Randomized Adaptive Sequential Allocation and the Limits of Bounded Exposure to Inferior Treatments
Abstract
A clinical trial must learn which treatment is superior while sparing its participants from the inferior one. Kundu et al. 2025 recently showed that a purely deterministic rule, which always assigns the treatment with the larger current sample mean, achieves a strong form of this ethical goal. The smaller allocation count stays almost surely finite and has all moments bounded uniformly in the sample size. Deterministic allocation is, however, predictable, and predictability invites selection bias. We ask whether this bounded-exposure phenomenon survives once the deterministic rule is replaced by a smooth randomized one. We show that it does not. Under any fixed non-degenerate symmetric randomization, both treatments are sampled infinitely often, the adaptive sample means are strongly consistent, and the inferior treatment draws a strictly positive limiting allocation proportion. Consequently, the inferior allocation count grows linearly and its moments are unbounded. We then study annealed randomization, in which the randomization scale cools with the accumulated information, and obtain summability and divergence criteria that separate finite, sublinear, and linear exposure. A forced-exploration variant identifies the superior treatment with probability one while keeping the inferior exposure sublinear, and thereby frees the annealed guarantees from any correct-stabilization event. We also treat normalized multi-treatment rules, ties, indifference zones, unequal variances, general response distributions, and local alternatives. Finally, we develop the post-trial asymptotic inference that each exposure regime permits. The results delineate the trade-off between allocation concealment, persistent exploration, and ethical concentration on the superior treatment.
Related
- Cost-Aware Iteration Allocation for Sequential Multi-Fidelity First-Order Optimization
- Improved Analysis of Adaptive Task Allocation and Non-Stationary Processing Times
- Control Allocation with Adaptive Augmentation for Aerodynamic Optimization of Trailing Edge Morphing Aircraft
- Mission Efficiency Optimization in Low-Altitude Economy: Adaptive Power Allocation for Coordinating Heterogeneous Aircraft Swarms
- Co-Adaptive Entropy Modeling and Rate-Distortion-Optimal Bit Allocation for Perceptual Audio Coding
BibTeX
@misc{kundu2026randomized,
title = {{Randomized Adaptive Sequential Allocation and the Limits of Bounded Exposure to Inferior Treatments}},
author = {Sampurna Kundu and Monitirtha Dey and Atanu Biswas},
howpublished = {arXiv},
year = {2026},
month = nov,
doi = {10.48550/arxiv.2611.02012},
eprint = {2611.02012},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2611.02012},
}
RIS
TY - UNPB
TI - Randomized Adaptive Sequential Allocation and the Limits of Bounded Exposure to Inferior Treatments
AU - Sampurna Kundu
AU - Monitirtha Dey
AU - Atanu Biswas
PY - 2026
DA - 2026/11/01
PB - arXiv
DO - 10.48550/arxiv.2611.02012
UR - https://arxiv.org/abs/2611.02012
AB - A clinical trial must learn which treatment is superior while sparing its participants from the inferior one. Kundu et al. 2025 recently showed that a purely deterministic rule, which always assigns the treatment with the larger current sample mean, achieves a strong form of this ethical goal. The smaller allocation count stays almost surely finite and has all moments bounded uniformly in the sample size. Deterministic allocation is, however, predictable, and predictability invites selection bias. We ask whether this bounded-exposure phenomenon survives once the deterministic rule is replaced by a smooth randomized one. We show that it does not. Under any fixed non-degenerate symmetric randomization, both treatments are sampled infinitely often, the adaptive sample means are strongly consistent, and the inferior treatment draws a strictly positive limiting allocation proportion. Consequently, the inferior allocation count grows linearly and its moments are unbounded. We then study annealed randomization, in which the randomization scale cools with the accumulated information, and obtain summability and divergence criteria that separate finite, sublinear, and linear exposure. A forced-exploration variant identifies the superior treatment with probability one while keeping the inferior exposure sublinear, and thereby frees the annealed guarantees from any correct-stabilization event. We also treat normalized multi-treatment rules, ties, indifference zones, unequal variances, general response distributions, and local alternatives. Finally, we develop the post-trial asymptotic inference that each exposure regime permits. The results delineate the trade-off between allocation concealment, persistent exploration, and ethical concentration on the superior treatment.
ER -