This repository presents the source code and examples of using Preconditioned Kaczmarz method, where preconditioner is constructed from QR decomposition of the sketched system matrix. We demonstrate numerically that even a small subset of rows of the system matrix can give sufficiently good preconditioner. The examples of using the presented method are also available here, in the folder examples
.
We provide numerical experiments to demonstrate the performance of the proposed approach:
- Random overdetermined ill-conditioned and consistent linear systems (GitHub, Nbviewer).
- Kernel regression with explicit feature map via Random Fourier Features (GitHub, Nbviewer).
- Image reconstruction from the tomography data (GitHub, Nbviewer). The used images and matrices are stored in
data
folder.
If you use this research in your work, we kindly ask you to cite the paper
@article{katrutsa2019preconditioning,
title={Preconditioning Kaczmarz method by sketching},
author={Katrutsa, Alexandr and Oseledets, Ivan},
journal={arXiv preprint arXiv:1903.01806},
year={2019}
}