Research Group of Boris Khoromskij (till 2016)

Tensor-structured Numerical Methods in High Dimensional Scientific Computing



Group members

      • PD DrSci. Boris Khoromskij, scientific group leader.

      • Dr. Venera Khoromskaia, research fellow.
        Tensor numerical methods, since 2006:
        multilevel canonical-to-Tucker algorithm,
        tensor-structured Hartree-Fock solvers (I, in 2009, II, in 2012),
        grid-based calculation of two-electron integrals,
        Tensor methods for large lattice-type and periodic systems.
        Summation of 3D lattice long-range potentials by assembled canonical/Tucker tensors

        Former Group members

      • Dr. Sergey Dolgov, as a PhD student in 2011 - 2014.
        The Focker-Planck dynamics, parametric PDEs, preconditioning.
        ALS for linear systems in higher dimensions. Two-level Tucker-QTT format.


      • Prof. Ivan Oseledets (01.10.2010 +6 months).


        Long-term and short-term visitors

      • Prof. Sergey Repin
        University of Jyvaskyla, Finland.
        Steklov Institute of Mathematics, St. Petersburg, Russia.


      • Dr. Angelos Mantzaflaris
        RICAM Institute, Linz, Austria

      • Prof. Eugene Tyrtyshnikov (2003-2011)
        Institute for Numerical Analysis, Russian Academy of Science, Moscow, Russia.

      • Prof. Ivan Oseledets (multiple visits in 2009-2012)
        Skoltech University, Moscow, Russia.

      • Dr. Dmitry Savostyanov (multiple visits in 2008-2013)
        University of Brighton, UK.

      • Dr. Sergey Dolgov
        University of Bath, UK.
        multiple visits as a student from Moscow in 2009-2011.

      • Dr. Vladimir Kazeev
        multiple visits as a student from Moscow in 2010-2011.


      Research

        The challenging problems of numerical analysis in higher dimensions (in Rd ) arise in various applications, for example
        • modelling of multi-particle interactions in large molecular systems such as proteins, biomolecules,
        • modelling of large atomic (metallic) clusters,
        • stochastic and parametric equations, uncertainty quantification,
        • multidimensional dynamical systems,
        • electronic structure calculations,
        • machine learning, data mining and information technologies.

        Numerical treatment of high-dimensional PDEs requires nonstandard approaches since the computational complexity of traditional numerical methods scales exponentially in the dimension refered as the "curse of dimensionality", O(nd ).

        The separable tensor product low parametric representations of multidimensional arrays in particular, the canonical and Tucker formats, since long have been used in the computer science community (F.L. Hitchcock (1927), L.R. Tucker (1966), ... L. De Lathauwer, et al (2000)).

        In 2006 it was shown and proved by B. Khoromskij that the Tucker tensor format has exceptional properties for approximating some multivariate functions discretized on d-dimensional Cartesian grid (Newton kernel, Slater function, etc). Next, the canonical-to-Tucker algorithm (C2T) and the reduced higher order singular value decomposition (RHOSVD) have been introduced in 2006, and 2008 by B. Khoromskij and V. Khoromskaia. In these papers it was shown that the canonical and Tucker formats can be efficiently used for the low-rank tensor representation of functions of several variables and for their numerical treatment. RHOSVD can be used for building the tensor formats "avoiding the curse of dimensionality".

        The tensor-structured methods have been initiated by the problem of the grid-based numerical solution of the 3D Hartree-Fock equation in a general basis (V. Khoromskaia, B. Khoromskij, H.-J. Flad). These methods, in particular, include fast and accurate tensor calculation of 3D and 6D convolution operators with the Newton kernel in 1D complexity, (2008), (2009) and (2010). Recent approach for factorized calculation of the two-electron integrals tensor is efficient in electronic structure calculations for large molecular systems. (2012). Our algorithms resulted in a package (3D grid-based "black-box" solver) for the numerical solution of the Hartree-Fock equation in a general basis (2013) .

        The tensor-train (TT) format of O( n d) complexity for the treatment of multidimensional tensors was developed by I. Oseledets and E. Tyrtyshnikov in 2009.

        The quantics N-d tensor approximation in high-dimensional numerical modeling (QTT) was invented by B. Khoromskij in 2009 as a tool to compress function related vectors/tensors with O(d log n ) complexity scaling. In this paper the QTT method was also justified by the proof of uniform rank estimates for a wide class of discretized functions.

        We focus on O(d log n ) -complexity numerical methods for the efficient solution of high-dimensional steady-state and dynamical PDEs. They are based on the QTT approximation of multivariate functions and integral-differential operators. In particular, tensor methods gainfully apply to the large scale problems in molecular dynamics, for the Focker-Planck and master equations, as well as for parametric/stochastic PDEs (B. Khoromskij, Ch. Schwab, I. Oseledets, S. Dolgov, V. Kazeev).

        Superfast QTT-Fourier and QTT-convolution transforms of O(d log n ) complexity open the way to the fast treatment of huge data arrays (B. Khoromskij, E. Tyrtyshnikov, D. Savostyanov, S. Dolgov, V. Kazeev).

        Numerical efficiency and accuracy of our tensor numerical algorithms is confirmed on the examples of real-life problems and in comparison with the existing alternative approaches (analytically based multidimensional integration in quantum chemistry, Monte Carlo and model reduction methods).


