# The numeraire e-variable and reverse information projection

arXiv: https://arxiv.org/abs/2402.18810 · HTML: https://mlsanity.com/paper/2402.18810/

## Key facts

- Cited by 12 papers in the map (data through 14 Oct 2026).
- First posted on arXiv 29 Feb 2024.

## Built on by (most cited first, 12 of 12)

- [Active, anytime-valid risk controlling prediction sets](https://mlsanity.com/paper/2406.10490/) (2024)
- [Backward Conformal Prediction](https://mlsanity.com/paper/2505.13732/) (2025)
- [On Stopping Times of Power-one Sequential Tests: Tight Lower and Upper Bounds](https://mlsanity.com/paper/2504.19952/) (2025)
- [Post-detection inference for sequential changepoint localization](https://mlsanity.com/paper/2502.06096/) (2025)
- [Efficient Sequential Evaluation of Large Language Models](https://mlsanity.com/paper/2607.17409/) (2026)
- [Fuzzy Prediction Sets: Conformal Prediction with E-values](https://mlsanity.com/paper/2509.13130/) (2025)
- [Power one sequential tests exist for weakly compact mathscr P against mathscr Pc](https://mlsanity.com/paper/2604.03218/) (2026)
- [Multi-Armed Sequential Hypothesis Testing by Betting](https://mlsanity.com/paper/2603.17925/) (2026)
- [The Type-II Error of Test Supermartingales: e-Power versus the Chernoff-Stein Exponent](https://mlsanity.com/paper/2609.27765/) (2026)
- [Optimal sequential tests yield log-optimal e-processes](https://mlsanity.com/paper/2605.12720/) (2026)
- [Always-On Experimentation](https://mlsanity.com/paper/2609.38695/) (2026)
- [textGrain watermarking is harder to detect: can sum-tests still do it optimally?](https://mlsanity.com/paper/2610.17655/) (2026)

## Abstract (arXiv)

We consider testing a composite null hypothesis P against a point alternative mathsfQ using e-variables, which are nonnegative random variables X such that mathbbE_mathsfP[X] ≤ 1 for every mathsfP ∈ P. This paper establishes a fundamental result: under no conditions whatsoever on P or mathsfQ, there exists a special e-variable X^* that we call the numeraire, which is strictly positive and satisfies mathbbE_mathsfQ[X/X^*] ≤ 1 for every other e-variable X. In particular, X^* is log-optimal in the sense that mathbbE_mathsfQ[log(X/X^*)] ≤ 0. Moreover, X^* identifies a particular sub-probability measure mathsfP^* via the density d mathsfP^*/d mathsfQ = 1/X^*. As a result, X^* can be seen as a generalized likelihood ratio of mathsfQ against P. We show that mathsfP^* coincides with the reverse information projection (RIPr) when additional assumptions are made that are required for the latter to exist. Thus mathsfP^* is a natural definition of the RIPr in the absence of any assumptions on P or mathsfQ. In addition to the abstract theory, we provide several tools for finding the numeraire and RIPr in concrete cases. We discuss several nonparametric examples where we can indeed identify the numeraire and RIPr, despite not having a reference measure. Our results have interpretations outside of testing in that they yield the optimal Kelly bet against P if we believe reality follows mathsfQ. We end with a more general optimality theory that goes beyond the ubiquitous logarithmic utility. We focus on certain power utilities, leading to reverse R'enyi projections in place of the RIPr, which also always exist.