You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
IN MEMORY OF ...<br />
Professor Emeritus Theodore S.<br />
Papatheodorou, a prominent<br />
faculty member of the Department<br />
of Computer Engineering &<br />
Informatics at the University of<br />
Patras at Patras-Greece, died on<br />
December 10, 2012. Theo, as we all<br />
called him, was an internationally<br />
renowned educator and<br />
researcher. He was a leading and<br />
recognized expert in Numerical<br />
Analysis.<br />
He was born in Patras-Greece in<br />
1945 and graduated with honors Theodore S. Papatheodorou<br />
from the University of Athens with a<br />
1945 - 2012<br />
B.S. degree in Mathematics in 1968.<br />
He entered the graduate program at<br />
Purdue University in 1969, and<br />
subsequently earned his M.S. in Applied Mathematics and in Civil<br />
Engineering, and Ph.D. in Science in 1971, 1975 and 1973,<br />
respectively under the guidance of Professor Robert Lynch. He<br />
joined Clarkson University in 1976 and served there until 1984<br />
when he was invited to join the newly formed Greek Ministry of<br />
Research & Development as scientific advisor. With his efforts<br />
the Computer Technology Institute was established and he<br />
served as its first Director until 1990. His numerous<br />
accomplishments over the years contributed immensely to<br />
building the reputation of the Institute as a leader in<br />
Computing research. Meanwhile he joined the faculty of the<br />
Department of Computer Engineering & Informatics at the<br />
University of Patras where he established the High Performance<br />
Information Systems Lab (HPCLab) and also served as Dean,<br />
Head, member of the University Senate, Director of graduate<br />
studies, etc.<br />
His research contributions covered various aspects of<br />
Computational Science and Computer Engineering. Seminal<br />
contributions include research and development work on system<br />
and application software for parallel and distributed<br />
computing, scientific computing, web and multimedia<br />
applications, 3D virtual reconstruction of monuments and other<br />
applications for culture and education, while, at the early years,<br />
his contribution in the development of a general methodology<br />
for generating arbitrary high order finite difference methods<br />
has influenced the invention of the so called HODIE methods<br />
and contributed significant fast methods that became part of<br />
ACM algorithms and ELLPACK libraries. He published over 150<br />
papers, including a book and he supervised and guided the<br />
doctoral work of fifteen students.
Theo was member of many international scientific committees<br />
and was awarded several prizes and awards.<br />
Theo taught a wide range of courses for his entire academic<br />
career. He was a gifted teacher with selfless service and with<br />
keen interest for his students. He influenced with his academic<br />
presence the lives of a great number of students and<br />
colleagues.<br />
Professor Emeritus Theodore S. Papatheodorou will be<br />
remembered as a gentleman, a scholar and an unassuming<br />
researcher. On a personal note, being his advisee, collaborator<br />
and close friend for over thirty years, he will always be<br />
remembered as “MY TEACHER” and beloved friend.<br />
Yiannis G. Saridakis<br />
Technical University of Crete.
NumAn2014 Book of Abstracts iv<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
Contents<br />
Invited speakers<br />
Houstis N. E.<br />
Remembering Theo Papatheodorou . . . . . . . . . . . . . . . . . . . . . . . . . 1<br />
Bai Z-Z.<br />
Scalable and Fast Iteration Methods for Complex Linear Systems . . . . . . . . 2<br />
Fokas S. A.<br />
The interplay of the concrete and general: from PDEs to medical imaging . . . 3<br />
Iserles A.<br />
Fast computation of there semiclassical Schrödinger equation . . . . . . . . . . 4<br />
Noutsos D.<br />
Perron-Frobenius Theory Some Extensions and Applications . . . . . . . . . . 5<br />
Vrahatis N. M.<br />
Sign Methods for Imprecise Problems . . . . . . . . . . . . . . . . . . . . . . . 6<br />
Contributed speakers<br />
Abhulimen E.C.<br />
A new class of second derivative methods for numerical integration of stiff initial<br />
value problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7<br />
Agaoglou M., Rothos V.M. and Susanto H.<br />
Homoclinic chaos in a pair of parametrically-driven coupled SQUIDs . . . . . . 8<br />
Alefragis P., Spyrou A. and Likothanassis S.<br />
Application of a hybrid parallel Monte Carlo PDE Solver on rectangular multidomains<br />
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9<br />
Antoniadou I. K. and Voyatzis G.<br />
Continuation and stability deduction of resonant periodic orbits in three dimensional<br />
systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10<br />
Antonopoulos C. and Bellas N.<br />
SOpenCL: An Infrastructure for Transparently Integrating FPGAs in Heterogeneous,<br />
Accelerator-Based Systems . . . . . . . . . . . . . . . . . . . . . . . . 11<br />
Antonopoulos C., Maroudas M. and Vavalis M.<br />
Software Platforms for Multi-Domain Multi-Physics Simulations . . . . . . . . 12<br />
Antonopoulos C. D. and Dougalis A. V.<br />
Error estimates for the standard Galerkin-Finite Element method for the shallow<br />
water equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13<br />
Antonopoulos G. C., Srivastava S., Pinto S. S., Baptista S. M.<br />
Do Brain Networks Evolve by Maximizing Flow of Information? . . . . . . . . 14<br />
Antonopoulou D.<br />
Finite elements for a class of nonlinear stochastic pdes from phase transition<br />
problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15<br />
v
NumAn2014 Book of Abstracts vi<br />
Antunes R. S. P.<br />
Numerical Solution of the Magnetic Laplacian Eigenvalue Problem using Radial<br />
Basis Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16<br />
Árciga A. M. P., Ariza H. F. J. and Sánchez O. J.<br />
Stochastic Riez-Fractional Partial Differential Equation with White Noise on<br />
the Half-Line . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17<br />
Ashton A.<br />
Functional Analytic Framework of the Fokas Method for Elliptic Boundary Value<br />
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18<br />
Athanasakis E. I., Papadopoulou P. E. and Saridakis G. Y.<br />
Discontinuous Hermite Collocation and Runge-Kutta schemes for multi-domain<br />
linear and non-linear brain tumor invasion models . . . . . . . . . . . . . . . . 19<br />
Athanasakis E. I., Vilanakis D. N., Mathioudakis N. E., Papadopoulou P. E. and Saridakis<br />
G. Y.<br />
Solving discontinuous collocation equations for a class of brain tumor models on<br />
GPUs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20<br />
Atisattapong W. and Maruphanton P.<br />
Obviating the Bin Width Effect of the 1/t Algorithm for Multidimensional Numerical<br />
Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21<br />
Bacigaluppi P. and Ricchiuto M.<br />
A 1D stabilized finite element model for non-hydrostatic wave breaking and run-up 22<br />
Barrera D., Ibáñez J. M., Roldán M. A., Roldán B. J. and Yáñez R.<br />
Parameter determination in MOSFETs transitors based on Discrete Orthogonal<br />
Chebyshev polynomials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23<br />
Bebiano N.<br />
Spectral inclusion regions for matrix pencils . . . . . . . . . . . . . . . . . . . . 24<br />
Belhah Z., Kouibia A., Pasadas M.<br />
Variational iterative method for solving a nonlinear partial differential equation.<br />
Application to the two-dimensional Bratu’s problem . . . . . . . . . . . . . . . 25<br />
Bellas N. and Antonopoulos C.<br />
Significance-Based Computing for Reliability and Power Optimization . . . . . 26<br />
Bellavia S., Governi L., Papini A. and Puggelli L.<br />
Quadratic Penalty Methods for Shape from Shading . . . . . . . . . . . . . . . 27<br />
Benmir M., Bessonov N., Boujena S. and Volpert V.<br />
Multi-scale hybrid model of cell differentiation propagation as traveling waves . 28<br />
Benzi M., Duff S. I. and Guo X-P.<br />
Preconditioned derivative-free globally convergent Newton-GMRES methods for<br />
large sparse nonlinear systems . . . . . . . . . . . . . . . . . . . . . . . . . . . 29<br />
Berisha F., Sadiku M. and Berisha N.<br />
Using an Euler type transform for accelerating convergence of series . . . . . . 30<br />
Bobolakis D.E., Delis A.I. and Mathioudakis E.N.<br />
Efficient Solution of the Two-Dimensional Shallow-Water Equations using GPUs 31<br />
Bountis T., Antonopoulos C. and Skokos H.<br />
Complex Statistics and Diffusion in Nonlinear Disordered Particle Chains . . . 32<br />
Bratsos G. A.<br />
A modified predictor-corrector method for the generalized BurgersHuxley equation 33<br />
Bueno I. M., Curlett K. and Furtado S.<br />
Structured Strong Linearizations obtained from Fiedler Pencils with Repetition 34<br />
Burde I. G., Nasibullayev Sh. I. and Zhalij A.<br />
Unified Semi-Analytical, Semi-Numerical Approach to Stability Analysis of Nonparallel<br />
Unsteady Flows . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts vii<br />
Bu Y-M. and Carpentieri B.<br />
A recursive multilevel approximate inverse-based preconditioner for solving general<br />
linear systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36<br />
Carpio J, Prieto L. J.<br />
A local anisotropic adaptive algorithm to solve time-dependent dominated convection<br />
problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37<br />
Charalampaki E. N. and Mathioudakis N. E.<br />
CPU-GPU computations for MultiGrid techniques coupled with Fourth-Order<br />
Compact Discretizations for Isotropic and Anisotropic Poisson problems . . . . 38<br />
Chaturantabut S.<br />
Nonlinear Model Reduction with Localized Basis for Two-Phase Miscible Flow<br />
in Porous Media . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39<br />
Chatzipantelidis P.<br />
On positivity preservation for finite element based methods for the heat equation 40<br />
Chollom P. J. and Kumleng M. G.<br />
Block Hybrid Numerical Integrators for the Solution of Stiff Equations . . . . . 41<br />
Dang D.-M., Christara C. and Jackson K.<br />
Efficient GPU pricing of interest rate derivatives: PDE formulation and ADI<br />
methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42<br />
Christodoulidi H., Cirto L., Bountis T. and Tsallis C.<br />
Dynamical and statistical behavior of the Fermi-Pasta-Ulam model with longrange<br />
interactions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43<br />
Crooks K.<br />
Two numerical implementations of the Fokas method for elliptic equations in a<br />
polygon . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44<br />
Crowdy G. D. and Luca E.<br />
Solving Wiener-Hopf problems without kernel factorisation . . . . . . . . . . . 45<br />
De Bonis M. C. and Occorsio D.<br />
Numerical evaluation of hypersingular integrals on the semiaxis . . . . . . . . . 46<br />
Demetriou C. I.<br />
A Characterization Theorem for the Discrete Best L 1 Monotonic Approximation<br />
Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47<br />
Dieci L., Elia C. and Lopez L.<br />
Numerical techniques for sliding motion in Filippov discontinuous systems . . . 48<br />
El-Gindy T.M., Salim M.S. and Ahmed A.I.<br />
A new filled function method applied to unconstrained global optimization . . 49<br />
Fernández L., Fortes A. M. and Rodríguez M. L.<br />
Multiresolution analysis for 3D scattered data sets . . . . . . . . . . . . . . . . 51<br />
Fevgas A., Tsompanopoulou P. and Bozanis P.<br />
Exploring the Performance of Out-of-Core Linear Algebra Algorithms in Flash<br />
based Storage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52<br />
Filelis-Papadopoulos K. C. and Gravvanis A. G.<br />
A comparative study on the effect of the ordering schemes for solving sparse<br />
linear systems, based on factored approximate sparse inverse matrix methods . 53<br />
Flouri E., Dougalis V. and Synolakis C.<br />
Tsunami hazard and inundation for the northern coast of Crete . . . . . . . . . 54<br />
Fokas S. A. and Kalimeris K.<br />
Eigenvalues and eigenfunctions for the Laplace Operator . . . . . . . . . . . . . 55<br />
Fortes A. M., González P., Palomares A. and Pasadas M.<br />
Filling holes with geometric constraints . . . . . . . . . . . . . . . . . . . . . . 56<br />
Fortes A. M., González P., Palomares A. and Pasadas M.<br />
Matrix-free resolution of PDEs using the Powell-Sabin FE . . . . . . . . . . . . 57<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts viii<br />
Gaitani M., Kazolea M. and Delis A.<br />
Numerical Solution for Sparse Linear Systems that occur from the discretization<br />
of Boussinesq-type equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58<br />
Georgieva I., Hofreither C. and Uluchev R.<br />
Approximations Using Radon Projection Data in the Unit Disc . . . . . . . . . 59<br />
González-Pinto S., Hernández-Abreu D.<br />
Splitting methods based on Approximate Matrix Factorization and Radau-IIA<br />
formulas for the time integration of advection diffusion reaction PDEs . . . . . 60<br />
Grylonakis G. E.N., Filelis-Papadopoulos K. C. and Gravvanis A. G.<br />
On the numerical modelling and solution of multi-asset Black-Scholes equation<br />
based on Generic Approximate Sparse Inverse Preconditioning . . . . . . . . . 61<br />
Gu C. and Zhang K.<br />
The Error Analysis of the Indirect Pade Method for Matrix Exponential . . . . 62<br />
Guedouar R., Bouzabia A. and Zarrad B.<br />
Optimization of pre-recontruction restoration filtering for filtered back projection<br />
reconstruction (FBP) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63<br />
Hadjidimos A. and Tzoumas M.<br />
On the Solution of the Linear Complementarity Problem by the Generalized<br />
Accelerated Overrelaxation Iterative Method . . . . . . . . . . . . . . . . . . . 64<br />
Hadjimichael Y. and Ketcheson I. D.<br />
Strong-stability-preserving additive linear multistep methods . . . . . . . . . . 65<br />
Hadjinicolaou M.<br />
Fokas method and Kelvin transformation applied to potential problems in non<br />
convex unbounded domains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66<br />
Hananel A., Pasadas M. and Rodríguez L. M.<br />
Construction and approximation of surfaces by smoothing meshless methods . 67<br />
Hashemzadeh P. and Fokas S A.<br />
The definitive estimation of the neuronal current via EEG and MEG using real<br />
data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68<br />
Hashemzadeh P. and Fokas S A.<br />
Numerical Solution of The Unified Transform For Linear Elliptic PDEs in Polygonal<br />
Domains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69<br />
Hassoun Y. and Othman H.<br />
Symmetric Key Cryptography Algorithms Based on Numerical Methods . . . . 70<br />
Hitzazis I.<br />
The Fokas Method and Initial-Boundary Value Problems for Multidimensional<br />
Integrable PDEs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72<br />
Hong X.-L., Meng L.-S. and Zheng B.<br />
Some new perturbation bounds of generalized polar decomposition . . . . . . . 73<br />
Huang Yu.-M. and Zhang X.-Y.<br />
On block preconditioners for PDE-constrained optimization problems . . . . . 74<br />
Kalosakas G.<br />
Modeling drug release kinetics . . . . . . . . . . . . . . . . . . . . . . . . . . . 75<br />
Kanellopoulos G. and van der Weele k.<br />
Granular Transport Dynamics: Numerics and Analysis . . . . . . . . . . . . . . 76<br />
Kastis A. G., Gaitanis A. and Fokas S A.<br />
Quantitative evaluation of SRT for PET imaging: Comparison with FBP and<br />
OSEM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77<br />
Kazolea M., Delis I. A. and Synolakis E. C.<br />
A wave breaking mechanism for an unstructured finite volume scheme . . . . . 78<br />
Khoshkhoo R. and Jahangirian A.<br />
Numerical Simulation of Flow Separation Control using Dielectric Barrier Discharge<br />
plasma actuator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts ix<br />
Kim M., Jung H.-K. and Park S.<br />
An effective approach on finite-difference-time-domain method for quasi-static<br />
electromagnetic field analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80<br />
Kim S. and Zhang H.<br />
Domain decomposition method with complete radiation boundary conditions for<br />
the Helmholtz equation in waveguides . . . . . . . . . . . . . . . . . . . . . . . 81<br />
Kincaid R. D., Chen J-Y. and Li Yu-C.<br />
Generalizations and Modifications of Iterative Methods for Solving Large Sparse<br />
Indefinite Linear Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82<br />
Kontogiorgos P., Sarri E., Vrahatis N. M. and Papavassilopoulos P. G.<br />
An energy market stackelberg game solved with particle swarm optimization . 83<br />
Korfiati A., Tsompanopoulou P. and Likothanassis S.<br />
Serial and Parallel Implementation of the Interface Relaxation Method GEO . 84<br />
Kouloukas T.<br />
A special class of integrable Lotka-Voltera systems and their Kahan discretization 85<br />
Kourounis D.<br />
Constraint handling for gradient-based optimization of compositional reservoir<br />
flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86<br />
Lisitsa V. and Tcheverda V.<br />
Combining discontinuous Galerkin and Finite Differences methods for simulation<br />
of seismic wave propagation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87<br />
Liu Z., Yamanea Y., Tsujib T. and Tanaka T.<br />
Decreasing Computational Load by Using Similarity for Lagrangian Approach<br />
to Gas-solid Two-phase Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88<br />
Makhanov S.<br />
Curvilinear Grids for Five-Axis Machining . . . . . . . . . . . . . . . . . . . . . 89<br />
Manapova A. and Lubyshev F.<br />
Numerical Solution of Optimization Problems for Semilinear Elliptic Equations<br />
with Discontinuous Coefficients and Solutions . . . . . . . . . . . . . . . . . . . 90<br />
Mandikas G. V., Mathioudakis N. E., Kozyrakis V. G., Ekaterinaris A. J. and Kampanis<br />
A. N.<br />
A MultiGrid accelerated high-order pressure correction compact scheme for incompressible<br />
Navier-Stokes solvers . . . . . . . . . . . . . . . . . . . . . . . . . 91<br />
Muslu M. G. and Borluk H.<br />
A Fourier Collocation Method for the Nonlocal Nonlinear Wave Equation . . . 92<br />
Mylonas K. I., Rothos M. V., Kevrekidis G. P. and Frantzeskakis J. D.<br />
Perturbation Theory of Dark-Bright solitons in Bose-Einstein condensates . . . 93<br />
Nikas A. I.<br />
Efficient Unconstrained Optimization Multistart Solvers Using a Self-Clustering<br />
Technique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94<br />
Noutsos D., Serra-Capizzano S. and Vassalos P.<br />
Essential spectral equivalence via multiple step preconditioning and applications<br />
to ill conditioned Toeplitz matrices . . . . . . . . . . . . . . . . . . . . . . . . . 95<br />
Occorsio D. and Russo G. M.<br />
Nyström methods for two-dimensional Fredholm integral equations on unbounded<br />
domains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96<br />
Okulicka-D̷lużewska F., Smoktunowicz A.<br />
Numerical stability of block direct methods for solving symmetric saddle point<br />
problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97<br />
Owolabi M. K. and Patidar C. K.<br />
Robust numerical simulation of reaction-diffusion models arising in Mathematical<br />
Ecology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts x<br />
Pelloni B. and Smith A. D.<br />
Unified Tranforms and classical spectral theory of operators . . . . . . . . . . . 99<br />
Petsounis K.<br />
MATLAB : Parallel and Distributed Computing using CPUs and GPUs . . . . 100<br />
Prusińska A. and Tretýakov A.A.<br />
Method for solving nonlinear singular problems . . . . . . . . . . . . . . . . . . 101<br />
Sablonniére P. a and Barrera D.<br />
Solving the Fredholm integral equation of the second kind by global spline quasiinterpolation<br />
of the kernel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102<br />
Sacconi A.<br />
On the comparison between fitted and unfitted finite element methods for the<br />
approximation of void electromigration . . . . . . . . . . . . . . . . . . . . . . . 103<br />
Sattarzadeh S. and Jahangirian A.<br />
A Numerical Mesh-Less Method for Solving Unsteady Compressible Flows . . . 104<br />
Sattarzadeh S., Jahangirian A. and Ebrahimi M.<br />
A Numerical Adaptive Mesh-Less Method for Solution of Compressible Flows . 105<br />
Schioppa Jr. E., Verkerke W., Visser J. and Koffeman E.<br />
Solving CT reconstruction with a particle physics tool (RooFit) . . . . . . . . . 106<br />
Shmerling E.<br />
Ziggurat algorithm for sampling from bivariate distributions . . . . . . . . . . . 107<br />
Sifalakis G. A., Papadomanolaki G. M., Papadopoulou P. E. and Saridakis G. Y.<br />
Fokas transform method for classes of advection-diffusion IBVPs . . . . . . . . 108<br />
Sintunavarat W.<br />
Approximate algorithm for single valued nonexpansive and multi-valued strictly<br />
pseudo contractive mappings in Hilbert . . . . . . . . . . . . . . . . . . . . . . 109<br />
Spanakis C., Marias K. and Kampanis A. N.<br />
Application of an image registration method based on maximization of mutual<br />
information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110<br />
Spivak A.<br />
Successive approximations for optimal control in some nonlinear systems with<br />
small parameter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111<br />
Stylianopoulos N.<br />
Inverse moment problems with applications in shape reconstruction . . . . . . 112<br />
Szczepanik E., Tretýakov A.<br />
Method for solving degenerate sub-definite nonlinear equations . . . . . . . . . 113<br />
Stratis P.N., Karatzas G.P., Papadopoulou E.P. and Saridakis Y.G.<br />
Stochastic optimization for a problem of saltwater intrusion in coastal aquifers<br />
with heterogeneous hydraulic conductivity . . . . . . . . . . . . . . . . . . . . . 114<br />
Taha T.<br />
Numerical simulations for 1+2 dimensional coupled nonlinear Schrödinger type<br />
equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115<br />
Tsakiri K. and Marsellos A.<br />
A Numerical Model for the prediction of flooding in Water Rivers . . . . . . . 116<br />
Tsompanopoulou P.<br />
Interface Rexation Methods for the solution of Multi-Physics Problems . . . . . 117<br />
Ukpebor A. L.<br />
An Order 19-rational integrator . . . . . . . . . . . . . . . . . . . . . . . . . . . 118<br />
Valtchev S S., Alves J. S. C. and Martins F. M. N.<br />
A Meshfree Method with Fundamental Solutions for Inhomogeneous Elastic<br />
Wave Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119<br />
Vavalis M. and Zimeris D.<br />
On the Numerical Solution of Power Flow Problems . . . . . . . . . . . . . . . 120<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts xi<br />
Venetis E. I., Kouris A., Nikoloutsakos N., Sobczyk A. and Gallopoulos E.<br />
Towards robust parallel solvers for tridiagonal systems for multiGPUs . . . . . 121<br />
Venetis E. I., Nikoloutsakos N., Gallopoulos E. and Ekaterinaris J.<br />
Local Stiffness Matrix Calculations for FSI Applications on multi-GPU Systems 122<br />
Wang Z.-Q.<br />
Chebyshev accelerated preconditioned MHSS iteration methods for a class of<br />
block two-by-two linear systems . . . . . . . . . . . . . . . . . . . . . . . . . . . 123<br />
Yang X.<br />
The WR-HSS Methods for Non-Self-Adjoint Positive Definite Linear Differential<br />
Equations and Applications to the Unsteady Discrete Elliptic Problem . . . . . 124<br />
Zaitseva A. and Lisitsa V.<br />
Sensitivity of the Domain Decomposition Method to Perturbation of the Transmission<br />
Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125<br />
Zakynthinaki M.<br />
An improved model of heart rate kinetics . . . . . . . . . . . . . . . . . . . . . 126<br />
Zambelli A.<br />
Normalizations of the Proposal Density in Markov Chain Monte Carlo Algorithms127<br />
Zhang G.-F. and Zheng Z.<br />
A local preconditioned alternating direction iteration method for generalized<br />
saddle point problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128<br />
Zhang Y. and Li Q.<br />
Katservich Algorithm Based on Spherical Detector for Cone-Beam CT and the<br />
Implementation on GPU . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129<br />
Zhao Z., Bai Z.-J. and Jin X.-Q.<br />
A Riemannian Newton Algorithm for Nonlinear Eigenvalue Problems . . . . . 130<br />
Zouraris E. G.<br />
Finite element approximations for a linear stochastic Cahn-Hilliard-Cook equation131<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts xii<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 1<br />
Remembering Theo Papatheodorou<br />
Elias N. Houstis<br />
Department of Electrical and Computer Engineering,<br />
University of Thessaly,<br />
Volos, Greece<br />
enh@inf.uth.gr<br />
”On the day I die, don’t say he’s gone<br />
Death has nothing to do with going away<br />
The sun sets, and the moon sets but theyre not gone<br />
Death is a coming together<br />
The human seed goes down into the ground like a bucket,<br />
and comes up with some unimagined beauty<br />
Your mouth closes here, and immediately opens<br />
with a shout of joy there”<br />
Rumi<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 2<br />
Conference in Numerical Analysis 2014 (NumAn 2014)<br />
September 2-5, 2014<br />
Chania,Greece<br />
Scalable and Fast Iteration Methods<br />
for Complex Linear Systems<br />
Zhong-Zhi Bai a ,<br />
a State Key Laboratory of Scientific/Engineering Computing<br />
Institute of Computational Mathematics and Scientific/Engineering Computing<br />
Academy of Mathematics and Systems Science<br />
Chinese Academy of Sciences, P.O. Box 2719, Beijing 100190, P.R. China<br />
bzz@lsec.cc.ac.cn<br />
Abstract<br />
Complex system of linear equations arises in many important applications. We further explore<br />
algebraic and convergence properties and present analytical and numerical comparisons among<br />
several available iteration methods such as C-to-R and PMHSS for solving such a class of linear<br />
systems. Theoretical analyses and computational results show that reformulating the complex<br />
linear system into an equivalent real form is a feasible and effective approach, for which we<br />
can construct, analyze and implement accurate, efficient and robust preconditioned iteration<br />
methods.<br />
Key words: complex symmetric linear system, real reformulation, PMHSS iteration, preconditioning,<br />
convergence theory, spectral properties.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 3<br />
Conference in Numerical Analysis 2014 (NumAn 2014)<br />
September 2-5, 2014<br />
Chania,Greece<br />
The interplay of the concrete and general:<br />
from PDEs to medical imaging<br />
Athanassios S. Fokas<br />
Department of Applied Mathematics and Theoretical Physics<br />
University of Cambridge, Cambridge, CB3 0WA, UK<br />
T.Fokas@damtp.cam.ac.uk<br />
Abstract<br />
The need to solve a concrete problem of physical significance occasionally leads to the development<br />
of a new mathematical technique. It is often realised that this technique can actually<br />
be used for the solution of a plethora of other problems, and thus it becomes a mathematical<br />
method. In this lecture a review will be presented on how a problem posed by the late Julian<br />
Cole led to the development of the so called unified transform, which provides a novel and<br />
powerful treatment to boundary value problems for linear and integrable non-linear PDEs. Interesting<br />
connections with the Riemann hypothesis, as well as the development of several effective<br />
algorithms for Medical Imaging, will also be reviewed.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 4<br />
Fast computation of the semiclassical Schrödinger equation<br />
Arieh Iserles<br />
Department of Applied Mathematics and Theoretical Physics,<br />
Centre for Mathematical Sciences,<br />
University of Cambridge,<br />
Cambridge CB3 OWA United Kingdom<br />
ai10@cam.ac.uk<br />
Abstract<br />
The computation of the semiclassical Schrödinger equation presents a number of difficult challenges<br />
because of the presence of high oscillation and the need to respect unitarity. Typical strategy<br />
involves a spectral method in space and Strang splitting in time, but it is of low accuracy and<br />
sensitive to high oscillation. In this talk we sketch an alternative strategy, based on high-order<br />
symmetric Zassenhaus splittings, combined with spectral collocation, which preserve unitarity<br />
and whose accuracy is immune to high oscillation. These splittings can be implemented with<br />
large time steps and allow for an exceedingly affordable computation of underlying exponentials.<br />
The talk will be illustrated by the computation of different quantum phenomena.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book Conference of Abstracts in Numerical Analysis 2014 (NumAn 2014) 5<br />
September 2-5, 2014<br />
Chania,Greece<br />
Perron-Frobenius Theory – Some Extensions and Applications<br />
Dimitrios Noutsos<br />
Department of Mathematics, University of Ioannina,<br />
Ioannina, Greece<br />
dnoutsos@uoi.gr<br />
Abstract<br />
The Perron-Frobenius theory on nonnegative matrices was introduced by Perron and Frobenius<br />
in the beginning of the 20th century. Since its construction the Perron-Frobenius theory has been<br />
developed and constituted a basic Linear Algebra tool to study and solve problems arising from<br />
applications in discretization of Differential and Integral Equations, Markov chains, Economics,<br />
Biosciences, etc. The class of M-matrices, which was introduced and studied in the meantime,<br />
appears in many of the aforementioned applications. Also, some classes of splittings (regular,<br />
weak regular, nonnegative, etc.) were proposed for the solution of large linear algebraic systems<br />
of equations by iterative methods.<br />
The Perron-Frobenius theory of nonnegative matrices was extended to matrices which have<br />
some negative entries, by D. Noutsos [Linear Algebra Appl., 412 (2006), no 2–3, 132–153].<br />
Some properties which give information, when a matrix possesses a Perron-Frobenius eigenpair,<br />
were presented and proved. This class of matrices is associated to eventually nonnegative<br />
matrices, namely matrices whose powers become and remain nonnegative. The class of<br />
Perron-Frobenius splitting was proposed and studied for the solution of linear systems by classical<br />
iterative methods. Some properties of the type of Stein-Rosenberg theorem were extended<br />
to the class of Perron-Frobenius splittings by D. Noutsos [Linear Algebra Appl., 429 (2008),<br />
1983–1996].<br />
Linear differential systems ẋ(t) = Ax(t), A ∈ R n,n , whose solutions become and remain<br />
nonnegative, were studied by D. Noutsos and M. J. Tsatsomeros [SIAM J. Matrix Anal. Appl.,<br />
30 (2008), no 2, 700–712]. Initial conditions that result to nonnegative states are shown to form<br />
a convex cone that is related to the matrix exponential e tA and its eventual nonnegativity.<br />
Further extension of the Perron-Frobenius theory of nonnegative matrices to certain complex<br />
matrices was proposed and proved by D. Noutsos and R. S. Varga [Linear Algebra Appl., 437<br />
(2012), 1071–1088].<br />
Recently, B. Iannazzo and D. Noutsos, in a forthcoming paper, have considered an extension<br />
of the Perron-Frobenius theory to matrices obtained by a suitable scaling (with positive and<br />
negative entries) applied to an M-matrix. This problem appears, for instance, in the study of<br />
Algebraic Riccati Equations arising in fluid queues, where one is interested in the invariant<br />
subspaces of the matrix<br />
[ ]<br />
A −B<br />
H =<br />
,<br />
−C −D<br />
obtained by changing the sign of the last m rows of an M-matrix.<br />
Key words: Nonnegative Matrices, Perron-Frobenius Theory, Eventually Nonnegative Matrices, M-Matrices,<br />
Riccati Equation.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book Conference of Abstracts in Numerical Analysis 2014 (NumAn 2014) 6<br />
September 2-5, 2014<br />
Chania, Greece<br />
Sign Methods for Imprecise Problems<br />
Michael N. Vrahatis<br />
Computational Intelligence Laboratory (CILab),<br />
Department of Mathematics, University of Patras,<br />
GR-26110 Patras, Greece<br />
vrahatis@math.upatras.gr<br />
Abstract<br />
Tackling problems with imprecise (not exactly known) information occur in different scientific<br />
fields including mathematics, physics, astronomy, meteorology, engineering, computer science,<br />
biomedical informatics, medicine and bioengineering among others. In many applications, precise<br />
function values are either impossible or time consuming to obtain. For example, when the<br />
function values depend on the results of numerical simulations, then it may be difficult or impossible<br />
to get very precise values. Or in other cases, it may be necessary to integrate numerically<br />
a system of differential equations in order to obtain a function value, so that the precision of the<br />
computed value is limited. Furthermore, in many problems the accurate values of the function<br />
are computationally expensive.<br />
Ideas from the topological degree theory and combinatorial topology (algebraic topology)<br />
have led to the introduction of iterative root-finding and fixed point methods as well as numerical<br />
optimization methods for tackling problems with impressions. We call these methods sign<br />
methods since the only computable information required is the algebraic sign of the function that<br />
is the smallest amount of information (one bit of information) necessary for the purpose needed,<br />
and not any additional information. In this contribution, some of these methods are reviewed<br />
and applications to computational mathematics and computational intelligence are presented.<br />
Key words: Sign Methods, Root-finding Methods, Fixed Point Methods, Numerical Optimization Methods,<br />
Imprecise Problems<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 7<br />
A new class of Second Derivative methods for numerical<br />
Integration of Stiff Initial Value Problems<br />
C. E. Abhulimen<br />
Department of Mathematics<br />
Ambrose Alli University<br />
Ekpoma, Nigeria<br />
cletusabhulimen@yahoo.co.uk<br />
Abstract<br />
In this paper, we construct a new class of four-step second derivative exponential fitting<br />
method of order six for the numerical integration of stiff initial-value problems of the type:<br />
y ′ = f(x, y); y(x 0 ) = y 0<br />
The implicit method which is based on the work of Cash [1], possess free parameters which allow<br />
it to be fitted automatically to exponential functions. For the purpose of effective implementation<br />
of the new proposed method, we adopt the mechanism in[1] by splitting the method into<br />
predictor and corrector schemes. The numerical analysis of the stability of the new method was<br />
discussed and some numerical experiments confirming theoretical expectations are provided.<br />
Finally, the numerical results show that the new method is A-stable and compete favorably with<br />
the existing methods in terms of efficiency and accuracy.<br />
Keywords: Second derivative four-step, exponentially fitted, A-stable, stiff initial value problems<br />
2010 MSC: 65L05, 65L2O<br />
References<br />
[1] J.R Cash (1981). ”On the exponential fitting of composite multiderivative linear multistep<br />
methods” SIAM J. Numer Annal 18(5), (1981), 808-821<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 8<br />
Homoclinic chaos in a pair of parametrically-driven coupled<br />
SQUIDs<br />
Makrina Agaoglou a , Vassilios M Rothos a and Hadi Susanto b<br />
a Department of Mechanical Engineering, Faculty of Engineering, Aristotle<br />
University of Thessaloniki, Thessaloniki 54124, Greece<br />
b Department of Mathematical Sciences, University of Essex, Wivenhoe Park,<br />
Colchester CO4 3SQ, United Kingdom<br />
rothos@auth.gr,hsusanto@essex.ac.uk<br />
Abstract<br />
An rf superconducting quantum interference device (SQUID) consists of a superconducting ring<br />
interrupted by a Josephson junction (JJ). When driven by an alternating magnetic field, the induced<br />
supercurrents around the ring are determined by the JJ through the celebrated Josephson<br />
relations. This system exhibits rich nonlinear behavior, including chaotic effects. We study<br />
the dynamics of a pair of parametrically-driven coupled SQUIDs arranged in series. We take<br />
advantage of the weak damping that characterizes these systems to perform a multiple-scales<br />
analysis and obtain amplitude equations, describing the slow dynamics of the system. This picture<br />
allows us to expose the existence of homoclinic orbits in the dynamics of the integrable<br />
part of the slow equations of motion. Using high-dimensional Melnikov theory, we are able to<br />
obtain explicit parameter values for which these orbits persist in the full system, consisting of<br />
both Hamiltonian and non-Hamiltonian perturbations, to form so-called Silnikov orbits, indicating<br />
a loss of integrability and the existence of chaos. Extensive numerical analysis requiring<br />
algorithms of rapid numerical integration are required to follow the solutions for long times and<br />
verify the accuracy of the analytical results.<br />
Key words: superconducting quantum interference device, Josephson junction, Near-Integrable Hamiltonian<br />
Systems, Silnikov Chaos, Numerical Simulations<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 9<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book Conference of Abstracts in Numerical Analysis 2014 (NumAn 2014) 10<br />
September 2-5, 2014<br />
Chania,Greece<br />
Continuation and stability deduction of resonant periodic orbits<br />
in three dimensional systems<br />
Kyriaki I. Antoniadou and George Voyatzis<br />
Section of Astrophysics, Astronomy and Mechanics, Department of Physics,<br />
Aristotle University of Thessaloniki,<br />
Thessaloniki, 54124, Greece<br />
kyant@auth.gr, voyatzis@auth.gr<br />
Abstract<br />
The general three body problem (GTBP) through the implementation of periodic orbits computed<br />
in a suitable rotating frame of reference can be used in order to describe the dynamics<br />
of planets locked in a mean motion resonance. The families of periodic orbits, either planar or<br />
spatial, derived by specific continuation processes can, also, constitute paths that can drive the<br />
planetary migration.<br />
In Hamiltonian systems, it is known that in phase the stable periodic orbits are surrounded<br />
by invariant tori, while in the neighbourhood of unstable periodic orbits chaotic regions exist.<br />
It has been shown numerically that in the vicinity of stable periodic orbits, where the motion<br />
is regular and bounded, exoplanetary systems can survive, whereas in case they are found near<br />
unstable ones, they will eventually destabilize and the planets may collide or even escape. The<br />
significance of periodic orbits is therefore taken for granted and the accuracy of their computation<br />
is apparently crucial.<br />
We herein depict examples of resonant periodic orbits, exploit analytic continuation and<br />
elaborate on matters of both horizontal and vertical stability. Particularly, we consider the spatial<br />
GTBP and study the dynamics of planetary systems consisting of a Star and two inclined<br />
Planets, which evolve into mean motion resonance. We attempt a comparative study between<br />
three methods used for the deduction of the stability of a periodic orbit: the computation of<br />
eigenvalues, stability indices and Fast Lyapunov Indicator. Finally, we construct maps of dynamical<br />
stability in the vicinity of periodic orbits, in order to identify the extent of stable regions<br />
in phase space.<br />
Key words: periodic orbits, horizontal and vertical stability, mean motion resonance<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 11<br />
SOpenCL: An Infrastructure for Transparently Integrating<br />
FPGAs in Heterogeneous, Accelerator-Based Systems 1<br />
Christos Antonopoulos and Nikolaos Bellas<br />
Department of Electrical and Computer Engineering, University of Thessaly,<br />
Volos, Greece<br />
cda, nbellas@uth.gr<br />
Abstract<br />
The use of heterogeneous parallel architectures appears as a promising approach in the HPC<br />
domain, due to both the absolute performance and the high performance/power ratio these architectures<br />
offer. Heterogeneous systems are typically organized as a number of computational<br />
accelerators, such as GPUs, DSPs etc., complementing one or more general purpose CPUs.<br />
Field Programmable Gate Arrays (FPGAs) are hardware devices that offer a sea of gates and<br />
memory islands which can be configured as digital circuits, thus implementing algorithms at<br />
the hardware level. FPGAs are an excellent accelerator choice, as they can often prove more<br />
power-efficient than conventional CPUs and even GPUs.<br />
Despite the favorable power/performance characteristics, the adoption of FPGAs in the HPC<br />
domain is rather limited. The implementation of algorithms at the hardware-level requires experience<br />
on hardware design and the use of specialized hardware-description languages (Verilog,<br />
VHDL, SystemC), thus remaining outside the realm of domain experts and software engineers.<br />
In this talk we present SOpenCL, a tool infrastructure that facilitates the wider use of FPGAs<br />
in reconfigurable systems. SOpenCL translates algorithmic descriptions at the software level<br />
to equivalent circuit descriptions in Verilog, which can then be directly implemented on an<br />
FPGA. We use OpenCL, a popular and industry supported parallel programming standard for<br />
heterogeneous systems, as the programming model of choice for the software-level algorithmic<br />
descriptions. This way, programs targeted at CPUs or GPUs can transparently and without any<br />
further development effort be executed on FPGA-based accelerator systems as well.<br />
Key words: Heterogeneous Systems, OpenCL, FPGAs, High-level synthesis.<br />
1 The present research work has been co-financed by the European Union (European Social Fund ESF) and Greek<br />
national funds through the Operational Program ”Education and Lifelong Learning” of the National Strategic Reference<br />
Framework (NSRF) - Research Funding Program: THALIS. Investing in knowledge society through the European<br />
Social Fund.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 12<br />
Software Platforms for Multi-Domain Multi-Physics<br />
Simulations 1<br />
Christos Antonopoulos, Manolis Maroudas and Manolis Vavalis<br />
Department of Electrical and Computer Engineering, University of Thessaly,<br />
Volos, Greece<br />
{cda,kapamaroo,mav}@uth.gr<br />
Abstract<br />
Advances in hardware and software technologies in the 1980s led to the modern era of scientific<br />
modeling and simulation. This era seems to come to an end. The simulation needs in both industry<br />
and academia mismatch with the existing software platforms and practices, which to a great<br />
extent have remained unchanged for the past several decades. We foresee that this mismatch,<br />
together with the emerging ICT advances and the cultural changes in scientific approaches will<br />
lead to a new generation of modeling and simulation.<br />
This paper proposes approaches for designing, analyzing, implementing and evaluating new<br />
simulation frameworks particularly suited to multi-domain and multi-physics (MDMP) problems<br />
that have Partial Differential Equations (PDEs) in their foundations. We focus on introducing<br />
software platforms that facilitate the numerical solution of PDEs associated with MDMP<br />
mathematical models.<br />
In particular, we propose an enhanced meta-computing environment which is based on: (a)<br />
scripting languages (like python) and their practices, and (b) on the Service Oriented Architecture<br />
(SOA) paradigm and the associated web services technologies.<br />
The proposed environment has been designed and engineered having in mind a set of characteristics<br />
particularly suited for MDMP problems. More specifically, it:<br />
• Allows domain experts to focus on expressing the models, rather than delving into implementation<br />
details.<br />
• Fully utilizes the plethora of PDE solving modules available.<br />
• Allows the programmer to effectively select the most appropriate available software module<br />
for the particular component of the problem, as this is defined by its associated single<br />
physics model and its simple/single domain<br />
• Transparently benefits from recent algorithmic advances (e.g. domain decomposition)<br />
• Allows users to efficiently deploy and run the MDMP computations on loosely coupled<br />
distributed and heterogeneous compute engines.<br />
Although our design is generic, covering a wide range of problems, our proof of concept<br />
implementation is restricted to elliptic PDEs in two or three dimensions. Furthermore, it clearly<br />
shows that our tool can easily exploit state of the art numerical solvers like those available in<br />
FENICS and deal.II, domain decomposition methods with or without overlapping, Monte Carlo<br />
based hybrid solvers, rectangular or curvilinear domains and interfaces and beyond.<br />
Key words: Numerical Solution of PDEs, Multi-domain, Multi-physics, Problem Solving Environments.<br />
1 The present research work has been co-financed by the European Union (European Social Fund ESF) and Greek<br />
national funds through the Operational Program Education and Lifelong Learning of the National Strategic Reference<br />
Framework (NSRF) - Research Funding Program: THALIS. Investing in knowledge society through the European<br />
Social Fund.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 13<br />
Error Estimates for the Standard Galerkin-Finite Element<br />
Method for the Shallow Water Equations<br />
D. C. Antonopoulos and V. A. Dougalis<br />
Institute of Computational and Applied Mathematics, FORTH, 70013 Heraklion,<br />
Greece<br />
dougalis@iacm.forth.gr<br />
Abstract<br />
We consider a simple initial-boundary-value problem for the shallow water equations on a finite<br />
interval, and also the analogous problem for a symmetric variant of the system that we justify<br />
for small-amplitude solutions. Assuming smoothness of solutions we discretize these problems<br />
in space using the standard Galerkin-finite element method and prove L 2 -error estimates for<br />
the semidiscrete problem for quasiuniform and uniform meshes. In particular we show that<br />
the semidiscretization with piecewise linear, continuous functions on a uniform mesh posseses<br />
optimal-order O(h 2 ) L 2 -error estimates. We also examine time-stepping of the semidiscrete<br />
problems with three explicit Runge-Kutta methods (the Euler, improved Euler, and the Shu-<br />
Osher scheme), and prove L 2 -error estimates for the resulting full discretizations that are of<br />
optimal order in the temporal variable. We also discuss the cases of periodic and absorbing<br />
boundary conditions.<br />
Key words: Shallow water equations, fully discrete Galerkin methods, error estimates.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 14<br />
Conference in Numerical Analysis 2014 (NumAn 2014)<br />
September 2-5, 2014<br />
Chania, Greece<br />
Do Brain Networks Evolve by Maximizing Flow of Information?<br />
Chris G. Antonopoulos, Shambhavi Srivastava, Sandro S. Pinto, Murilo S.<br />
Baptista<br />
Department of Physics, University of Aberdeen, Institute for Complex Systems<br />
and Mathematical Biology (ICSMB),<br />
Aberdeen, AB24 3UE, UK<br />
chris.antonopoulos@abdn.ac.uk<br />
Abstract<br />
In this talk, I will first present unexpected structural and functional similarities we have, recently,<br />
been able to find in the C.elegans and human brain networks. Based on these findings, we then<br />
propose an appropriately constructed model for the evolution of such networks by adding and<br />
retaining new connections between neurons of the network that lead to a subsequent increase<br />
of the upper bound of Mutual Information Rate (MIR), a quantity related to the amount of<br />
information the brain network can process. This idea is reminiscent of the Hebbian rule of<br />
learning and synaptic plasticity. I will show the ability of our model for brain evolution to<br />
capture important properties, such as synchronization and upper bound of MIR patterns, of<br />
realistic already evolved brain networks. Finally, I will comment on some of the computational<br />
aspects arising in this study, like numerical integration methods, accuracy and computation time<br />
regarding the serial or parallel implementation of the model.<br />
Key words:<br />
Brain networks, Evolutionary process, Hindmarsh-Rose dynamics, Synchronization, Mutual Information<br />
Rate (MIR), Upper bound of MIR<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 15<br />
Conference in Numerical Analysis 2014 (NumAn 2014)<br />
September 2-5, 2014<br />
Chania,Greece<br />
Finite elements for a class of nonlinear stochastic pdes<br />
from phase transition problems<br />
Dimitra Antonopoulou a ,<br />
a Department of Mathematics and Applied Mathematics, University of Crete,<br />
GR-714 09 Heraklion, Greece, and Institute of Applied and Computational<br />
Mathematics, FORTH, GR-711 10 Heraklion, Greece<br />
danton@tem.uoc.gr<br />
Abstract<br />
We construct Galerkin numerical schemes with possible discontinuities in time for a class of<br />
nonlinear evolutionary pdes with additive noise. These equations appear in phase transitions<br />
problems and may involve a positive parameter ε which stands as a measure for the inner interfacial<br />
regions width. Our goal is to establish existence of numerical solution and derive optimal<br />
error estimates even for the discontinuous Galerkin case in the presence of noise.<br />
Key words: Finite elements, nonlinear stochastic pdes, dG methods.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book Conference of Abstracts in Numerical Analysis 2014 (NumAn 2014) 16<br />
September 2-5, 2014<br />
Chania,Greece<br />
Numerical Solution of the Magnetic Laplacian Eigenvalue<br />
Problem using Radial Basis Functions<br />
Pedro R. S. Antunes a,b<br />
a Group of Mathematical Physics of the University of Lisbon,<br />
Lisbon, Portugal<br />
b Department of Mathematics, Lusophone University of Humanities and<br />
Technologies,<br />
Lisbon, Portugal<br />
pant@cii.fc.ul.pt<br />
Abstract<br />
We consider the numerical solution of the Magnetic Laplacian eigenvalue problem,<br />
{<br />
(i∇ + F ) 2 u = λu in Ω,<br />
u = 0 on ∂Ω.<br />
(1)<br />
where u(x) is complex-valued and the vector potential is F (x) = β(−x 2 , x 1 ). The magnetic<br />
field is ∇ × F = (0, 0, 2β), where β ∈ R is constant. In this work we study the application<br />
of a numerical method based on radial basis functions. It is well known that the Kansa method<br />
allows for the numerical solution of boundary value problems using radial basis functions and<br />
the efficiency has been verified in a wide range of problems. It is a meshfree method, which<br />
can have high accuracy, provided an appropriate shape parameter is chosen. On the other hand,<br />
a disadvantage of the method is that the matrices involved tend to become progressively more<br />
ill-conditioned as the rank increases. In this work we propose a numerical algorithm based<br />
on the Generalized Singular Value Decomposition to circumvent the ill-conditioning. Several<br />
numerical simulations are presented to illustrate the good performance of the method.<br />
Key words: Magnetic Laplacian, eigenvalue problem, radial basis functions.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 17<br />
Conference in Numerical Analysis 2014 (NumAn 2014)<br />
September 2-5, 2014<br />
Chania,Greece<br />
Stochastic Riez-Fractional Partial Differential Equation with<br />
White Noise on the Half-Line<br />
Árciga A. Martín P. a , Ariza H. Francisco J. a and Sánchez O. Jorge a<br />
a Unidad Académica de Matemáticas, Universidad Autónoma de Guerrero,<br />
Chilpancingo, Guerrero, México<br />
mparcigae@gmail.com,aarizahfj@gmail.com,jsanchezmate@gmail.com<br />
Abstract<br />
We consider a Bayesian numerical solution of a stochastic Riesz-fractional partial differential<br />
equation with white noise on the half-line. This equation is given by<br />
u t = Dx α u + N u + Ḃ(x, t), x, t > 0 (1)<br />
where Dx α is the Riesz-fractional derivative, N is a nonlinear operator and Ḃ(x, t) is the white<br />
noise. To construct the integral representation of solution we use the Fokas’ Method.<br />
Key words: Fractional derivative, Fokas’ Method, Brownian motion.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book Conference of Abstracts in Numerical Analysis 2014 (NumAn 2014) 18<br />
September 2-5, 2014<br />
Chania,Greece<br />
Functional Analytic Framework of the Fokas Method<br />
for Elliptic Boundary Value Problems<br />
Anthony Ashton a<br />
a DAMTP, University of Cambridge,<br />
United Kingdom<br />
acla2@damtp.cam.ac.uk<br />
Abstract<br />
We give an overview of the functional analytic framework for the Fokas approach to elliptic<br />
boundary value problems in convex planar domains. The global relation can be interpreted as<br />
an operator equation of the form Ax = y, where y depends on the known data of a given boundary<br />
value problem and x corresponds to the unknown boundary values. We study the functional<br />
analytic properties of the operator A : X → Y where X, Y are Banach spaces of complex<br />
analytic functions that are similar to the classical Paley-Wiener spaces. These results are important,<br />
not only from a theoretical perspective, but are essential for establishing convergence and<br />
stability results for the numerical implementation of this approach to boundary value problems.<br />
Finally, we give a brief account of some recent results that are applicable to elliptic boundary<br />
value problems in three dimensional polyhedra.<br />
Key words: Fokas Method, Boundary Value Problems, Functional Analysis, Operator Theory.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 19<br />
Discontinuous Hermite Collocation and Runge-Kutta schemes<br />
for multi-domain linear and non-linear brain tumor invasion<br />
models 1<br />
I.E. Athanasakis ∗ , E.P. Papadopoulou and Y.G. Saridakis<br />
Applied Mathematics and Computers Laboratory (AMCL)<br />
Technical University of Crete<br />
Chania 73100, Greece<br />
∗ g.athanasakis@amcl.tuc.gr<br />
Abstract<br />
Growth simulation models of aggressive forms of malignant brain tumors have been well developed<br />
over the past years. In our recent works we have considered both novel analytical and<br />
numerical methods for the efficient treatment of brain tumor models that, apart from proliferation<br />
and diffusion, are being characterized by a discontinuous diffusion coefficient to incorporate<br />
the heterogeneity of the brain tissue. In this direction we have recently introduced a Discontinuous<br />
Hermite Collocation (DHC) finite element method, with appropriately discontinuous basis<br />
functions associated with the discontinuity nodes. The method was coupled with Diagonally<br />
Implicit (DI) Runge-Kutta schemes and studied for a three region linear model to reveal its high<br />
order approximation properties. In this work, we consider extending our results in the following<br />
directions:<br />
• Employment of both linear and non-linear multi-domain brain tumor models<br />
• Coupling of the DHC with both DI and Strong Stability Preserving (SSP) Runge-Kutta<br />
schemes.<br />
Their behavior is being examined and several experiments are included to demonstrate their<br />
performance.<br />
1 The present research work has been co-financed by the European Union (European Social Fund ESF) and Greek<br />
national funds through the Operational Program Education and Lifelong Learning of the National Strategic Reference<br />
Framework (NSRF) - Research Funding Program: THALIS. Investing in knowledge society through the European<br />
Social Fund.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 20<br />
Conference in Numerical Analysis 2014 (NumAn 2014)<br />
September 2-5, 2014<br />
Chania,Greece<br />
Solving discontinuous collocation equations for a class of<br />
brain tumor models on GPUs 1<br />
I.E. Athanasakis, N.D. Vilanakis ∗ , E.N. Mathioudakis,<br />
E.P. Papadopoulou and Y.G. Saridakis<br />
Applied Mathematics and Computers Laboratory<br />
Technical University of Crete<br />
Chania, Crete, Greece<br />
∗ nivilanakis@amcl.tuc.gr<br />
Abstract<br />
Brain tumor models, that incorporate brain’s heterogeneity, have been well developed in the last<br />
decades. The core PDE, that models tumor’s cell diffusion and proliferation properties, is being<br />
characterized by a discontinuous diffusion coefficient, since tumor cells migrate with different<br />
