γ = 1 c X , 0 + O ( D δ / T ) . \frac{1+\epsilon}{(1-\gamma)c_{X,\gamma}}=\frac{1}{c_{X,0}}+O\left(\sqrt{D_\delta/T}\right). ( 1 − γ ) c X , γ 1 + ϵ = c X , 0 1 + O ( D δ / T ) .
An LP approximation with ϵ ′ = O ( ϵ + γ ) \epsilon'=O(\epsilon+\gamma) ϵ ′ = O ( ϵ + γ ) has the same order of additional loss. At δ = T \delta=\sqrt{T} δ = T , this delay cost is O ( T − 1 / 4 log T ) O(T^{-1/4}\sqrt{\log T}) O ( T − 1/4 log T ) and tends to zero.
GM13’s delayed factor is 2 ( 1 + ϵ ) + 32 ( y + y 2 ) 2(1+\epsilon)+32(y+y^2) 2 ( 1 + ϵ ) + 32 ( y + y 2 ) , where y = δ / T y=\delta/\sqrt{T} y = δ / T [17 ] , Theorem 16, relative to its delay-respecting relaxation; at δ = T \delta=\sqrt{T} δ = T it tends to 66 as ϵ → 0 \epsilon\to0 ϵ → 0 .
Corollary B.5 (The delay cost vanishes). Let δ ≥ 1 \delta\ge1 δ ≥ 1 be the largest delay of a fast arm, put ϵ δ = 5 D δ / T \epsilon_\delta=\sqrt{5D_\delta/T} ϵ δ = 5 D δ / T , and let ϵ δ < 1 \epsilon_\delta<1 ϵ δ < 1 . (i) Suppose every unknown arm is fast. The policy of Corollary 3.10 with ϵ = γ = ϵ δ \epsilon=\gamma=\epsilon_\delta ϵ = γ = ϵ δ , applied to an optimal solution, earns
EV ≥ 1 − ϵ δ 1 + ϵ δ c X , ϵ δ LP ( P ) ≥ ( 1 − ϵ δ ) 2 2 ( 1 + ϵ δ ) LP ( P ) , \operatorname{EV}\ge\frac{1-\epsilon_\delta}{1+\epsilon_\delta}c_{X,\epsilon_\delta}\operatorname{LP}(P)\ge\frac{(1-\epsilon_\delta)^2}{2(1+\epsilon_\delta)}\operatorname{LP}(P), EV ≥ 1 + ϵ δ 1 − ϵ δ c X , ϵ δ LP ( P ) ≥ 2 ( 1 + ϵ δ ) ( 1 − ϵ δ ) 2 LP ( P ) ,
the second bound for arbitrary caps and general bounded rewards. The factor is 2 ( 1 + ε δ ) / ( 1 − ε δ ) 2 = 2 + O ( D δ / T ) 2(1+\varepsilon_\delta)/(1-\varepsilon_\delta)^2 = 2+O(\sqrt{D_\delta/T}) 2 ( 1 + ε δ ) / ( 1 − ε δ ) 2 = 2 + O ( D δ / T ) , and its limit 2 is tight by Proposition 3.12 . (ii) Run Algorithm 1 with ϵ = ε δ ≤ 1 \epsilon=\varepsilon_\delta\le1 ϵ = ε δ ≤ 1 , w = 5 D δ / ϵ w=5D_\delta/\epsilon w = 5 D δ / ϵ and the uniform reserve s < T − w s<T-w s < T − w of Theorem 3.7 . Then
E ∑ e V e ≥ LP N ( T , B ′ ) − N ( 2 L ˉ T Λ e x e c + 2 3 L ˉ Λ e x e c + 2 5 D δ T − 1 ) . \mathbb{E}\sum_e V_e \ge\operatorname{LP}_N(T,B')-N\left(\sqrt{2\bar{L}T\Lambda_{\mathrm{exec}}+\frac{2}{3}\bar{L}\Lambda_{\mathrm{exec}}+2\sqrt{5D_\delta T}}-1\right). E e ∑ V e ≥ LP N ( T , B ′ ) − N ( 2 L ˉ T Λ exec + 3 2 L ˉ Λ exec + 2 5 D δ T − 1 ) .
where LP N ( T , B ′ ) ≥ OPT N ( B ′ ) \operatorname{LP}_N(T,B')\ge\operatorname{OPT}_N(B') LP N ( T , B ′ ) ≥ OPT N ( B ′ ) . For N = 1 N=1 N = 1 and B i ′ ≥ L i B'_i\ge L_i B i ′ ≥ L i , LP N ( T , B ′ ) ≥ OPT δ ( P ) \operatorname{LP}_N(T,B')\ge\operatorname{OPT}_\delta(P) LP N ( T , B ′ ) ≥ OPT δ ( P ) . It equals LP ( P ) \operatorname{LP}(P) LP ( P ) when every unknown arm is fast; with slow arms it is the single-episode LP that restricts them to nonadaptive local policies.
Proof. (i) Corollary 3.10 applies because W i ≤ 5 D δ / ε δ = ε δ T W_i\le5D_\delta/\varepsilon_\delta=\varepsilon_\delta T W i ≤ 5 D δ / ε δ = ε δ T . It gives the first bound, and c X , ε δ ≥ ( 1 − ε δ ) / 2 c_{X,\varepsilon_\delta}\ge(1-\varepsilon_\delta)/2 c X , ε δ ≥ ( 1 − ε δ ) /2 gives the second. (ii) Lemma 3.4 bounds every prefix by w w w . For an optimal solution, Proposition 2.8(ii) gives F ≥ ( C / T ) LP N ( T , B ′ ) F\ge(C/T)\operatorname{LP}_N(T,B') F ≥ ( C / T ) LP N ( T , B ′ ) , and C / T ≥ 1 − ( w + s ) / T − ϵ C/T\ge1-(w+s)/T-\epsilon C / T ≥ 1 − ( w + s ) / T − ϵ . Since LP N ( T , B ′ ) ≤ T N \operatorname{LP}_N(T,B')\le TN LP N ( T , B ′ ) ≤ T N , Theorem 3.7 gives E ∑ e V e ≥ LP N ( T , B ′ ) − N ( w + s + ϵ T ) − 1 \mathbb{E}\sum_e V_e\ge\operatorname{LP}_N(T,B')-N(w+s+\epsilon T)-1 E ∑ e V e ≥ LP N ( T , B ′ ) − N ( w + s + ϵ T ) − 1 . The choice of ϵ \epsilon ϵ makes w + ϵ T = 2 5 D δ T w+\epsilon T=2\sqrt{5D_\delta T} w + ϵ T = 2 5 D δ T .
Several plays per step
We prove Theorem 3.13 and its dependence on K K K . Each arm remains serial, while different arms share K K K lanes.
Proof of Theorem 3.13 . If W = 0 W=0 W = 0 , then J = 0 J=0 J = 0 and the claim is immediate. Otherwise discard policies with zero workload and index the rest in priority order. Let N j N_j N j be the plays of policy j j j when run alone, τ j = E N j \tau_j=\mathbb{E}N_j τ j = E N j , R j = R ( π j ) R_j=R(\pi_j) R j = R ( π j ) and S i = ∑ j < i N j S_i=\sum_{j<i}N_j S i = ∑ j < i N j .
Blocking. Higher-priority policies do not depend on policy i i i . At a step where policy i i i has not stopped but is not played, K K K higher-priority policies are played; such steps use K K K of at most S i S_i S i higher-priority plays, so there are at most S i / K S_i/K S i / K of them. Policy i i i is played at every other step until it stops. Its executed plays therefore satisfy
M i ≥ min { N i , ( T − S i / K ) + } ≥ ϕ i N i , ϕ i = min { 1 , ( T − S i / K ) + / L } . M_i\ge\min\{N_i,(T-S_i/K)^+\}\ge\phi_iN_i,\qquad\phi_i=\min\{1,(T-S_i/K)^+/L\}. M i ≥ min { N i , ( T − S i / K ) + } ≥ ϕ i N i , ϕ i = min { 1 , ( T − S i / K ) + / L } .
The factor ϕ i \phi_i ϕ i is determined by the tapes and coins of policies j < i j<i j < i , which are independent of policy i i i , so Lemma 3.1 gives expected reward at least R i E ϕ i R_i\mathbb{E}\phi_i R i E ϕ i .
Triangle bound. Since L ≤ T L\le T L ≤ T , ϕ i ≥ 1 − S i / ( K T ) \phi_i\ge1-S_i/(KT) ϕ i ≥ 1 − S i / ( K T ) and E ϕ i ≥ 1 − ∑ j < i τ j / ( K T ) \mathbb{E}\phi_i\ge1-\sum_{j<i}\tau_j/(KT) E ϕ i ≥ 1 − ∑ j < i τ j / ( K T ) . The densities R i / τ i R_i/\tau_i R i / τ i are nonincreasing and ∑ j < i τ j \sum_{j<i}\tau_j ∑ j < i τ j is nondecreasing in i i i , so Chebyshev’s sum inequality with weights τ i \tau_i τ i gives
∑ i R i ∑ j < i τ j ≤ J W ∑ i τ i ∑ j < i τ j ≤ J W 2 . \sum_i R_i\sum_{j<i}\tau_j\le\frac{J}{W}\sum_i\tau_i\sum_{j<i}\tau_j\le\frac{JW}{2}. i ∑ R i j < i ∑ τ j ≤ W J i ∑ τ i j < i ∑ τ j ≤ 2 J W .
Hence ∑ i R i E ϕ i ≥ ( 1 − W / ( 2 K T ) ) J \sum_iR_i\mathbb{E}\phi_i\ge(1-W/(2KT))J ∑ i R i E ϕ i ≥ ( 1 − W / ( 2 K T )) J .
