Matrix Computations Seminar: Tensor approximation tools free of the curse of dimensionality

Seminar: LAPACK | November 4 | 11:10 a.m.-12 p.m. | 380 Soda Hall

Speaker/Performer: Prof. Eugene Tyrtyshnikov, Institute of Numerical Mathematics, Russian Academy of Sciences

Sponsor: Electrical Engineering and Computer Sciences (EECS)

Even "simple" case in higher dimensions may require data elements
as many as atoms in the universe. Structure in data in such case
is the key issue. However, existing tensor representations suffer from
various drawbacks. We propose new tensor decompositions called
TENSOR-TRAIN DECOMPOSITIONS and the corresponding numerical algorithms
with then complexity linear in the number of axes. Applications include
interpolation of multi-variate functions, computation of multi-dimensional
integrals, solving PDEs, fast inversion of tensor structured matrices etc.

Event Contact: mgu@math.berkeley.edu, 510-642-3145