rates in brain’s white and gray matter. In recent years, working towards the development of<br />
high order approximation methods, we have introduced and studied Discontinuous Hermite<br />
Collocation (DHC) methods coupled with traditional as well as high order semi implicit and<br />
strongly stable Runge-Kutta (RK) time discretization schemes. In this work the problem at<br />
hand is the efficient solution of the linear model tumor invasion problem in 1+2 dimensions.<br />
Tensor product formulated fourth order DHC method is used as spatial discretization to produce<br />
a system of ODEs, to be solved, in the sequel, by third order Diagonally-Implicit RK (DIRK)<br />
schemes. Therefore, in each time step a large linear system of order O(N 2 ), where N is the<br />
number of elements in each dimension, has to be solved demanding quite intense computational<br />
effort. Its efficient solution by incomplete factorization preconditioned BiCG stabilized iterative<br />
method (as the eigenvalue topology suggests) in GPU computational environments is presented<br />
and several numerical experiments are used to demonstrate its performance.<br />
1 The present research work has been co-financed by the European Union (European Social Fund ESF) and Greek<br />
national funds through the Operational Program ’Education and Lifelong Learning’ of the National Strategic Reference<br />
Framework (NSRF) - Research Funding Program: THALIS. Investing in knowledge society through the European<br />
Social Fund.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book Conference of Abstracts in Numerical Analysis 2014 (NumAn 2014) 21<br />
September 2-5, 2014<br />
Chania,Greece<br />
Obviating the Bin Width Effect of the 1/t Algorithm<br />
for Multidimensional Numerical Integration<br />
Wanyok Atisattapong a and Pasin Maruphanton a<br />
a Department of Mathematics and Statistics, Faculty of Science and Technology,<br />
Thammasat University, Pathum Thani, Thailand 12120<br />
wanyok@mathstat.sci.tu.ac.th, oporkabbb@hotmail.com<br />
Abstract<br />
In this work we improve the accuracy and the convergence of the 1/t algorithm [1] for multidimensional<br />
numerical integration. The 1/t algorithm has been proposed as an improved version<br />
of the Wang-Landau algorithm [2] which belongs to the class of Monte Carlo methods. After<br />
the lower bound y min and the upper bound y max of the integral are determined by a domain<br />
sampling run [3], the integral can then be approximated by<br />
I =<br />
∫ b<br />
a<br />
y∑<br />
max<br />
y(x)dx ≃ g(y).y, (1)<br />
y min<br />
where g(y) ≡ {x|x ∈ [a, b], y ≤ y(x) ≤ y + dy} and dy is the bin width of y. The distribution<br />
g(y) can be obtained from the 1/t algorithm. However, the errors of estimated integrals saturate<br />
because of the bin width effect. To obviate this effect, we introduce a new approximation<br />
method based on the simple sampling Monte Carlo method by using the average of y values<br />
in the subinterval [y, y + dy], which varies as the number of Monte carlo trials changes, instead<br />
of the fixed value of y.<br />
The non-convergence of the 1/t algorithm [4, 5] and the convergence of the new method are<br />
proved by theoretical analysis. A potential of the method is illustrated by the evaluation of one-,<br />
two- and multi- dimensional integrals up to six dimensions. The dynamic behavior of accuracy<br />
shows that the numerical estimates from our method converge to their exact values without<br />
either error saturation or the bin with effect in contrast with the conventional 1/t algorithm.<br />
Key words: Monte Carlo method, Numerical integration, the 1/t algorithm, Bin width effect<br />
References<br />
[1] R. E. Belardinelli, S. Manzi, and V. D. Pereyra (2008), Phys. Rev. E. 78, 067701.<br />
[2] Y. W. Li, Wüst, D. P. Landau, and H. Q. Lin (2007), Comput. Phys. Commum. 177, 524.<br />
[3] A. Tröster and C. Dellago (2005), Phys. Rev. E. 71, 066705.<br />
[4] C. Zhou and J. Su (2008), Phys. Rev. E. 78, 046705.<br />
[5] Y. Komura and Y. Okabe (2012), Phys. Rev. E. 85, 010102 (R).<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
Conference in Numerical Analysis 2014 (NumAn 2014)<br />
NumAn2014 Book of Abstracts September 2-5, 2014 22<br />
Chania,Greece<br />
A 1D stabilized finite element model for<br />
non-hydrostatic wave breaking and run-up<br />
Paola Bacigaluppi a and Mario Ricchiuto b<br />
a Department of Mathematics, Universität Zürich,<br />
Zürich, Switzerland<br />
b Inria Bordeaux Sud-Ouest,<br />
Talence cedex, France<br />
paola.bacigaluppi@gmail.com,Mario.Ricchiuto@inria.fr<br />
Abstract.<br />
A new methodology is presented to model the propagation, wave breaking and run-up of waves in coastal<br />
zones. Propagation is modelled by a form of the enhanced Boussinesq equations (Madsen and Sorensen,<br />
Coast.Eng. 1992), while the forming of a roller in breaking regions is captured by reverting to the shallow<br />
water equations and allowing waves to locally converge into discontinuities. The switch between the two<br />
models is defined by a wave breaking criterion that depends on several physical parameters, including<br />
the shape and celerity of the wave and the presence of dry areas. To discretize the system we propose<br />
a non-linear variant of the stabilized finite element method of (Ricchiuto and Filippini, J.Comput.Phys.<br />
2014). To guarantee monotone shock capturing, a technique based on a non-linear mass-lumping allows<br />
to provide local non-oscillatory approximations of discontinuities reverting from a third order scheme in<br />
smooth regions to a first order upwind scheme. The local character of the mass-lumping is guaranteed by<br />
the use of limiters, or of properly defined smoothness sensors. The presented scheme guarantees positivity<br />
preservation, well balancedness and the treatment of wet/dry fronts. The wave breaking is triggered<br />
by means of three different criteria, including a local implementation of the theoretical convective criterion<br />
of (Bjørkavåg and Kalisch, Phys.Letters A 2011), which have been thoroughly analysed and tested.<br />
The model obtained is validated on several benchmarks showing excellent agreement with the available<br />
experimental data. As an example the figure below shows the run-up of a periodic wave over a constant<br />
slope. In particular the solution between the black lines corresponds to the detected breaking area and is<br />
computed through the shallow water model.<br />
Key words: Wave propagation, wave breaking, shock-capturing, stabilized finite elements, SUPG scheme,<br />
Boussinesq equations, shallow water equations, wet/dry fronts, wave breaking model.<br />
Fig.: Snapshots of the first and last breaking instants for a periodic wave run up on a constant slope.<br />
Result obtained with the local variant of the convective criteria.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 23<br />
Conference in Numerical Analysis 2014 (NumAn 2014)<br />
September 2-5, 2014<br />
Chania, Greece<br />
Parameter determination in MOSFETs transitors based<br />
on Discrete Orthogonal Chebyshev polynomials<br />
D. Barrera a , M. J. Ibáñez a , A. M. Roldán b , J. B. Roldán b and R. Yáñez a<br />
a Department of Applied Mathematics, University of Granada,<br />
Granada, Spain<br />
b Department of Electronics, University of Granada,<br />
Granada, Spain<br />
{dbarrera,mibanez,amroldan,jroldan,ryanez}@ugr.es<br />
Abstract<br />
Transistors, and in particular MOSFETs (Metal Oxide Semiconductor Field Effect Transistors),<br />
are the most used basic building blocks of integrated circuits (ICs). The complexity of current<br />
chips makes essential their accurate characterization to use them for circuit design purposes.<br />
For each generation of transistors the main electrical features have to be modeled in order to<br />
reproduce them as a function of the voltages differences applied between their terminals. The<br />
models (usually known as compact models) consist of a set of analytical equations and a set<br />
of parameters to include in those equations. A different set of parameters is used for each<br />
fabrication technology. These models are used in TCAD circuit simulation tools and also for<br />
hand-calculations used at the first stages of circuit design.<br />
The extraction of the parameters of new technologies is essential since the capacities of<br />
circuit designers are dependant on the accuracy of model parameters that in many cases are<br />
linked to important physical effects.<br />
Each parameter is obtained in a different way. However, few of them share some features in<br />
common, at least from the numerical viewpoint. In this respect, several parameters are obtained<br />
by means of extrapolation methods (for example threshold voltage calculation), linear regression<br />
(determination of the body factor), slope calculations (extraction of the DIBL parameter), etc.<br />
In all these procedures, the determination of portions of curves that can be approximated by a<br />
straight line is crucial. In this work we just deal with this issue trying to shed light by means of<br />
advanced numerical techniques.<br />
We have developed a method to determine the number of straight line portions contained<br />
in a curve in an automatic manner. The algorithm developed, based on discrete orthogonal<br />
polynomials, can be used for parameter extraction purposes. It consist on the isolation of straight<br />
line portions in experimental or simulated data and the determination of the slope of those curve<br />
sections to calculate one or more parameters of a compact model.<br />
Key words: Discrete orthogonal polynomials, straight line portion, MOSFET.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book Conference of Abstracts in Numerical Analysis 2014 (NumAn 2014) 24<br />
September 2-5, 2014<br />
Chania,Greece<br />
Spectral inclusion regions for matrix pencils<br />
Natalia Bebiano<br />
Department of Mathematics, University of Coimbra,<br />
Coimbra, Portugal<br />
bebiano@mat.uc.pt<br />
Abstract<br />
Consider the linear pencil A − λB, where A and B are n × n complex matrices and λ ∈ C.<br />
Our main purpose is to obtain spectral inclusion regions for the pencil based on certain fields<br />
of values. Namely, we propose efficient methods for the numerical approximation of the field<br />
of values of A − λB denoted by W (A, B). Our approach builds on the fact that the field of<br />
values can be reduced under compressions to the bidimensional case, in which case these sets<br />
can be exactly determined. The obtained results are illustrated by numerical examples. We point<br />
out that the given procedure to approximate W (A, B) compares well with those existing in the<br />
literature.<br />
Key words: Field of values, Numerical range, Linear pencil, Eigenvalue, Compression.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 25<br />
Conference in Numerical Analysis 2014 (NumAn 2014)<br />
September 2-5, 2014<br />
Chania,Greece<br />
Variational iterative method for solving a nonlinear partial<br />
differential equation. Application to the two-dimensional<br />
Bratu’s problem<br />
Z. Belhah a , A. Kouibia b , and Miguel Pasadas b<br />
a LERMA–Engineering Mohammedia School,Rabat, Morocco.<br />
b Department of Applied Mathematics, University of Granada,<br />
Granada, Spain<br />
z.belhaj@gmail.com, kouibia@ugr.es, mpasadas@ugr.es<br />
Abstract<br />
In this paper we present a variational approximation method for solving the two–dimensional<br />
Bratu’s problem. The existence and the uniqueness of this problem are shown. Moreover, we<br />
construct a sequence of bicubic splines approximating the problem solution and depending of<br />
the knots number of a sequence of partitions of the domain. Such sequence converge to the<br />
exact solution of the problem. Finally, we analyze some numerical examples in order to show<br />
the efficiency of our method.<br />
Key words: Bratu’s problem, PDE, variational method, bicubic splines.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 26<br />
Significance-Based Computing for Reliability and Power<br />
Optimization 1<br />
N¯ ikolaos Bellas and Christos D. Antonopoulos<br />
Department of Electrical and Computer Engineering, University of Thessaly,<br />
Volos, Greece<br />
{nbellas, cda}@uth.gr<br />
Abstract<br />
Manufacturing process variability at low geometries and energy dissipation are the most challenging<br />
problems in the design of future computing systems. Currently, manufacturers go to<br />
great lengths to guarantee fault-free operation of their products by introducing redundancy in<br />
voltage margins, conservative layout rules, and extra protection circuitry. However, such design<br />
redundancy leading to significant energy overheads may not be really required, given that<br />
many modern workloads, such as multimedia, machine learning, visualization, etc. can tolerate<br />
a degree of imprecision in computations and data.<br />
In this talk, I will introduce an approach which seeks to exploit this observation and to relax<br />
reliability requirements for the hardware layer by allowing a controlled degree of imprecision to<br />
be introduced to computations and data. It proposes to research methods that allow the systemand<br />
application-software layers to synergistically characterize the significance of various parts<br />
of the program for the quality of the end result, and their tolerance to faults. Based on this<br />
information, extracted automatically or manually, the system software will steer computations<br />
and data to either low-power, yet unreliable or higher-power and reliable functional and storage<br />
components. In addition, the system will be able to aggressively reduce its power footprint by<br />
opportunistically powering hardware modules below nominal values. Significance-based computing<br />
lays the foundations for not only approaching the theoretical limits of energy reduction<br />
of CMOS technology, but also moving beyond those limits by accepting hardware faults in a<br />
controlled manner.<br />
Key words: Computational significance, Low-power design, Reliable Design.<br />
1 The present research work has been co-financed by the European Union (European Social Fund ESF) and Greek<br />
national funds through the Operational Program ”Education and Lifelong Learning” of the National Strategic Reference<br />
Framework (NSRF) - Research Funding Program: THALIS. Investing in knowledge society through the European<br />
Social Fund.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 27<br />
Quadratic Penalty Methods<br />
for Shape from Shading<br />
Stefania Bellavia, Lapo Governi, Alessandra Papini and Luca Puggelli<br />
Department of Industrial Engineering, University of Florence,<br />
Florence, Italy<br />
stefania.bellavia@unifi.it,lapo.governi@unifi.it,<br />
alessandra.papini@unifi.it,luca.puggelli@unifi.it<br />
Abstract<br />
“Shape from shading” (SFS) denotes the problem of reconstructing a 3D surface, starting from<br />
only one image showing a shaded representation of the surface itself. Minimization techniques<br />
are commonly used for solving the SFS problem, where the functional that must be optimized<br />
is a weighted combination of the brightness functional plus one or more regularization terms.<br />
A critical role in this context is played by the weights used in the functional, which markedly<br />
affect the possibility of obtaining a good reconstruction. However the choice of these weights<br />
in not trivial.<br />
In this work we present a quadratic penalty method where an a-priori choice of the weights<br />
is not needed. In this approach the SFS problem is formulated as a constrained minimization<br />
problem, where the objective function is given by a term accounting for the smoothness of the<br />
reconstructed surface, and the constraints consist of the image irradiance equation (representing<br />
how well the reconstructed surface reproduces the original image) and of an integrability term.<br />
Using a quadratic penalty strategy the original constrained problem is replaced by a sequence<br />
of unconstrained subproblems, which are solved by a Barzilai-Borwein method. The<br />
results obtained on a set of case studies show the effectiveness of the proposed approach.<br />
Key words: Shape from shading, equality constrained minimization, quadratic penalty methods, Barzilai-<br />
Borwein method.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book Conference of Abstracts in Numerical Analysis 2014 (NumAn 2014) 28<br />
September 2-5, 2014<br />
Chania,Greece<br />
Multi-scale hybrid model of cell differentiation propagation<br />
as traveling waves<br />
Mohammed Benmir a , Nikolai Bessonov b , Soumaya Boujena a and Vitaly<br />
Volpert c<br />
a Faculty of Sciences, University Hassan II,<br />
Casablanca 20100, Maroc<br />
b Institute of Problems of Mechanical Engineering, Russian Academy of Sciences,<br />
199178 Saint Petersburg, Russia<br />
c Institut Camille Jordan,UMR 5208 CNRS,University Lyon 1,69622<br />
Villeurbanne,France<br />
mohammed.benmir05@etude.univcasa.ma,boujena@gmail.com,<br />
bessonov@bess.ipme.ru,volpert@math.univ-lyon1.fr,<br />
Abstract<br />
Multi-scale and hybrid models are well adapted for the description of complex physiological<br />
processes. They represent an interesting class of models whose properties are not yet sufficiently<br />
well studied. In particular, they can show unusual nonlinear dynamics. In this work<br />
we will study propagation of reaction-diffusion waves in the medium composed of unmovable<br />
cells. Dynamics of cell population is determined by complex intracellular and extracellular regulations.<br />
If cell differentiation is initiated locally in space in the population of undifferentiated<br />
cells, it propagates as a travelling wave converting undifferentiated cells into differentiated ones.<br />
We suggest a model of this process which takes into account intracellular regulation, extracellular<br />
regulation and different cell types. They include undifferentiated cells and two types of<br />
differentiated cells. When a cell differentiates, its choice between two types of differentiated<br />
cells is determined by the concentrations of intracellular proteins. Differentiated cells can either<br />
stimulate differentiation into their own cell lineage or into another cell lineage. Periodic spatial<br />
patterns can emerge behind the propagating wave.<br />
Key words: multi-scale, hybrid, models, traveling waves, extracellular, intracellular, cells, differentiation.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 29<br />
Conference in Numerical Analysis 2014 (NumAn 2014)<br />
September 2-5, 2014<br />
Chania,Greece<br />
Preconditioned derivative-free globally convergent<br />
Newton-GMRES methods for large sparse nonlinear systems<br />
Michele Benzi a , Iain S. Duff b and Xue-Ping Guo c<br />
a Department of Mathematics and Computer Science, Emory University,<br />
Atlanta, GA 30322, USA<br />
b RAL, Oxfordshire, England. CERFACS, 42 av. Gaspard Coriolis, 31057,<br />
Toulouse, Cedex 1, France.<br />
a Department of Mathematics, East China Normal University,<br />
Shanghai, 200241, P. R. China.<br />
iain.duff@stfc.ac.uk,benzi@mathcs.emory.edu,xpguo@math.ecnu.edu.cn<br />
Abstract<br />
Jocobian-free Newton GMRES(m) methods (JFNG) solve systems of nonlinear equations without<br />
computing matrix vector product and forming derivatives of functions. In this paper, we<br />
introduce the derivative-free HSS preconditioner in JFNG methods and obtain preconditioned<br />
derivative-free globally convergent Newton-GMRES methods (PDFNG) for solving large s-<br />
parse system of nonlinear equations. The convergence is also given. Finally, numerical tests are<br />
shown to illustrate the efficiency of PDFNG methods.<br />
Key words: preconditioner, derivative-free, system of nonlinear equations.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 30<br />
Using an Euler type transform for accelerating convergence of<br />
series<br />
Faton Berisha a , Murat Sadiku b and Nimete Berisha c<br />
a Department of Mathematics, University of Prishtina, Prishtina, Kosovo<br />
b Faculty of Economics and Business, South East European University,<br />
Tetovo, FYROM<br />
c Faculty of Economics, University of Prishtina, Prishtina, Kosovo<br />
faton.berisha@uni-pr.edu,m.sadiku@seeu.edu.mk<br />
nimete.berisha@gmail.com<br />
Abstract<br />
Let us use the following operator of generalised difference, linear on a set of sequences:<br />
∆ 1 r 1<br />
(a n ) = ∆ r1<br />
(a n ) = a n+1 − r 1 a n ,<br />
∆ m+1<br />
r 1 r 2 ...r m+1<br />
(a n ) = ∆ 1 r m+1<br />
(∆ m r 1 r 2 ...r m<br />
(a n )) (m = 1, 2, . . .),<br />
where {r m } ∞ m=1 is a given sequence of real numbers.<br />
In the present paper, we give a property of the operator ∆ k r 1 r 2 ...r k<br />
when applied on an alternating<br />
sequence {(−1) n a n } ∞ n=0 . Then, we use this property in order to establish modified Euler<br />
transforms for alternating number, power and trigonometric series. We present algorithm analysis<br />
for all cases of computing the n-th partial sum of transformed series by using the operator<br />
of generalised difference of order p, and prove that its order of complexity is O(p 2 n), i.e. that<br />
its complexity is linear with respect to the number of terms computed. Finally, we give two examples<br />
illustrating, for instance, that in order to calculate the approximate sum of a given series<br />
with an error not greater than 10 −6 we must compute the sum of the first 19 terms. To obtain<br />
this accuracy for the classical Euler transform we need 18 summands. Applying the modified<br />
transform considered in the paper, the same accuracy is obtained by computing the sum of the<br />
first 11 terms for p = 1, 7 terms for p = 2, and 4 terms for p = 3.<br />
Key words: Accelerating convergence of series, Euler transform.<br />
References<br />
[1] F. M. Berisha and M. H. Filipović, On some transforms of trigonometric series, Publ. Inst.<br />
Math. (Beograd) (N.S.) 61(75) (1997), 53–60. MR 1472937 (98g:42012)<br />
[2] I. Ž. Milovanović, M. A. Kovačević, S. D. Cvejić, and J. Klippert, A modification of the Euler-<br />
Abel transform for convergent series, J. Natur. Sci. Math. 29 (1989), no. 1, 1–9. MR 91g:40008<br />
[3] G. A. Sorokin, O nekotorykh preobrazovanyakh ryadov, Izv. Vyssh. Uchebn. Zaved. Mat.<br />
(1984), no. 11, 34–40, 83. MR 86f:40003<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 31<br />
Efficient Solution of the Two-Dimensional<br />
Shallow-Water Equations using GPUs<br />
D.E. Bobolakis ∗ , A.I. Delis and E.N. Mathioudakis<br />
Applied Mathematics and Computers Laboratory<br />
Technical University of Crete, Chania, Greece<br />
∗ dbobolakis@isc.tuc.gr<br />
Abstract<br />
A parallel algorithm for solving two-dimensional shallow-flow problems that takes advantage<br />
of modern computing accelerators such as graphics processing units (GPUs) is presented. The<br />
high-resolution Godunov-type explicit scheme is used, with Roe’s approximate Riemman solver<br />
to create a numerical method suitable for different types of flood simulation. The performance<br />
of a real-world dam collapse test-case using a massive grid with more than 1.9 million cells<br />
is investigated. The application is developed in double precision Fortran code using the OpenACC<br />
standard and the realization of the algorithm takes place on a HP SL390s G7 multicore<br />
system with Tesla M2070 GPUs and PGI’s compilers. Numerical results reveal an impressive<br />
agreement with the post-event survey. Solver’s parallel algorithm is designed to perform<br />
the total computation on the GPU, significantly expediting simulations when compared to the<br />
conventional Central Processing Unit (CPU) - Open Multi-Processing (OpenMP) performance.<br />
Scientific computations for GPU technology at shallow-water simulations yield to performance<br />
acceleration, when compared to classical parallel CPU - OpenMP realizations.<br />
Key words: Shallow-Water equations, Roe’s solver, CPU-GPU computations, OpenMP, OpenACC.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 32<br />
Conference in Numerical Analysis 2014 (NumAn 2014)<br />
September 2-5, 2014<br />
Chania,Greece<br />
Complex Statistics and Diffusion in Nonlinear Disordered<br />
Particle Chains<br />
Tassos Bountis 1 Christos Antonopoulos 2 and Haris Skokos 3,4<br />
1 Center of Research and Applications of Nonlinear Systems (CRANS),<br />
Department of Mathematics, University of Patras, 26500, Greece<br />
2 Institute for Complex Systems and Mathematical Biology<br />
University of Aberdeen, AB24 3UE Aberdeen, UK<br />
3 Department of Mathematics and Applied Mathematics<br />
University of Cape Town, Rondebosch, 7701, South Africa<br />
4 Department of Physics, University of Thessaloniki, 54124 Thessaloniki, Greece<br />
tassos50@otenet.gr, chris.antonopoulos@abdn.ac.uk, haris.skokos@uct.ac.za<br />
Abstract<br />
We perform a numerical study to investigate dynamically and statistically diffusive motion in a<br />
Klein-Gordon particle chain in the presence of disorder. In particular, we examine a low energy<br />
(subdiffusive) and a higher energy (self-trapping) case and verify that subdiffusive spreading is<br />
always observed. We then carry out a statistical analysis of the motion in both cases in the sense<br />
of the Central Limit Theorem and present evidence of different chaos behaviors, for various<br />
groups of particles. Integrating the equations of motion for times as long as 10 9 , our probability<br />
distribution functions always tend to Gaussians and show that the dynamics does not relax onto<br />
a quasi-periodic KAM torus and that diffusion continues to spread chaotically for arbitrarily<br />
long times. We also discuss some particular features that concern our numerical computations.<br />
Key words: Complex Statistics, Hamiltonian Systems, Klein-Gordon, q-Gaussians, Tsallis Entropy, Diffusive<br />
Motion, Weak and Strong Chaos<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book Conference of Abstracts in Numerical Analysis 2014 (NumAn 2014) 33<br />
September 2-5, 2014<br />
Chania, Greece<br />
A modified predictor-corrector method for the generalized<br />
Burgers–Huxley equation<br />
A. G. Bratsos<br />
Department of Naval Architecture,<br />
Technological Educational Institution (T.E.I.) of Athens,<br />
Athens, Greece.<br />
e-mail: bratsos@teiath.gr<br />
URL: http://users.teiath.gr/bratsos/<br />
Abstract<br />
A third-order in time modified predictor-corrector method is proposed for the numerical solution<br />
of the generalized Burgers–Huxley (BgH) equation, which is given by<br />
(<br />
u t + αu δ u x − u xx = βu 1 − u δ) ( )<br />
u δ − γ ; 0 ≤ x ≤ 1 , t > 0, (1)<br />
u = u (x, t) is a sufficiently differentiable function, with α a real parameter, β ≥ 0, γ ∈ (0, 1),<br />
δ > 0, initial condition u (x, 0) = f(x); x ∈ [0, 1] and boundary conditions u x | x=0, 1<br />
= g(t);<br />
t > 0. Eq. (1) is the modified Burgers equation for β = 0, is the Huxley equation for α = 0,<br />
δ = 1 and is the Fitzhugh-Nagoma equation for α = 0.<br />
Many researchers have used various methods to solve the BgH equation. A theoretical study<br />
of the BgH equation was found in [1], while as far as the numerical methods among others in<br />
[2] etc.<br />
The main aim of this paper is to solve the BgH equation explicitly with a direct method. To<br />
this attempt, the solution of the resulting nonlinear system is given by expressing the unknown<br />
vector component wise and updating each component as soon as its value becomes available.<br />
This process, which is known as a modified predictor-corrector method (see, e.g., [3] and references<br />
therein), opposite to the iterative classical predictor-corrector one is always explicit and<br />
is applied once, has also been examined successfully with various other approximations in time<br />
giving an improvement in the accuracy over the classical method.<br />
References<br />
[1] X. Deng, Travelling wave solutions for the generalized Burgers-Huxley equation, Appl<br />
Math Comput 204 (2008) 733-737.<br />
[2] M. Javidi, A numerical solution of the generalized Burgers-Huxley equation by spectral<br />
collocation method, Appl Math Comput 178(2006) 338–344.<br />
[3] A. G. Bratsos, An improved second-order numerical method for the generalized Burgers-<br />
Fisher equation, ANZIAM Journal 54(3) (2013) 181–199.<br />