Poisson bound. Write 1 − ϕ i = h ( S i / K ) 1-\phi_i=h(S_i/K) 1 − ϕ i = h ( S i / K ) with h ( s ) = min { 1 , ( s − T + L ) + / L } h(s)=\min\{1,(s-T+L)^+/L\} h ( s ) = min { 1 , ( s − T + L ) + / L } , and put H ( x ) = ∫ 0 x h H(x)=\int_0^x h H ( x ) = ∫ 0 x h and S = ∑ i N i S=\sum_iN_i S = ∑ i N i . Since h h h is nondecreasing and S i + 1 = S i + N i S_{i+1}=S_i+N_i S i + 1 = S i + N i ,
∑ i N i h ( S i / K ) ≤ K ∑ i ∫ S i / K S i + 1 / K h ( y ) d y = K H ( S / K ) . \sum_iN_ih(S_i/K)\le K\sum_i\int_{S_i/K}^{S_{i+1}/K}h(y)\,\mathrm{d}y=KH(S/K). i ∑ N i h ( S i / K ) ≤ K i ∑ ∫ S i / K S i + 1 / K h ( y ) d y = K H ( S / K ) .
By independence of N i N_i N i and S i S_i S i , ∑ i τ i E h ( S i / K ) ≤ K E H ( S / K ) \sum_i\tau_i\mathbb{E}h(S_i/K)\le K\mathbb{E}H(S/K) ∑ i τ i E h ( S i / K ) ≤ K E H ( S / K ) . As in the proof of Lemma B.2 , S ≤ c x L Z ν 0 S\le_{\mathrm{cx}}LZ_{\nu_0} S ≤ cx L Z ν 0 with ν 0 = W / L ≤ K ν \nu_0=W/L\le K\nu ν 0 = W / L ≤ K ν . The map x ↦ H ( x / K ) x\mapsto H(x/K) x ↦ H ( x / K ) is convex and H ( L x ) = L H ν ( x ) H(Lx)=LH_\nu(x) H ( Lx ) = L H ν ( x ) , so E H ( S / K ) ≤ L ψ ( ν 0 ) \mathbb{E}H(S/K)\le L\psi(\nu_0) E H ( S / K ) ≤ L ψ ( ν 0 ) with ψ ( u ) = E H ν ( Z u / K ) \psi(u)=\mathbb{E}H_\nu(Z_u/K) ψ ( u ) = E H ν ( Z u / K ) . The second derivative of ψ \psi ψ is the expected second difference of the convex function z ↦ H ν ( z / K ) z \mapsto H_\nu(z/K) z ↦ H ν ( z / K ) at Z ν Z_\nu Z ν , so ψ \psi ψ is convex with ψ ( 0 ) = 0 \psi(0)=0 ψ ( 0 ) = 0 , and ψ ( ν 0 ) ≤ ( ν 0 / K ν ) ψ ( K ν ) \psi(\nu_0) \le(\nu_0/K\nu)\psi(K\nu) ψ ( ν 0 ) ≤ ( ν 0 / K ν ) ψ ( K ν ) . Therefore
∑ i τ i E h ( S i / K ) ≤ L ν 0 ν ψ ( K ν ) = ( 1 − ρ P , K ( ν ) ) W . \sum_i \tau_i \mathbb{E}h(S_i/K) \le\frac{L\nu_0}{\nu}\psi(K\nu)=(1-\rho_{P,K}(\nu))W. i ∑ τ i E h ( S i / K ) ≤ ν L ν 0 ψ ( K ν ) = ( 1 − ρ P , K ( ν )) W .
Both R i / τ i R_i/\tau_i R i / τ i and 1 − E h ( S i / K ) 1-\mathbb{E}h(S_i/K) 1 − E h ( S i / K ) are nonincreasing in i i i , and Chebyshev’s sum inequality gives ∑ i R i ( 1 − E h ( S i / K ) ) ≥ ρ P , K ( ν ) J \sum_i R_i(1-\mathbb{E}h(S_i/K)) \ge\rho_{P,K}(\nu)J ∑ i R i ( 1 − E h ( S i / K )) ≥ ρ P , K ( ν ) J .
Dependence on K K K
Let ν ≥ 1 \nu\ge1 ν ≥ 1 . (i) ρ P , K ( ν ) \rho_{P,K}(\nu) ρ P , K ( ν ) is nondecreasing in K K K . (ii) For every K K K ,
1 − 1 2 ν − E ( Z K ν − K ν ) + K ν ≤ ρ P , K ( ν ) ≤ 1 − 1 2 ν . 1-\frac{1}{2\nu}-\frac{\mathbb{E}(Z_{K\nu}-K\nu)^{+}}{K\nu}\le\rho_{P,K}(\nu)\le1-\frac{1}{2\nu}. 1 − 2 ν 1 − K ν E ( Z K ν − K ν ) + ≤ ρ P , K ( ν ) ≤ 1 − 2 ν 1 .
The subtracted term is at most ( 4 K ν ) − 1 / 2 (4K\nu)^{-1/2} ( 4 K ν ) − 1/2 , and equals Pr ( Z K ν = K ν ) ≤ ( 2 π K ν ) − 1 / 2 \Pr(Z_{K\nu}=K\nu)\le(2\pi K\nu)^{-1/2} Pr ( Z K ν = K ν ) ≤ ( 2 π K ν ) − 1/2 when K ν K\nu K ν is an integer. In particular ρ P , K ( ν ) → 1 − 1 / ( 2 ν ) = 1 − L / ( 2 T ) \rho_{P,K}(\nu)\to1-1/(2\nu)=1-L/(2T) ρ P , K ( ν ) → 1 − 1/ ( 2 ν ) = 1 − L / ( 2 T ) as K → ∞ K\to\infty K → ∞ .
Let X 1 , X 2 , … X_1,X_2,\ldots X 1 , X 2 , … be i.i.d. Poisson( ν ) (\nu) ( ν ) and X ˉ K = K − 1 ∑ k ≤ K X k \bar{X}_K=K^{-1}\sum_{k\le K}X_k X ˉ K = K − 1 ∑ k ≤ K X k ; then Z K ν / K Z_{K\nu}/K Z K ν / K has the law of X ˉ K \bar{X}_K X ˉ K . (i) For k ≤ K + 1 k\le K+1 k ≤ K + 1 let X ˉ K ( k ) \bar{X}^{(k)}_K X ˉ K ( k ) be the mean of X 1 , … , X K + 1 X_1,\ldots,X_{K+1} X 1 , … , X K + 1 without X k X_k X k . By exchangeability the conditional expectations E [ X ˉ K ( k ) ∣ X ˉ K + 1 ] \mathbb{E}[\bar{X}^{(k)}_K\mid\bar{X}_{K+1}] E [ X ˉ K ( k ) ∣ X ˉ K + 1 ] coincide, and their average is X ˉ K + 1 \bar{X}_{K+1} X ˉ K + 1 . Conditional Jensen gives E f ( X ˉ K + 1 ) ≤ E f ( X ˉ K ( 1 ) ) = E f ( X ˉ K ) \mathbb{E}f(\bar{X}_{K+1})\le\mathbb{E}f(\bar{X}^{(1)}_K)=\mathbb{E}f(\bar{X}_K) E f ( X ˉ K + 1 ) ≤ E f ( X ˉ K ( 1 ) ) = E f ( X ˉ K ) for convex f f f ; take f = H ν f=H_\nu f = H ν . (ii) Since ν ≥ 1 \nu\ge1 ν ≥ 1 , H ν ( ν ) = ∫ ν − 1 ν ( y − ν + 1 ) d y = 1 2 H_\nu(\nu)=\int_{\nu-1}^{\nu}(y-\nu+1)\,dy=\frac12 H ν ( ν ) = ∫ ν − 1 ν ( y − ν + 1 ) d y = 2 1 , and Jensen gives E H ν ( X ˉ K ) ≥ 1 2 \mathbb{E}H_\nu(\bar{X}_K)\ge\frac12 E H ν ( X ˉ K ) ≥ 2 1 , the upper bound. Also H ν ( x ) ≤ 1 2 + ( x − ν ) + H_\nu(x)\le\frac12+(x-\nu)^+ H ν ( x ) ≤ 2 1 + ( x − ν ) + : for x ≤ ν x\le\nu x ≤ ν because H ν H_\nu H ν is nondecreasing, and for x ≥ ν x\ge\nu x ≥ ν with equality. Hence E H ν ( X ˉ K ) ≤ 1 2 + E ( Z K ν − K ν ) + / K \mathbb{E}H_\nu(\bar{X}_K)\le\frac12+\mathbb{E}(Z_{K\nu}-K\nu)^+/K E H ν ( X ˉ K ) ≤ 2 1 + E ( Z K ν − K ν ) + / K , the lower bound. Finally E ( Z m − m ) + = 1 2 E ∣ Z m − m ∣ ≤ 1 2 m \mathbb{E}(Z_m-m)^+=\frac12\mathbb{E}|Z_m-m|\le\frac12\sqrt{m} E ( Z m − m ) + = 2 1 E ∣ Z m − m ∣ ≤ 2 1 m , and for integer m m m , E ( Z m − m ) + = m Pr ( Z m = m ) ≤ m ( 2 π m ) − 1 / 2 \mathbb{E}(Z_m-m)^+=m\Pr(Z_m=m)\le m(2\pi m)^{-1/2} E ( Z m − m ) + = m Pr ( Z m = m ) ≤ m ( 2 π m ) − 1/2 .
