Publications 

Preconditioned norms: A unified framework for steepest descent, quasi-Newton, and adaptive methods
Veprikov, Bolatov, Horváth, Beznosikov, Takáč, Hanzely; arXiv, 2025

Polyak stepsize: estimating optimal functional values without parameters or prior knowledge
Abdukhakimov, Pham, Horváth, Gasnikov, Takáč, Hanzely; arXiv, 2025

Loss-transformation invariance in the damped Newton method
Shestakov, Bohara, Horváth, Takáč, Hanzely; arXiv, 2025

Simple stepsize for quasi-Newton methods with global convergence guarantees
Agafonov, Ryspayev, Horváth, Gasnikov, Takáč, Hanzely; arXiv, 2025

Sketch-and-project meets Newton method: global O(k^−2) convergence with low-rank updates
Hanzely; AISTATS 2025

Newton method revisited: global convergence rates up to O(k^−3) for stepsize schedules and linesearch procedures
Hanzely, Abdukhakimov, Takáč; arXiv, 2024

ψDAG: projected stochastic approximation iteration for DAG structure learning
Ziu, Hanzely, Li, Zhang, Takáč, Kamzolov; arXiv, 2024

Adaptive optimization algorithms for machine learning (dissertation thesis)
Hanzely; KAUST, 2023

Convergence of first-order algorithms for meta-learning
Mishchenko, Hanzely, Richtárik; FL Workshop, ICML 2023

Distributed Newton-type methods with communication compression
Islamov, Qian, Hanzely, Safaryan, Richtárik; TMLR, 2023

A damped Newton method achieves a global O(k^−2) convergence rate
Hanzely, Kamzolov, Pasechnyuk, Gasnikov, Richtárik, Takáč; NeurIPS 2022

ZeroSARAH: nonconvex optimization with zero full gradient computation
Li, Hanzely, Richtárik; arXiv, 2021

Lower bounds and optimal algorithms for personalized federated learning
Hanzely F., Hanzely S., Horváth, Richtárik; NeurIPS 2020

Adaptive learning of the optimal mini-batch size of SGD
Alfarra, Hanzely, Albasyoni, Ghanem, Richtárik; OPT-ML, NeurIPS 2020