On the Lower Bound of Minimizing Polyak-Łojasiewicz functions

Pengyun Yue, Cong Fang, Zhouchen Lin
Proceedings of Thirty Sixth Conference on Learning Theory, PMLR 195:2948-2968, 2023.

Abstract

Polyak-Łojasiewicz (PL) (Polyak, 1963) condition is a weaker condition than the strong convexity but suffices to ensure a global convergence for the Gradient Descent algorithm. In this paper, we study the lower bound of algorithms using first-order oracles to find an approximate optimal solution. We show that any first-order algorithm requires at least Ω(Lμlog1ϵ) gradient costs to με find an ϵ-approximate optimal solution for a general L-smooth function that has an μ-PL constant. This result demonstrates the optimality of the Gradient Descent algorithm to minimize smooth PL functions in the sense that there exists a “hard” PL function such that no first-order algorithm can be faster than Gradient Descent when ignoring a numerical constant. In contrast, it is well-known that the momentum technique, e.g. Nesterov (2003, chap. 2), can provably accelerate Gradient Descent to O(√Lˆμlog1ϵ) gradient costs for functions that are L-smooth and ˆμ-strongly convex. Therefore, our result distinguishes the hardness of minimizing a smooth PL function and a smooth strongly convex function as the complexity of the former cannot be improved by any polynomial order in general.

Cite this Paper

BibTeX
@InProceedings{pmlr-v195-yue23a, title = {On the Lower Bound of Minimizing Polyak-Łojasiewicz functions}, author = {Yue, Pengyun and Fang, Cong and Lin, Zhouchen}, booktitle = {Proceedings of Thirty Sixth Conference on Learning Theory}, pages = {2948--2968}, year = {2023}, editor = {Neu, Gergely and Rosasco, Lorenzo}, volume = {195}, series = {Proceedings of Machine Learning Research}, month = {12--15 Jul}, publisher = {PMLR}, pdf = {https://proceedings.mlr.press/v195/yue23a/yue23a.pdf}, url = {https://proceedings.mlr.press/v195/yue23a.html}, abstract = {Polyak-Łojasiewicz (PL) (Polyak, 1963) condition is a weaker condition than the strong convexity but suffices to ensure a global convergence for the Gradient Descent algorithm. In this paper, we study the lower bound of algorithms using first-order oracles to find an approximate optimal solution. We show that any first-order algorithm requires at least ${\Omega}\left(\frac{L}{\mu}\log\frac{1}{\epsilon}\right)$ gradient costs to με find an $\epsilon$-approximate optimal solution for a general $L$-smooth function that has an $\mu$-PL constant. This result demonstrates the optimality of the Gradient Descent algorithm to minimize smooth PL functions in the sense that there exists a “hard” PL function such that no first-order algorithm can be faster than Gradient Descent when ignoring a numerical constant. In contrast, it is well-known that the momentum technique, e.g. Nesterov (2003, chap. 2), can provably accelerate Gradient Descent to ${O}\left(\sqrt{\frac{L}{\hat{\mu}}}\log\frac{1}{\epsilon}\right)$ gradient costs for functions that are $L$-smooth and $\hat{\mu}$-strongly convex. Therefore, our result distinguishes the hardness of minimizing a smooth PL function and a smooth strongly convex function as the complexity of the former cannot be improved by any polynomial order in general.} }
Endnote
%0 Conference Paper %T On the Lower Bound of Minimizing Polyak-Łojasiewicz functions %A Pengyun Yue %A Cong Fang %A Zhouchen Lin %B Proceedings of Thirty Sixth Conference on Learning Theory %C Proceedings of Machine Learning Research %D 2023 %E Gergely Neu %E Lorenzo Rosasco %F pmlr-v195-yue23a %I PMLR %P 2948--2968 %U https://proceedings.mlr.press/v195/yue23a.html %V 195 %X Polyak-Łojasiewicz (PL) (Polyak, 1963) condition is a weaker condition than the strong convexity but suffices to ensure a global convergence for the Gradient Descent algorithm. In this paper, we study the lower bound of algorithms using first-order oracles to find an approximate optimal solution. We show that any first-order algorithm requires at least ${\Omega}\left(\frac{L}{\mu}\log\frac{1}{\epsilon}\right)$ gradient costs to με find an $\epsilon$-approximate optimal solution for a general $L$-smooth function that has an $\mu$-PL constant. This result demonstrates the optimality of the Gradient Descent algorithm to minimize smooth PL functions in the sense that there exists a “hard” PL function such that no first-order algorithm can be faster than Gradient Descent when ignoring a numerical constant. In contrast, it is well-known that the momentum technique, e.g. Nesterov (2003, chap. 2), can provably accelerate Gradient Descent to ${O}\left(\sqrt{\frac{L}{\hat{\mu}}}\log\frac{1}{\epsilon}\right)$ gradient costs for functions that are $L$-smooth and $\hat{\mu}$-strongly convex. Therefore, our result distinguishes the hardness of minimizing a smooth PL function and a smooth strongly convex function as the complexity of the former cannot be improved by any polynomial order in general.
APA
Yue, P., Fang, C. & Lin, Z.. (2023). On the Lower Bound of Minimizing Polyak-Łojasiewicz functions. Proceedings of Thirty Sixth Conference on Learning Theory, in Proceedings of Machine Learning Research 195:2948-2968 Available from https://proceedings.mlr.press/v195/yue23a.html.

Related Material