Example B.6 (Finding time with K K K plays per step). For every K ≥ 1 K\ge1 K ≥ 1 and integers 1 ≤ L ≤ T 1\le L\le T 1 ≤ L ≤ T with K T ≥ 2 KT\ge2 K T ≥ 2 there are undelayed one-shot instances with caps L L L and
OPT K ( P ) LP K ( P ) ≤ ( 1 − L − 1 2 T ) K T K T − 1 . \frac{\operatorname{OPT}_K(P)}{\operatorname{LP}_K(P)}\le\left(1-\frac{L-1}{2T}\right)\frac{KT}{KT-1}. LP K ( P ) OPT K ( P ) ≤ ( 1 − 2 T L − 1 ) K T − 1 K T . Take n = K T − 1 n=KT-1 n = K T − 1 arms whose episode mean is 1 1 1 with probability p ∈ ( 0 , 1 ) p\in(0,1) p ∈ ( 0 , 1 ) and 0 0 0 otherwise, where p ( L − 1 ) ≤ 1 / ( K T − 1 ) p(L-1)\le1/(KT-1) p ( L − 1 ) ≤ 1/ ( K T − 1 ) . Probing every arm and continuing to the cap after a success uses n ( 1 + p ( L − 1 ) ) ≤ K T n(1+p(L-1))\le KT n ( 1 + p ( L − 1 )) ≤ K T plays in expectation, so LP K ( P ) ≥ n p L \operatorname{LP}_K(P)\ge npL LP K ( P ) ≥ n p L . Fix an admissible policy and let t i t_i t i be the step of arm i i i ’s first play (t i = ∞ t_i=\infty t i = ∞ if none). That decision uses no outcome of arm i i i , so arm i i i ’s mean is independent of t i t_i t i , and arm i i i earns at most its mean times min { L , ( T − t i + 1 ) + } \min\{L,(T-t_i+1)^+\} min { L , ( T − t i + 1 ) + } . At most K K K arms are first played at any step, and min { L , T − t + 1 } \min\{L,T-t+1\} min { L , T − t + 1 } is nonincreasing in t t t , so
OPT K ( P ) ≤ p K ∑ t = 1 T min { L , T − t + 1 } = p K L ( T − L − 1 2 ) . \operatorname{OPT}_K(P)\le pK\sum_{t=1}^{T}\min\{L,T-t+1\}=pKL\left(T-\frac{L-1}{2}\right). OPT K ( P ) ≤ p K t = 1 ∑ T min { L , T − t + 1 } = p K L ( T − 2 L − 1 ) . Dividing by n p L npL n p L proves the claim.
As T → ∞ T\to\infty T → ∞ with L / T → x ∈ ( 0 , 1 ] L/T\to x\in(0,1] L / T → x ∈ ( 0 , 1 ] , this bound tends to 1 − x / 2 1-x/2 1 − x /2 , matching the many-lane limit above. The uniform factor of Theorem 3.13 is therefore tight relative to LP K \operatorname{LP}_K LP K in the limit of many lanes. For each fixed K K K , taking L = T → ∞ L=T\to\infty L = T → ∞ makes the displayed upper bound tend to 1 / 2 1/2 1/2 ; the theorem attains that uniform factor for every K K K . For K = 1 K=1 K = 1 and L = T L=T L = T , this recovers the limiting factor 1 / 2 1/2 1/2 of Proposition 3.12. The common value 1 − x / 2 = ∫ 0 1 min { 1 , ( 1 − s ) / x } d s 1-x/2=\int_0^1\min\{1,(1-s)/x\}\,ds 1 − x /2 = ∫ 0 1 min { 1 , ( 1 − s ) / x } d s is the fraction a job of length x T xT x T keeps when it starts at a uniform time: GM13’s triangle, with the job’s own length in place of the work ahead of it.
Statistical bounds from reported experiments
We treat two drift models: a known affine path between unknown endpoint priors (Theorem C.4 ), and general drift that is slow in the laws of short probes. Throughout this appendix, rewards have finite alphabets Y i Y_i Y i of sizes d i d_i d i . Fix caps 1 ≤ L i ≤ L ‾ 1 \le L_i \le\overline{L} 1 ≤ L i ≤ L and adaptation depths k i k_i k i across episodes. Slow arms have k i = 0 k_i=0 k i = 0 . Write k = max i k i k = \max_i k_i k = max i k i and use L P ( k ) \mathrm{LP}^{(k)} LP ( k ) for these arm-specific classes. Put r i = min { L i , k i + 1 } r_i = \min\{L_i,k_i+1\} r i = min { L i , k i + 1 } . The comparison budgets do not bind: B i ≥ N L i B_i \ge NL_i B i ≥ N L i . The drift law is known, as in Assumption 2.1 . The realized priors are learned from reports. Earlier reports arrive only at episode starts, with lag at most g g g .
The supplied single-episode planner has a uniform completion certificate κ L P ( k ) ( P ) \kappa\mathrm{LP}^{(k)}(P) κ LP ( k ) ( P ) for every product prior P P P . Theorem 3.9 and Corollary 3.10 supply such certificates. Any valid single-episode factor can be used. Restricting a slow arm to nonadaptive local policies is essential to that certificate. The oracle knows the current priors, not the realized qualities θ i ( e ) \theta_i^{(e)} θ i ( e ) .
Stationary comparisons
Proof of Corollary 4.6. Let q ^ e , i \widehat{q}_{e,i} q e , i be the empirical law of arm i i i ’s probe histograms from episodes 1 , … , ( e − g ) + 1,\ldots,(e-g)^+ 1 , … , ( e − g ) + , using a fixed probability vector if there are none. All these reports are available. Conditional on P P P , arm i i i ’s probe histograms are independent with law q r i ( P i ) q_{r_i}(P_i) q r i ( P i ) . For m > 0 m>0 m > 0 samples, the coordinate variances sum to at most 1 / m 1/m 1/ m , giving expected total-variation error at most 1 2 D i / m \frac{1}{2}\sqrt{D_i/m} 2 1 D i / m . Among h h h earlier episodes whose reports have arrived, the sample count is M ∼ Binomial ( h , p ) M \sim\operatorname{Binomial}(h,p) M ∼ Binomial ( h , p ) . Including M = 0 M=0 M = 0 and using E [ 1 / ( M + 1 ) ] ≤ 1 / [ p ( h + 1 ) ] \mathbb{E}[1/(M+1)] \le1/[p(h+1)] E [ 1/ ( M + 1 )] ≤ 1/ [ p ( h + 1 )] bounds the error by D i / [ p ( h + 1 ) ] \sqrt{D_i/[p(h+1)]} D i / [ p ( h + 1 )] . Summing over episodes, with ∑ h = 0 N − 1 ( h + 1 ) − 1 / 2 ≤ 2 N \sum_{h=0}^{N-1}(h+1)^{-1/2} \le2\sqrt{N} ∑ h = 0 N − 1 ( h + 1 ) − 1/2 ≤ 2 N , gives
∑ e , i L i E ∥ q r i ( P i ) − q ^ e , i ∥ T V ≤ 2 C e s t N / ξ + ( g − 1 ) L Σ . \sum_{e,i} L_i\mathbb{E}\lVert q_{r_i}(P_i)-\widehat{q}_{e,i}\rVert_{\mathrm{TV}} \le2C_{\mathrm{est}}\sqrt{N/\xi}+(g-1)L_\Sigma. e , i ∑ L i E ∥ q r i ( P i ) − q e , i ∥ TV ≤ 2 C est N / ξ + ( g − 1 ) L Σ .
The planning branch has the conditional floor ( 1 − ξ ) κ L P ( k ) ( P ‾ e ) (1-\xi)\kappa\mathrm{LP}^{(k)}(\overline{P}_e) ( 1 − ξ ) κ LP ( k ) ( P e ) . Apply Corollary 4.3, and charge at most κ ξ N T \kappa\xi NT κ ξ N T to restore the factor κ \kappa κ . The resulting bound is
κ [ ξ N T + 4 C e s t N / ξ + 2 ( g − 1 ) L Σ ] . \kappa\left[\xi NT+4C_{\mathrm{est}}\sqrt{N/\xi}+2(g-1)L_\Sigma\right]. κ [ ξ N T + 4 C est N / ξ + 2 ( g − 1 ) L Σ ] .
The stated choice of ξ \xi ξ gives (32) when unclipped. If clipping occurs, ( T C e s t 2 ) 1 / 3 N 2 / 3 > N T / 2 (TC_{\mathrm{est}}^2)^{1/3}N^{2/3}>NT/2 ( T C est 2 ) 1/3 N 2/3 > N T /2 , so the trivial bound κ N T \kappa NT κ N T suffices.
For the protected learner, use the same calculation with one probe group, C e s t = ∑ i L i D i C_{\mathrm{est}}=\sum_i L_i\sqrt{D_i} C est = ∑ i L i D i . Theorem 4.5 gives
κ p l ∑ e E L P ( k ) ( P ) − V A ≤ 5 κ p l ( T C e s t 2 ) 1 / 3 N 2 / 3 + 2 κ p l ( g − 1 ) L Σ . (50) \kappa_{\mathrm{pl}}\sum_e \mathbb{E}\mathrm{LP}^{(k)}(P)-V_A \le5\kappa_{\mathrm{pl}}(TC_{\mathrm{est}}^2)^{1/3}N^{2/3}+2\kappa_{\mathrm{pl}}(g-1)L_\Sigma. \tag*{(50)} κ pl e ∑ E LP ( k ) ( P ) − V A ≤ 5 κ pl ( T C est 2 ) 1/3 N 2/3 + 2 κ pl ( g − 1 ) L Σ . (50)
Use the empirical probe estimates above, bound 1 − ξ 1-\xi 1 − ξ by one in the error term, charge κ p l ξ N T \kappa_{\mathrm{pl}}\xi NT κ pl ξ N T to restore the benchmark, and optimize ξ \xi ξ as before. A planning gap ζ \zeta ζ adds c ζ c\zeta c ζ .
Reported probes and estimator comparisons
Use Algorithm 3, with probe probability p = ξ / J g r p p=\xi/J_{\mathrm{grp}} p = ξ / J grp and constants from (30) . The following bounds specialize Corollary 4.3 to reported probe data.
