Beyond the Kullback–Leibler divergence: the likelihood-ratio distribution as a detectability measure
First author Lead authorPreprint
arXiv preprint · Submitted to arXiv 12 Oct 2026
- In brief
- Shows that the KL divergence, the standard forecast of whether a model is detectable, does not by itself give the probability of detecting it.
- Details
- For multivariate Gaussian data, derives the distribution of the log-likelihood ratio and its cumulants under both hypotheses, and examines discovery potential and parameter-estimation accuracy when the null is a point and the alternative a continuous family. Case studies: a toy Gaussian model, the cubic three-torus, the tensor-to-scalar ratio from B-mode polarization, and dark energy from supernova distances. Detectability also depends on the size of the parameter space, the likelihood-ratio distributions, and the distance from parameter-space boundaries (the look-elsewhere effect), so the KL–detectability relation has to be calibrated for each model family and test statistic.
- Why it matters
- Lets detectability forecasts be calibrated for the model family and test statistic actually used, rather than read off the KL divergence alone.
- My contribution
- Conceived the project and carried out the full analysis, including the analytical derivations, numerical implementation, and interpretation of the results.
- Keywords
- KL divergence · likelihood-ratio tests · model-selection forecasting · statistical detectability · look-elsewhere effect
BibTeX
@article{Samandar:2026beyondKL,
author = {Samandar, Amirhossein and {Martin Barandiaran}, Mikel and Lu, Linn Htat and Jaffe, Andrew H. and Mihaylov, Deyan P. and Starkman, Glenn D. and Trendafilova, C. and Akrami, Yashar and Anselmi, Stefano and {Carr{\'o}n Duque}, Javier and Copi, Craig J. and Cornet-Gomez, Fernando and Kosowsky, Arthur and Negro, Anna and Rodrigues, R. G. and Tamosiunas, Andrius},
title = {Beyond the {Kullback--Leibler} divergence: the likelihood-ratio distribution as a detectability measure},
year = {2026},
eprint = {2610.17186},
archivePrefix = {arXiv},
primaryClass = {astro-ph.CO}
}