Pregled bibliografske jedinice broj: 54672
Accuracy and stability in numerical linear algebra
Accuracy and stability in numerical linear algebra // Proceedings of the Conference Applied Mathematics and Computation, Dubrovnik, Croatia, September 13--18, 1999. / Rogina, Mladen ; Hari, Vjeran ; Limić, Nedžad ; Tutek, Zvonimir (ur.).
Zagreb: Matematički odsjek Prirodoslovno-matematičkog fakulteta Sveučilišta u Zagrebu, 2001. str. 1-8 (predavanje, međunarodna recenzija, cjeloviti rad (in extenso), pregledni)
CROSBI ID: 54672 Za ispravke kontaktirajte CROSBI podršku putem web obrasca
Naslov
Accuracy and stability in numerical linear algebra
Autori
Drmač, Zlatko
Vrsta, podvrsta i kategorija rada
Sažeci sa skupova, cjeloviti rad (in extenso), pregledni
Izvornik
Proceedings of the Conference Applied Mathematics and Computation, Dubrovnik, Croatia, September 13--18, 1999.
/ Rogina, Mladen ; Hari, Vjeran ; Limić, Nedžad ; Tutek, Zvonimir - Zagreb : Matematički odsjek Prirodoslovno-matematičkog fakulteta Sveučilišta u Zagrebu, 2001, 1-8
Skup
Applied Mathematics and Computation
Mjesto i datum
Dubrovnik, Hrvatska, 13.09.1999. - 18.09.1999
Vrsta sudjelovanja
Predavanje
Vrsta recenzije
Međunarodna recenzija
Ključne riječi
accuracy; eigenvalues; singular values
Sažetak
We give an overview of our recent work on ccurate computation of generalized eigenvalue and ingular value decompositions of matrix pairs and triplets. Our goal is to develop efficient and highly accurate algorithms and to produce high quality mathematical software. Using error analysis and perturbation theory, we develop templates for accurate floating point computation of the product and quotient induced singular value decompositions, canonical correlations and diagonalization of symmetric pencils $H-\lambda M$, $HM-\lambda I$ with positive definite $H$, $M$. The new algorithms are numerically robust. For instance, the eigenvalues of $H-\lambda M$ and $HM-\lambda I$ are computed with optimal relative error bound: each eigenvalue $\lambda$ is computed with relative error $|\delta\lambda|/\lambda$ of order (up to a factor of the dimension) ${\bf u}\{ \min_{\Delta\in{\cal D}}\kappa_2(\Delta H\Delta) + \min_{\Delta\in{\cal D}}\kappa_2(\Delta M\Delta)\}$, where ${\bf u}$ is the roundoff unit and ${\cal D}$ is the set of nonsingular diagonal matrices. Moreover, the backward error is element-wise small in the sense that finite precision computation corresponds to an exact computation with $H+\delta H$, $M+\delta M$, where for all $i, j$ and some moderate functions $f()$ and $g()$, $|\delta H_{ij}|\leq f(n){\bf u} \sqrt{H_{ii}H_{jj}}$, $|\delta M_{ij}|\leq g(n) {\bf u}\sqrt{M_{ii}M_{jj}}$.
Izvorni jezik
Engleski
Znanstvena područja
Matematika
POVEZANOST RADA
Projekti:
037012
Ustanove:
Prirodoslovno-matematički fakultet, Matematički odjel, Zagreb
Profili:
Zlatko Drmač
(autor)