Work

When early information pays: sharp thresholds, optimal resource allocation and certified decisions

Year: 2026
Type: preprint
Abstract: This work studies how early information, observation order and resource constraints affect finite sequential decision problems. Calibration and deployment measurements share an expected-cost budget, and several actions may be correct under the same model. The analysis establishes exact excess-risk exponents for the specified acquisition protocols. In a compatible rare-context experiment, a cost-dependent likelihood-ratio bit preserves the preview exponent, and two logarithmic sampling thresholds characterize the transition from negligible benefit to the full perfect-preview advantage. The work also identifies the optimal symmetric cost criterion for minimizing the worst relative penalty of precommitting to calibration: its spectral weights are geometric, with the golden ratio arising in the two-profile case. A separate finite-sample audit compares a frozen policy with the optimum at the same actual cost, using convex tuning and a dual certificate supported on at most two comparator trees. This record archives the same scientific work submitted to the Journal of Statistical Planning and Inference as Designing early information: sharp resource laws and certified sequential decisions, under an alternative repository title. It is an author manuscript; journal peer review has not been completed. Supporting proofs, LaTeX sources and reproducibility materials accompany the manuscript. Version V13.1. The files are restricted; no automatic public-release date is set. The title and record metadata are public.
Language: English
Open Access status: green
OpenAlex ID: W7233079182
Cites:
Cited by:
Related to: 10