Journal article icon

Journal article

The computational complexity of finding stationary points in non-convex optimization

Abstract:

Finding approximate stationary points, i.e., points where the gradient is approximately zero, of non-convex but smooth objective functions f over unrestricted d-dimensional domains is one of the most fundamental problems in classical non-convex optimization. Nevertheless, the computational and query complexity of this problem are still not well understood when the dimension d of the problem is independent of the approximation error. In this paper, we show the following computational and query complexity results:

  1. The problem of finding approximate stationary points over unrestricted domains is PLScomplete.
  2. For d = 2, we provide a zero-order algorithm for finding ε-approximate stationary points that requires at most O(1/ε) value queries to the objective function.
  3. We show that any algorithm needs at least Ω(1/ε) queries to the objective function and/or its gradient to find ε-approximate stationary points when d = 2. Combined with the above, this characterizes the query complexity of this problem to be Θ(1/ε).
  4. For d = 2, we provide a zero-order algorithm for finding ε-KKT points in constrained optimization problems that requires at most O(1/ √ ε) value queries to the objective function. This closes the gap between the works of Bubeck and Mikulincer [2020] and Vavasis [1993] and characterizes the query complexity of this problem to be Θ(1/ √ ε).
  5. Combining our results with the recent result of Fearnley et al. [2022], we show that finding approximate KKT points in constrained optimization is reducible to finding approximate stationary points in unconstrained optimization but the converse is impossible.
Publication status:
Published
Peer review status:
Peer reviewed

Actions

Access Document

Publisher copy:
10.1007/s10107-024-02139-3

Authors

More by this author
Institution:
University of Oxford
Division:
MPLS
Department:
Computer Science
Oxford college:
All Souls College
Role:
Author
ORCID:
0000-0001-5255-9349


Publisher:
Springer
Journal:
Mathematical Programming More from this journal
Volume:
213
Issue:
1
Pages:
281-341
Publication date:
2024-09-27
Acceptance date:
2024-08-12
DOI:
EISSN:
1436-4646
ISSN:
0025-5610


Language:
English
Pubs id:
2026925
Local pid:
pubs:2026925
Deposit date:
2024-09-11
ARK identifier:

Terms of use


Views and Downloads






If you are the owner of this record, you can report an update to it here: Report update to this record

TO TOP