AI Paper Atlas

Research map / Optimization / Nonconvex and smooth optimization

Only tails matter: Average-Case Universality and Robustness in the Convex Regime

Leonardo Cunha, Gauthier Gidel, Fabian Pedregosa, Damien Scieur, Courtney Paquette

ICML 2022 · Optimization › Nonconvex and smooth optimization · approach: Stochastic gradient last-iterate analysis · 1 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
Use GCM with high beta (beta=5/2 or 3/2) instead of gradient descentGeneralized Chebyshev Method · Marchenko-Pastur-type ESD, asymptotic function-value rateconvergence exponent: t^{-3} vs t^{-3/2} at xi=-1/2Improved
Replace gradient descent with Nesterov's methodfirst-order methods on random quadratics · Marchenko-Pastur-type ESD, asymptotic function-value rateconvergence exponent: t^{-3}log t vs t^{-3/2}Improved
Overestimate beta / underestimate alpha in GCM (robustness to parameter mis-tuning)Generalized Chebyshev Method · Beta-distributed ESD, asymptotic ratefunction-value rateNo clear effect
Use Nesterov's method on problems with concentration near 0 (low xi)Nesterov method · Beta-distributed ESD, function-value rategap to optimal rateNo clear effect
Overestimating L in GCMGeneralized Chebyshev Method · Beta-distributed ESDrateNo clear effect
Use beta below optimal beta* (beta<beta*) or low xi in GCM (ablation)Generalized Chebyshev Method · Synthetic Beta / Marchenko-Pastur, finite dimensionvariance of function value across initializationsHurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy