Volume 39, Issue 2
Nonnegative Low Rank Matrix Completion by Riemannian Optimalization Methods

Guang-Jing Song & Michael K. Ng

Ann. Appl. Math., 39 (2023), pp. 181-205.

Published online: 2023-06

[An open-access article; the PDF is free to any online user.]

Export citation
  • Abstract

In this paper, we study Riemannian optimization methods for the problem of nonnegative matrix completion that is to recover a nonnegative low rank matrix from its partial observed entries. With the underlying matrix incoherence conditions, we show that when the number $m$ of observed entries are sampled independently and uniformly without replacement, the inexact Riemannian gradient descent method can recover the underlying $n_{1}$-by-$n_{2}$ nonnegative matrix of rank $r$ provided that $m$ is of $\mathcal{O}(r^{2} s \log^2s )$, where $s = \max \{n_{1},n_{2} \}$. Numerical examples are given to illustrate that the nonnegativity property would be useful in the matrix recovery. In particular, we demonstrate the number of samples required to recover the underlying low rank matrix with using the nonnegativity property is smaller than that without using the property.

  • AMS Subject Headings

15A23, 65f22

  • Copyright

COPYRIGHT: © Global Science Press

  • Email address
  • BibTex
  • RIS
  • TXT
@Article{AAM-39-181, author = {Song , Guang-Jing and Ng , Michael K.}, title = {Nonnegative Low Rank Matrix Completion by Riemannian Optimalization Methods}, journal = {Annals of Applied Mathematics}, year = {2023}, volume = {39}, number = {2}, pages = {181--205}, abstract = {

In this paper, we study Riemannian optimization methods for the problem of nonnegative matrix completion that is to recover a nonnegative low rank matrix from its partial observed entries. With the underlying matrix incoherence conditions, we show that when the number $m$ of observed entries are sampled independently and uniformly without replacement, the inexact Riemannian gradient descent method can recover the underlying $n_{1}$-by-$n_{2}$ nonnegative matrix of rank $r$ provided that $m$ is of $\mathcal{O}(r^{2} s \log^2s )$, where $s = \max \{n_{1},n_{2} \}$. Numerical examples are given to illustrate that the nonnegativity property would be useful in the matrix recovery. In particular, we demonstrate the number of samples required to recover the underlying low rank matrix with using the nonnegativity property is smaller than that without using the property.

}, issn = {}, doi = {https://doi.org/10.4208/aam.OA-2023-0010}, url = {http://global-sci.org/intro/article_detail/aam/21833.html} }
TY - JOUR T1 - Nonnegative Low Rank Matrix Completion by Riemannian Optimalization Methods AU - Song , Guang-Jing AU - Ng , Michael K. JO - Annals of Applied Mathematics VL - 2 SP - 181 EP - 205 PY - 2023 DA - 2023/06 SN - 39 DO - http://doi.org/10.4208/aam.OA-2023-0010 UR - https://global-sci.org/intro/article_detail/aam/21833.html KW - Manifolds, tangent spaces, nonnegative matrices, low rank. AB -

In this paper, we study Riemannian optimization methods for the problem of nonnegative matrix completion that is to recover a nonnegative low rank matrix from its partial observed entries. With the underlying matrix incoherence conditions, we show that when the number $m$ of observed entries are sampled independently and uniformly without replacement, the inexact Riemannian gradient descent method can recover the underlying $n_{1}$-by-$n_{2}$ nonnegative matrix of rank $r$ provided that $m$ is of $\mathcal{O}(r^{2} s \log^2s )$, where $s = \max \{n_{1},n_{2} \}$. Numerical examples are given to illustrate that the nonnegativity property would be useful in the matrix recovery. In particular, we demonstrate the number of samples required to recover the underlying low rank matrix with using the nonnegativity property is smaller than that without using the property.

Song , Guang-Jing and Ng , Michael K.. (2023). Nonnegative Low Rank Matrix Completion by Riemannian Optimalization Methods. Annals of Applied Mathematics. 39 (2). 181-205. doi:10.4208/aam.OA-2023-0010
Copy to clipboard
The citation has been copied to your clipboard