Approximate Euclidean lengths and distances beyond Johnson-Lindenstrauss

Authors: Aleksandros Sobczyk, Mathieu Luisier

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Proof-of-concept numerical experiments are presented to validate the theoretical analysis.
Researcher Affiliation Collaboration Aleksandros Sobczyk IBM Research and ETH Zürich Zürich, Switzerland obc@zurich.ibm.com Mathieu Luisier ETH Zürich Zürich, Switzerland mluisier@iis.ee.ethz.ch
Pseudocode Yes Algorithm 1 Adaptive Euclidean Norm Estimation
Open Source Code No No because the code is proprietary.
Open Datasets No The paper discusses generating 'synthetic matrices' and provides details on their construction but does not refer to any publicly available datasets or provide access information for the generated data.
Dataset Splits No The paper describes numerical experiments, but it does not specify any training/test/validation splits. It discusses generating synthetic matrices and running experiments, but no explicit data partitioning is mentioned.
Hardware Specification No Our small scale indicative experiments were ran on a small laptop.
Software Dependencies No Algorithm 1 was implemented in Python using Num Py.
Experiment Setup Yes Specifically, d d matrices A, with d = 5000, were created as follows. We drew a random orthogonal d d matrix Q. We then fixed a diagonal d d matrix Λ which defines the eigenvalues of the matrix. Each element Λi,i, i [d] is set to i c for a given c 0. The larger the c, the faster the spectral decay. We finally constructed the symmetric A = QΛQ which were used in the numerical experiments. Following [34], we applied four different decay factors, specifically c = {0.5, 1, 1.5, 2}. Therefore, we set G, S, and Q in Algorithm 1 to have size d m/4, so that both algorithms are tested with the same number of matrix-vector products.