Key words: Burgers–Huxley; Modified Predictor-Corrector; Finite-difference method<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book Conference of Abstracts in Numerical Analysis 2014 (NumAn 2014) 34<br />
September 2-5, 2014<br />
Chania,Greece<br />
Structured Strong Linearizations obtained from Fiedler Pencils<br />
with Repetition<br />
Maria Isabel Bueno a , K. Curlett b , and Susana Furtado c<br />
a Department of Mathematics and College of Creative Studies, University of<br />
California<br />
Santa Barbara, CA, USA<br />
b College of Creative Studies, University of California<br />
Santa Barbara CA, USA<br />
c Faculdade de Economia da Universidade do Porto and Centro de Estruturas<br />
Lineares e Combinatrias da Universidade de Lisboa<br />
Portugal<br />
mbueno@math.ucsb.edu,curlett@umail.ucsb.edu,sbf@fep.up.pt<br />
Abstract<br />
Let P (λ) be a matrix polynomial of degree k ≥ 2 whose coefficients are n × n matrices with<br />
entries in a field F. A matrix pencil L(λ) = λL 1 −L 0 , with L 1 , L 0 ∈ M kn (F), is a linearization<br />
of P (λ) if there exist two unimodular matrix polynomials (i.e. matrix polynomials with constant<br />
nonzero determinant), U(λ) and V (λ), such that<br />
U(λ)L(λ)V (λ) =<br />
[<br />
I(k−1)n 0<br />
0 P (λ)<br />
Beside other applications, linearizations of matrix polynomials are used in the study of the polynomial<br />
eigenvalue problem. For each matrix polynomial P (λ), many different linearizations can<br />
be constructed but, in practice, those sharing the structure of P (λ) are the most convenient from<br />
the theoretical and computational point of view, since the structure of P (λ) often implies some<br />
symmetries in its spectrum.<br />
In this talk we present nk × nk matrix pencils obtained from the family of Fiedler pencils<br />
with repetition, introduced by S. Vologiannidis and E. N. Antoniou (2011), preserving the<br />
structure of P (λ), when P (λ) is symmetric, skew-symmetric or T-alternating. Under certain<br />
conditions, these pencils are strong linearizations of P (λ). These linearizations are companion<br />
forms in the sense that, if their coefficients are viewed as k-by-k block matrices, each n × n<br />
block is either 0 n , ±I n , or ±A i , where A i , i = 0, . . . , k, are the coefficients of P (λ).<br />
]<br />
.<br />
Key words: Structured linearization, Fiedler pencils with repetition, matrix polynomial, companion form,<br />
polynomial eigenvalue problem.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 35<br />
Unified Semi-Analytical, Semi-Numerical Approach<br />
to Stability Analysis of Nonparallel Unsteady Flows<br />
Georgy I. Burde a , Ildar Sh. Nasibullayev b and Alexander Zhalij c<br />
a Jacob Blaustein Institutes for Desert Research, Ben-Gurion University,<br />
Sede-Boker Campus, Israel<br />
b Institute of Mechanics, Russian Academy of Sciences,<br />
Ufa, Russia<br />
c Institute of Mathematics of the Academy of Sciences of Ukraine,<br />
Kyiv, Ukraine<br />
georg@bgu.ac.il,ildar@bgu.ac.il,zhaliy@imath.kiev.ua<br />
Abstract<br />
Stability of some unsteady nonparallel three-dimensional flows (exact solutions of the Navier-<br />
Stokes equations) is studied via separation of variables using a semi-analytical, semi-numerical<br />
approach. In this approach, a new coordinate system, which allows solution with separated variables,<br />
is defined together with the solution form. This part of the method involves complicated<br />
analytical calculations which can be implemented only using symbolic manipulating programs.<br />
The resulting eigenvalue problems are solved numerically with the help of the spectral collocation<br />
method based on Chebyshev polynomials. Such computational synthesis of analytical and<br />
numerical calculations allows to extend the physically significant concept of normal modes to<br />
the case of non-steady nonparallel basic flows for which this concept in its traditional form is not<br />
applicable at all. This may provide a basis for a well-grounded discussion of some problematic<br />
points of hydrodynamic stability analysis and a useful test for methods used in the hydrodynamic<br />
stability theory, in general. The basic flows whose stability is studied in the paper are<br />
of significant interest for fluid dynamics theory and have received considerable attention in the<br />
literature due to their relevance in a number of engineering applications.<br />
Key words: Linear stability, Nonparallel unsteady flows, Separation of variables, Direct method.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book Conference of Abstracts in Numerical Analysis 2014 (NumAn 2014) 36<br />
September 2-5, 2014<br />
Chania,Greece<br />
A recursive multilevel approximate inverse-based<br />
preconditioner for solving general linear systems<br />
Yi-Ming Bu a,b and Bruno Carpentieri b<br />
a School of Mathematical Sciences, University of Electronic Science and<br />
Technology of China, Chengdu, Sichuan 611731, China<br />
b Institute of Mathematics and Computer Science, University of Groningen,<br />
Groningen, The Netherlands<br />
y.bu@rug.nl,b.carpentieri@gmail.com<br />
Abstract<br />
We consider multilevel approximate inverse-based preconditioning techniques for solving systems<br />
of linear equations<br />
Ax = b (1)<br />
where A ∈ R n×n is a typically large, sparse, nonsymmetric matrix arising from the discretization<br />
of partial differential equations. Approximate inverse methods directly approximate A −1<br />
as the product of sparse matrices, so that the preconditioning operation reduces to forming one<br />
(or more) sparse matrix-vector product. Due to their inherent parallelism and numerical robustness,<br />
this class of methods is receiving renewed consideration for iterative solutions of large<br />
linear systems on emerging massively parallel computer systems. In practice, however, some<br />
questions need to be addressed. First of all the computed preconditioner could be singular. In<br />
the second place, these techniques usually require more CPU-time to compute the preconditioner<br />
than ILU-type methods. Third, the computation of the sparsity pattern of the approximate<br />
inverse can be problematic, as the inverse of a general sparse matrix is typically fairly dense.<br />
This leads to prohibitive computational and storage costs.<br />
We present an algebraic recursive multilevel inverse-based preconditioner, based on a distributed<br />
Schur complement formulation, that attempts to remedy these problems. The proposed<br />
solver uses recursive combinatorial algorithms to preprocess the structure of A and to produce<br />
a suitable permutation of the linear system that can maximize sparsity in the approximate inverse.<br />
An efficient tree-based recursive data structure is generated to compute and apply the<br />
approximate inverse fast and efficiently. We report on numerical experiments on matrix problems<br />
arising in different application areas to illustrate the potential of the proposed solver to<br />
reduce significantly the number of iterations of Krylov methods at low memory costs, also compared<br />
to other sparse approximate inverses and multilevel Schur-complement based incomplete<br />
LU factorization methods and software. Finally, we discuss block generalizations of our method<br />
that can exploit available block structure in the matrix to maximize computational efficiency.<br />
Key words: Krylov subspace methods, Approximate inverse preconditioners, combinatorial algorithms.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 37<br />
Conference in Numerical Analysis 2014 (NumAn 2014)<br />
September 2-5, 2014<br />
Chania,Greece<br />
A local anisotropic adaptive algorithm<br />
to solve time-dependent dominated convection problems<br />
Jaime Carpio, Juan Luis Prieto<br />
Departamento de Ingeniería Energética y Fluidomecánica<br />
Universidad Politécnica de Madrid, Spain<br />
jaime.carpio@upm.es,juanluis.prieto@upm.es<br />
Abstract<br />
In this work we present a local, anisotropic adaptive algorithm useful to solve scientific and<br />
engineering time-dependent problems encompassing multiple scales. The algorithm is derived<br />
in the context of semi-Lagrangian schemes within a finite element framework, being suitable<br />
for higher-order finite elements. Convection-dominated equations, like those present in Fluid<br />
Dymanics, are ideal to employ anisotropic refinement due to the ‘directional features’ present<br />
in flows such jets, mixing layers, vortices... The size, shape and orientation of the anisotropic<br />
elements which define the optimal triangulation are provided by a metric tensor based on an a<br />
posteriori error indicator of the local or truncated error incurred at each time step.<br />
We illustrate the good performance of the algorithm with a convection-dominated problem<br />
taken from the Combustion field of knowledge. Simulation in 2D and 3D is also considered to<br />
address the interaction between a diffusion flame and a vortex generated by a turbulent flow.<br />
Finally, we include a comparison with actual experimental data.<br />
Key words: Semi-Lagrangian schemes, finite element method, a posteriori error indicator, local anisotropic<br />
refinement, combustion problems.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 38<br />
Conference in Numerical Analysis 2014 (NumAn 2014)<br />
September 2-5, 2014<br />
Chania,Greece<br />
CPU-GPU computations for MultiGrid techniques coupled with<br />
Fourth-Order Compact Discretizations for Isotropic and<br />
Anisotropic Poisson problems<br />
N. E. Charalampaki and E. N. Mathioudakis<br />
Applied Mathematics and Computers Laboratory<br />
Technical University of Crete, Chania, Greece<br />
ncharalabaki@isc.tuc.gr manolis@amcl.tuc.gr<br />
Abstract<br />
A CPU-GPU parallel algorithm for a fourth-order compact finite difference scheme with unequal<br />
mesh size in different coordinate directions, is designed to discretize a two dimensional<br />
isotropic or anisotropic Poisson equation in a rectangular domain. A multigrid technique with<br />
partial semi-coarsening strategy is used to iteratively solve the sparse linear system derived.<br />
Numbering the unknowns and equations according to the line red-black fashion, the coefficient<br />
matrix obtains a block structure suitable for parallel computations. These blocks consist of<br />
Toeplitz matrices with known inverses, allowing the efficient solution of inner linear systems<br />
on parallel computing environments with accelerators. The realization of the algorithm takes<br />
place on a HP SL390s G7 multicore system with Tesla M2070 GPUs and the application is<br />
developed in double precision Fortran code using the OpenACC standard with PGI’s compilers.<br />
The performance investigation reveals that the solution of fine discretization problems can be<br />
accelerated, although multigrid techniques usually yield poor efficiency on parallel computing<br />
architectures due to solution approximations of decreased size problems.<br />
Key words: Multigrid techniques, Compact finite difference schemes, GPU computations, OpenACC<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book Conference of Abstracts in Numerical Analysis 2014 (NumAn 2014) 39<br />
September 2-5, 2014<br />
Chania,Greece<br />
Nonlinear Model Reduction with Localized Basis for<br />
Two-Phase Miscible Flow in Porous Media<br />
Saifon Chaturantabut a<br />
a Department of Mathematics and Statistics, Thammasat University,<br />
Pathumthani, Thailand<br />
saifon@mathstat.sci.tu.ac.th<br />
Abstract<br />
This work presents an application of a model reduction approach to substantially decrease<br />
the simulation time for two-phase nonlinear miscible flow in porous media. Since this type of<br />
flow often contains detailed features in fingering displacement, the approach considered here<br />
employs the localized basis sets from Proper Orthogonal Decomposition (POD) in the Galerkin<br />
projection procedure to accurately capture the important dynamics of the system. Discrete Empirical<br />
Interpolation Method (DEIM) with corresponding localized basis is then applied to efficiently<br />
compute the projected nonlinear terms in the POD reduced system. The related theoretical<br />
aspect of this approach is discussed. This work also proposes an adaptive scheme based<br />
on an error estimate indicator to choose a subdivision of the localized basis, together with an<br />
efficient procedure for updating each localized basis during the online simulation. This localized<br />
model reduction approach is shown to construct a system that can accurately capture the<br />
characteristics of the original miscible flow in the 2D finite-difference discretized setting, with<br />
the dimension reduced by a factor of O(10 2 ) and the CPU time decreased by a factor of O(10 3 ).<br />
Key words: Nonlinear Model Reduction, Proper Orthogonal Decomposition, Empirical Interpolation<br />
Methods, Nonlinear Partial Differential Equations, Miscible Viscous Fingering in Porous Media.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 40<br />
On positivity preservation for finite element based methods for<br />
the heat equation<br />
Panagiotis Chatzipantelidis a<br />
a Department of Mathematics and Applied Mathematics, University of Crete,<br />
Greece<br />
chatzipa@math.uoc.gr<br />
Abstract<br />
We consider the model initial–boundary value problem<br />
u t − ∆u = 0, in Ω, u = 0, on ∂Ω, for t ≥ 0, u(0) = v, in Ω, (1)<br />
where Ω is a bounded convex polygonal domain in R 2 . By the maximum-principle, we have<br />
v ≥ 0 in Ω implies u(t, ·) ≥ 0 in Ω, for t ≥ 0. (2)<br />
Our purpose is to discuss analogues of this property for some finite element methods, based on<br />
piecewise linear finite elements, including, in particular, the Standard Galerkin (SG) method,<br />
the Lumped Mass (LM) method, and the Finite Volume Element (FVE) method.<br />
We consider the analog semidiscrete problem of (1), where we discretize space using either<br />
the SG, LM or FVE method. It is known that for the semidiscrete SG the analog of (2) does<br />
not hold for all t ≥ 0. However, in the case of the LM method, this holds if and only if the<br />
triangulation is of Delaunay type. For the FVE method we will show here that the situation is<br />
the same as for the SG method.<br />
However, when the solution is not positive for all t > 0, it may be positive for all t sufficiently<br />
large. We shall study this and approximate a corresponding minimum t 0 such that for<br />
all t > t 0 the solution of the semidiscrete problem is positive. Also we consider similar results<br />
for the corresponding fully discrete problems, when we discretize time with the backward Euler<br />
method. Finally we provide numerical results in 1 and 2 dimensions.<br />
Key words: positivity, finite element method, finite volume method, lumped mass method, Delaunay<br />
triangulation.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 41<br />
Conference in Numerical Analysis 2014 (NumAn 2014)<br />
September 2-5, 2014<br />
Chania,Greece<br />
Block Hybrid Numerical Integrators for the<br />
Solution of Stiff Equations<br />
J. P. Chollom and G.M.Kumleng<br />
Department of Mathematics<br />
University of Jos<br />
Jos, Nigeria<br />
chollomp@gmail.com<br />
Abstract<br />
Stiff equations occur in a wide variety of applications including springs, damping systems<br />
and chemical reactions. The stiffness occurs due to the great difference among the reaction<br />
constants.Due to step size restriction, it becomes necessary to search for numerical methods<br />
with large regions of absolute stability . In this paper, new block hybrid linear integrators of the<br />
Adams Moulton class for the solution of stiff systems are constructed. This is achieved through<br />
the multistep collocation approach which yielded discrete schemes used simultaneously in block<br />
form as block integrators. This approach eliminates the use of starting values and overlap of<br />
pieces of solutions. The stability analysis of the new methods carried shows that they A-stable,<br />
a property desirable of any numerical method suitable for the solution of stiff systems. The<br />
new methods are tested on circular reaction equations, conserved systems, Robertson problem<br />
and a chemical reaction problem. The results shows that the new methods are efficient as they<br />
compare favorably with the state of the art Mat lab ode solver, ode23s.<br />
Key words: Block Linear integrators ,multi step collocation, A-stability, stiff systems, chemical reactions.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 42<br />
Efficient GPU pricing of interest rate derivatives:<br />
PDE formulation and ADI methods<br />
Duy-Minh Dang a , Christina Christara b and Kenneth Jackson b<br />
a Department of Computer Science, University of Waterloo,<br />
Waterloo, Ontario, Canada<br />
b Department of Computer Science, University of Toronto,<br />
Toronto, Ontario, Canada<br />
dm2dang@uwaterloo.ca,ccc@cs.toronto.edu,krj@cs.toronto.edu<br />
Abstract<br />
We study the parallel implementation on a Graphics Processing Unit (GPU) of Alternating Direction<br />
Implicit (ADI) time discretization methods for solving three-dimensional time-dependent<br />
parabolic Partial Differential Equations (PDEs) with mixed spatial derivatives, and investigate<br />
the performance of the resulting parallel methods in pricing foreign exchange (FX) interest rate<br />
hybrids, namely Power Reverse Dual Currency (PRDC) swaps with various exotic features.<br />
A model for pricing PRDC swaps involves three stochastic factors, namely the FX rate, and<br />
the interest rates in the two currencies. By certain financial and mathematical arguments, the resulting<br />
model is a three-dimensional in space parabolic PDE, which includes all cross-derivative<br />
terms and, assuming a local volatility model, has variable coefficients. We use standard centered<br />
Finite Differences (FDs) for the space discretization and the Hundsdorfer-Verwer (HV) ADI<br />
method for the timestepping. We discuss the parallelization on a GPU of the computational<br />
requirements of the ADI method, such as the multiple tridiagonal solutions along each of the<br />
problem’s spatial dimensions and the matrix-vector products, with special attention to coalesced<br />
memory access.<br />
Furthermore, we consider the highly popular Target Redemption (TARN) feature for PRDC<br />
swaps, which adds path-dependency and, therefore, complexity to the PDE problem. The pricing<br />
of the FX-TARN PRDC swap is handled by breaking it down into several independent<br />
pricing subproblems over each period of the tenor structure. Each of the subproblems is solved<br />
on an individual GPU, with communication at the end of each period of the tenor structure taken<br />
care by MPI.<br />
We present numerical experiments that indicate considerable speedup, when comparing the<br />
CPU versus the GPU implementations, as well as the implementations on one versus multiple<br />
GPUs.<br />
Key words: Power Reverse Dual Currency (PRDC) swaps, local volatility, Target Redemption (TARN),<br />
three-dimensional parabolic PDE, Hundsdorfer-Verwer ADI method, Graphics Processing Unit (GPU), MPI.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book Conference of Abstracts in Numerical Analysis 2014 (NumAn 2014) 43<br />
September 2-5, 2014<br />
Chania, Greece<br />
Dynamical and statistical behavior of the Fermi-Pasta-Ulam<br />
model with long-range interactions<br />
H. Christodoulidi, L. Cirto, T. Bountis and C. Tsallis<br />
Center for Research and Applications of Nonlinear Systems (CRANS),<br />
Department of Mathematics, University of Patras, GR–26500, Patras, Greece<br />
hchrist@master.math.upatras.gr<br />
Abstract<br />
In this talk we describe the results of a recent study on a long-range-interaction generalisation of<br />
the one-dimensional Fermi-Pasta-Ulam (FPU) β− model. In particular, we have used a coupling<br />
constant of the quartic interactions that decays as 1/r α and controls the range of interaction<br />
(α ≥ 0). We demonstrate that: (i) For α ≥ 1 the maximal Lyapunov exponent remains finite<br />
and positive for increasing number of oscillators N whereas, for 0 ≤ α < 1, it asymptotically<br />
decreases as N −κ(α) ; (ii) The distribution of time-averaged velocities is Maxwellian for α large<br />
enough, whereas it is well approached by a q-Gaussian, with the index q(α) monotonically<br />
decreasing from about 1.5 to 1 (Gaussian) when α increases from zero to close to one. To<br />
achieve these results for very large numbers of particles and very long integration times we made<br />
use of a number of numerical methods and strategies which will be discussed in the present talk.<br />
Key words:<br />
q–Gaussian Distributions, Lyapunov Exponents, Hamiltonian Lattices, Long Range Dynamics, Symplectic<br />
Integrators<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 44<br />
Conference in Numerical Analysis 2014 (NumAn 2014)<br />
September 2-5, 2014<br />
Chania,Greece<br />
Two numerical implementations of the Fokas method<br />
for elliptic equations in a polygon<br />
Kevin Crooks a<br />
a DAMTP, University of Cambridge,<br />
United Kingdom<br />
k.m.crooks@maths.cam.ac.uk<br />
Abstract<br />
We consider the Dirichlet boundary value problem for the Helmholz and modified-Helmholz<br />
equations in a convex polygonal domain. Recent work has used the Fokas method to derive a<br />
Dirichlet to Neumann map for Laplace’s equation on the polygon: given Dirichlet data this map<br />
recovers our unknown Neumann data. These data are coupled by an integral equation known as<br />
the global relation. By reformulating the global relation as a linear operator equation of the form<br />
T Φ = Ψ, it was shown that T is a semi-Fredholm operator between Banach spaces. Analogous<br />
results may be obtained for the class of Helmholz equations by considering the resulting linear<br />
operators as perturbations, T β , of T .<br />
We analyse two numerical implementations: a Galerkin method and a pointwise method.<br />
For β not an eigenvalue for a domain, we use these approaches to solve the Dirichlet to Neumann<br />
map. Secondly, we demonstrate their use to search for eigenvalues, with the pointwise method<br />
seen to be particularly effective. Finally we discuss the numerical accuracy and difficulty of<br />
these implementations.<br />
Key words: Fokas Method, Helmholz, Boundary Value Problems, Numerical Approach.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 45<br />
Solving Wiener-Hopf problems without kernel factorisation<br />
Darren G. Crowdy and Elena Luca<br />
Department of Mathematics, Imperial College London, UK<br />
d.crowdy@imperial.ac.uk, el1710@imperial.ac.uk<br />
Abstract<br />
We present a new approach for solving Wiener-Hopf problems by showing its implementation<br />
in two typical examples from fluid mechanics. The new method adapts various mathematical<br />
ideas underlying the so-called unified transform method due to A.S. Fokas and collaborators in<br />
recent years. The method has the key advantage of avoiding what is usually the most challenging<br />
part of the usual Wiener-Hopf approach: the factorisation of kernel functions into sectionally<br />
analytic functions. We show that the new approach leads naturally to fast and accurate schemes<br />
for evaluation of the solutions.<br />
Key words: Fokas transform method, Wiener-Hopf, complex analysis<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 46<br />
Numerical evaluation of hypersingular integrals on the semiaxis<br />
Maria Carmela De Bonis a and Donatella Occorsio a<br />
a Department of Mathematics, Computer Science and Economics,<br />
University of Basilicata,<br />
Potenza, Italy<br />
mariacarmela.debonis@unibas.it, donatella.occorsio@unibas.it<br />
Abstract<br />
We consider hypersingular integrals of the following type<br />
H p (g, t) = =<br />
∫ +∞<br />
0<br />
g(x)<br />
dx,<br />
(x − t) p+1<br />
where 0 < t < +∞, p ≥ 1 is an integer and the function g behaves like x α for x → 0 and has<br />
an exponential or algebraic decay for x → +∞, i.e., it can be written in one of the following<br />
forms<br />
• g(x) = f(x)w α (x), w α (x) = x α e −x , α ≥ 0;<br />
• g(x) = f(x)u α,β (x), u α,β (x) = xα<br />
(1+x) β , α, β ≥ 0.<br />
They appear in different contexts and, in particular, in some problems of mathematical theory<br />
of elasticity (see [?]) and in hypersingular integral equations coming from Neumann twodimensional<br />
elliptic problems defined on a half-plane by using a Petrov-Galerkin infinite BEM<br />
approach as discretization technique (see [?]). To our knowledge, the literature dealing with the<br />
approximation of hypersingular integrals on unbounded intervals is very poor.<br />
In this talk we propose some numerical procedures for the pointwise approximation of the<br />
integrals H p (g, t) that are based on Gaussian-type quadrature formulas. We prove that such<br />
procedures are stable and convergent in suitable weighted uniform spaces and, for each of them,<br />
we give error estimates. Finally, we show that the theoretical results are confirmed by the<br />
numerical tests.<br />
Key words: hypersingular integrals, Gaussian-type quadrature rules.<br />
References<br />
[1] A. Aimi, M. Diligenti, Numerical integration schemes for hypersingular integrals on the real<br />
line, Communications to SIMAI Congress, doi: 10.1685/CSC06003, ISSN 1827-9015, Vol. 2<br />
(2007).<br />
[2] A.I. Kalandya, Mathematical methods of two-dimensional elasticity, Mir Publisher, Moskow,<br />
1975.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 47<br />
Conference in Numerical Analysis 2014 (NumAn 2014)<br />
September 2-5, 2014<br />
Chania, Greece<br />
A Characterization Theorem for the Discrete Best L 1 Monotonic<br />
Approximation Problem<br />
Ioannis C. Demetriou<br />
Department of Economics, University of Athens,<br />
Athens, Greece<br />
demetri@econ.uoa.gr<br />
Abstract<br />
Let n measurements of a real valued function of one variable be given. If the function is monotonic<br />
but the data have lost monotonicity due to measuring errors, then the least sum of the<br />
moduli of the errors that provides nonnegative first divided differences may be required. A<br />
characterization theorem is obtained for the solution of this problem in terms of Lagrange multipliers.<br />
Key words: first divided differences, monotonic approximation, L 1 -norm.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book Conference of Abstracts in Numerical Analysis 2014 (NumAn 2014) 48<br />
September 2-5, 2014<br />
Chania,Greece<br />
Numerical techniques for sliding motion<br />
in Filippov discontinuous systems<br />
Luca Dieci a , Cinzia Elia b and Luciano Lopez a<br />
a School of Mathematics, Georgia Tech Institute, Atlanta, GA 30332-0160, USA<br />
b Department of Mathematics, University of Bari, 70125 Bari, Italy<br />
dieci@math.gatech.edu,cinzia.elia@uniba.it,luciano.lopez@uniba.it<br />
Abstract<br />
In this talk we present numerical techniques to approximate the solution of a discontinuous<br />
differential system of Filippov type during sliding motion. Namely, for a given surface Σ defined<br />
as the 0-set of a smooth scalar function h: Σ = {x : h(x) = 0}, we have the problem<br />
x ′ = f(x), where f(x) = f 1 (x) when h(x) < 0, and f(x) = f 2 (x) when h(x) > 0. Further,<br />
(∇h) T f 1 > 0 and (∇h) T f 2 < 0, on and near Σ, so that trajectories are attracted to Σ and must<br />
remain there.<br />
So the main steps of a numerical procedure for solving such kind of problems are: reach Σ<br />
at x 0 (by an event location procedure) and, starting with x 0 ∈ Σ, solve the differential system<br />
x ′ = (1 − α)f 1 + αf 2 , where α has to be found so that x ′ is tangent to Σ: sliding motion.<br />
Here we propose an event location procedure, which determines the event point on Σ in a<br />
finite number of steps, and compare different numerical procedures to integrate our piecewise<br />
differential system during the sliding motion.<br />
It is well understood that when one integrates the differential system on Σ typically the numerical<br />
solution does not remain on Σ. Thus, the main feature of an effective numerical method<br />
is to require that the the numerical solution also remains on Σ. To achieve this, projection<br />
techniques can be used.<br />
We will consider the following three different flavors of projection techniques.<br />
(i) Classical projection of the numerical solution obtained by any method, say an explicit<br />
