Séminaire Jeunes chercheurs (Dana Naderi, jeudi 21 mai 2026)

La prochaine séance du séminaire Jeunes chercheurs du CEREMADE aura lieu le jeudi 21 mai 2026 à 17h en salle A707. Nous aurons le plaisir d'écouter Dana Naderi (CEREAMDE), qui nous parlera de

Approximating evidence via bounded harmonic means.

Abstract
Efficient Bayesian model selection relies on the model evidence (marginal likelihood), whose computation involves an intractable integral. The harmonic mean estimator (HME) has long been used as a simple approach to approximate the evidence, but the original version introduced by Newton and Raftery (1994) may suffer from infinite variance. To address this issue, Gelfand and Dey (1994) proposed a generalized framework based on instrumental distributions, and Robert and Wraith (2009) further improved this idea using higher posterior density (HPD) regions as instrumental functions. Building on this line of work, we propose a practical estimator based on an elliptical covering of the HPD region using non-overlapping ellipsoids. The resulting method, called the Elliptical Covering Marginal Likelihood Estimator (ECMLE), avoids the infinite-variance problem of the classical HME, allows exact volume computation, and extends naturally to multimodal settings. Through several examples, we show that ECMLE outperforms recent approaches such as THAMES and its improved variants (Metodiev et al., 2024, 2025). In addition, ECMLE provides lower variance and more stable evidence estimates, even in challenging scenarios.