For general drift, define
V = E ∑ e = 1 N − 1 ∑ i L i d r i ( P i ( e + 1 ) , P i ( e ) ) . (51) \mathcal{V}=\mathbb{E}\sum_{e=1}^{N-1}\sum_i L_i\,d_{r_i}\left(P_i^{(e+1)},P_i^{(e)}\right). \tag*{(51)} V = E e = 1 ∑ N − 1 i ∑ L i d r i ( P i ( e + 1 ) , P i ( e ) ) . (51)
We call the drift slow when V = o ( N ) \mathcal{V}=o(N) V = o ( N ) , with the other parameters fixed. This definition allows occasional jumps. Changes that preserve the r i r_i r i -play probe law contribute zero. For a slow arm ( δ i ≥ T ) (\delta_i \ge T) ( δ i ≥ T ) , r i = 1 r_i=1 r i = 1 and the one-reward law controls its current reward and workload. Longer reported batches can still reveal features of the recurring law that matter to later decisions.
Write Δ i , u = d r i ( P i ( u + 1 ) , P i ( u ) ) \Delta_{i,u}=d_{r_i}(P_i^{(u+1)},P_i^{(u)}) Δ i , u = d r i ( P i ( u + 1 ) , P i ( u ) ) , so V = ∑ i L i ∑ u < N E Δ i , u \mathcal{V}=\sum_i L_i\sum_{u<N}\mathbb{E}\Delta_{i,u} V = ∑ i L i ∑ u < N E Δ i , u . The proofs below use the history factorization and conditional predictive prior from § 3 . The experiment law and its planning-value bound are (27) and Corollary 4.3.
Put κ ξ = ( 1 − ξ ) κ \kappa_{\xi}=(1-\xi)\kappa κ ξ = ( 1 − ξ ) κ and define the scaled comparison loss
Reg N = E ∑ e = 1 N ( κ ξ LP ( k ) ( P ( e ) ) − V e ) . \operatorname{Reg}_N=\mathbb{E}\sum_{e=1}^{N}\left(\kappa_{\xi}\operatorname{LP}^{(k)}(P^{(e)})-V_e\right). Reg N = E e = 1 ∑ N ( κ ξ LP ( k ) ( P ( e ) ) − V e ) .
This is an approximation regret in the sense of Chen et al. [11 ] . All expectations average under the meta-prior and the policy’s coins.
Theorem C.1 (Floor and comparison with any estimator). Algorithm 3 respects every cap and budget, and in every episode E [ V e ∣ H e − 1 ] ≥ κ ξ LP ( k ) ( P ‾ e ) \mathbb{E}[V_e\mid H_{e-1}]\ge\kappa_{\xi}\operatorname{LP}^{(k)}(\overline{P}_e) E [ V e ∣ H e − 1 ] ≥ κ ξ LP ( k ) ( P e ) . For any history-measurable vectors q ^ e , i ∈ R D i \widehat{q}_{e,i}\in\mathbb{R}^{D_i} q e , i ∈ R D i ,
Reg N ≤ 2 κ ξ ∑ i L i ∑ e = 1 N E ∥ q r i ( P i ( e ) ) − q ^ e , i ∥ T V . \operatorname{Reg}_N\le2\kappa_{\xi}\sum_i L_i\sum_{e=1}^{N}\mathbb{E}\lVert q_{r_i}(P_i^{(e)})-\widehat{q}_{e,i}\rVert_{\mathrm{TV}}. Reg N ≤ 2 κ ξ i ∑ L i e = 1 ∑ N E ∥ q r i ( P i ( e ) ) − q e , i ∥ TV . The estimators need not correspond to realizable priors.
Proof. Both branches respect the caps and report calendar, and comparison budgets do not bind. The planning branch and Theorem 4.4 give the conditional floor. Apply Corollary 4.3 to the available history, multiply by κ ξ \kappa_{\xi} κ ξ and sum over episodes.
Estimating moving experiment laws
Lemma C.2 (Windowed estimate). Fix an integer 1 ≤ τ ≤ N 1\le\tau\le N 1 ≤ τ ≤ N . In episode e e e , use the window I e = { max { 1 , e − g − τ + 1 } , … , e − g } I_e=\{\max\{1,e-g-\tau+1\},\ldots,e-g\} I e = { max { 1 , e − g − τ + 1 } , … , e − g } of h e = ∣ I e ∣ = min { τ , ( e − g ) + } h_e=|I_e|=\min\{\tau,(e-g)^+\} h e = ∣ I e ∣ = min { τ , ( e − g ) + } episodes. The window is empty when e ≤ g e\le g e ≤ g . Let M e , i M_{e,i} M e , i be the number of probes of arm i i i in I e I_e I e . Let q ^ e , i \widehat{q}_{e,i} q e , i be the empirical distribution of their histograms, or a fixed probability vector if M e , i = 0 M_{e,i}=0 M e , i = 0 . These quantities are history-measurable. For e ≥ g e\ge g e ≥ g ,
E ∥ q r i ( P i ( e ) ) − q ^ e , i ∥ T V ≤ D i p ( h e + 1 ) + b e , i , b e , i = E max t ∈ I e d r i ( P i ( t ) , P i ( e ) ) . \mathbb{E}\lVert q_{r_i}(P_i^{(e)})-\widehat{q}_{e,i}\rVert_{\mathrm{TV}}\le\sqrt{\frac{D_i}{p(h_e+1)}}+b_{e,i},\qquad b_{e,i}=\mathbb{E}\max_{t\in I_e}d_{r_i}(P_i^{(t)},P_i^{(e)}). E ∥ q r i ( P i ( e ) ) − q e , i ∥ TV ≤ p ( h e + 1 ) D i + b e , i , b e , i = E t ∈ I e max d r i ( P i ( t ) , P i ( e ) ) . Here b e , i = 0 b_{e,i}=0 b e , i = 0 when I e I_e I e is empty, and b e , i ≤ E ∑ u = max { 1 , e − g − τ + 1 } e − 1 Δ i , u b_{e,i}\le\mathbb{E}\sum_{u=\max\{1,e-g-\tau+1\}}^{e-1}\Delta_{i,u} b e , i ≤ E ∑ u = m a x { 1 , e − g − τ + 1 } e − 1 Δ i , u .
Proof. Condition on the true prior sequences and probe times. The episode qualities are conditionally independent, and the probes are nonadaptive. The probe histograms are therefore independent with categorical laws q r i ( P i ( t ) ) q_{r_i}(P_i^{(t)}) q r i ( P i ( t ) ) . Suppose M e , i = m > 0 M_{e,i}=m>0 M e , i = m > 0 . The coordinate variances of the empirical distribution sum to at most 1 / m 1/m 1/ m , so by Cauchy–Schwarz it differs from the average of these laws by at most 1 2 D i / m \frac{1}{2}\sqrt{D_i/m} 2 1 D i / m in expected total variation. The average differs from q r i ( P i ( e ) ) q_{r_i}(P_i^{(e)}) q r i ( P i ( e ) ) by at most max t ∈ I e d r i ( P i ( t ) , P i ( e ) ) \max_{t\in I_e}d_{r_i}(P_i^{(t)},P_i^{(e)}) max t ∈ I e d r i ( P i ( t ) , P i ( e ) ) . Including m = 0 m=0 m = 0 , where the error is at most 1, the statistical term is at most D i / ( m + 1 ) \sqrt{D_i/(m+1)} D i / ( m + 1 ) . Each episode probes arm i i i with probability p p p , independently of the sequences. Thus M e , i ∼ Binomial ( h e , p ) M_{e,i}\sim\operatorname{Binomial}(h_e,p) M e , i ∼ Binomial ( h e , p ) . The first bound follows from Jensen’s inequality and the identity
E 1 M e , i + 1 = 1 − ( 1 − p ) h e + 1 p ( h e + 1 ) ≤ 1 p ( h e + 1 ) . \mathbb{E}\frac{1}{M_{e,i}+1}=\frac{1-(1-p)^{h_e+1}}{p(h_e+1)}\le\frac{1}{p(h_e+1)}. E M e , i + 1 1 = p ( h e + 1 ) 1 − ( 1 − p ) h e + 1 ≤ p ( h e + 1 ) 1 . For t ∈ I e t\in I_e t ∈ I e the triangle inequality gives d r i ( P i ( t ) , P i ( e ) ) ≤ ∑ u = t e − 1 Δ i , u d_{r_i}(P_i^{(t)},P_i^{(e)})\le\sum_{u=t}^{e-1}\Delta_{i,u} d r i ( P i ( t ) , P i ( e ) ) ≤ ∑ u = t e − 1 Δ i , u , and t ≥ e − g − τ + 1 t\ge e-g-\tau+1 t ≥ e − g − τ + 1 .
Theorem C.3 (Tracking with a reward floor). Under the conditions above, use the execution factor κ \kappa κ and the scaled comparison Reg N \operatorname{Reg}_{N} Reg N defined above. Algorithm 3 satisfies
E [ V e ∣ H e − 1 ] ≥ κ ξ L P ( k ) ( P ˉ e ) \mathbb{E}[V_e \mid H_{e-1}] \ge\kappa\xi L^{\mathrm{P}}(k)(\bar{P}_e) E [ V e ∣ H e − 1 ] ≥ κ ξ L P ( k ) ( P ˉ e ) in every episode. For every integer 1 ≤ τ ≤ N 1 \le\tau\le N 1 ≤ τ ≤ N ,
Reg N ≤ κ ξ c ξ [ N τ + τ V + ( g − 1 ) ( 1 + V ) ] . \operatorname{Reg}_{N} \le\kappa\xi c_{\xi}\left[\frac{N}{\sqrt{\tau}}+\tau\mathcal{V}+(g-1)(1+\mathcal{V})\right]. Reg N ≤ κ ξ c ξ [ τ N + τ V + ( g − 1 ) ( 1 + V ) ] . where c ξ = 2 max { 2 C e s t / ξ , L Σ , 1 } c_{\xi}=2\max\{2C_{\mathrm{est}}/\sqrt{\xi},L_{\Sigma},1\} c ξ = 2 max { 2 C est / ξ , L Σ , 1 } . Thus c ξ = O ( ξ − 1 / 2 ) c_{\xi}=O(\xi^{-1/2}) c ξ = O ( ξ − 1/2 ) . Its other dependence is explicit; it does not depend on N N N , V \mathcal{V} V , g g g or τ \tau τ .