method. In particular, we will discus two ways to perform this projection.<br />
(ii) A change of variable technique, in case one can write explicitly h(x) = 0 ⇔ x k =<br />
g(x 1 , . . . , x k−1 , x k+1 , . . . , x n ).<br />
(iii) Reverse projection technique, whereby rather than starting with x 0 on Σ we seek a perturbed<br />
initial condition ˜x 0 ≈ x 0 such that the value of x 1 computed by one step of a<br />
numerical method starting at ˜x 0 is on Σ. Even for this technique, we will present different<br />
implementations.<br />
We will compare the above projection techniques on several examples.<br />
Key words: Discontinuous ODEs, Filippov systems, sliding motion, one-step methods, projection methods.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 49<br />
A new filled function method applied to unconstrained<br />
global optimization<br />
T. M. El-Gindy, M. S. Salim and A. I. Ahmed<br />
Department of Mathematics, Faculty of Science<br />
Assiut University<br />
Assiut, Egypt<br />
Taha.elgindy@gmail.com, salim@yahoo.com, ibrahim@yahoo.com<br />
Abstract<br />
The filled function method is an efficient approach to find the global minimizer of multidimensional functions.<br />
A number of filled functions were proposed recently, most of which have one or two adjustable<br />
parameters. The idea behind the filled function methods is to construct an auxiliary function that allows<br />
us to escape from a given local minimum of the original objective function. It consists of two phases: local<br />
minimization and filling. So a global optimization problem can be solved via a two-phase cycle:<br />
In phase 1, we start from a given point and use any local minimization method to find a local minimizer<br />
x1 ∗ of f (x).<br />
In phase 2, we construct a filled function at x1 ∗ and minimize the filled function in order to identify a point<br />
x ′ with f (x ′ ) < f (x1 ∗).<br />
If such a point x ′ is found, x ′ is certainly in a lower basin than B ∗ 1 . Then we can use x′ as the initial<br />
point in phase 1 again, and hence we can find a better minimizer x2 ∗ of f (x) with f (x∗ 2 ) < f (x∗ 1<br />
). This<br />
process repeats until the time when minimizing a filled function does not yield a better solution. The<br />
current local minimum will be taken as a global minimizer of f (x).<br />
In this paper, we Consider the following unconstrained optimization problem:<br />
(P)<br />
min f (x)<br />
s.t. x ∈ R n .<br />
(1)<br />
let L(P) stands for the set of local minimizers of f (x). The new two-parameters filled function for problem<br />
(P) at the local minimizer x1 ∗ has the following form:<br />
(<br />
F(x, x1 ∗ , r, q) = 1<br />
|f (x) − f (x<br />
∗ )<br />
1 + ‖x − x1 ∗‖ arctan 1<br />
) + r|<br />
|f (x) − f (x1 ∗)| + q , (2)<br />
where 0 < q < r and r satisfies<br />
0 < r < max<br />
x ∗ , x ∗ 1 ∈L(P)<br />
f (x ∗ )
NumAn2014 Book of Abstracts 50<br />
References<br />
[1] T.M. El-Gindy, M.S. Salim, Abdel-Rahman Ibrahim, A Modified partial quadratic interpolation method for unconstrained<br />
optimization, Journal of Concrete and Applicable Mathematics−JCAAM 11(1) (2013) 136−146.<br />
[2] C. Wang, Y. Yang, J. Li, A new filled function method for unconstrained global Optimization, J. Comput. Appl.<br />
Math. 225 (2009) 68−79.<br />
[3] Y. Yang, Y. Shang, A new filled function method for unconstrained global Optimization, Appl. Math. Comput. 173<br />
(2006) 501−512.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 51<br />
Conference in Numerical Analysis 2014 (NumAn 2014)<br />
September 2-5, 2014<br />
Chania,Greece<br />
Multiresolution analysis for 3D scattered data sets<br />
L. Fernández a , M. A. Fortes a and M. L. Rodríguez a<br />
a Department of Applied Mathematics, University of Granada,<br />
Granada, Spain<br />
lidiafr@ugr.es,mafortes@ugr.es,miguelrg@ugr.es<br />
Abstract<br />
In [1], the authors develop a multiresolution analysis in a one-dimensional context based on<br />
Harten’s multiscale representation (see e.g. [2, 3]). In the present work we propose to generalize<br />
[1] to a three-dimensional context in order to handle with clouds of 3D datasets: we obtain<br />
the decomposition and reconstruction algorithms associated to different interpolatory schemes,<br />
such as the one considering just function values, or the one considering function and first derivative<br />
values. Different interpolatory schemes will lead to consider different interpolatory spaces<br />
where to develop the algorithms. As an application of the developed theory, we will consider<br />
some examples regarding data compression and discontinuities detection.<br />
Key words: Multiresolution analysis, decomposition-reconstruction algorithms, compression data, discontinuities<br />
detection.<br />
References<br />
[1] R. M. Beam and R. F. Warming, Discrete multiresolution analysis using Hermite interpolation:<br />
Biorthogonal multiwavelets, SIAM J. Sci. Comput. 22(4), (2000) 1269–1317.<br />
[2] A. Harten, Discrete multi-resolution analysis and generalized wavelets, Applied Numerical<br />
Mathematics 12 (1993), 153–192.<br />
[3] A. Harten, Multiresolution representation and numerical algorithms: A brief review, ICASE<br />
Report No. 94-59, 1994.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 52<br />
Exploring the Performance of Out-of-Core Linear Algebra<br />
Algorithms in Flash based Storage 1<br />
Athanasios Fevgas, Panagiota Tsompanopoulou and Panayiotis Bozanis<br />
Department of Electrical and Computer Engineering, University of Thesssaly,<br />
Volos, Thessaly, Greece<br />
fevgas@inf.uth.gr,yota@inf.uth.gr,pbozanis@inf.uth.gr<br />
Abstract<br />
In the recent years, flash memory has been widely utilized as storage medium to mobile<br />
and embedded systems, laptops and servers. The outstanding efficiency of flash based storages<br />
motivated us to study the performance of out-of-core linear algebra algorithms in flash SSDs.<br />
Flash memory is a non-volatile electronic storage that can be electrically erased and reprogrammed.<br />
There are two types of flash, NOR and NAND with the later utilized as mass storage<br />
medium. In the rest of this document the term flash denotes the NAND flash. Storages based on<br />
flash lack of mechanical and moving parts, providing low power consumption, shock resistance<br />
and high read/write performance. Flash consists of cells which store one or more bits. Cells are<br />
organized to pages and pages to blocks. Reads and writes are performed at page level, while<br />
erases at block level. Write operations are slower than reads and erases are even slower. Moreover,<br />
pages have to be erased before are re-written and flash endurance is limited by a finite<br />
number of write/erase cycles (wear out). Solid state drives (SSDs) are block devices compatible<br />
with traditional hard disk drives (HDDs) relying in flash memories. The main components of an<br />
SSD are the flash memory chips and a controller which emulates the block interface using FTL<br />
(Flash Translation Layer). FTL remaps logical addresses, used by the upper layers, to physical<br />
addresses in flash chips. It incorporates out-of-place-updates, wear leveling and garbage collection<br />
mechanisms aiming to improve write performance and prevent wear out. All the mentioned<br />
flash characteristics make data structures and algorithms designed for hard disks not performing<br />
well in it. Many recent studies, mostly in databases, aim to design new approaches suitable for<br />
flash. Some of them are using deltas instead of performing expensive page rewrites while others<br />
deferring operations in the future in order to reduce random writes.<br />
The development of efficient external memory (out-of-core) algorithms for solving linear equations<br />
systems or calculating eigenvalues of large matrices has been a popular research topic.<br />
Several algorithms have been proposed aiming to accelerate calculations by efficiently partitioning<br />
and managing large disk-resident datasets (matrices) into main memory blocks (submatrices).<br />
Alternative approaches require clusters with distributed memory, large enough for<br />
the entire dataset, and high bandwidth interconnections. Nowadays, the emergence of multiprocessor,<br />
multi-core and GPU accelerated computers provides high processing power at low<br />
cost. On the other hand, flash storages are capable to accelerate the storage layer. Considering<br />
the specifics of the flash memory, we present a study of the performance of few out-of-core<br />
algorithms for numerical linear algebra problems in flash based storages.<br />
Key words: Out-of-Core algorithms, linear algebra, scientific data, flash memory, SSD<br />
1 The present research work has been co-financed by the European Union (European Social Fund ESF) and Greek<br />
national funds through the Operational Program Education and Lifelong Learning of the National Strategic Reference<br />
Framework (NSRF) - Research Funding Program: THALIS. Investing in knowledge society through the European<br />
Social Fund.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 53<br />
A comparative study on the effect of the ordering schemes for<br />
solving sparse linear systems, based on factored approximate<br />
sparse inverse matrix methods<br />
Christos K. Filelis-Papadopoulos, George A. Gravvanis<br />
Department of Electrical and Computer Engineering,<br />
School of Engineering, Democritus University of Thrace,<br />
University Campus, Kimmeria, GR 67100 Xanthi, Greece<br />
cpapad@ee.duth.gr,ggravvan@ee.duth.gr<br />
Abstract<br />
Preconditioned iterative schemes have been used extensively in many scientific disciplines, during<br />
the last decades for solving sparse linear systems. The effectiveness of the Preconditioning<br />
methods relies on the construction and use of efficient preconditioners, in the sense that are<br />
close approximants to the coefficient matrix of the linear system, suitable for modern computer<br />
systems.<br />
Recently, a class of Generic Approximate Inverses has been proposed that can handle any sparsity<br />
pattern of the coefficient matrix. A class of Generic Approximate Sparse Inverse matrix in<br />
conjunction with approximate inverse sparsity patterns, based on powers of sparsified matrices,<br />
has been proposed, that presented improved convergence behavior than existing Generic Approximate<br />
Banded Inverses schemes. Moreover, a factored approach, namely Generic Factored<br />
Approximate Sparse Inverse has been proposed, that further improved the convergence rate and<br />
further reduces the computational complexity and memory requirements. The Modified Generic<br />
Factored Approximate Sparse Inverse is a column wise variant that increases the performance<br />
by reducing the searches for nonzero elements required in the row-wise approach. The reordering<br />
schemes have been used to reduce fill-in for computing the decomposition factors of the<br />
coefficient matrix. Additionally, the reordering schemes have been used to increase the quality<br />
of incomplete factorization used in conjunction with preconditioned iterative methods. The various<br />
reordering schemes, namely Approximate Minimum Degree, the Reverse Cuthill-McKee<br />
and the Block Breadth First Search, affect the number of nonzeros and the quality of Modified<br />
Generic Factored Approximate Sparse Inverse. Moreover, the reordering schemes affect<br />
the sparsity pattern of the resulting approximate sparse inverse preconditioners and the convergence<br />
behavior of the proposed schemes.<br />
Finally, we examine the effectiveness and applicability of the various ordering schemes on the<br />
computation of the Modified Generic Factored Approximate Sparse Inverse (MGenFAspI) matrix<br />
in conjunction with the preconditioned Bi-Conjugate Gradient STABilized method for solving<br />
various problems from Matrix Market collection and numerical results are given, which are<br />
comparatively better than existing ones.<br />
Key words: Modified Generic Factored Approximate Sparse Inverses, Reordering schemes, Preconditioned<br />
iterative methods, Sparsity patterns.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book Conference of Abstracts in Numerical Analysis 2014 (NumAn 2014) 54<br />
September 2-5, 2014<br />
Chania,Greece<br />
Tsunami hazard and inundation<br />
for the northern coast of Crete<br />
Flouri Evangelia a,b , Vassilios Dougalis a , and Costas Synolakis b<br />
a Institute of Applied and Computational Mathematics,<br />
Foundation for Research and Technology Hellas, Heraklion, Crete<br />
b Department of Environmental Engineering,<br />
Technical University of Crete, Chania, Crete<br />
flouri@iacm.forth.gr<br />
Abstract<br />
Tsunamis are rare events compared to other natural hazards, but population growth along shorelines<br />
has increased their potential impact. Tsunamis are usually generated by an earthquake–<br />
induced dislocation of the seabed which displaces a large mass of water. They can be simulated<br />
effectively as long waves whose propagation and inundation are modeled by the nonlinear shallow<br />
water equations.<br />
In this work, we present a systematic assessment of earthquake-generated tsunami hazards<br />
for the northern coastal areas of the island of Crete. Our approach is based on numerical hydrodynamic<br />
simulations, including inundation computations, with the model MOST, using accurate<br />
bathymetry and topography data of the study area. MOST implements a splitting method<br />
in space to reduce the hyperbolic system of shallow water equations in two successive systems,<br />
one for each spatial variable, and uses a dispersive, Godunov–type finite difference method to<br />
solve the equations in characteristic form.<br />
In the present study we consider hypothetical, but credible, ‘worst case’ scenarios based on<br />
the unit sources methodology of NOAA, and, present inundation results, associated with seismic<br />
events of magnitude 8.5 originated in the Hellenic Arc, and 7.5 due to the seismic sources of the<br />
central Aegean sea. We also implement a probabilistic scenario in which we assess the influence<br />
of the epicenter location on the tsunami hazard, for time windows of 100, 500 and 1000 years.<br />
Our results include calculations of the maximum inundation and the maximum wave elevation<br />
for the two largest cities of the northern coast of Crete, Chania and Heraklion. We illustrate our<br />
findings superimposed on satellite images as maps indicating the estimated maximum values.<br />
Key words: tsunami hazard, inundation, Crete.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 55<br />
Conference in Numerical Analysis 2014 (NumAn 2014)<br />
September 2-5, 2014<br />
Chania,Greece<br />
Eigenvalues and eigenfunctions for the Laplace Operator<br />
Athanassios S. Fokas a and Konstantinos Kalimeris b<br />
a DAMTP, University of Cambridge,<br />
Cambridge, UK<br />
b RICAM, Austrian Academy of Sciences,<br />
Linz, Austria<br />
konstantinos.kalimeris@ricam.oeaw.ac.at<br />
Abstract<br />
The eigenvalues of the Laplace operator for the Dirichlet, Neumann and Robin problems<br />
in the interior of an equilateral triangle were first obtained by Lamé. Here, we present a simple,<br />
unified approach for deriving the relevant eigenvalues for several types of Boundary Value<br />
Problems (BVPs). Among these results the most general one consists of a system of explicit algebraic<br />
equations which give the eigenvalues for the Poincaré type BVP. These formulae for the<br />
Poincaré eigenvalues yield, via appropriate limits, the relevant formulae for the oblique Robin,<br />
Robin, Neumann and Dirichlet eigenvalues. The latter three give exactly the above mentioned<br />
results of Lamé. The method introduced here is based on the analysis of the so-called global<br />
relation, which as shown recently in the literature provides an effective tool for the study of<br />
BVPs. Moreover, we illustrate results considering the relevant eigenfunctions and some ideas<br />
related to other convex and bounded regular domains<br />
Key words: Eigenvalues, Laplace operator, global relation.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 56<br />
Conference in Numerical Analysis 2014 (NumAn 2014)<br />
September 2-5, 2014<br />
Chania,Greece<br />
Filling holes with geometric constraints<br />
M. A. Fortes a , P. González a , A. Palomares a and M. Pasadas a<br />
a Department of Applied Mathematics, University of Granada,<br />
Granada, Spain<br />
mafortes@ugr.es,prodelas@ugr.es,anpalom@ugr.es,mpasadas@ugr.es<br />
Abstract<br />
Let D ⊂ R 2 be a polygonal domain, H be a subdomain of D and f : D − H −→ R be<br />
a function. In this paper we propose a method to reconstruct the ‘hole’ of f over H using a<br />
technique based on the minimization of an energy functional. More precisely, we construct a<br />
C 1 -Powell-Sabin spline function f ∗ over the whole D that approximates f outside H, and fills<br />
the hole of f inside H by respecting some geometric constraints.We present some graphical and<br />
numerical examples.<br />
Key words: Filling, approximation, finite element, Powell-Sabin, minimal energy.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book Conference of Abstracts in Numerical Analysis 2014 (NumAn 2014) 57<br />
September 2-5, 2014<br />
Chania,Greece<br />
Matrix-free resolution of PDEs using the Powell-Sabin FE<br />
Miguel A. Fortes a , Pedro González a , Antonio Palomares a and Miguel<br />
Pasadas a<br />
a Department of Applied Mathematics, University of Granada,<br />
Granada, Spain<br />
mafortes@ugr.es,prodelas@ugr.es,anpalom@ugr.es,mpasadas@ugr.es<br />
Abstract<br />
Let consider the following general boundary-value problem of second or fourth-order (depending<br />
on the values of τ 1 , τ 2 ∈ R, not vanishing simultaneously),<br />
⎧<br />
⎪⎨<br />
⎪⎩<br />
−l ∂ (l)<br />
t u − τ 1 ∆u + τ 2 ∆ 2 u = f, t > 0 in Ω<br />
u (l t, ·) = ϕ (l t, ·) , τ 2<br />
∂u<br />
∂n (l t, ·) = τ 2 ψ (l t, ·) , t ≥ 0 on Γ<br />
l u (0, ·) = l u 0 (·) , (l − 1) ∂ ∂t u (0, ·) = (l − 1)u 1 (·) , on Ω<br />
where, depending on the value of l ∈ {0, 1, 2}, ∂ (l)<br />
t u will denote just u (for l = 0), ∂u<br />
∂t (for<br />
l = 1) or ∂2 u<br />
(for l = 2) and the problem considered may be elliptic, parabolic or hyperbolic.<br />
∂t 2<br />
For solving numerically any of these problems, depending on whether it is a transient<br />
parabolic or hyperbolic PDE problem, in a finite temporal interval [0, T ] ⊂ R (with T > 0), or<br />
just a stationary elliptic one (independent of time, or just considered only for t = 0), we will<br />
apply a general Galerkin procedure to their corresponding variational formulation.<br />
In this work we present a procedure to obtain a C 1 -surface on a polygonal domain Ω ⊂ R 2 ,<br />
depending or not on time, that also solves the corresponding Galerkin variational formulation<br />
of a transient or stationary PDE problem up to fourth-order (1). The approximation space is that<br />
of C 1 -quadratic splines constructed from the Powell-Sabin subtriangulation associated with an<br />
α-triangulation of Ω.<br />
The main idea also is to try to avoid the resolution of any large linear system, or even to consider<br />
matrix-free formulations of the problems, using special triangulations with not too many<br />
(or even without any) nodes on the interior of the domain, and by using the appropriate interpolation<br />
conditions over some points in the boundary in order to take into account the boundary<br />
conditions for each of these problems. We study the actual feasibility of such procedure for these<br />
prototype problems, and give some numerical and graphical examples to assess their efficiency<br />
and reliability.<br />
(1)<br />
Key words: Powell-Sabin FE, interpolating PS-splines, matrix-free formulation.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 58<br />
Conference in Numerical Analysis 2014 (NumAn 2014)<br />
September 2-5, 2014<br />
Chania,Greece<br />
Numerical Solution for Sparse Linear Systems that occur from<br />
the discretization of Boussinesq-type equations<br />
Maria Gaitani a , Maria Kazolea b and Argiris Delis a<br />
a School of Production Engineering and Management, Technical University of<br />
Crete, Chania, Crete, Greece<br />
b School of Environmental Engineering, Technical University of Crete, Chania,<br />
Crete, Grece<br />
mgaitani1@isc.tuc.gr,mkazolea@isc.tuc.gr, adelis@science.tuc.gr<br />
Abstract<br />
This work investigates preconditioned iterative techniques for the solution for sparse linear systems<br />
that occurs from the discretization of Boussinesq-type (BT) models using a finite volume<br />
scheme on unstructured meshes. The past few years enhanced Boussinesq-type (BT) models<br />
and their numerical solutions have evolved as predictive tools in the modeling of wave propagation<br />
and transformations. Recently, a novel high-order FV scheme on unstructured meshes<br />
for the extended 2D BT equations of Nwogu was developed. The equations of Nwogu are recasted<br />
in the form of a system of balance laws and are then numerically solved using a novel<br />
high-order well-balanced FV numerical method in unstructured meshes. In each time step the<br />
solution of a large sparse linear system (with a mesh depended matrix, M that occurs from the<br />
discretization of the dispersion terms) is mandatory to recover the velocity field. Matrix M is<br />
sparse, un-symmetric and often ill-conditioned. The properties of the matrix also vary on the<br />
physical situation of the problem examined. Various preconditioned and reordering strategies<br />
are investigated, including the ILU factorization the ILUT factorization and the CMK and RCM<br />
reordering techniques. Two iterative methods, BicGstab and GMRES, are tested for the solution<br />
process. A detailed comparison of the methods is given and their strengths and limitations of<br />
each are discussed. Furthermore, the performance of the various strategies is tested versus the<br />
most important parameters of the problem examined.<br />
Key words: sparse matrix, finite volumes, Boussinesq-type equations.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book Conference of Abstracts in Numerical Analysis 2014 (NumAn 2014) 59<br />
September 2-5, 2014<br />
Chania, Greece<br />
Approximations Using Radon Projection Data in the Unit Disc<br />
Irina Georgieva a , Clemens Hofreither b and Rumen Uluchev c<br />
a Institute of Mathematics and Informatics, Bulgarian Academy of Sciences,<br />
Sofia, Bulgaria<br />
b “Computational Mathematics”, Johannes Kepler University Linz<br />
Linz, Austria<br />
c Department of Mathematics and Informatics, University of Transport<br />
Sofia, Bulgaria<br />
irina@math.bas.bg, chofreither@numa.uni-linz.ac.at, rumenu@vtu.bg<br />
Abstract<br />
Noninvasive methods using line integrals for 2D object reconstruction have their theoretical<br />
foundation in the work of Johann Radon in the early twentieth century and have important<br />
practical applications in medicine, geology, radiology, astronomy, etc.<br />
In our survey we present recent results on various approximation problems where the basic<br />
information consists of line integrals in the unit disc. For instance, in 2D computing tomography,<br />
the data on which the reconstruction is based, comes as Radon projections along fixed directions.<br />
More generally, sometimes our data includes function values on the unit circle, in addition to<br />
line integrals. Our methods stay nondestructive in such a case and analytical as well.<br />
For Radon projection data we have studied:<br />
• interpolation and fitting by bivariate polynomials;<br />
• interpolation by quadratic bivariate splines;<br />
• interpolation and fitting by harmonic polynomials;<br />
• interpolation problem for the Poisson equation;<br />
• cubatures for harmonic functions.<br />
Questions under consideration were to determine sets of chords on the unit disc for which<br />
the relevant problem has a unique solution, developing numerical algorithms, error estimation,<br />
etc.<br />
Key words: Interpolation, fitting, cubature, bivariate polynomial, harmonic function, Poisson equation,<br />
Radon transform.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book Conference of Abstracts in Numerical Analysis 2014 (NumAn 2014) 60<br />
September 2-5, 2014<br />
Chania,Greece<br />
Splitting methods based on Approximate Matrix Factorization<br />
and Radau-IIA formulas for the time integration of advection<br />
diffusion reaction PDEs.<br />
Severiano González-Pinto a , Domingo Hernández-Abreu a<br />
a Departamento de Análisis Matemático. Universidad de La Laguna,<br />
Santa Cruz de Tenerife, Spain.<br />
spinto@ull.es,dhabreu@ull.edu.es<br />
Abstract<br />
A family of methods for the time integration of evolutionary Advection Diffusion Reaction<br />
Partial Differential Equations (PDEs) semi-discretized in space is introduced. The methods are<br />
obtained by combining a splitting J h = ∑ d<br />
j=1 J h,d of the Jacobian matrix J of the resulting<br />
ODE -where h is a small positive parameter related to the spatial resolution, such as the meshwidth-<br />
and a number of inexact Newton Iterations applied to the two-stage Radau IIA method.<br />
The overall process reduces the storage and the algebraic cost involved in the numerical solution<br />
of the multidimensional linear systems to the level of one-dimensional linear systems with small<br />
bandwidths.<br />
The local error of AMF-Radau methods when applied to semi-linear equations is described.<br />
From here, since the order in time is at most three, some specific methods considering up to<br />
three inexact Newton Iterations are selected. Furthermore, linear stability properties for the<br />
selected methods are established, in such a way that the wedges of stability depend on the<br />
number of terms d considered in the splitting J h . In particular, A(α d )-stability is shown for<br />
1 ≤ d ≤ 4, where α d := min{ π 2 , π<br />
2(d−1)<br />
}, and A(0)-stability for any d ≥ 1.<br />
Numerical experiments on 2D and 3D problems are presented, and they show that the methods<br />
compare well with standard classical methods in parabolic problems and can also be successfully<br />
used for advection dominated problems whenever some diffusion or stiff reactions are<br />
present.<br />
Key words: Evolutionary Advection-Diffusion-Reaction Partial Differential Equations, Approximate Matrix<br />
Factorization, Runge-Kutta Radau IIA methods, Stability.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 61<br />
On the numerical modelling and solution of multi-asset<br />
Black-Scholes equation based on Generic Approximate Sparse<br />
Inverse Preconditioning<br />
Eleftherios-Nektarios G. Grylonakis, Christos K. Filelis-Papadopoulos,<br />
George A. Gravvanis<br />
Department of Electrical and Computer Engineering,<br />
School of Engineering, Democritus University of Thrace,<br />
University Campus, Kimmeria, GR 67100 Xanthi, Greece<br />
elevgryl@ee.duth.gr,cpapad@ee.duth.gr,ggravvan@ee.duth.gr<br />
Abstract<br />
One of the most important topics in the area of financial mathematics is the study of the multiasset<br />
Black-Scholes equation for the pricing of options. While there is a closed-form solution<br />
in one dimension for pricing European vanilla options, in higher dimensions the finite difference<br />
method allows the consideration of a wider range of parameters (coefficients of the partial<br />
differential equation, initial and boundary conditions). Hence, research efforts have been directed<br />
towards finding accurate prices for options with two or more underlying assets. In this<br />
paper, we present a fourth order accurate discretization scheme for the numerical solution of<br />
Black-Scholes equation in two space variables.<br />
The purpose of this work is to derive fourth order accurate option pricing methods while maintaining<br />
low computational complexity. For the space discretization we use a fourth order finite<br />
difference scheme combined with Richardson’s extrapolation method while for the time integration<br />
high order Backward Differences along with fourth order Gauss-Legendre Runge-Kutta<br />
scheme was used. The resulting sparse linear system of algebraic equations is solved by preconditioned<br />
iterative techniques based on generic approximate sparse inverses. Herewith, the<br />
Preconditioned Induced Dimension Reduction (PIDR(s)) method in conjunction with Generic<br />
Approximate SParse Inverse (GenAspI) is used for the efficient solution of the sparse linear<br />
