no. 8 · Wednesday 4 November 2026
The Exact Ville Identity with Applications to E-Values
Abstract
Let $M_n$ be a nonnegative supermartingale with $M_0=1$, and let $T_b$ be the first time $M_n$ crosses $b>1$. Ville's inequality gives $P(T_b<\infty)\leq 1/b$. We show that, at every finite time horizon $N$, $$ 1=bP(T_b\leq N)+V_{b,N}+D_{b,N}+R_{b,N}, $$ where $V_{b,N}$ is the expected overshoot at crossing, $D_{b,N}$ is the cumulative predictable decrease before crossing, and $R_{b,N}$ is the expected value of paths that up to the time horizon $N$ remained below the boundary $b$. The identity follows from telescoping and can be made horizon free without additional integrability assumption. It gives a formula for the slack in Ville's inequality and thus the conservatism of the threshold test $M_n\geq b$. For likelihood ratio martingales, the same terms also describe power: $R_{b,N}$ is the probability, under the alternative, of not crossing by $N$, whereas $bP(T_b\leq N)+V_{b,N}$ is the power. The identity also allows us to verify horizon-free guarantees from a finite computation: $P(T_b<\infty)$ lies between $P(T_b\leq N)$ and $P(T_b\leq N)+R_{b,N}/b$, so a proposed threshold test $M_n\geq b_α$ can be shown to be of level $α\in(0,1)$ for every time horizon based on computations at the single horizon $N$. We illustrate how Ville's identity yields sharper thresholds for Bernoulli and exponential likelihood-ratio tests, and for a conformal test of exchangeability, demonstrating Monte Carlo calibration under a composite, nonparametric null. Lastly, we also establish a continuous-time analogue of Ville's identity.
- fields
- Probability