<html aria-label="message body">
<head>
<meta http-equiv="Content-Type" content="text/html; charset=utf-8">
</head>
<body style="overflow-wrap: break-word; -webkit-nbsp-mode: space; line-break: after-white-space;">
<div>We are glad to announce the following joint talk in the ETH/UZH ZüKoSt Seminar on Applied Statistics and ETH Research Seminar on Statistics: </div>
<div><br>
</div>
<div>"<b>On the feature space decomposition of several estimators</b>" </div>
<div><br>
</div>
<div>by <b>Guillaume Lecué</b>, ESSEC Business School, Cergy, France</div>
<div><br>
</div>
<div><b>Date and Time</b>: Thursday, 28 May 2026 at 16.15 h</div>
<div><b>Place</b>: ETH Zurich, HG G 19.1</div>
<div><br>
</div>
<div><b>Abstract</b>: In this talk, I will present a methodology called the feature space decomposition and apply it to the statistical analysis of several estimators such as min ell2 norm interpolant estimators, ridge estimators, spectral methods, minimum
ellq-norm interpolant estimators and max-margin classifiers. We establish high-probability non-asymptotic upper bounds for the excess risk of these estimators. In the case of min ell2 norm interpolant estimators, ridge estimators and spectral methods, we show
that these bounds are sharp. In particular, we obtain necessary and sufficient conditions for benign overfitting behavior of the min ell2 norm interpolant estimators. Our results rely on a method called feature space decomposition where the self-regularization
properties of minimum norm interpolant estimator is highlighted. Technically, we circumvent tools such as the convex min-max theorem, instead employing tools from Geometric Aspects of Functional Analysis, in particular, the Dvoretzky-Milman theorem plays a
central role in our analysis. This provides a geometric perspective on benign overfitting, and crucially, our techniques remain valid beyond the Gaussian case. Consequently, we obtain benign overfitting results with high probability when the design vector
is not necessarily Gaussian. We particularly emphasize that the feature space decomposition method may potentially refine the uniform convergence approach, suggesting its promise as a new methodology in mathematical statistics. Based on joint works with Radoslaw
Adamczak, George Gavrilopoulos, Zhifan Li, Zong Shang and Marta Strzelecka.</div>
<div><br>
</div>
<div>Seminar website: https://math.ethz.ch/sfs/news-and-events/seminar-overview.html</div>
<div><br>
</div>
<div>Organisers: F. Balabdaoui, A. Bandeira, P. Bühlmann, Y. Chen, R. Furrer, L. Held, T. Hothorn, M. Kalisch, J. B. H. Koh, L. Meier, J. Peters, M. Robinson, F. Sigrist, C. Strobl, J. Ziegel</div>
<div><br>
</div>
<div><br>
</div>
</body>
</html>