systems. The GenAspI is computed through an incomplete factorization of the coefficient matrix<br />
to a predefined sparsity pattern acquired from Powers of Sparsified Matrices (PSM’s), thus<br />
handling any sparsity pattern.<br />
Numerical results are presented along with discussions for the proposed schemes in order to<br />
highlight the applicability and efficiency for solving the Black-Scholes equation in two space<br />
variables. The implementation issues of the proposed method are also discussed.<br />
Key words: Multi-Asset Black-Scholes equation, high order finite difference schemes, sparse linear systems,<br />
generic approximate sparse inverses, preconditioned induced dimension reduction method.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 62<br />
The Error Analysis of the Indirect Padé Method for Matrix Exponential 1<br />
Chuanqing Gu, Ke Zhang<br />
Department of Mathematics, Shanghai University, Shanghai 200444, China<br />
cqgu@staff.shu.edu.cn,xznuzk123@126.com<br />
Abstract<br />
One of the most frequently discussed matrix function is the matrix exponential. Compared with<br />
other methods for computing the matrix exponential, the scaling and squaring method highlights<br />
itself not least for its implementation in MATLAB function expm. Najfeld and Havel in [Adv. in<br />
Appl. Math., 16 (1995)] presented an efficient algorithm using the scaling and squaring method<br />
as well as indirect Padé approximation which is a little different from the traditional method<br />
for the computation of the matrix exponential. This method is known for its lower computation<br />
cost, however, as pointed out by Higham in [SIAM J.Matrix Anal. Appl., 26 (2005)], it is lack<br />
of sufficient error analysis. In this paper we give an analysis of the sensitivity and conditioning<br />
of the matrix polynomial function H(B) by Najfeld and Havel given and conclude that H(B)<br />
is a well-conditioned matrix. We also present the relative error bounds of the approximating<br />
function (H 2m (B) + B)(H 2m (B) − B) −1 and exploit the impact of the condition number of<br />
H(B)−B and the scaling times d on the relative error bounds in a heuristic way. Our numerical<br />
result shows that the algorithm given by Najfeld and Havel generally provides accuracy almost<br />
the same as the MATLAB 7.6 functions expm and funm with a lower cost and proves to be<br />
stable.<br />
Key words: Matrix exponential, error analysis, Padé approximation, scaling and squaring, overscaling,<br />
MATLAB.<br />
1 The work are supported by National Natural Science Foundation (11371243), by Innovation major project of Shanghai<br />
Municipal Education Commission (13ZZ068) and by Key Disciplines of Shanghai Municipality (S30104).<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book Conference of Abstracts in Numerical Analysis 2014 (NumAn 2014) 63<br />
September 2-5, 2014<br />
Chania,Greece<br />
Optimization of pre-recontruction restoration filtering for<br />
filtered back projection reconstruction (FBP)<br />
R. Guedouar 1 , A. Bouzabia 2 , B. Zarrad 3,4<br />
1 Biophysics department, faculty of pharmacy, University of Monastir, Tunisia<br />
2 Electronics and automatics department, High Institute of Informatics and Math,<br />
University of Monastir, Tunisia<br />
3 Biophysics and medical imaging department, Higher school of health sciences<br />
and technicals, University of Monastir, Tunisia<br />
4 Biophysics laboratory, Higher institute of medical technologies, University of<br />
Tunis-Elmanar, Tunisia<br />
raja guedouar@yahoo.fr<br />
Abstract<br />
Tomographic reconstruction is the technique underlying nearly all of the key diagnostic imaging<br />
modalities since they generated 2D/3D representation from a set of 2D projections acquired<br />
by a topographic system. For a long time, and despite the advantages of iterative algorithms,<br />
the FBP algorithm was preferred because it was computationally faster and more practical for<br />
routine use. However, in clinical practice, the FBP algorithm performances depend on several<br />
parameters effecting seriously final reconstruction results: The use of a smoothing filter to reduce<br />
the noise results in a loss of resolution and the choice of an optimal pre-reconstruction filter<br />
is necessary to provide the best trade-off between image noise and image resolution. A significant<br />
improvement in the quality of SPECT images has been demonstrated through the use of 2D<br />
pre-reconstruction restoration filtering of the projection images with FBP techniques. However,<br />
these filters should be designed to account for the image blurring, the noise level, and the imaged<br />
object to obtain the maximum restoration of image quality. In this work, we propose a userfriendly<br />
Interface for interactive optimization of FBP pre-construction filtering with emphases<br />
on Image-dependent restoration filters. The framework proposes an interactive visual algorithm<br />
for implementation of digital smoothing (hanning and Butterworth) and semi-automatically implementation<br />
of Metz restoration filters in frequency domain. To more objectively determine the<br />
optimum cut-off frequency, the user is assisted in visual feed-back optimization, by displaying<br />
the calculated of the power spectrum of both projection and estimated noise, and the filtered<br />
reconstructed images. A comparative study using 6464 2D-noiseless-numerical simulated and<br />
myocardial perfusion SPECT data, was conducted to investigated the performance of optimized<br />
filters on a pre-reconstruction task using FBP in terms of visual assessment, mean standard deviation<br />
and contrast. The reconstruction was done by a conventional FBP method using a ramp<br />
filter with no attenuation or scatter correction. Results show that optimized Metz restoration filtering<br />
provides reconstructed data with reduced noise without unduly penalizing resolution. It is<br />
also giving more improvement in clinical SPECT image contrast than Butterworth and Hanning.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 64<br />
Conference in Numerical Analysis 2014 (NumAn 2014)<br />
September 2-5, 2014<br />
Chania,Greece<br />
On the Solution of the Linear Complementarity<br />
Problem by the Generalized Accelerated<br />
Overrelaxation Iterative Method 1<br />
Apostolos Hadjidimos a and Michael Tzoumas b<br />
a Department of Electrical and Computer Engineering<br />
University of Thessaly, 382 21 Volos, Greece<br />
b Department of Mathematics, University of Ioannina<br />
451 10 Ioannina, Greece<br />
hadjidim@inf.uth.gr,mtzoumas@sch.gr<br />
Abstract<br />
In the present work, we determine intervals of convergence for the various parameters involved<br />
for what is known as the Generalized Accelerated Overrelaxation iterative method for the solution<br />
of the Linear Complementarity Problem. The convergence intervals found constitute<br />
sufficient conditions for the Generalized Accelerated Overrelaxation method to converge and<br />
are better than what have been known so far.<br />
1 J. Optim. Theory Appl., in press, DOI 10.1007/s10957-014-0589-4<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 65<br />
Conference in Numerical Analysis 2014 (NumAn 2014)<br />
September 2-5, 2014<br />
Chania,Greece<br />
Strong-stability-preserving additive linear multistep methods<br />
Yiannis Hadjimichael a and David I. Ketcheson a<br />
a Computer, Electrical and Mathematical Sciences and Engineering Division,<br />
King Abdullah University of Science and Technology (KAUST),<br />
P.O. Box 4700, Thuwal 23955, Saudi Arabia<br />
yiannis.hadjimichael@kaust.edu.sa, david.ketcheson@kaust.edu.sa<br />
Abstract<br />
Semi-discretization of a variety of partial differential equations results in ordinary differential<br />
systems containing terms with different stiffness properties. In such cases additive methods can<br />
be used to make the most of the special structure of the resulting system. We study the monotonicity<br />
properties of additive linear multistep methods. We show that for a fixed number of<br />
steps and order of accuracy, optimal strong-stability-preserving (SSP) additive methods attain<br />
the same time-step restriction as the optimal SSP linear multistep methods, regardless of the<br />
stiffness of the problem. The concept of SSP linear multistep methods is also extended to problems<br />
for which the upwind- and downwind-biased operators have different stiffness properties.<br />
Key words: strong-stability-preservation (SSP), monotonicity, linear multistep methods, time integration<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book Conference of Abstracts in Numerical Analysis 2014 (NumAn 2014) 66<br />
September 2-5, 2014<br />
Chania,Greece<br />
Fokas method and Kelvin transformation applied to potential<br />
problems in non convex unbounded domains.<br />
Maria Hadjinicolaou<br />
School of Science and Technology, Hellenic Open University,<br />
Patras, Greece<br />
hadjinicolaou@eap.gr<br />
Abstract<br />
In this presentation Fokas integral method is combined with Kelvin transformation to develop<br />
a new method for solving Dirichlet or Neumann problems in non-convex unbounded<br />
domains. A key aspect in Fokas method is the coupling of all boundary values in one equation,<br />
which has been termed global relation. Through this, any missing data on a boundary value<br />
problem can be derived, as Dassios and Fokas have shown. On the other hand, Kelvin transformation<br />
preserves harmonicity, and thus, by applying it to an exterior potential problem, the<br />
solution of the equivalent interior problem can be established, in the domain which is the Kelvin<br />
image of the original exterior one. In the present work, these two methods have been employed<br />
in order to derive integral representations for the Dirichlet and the Neumann problem in a nonconvex<br />
domain which is the Kelvin image of an equilateral triangle. The proposed methodology<br />
for the case of a Neumann exterior problem is given below. Physically, this could be explained<br />
as the construction of a potential for a vector field of which the effect of its normal derivative is<br />
known along its boundary that is assumed to be the image of an equilateral triangle under the<br />
Kelvin transformation.<br />
First, we apply the Kelvin inversion and thus the corresponding Neumann data on the boundary<br />
of the equilateral triangle are obtained. By then employing the Neumann to Dirichlet map,<br />
the Dirichlet data on the perimeter of the triangle are extracted. Subsequently, an integral representation<br />
of the solution of the Neumann problem in the interior of the triangle is accomplished.<br />
Applying again the Kelvin transformation to the attained Dirichlet data we derive the<br />
corresponding Dirichlet data on the initial boundary. By employing Kelvins 2-D theorem, we<br />
eventually obtain an integral representation of the solution of the Neumann problem in the given<br />
exterior non convex domain. Furthermore, we derive the Neumann to Dirichlet map for every<br />
Fourier component of some arbitrary data. This way, we provide a basis for representing the<br />
Dirichlet data on the boundary and thus we can obtain an integral representation of the solution<br />
of a large class of potential problems regarding non convex domains, encounted in many fields<br />
of science and engineering, that they would not be possible otherwise.<br />
Alternatively, in the case of a Dirichlet problem, we pursue the proposed methodology modified<br />
appropriately to obtain analogous results.<br />
Key words: Fokas method, Kelvin inversion integral representation , potential problems<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 67<br />
Conference in Numerical Analysis 2014 (NumAn 2014)<br />
September 2-5, 2014<br />
Chania,Greece<br />
Construction and approximation of surfaces by smoothing<br />
meshless methods.<br />
A. Hananel a , M. Pasadas a , and M. L. Rodríguez a<br />
a Department of Applied Mathematics, University of Granada,<br />
Granada, Spain<br />
ahananel@ugr.es, mpasadas@ugr.es, miguelrg@ugr.es<br />
Abstract<br />
In Earth science, especially Geology and other Sciences and Technologies, the reconstruction<br />
of surfaces from some scattered data set is a commonly encountered problem.<br />
In this work, under a generic schema, we enrich the theory of the discrete variational spline<br />
functions by minimizing some quadratic functional in a suitable space which can be a fairness<br />
functional, the flexion energy of a thin plate or others.<br />
It is essential to consider a finite dimension space of functions, where the minimization<br />
problem can be solved, and then a variational problem will be formulated. The discrete finite<br />
dimension space of functions that we propose, in this case, is a parametric finite dimensional<br />
space generated by a radial function basis. Then, we describe a smoothing meshless method of<br />
surfaces. The convergence of the problem is shown and finally, we analyze some numerical and<br />
graphical examples.<br />
Key words: Surfaces, approximation, smoothing, meshless methood.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 68<br />
Conference in Numerical Analysis 2014 (NumAn 2014)<br />
September 2-5, 2014<br />
Chania,Greece<br />
The definitive estimation of the neuronal current via<br />
EEG and MEG using real data<br />
P. Hashemzadeh and A.S Fokas<br />
Department of Applied Mathematics and Theoretical Physics<br />
University of Cambridge, UK<br />
hashemzadeh@damtp.cam.ac.uk T.Fokas@damtp.cam.ac.uk<br />
Abstract<br />
The medical significance of Electroencephalography (EEG), and Magneto-Electroencephalography<br />
is well established, see for examples [1, 2, 3, 5]. EEG and MEG are considered two of<br />
the most important imaging techniques for real time brain imaging. In order to generate images<br />
of the brain activation using either EEG or MEG, it is necessary to analyse certain mathematical<br />
inverse problems. The definitive answer to the inverse source problem for the case of EEG<br />
and MEG was finally obtained by [4]. Here, we present reconstructions of the current using real<br />
data via the formulation proposed by [4]. The data was provided by the medical research council<br />
(MRC) Cambridge, UK. It involves both auditory and visual stimulus. We show comparisons of<br />
the reconstructed irrotational component of the neuronal current using EEG measurements and<br />
the radial component of the neuronal current using MEG measurements. Based on the results,<br />
we argue that EEG imaging technology has the potential to become the dominant real time, low<br />
cost brain imaging tool.<br />
References<br />
[1] Ribary U Ionannides A A Singh K D Hasson R Bolton J P R Lado F Mogilner A and LLinas<br />
R. Magnetic field tomography of coherent thalamocortical 40-hz oscillations in humans. Proc.<br />
Natl Acad. Sci. USA, 8,11 037-11 041, 1991.<br />
[2] Hauk O Rockstroth B Eulitz C. Gapheme monitoring in picture naming: an electrophysiological<br />
study of language production. Brain Topogr., 14:3–13, 2001.<br />
[3] Papanicolaou A C. The amensias: a clinical textbook of memory disorders. Oxford, UK: Oxford<br />
University Press., 2006.<br />
[4] A S Fokas. Electro-magneto-encephalography for a three-shell model: distributed current in<br />
arbitrary, spherical and ellipsoidal geometries. J.R.Soc. Interface, 6:479–488, 2009.<br />
[5] Langheim F J Leuthold A C Georgopolous A P. Synchronous dynamic brain networks revealed<br />
by magnetoencephalography. Proc, 103:455–459, 2006.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 69<br />
Conference in Numerical Analysis 2014 (NumAn 2014)<br />
September 2-5, 2014<br />
Chania,Greece<br />
Numerical Solution of the Unified Transform for<br />
Linear Elliptic PDEs in Polygonal Domains<br />
P. Hashemzadeh and A.S Fokas<br />
Department of Applied Mathematics and Theoretical Physics<br />
University of Cambridge, UK<br />
hashemzadeh@damtp.cam.ac.uk T.Fokas@damtp.cam.ac.uk<br />
Abstract<br />
Integral representations for the solution of linear elliptic partial differential equations (PDEs)<br />
can be obtained using Green’s theorem. A new transform method for solving BVPs for linear<br />
and integrable nonlinear PDEs usually referred to as the Unified Transform or (Fokas Transform)<br />
was introduced by the second author [1]. The numerical implementation of this method has led<br />
to new numerical techniques for both evolution and elliptic PDEs, see for example [4, 5, 2, 3].<br />
Here, we consider Laplace, Helmholtz, and modified Helmholtz equations in polygonal domains<br />
with a Robin boundary condition. We validate and compare the numerical solution obtained by<br />
Unified Transform to the solution obtained via the finite element method (FEM). We present a<br />
simple rule for choosing collocation points-i.e points in the Complex Fourier plane where the<br />
so called global relations are evaluated which guarantees a low condition number the matrix of<br />
the associated linear system.<br />
References<br />
[1] Fokas A S. A unified transform method for solving linear and certain nonlinear PDEs. Proc.<br />
R. Soc. A, 453:1411–1443, 1997.<br />
[2] Bengt Fornberg and Nathasha Flyer, A numerical implementation of Fokas boundary integral<br />
approach: Laplace’s equation on a polygonal domain, Proc. R. Soc.A, 467:2083–3003, 2011.<br />
[3] C.I. Davis and Bengt Fornberg, A spectrally accurate numerical implementation of the Fokas<br />
transform method for Helmholtz-type PDEs. Complex Variables and Elliptic Equations, 59:<br />
Issue 4:564–577, 2014.<br />
[4] Sifalakis A.G, Papadopoulou E.P and Saridakis Y G, Numerical study of iterative methods<br />
for the solution of the Dirichlet-Neumann map for linear elliptic PDEs on regular polygon<br />
domains, Int. J. Appl. Math. Comput. Sci , 4:173-178, 2007.<br />
[5] Sifalakis A.G, Fulton S.R, Saridakis Y.G. Direct and iterative solution of the generalized<br />
Dirichlet-Neumann map for elliptic PDEs on square domains. J. Comput. Appl. Math,<br />
227:171-184, 2009.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 70<br />
Conference in Numerical Analysis 2014 (NumAn 2014)<br />
September 2-5, 2014<br />
Chania,Greece<br />
Symmetric Key Cryptography Algorithms Based on Numerical<br />
Methods<br />
Youssef Hassoun a and Hiba Othman b<br />
a,b Department of Mathematics, American University of Science and Technology,<br />
Beirut, Lebanon<br />
yhassoun@aust.edu.lb,hothman@aust.edu.lb<br />
Abstract<br />
Cryptography is used to protect information content communicated over a network from being<br />
accessed by adversaries. This is achieved by transforming (encrypting) plaintext before transmission<br />
in such a way that its contents can only be disclosed upon application of a reverse<br />
transformation (decryption). Both transformations involve a secret component which can be<br />
either the transformations themselves or some key used in the process. This paper focuses on<br />
implementing symmetric-key cryptography algorithms based on numerical methods. An empirical<br />
study is performed investigating the correlation of encryption and decryption efficiency of<br />
different root-finding numerical methods to the size of the plaintext and to key parameters.<br />
There are two categories of key-based cryptographic algorithms 1 , symmetric-key and asymmetrickey<br />
or public-key algorithms [1, 2, 3]. In the first category, sender and recipient share a private<br />
key known only to both; the same key is used for encryption and decryption. By contrast, in<br />
public-key cryptography two keys are used, one key is made publicly available to senders for<br />
encrypting plaintexts while a second key is kept secret and is used by the recipient for decrypting<br />
ciphered texts.<br />
Depending on the plaintext chunks on which an algorithm operates, symmetirc encryption algorithms<br />
are classified as stream and block ciphers. Stream ciphers operate on individual characters<br />
one at a time using time-varying encryption transformation. Block ciphers, on the other<br />
hand, operate on blocks of characters (n ≥ 64 bits) using fixed encryption transformation.<br />
Menzes et. al [3] defines a block cipher as an encryption function which maps n-bit plaintext<br />
block into an n-bit ciphertext block, where n represents the blocklength- a substitution cipher<br />
with a large character size. The function is bijective and is parametrized by a k-bit key. DES is<br />
an example of a 64-bit block cipher with a 56-bit key. Caesar cipher can be classified as a block<br />
cipher with one character block and a shift of k characters as key.<br />
We propose a symmmetric-key encryption algorithm based on solving a system of linear equations.<br />
It is a block cipher that maps n-characters of plaintext into n-characters of ciphertext. The<br />
1 Hash functions are one-way and do not fall into these categories, since there is no decryption<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 71<br />
key consists of (n×1) vector b j and an (n×n) matrix (a ij ). Encrypting a block of n characters,<br />
represented by (n × 1) vector (c j ), is achieved by solving the systems of linear equations:<br />
n∑<br />
a ij x j = b j − c j (I)<br />
i=1<br />
Provided that det(a ij ) ≠ 0, the solution vector (x ∗ j ) exists; it represents the cipher text and is of<br />
dimension (n × 1). The condition on (a ij ) guarantees that encryption function is bijective and,<br />
consequently, has an inverse- the decryption function. Decrypting the ciphered text is achieved<br />
by substituting solution vectors into equation (I) giving rise to c j = ∑ n<br />
i=1 a ij x ∗ j − b j.<br />
Another algorithm proposed in [4] and based on solving non-linear equations can also be classified<br />
as a 1-character block cipher. Any non-linear function with one variable can be defined<br />
as a key 2 . The encryption function is defined as finding the solution of the equation:<br />
f(x) − c i = 0<br />
Here, (c i ) represents the numerical code of the i th character in the plaintext, e.g., the ascii-code,<br />
and f(x) is an arbitrary non-linear function, polynomial or otherwise. To guarantee that encryption<br />
function has an inverse, numerical encoding of plaintext together with f(x) must be<br />
chosen in such a way that equation (II) has at least one real root. The set of resulting roots {x ∗ i }<br />
represents the ciphertext. On the recipient side, each entry (x ∗ i ) in the ciphered text is decrypted<br />
by substituting it into f(x) giving rise to c i = f(x ∗ i ).<br />
To conclude, we propose symmetric encryption algorithms based on solving a system of linear<br />
equations as well as solving non-linear equations using numerical methods [5]. The proposed<br />
cryptosystems are block ciphers with matrices and non-linear functions as private keys. To make<br />
encryption functions invertible, and thereby guarantee decryption, the keys are constrained to<br />
satisfy some conditions. We implement the algorithms and perform an empirical study investigating<br />
the efficiency of encryption and decryption functions in terms of the functions’ parameters<br />
and in terms of plaintext message size.<br />
References<br />
[1] N. Ferguson, B. Schneier and T. Kohno (2010). Cryptography Engineering. John Wiley &<br />
Sons. ISBN: 9780470474242<br />
[2] B. Schneier (1996). Applied Cryptography: Protocols, Algorithms, and Source Code in C.<br />
John Wiley & Sons. ISBN: 0471117099<br />
[3] A. J. Menezes, P. C. van Oorschot and S. A. Vanstone (1996). Handbook of Applied Cryptography.<br />
CRC Press. ISBN: 0849385237<br />
[4] A. Ghosh and A. Saha (2013). A Numerical Method Based Encryption Algorithm with<br />
Steganography. R. Bhattacharyya et al. (Eds) : ACER 2013, pp. 149157. CS & IT-CSCP<br />
[5] R. L. Burden and J. D. Faires(2010). Numerical Analysis, 9 th Edition. Cengage Learning.<br />
ISBN-13: 9780538733519<br />
(II)<br />
Key words: Cryptography, Encryption, Numerical Methods, Symmetric-Key.<br />
2 The authors in [4] proposed a polynomial function of degree 3 as a key<br />
2<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 72<br />
Conference in Numerical Analysis 2014 (NumAn 2014)<br />
September 2-5, 2014<br />
Chania,Greece<br />
The Fokas Method and Initial-Boundary Value Problems<br />
for Multidimensional Integrable PDEs<br />
Iasonas Hitzazis<br />
Department of Mathematics, University of Patras,<br />
Rion, Greece<br />
hitzazis@math.upatras.gr<br />
Abstract<br />
The Fokas Method, or Unified Transform Method, introduced in (A. S. Fokas, Proc. Roy.<br />
Soc. Lond. A 453 (1997), 1411-1443), is the appropriate generalization of the classical inverse<br />
scattering method which renders it applicable to the far more rich context of initial-boundary<br />
value problems (IBVPs) for integrable evolution PDEs in one spatial dimension.<br />
It is, however, interesting that there do also exist physically significant evolution equations<br />
in 2 or more spatial dimensions which share the property of integrability, i.e. that of admitting<br />
a Lax pair formulation. The most well-known integrable nonlinear PDEs in 2+1 (2 spatial and<br />
1 temporal) dimensions are the so-called Davey-Stewartson (DS) and Kadomtsev-Petviashvili<br />
(KP) equations. Recently ( A. S. Fokas, Commun. Math. Phys. 289 (2009), 957-993) took<br />
the first step towards the extension of his method to the case of multidimensions. In particular,<br />
the problem treated therein was the IBVP for the DS equation - as well as for its linearized<br />
version - posed on the half-plane. Soon thereafter, the half-plane case of the KP equation was<br />
also analyzed.<br />
In the present work we attempt to generalize this methodology so as to cover cases of more<br />
general domains, again in the context of (2+1)-dimensional PDEs. The linearized version of<br />
the DS equation is used as a prototypical example. In particular, we analyze the following two<br />
problems:(i) the quarter-plane IBVP, and (ii) the IBVP in a rectangular domain, both problems<br />
under smooth, temporally-decaying, non-homogeneous boundary conditions.<br />
The approach is totally based on the Lax pair formulation of the given PDE, and thus is the<br />
first step towards the construction of a formalism for the nonlinear case, i.e., for the DS equation<br />
itself, in each one of the geometries (i) and (ii). It is shown how both two eigenvalue equations<br />
constituting the Lax pair can undergo a simultaneous spectral analysis associated to any of the<br />
given domains (i) and (ii). Thus, in any one of the two cases, we achieve an appropriate d-bar<br />
problem for a sectionally non-analytic (in fact, sectionally generalized-analytic) function, i.e.,<br />
for a function that has different generalized-analytic representations in different regions of the<br />
complex plane.<br />
In the presentation, if time permits, we will also briefly refer to the nonlinear case, i.e., to<br />
the DS equation itself in each one of the geometries (i) and (ii).<br />
Key words: Lax pairs, initial-boundary value problems, multidimensions.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book Conference of Abstracts in Numerical Analysis 2014 (NumAn 2014) 73<br />
September 2-5, 2014<br />
Chania,Greece<br />
Some new perturbation bounds of generalized polar<br />
decomposition<br />
X.-L. Hong, L.-S. Meng and B. Zheng<br />
School of Mathematics and Statistics, Lanzhou University<br />
Lanzhou, Gansu Province, P.R.China<br />
hongxiaoli2007@163.com,menglsh07@lzu.edu.cn,bzheng@lzu.edu.cn<br />
Abstract<br />
Let A, Ã = A + E ∈ Cm×n have the (generalized) polar decompositions<br />
A = QH and à = ˜Q ˜H, (1)<br />
where Q is subunitary and H is Hermitian positive semi-definite. We present the following<br />
new bounds of the positive (semi-)definite polar factor and the (sub) unitary polar factor for<br />
the (generalized) polar decomposition under the general unitarily invariant norm ∥ · ∥ and the<br />
spectral norm ∥ · ∥ 2 , which are stated as in the following theorem.<br />
Theorem. Let A, Ã = A + E ∈ Cm×n r have the (generalized) polar decompositions in (1).<br />
(1). When r and N ∈ C n×n<br />
> . Hence, all perturbation bounds in the above theorem<br />
can be naturally extended to the case of the weighted polar decomposition of A, which also<br />
improved the known perturbation bounds for the weighted polar decomposition.<br />
