ICM 2026 Invited Section Lecture - Judith Rousseau
Keywords
Summary
191 words
Critical Evaluation
Value of the Information & Strength of the Argument
The value of the information is substantial: it presents novel theoretical guarantees for Bayesian procedures in a high-dimensional setting that was previously only studied for fixed dimension. The argumentation is rigorous, built on probabilistic inequalities and careful control of posterior concentrations. The speaker clearly explains the assumptions and their role, and the proof outlines are given for the main results. The distinction between estimating intensities and parameters is well-motivated, and the discussion on prior elicitation provides practical guidance. The argumentation would benefit from more explicit comparisons with existing frequentist methods and a more detailed explanation of the technical conditions, but overall it is compelling and technically sound.
Scientific Rigor, Source Quality, Title Accuracy
The lecture is based on a specific research paper (implied) and refers to prior work by the author and others. The sources are not explicitly listed with URLs, but the speaker mentions relevant literature (e.g., Hansen et al. for frequentist lasso methods). The title is perfectly aligned with the content. The presentation is rigorous, with clear definitions and theorem statements. The lack of a formal reference list is a limitation, but the talk is intended as an expert-level lecture, not a literature review. The adequacy between title and content is excellent.
213 words
Title / Content Match
Title accurately describes the content: Bayesian estimation in linear high-dimensional Hawkes processes.
Quality & Reliability
8/10
Lecture by a leading statistician presenting original peer-reviewed research; rigorous mathematical arguments, explicit model assumptions, and clear statements of results.
Key Moments
Markers derived by PSI from the transcript: the creator did not define chapters.
- Introduction and motivation: recovering causal interactions among neurons from point process data.
- Definition of point processes and linear Hawkes processes with intensity functions.
- High-dimensional setting: estimating K^2 interaction functions and K background rates as K grows.
- Assumptions: sparsity, stationarity, and spectral radius conditions.
- Prior construction: hierarchical selection prior for active sets and functions.
- Main results: posterior contraction rates for intensity and parameter metrics; oracle rates and exponential penalties.
- Discussion of spike-and-slab vs selection priors and why parameter estimation is harder.
Cited Sources
- Estimation in Linear High Dimensional Hawkes Processes: A Bayesian Approach — Main paper referenced in the description, presenting the work discussed in the lecture.
- Hansen et al. (frequentist lasso for Hawkes processes) — Mentioned as foundational work on multivariate Hawkes processes for finite dimension.
Concurring Sources
- Hansen et al. (frequentist lasso for Hawkes processes) — Provides the frequentist counterpart and motivates the Bayesian approach.
Contribution & Novelties
The lecture provides original theoretical contributions to Bayesian inference for high-dimensional Hawkes processes, a setting that had not been previously addressed. The key novelty is the derivation of posterior contraction rates under sparsity, allowing the dimension K to grow nearly exponentially with time. The hierarchical selection prior is shown to be superior to standard spike-and-slab priors in achieving good rates. The separation between intensity estimation and parameter estimation highlights a fundamental challenge and explains when exponential penalties arise.
Pour aller plus loin :
- Hawkes process — Definition and properties, relevant to the process studied.
- Bayesian inference — General framework for prior-to-posterior updating.
- Bernstein–von Mises theorem — Asymptotic normality of posterior, related to contraction rates.
114 words
Radar Profile
The profile shows high technical level and strong information quality, with slightly lower scores on quantity of information and overall reliability due to the niche nature and lack of explicit citations. The lecture is dense but well-structured.