Sharp Class-Count Bounds for Calibration Reuse in Temperature-Scaled Conformal Prediction
How much coverage can a conformal classifier lose when it uses the same observations to select a temperature and calibrate its prediction sets? For nonrandomized least-ambiguous classification with a fixed K-class softmax model, the temperature-and-threshold family has growth function at most m + 1 + (K − 2)m(m − 1) and VC dimension asymptotic to log₂ K. One fixed classifier can realize K − 1 independent score coordinates at prescribed temperatures. This gives an exact beta order-statistic law for the coverage of an adversarial calibration-only selector. At interior finite-rank coverage levels, the worst-selector marginal deficit has order min{1, √(log K / n)}, where n is the calibration sample size. The lower construction allows bounded logits and temperatures in any nonempty open interval of positive values. A three-class example compares adversarial selection with likelihood tuning. The results characterize worst-case calibration reuse for unmodified LAC sets.
Authors
- S. M. Abdulla Al Mamun
Institutions
- University of Manchester (GB)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-19
- DOI
- https://doi.org/10.5281/zenodo.24374512
- Primary Topic
- Machine Learning and Algorithms
- Type
- preprint