Type of research: basic Duration from: 01/01/91. to 12/31/93. Papers on project (total): 28
Papers on project quoted in Current Contents: 11
Institution name: Prirodoslovno-matematički fakultet, Matematički odjel, Zagreb (37) Department/Institute: Department of Mathematics, University of Zagreb Address: Bijenička cesta 30 City: 10000 - Zagreb, Croatia
Communication
Phone: (385 1) 4555720/117
E-mail: hari@math.hr
Fax: (385 1) 432484
Summary: The work is devoted to matrix algorithms. The main field of
research is diagonalization methods for eigenvalue and singular value
problems. Special attention is paid to derivation, construction and
investigation of algorithms which can compute eigendecompositions and
singular value decompositions with high relative accuracy. It is known
that only a few of the standard methods and under special constraints
have that property. The task was to find new (primarily diagonalization)
methods that will have this property under more general constraints. From
this area, as part of the project, two Ph. D. theses (I. Slapničar, Z.
Drmač), two M.S. theses (from project only Sanja Singer) and a series of
notable articles (some of them already published) have been written. The
second field of research is connected with asymptotical behavior of
diagonalization methods. Rate of convergence described in terms of linear,
quadratic and cubic convergence essentially determines numerical
efficiency of algorithms. Very sharp quadratic and cubic convergence bounds
for the symmetric and skew-symmetric Jacobi methods have been proved
(Hari). Sharp quadratic convergence bounds for J-symmetric method (Drmač,
Hari) and for method for the simultaneous diagonalization of a pair of
symmetric matrices (Slapničar, Hari) have been proved. Special
investigations have been devoted to (refers to published articles only):
1. Studying structure of a pair of almost diagonal Hermitian matrices with
multiple eigenvalues (Hari), 2. Construction of a modified Jacobi method
for Hermitian matrices which requires 25% less flops per cycle than the
existing Jacobi methods (Hari), 3. Construction of a parallel SVD
algorithm (Manger, Slapničar), 4. Implementation of parallel "for" loops on
multiprocessors (Manger), 5. Relative perturbations of Hermitian matrices
and perturbations of spectral projectors (Slapničar).
Keywords: Eigenvalues and eigenvectors, SVD, diagonalization methods, global and asymptotic convergence, relative accuracy.
Research goals: As was said in the proposal of the project five
years ago, the aim of the project was to form a group of researchers
recognized in the European (or maybe in the world) community of numerical
linear algebraists. This demanding task presumed presence of several
devoted young researchers and a good leadership in scientific research.
To this end a good collaboration was established with the known expert
Professor K. Veselic from Hagen, Germany. The idea behind was that young
researchers finish grade level study (which ends with an M.S. thesis) in
Zagreb and then continue their research in Hagen by dr Veselic. Having
more free time (no lectures to attend, less assistant duties), better
computing facilities and an excellent supervisor, they could finish their
Ph. D. theses in Hagen. After comeback they continue their work, publish
articles and with time show that there is a recognized group of numerical
linear algebraists in Croatia. As such, the members of the group will be
visiting foreign scientific centers and come back with even better
research experience and knowledge. Since all the researchers included in
the project are working at Croatian universities this will also revive
or/and enhance the level of scientific work at domestic universities.
OTHER ACHIEVEMENTS
Name
: I. Slapničar je za svoj doktorat 1992. dobio nagradu
"Preis zur Forderung junger Wissenschaftler der Fernuniversitat Hagen". Other information about the project.