Chair: Matteo Gianella

Speakers:

• Cecilia Secchi (Università commerciale Luigi Bocconi)
  Spectral gap of Metropolis-within-Gibbs under log-concavity

• Nicola Branchini (University of Warwick)
  Understanding the behaviour of self-normalized importance sampling in the infinite variance setting

• Laura Battaglia (University of Oxford)
  Variational predictive resampling

• Simone Panzeri (Politecnico di Milano)
  Physics-informed statistical learning on non-standard spatial domains with fdaPDE

Chair: Alice Giampino

Speakers:

• Jaromir Sant (Universidad Carlos III de Madrid)
  Gamma duality and a tractable transition density for the Wright-Fisher diffusion with selection

• Marco dalla Pria (Università di Torino)
  Tracking diversity in time

• Francesco Furlan (Università di Torino)
  Scalable computation of Fleming–Viot filtering and smoothing weights with Gibbs sampling

• Ylenia Francesca Buttigliero (Università di Torino)
  Bayesian bootstrap beyond observation times

Chair: Rocco Mazza

Speakers:

• Erika Banzato (Università di Padova)
  Sex-specific patterns in multimorbidity networks: a differential network approach

• Erika Dicorato (Università degli Studi di Bari Aldo Moro)
  Foreign residents effect on population ageing assessment over space and time

• Amin Gino Fabbrucci Barbagli (Università degli Studi di Trieste)
  Modeling networks of tripartite hyperevents using the relational hyperevent model

• Erika Grammatica (Università degli Studi di Milano-Bicocca)
  Mapping the progress of gender equality in the EU: a multidimensional longitudinal approach through dynamic clustering

Speaker: Francesca Romana Crucinio (Università di Torino & Collegio Carlo Alberto)
Chair: Alice Giampino

Title:  Particle methods for empirical Bayes

Abstract: Latent variable models are widely used to describe complex data, but learning their parameters is challenging because it requires integrating over unobserved variables. Empirical Bayes offers a practical solution by estimating model parameters through marginal likelihood maximization, typically using the expectation–maximization (EM) algorithm. However, EM relies on alternating updates and can struggle when posterior distributions are difficult to compute. In this talk, we present a different perspective that treats parameter estimation and posterior inference as a single optimization problem. This viewpoint leads naturally to a class of algorithms where we update both the parameters and our approximation of the latent variable distribution simultaneously. To make this approach practical, we use particle methods to represent the evolving posterior. In particular, we explore two complementary strategies. The first relies on gradient-based sampling using Langevin dynamics, leading to simple algorithms such as the Unadjusted Langevin Algorithm (ULA), which efficiently approximates the posterior when gradients are available. The second uses importance sampling based on Fisher–Rao geometry, which is more flexible and can handle settings where gradients are unavailable or the latent variables are discrete. This combination results in a flexible framework that works across a wide range of models, including those with discrete or non-differentiable latent variables where standard gradient-based methods are not applicable. We illustrate the approach on several examples, showing improved convergence over EM and competitive performance with existing methods. Overall, the proposed framework provides an intuitive and versatile way to combine optimization and sampling for empirical Bayes inference.

Chair: Martina Amongero

Speakers:

• Giulia Bertagnolli (Libera Università di Bolzano)
  Statistical data depths for network centrality and beyond

• Sara Geremia (Università degli Studi di Trieste)
  Community-level core-periphery structures in co-authorship networks

• Noemi Corsini (University of Cambridge)
  A Bayesian latent space approach for modeling social influence on binary outcomes

• Francesco Gaffi (Università degli Studi di Bergamo)
  Exchangeable random permutations with an application to Bayesian graph matching

Chair: Giulia Capitoli

Speakers:

• Daniele Giardiello (Università degli Studi di Milano-Bicocca)
  Dynamic prediction methods and prediction performance assessment using landmarking with an application in traumatic brain injury

• Niccolò Cao (Alma Mater Studiorum - Università di Bologna)
  Joint modelling mixed-type multivariate longitudinal data and a time-to-event outcome: an application to ROI dataset

• Alessandra Ragni (Politecnico di Milano)
  Evaluating treatment effects for recurrent events with terminal events: estimating the patient-weighted while-alive estimand

• Salvatore Battaglia (Università degli Studi di Palermo)
  A changepoint-based Cox model for interpretable time-varying covariate effects