Adaptive Endogenous Detection in Gaussian Experiments
Higher criticism adapts to unknown sparsity and signal strength in classical Gaussian detection, where the alternative law is independent of the test. Here a level-$\alpha$ test is announced before independent, privately informed agents make hidden decisions to produce costly evidence. Production leaves the Gaussian null unchanged but determines the alternative law at equilibrium. This paper constructs a single observation-only test that uses no sparsity, signal-strength, or private-cost parameters. Whenever the signal is detectable under full production and the private-cost exponent is strictly above the sharp information-incentive frontier, its power tends to one at every independent Bayesian Nash equilibrium. Strictly below the frontier, every level-$\alpha$ test admits an independent equilibrium with power at most $\alpha+o(1)$. The construction combines overlapping threshold counts on publicly labeled nested panels with maximum tests on separate blocks. An exact many-threshold pivotality identity and uniform bounds for conditional Bernoulli sums establish the all-equilibrium guarantee. Equilibrium potential budgets, Hellinger information, pivotality energy, and truncated Gaussian scores yield the unrestricted converse. Exact binomial count calibration preserves the sharp strict frontier and gives a conservative global $p$-value. The study also determines the complete strict equilibrium frontier among permutation-invariant Gaussian tests. The same untrimmed, one-sided ordinary higher-criticism test attains this frontier at every independent equilibrium throughout the classically detectable region, while publicly labeled tests achieve a strictly better frontier in precisely characterized cost and signal regimes.
Authors
- K. C. Lalropuia (ORCID: https://orcid.org/0009-0008-4861-602X)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-21
- DOI
- https://doi.org/10.5281/zenodo.24417395
- Primary Topic
- Statistical Methods and Inference
- Type
- preprint