Key words: Perturbation bounds; Positive semi-definite polar factor; Subunitary polar factor; Generalized<br />
polar decomposition; Weighted polar decomposition; Unitarily invariant norm; Spectral norm.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 74<br />
Conference in Numerical Analysis 2014 (NumAn 2014)<br />
September 2-5, 2014<br />
Chania,Greece<br />
On block preconditioners for PDE-constrained optimization<br />
problems<br />
Yu-Mei Huang a and Xiao-Ying Zhang a<br />
a School of Mathematics and Statistics, Lanzhou University,<br />
Lanzhou 730000, Gansu Province, P.R. China<br />
huangym@lzu.edu.cn, aiqian21921@163.com<br />
Abstract<br />
Recently, Bai proposed a block-counter-diagonal and a block-counter-triangular preconditioning<br />
matrices to precondition the GMRES method for solving the structured system of linear<br />
equations arising from the Galerkin finite-element discretizations of the distributed control problems<br />
in (Computing 91 (2011) 379-395). He analyzed the spectral properties and derived explicit<br />
expressions of the eigenvalues and eigenvectors for the preconditioned matrices. By applying<br />
the special structures and properties of the eigenvector matrices of the preconditioned matrices,<br />
we derive upper bounds for the 2-norm condition numbers of the eigenvector matrices and give<br />
the asymptotic convergence factors for the preconditioned GMRES methods with the blockcounter-diagonal<br />
and the block-counter-triangular preconditioners. Experimental results show<br />
that the convergence analyses match well with the numerical results.<br />
Key words: PDE-constrained optimization, the GMRES method, preconditioner, condition number, asymptotic<br />
convergence factor.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 75<br />
Conference in Numerical Analysis 2014 (NumAn 2014)<br />
September 2-5, 2014<br />
Chania,Greece<br />
Modeling drug release kinetics<br />
George Kalosakas<br />
University of Patras, Materials Science Dept., Rio Gr-26504, Greece, &<br />
Crete Center for Quantum Complexity and Nanotechnology, Physics Dept., Univ.<br />
of Crete, Greece<br />
georgek@upatras.gr<br />
Abstract<br />
We numerically calculate drug release profiles from simple or composite spherical devices, as<br />
well as from slabs of inhomogeneous thickness, using Monte Carlo simulations, when diffusion<br />
is the dominant release mechanism. In the case of spherical matrices the numerical results are<br />
compared with analytical solutions of Ficks second law of diffusion.<br />
Release curves are accurately described in all cases by the stretched exponential function.<br />
The dependence of the two stretched exponential parameters on the device characteristics is<br />
investigated and simple analytical relations are provided. Release kinetics does not depend on<br />
the initial drug concentration.<br />
We discuss in detail the numerical procedure followed in the Monte Carlo simulations. Particular<br />
emphasis is given in the techniques used to simulate regions of different drug diffusion<br />
coefficients when composite spheres are considered, as well as to construct slabs with inhomogeneous<br />
boundaries.<br />
Key words: drug release, diffusion, Monte Carlo simulations<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 76<br />
Conference in Numerical Analysis 2014 (NumAn 2014)<br />
September 2-5, 2014<br />
Chania,Greece<br />
Granular Transport Dynamics: Numerics and Analysis<br />
Giorgos Kanellopoulos and Ko van der Weele<br />
Center of Research and Applications of Nonlinear Systems (CRANS),<br />
Department of Mathematics, University of Patras,<br />
Rio, 26500, Greece<br />
kanellop@master.math.upatas.gr, weele@math.upatras.gr<br />
Abstract<br />
We study numerically the transport of granular matter on a compartmentalized conveyor belt,<br />
being a representative model for numerous applications both in industry and the natural environment,<br />
and a prime example of an open multi-particle system prone to spontaneous pattern<br />
formation. When the inflow rate exceeds a certain critical threshold value, a cluster is formed<br />
at the entrance of the conveyor belt and the flow is obstructed. This behavior can be understood<br />
by a dynamical flux model, in which the flow from one compartment to the next is described<br />
by a flux function. We show how the detailed form of the flux function can be reconstructed<br />
from Molecular Dynamics simulations, using a least-squares method. We then investigate the<br />
relation between the form of the flux function and the precise way in which the transition from<br />
free flow to the clustered state takes place. In particular, we find that – depending on the reconstructed<br />
parameter values – this transition can either take place via a reverse or a forward period<br />
doubling bifurcation.<br />
Key words: Clustering, flux function analysis, molecular dynamics simulations<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 77<br />
Quantitative evaluation of SRT for PET imaging:<br />
Comparison with FBP and OSEM<br />
George A Kastis a , Anastasios Gaitanis b and Athanasios S Fokas a,c<br />
a Research Center of Mathematics, Academy of Athens,<br />
Soranou Efessiou 4, Athens 11527, Greece<br />
b Biomedical Research Foundation of the Academy of Athens (BRFAA),<br />
Soranou Efessiou 4, Athens 11527, Greece<br />
c Department of Applied Mathematics and Theoretical Physics, University of<br />
Cambridge, Cambridge, CB30WA, UK<br />
gkastis@academyofathens.gr,agaitanis@bioacademy.gr,t.fokas@damtp.cam.ac.uk<br />
Abstract<br />
SRT is a new, fast, algorithm for PET imaging based on an analytic formula for the inverse<br />
Radon transform. Its mathematical formulation involves the numerical evaluation of the Hilbert<br />
transform of the sinogram via an approximation in terms of ‘custom made’ cubic splines.<br />
Here, we present a comparison between SRT, filtered backprojection (FBP) and ordered-subsets<br />
expectation-maximization (OSEM) with 21 subsets at various iteration numbers (1, 2, 4, 6 and<br />
10) using simulated and real PET projection data.<br />
For the simulation studies, we have simulated sinograms of an image quality (IQ) phantom<br />
and a Hoffman phantom with an implanted tumor of various tumor-to-background ratios. Using<br />
these sinograms, we have created realizations of Poisson noise at five noise levels. In addition to<br />
visual comparisons of the reconstructed images, we have determined contrast, bias and radioactivity<br />
concentration ratios for different regions of the phantoms as a function of noise level. For<br />
the real-data studies, sinograms of an IQ phantom has been acquired from a commercial PET<br />
system. We have determined RCR and contrast for the various lesions of the IQ phantom.<br />
In all simulated phantoms, the SRT exhibits higher contrast and lower bias than FBP and<br />
OSEM at 2 iterations (clinical protocol) at all noise levels. The contrast and bias of OSEM<br />
approach the values of SRT after 6 iterations. However, the SRT reconstructions exhibit higher<br />
coefficient of variations (COV). In real studies, SRT exhibits better contrast and RCR in all<br />
spheres over both FBP and OSEM at the clinical protocol of 21 subsets and 2 iterations. This<br />
improvement increases as the diameter of the relevant spheres in the phantom decrease.<br />
In conclusion, SRT is an analytical algorithm with clearly improved quantification characteristics<br />
over FBP and the clinical protocol of OSEM. Since SRT increases the noise in the<br />
reconstructed image, further investigations are needed to determine appropriate applications for<br />
the algorithm.<br />
Key words: Image reconstruction-analytical methods, PET, filtered backprojection, ordered-subsets expectationmaximization,<br />
OSEM.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 78<br />
A wave breaking mechanism for an unstructured finite volume<br />
scheme<br />
Maria Kazolea a , Argiris I. Delis b and Costas E. Synolakis a<br />
a School of Environmental Engineering, Technical University of Crete, Chania,<br />
Crete, Grece<br />
b School of Production Engineering and Management, Technical University of<br />
Crete, Chania, Crete, Greece<br />
mkazolea@isc.tuc.gr, adelis@science.tuc.gr, costas@usc.edu<br />
Abstract<br />
Wave breaking is a natural phenomenon of fundamental significance in the near-shore and one<br />
of the most important issues once have to consider in the numerical modeling of non-linear wave<br />
transformations. In this work a new methodology is presented and incorporated to TUCWave<br />
code, as to handle wave breaking over complex bathymetries in extended two-dimensional<br />
Boussinesq-type (BT) models. In the TUCWave code the 2D BT equations of Nwogu(1993),<br />
are solved using a novel high-order well balancing finite volume (FV) numerical method in unstructured<br />
meshes following the median dual node-centered approach. The novel wave breaking<br />
mechanism is of a hybrid type and consists of to parts. We first estimate the location of breaking<br />
waves using certain explicit criteria. Once breaking waves are recognized we switch locally<br />
in the computational domain from BT equations to the Non-linear Shallow Water Equations<br />
(NSWE) by suppressing the dispersive terms in the vicinity of the wave fronts. An additional<br />
methodology is presented on how to perform a stable switching between the BT and the NSWE<br />
equations within the unstructured FV framework. Comparison with laboratory data reveals that<br />
the proposed mechanism can accurately predict wave’s breaking position along with wave’s<br />
height decay and mean water level for both regular and solitary waves propagation on sloping<br />
beaches and submerged shoals.<br />
Key words: wave breaking, unstructured, Boussinesq-type equations<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book Conference of Abstracts in Numerical Analysis 2014 (NumAn 2014) 79<br />
September 2-5, 2014<br />
Chania,Greece<br />
Numerical Simulation of Flow Separation Control using<br />
Dielectric Barrier Discharge plasma actuator<br />
R. Khoshkhoo and A. Jahangirian<br />
Aerospace Engineering Department Amirkabir University of Technology, Tehran,<br />
Iran.<br />
r.khoshkhoo@aut.ac.ir and ajahan@aut.ac.ir<br />
Abstract<br />
A numerical simulation method is employed to investigate the effect of the steady plasma body<br />
force over the stalled NACA 0015 airfoil on flow field at low Reynolds number flow condition.<br />
The plasma body force created by a Dielectric Barrier Discharge (DBD) actuator modeled with<br />
a phenomenological plasma method is coupled with 2-dimensional compressible Navier- Stokes<br />
equations. The body force distribution is assumed to vary linearly in the triangular region, and<br />
the body force decreases by going away from the surface.The equations are solved using an<br />
implicit finite volume method on unstructured grids.The responses of the separated flow field<br />
to the effects of a steady body force in various angles of attack are studied; also the effect of<br />
single- and multi-actuator and the positioning of the actuator on the flow field are investigated.<br />
It is shown that the DBD have significant effect on flow separation control in low Reynolds<br />
number aerodynamics.<br />
Key words: Flow control, Plasma actuator, Numerical simulation, Low Reynolds number.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book Conference of Abstracts in Numerical Analysis 2014 (NumAn 2014) 80<br />
September 2-5, 2014<br />
Chania,Greece<br />
An effective approach on finite-difference-time-domain method<br />
for quasi-static electromagnetic field analysis<br />
Minhyuk Kim a , Hyun-Kyo Jung a and SangWook Park b<br />
a Department of Electrical and Computer Engineering, Seoul National University,<br />
Seoul, Korea<br />
b ICT Convergence Research Team, EMI/EMC R&D Center, Corporation Support<br />
& Reliability Division, Korea Automotive Technology Institute,<br />
Chon-Ahn, Korea<br />
ejnp@snu.ac.kr, hkjung@snu.ac.kr, parksw@katech.re.kr<br />
Abstract<br />
This paper deals with an effective computational electromagnetic numerical method for the<br />
quasi-static field problems. There are lots of numerical technique to simulate electromagnetic<br />
problems. Among them, the finite-difference-time-domain (FDTD) is a popular method to analyze<br />
huge computational complexity problems. However, it is realistically impossible to apply<br />
directly standard FDTD method to the near-field analysis under a few MHz because of the time<br />
step problem. We overcome this by approximating the current source to have quasi-static behavior<br />
on the arbitrary surface at first. Then, the surfaces are employed in the FDTD method as the<br />
source excitation by the surface equivalence theorem. The time consuming computation problems<br />
are treated efficiently and the results of our method are in good agreement with full-wave<br />
analysis electromagnetic commercial solver.<br />
Key words: Finite-difference-time-domain, quasi-static electromagnetic field, surface equivalence theorem.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 81<br />
<br />
Domain decomposition method with complete radiation<br />
boundary conditions for the Helmholtz equation in waveguides<br />
Seungil Kim a and Hui Zhang b<br />
a Department of Mathematics, Kyung Hee University,<br />
Seoul, South Korea<br />
b Section of Mathematics, University of Geneva,<br />
Geneva, Switzerland<br />
sikim@khu.ac.kr, mike.hui.zhang@hotmail.com<br />
Abstract<br />
In this paper, we present a nonoverlapping domain decomposition algorithm with a high-order<br />
transmission condition for the Helmholtz equation posed in a waveguide. We introduce the<br />
new high-order transmission conditions based on the complete radiation boundary conditions<br />
(CRBCs) that have been developed for high-order absorbing boundary conditions [3, 4]. We<br />
verify the rapid convergence of the Schwarz algorithm in terms of the order of CRBCs. It will<br />
be shown that damping parameters involved in the transmission conditions can be selected in an<br />
optimal way for enhancing the convergence of the Schwarz algorithm. This algorithm can also<br />
be employed efficiently for a preconditioner in GMRES implementations as recently developed<br />
sweeping preconditioners [1, 2, 5]. Finally, numerical examples confirming the theory will be<br />
presented.<br />
Key words: Helmholtz equation, domain decomposition, complete radiation boundary condition.<br />
References<br />
[1] Z. Chen and X. Xiang. A source transfer domain decomposition method for Helmholtz equations<br />
in unbounded domain. SIAM J. Numer. Anal., 51(4):2331–2356, 2013.<br />
[2] B. Engquist and L. Ying. Sweeping preconditioner for the Helmholtz equation: moving perfectly<br />
matched layers. Multiscale Model. Simul., 9(2):686–710, 2011.<br />
[3] T. Hagstrom and T. Warburton. Complete radiation boundary conditions: minimizing the long<br />
time error growth of local methods. SIAM J. Numer. Anal., 47(5):3678–3704, 2009.<br />
[4] S. Kim. Analysis of the convected Helmholtz equation with a uniform mean flow in a waveguide<br />
with complete radiation boundary conditions. J. Math. Anal. Appl., 410(1):275–291, 2014.<br />
[5] C. C. Stolk. A rapidly converging domain decomposition method for the helmholtz equation.<br />
Journal of Computational Physics, 241(0):240 – 252, 2013.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 82<br />
Conference in Numerical Analysis 2014 (NumAn 2014)<br />
September 2–5, 2014<br />
Chania, Greece<br />
Generalizations and Modifications of Iterative Methods for<br />
Solving Large Sparse Indefinite Linear Systems<br />
David R. Kincaid a , Jen-Yuan Chen b , and Yu-Chien Li b<br />
a Institute for Computational Engineering and Sciences,<br />
The University of Texas at Austin, Austin, Texas 78712 USA<br />
b Department of Mathematics, National Kaohsiung Normal University,<br />
Kaohsiung, TAIWAN,<br />
kincaid@cs.utexas.edu jchen@nknucc.nknu.edu.tw<br />
Abstract<br />
An overview of generalizations and modifications of iterative methods for solving large sparse<br />
indefinite linear systems with both symmetric and nonsymmetric coefficients matrices.<br />
Keywords: iterative methods, large sparse indefinite linear systems, generalizations and modifications of<br />
iterative methods, Arnoldi process, GMRES, SYMMLQ/SYMMQR, Lanczos SYMMLQ/SYMMQR.<br />
Frequently, when computing numerical solutions of partial differential equations, we need to<br />
solve systems of very large sparse linear algebraic equations of the form<br />
Ax = b<br />
where A is a given n × n matrix, b the given righthand side vector, and we seek a numerical<br />
solution vector x or a good approximation of it. Particularly for large linear systems arising from<br />
partial differential equations in three dimension, well-known direct methods based on Gaussian<br />
elimination may become prohibitively expensive in terms of both computer memory and computer<br />
time. On the other hand, iterative methods may avoid these difficulties.<br />
While the conjugate gradient (CG) method (and variations of it) may work well, for linear systems<br />
with a symmetric positive definite (SPD) coefficient matrices A, the choice of a suitable iterative<br />
method is not at all clear, when the linear system has a symmetric indefinite coefficient matrix.<br />
We discuss a variety of iterative methods that are based on the Arnoldi Process for solving large<br />
sparse symmetric indefinite linear systems. We describe the SYMMLQ and SYMMQR methods,<br />
as well as, generalizations and modifications of them. Then, we cover the Lanczos/MSYMMLQ<br />
and Lanczos/MSYMMQR methods, which arise from a double linear system. We present some<br />
pseudocodes for these algorithms.<br />
Finally, we mention some additional generalizations and modifications of iterative methods for<br />
solving large sparse symmetric and nonsymmetric indefinite systems of linear equations such as<br />
GMRES, MGMRES, MINRES, LQ-MINRES, QR-MINRES, MMINRES, and others.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 83<br />
An energy market stackelberg game solved with particle swarm<br />
optimization<br />
Panagiotis Kontogiorgos, 1 Elena Sarri, 1 Michael N. Vrahatis, 2 George P.<br />
Papavassilopoulos 1<br />
1 Department of Electrical and Computer Engineering, National Technical<br />
University of Athens, Athens, Greece,<br />
2 Department of Mathematics, University of Patras, GR-26110 Patras, Greece,<br />
panko09@hotmail.com, elena@netmode.ntua.gr, vrahatis@math.upatras.gr,<br />
yorgos@netmode.ece.ntua.gr<br />
Abstract<br />
Complex interactions between stakeholders in deregulated markets are formulated using game<br />
theory notions. This study is motivated by energy markets and addresses Stackelberg games with<br />
a leader that decides first his strategy and many followers, each one with his own characteristics.<br />
A static Stackelberg game corresponding to a Voluntary Load Curtailment (VLC) program for<br />
energy consumers is formulated. This leads to a bilevel programming problem that is generally<br />
difficult to solve, due to nonlinearities, nonconvexities that arise and the large dimensionality of<br />
the problem due to the existence of many followers. In these problems metaheuristic algorithms<br />
become attractive. In the present study an algorithm for solving such problems is developed,<br />
using Particle Swarm Optimization (PSO), which is based on collective intelligent behaviors in<br />
nature and has gained wide recognition in recent years. Some examples are then solved using<br />
the proposed algorithm in order to evaluate its efficiency and examine the interactions between<br />
the players of the game.<br />
Key words: Complex Systems, Energy Market, Stackelberg, Particle Swarm Optimization<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 84<br />
Serial and Parallel Implementation of the<br />
Interface Relaxation Method GEO 1<br />
Aigli Korfiati a , Panagiota Tsompanopoulou b and Spiros Likothanassis a<br />
a Department of Computer Engineering and Informatics,<br />
University of Patras, Patras, Greece<br />
b Department of Computer and Communications Engineering,<br />
University of Thessaly, Volos, Greece<br />
korfiati@ceid.upatras.gr, yota@inf.uth.gr,<br />
likothan@ceid.upatras.gr<br />
Abstract<br />
Interface Relaxation (IR) methods are an interesting approach for the solution of multiphysics /<br />
multidomain problems. Assuming initial guesses on the interfaces of the original problem, IR<br />
methods iteratively solve the subproblems and relax for new values on the interfaces until convergence<br />
is succeed. Their main advantages are that their rates of convergence only depend on<br />
the parameters of the problem itself, the parameters related to its decomposition into subproblems<br />
and the parameters related to the operator imposed on the interfaces. In this paper a new<br />
implementation of an IR method named GEO is presented. GEO is based on a simple geometric<br />
coorrection mechanism and acts iteratively so as to relax the values of the solution on the interfaces.<br />
In particular, it adds to the old interface values a geometrically weighted combination of<br />
the normal boundary derivatives of the adjacent subdomains.<br />
In this paper GEO is implemented in FEniCS. The FEniCS project is a collection of free software<br />
for automated, efficient solution of differential equations. In order to evaluate the GEO<br />
implementation, it is applied on two different PDE problems with the same differential equation<br />
and boundary conditions and different domains. FEniCS methods are used to specify the problem’s<br />
subdomains properties (i.e. geometry, PDE operator and boundary/interface conditions).<br />
They are also used to generate and/or refine meshes (triangular elements) for each subdomain,<br />
solve the local PDE problems and show the computed results in the global domain and on the<br />
interfaces. Getting values of the solutions on the interface (boundaries of the subproblems) and<br />
passing the new relaxed values back to the subproblems as updated values for the boundary<br />
conditions is the main challenge of the IR methodology implementation and contribution of this<br />
paper.<br />
The experiments are performed for 2-dimensional elliptic partial differential model problems<br />
with partitions in multiple subdomains and the results are examined in terms of the method’s<br />
applicability and convergence. The exact solution and the computed approximations on the<br />
whole domain and on interface points, are depicted per iteration in appropriate graphs for applicability<br />
and convergence evaluation. A parallel implementation of the GEO method using<br />
FEniCS is also presented, as well as its performance comparison to the serial implementation.<br />
Key words: Interface relaxation, GEO method, multiphysics problems, parallel implementation, FEniCS.<br />
1 The present research work has been co-financed by the European Union (European Social Fund ESF) and Greek<br />
national funds through the Operational Program Education and Lifelong Learning of the National Strategic Reference<br />
Framework (NSRF) - Research Funding Program: THALIS. Investing in knowledge society through the European<br />
Social Fund.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 85<br />
Conference in Numerical Analysis 2014 (NumAn 2014)<br />
September 2-5, 2014<br />
Chania,Greece<br />
A special class of integrable Lotka-Voltera systems and their<br />
Kahan discretization<br />
Theodoros Kouloukas<br />
Department of Mathematics, La Trobe University, Melbourne VIC 3086, Australia<br />
T.Kouloukas@latrobe.edu.au<br />
Abstract<br />
We present a family of integrable systems associated with a special set of polynomials z (n)<br />
i .<br />
The quadratic vector fields associated with z (n)<br />
3 are closely related to a class of Lotka-Voltera<br />
systems. We prove that they are superintegrable when n is odd and non-commutative integrable<br />
(of rank 2) when n is even. We also apply the Kahan-Hirota-Kimura discretization (a special<br />
Runge-Kutta method) to these quadratic vector fields, restricted to a subspace, and show that<br />
Liouville integrability and superintegrability are preserved. In the more general case of full,<br />
non-restricted, quadratic vector fields, numerical computations indicate integrability as well<br />
and therefore generate new questions for further investigation.<br />
Key words: Integrable systems, Kahan-Hirota-Kimura discretization<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book Conference of Abstracts in Numerical Analysis 2014 (NumAn 2014) 86<br />
September 2-5, 2014<br />
Chania,Greece<br />
Constraint handling for gradient-based optimization of<br />
compositional reservoir flow<br />
Drosos Kourounis a<br />
a Institute of Computational Science, Università della Svizzera italiana<br />
CH-6904 Lugano, Switzerland drosos.kourounis@usi.ch<br />
Abstract<br />
The development of adjoint gradient-based optimization techniques for general compositional<br />
flow problems is much more challenging than for oil-water problems due to the increased complexity<br />
of the code and the underlying physics. An additional challenge is the treatment of non<br />
smooth constraints, an example of which is a maximum gas rate specification in injection or<br />
production wells, when the control variables are well bottom-hole pressures. Constraint handling<br />
through lumping is a popular and efficient approach. It introduces a smooth function that<br />
approximates the maximum of the specified constraints over the entire model or on a well-bywell<br />
basis. However, it inevitably restricts the possible solution paths the optimizer may follow<br />
preventing it to converge to feasible solutions exhibiting higher optimal values. A simpler way<br />
to force feasibility, when the constraints are upper and lower bounds on output quantities, is to<br />
satisfy these constraints in the forward model. This heuristic treatment has been demonstrated<br />
to be more efficient than lumping and at the same time it obtained better feasible optimal solutions<br />
for several models of increased complexity. In this work a new formal constraint handling<br />
approach is presented. Necessary modifications of the nonlinear solver used at every timestep<br />
during the forward simulation are also discussed. All these constrained handling approaches are<br />
applied in a gradient-based optimization framework for exploring optimal CO2 injection strategies<br />
that enhance oil recovery for a realistic offshore field, the Norne field. The new approach<br />
recovers 4% more oil than the best of the optimal solutions obtain by its competitors.<br />
Key words: Constraint-handling, production optimization, recovery-optimization, discrete adjoint.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 87<br />
Combining Discontinuous Galerkin and Finite Differences<br />
Methods for Simulation of Seismic Wave Propagation<br />
Vadim Lisitsa a and Vladimir Tcheverda a<br />
a Institute of Petroleum Geology and Geophysics of SB RAS,<br />
Novosibirsk, Russia<br />
lisitsavv@ipgg.sbras.ru,tcheverdava@ipgg.sbras.ru<br />
Abstract<br />
In this paper we present an original approach to combination of standard staggered grid finitedifference<br />
scheme with interior penalty discontinuous Galerkin (DG) method for simulation of<br />
seismic wave propagation in presence of sharp interfaces with complex topography. The approach<br />
takes the advantages of the two numerical techniques. Discontinuous Galerkin methods<br />
is applied in the vicinity of the free-surface or sea-bed ensuring accurate approximation of the<br />
surface by the triangular (tetrahedral) mesh. However, DG is typically more computationally<br />
intense than finite differences (FD). So, in the major part of the model the FD are applied to<br />
reduce the overall computational cost of the algorithm. As the result the designed approach<br />
combines high accuracy of the DG with computational intensity of the FD.<br />
The idea of the approach is to combine the two approaches via transition zone where P0 discontinuous<br />
Galerkin method on a regular rectangular grid is used. This formulation is equivalent<br />
to the conventional (nonstaggered grid) scheme, approximating the elastic wave equation with<br />
a second order. So, the coupling of FD with the DG on arbitrary triangular mesh is reduced to<br />