Proof of Theorem C.3. The floor is Theorem C.1. Use q ^ e , i \hat{q}_{e,i} q ^ e , i of Lemma C.2 in Theorem C.1. The first g − 1 g-1 g − 1 episodes contribute at most 1 each to the sum of total variations, hence ( g − 1 ) L Σ (g-1)L_{\Sigma} ( g − 1 ) L Σ after weighting. For the others put e ′ = e − g + 1 e'=e-g+1 e ′ = e − g + 1 , so that h e = min { τ , ( e ′ − 1 ) + } h_e=\min\{\tau,(e'-1)^{+}\} h e = min { τ , ( e ′ − 1 ) + } and
∑ e ′ = 1 N 1 h e + 1 ≤ 2 τ + N − τ τ + 1 ≤ 2 N τ , \sum_{e'=1}^{N}\frac{1}{\sqrt{h_e+1}}\le2\sqrt{\tau}+\frac{N-\tau}{\sqrt{\tau+1}}\le\frac{2N}{\sqrt{\tau}}, e ′ = 1 ∑ N h e + 1 1 ≤ 2 τ + τ + 1 N − τ ≤ τ 2 N ,
which with p = ξ / J g r p p=\xi/J_{\mathrm{grp}} p = ξ / J grp bounds the statistical terms by 2 N C e s t / ξ τ 2NC_{\mathrm{est}}/\sqrt{\xi\tau} 2 N C est / ξ τ after weighting by L i L_i L i . For the bias, each increment Δ i , u \Delta_{i,u} Δ i , u appears in b e , i b_{e,i} b e , i for at most τ + g − 1 \tau+g-1 τ + g − 1 values of e e e , so the weighted bias sums to at most ( τ + g − 1 ) V (\tau+g-1)\mathcal{V} ( τ + g − 1 ) V . Multiplying by 2 κ ξ 2\kappa\xi 2 κ ξ gives the explicit window bound
Reg N ≤ 2 κ ξ [ 2 N C e s t ξ τ + τ V + ( g − 1 ) ( L Σ + V ) ] . (52) \operatorname{Reg}_{N}\le2\kappa\xi\left[\frac{2NC_{\mathrm{est}}}{\sqrt{\xi\tau}}+\tau\mathcal{V}+(g-1)(L_{\Sigma}+\mathcal{V})\right]. \tag*{(52)} Reg N ≤ 2 κ ξ [ ξ τ 2 N C est + τ V + ( g − 1 ) ( L Σ + V ) ] . (52)
Taking c ξ = 2 max { 2 C e s t / ξ , L Σ , 1 } c_{\xi}=2\max\{2C_{\mathrm{est}}/\sqrt{\xi},L_{\Sigma},1\} c ξ = 2 max { 2 C est / ξ , L Σ , 1 } proves Theorem C.3 in its stated form. Optimizing over τ \tau τ , with τ ≍ ( N C e s t / ( ξ V ) ) 2 / 3 \tau\asymp(NC_{\mathrm{est}}/(\sqrt{\xi}\mathcal{V}))^{2/3} τ ≍ ( N C est / ( ξ V ) ) 2/3 rounded and clipped to [ 1 , N ] [1,N] [ 1 , N ] , and τ = N \tau=N τ = N when V = 0 \mathcal{V}=0 V = 0 , gives
Reg N = O ( κ ξ [ C e s t N / ξ + ( N C e s t / ξ ) 2 / 3 V 1 / 3 + ( g − 1 ) ( L Σ + V ) ] ) . \operatorname{Reg}_{N}=O\left(\kappa\xi\left[C_{\mathrm{est}}\sqrt{N/\xi}+(NC_{\mathrm{est}}/\sqrt{\xi})^{2/3}\mathcal{V}^{1/3}+(g-1)(L_{\Sigma}+\mathcal{V})\right]\right). Reg N = O ( κ ξ [ C est N / ξ + ( N C est / ξ ) 2/3 V 1/3 + ( g − 1 ) ( L Σ + V ) ] ) .
The lower endpoint never binds, since V ≤ ( N − 1 ) L Σ ≤ N C e s t / ξ \mathcal{V}\le(N-1)L_{\Sigma}\le NC_{\mathrm{est}}/\sqrt{\xi} V ≤ ( N − 1 ) L Σ ≤ N C est / ξ ; when the upper endpoint binds, the square-root term covers the bound with τ = N \tau=N τ = N .
The window appears only in the proof; the learner uses its full posterior and needs neither a window nor a variation budget. For fixed parameters and ξ \xi ξ , the bound is O ( N ) O(\sqrt{N}) O ( N ) under stationarity, O ( N 2 / 3 ) O(N^{2/3}) O ( N 2/3 ) for bounded V \mathcal{V} V , and sublinear whenever V = o ( N ) \mathcal{V}=o(N) V = o ( N ) .
Known affine evolution. Local averaging incurs first-order error even when drift has a stable direction. A local linear estimate removes this error. We use that estimate only in the analysis. Algorithm 3 continues to use its full posterior under the known drift law. The realized endpoints or trajectory remain unknown. Take g = 1 g=1 g = 1 and 2 ≤ τ ≤ N 2\le\tau\le N 2 ≤ τ ≤ N . For h = min { τ , e − 1 } ≥ 2 h=\min\{\tau,e-1\}\ge2 h = min { τ , e − 1 } ≥ 2 give the last h h h episodes, of ages ℓ = 1 , … , h \ell=1,\ldots,h ℓ = 1 , … , h , the weights for linear extrapolation ν h , ℓ = ( 4 h + 2 − 6 ℓ ) / ( h ( h − 1 ) ) \nu_{h,\ell}=(4h+2-6\ell)/(h(h-1)) ν h , ℓ = ( 4 h + 2 − 6 ℓ ) / ( h ( h − 1 )) , which satisfy
∑ ℓ ν h , ℓ = 1 , ∑ ℓ ℓ ν h , ℓ = 0 , ∑ ℓ ν h , ℓ 2 = 2 ( 2 h + 1 ) h ( h − 1 ) ≤ 10 h . \sum_{\ell}\nu_{h,\ell}=1,\qquad\sum_{\ell}\ell\nu_{h,\ell}=0,\qquad\sum_{\ell}\nu_{h,\ell}^{2}=\frac{2(2h+1)}{h(h-1)}\le\frac{10}{h}. ℓ ∑ ν h , ℓ = 1 , ℓ ∑ ℓ ν h , ℓ = 0 , ℓ ∑ ν h , ℓ 2 = h ( h − 1 ) 2 ( 2 h + 1 ) ≤ h 10 .
For h < 2 h<2 h < 2 , use a fixed probability vector. Let A t , i A_{t,i} A t , i indicate a probe of arm i i i in episode t t t and S t , i S_{t,i} S t , i its outcome histogram, set to an arbitrary fixed histogram when there is no probe. For h ≥ 2 h\ge2 h ≥ 2 , set q ^ e , i = ∑ ℓ = 1 h ν h , ℓ ( A e − ℓ , i / p ) e S e − ℓ , i \hat{q}_{e,i}=\sum_{\ell=1}^{h}\nu_{h,\ell}(A_{e-\ell,i}/p)e_{S_{e-\ell,i}} q ^ e , i = ∑ ℓ = 1 h ν h , ℓ ( A e − ℓ , i / p ) e S e − ℓ , i , with e s e_s e s a coordinate vector. The estimate need not be a probability vector. Theorem C.1 permits such centers. Given the sequence of priors, its mean is ∑ ℓ υ h , ℓ α r i ( P i ( e − ℓ ) ) \sum_{\ell} \upsilon_{h,\ell}\boldsymbol{\alpha}_{r_i}(P_i^{(e-\ell)}) ∑ ℓ υ h , ℓ α r i ( P i ( e − ℓ ) ) . One inverse-probability weighted categorical vector has summed coordinate variances at most 1 / p 1/p 1/ p , and the episodes are independent given the sequence. Hence
E [ ∥ q ^ e , i − E [ q ^ e , i ∣ ( P i ( t ) ) t ] ∥ T V ] ≤ 1 2 D i p ∑ ℓ υ h , ℓ 2 . \mathbb{E}\left[\left\|\widehat{\mathbf{q}}_{e,i}-\mathbb{E}[\widehat{\mathbf{q}}_{e,i}\mid(P_i^{(t)})_t]\right\|_{\mathrm{TV}}\right]\leq\frac{1}{2}\sqrt{\frac{D_i}{p}}\sum_{\ell}\upsilon_{h,\ell}^{2}. E [ q e , i − E [ q e , i ∣ ( P i ( t ) ) t ] TV ] ≤ 2 1 p D i ℓ ∑ υ h , ℓ 2 .
Theorem C.4 (Unknown affine mixtures). Suppose that for each arm there are unknown random endpoint priors P i L , P i R P_i^{\mathrm{L}},P_i^{\mathrm{R}} P i L , P i R such that P i ( e ) = ( 1 − e / N ) P i L + ( e / N ) P i R P_i^{(e)}=(1-e/N)P_i^{\mathrm{L}}+(e/N)P_i^{\mathrm{R}} P i ( e ) = ( 1 − e / N ) P i L + ( e / N ) P i R , with a known meta-prior on the endpoint pairs, independent across arms. For g = 1 g=1 g = 1 and N ≥ 2 N\geq2 N ≥ 2 .
