A Krylov projection algorithm for large symmetric matrices with dense spectra
Fuente:
arXiv
Saved in:
| Main Authors: | , |
|---|---|
| Format: | Preprint |
| Published: |
2025
|
| Subjects: | |
| Online Access: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866908324288528384 |
|---|---|
| author | Druskin, Vladimir Zimmerling, Jörn |
| author_facet | Druskin, Vladimir Zimmerling, Jörn |
| contents | We consider the approximation of $B^T (A+sI)^{-1} B$ for large s.p.d. $A\in\mathbb{R}^{n\times n}$ with dense spectrum and $B\in\mathbb{R}^{n\times p}$, $p\ll n$. We target the computations of Multiple-Input Multiple-Output (MIMO) transfer functions for large-scale discretizations of problems with continuous spectral measures, such as linear time-invariant (LTI) PDEs on unbounded domains. Traditional Krylov methods, such as the Lanczos or CG algorithm, are known to be optimal for the computation of $(A+sI)^{-1}B$ with real positive $s$, resulting in an adaptation to the distinctively discrete and nonuniform spectra. However, the adaptation is damped for matrices with dense spectra. It was demonstrated in [Zimmerling, Druskin, Simoncini, Journal of Scientific Computing 103(1), 5 (2025)] that averaging Gauß and Gauß-Radau quadratures computed using the block-Lanczos method significantly reduces approximation errors for such problems. Here, we introduce an adaptive Kreĭn-Nudelman extension to the (block) Lanczos recursions, allowing further acceleration at negligible $o(n)$ cost. Similar to the Gauß-Radau quadrature, a low-rank modification is applied to the (block) Lanczos matrix. However, unlike the Gauß-Radau quadrature, this modification depends on $\sqrt{s}$ and can be considered in the framework of the Hermite-Padé approximants, which are known to be efficient for problems with branch-cuts, that can be good approximations to dense spectral intervals. Numerical results for large-scale discretizations of heat-diffusion and quasi-magnetostatic Maxwell's operators in unbounded domains confirm the efficiency of the proposed approach. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2504_06998 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | A Krylov projection algorithm for large symmetric matrices with dense spectra Druskin, Vladimir Zimmerling, Jörn Numerical Analysis 65F10, 65N22, 65F50, 65F60 We consider the approximation of $B^T (A+sI)^{-1} B$ for large s.p.d. $A\in\mathbb{R}^{n\times n}$ with dense spectrum and $B\in\mathbb{R}^{n\times p}$, $p\ll n$. We target the computations of Multiple-Input Multiple-Output (MIMO) transfer functions for large-scale discretizations of problems with continuous spectral measures, such as linear time-invariant (LTI) PDEs on unbounded domains. Traditional Krylov methods, such as the Lanczos or CG algorithm, are known to be optimal for the computation of $(A+sI)^{-1}B$ with real positive $s$, resulting in an adaptation to the distinctively discrete and nonuniform spectra. However, the adaptation is damped for matrices with dense spectra. It was demonstrated in [Zimmerling, Druskin, Simoncini, Journal of Scientific Computing 103(1), 5 (2025)] that averaging Gauß and Gauß-Radau quadratures computed using the block-Lanczos method significantly reduces approximation errors for such problems. Here, we introduce an adaptive Kreĭn-Nudelman extension to the (block) Lanczos recursions, allowing further acceleration at negligible $o(n)$ cost. Similar to the Gauß-Radau quadrature, a low-rank modification is applied to the (block) Lanczos matrix. However, unlike the Gauß-Radau quadrature, this modification depends on $\sqrt{s}$ and can be considered in the framework of the Hermite-Padé approximants, which are known to be efficient for problems with branch-cuts, that can be good approximations to dense spectral intervals. Numerical results for large-scale discretizations of heat-diffusion and quasi-magnetostatic Maxwell's operators in unbounded domains confirm the efficiency of the proposed approach. |
| title | A Krylov projection algorithm for large symmetric matrices with dense spectra |
| topic | Numerical Analysis 65F10, 65N22, 65F50, 65F60 |
| url | https://arxiv.org/abs/2504.06998 |