      Main results (in chronological order)


      • Low-rank tensor-product representation of a class of multi-dimensional nonlocal operators:
        constructive approximation by sinc-interpolation and quadratures, error analysis, analytic rank optimization.
        [Gavrilyuk, Hackbusch, Khoromskij '02, 03, '05],
        [Hackbusch, Khoromskij, Tyrtyshnikov '03, '04', '05],
        [Bertoglio, Khoromskij '08]

      • Structured hierarchical dimension splitting and Tucker decomposition of function related tensors.
        [Khoromskij '06]

      • Functional Tucker approximation, canonical-to-Tucker decomposition.
        Starting the development of tensor algebra MATLAB library for the calculation of 3D integral operators .

        [Khoromskij '06]
        [Khoromskij, Khoromskaia '06]

      • Theory and numerical algorithms for multidimensional tensor convolution
        in canonical and Tucker formats with applications.
        [Khoromskij, '08]

      • Efficient multigrid canonical-to-Tucker-to-canonical (C2T+T2C) tensor rank reduction for 3D tensors in 1D complexity
        based on the two-level (combined) Tucker-to-canonical decomposition on a sequence of grids,
        reduced higher order SVD for the canonical tensors and concept of most important fibers.

        [Khoromskij, Khoromskaia '08]

      • Tensor methods in applications to Boltzmann and Kohn-Sham equations.
        [Khoromskij'04]
        [Khoromskij'06]

      • Tensor-structured calculation of 3D and 6D integrals in the Fock operator in 1D complexity.
        [Khoromskij, Khoromskaia '08]

      • Computation of the Hartree-Fock (HF) exchange by the tensor-structured methods.
        [Khoromskaia '09]

      • Grid-based solution of the Hartree-Fock equation by multilevel tensor-structured methods.
        3D "black-box" EVP solver for the HF euqation, in 1D complexity.

        [Khoromskij, Khoromskaia, Flad '09]
        [Khoromskaia, '10]

      • Tensor methods for stochastic and parametric PDEs.
        [Khoromskij, Litvinenko, Matthias, '08]
        [Khoromskij, Schwab, '09]
        [Khoromskij, Oseledets, '10]
        [Dolgov, Kazeev, Khoromskij, '12]

      • Novel concept of quantized vector/tensor format of O(d log n ) (log-volume) complexity.
        Approximation theory by low-rank quantized-canonical and quantized-TT (QTT) representation of functional tensors.

        [Khoromskij '09]

      • Super-compressed representation to multivariate polynomials,
        operator exponential, and QTT-DMRG spectral solvers.

        [Khoromskij,'09]
        [Khoromskij, Oseledets '09, '10]

      • Tensor-structured preconditioning in Rd .
        [Khoromskij '09]
        [Dolgov, Khoromskij, Oseledets, Tyrtyshnikov '09]

      • QTT-compression method applied to the d-dimensional elliptic operator inverse,
        convolution transform, FFT and other operations in quantized tensor spaces
        providing log-scaling complexity.

        [Kazeev, Khoromskij '10],
        [Dolgov, Khoromskij, Savostyanov '11],
        [Kazeev, Khoromskij, Tyrtyshnikov '11]

      • Quantum molecular dynamics on tensor manifolds and related spectral calculations by QTT-DMRG and QTT-Cayley-transform iteration.
        Dynamics on tensor manifold by the Dirac-Frenkel variational principle.

        [Khoromskij, Oseledets '10]
        [Gavrilyuk, Khoromskij '11]

      • Time-dependent Fokker-Planck and master equations.
        [Dolgov, Khoromskij, Oseledets '11]
        [Dolgov, Khoromskij '13]

      • Tensor-structured grid-based calculation of the 3D Laplace and nuclear potential parts in the Fock operator.
        [Khoromskaia, Andrae, Khoromskij '12]

      • Efficient grid-based tensor calculation of the two-electron integrals and the Fock operator in a general basis.
        [Khoromskaia, Khoromskij, Schneider '12]

      • Moller-Plesset (MP2) Energy Correction using Tensor Factorization of the Grid-Based Two-Electron Integrals.
        [Khoromskaia, Khoromskij '13]

      • Black-box Hartree-Fock solver by two-electron integrals and the Fock operator in a general basis.
        [Khoromskaia '13]

      • Tensor-structured Hartree-Fock for extended systems.
        Fast Tensor Method for Summation of Long-Range Potentials on 3D Lattices with Defects.

        [Khoromskaia, Khoromskij '13,'14]

      • A Reduced Basis Approach for Calculation of the Bethe-Salpeter Excitation Energies using Low-Rank Tensor Factorizations.
        [Benner, Khoromskaia, Khoromskij '15]

      •  Since 2006: development of the MATLAB library implementing  multi-linear algebra and 
        the rank-structured tensor approximations to a class of multivariate functions, 
        integral transforms and Hamiltonians (electronic structure calculations), many-particle 
        modeling, as well as  the fast tensor-truncated iterative solution methods 
        for high-dimensional stationary and dynamical equations in scientific computing.



    last modified 12.06.2017
    Webdesign V. Khoromskaia