Reg N ≤ 4 κ ξ L Σ + 2 10 κ ξ N / ξ C e s t . \operatorname{Reg}_N \leq4\kappa_{\xi}L_{\Sigma}+2\sqrt{10}\,\kappa_{\xi}\sqrt{N/\xi}\,C_{\mathrm{est}}. Reg N ≤ 4 κ ξ L Σ + 2 10 κ ξ N / ξ C est . In particular, for J g r p = 1 J_{\mathrm{grp}}=1 J grp = 1 this is O ( L ‾ n D max N / ξ + L ‾ n ) O(\overline{L}n\sqrt{D_{\max}N/\xi}+\overline{L}n) O ( L n D m a x N / ξ + L n ) , where D max = max i D i D_{\max}=\max_i D_i D m a x = max i D i .
Proof. By linearity, α r i ( P i ( e ) ) = ( 1 − e / N ) α r i ( P i L ) + ( e / N ) α r i ( P i R ) \boldsymbol{\alpha}_{r_i}(P_i^{(e)})=(1-e/N)\boldsymbol{\alpha}_{r_i}(P_i^{\mathrm{L}})+(e/N)\boldsymbol{\alpha}_{r_i}(P_i^{\mathrm{R}}) α r i ( P i ( e ) ) = ( 1 − e / N ) α r i ( P i L ) + ( e / N ) α r i ( P i R ) , and the first two identities for the weights make the estimator unbiased for the current outcome law, whatever the endpoint priors. Take τ = N \tau=N τ = N , bound the first two episodes by total variation one each, and use the variance bound above. For e ≥ 3 e\geq3 e ≥ 3 , h = e − 1 h=e-1 h = e − 1 and ∑ h = 2 N − 1 h − 1 / 2 ≤ 2 N \sum_{h=2}^{N-1}h^{-1/2}\leq2\sqrt{N} ∑ h = 2 N − 1 h − 1/2 ≤ 2 N , which gives a statistical sum of at most 10 D i N / p \sqrt{10D_iN/p} 10 D i N / p per arm. Multiply by 2 κ ξ L i 2\kappa_{\xi}L_i 2 κ ξ L i as in Theorem C.1 and sum over arms.
The priors may travel a nonvanishing distance; the rate comes from the known linear form of the probe laws, not from a vanishing total distance.
Transfer bounds
Algorithm 1 plans for cumulative reward. Its Bayes-optimal comparison value is at least Algorithm 3’s reward for every ξ \xi ξ . To restore the benchmark factor from κ ξ \kappa_{\xi} κ ξ to κ \kappa κ , charge at most ξ N T \xi NT ξ N T for exploration. Optimize this charge together with the estimation and drift costs.
Corollary C.5 (The learning-LP learner under slow drift). Run Algorithm 1 with an optimal solution. Suppose the uniform completion certificate of Theorem 3.7 holds, B i ′ ≥ N L i B_i'\geq NL_i B i ′ ≥ N L i , and 1 − ω = C / T > 0 1-\omega=C/T>0 1 − ω = C / T > 0 as in Theorem 3.7. Its expected cumulative regret against ( 1 − ω ) κ L P ( k ) ( P ( e ) ) (1-\omega)\kappa\mathrm{LP}^{(k)}(P^{(e)}) ( 1 − ω ) κ LP ( k ) ( P ( e ) ) satisfies
R N ( 1 ) ≤ ( 1 − ω ) κ c i n s t [ N 2 / 3 + V 1 / 4 N 3 / 4 ] + 1. \mathcal{R}_N^{(1)}\leq(1-\omega)\kappa c_{\mathrm{inst}}\left[N^{2/3}+\mathcal{V}^{1/4}N^{3/4}\right]+1. R N ( 1 ) ≤ ( 1 − ω ) κ c inst [ N 2/3 + V 1/4 N 3/4 ] + 1. Constant c i n s t c_{\mathrm{inst}} c inst depends on horizon, arms, caps, depths and report lag. It does not depend on N N N or V \mathcal{V} V . For the affine model with g = 1 g=1 g = 1 , the bound improves to O ( N 2 / 3 ) + 1 O(N^{2/3})+1 O ( N 2/3 ) + 1 for fixed parameters.
For every fixed ξ \xi ξ the probing learner is admissible for the budgets B ′ B' B ′ , so OPT N ( B ′ ) ≥ E ∑ e V e p r o b e \operatorname{OPT}_N(B')\geq\mathbb{E}\sum_e V_e^{\mathrm{probe}} OPT N ( B ′ ) ≥ E ∑ e V e probe , and with Theorem 3.7 a regret bound R N R_N R N for it gives
E ∑ e ( ( 1 − ω ) κ ξ L P ( k ) ( P ( e ) ) − V e A l g . 1 ) ≤ ( 1 − ω ) R N + 1. \mathbb{E}\sum_e\left((1-\omega)\kappa_{\xi}\mathrm{LP}^{(k)}(P^{(e)})-V_e^{\mathrm{Alg.\ 1}}\right)\leq(1-\omega)R_N+1. E e ∑ ( ( 1 − ω ) κ ξ LP ( k ) ( P ( e ) ) − V e Alg. 1 ) ≤ ( 1 − ω ) R N + 1.
Algorithm 1 is not modified to use probes or windows. With R N ( 1 ) = E ∑ e ( ( 1 − ω ) κ L P ( k ) ( P ( e ) ) − V e ) \mathcal{R}_N^{(1)}=\mathbb{E}\sum_e((1-\omega)\kappa\mathrm{LP}^{(k)}(P^{(e)})-V_e) R N ( 1 ) = E ∑ e (( 1 − ω ) κ LP ( k ) ( P ( e ) ) − V e ) , the exploration/window tradeoff underlying Corollary C.5 is
R N ( 1 ) ≤ ( 1 − ω ) κ inf 0 < ξ < 1 1 ≤ τ ≤ N { ξ N T + 2 ( g − 1 ) L Σ + 4 N C e s t ξ τ + 2 ( τ + g − 1 ) V } + 1. (53) \mathcal{R}_N^{(1)}\leq(1-\omega)\kappa\inf_{\substack{0<\xi<1\\1\leq\tau\leq N}}\left\{\xi NT+2(g-1)L_{\Sigma}+\frac{4NC_{\mathrm{est}}}{\sqrt{\xi\tau}}+2(\tau+g-1)\mathcal{V}\right\}+1. \tag*{(53)} R N ( 1 ) ≤ ( 1 − ω ) κ 0 < ξ < 1 1 ≤ τ ≤ N inf { ξ N T + 2 ( g − 1 ) L Σ + ξ τ 4 N C est + 2 ( τ + g − 1 ) V } + 1. (53)
where τ \tau τ is an integer. The coefficients remain explicit here; the corollary absorbs them into constants to display the rates.
Proof of Corollary C.5. The relaxation satisfies L P ( k ) ( P ) ≤ T \mathrm{LP}^{(k)}(P) \le T LP ( k ) ( P ) ≤ T , since expected reward is at most expected plays. Restoring the benchmark from κ ξ L P ( k ) \kappa_{\xi}\mathrm{LP}^{(k)} κ ξ LP ( k ) to κ L P ( k ) \kappa\mathrm{LP}^{(k)} κ LP ( k ) therefore costs at most κ ξ N T \kappa\xi NT κ ξ N T . Add this charge to (52), bound 1 − ξ 1-\xi 1 − ξ by one in its nonnegative upper bound, and transfer through Theorem 3.7. This proves (53). Every fixed ξ \xi ξ defines an admissible comparison learner, so the infimum changes only the analysis.
The term ξ N T \xi NT ξ N T charges for exploration, and N C e s t / ξ τ NC_{\mathrm{est}}/\sqrt{\xi\tau} N C est / ξ τ measures sampling error. Drift over the window costs τ V \tau\mathcal{V} τ V . The remaining terms account for report lag. Optimizing the window as above leaves, apart from lag terms,
O ( ξ N T + C e s t N / ξ + N 2 / 3 C e s t 2 / 3 ξ − 1 / 3 ) . O\left(\xi NT+C_{\mathrm{est}}\sqrt{N/\xi}+N^{2/3}C_{\mathrm{est}}^{2/3}\xi^{-1/3}\right). O ( ξ N T + C est N / ξ + N 2/3 C est 2/3 ξ − 1/3 ) .
Define the two exploration scales
ξ s t a t = ( C e s t 2 T 2 N ) 1 / 3 , ξ d r i f t = ( C e s t 2 V T 3 N ) 1 / 4 . \xi_{\mathrm{stat}}=\left(\frac{C_{\mathrm{est}}^2}{T^2N}\right)^{1/3},\qquad\xi_{\mathrm{drift}}=\left(\frac{C_{\mathrm{est}}^2\mathcal{V}}{T^3N}\right)^{1/4}. ξ stat = ( T 2 N C est 2 ) 1/3 , ξ drift = ( T 3 N C est 2 V ) 1/4 .
