Matrix Denoising with Doubly Heteroscedastic Noise: Fundamental Limits and Optimal Spectral Methods

Authors: Yihan Zhang, Marco Mondelli

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Numerical experiments demonstrate the significant advantage of our theoretically principled method with the state of the art.
Researcher Affiliation Academia Yihan Zhang Institute of Science and Technology Austria zephyr.z798@gmail.com Marco Mondelli Institute of Science and Technology Austria marco.mondelli@ist.ac.at
Pseudocode No The paper describes the AMP algorithm in (5.11) using mathematical equations, but it is not presented as a formal pseudocode block or labeled as an algorithm.
Open Source Code No All experiments use synthetic data and can be reproduced given the instructions in Section 5. Data and code are not released.
Open Datasets No The paper uses synthetic data generated with specified parameters (n=4000, d=2000, P=Q=N(0,1)) and does not refer to a publicly available dataset.
Dataset Splits No The paper uses synthetic data and discusses running '20 i.i.d. trials' for data points, but does not specify training, validation, or test splits in the traditional sense for a dataset.
Hardware Specification No All experiments are synthetic and can be run efficiently on standard personal computers. No specific hardware details (e.g., CPU/GPU models, memory) are provided.
Software Dependencies No The paper mentions 'standard SVD algorithms or power iteration' but does not specify any software names with version numbers.
Experiment Setup Yes In both figures, n = 4000, d = 2000 (so  = 2), and P = Q = N(0, 1). Each data point is computed from 20 i.i.d. trials and error bars are reported at 1 standard deviation. We let  be either the identity or a Toeplitz matrix [77, 41, 19], i.e., i,j = |i j| with  = 0.9. We let  be a circulant matrix [40, 39]: the first row has 1 in the first position, c = 0.0078 in the second through (+ 1)-st position and in the last positions (= 300), with the remaining entries being 0; for 2 i d, the i-th row is a cyclic shift of the (i 1)-st row to the right by 1 position.