Tight Bounds for Volumetric Spanners and Applications

Authors: Aditya Bhaskara, Sepideh Mahabadi, Ali Vakilian

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Theoretical Our main contribution in this paper is to show that a simple local search algorithm yields volumetric spanners with parameters that improve both lines of prior work by Hazan and Karnin [2016] and Woodruff and Yasuda [2023]. Our arguments also allow us to study the case of having a general ℓp norm bound on the coefficients. Thus, we obtain a common generalization with the results of Awerbuch and Kleinberg [2008] on barycentric spanners (which correspond to the case p = ). Further our results can be plugged in, following the same approach of Woodruff and Yasuda [2023], to obtain improvements on a range of matrix approximation problems in these settings.One application we highlight is the following. Volumetric spanners turn out to be closely related to another well-studied problem, that of finding the minimum volume enclosing ellipsoid (MVEE) for a given set of points, or more generally, for a given convex body K. This is a classic problem in geometry [Welzl, 1991, Khachiyan and Todd, 1990]. The celebrated result of Fitz John (e.g., see [Ball, 1992]) characterized the optimal solution for general K. Computationally, the MVEE can be computed using a semidefinite programming relaxation [Boyd et al., 2004], and more efficient algorithms have subsequently been developed; see [Cohen et al., 2019]. Coresets for MVEE (defined formally below) were used to construct well-conditioned spanning subsets in the recent work of Woodruff and Yasuda [2023]. We give a result in the opposite direction, and show that the local search algorithm for finding well-conditioned spanning sets can be used to obtain a coreset of size O(d/ϵ). This quantitatively improves upon prior work, as we now discuss.
Researcher Affiliation Collaboration Aditya Bhaskara University of Utah bhaskaraaditya@gmail.com Sepideh Mahabadi Microsoft Research Redmond smahabadi@microsoft.com Ali Vakilian Toyota Technological Institute at Chicago (TTIC) vakilian@ttic.edu
Pseudocode Yes Algorithm 1 Procedure Local Search-NR
Open Source Code No The paper does not include any explicit statement about releasing code or a link to a source-code repository.
Open Datasets No The paper is theoretical and does not use or describe any specific datasets for training or evaluation. It refers to abstract 'sets of vectors' but not concrete public datasets.
Dataset Splits No The paper is theoretical and does not mention any data splits for validation or other purposes.
Hardware Specification No The paper is theoretical and does not describe any specific hardware used for experiments.
Software Dependencies No The paper does not specify any software dependencies with version numbers.
Experiment Setup No The paper is theoretical and focuses on algorithmic properties and bounds. It does not describe an experimental setup with hyperparameters or system-level training settings.