and take ξ = min { 1 / 2 , max { ξ s t a t , ξ d r i f t } } \xi=\min\{1/2,\max\{\xi_{\mathrm{stat}},\xi_{\mathrm{drift}}\}\} ξ = min { 1/2 , max { ξ stat , ξ drift }} . Without clipping, each sampling or drift term is at most ξ N T \xi NT ξ N T . The cost is therefore O ( ( T C e s t 2 ) 1 / 3 N 2 / 3 + ( T C e s t 2 ) 1 / 4 V 1 / 4 N 3 / 4 ) O((TC_{\mathrm{est}}^2)^{1/3}N^{2/3}+(TC_{\mathrm{est}}^2)^{1/4}\mathcal{V}^{1/4}N^{3/4}) O (( T C est 2 ) 1/3 N 2/3 + ( T C est 2 ) 1/4 V 1/4 N 3/4 ) . If clipping is active, this expression is at least N T / 2 NT/2 N T /2 , so the trivial regret bound suffices. Consequently, for a universal constant c c c ,
R N ( 1 ) ≤ ( 1 − ω ) κ c [ ( T C e s t 2 ) 1 / 3 N 2 / 3 + ( T C e s t 2 ) 1 / 4 V 1 / 4 N 3 / 4 + ( g − 1 ) ( L Σ + V ) ] + 1. (54) \mathcal{R}^{(1)}_N\le(1-\omega)\kappa c\left[(TC_{\mathrm{est}}^2)^{1/3}N^{2/3}+(TC_{\mathrm{est}}^2)^{1/4}\mathcal{V}^{1/4}N^{3/4}+(g-1)(L_\Sigma+\mathcal{V})\right]+1. \tag*{(54)} R N ( 1 ) ≤ ( 1 − ω ) κ c [ ( T C est 2 ) 1/3 N 2/3 + ( T C est 2 ) 1/4 V 1/4 N 3/4 + ( g − 1 ) ( L Σ + V ) ] + 1. (54)
To obtain the stated form, use 0 ≤ V ≤ ( N − 1 ) L Σ 0\le\mathcal{V}\le(N-1)L_\Sigma 0 ≤ V ≤ ( N − 1 ) L Σ and N ≥ 1 N\ge1 N ≥ 1 , so that ( g − 1 ) L Σ ≤ ( g − 1 ) L Σ N 2 / 3 (g-1)L_\Sigma\le(g-1)L_\Sigma N^{2/3} ( g − 1 ) L Σ ≤ ( g − 1 ) L Σ N 2/3 and ( g − 1 ) V ≤ ( g − 1 ) L Σ 3 / 4 V 1 / 4 N 3 / 4 (g-1)\mathcal{V}\le(g-1)L_\Sigma^{3/4}\mathcal{V}^{1/4}N^{3/4} ( g − 1 ) V ≤ ( g − 1 ) L Σ 3/4 V 1/4 N 3/4 . All coefficients in (54) are then absorbed into c i n s t c_{\mathrm{inst}} c inst , which depends only on the fixed horizon, arms, caps, depths and lag. For a feasible solution with objective F F F , the feasible-solution form of Theorem 3.7 loses at most L P N ( C , B ′ ) − F \mathrm{LP}_N(C,B')-F LP N ( C , B ′ ) − F relative to this comparison.
Affine transfer. Under Theorem C.4’s conditions and the completion certificate of Theorem 3.7, suppose B i ′ ≥ N L i B'_i\ge NL_i B i ′ ≥ N L i and ω < 1 \omega<1 ω < 1 . Comparing the LP learner with Algorithm 3 gives
R N ( 1 ) ≤ ( 1 − ω ) κ inf 0 < ξ < 1 { ξ N T + 4 L Σ + 2 10 C e s t N / ξ } + 1. \mathcal{R}^{(1)}_N\le(1-\omega)\kappa\inf_{0<\xi<1}\left\{\xi NT+4L_\Sigma+2\sqrt{10}\,C_{\mathrm{est}}\sqrt{N/\xi}\right\}+1. R N ( 1 ) ≤ ( 1 − ω ) κ 0 < ξ < 1 inf { ξ N T + 4 L Σ + 2 10 C est N / ξ } + 1.
Taking ξ = min { 1 / 2 , ( C e s t 2 / ( T 2 N ) ) 1 / 3 } \xi=\min\{1/2,(C_{\mathrm{est}}^2/(T^2N))^{1/3}\} ξ = min { 1/2 , ( C est 2 / ( T 2 N ) ) 1/3 } gives O ( ( 1 − ω ) κ [ ( T C e s t 2 ) 1 / 3 N 2 / 3 + L Σ ] ) + 1 O((1-\omega)\kappa[(TC_{\mathrm{est}}^2)^{1/3}N^{2/3}+L_\Sigma])+1 O (( 1 − ω ) κ [( T C est 2 ) 1/3 N 2/3 + L Σ ]) + 1 ; if clipping is active, the displayed rate is of order N T NT N T and the trivial bound suffices. The N 2 / 3 N^{2/3} N 2/3 term pays for exploration while restoring κ \kappa κ ; it is an upper bound for this comparison, not a lower bound.
Protected-report execution. Algorithm 2 needs no reserve. Its comparison runs through a planned version of the probing policy that learns only from the first r i r_i r i plays of each arm in each episode.
Corollary C.6 (Rates without a reserve). Assume the conditions of this appendix, with B i ′ ≥ N L i B'_i\ge NL_i B i ′ ≥ N L i . Take prediction depths r e , i = r i r_{e,i}=r_i r e , i = r i and a window with w + ( 1 + ϵ ) ∑ i r i ≤ T p < T w+(1+\epsilon)\sum_i r_i\le T_p<T w + ( 1 + ϵ ) ∑ i r i ≤ T p < T and X 2 = ( T − T p ) / L ‾ ≥ 1 X_2=(T-T_p)/\overline{L}\ge1 X 2 = ( T − T p ) / L ≥ 1 . Suppose Algorithm 2 uses an optimal solution. Put
κ p l = c X 2 , 0 ( 1 − T p T ) > 0. \kappa_{\mathrm{pl}}=c_{X_2,0}\left(1-\frac{T_p}{T}\right)>0. κ pl = c X 2 , 0 ( 1 − T T p ) > 0. Then, for every N N N ,
E ∑ e ( κ p l L P ( k ) ( P ( e ) ) − V e ) ≤ κ p l inf 0 < ξ < 1 1 ≤ τ ≤ N { ξ N T + 2 ( g − 1 ) L Σ + 4 N C e s t ξ τ + 2 ( τ + g − 1 ) V } , \mathbb{E}\sum_e\left(\kappa_{\mathrm{pl}}\mathrm{LP}^{(k)}(P^{(e)})-V_e\right)\le\kappa_{\mathrm{pl}}\inf_{\substack{0<\xi<1\\1\le\tau\le N}}\left\{\xi NT+2(g-1)L_\Sigma+\frac{4NC_{\mathrm{est}}}{\sqrt{\xi\tau}}+2(\tau+g-1)\mathcal{V}\right\}, E e ∑ ( κ pl LP ( k ) ( P ( e ) ) − V e ) ≤ κ pl 0 < ξ < 1 1 ≤ τ ≤ N inf { ξ N T + 2 ( g − 1 ) L Σ + ξ τ 4 N C est + 2 ( τ + g − 1 ) V } ,
with C e s t C_{\mathrm{est}} C est from (30) at J g r p = 1 J_{\mathrm{grp}}=1 J grp = 1 . This is (53) with ( 1 − ω ) κ (1-\omega)\kappa ( 1 − ω ) κ replaced by κ p l \kappa_{\mathrm{pl}} κ pl and without the additive 1. The rates of Corollary C.5, including the affine case, therefore hold for every N N N , with κ p l \kappa_{\mathrm{pl}} κ pl in place of ( 1 − ω ) κ (1-\omega)\kappa ( 1 − ω ) κ .
Proof. Use Theorem 4.5 with ζ = 0 \zeta=0 ζ = 0 . The window estimates use probe outcomes only, which are predicted plays of the planned comparison. With J g r p = 1 J_{\mathrm{grp}}=1 J grp = 1 , the calculation of (52) gives
2 E N ( ξ ) ≤ 2 ( g − 1 ) L Σ + 4 N C e s t ξ τ + 2 ( τ + g − 1 ) V . 2\mathbb{E}_N(\xi) \le2(g-1)L_\Sigma+ \frac{4NC_{\mathrm{est}}}{\sqrt{\xi\tau}} + 2(\tau+g-1)V. 2 E N ( ξ ) ≤ 2 ( g − 1 ) L Σ + ξ τ 4 N C est + 2 ( τ + g − 1 ) V .
Charge at most ξ N T \xi NT ξ N T for restoring the factor 1 − ξ 1-\xi 1 − ξ in (31) and take the infimum over ξ , τ \xi,\tau ξ , τ . The affine estimate gives the corresponding affine bound in the same way.
Proofs for the price of prior uncertainty
The information-price bounds use exact episode optimization to compare with V p r i o r V_{\mathrm{prior}} V prior . The posterior-sampling argument below also applies to a fixed approximate planner and executor. It compares with that reference policy’s value, using a quadratic information bound and a telescoping entropy budget. Whole-tape histogram dimensions control the coefficient.
Lemma D.1 (Sampling a reference episode policy). Assume the stationary conditions of Proposition 5.4. Let A ( P , U ) A(P,U) A ( P , U ) be a fixed map to deterministic feasible episode policies, with U U U independent of P P P , and put π A = A ( P , U 0 ) \pi^A=A(P,U_0) π A = A ( P , U 0 ) for an independent reference seed. Use g g g threads, each drawing P ~ \widetilde{P} P from its own posterior and executing A ( P ~ , U e ) A(\widetilde{P},U_e) A ( P , U e ) with fresh coins. With D t a p e = ∏ i ( L i + d i − 1 d i − 1 ) D_{\mathrm{tape}}=\prod_i\binom{L_i+d_i-1}{d_i-1} D tape = ∏ i ( d i − 1 L i + d i − 1 ) , its reward satisfies
N E V P ( π A ) − V P S ≤ min { T , L Σ } 1 2 g D t a p e N H ( π A ) . N\mathbb{E}V^P(\pi^A)-V^{\mathrm{PS}} \le\min\{T,L_\Sigma\}\sqrt{\frac{1}{2}gD_{\mathrm{tape}}NH(\pi^A)}. N E V P ( π A ) − V PS ≤ min { T , L Σ } 2 1 g D tape N H ( π A ) . Proof. Let b 0 = min { T , L Σ } b_0=\min\{T,L_\Sigma\} b 0 = min { T , L Σ } . Every episode reward lies in [ 0 , b 0 ] [0,b_0] [ 0 , b 0 ] . The complete reported trajectory determines that reward. Given P P P , the arms’ tape histograms are independent with laws q L i ( P i ) q_{L_i}(P_i) q L i ( P i ) , and their orders given those histograms are uniform. Thus every fixed episode policy has mean reward
V P ( π ) = ⟨ f π , q ( P ) ⟩ , q ( P ) = ⨂ i q L i ( P i ) ∈ R D t a p e , V^P(\pi)=\langle f_\pi,\mathbf{q}(P)\rangle,\qquad\mathbf{q}(P)=\bigotimes_i q_{L_i}(P_i)\in\mathbb{R}^{D_{\mathrm{tape}}}, V P ( π ) = ⟨ f π , q ( P )⟩ , q ( P ) = i ⨂ q L i ( P i ) ∈ R D tape , where f π f_\pi f π averages its reward over the uniform tape orders.
