Improved Sample Complexity for Private Nonsmooth Nonconvex Optimization

Authors: Guy Kornowski, Daogao Liu, Kunal Talwar

ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Theoretical We study differentially private (DP) optimization algorithms for stochastic and empirical objectives which are neither smooth nor convex, and propose methods that return a Goldstein-stationary point with sample complexity bounds that improve on existing works. We start by providing a singlepass (ε, δ)-DP algorithm that returns an (α, β)stationary point as long as the dataset is of size eΩ(d/αβ3 + d/εαβ2), which is Ω(d) times smaller than the algorithm of Zhang et al. (2024) for this task, where d is the dimension. We then provide a multi-pass polynomial time algorithm which further improves the sample complexity to eΩ d/β2 + d3/4/εα1/2β3/2 , by designing a sample efficient ERM algorithm, and proving that Goldstein-stationary points generalize from the empirical loss to the population loss.
Researcher Affiliation Collaboration 1Weizmann Institute of Science; work done while interning at Apple. 2Google; work done while at University of Washington and Apple. 3Apple. Correspondence to: Guy Kornowski <EMAIL>, Daogao Liu <EMAIL>.
Pseudocode Yes Algorithm 1 Tree Mechanism Algorithm 2 Nonsmooth Nonconvex Algorithm (based on O2NC (Cutkosky et al., 2023)) Algorithm 3 Single-pass instantiation of O(zt) in Line 7 of Algorithm 2 Algorithm 4 Multi-pass instantiation of O(zt) in Line 7 of Algorithm 2 Algorithm 5 First-order instantiation of O(zt) in Line 7 of Algorithm 2
Open Source Code No The paper does not contain any explicit statements about releasing code, nor does it provide links to code repositories.
Open Datasets No The paper discusses theoretical objectives like F(x) := E ξ P [f(x; ξ)] (stochastic) and b F D(x) := 1 Pn i=1 f(x; ξi) (ERM), and refers to a dataset D = (ξ1, . . . , ξn) Pn as a theoretical construct for analysis. It does not mention any specific named datasets (e.g., MNIST, CIFAR-10) that were used for experiments, nor does it provide any links, DOIs, or citations to actual datasets for public access.
Dataset Splits No The paper does not describe any empirical experiments involving specific datasets, and therefore, it does not mention any training/test/validation dataset splits.
Hardware Specification No The paper focuses on theoretical contributions regarding algorithms and sample complexity. It does not describe any experiments that would require specific hardware, and thus, no hardware specifications are provided.
Software Dependencies No The paper presents theoretical algorithms and their complexity analysis. It does not describe any implemented software or provide specific version numbers for software dependencies.
Experiment Setup No The paper details theoretical algorithms and proves their properties but does not describe an experimental setup, including hyperparameters or system-level training settings, as it does not present empirical results.