Fix a thread’s history h h h and write p α = Pr { π A = α ∣ h } p_\alpha=\Pr\{\pi^A=\alpha\mid h\} p α = Pr { π A = α ∣ h } for α ∈ Π \alpha\in\Pi α ∈ Π . The selected policy π e \pi_e π e has these probabilities and is conditionally independent of P P P and U 0 U_0 U 0 . Posterior matching gives
Δ h : = E [ V P ( π A ) − V P ( π e ) ∣ h ] = E [ V P ~ ( π e ) − V P ( π e ) ∣ h ] . \Delta_h:=\mathbb{E}[V^P(\pi^A)-V^P(\pi_e)\mid h]=\mathbb{E}[V^{\widetilde P}(\pi_e)-V^P(\pi_e)\mid h]. Δ h := E [ V P ( π A ) − V P ( π e ) ∣ h ] = E [ V P ( π e ) − V P ( π e ) ∣ h ] . On policies of positive conditional probability, form the matrix
M α β = p α p β ( E [ V P ( α ) ∣ π A = β , h ] − E [ V P ( α ) ∣ h ] ) . M_{\alpha\beta}=\sqrt{p_\alpha p_\beta}\left(\mathbb{E}[V^P(\alpha)\mid\pi^A=\beta,h]-\mathbb{E}[V^P(\alpha)\mid h]\right). M α β = p α p β ( E [ V P ( α ) ∣ π A = β , h ] − E [ V P ( α ) ∣ h ] ) . Then Δ h = tr M \Delta_h=\operatorname{tr}M Δ h = tr M . The linear representation gives rank M ≤ D t a p e \operatorname{rank}M\le D_{\mathrm{tape}} rank M ≤ D tape . If Y e Y_e Y e denotes the episode’s complete reports, Pinsker’s inequality and the reward range give
I ( π A ; π e , Y e ∣ h ) ≥ 2 b 0 2 ∥ M ∥ F 2 , Δ h 2 ≤ D t a p e ∥ M ∥ F 2 ≤ b 0 2 D t a p e 2 I ( π A ; π e , Y e ∣ h ) . I(\pi^A;\pi_e,Y_e\mid h)\ge\frac{2}{b_0^2}\lVert M\rVert_F^2,\qquad\Delta_h^2\le D_{\mathrm{tape}}\lVert M\rVert_F^2\le\frac{b_0^2D_{\mathrm{tape}}}{2}I(\pi^A;\pi_e,Y_e\mid h). I ( π A ; π e , Y e ∣ h ) ≥ b 0 2 2 ∥ M ∥ F 2 , Δ h 2 ≤ D tape ∥ M ∥ F 2 ≤ 2 b 0 2 D tape I ( π A ; π e , Y e ∣ h ) . This is the matrix argument of Russo and Van Roy [34 Proposition 5] , valid for an arbitrary reference target [35 Section 6] ; Δ h \Delta_h Δ h may be signed.
Split episodes by their residue modulo g g g . Each thread uses only its own earlier reports, all of which have arrived before its next episode. Conditional on P P P , fresh outcome tapes are independent across episodes. The information chain rule bounds its total expected information gain about π A \pi^A π A by H ( π A ) H(\pi^A) H ( π A ) . Cauchy–Schwarz therefore bounds a thread of n r n_r n r episodes by b 0 D t a p e n r H ( π A ) / 2 b_0\sqrt{D_{\mathrm{tape}}n_rH(\pi^A)/2} b 0 D tape n r H ( π A ) /2 . Summing and using ∑ r n r ≤ g N \sum_r\sqrt{n_r}\le\sqrt{gN} ∑ r n r ≤ g N proves the result.
Posterior sampling with episode LPs
Posterior sampling gives a second comparison without prescribed probing episodes. Sample the recurring priors, solve their episode LP, and execute it with the known-prior guarantee. Fix the planner, its tie-breaking and its executor as a map A ( P , U ) A(P,U) A ( P , U ) , where U U U supplies all policy coins. Fixing U U U gives a deterministic episode policy, feasible under every recurring prior. Its expected reward guarantee is required only under the supplied prior P P P .
Theorem D.2 (Posterior sampling with episode LPs). Assume the stationary conditions of Corollary 4.6. Let V P ( π ) V^{P}(\pi) V P ( π ) be the expected episode reward of policy π \pi π under P P P , and suppose
E U V P ( A ( P , U ) ) ≥ κ LP ( k ) ( P ) \mathbb{E}_{U}V^{P}(A(P,U)) \ge\kappa\operatorname{LP}^{(k)}(P) E U V P ( A ( P , U )) ≥ κ LP ( k ) ( P ) for every P P P .
Use g g g inference threads, each updating only from its own episodes. In each episode, draw P ~ \widetilde{P} P from the thread’s posterior and execute A ( P ~ , U e ) A(\widetilde{P},U_e) A ( P , U e ) with fresh coins. Put
D t a p e = ∏ i ( L i + d i − 1 d i − 1 ) , π A = A ( P , U 0 ) , D_{\mathrm{tape}} = \prod_i \binom{L_i+d_i-1}{d_i-1}, \qquad\pi^{A}=A(P,U_0), D tape = i ∏ ( d i − 1 L i + d i − 1 ) , π A = A ( P , U 0 ) , where U 0 U_0 U 0 is an independent reference seed, unused by the learner. Then this policy’s expected total reward V P S − L P V_{\mathrm{PS-LP}} V PS − LP satisfies
κ N E Q LP ( k ) ( P ) − V P S − L P ≤ min { T , L Σ } 1 2 g D t a p e N H ( π A ) . (55) \kappa N\mathbb{E}_{Q}\operatorname{LP}^{(k)}(P)-V_{\mathrm{PS-LP}} \le\min\{T,L_{\Sigma}\}\sqrt{\frac{1}{2}gD_{\mathrm{tape}}N H(\pi^{A})}. \tag*{(55)} κ N E Q LP ( k ) ( P ) − V PS − LP ≤ min { T , L Σ } 2 1 g D tape N H ( π A ) . (55) The entropy is under Q Q Q and U 0 U_0 U 0 , with natural logarithms. It is at most log ∣ Π ∣ \log|\Pi| log ∣Π∣ , where Π \Pi Π is the finite set of deterministic feasible episode policies.
Proof. The selected policy has the conditional distribution of π A \pi^{A} π A given the thread’s history. Apply Lemma D.1 to this fixed reference map, then use its known-prior guarantee. The lemma uses whole-tape histograms to represent every episode policy’s expected reward linearly; it does not require the reference policy to be optimal.
Under the completion conditions of Theorem 3.7 and nonbinding comparison budgets B i ′ ≥ N L i B'_i \ge NL_i B i ′ ≥ N L i , the posterior-sampled comparison gives
( 1 − ω ) κ N E Q LP ( k ) ( P ) − V A ≤ ( 1 − ω ) R ‾ P S + 1. (1-\omega)\kappa N\mathbb{E}_{Q}\operatorname{LP}^{(k)}(P)-V_A \le(1-\omega)\overline{R}_{\mathrm{PS}}+1. ( 1 − ω ) κ N E Q LP ( k ) ( P ) − V A ≤ ( 1 − ω ) R PS + 1.
where R ‾ P S \overline{R}_{\mathrm{PS}} R PS is the right side of (55) . Indeed, the learning LP at budgets B ′ B' B ′ upper-bounds this admissible comparison, and Theorem 3.7 transfers its value. This argument uses full completion.
Proof of Proposition 5.4. Apply Lemma D.1 with A ( P , U ) = π ∗ ( P ) A(P,U)=\pi^{*}(P) A ( P , U ) = π ∗ ( P ) . Nonbinding budgets give V p r i o r = N E Q V P ( π ∗ ( P ) ) V_{\mathrm{prior}}=N\mathbb{E}_{Q}V^{P}(\pi^{*}(P)) V prior = N E Q V P ( π ∗ ( P )) , and V l e a r n V_{\mathrm{learn}} V learn is at least the value of this admissible posterior-sampling policy. Finally H ( π ∗ ) ≤ log ∣ Π ∣ H(\pi^{*})\le\log|\Pi| H ( π ∗ ) ≤ log ∣Π∣ .
The information-price bounds also allow the specified safe alternative of § 3.6 . Fill otherwise unused steps with that action. Episode rewards then lie in [ a ( T − b 0 ) , a T + ( 1 − a ) b 0 ] [a(T-b_0),aT+(1-a)b_0] [ a ( T − b 0 ) , a T + ( 1 − a ) b 0 ] , still an interval of length b 0 b_0 b 0 , and remain linear in the whole-tape law. The same range and information arguments apply to these policies.
The exact reference planner bounds the information price; the LP planner and executor in Theorem D.2 give a scaled LP comparison by the same argument. Both coefficients include the product of whole-tape histogram dimensions. Even for one unknown arm, this dimension, the report lag and the reference-policy entropy remain. If Q Q Q has finite support, then H ( π ∗ ) ≤ log ∣ supp Q ∣ H(\pi^{*})\le\log|\operatorname{supp} Q| H ( π ∗ ) ≤ log ∣ supp Q ∣ .
AI Use Statement
We used multiple generative AI tools for improving writing and searching related work. We have reviewed all AI-assisted work.