the two independent problems. First, standard staggered grid scheme should be combined with<br />
a conventional scheme. This is done on the base of approximation of the reflection coefficients,<br />
either physical and artificial (corresponding to the spurious modes). Second, the DG on an arbitrary<br />
mesh is coupled with DG on a regular rectangular grid (conventional finite difference<br />
scheme). This is done on the base of hp-adaptivity of the DG method.<br />
Key words: discontinuous Galerkin, finite differences, elastic wave quation.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 88<br />
Decreasing Computational Load by Using Similarity for<br />
Lagrangian Approach to Gas-solid Two-phase Flow<br />
Zhihong Liu a , Yoshiyuki Yamane a , Takuya Tsuji b and Toshitsugu Tanaka b<br />
a Heat and Fluid Dynamics Department, Research Laboratory,<br />
IHI Corporation, Yokohama,235-8501 Japan<br />
b Department of Mechanical Engineering, Osaka University,<br />
Osaka, 565-0871 Japan<br />
shikou ryuu@ihi.co.jp,yoshiyuki yamane@ihi.co.jp<br />
tak@mech.eng.osaka-u.ac.jp,tanaka@mech.eng.osaka-u.ac.jp<br />
Abstract<br />
Gas-solid two-phase flow constitutes a continuous-discrete phase system. It is natural to employ<br />
the Lagrangian approach to numerically analyze the discrete phase (solid particles); However<br />
the computational load becomes very heavy, when the number of particles increases. In this paper,<br />
we employ similarity rules to decrease the computational load. An imaginary system with<br />
imaginary gas and particles is used to describe the real gas-particle system. Each imaginary<br />
particle with a diameter K times of that of the real particle, replaces a group of real particles.<br />
By dimensionless analysis of the conservation equations of gas-solid two- phase flow, it is show<br />
that the solution of the imaginary system is similar to that of the real system, if the physical<br />
properties of the imaginary system are adjusted such that Reynolds number Re and Archimedes<br />
number Ar euqal those of the real system. Enlarging the imaginary particles by a factor K can<br />
decrease the number of particles and hence the computational load by a factor of K −4.5 .<br />
In order to validate the similarity rules, the behavior of a bubble in a fluidized bed is simulated<br />
with various factors K. It is shown that the similarities show reasonable accuracy.<br />
Nomenclature<br />
Re = |V − U|ρ f εD p<br />
µ f<br />
Ar = D3 pρ f (ρ p − ρ f )g<br />
µ 2 f<br />
D p particle diameter ε void fraction<br />
g gravity acceleration ρ f gas density<br />
U particle velocity ρ p particle density<br />
V gas velocity µ f gas viscosity<br />
Key words: Similarity, Lagrangian approach, computation load, two-phase flow simulation<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 89<br />
Curvilinear Grids for Five-Axis Machining<br />
Stanislav Makhanov<br />
School of Information and Computer Technology, Sirindhorn International<br />
Institute of Technology, Thammasat University, Thailand<br />
makhanov@siit.tu.ac.th<br />
Abstract<br />
Machining large complex industrial parts with a high accuracy often requires tens, hundreds of<br />
thousands or even millions of cutter location points and hundreds hours of machining. That<br />
is why reducing the machining time is one of the most important topics in the optimization of<br />
CNC codes for five axis milling machines.<br />
We propose and analyze a new method of constructing curvilinear tool paths which partly or<br />
even entirely align with the direction of the maximum material removal rate. The alignment<br />
based on the curvilinear elliptic grid generation allows to minimize the machining time while<br />
keeping the convenient zigzag-like topology of the path. The method is applicable to a variety of<br />
cost functions such as the length of the path, the machining speed, the material removal rate, the<br />
kinematic error, etc., generating different machining strategies. The method has been combined<br />
with a new version of the adaptive space filling curves.<br />
The approach has been tested against the standard iso-parametric zigzag, MasterCam X5 and<br />
the conventional space filling curves. The material removal rate cost function has been tested<br />
against the tool path length criteria. The numerical experiments, the real machining as well the<br />
accuracy measurements demonstrate a considerable advantage of the proposed method.<br />
Key words: numerical grid generation, milling machine, error minimization, tool path planning.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 90<br />
Numerical Solution Of Optimization Problems for Semilinear<br />
Elliptic Equations with Discontinuous Coefficients and Solutions<br />
Aigul Manapova a , Fedor Lubyshev b<br />
a Department of Mathematics and IT, Bashkir State University,<br />
Ufa, Republic of Bashkortostan, Russian Federation<br />
b Department of Mathematics and IT, Bashkir State University,<br />
Ufa, Republic of Bashkortostan, Russian Federation<br />
aygulrm@yahoo.com<br />
Abstract<br />
Mathematical models of optimization for systems with distributed parameters (described by<br />
equations of mathematical physics) are the most difficult class of problems in optimization,<br />
especially for nonlinear optimal control problems. By ”non-linear optimization problems” for<br />
equations of mathematical physics we understand those in which the mapping g → u(g) from<br />
the set of admissible controls U to the space of states W is a nonlinear. Particular formulations<br />
of optimization problems for distributed parameter systems depend substantially on whether the<br />
controls enter into the free terms of the equations of state or in the equation coefficients and on<br />
whether linear or nonlinear PDEs describe the states of the control systems. Linear control systems<br />
with sufficiently smooth input data and control state functions have been most thoroughly<br />
studied and nonlinear optimization problem have been least studied to this day. Especially interesting<br />
from theoretical and practical points of view are optimal control problems in which the<br />
states are described by nonlinear PDEs with discontinuous coefficients and the solutions can be<br />
discontinuous due to the character of the physics process under study.<br />
Before solving optimal control problems numerically, they have to be approximated by problems<br />
of a simpler nature, specifically, by ”finite-dimensional problems”. One of the most convenient<br />
and universal techniques for finite-dimensional approximation as applied to optimization problems<br />
is the grid method. Also relevant is the question of the development of efficient numerical<br />
methods for solving constructed finite-dimensional grid optimal control problems, which requires<br />
effective procedures for calculating the gradient of the minimized functional.<br />
In this work we consider optimization problems for processes described by semilinear partial<br />
differential equations of elliptic type with discontinuous coefficients and solutions (with imperfect<br />
contact matching conditions), with controls involved in the coefficients. Finite difference<br />
approximations of optimization problems are constructed. For the numerical implementation of<br />
finite optimization problems differentiability and Lipshitz-continuity of the grid functional of<br />
the approximating grid problems are proved. Effective procedures for calculating gradients of<br />
minimized functionals using the solutions of direct problems for the state and adjoint problems<br />
are obtained.<br />
Key words: semilinear elliptic equations, optimization problem, numerical method.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 91<br />
Conference in Numerical Analysis 2014 (NumAn 2014)<br />
September 2-5, 2014<br />
Chania,Greece<br />
A MultiGrid accelerated high-order pressure correction<br />
compact scheme for incompressible Navier-Stokes solvers<br />
V.G. Mandikas a , E.N. Mathioudakis a , G.V. Kozyrakis b,c ,<br />
J.A. Ekaterinaris d and N.A. Kampanis b<br />
a Applied Mathematics and Computers Laboratory, Technical University of Crete,<br />
University Campus, 73132 Chania, Hellas<br />
b Institute of Applied and Computational Mathematics, Foundation for<br />
Research and Technology - Hellas, 70013 Heraklion, Crete, Hellas<br />
c Department of Marine Sciences, University of the Aegean,<br />
University Hill, Mytilene 81100, Hellas<br />
d Department of Aerospace Engineering, Daytona Beach College of Engineering,<br />
Embry-Riddle Aeronautical University, 600 S. Clyde Morris Blvd., Daytona<br />
Beach FL 32114, USA<br />
bmandikas@science.tuc.gr, manolis@amcl.tuc.gr<br />
gkoz@iacm.forth.gr, ekaterin@iacm.forth.gr<br />
kampanis@iacm.forth.gr<br />
Abstract<br />
A high-order accurate compact finite-difference numerical scheme, based on multigrid techniques,<br />
is constructed on staggered grids in order to develop an efficient incompressible Navier-<br />
Stokes solver. The enforcement of the incompressibility condition by solving a Poisson-type<br />
equation at each time step is commonly accepted to be the most computationally demanding<br />
part of the global pressure correction procedure of a numerical method. Since the efficiency<br />
of the overall algorithm depends on the Poisson solver, a multigrid acceleration technique coupled<br />
with compact high-order descretization scheme is implemented to accelerate the iterative<br />
procedure of the pressure updates and enhance computational efficiency. The employment of<br />
geometric multigrid techniques on staggered grids has an intrinsic difficulty, since the coarse<br />
grids do not constitute part of the finer grids. Appropriate boundary closure formulas are developed<br />
for the cell-centered pressure approximations of the boundary conditions. Performance<br />
investigations demonstrate that the proposed multigrid algorithm can significantly accelerate the<br />
numerical solution process, while retaining the high order of accuracy of the numerical method<br />
even for high Reynolds number flows.<br />
Key words: Global pressure correction, Poisson type equation, Incompressible Navier-Stokes equations,<br />
High-order compact schemes, staggered grids, Geometric MultiGrid techniques.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 92<br />
A Fourier Collocation Method for the Nonlocal Nonlinear Wave<br />
Equation<br />
Gulcin M. Muslu a and Handan Borluk b<br />
a Istanbul Technical University, Department of Mathematics,<br />
Maslak 34469, Istanbul, Turkey.<br />
b Isik University, Department of Mathematics,<br />
Sile 34980, Istanbul, Turkey.<br />
gulcin@itu.edu.tr, hborluk@isikun.edu.tr<br />
Abstract<br />
We consider a general class of nonlinear nonlocal wave equation arising in one-dimensional<br />
nonlocal elasticity [1]. The model involves a convolution operator with a general kernel function<br />
whose Fourier transform is nonnegative. We propose a Fourier collocation numerical method for<br />
the nonlinear nonlocal wave equation. We first test our scheme for some well-known examples<br />
of nonlinear nonlocal wave equation, such as Boussinesq-type equations which arise from the<br />
suitable choices of the kernel function. To understand the structural properties of the solutions<br />
of nonlocal nonlinear wave equation, we present some numerical results illustrating the effects<br />
of both the smoothness of the kernel function and the strength of the nonlinear term on the<br />
solutions.<br />
This work has been supported by the Scientific and Technological Research Council of Turkey<br />
(TUBITAK) under the project MFAG-113F114.<br />
References<br />
[1] N. Duruk, H. A. Erbay, A. Erkip, Global existence and blow-up for a class of nonlocal nonlinear Cauchy<br />
problems arising in elasticity, Nonlinearity 23, 107–118, (2010).<br />
Key words: Nonlocal nonlinear wave equation, Boussinesq-type equations, Solitary waves.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 93<br />
Conference in Numerical Analysis 2014 (NumAn 2014)<br />
September 2-5, 2014<br />
Chania,Greece<br />
Perturbation Theory of Dark-Bright solitons<br />
in Bose-Einstein condensates<br />
I.K.Mylonas a V.M.Rothos a , P.G.Kevrekidis b and D.J.Frantzeskakis c<br />
a Department of Mechanical Engineering, Faculty of Engineering Aristotle<br />
University of Thessaloniki GR54124 Thessaloniki, Greece<br />
b Department of Mathematics and Statistics, University of Massachusetts,<br />
Amherst, Massachusetts 01003-4515, USA<br />
c Department of Physics, University of Athens, Panepistimiopolis, Zografos,<br />
Athens 157 84, Greece<br />
imylon1986@gmail.com<br />
Abstract<br />
We develop a direct perturbation theory for a coupled dark-bright solitons and we derive the<br />
equation of motion for the soliton parameters. In this method, we solve the linearized wave<br />
equation around the solitons by expanding its solution into a set of complete eigenfunctions of<br />
the linearization operator. Suppression of secular growth in the linearized solution gives the<br />
evolution equations of soliton parameters. This method does not rely explicitly on the inverse<br />
scattering transform but its connection to the integrable theory is still visible since these eigenfunctions<br />
of the linearized equation are simply the squared eigenfunctions of the underlying<br />
scattering operator. Moreover, we study the stability of dark-bright solitons. Our analytical results<br />
for the small-amplitude oscillations of solitons is in good agreement with results obtained<br />
via a Bogoliubov-de Gennes analysis and compares very well with direct numerical computations.<br />
Key words: Solitons, Near-Integrable PDEs, BEC<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book Conference of Abstracts in Numerical Analysis 2014 (NumAn 2014) 94<br />
September 2-5, 2014<br />
Chania, Greece<br />
Efficient Unconstrained Optimization Multistart Solvers<br />
Using a Self-Clustering Technique<br />
Ioannis A. Nikas<br />
Department of Tourism Management,<br />
TEI of Western Greece,<br />
Patras, Greece<br />
nikas@teipat.gr<br />
Abstract<br />
One of the commonly occurring drawbacks in multistart solvers in unconstrained optimization<br />
problems is their inability to determine the quality and the quantity of local minima regions of<br />
attraction. Thus, when a sample of random points is generated it is not feasible the correspondence<br />
of each point to a single region of attraction, and consequently to a single local minimum.<br />
This results into a repeated application of a local search method on all sample points, finding,<br />
in general, multiple times the same local minima.<br />
In this work, motivated mostly on the above mentioned weakness of multistart algorithms<br />
and having in mind the local minimum definition, a new technique is proposed to correspond<br />
each sample point to a single candidate region of attraction. Specifically, each point of the<br />
sample is moved towards the best nearest neighbor point, which has the best functional value in<br />
this neighborhood. Through this process the sample points are concentrated around these best<br />
points, creating clusters that constitute candidate region of attractions. Then, it is assumed that<br />
each point inside a candidate region of attraction will drive a multistart algorithm to the same<br />
local minimum. For this reason, a new set of points is created. The new set will contain the best<br />
points inside each candidate region of attraction, namely the center of each self-clustered area,<br />
and these points will feed the multistart algorithm.<br />
It is noted that the number of created clusters depends on the sample size and the overall<br />
morphology of the objective function, that is the actual number of local minima. Furthermore,<br />
the proposed technique is a first-order process, that is, only functional values are necessary to<br />
determine the clusters and their corresponding centers.<br />
Finally, the proposed technique is utilized in a classic multistart solver, using a local search<br />
algorithm, and is tested on a set of well-known, one-dimensional test function. The results of<br />
these experiments show that in most cases the proposed technique makes the classic multistart<br />
algorithm efficient in finding uniquely many local minima. In addition, the experimental results<br />
showed that in most of the cases the global minimum is also found and as the grid size grows<br />
up, the number of clusters tends rapidly to the total number of local and global minima of the<br />
objective function.<br />
Key words: multistart algorithm, unconstrained optimization, self-clustered technique.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 95<br />
Conference in Numerical Analysis 2014 (NumAn 2014)<br />
September 2-5, 2014<br />
Chania,Greece<br />
Essential spectral equivalence via multiple step preconditioning<br />
and applications to ill conditioned Toeplitz matrices<br />
Dimitrios Noutsos a , Stefano Serra-Capizzano b and Paris Vassalos c<br />
a Department of Mathematics, University of Ioaninna,<br />
Ioannina, Greece<br />
b Department of Science and high Technology,<br />
University of Iunsubria, Como, Italy.<br />
c Department of Informatics, Athens University of Economics and Business,<br />
Athens,Greece.<br />
dnoutsos@uoi.gr, stefano.serrac@uninsubria.it, pvassal@aueb.gr<br />
Abstract<br />
We are concerned with the fast solution of Toeplitz linear systems with coefficient matrix T n (f),<br />
where the generating function f is nonnegative and has a unique zero at zero of any real positive<br />
order θ. As preconditioner we choose a matrix τ n (f) belonging to the so-called τ algebra,<br />
which is diagonalized by the sine transform associated to the discrete Laplacian. In previous<br />
works, the spectral equivalence of the matrix sequences {τ n (f)} n and {T n (f)} n was proven<br />
under the assumption that the order of the zero is equal to 2: in other words the preconditioned<br />
matrix sequence {τn<br />
−1 (f)T n (f)} n has eigenvalues, which are uniformly away from zero and<br />
from infinity. Here we prove a generalization of the above result when θ < 2. Furthermore,<br />
by making use of multiple step preconditioning, we show that the matrix sequences {τ n (f)} n<br />
and {T n (f)} n are essentially spectrally equivalent for every θ > 2, i.e., for every θ > 2,<br />
there exist m θ and a positive interval [α θ , β θ ] such that all the eigenvalues of {τn<br />
−1 (f)T n (f)} n<br />
belong to this interval, except at most m θ outliers larger than β θ . Such a nice property, already<br />
known only when θ is an even positive integer greater than 2, is coupled with the fact that the<br />
preconditioned sequence has an eigenvalue cluster at one, so that the convergence rate of the<br />
associated preconditioned conjugate gradient method is optimal. As a conclusion we discuss<br />
possible generalizations and we present selected numerical experiments.<br />
Key words: Toeplitz systems, τ algebra, preconditioning.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 96<br />
Nyström methods for two-dimensional<br />
Fredholm integral equations on unbounded domains<br />
Donatella Occorsio1 a , and Maria Grazia Russo a<br />
a Department of Mathematics, Computer Science and Economics,<br />
University of Basilicata, Potenza, Italy<br />
donatella.occorsio@unibas.it, mariagrazia.russo@unibas.it<br />
Abstract<br />
We investigate the numerical solution of two-dimensional Fredholm integral equations defined<br />
on the set S = [a, b] × [c, d], −∞ ≤ a < b ≤ ∞, −∞ ≤ c < d ≤ ∞,<br />
∫<br />
f(x, y) − µ k(x, y, s, t)f(s, t)w(s, t) ds dt = g(x, y), (x, y) ∈ S, (1)<br />
S<br />
where w(x, y) := w 1 (x)w 2 (y) and w 1 , w 2 are suitable weight functions defined on [a, b], [c, d]<br />
respectively, µ is a real number. k and g are given functions defined on ([a, b] × [c, d]) 2 and<br />
[a, b]×[c, d] respectively, which are sufficiently smooth on the open sets but can have (algebraic)<br />
singularities on the finite boundaries and an exponential growth at ±∞ at most w.r.t. each<br />
variable. f is the unknown function.<br />
Therefore S is intended to be an unbounded domain, for instance a quarter of the plane, a<br />
strip etc.<br />
We introduce some Nyström methods based on cubature formulas obtained as tensor products<br />
of two Gaussian quadrature formulas w.r.t. the weights w 1 , w 2 . Due to the “unboundedness”<br />
of the domain we need to “truncate” the quadrature rules. The convergence, stability and<br />
well conditioning of the methods are proved in suitable weighted spaces of continuous functions.<br />
Some numerical examples illustrate the efficiency of the methods.<br />
Key words: Fredholm integral equation, Nyström method, spectral methods<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 97<br />
Numerical stability of block direct methods for solving<br />
symmetric saddle point problem<br />
Felicja Okulicka-Dłużewska, Alicja Smoktunowicz<br />
Faculty of Mathematics and Information Science, Warsaw University of<br />
Technology, ul. Koszykowa 75, Warsaw, 00-662, Poland<br />
F.Okulicka@mini.pw.edu.pl, A.Smoktunowicz@mini.pw.edu.pl<br />
Abstract<br />
We study the numerical properties of some block direct methods for solving the following saddle<br />
point problem (quasidefinite case)<br />
( ) ( )<br />
A B x<br />
Mz = f ⇔<br />
B T −C y<br />
=<br />
(<br />
b<br />
c<br />
)<br />
, (1)<br />
where A ∈ R m×m , C ∈ R n×n are symmetric positive definite and B ∈ R m×n , n ≤ m.<br />
Then M is nonsingular and there is a unique solution (x ∗ , y ∗ ) of (??). Such problems arise in<br />
many applications, e.g., in optimization, in the solution of PDEs, weighted least squares (image<br />
restoration), FE formulations of consolidation problem. The matrices A and B are usually<br />
large, sparse and ill-conditioned. Structure of the problem leads to the application of block<br />
methods which operate on groups of columns of M and allow to apply the BLAS-3 compatible<br />
algorithms. It is known that the block LU methods are not numerically stable, in general. The<br />
block factorization<br />
( ) (<br />
A B<br />
M =<br />
B T =<br />
−C<br />
I 0<br />
B T A −1 I<br />
) ( A 0<br />
0 −(C + B T A −1 B)<br />
) ( I A −1 )<br />
B<br />
0 I<br />
is commonly used to solve the equation (??). We analyze the methods avoiding computing the<br />
Schur complement S = −(C + B T A −1 B). If C is ill-conditioned then the computed Schur<br />
complement S may be singular in working precision. We propose and analyze algorithms for<br />
solving symmetric saddle point problem which are based upon the block Cholesky decomposition<br />
and the block Gram-Schmidt method. In particular, we prove that the algorithm BCGS2<br />
(Reorthogonalized Block Classical Gram-Schmidt) using Householder Q-R decomposition implemented<br />
in the floating point arithmetic is backward stable, under a mild assumption on the<br />
matrix M.<br />
Extensive numerical testing was done in MATLAB to compare the performance of some<br />
direct methods for solving linear system of equations of special block matrices.<br />
Key words: symmetric quasidefinite (sqd) systems, saddle point problem, QR decomposition, numerical<br />
stability, condition number, iterative refinement.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 98<br />
Robust numerical simulation of reaction-diffusion models<br />
arising in Mathematical Ecology<br />
Kolade M. Owolabi a and Kailash C. Patidar b<br />
a Department of Mathematics and Applied Mathematics, University of the Western<br />
Cape, Private Bag X17, Bellville 7535, South Africa<br />
b Department of Mathematics and Applied Mathematics, University of the Western<br />
Cape, Private Bag X17, Bellville 7535, South Africa<br />
Speaker: Kailash C. Patidar, E-mail: kpatidar@uwc.ac.za<br />
Abstract<br />
In this work, we consider numerous types of reaction-diffusion models arising in mathematical<br />
ecology. Using local analysis theory applied to ecological modeling, we study four important<br />
ecological systems describing some prey-predator models. We address the interaction between<br />
two species in terms of predator-prey systems and the biological system displaying the formation<br />
of chaotic spatiotemporal patterns arising from a community of three competitive species.<br />
Then we design and analyze robust time-integration techniques to simulate these models. Two<br />
competing exponential time-differencing methods that are of order-four are used as the major<br />
time stepping methods. We justify the supremacy of these two schemes when applied to above<br />
mentioned dynamical systems and compared our results with those obtained by other existing<br />
multistep exponential integrators of orders four, five and six.<br />
Key words: Reaction-diffusion models, Mathematical Ecology, Exponential time-differencing methods.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book Conference of Abstracts in Numerical Analysis 2014 (NumAn 2014) 99<br />
September 2-5, 2014<br />
Chania,Greece<br />
Unified Tranforms and classical spectral theory of operators<br />
Beatrice Pelloni a and David A Smith b<br />
a Department of Mathematics, University of Reading,<br />
Reading, UK<br />
b Department of Mathematics, University of Cincinnati,<br />
Cincinnati, OH, USA<br />
b.pelloni@reading.ac.uk,david.smith2@uc.edu<br />
Abstract<br />
We describe how the application of the unified method of Fokas to the study of linear evolution<br />
PDEs sheds light on the spectral theory of non self-adjoint linear differential operators.<br />
Key words: unified transform, fokas method, spectral theory.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 100<br />
MATLAB : Parallel and Distributed Computing<br />
using CPUs and GPUs<br />
K. Petsounis<br />
Mentor Hellas Ltd<br />
Greece<br />
costas@mentorhellas.com<br />
Abstract<br />
MATLAB is a high level structured language and an interactive development environment for<br />
technical computing and algorithm development. It has enabled scientists and engineers to efficiently<br />
process and analyze data, develop and deploy algorithms and applications. Furthermore,<br />
Parallel and Distributed Computing capabilities in MATLAB, allow users to solve computationally<br />
and data intensive problems by taking advantage of the latest multiprocessing systems:<br />
multicore desktops, computer clusters, GPUs, grid and cloud computing services. It is now possible<br />
to interactively prototype and develop distributed and parallel applications, briefly touch<br />
upon the parallel data structures, such as distributed arrays, and programming constructs such<br />
as parallel for loops, parallel numeric algorithms and message passing functions. Using typical<br />
numerical computing problems as examples, this workshop describes how to use MATLAB<br />
parallel tools to take full advantage of the performance enhancements offered by multicore /<br />
multiprocessor computing environments. In addition, you will learn how you can leverage the<br />
computing power of NVIDIA CUDA-enabled GPUs to accelerate your MATLAB applications<br />
with minimal programming effort using GPU arrays and GPU enabled MATLAB functions.<br />
Sample codes, differences in CPU and GPU implementations as well as benchmark results for<br />
some typical numerical computing problems will be presented.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 101<br />
Method for solving nonlinear singular problems<br />
Agnieszka Prusińska a and Alexey A. Tretýakov b,c<br />
a Department of Mathematics and Physics, Siedlce University of Natural Sciences<br />
and Humanities, Siedlce, Poland<br />
b System Research Institute, Polish Academy of Sciences<br />
Warsaw, Poland<br />
c Computing Center, Russian Academy of Sciences,<br />
Moscow, Russia<br />
aprus@uph.edu.pl, tret@uph.edu.pl<br />
Abstract<br />
The aim of our work is to investigate conditions for existence of local solutions to nonlinear<br />
equations of the form<br />
F (x) = 0, F : R n → R m , (1)<br />
in degenerate case, i.e. when Im F ′ (x 0 ) ≠ R m and x 0 is a chosen initial point. We propose an<br />
algorithm of the numerical method that is convergent to a solution point in the mentioned case.<br />
In our approach we use p-factor operator and some elements of p-regularity theory [1]. Main<br />
result of this theory is a description of the tangent cone to the solution set in degenerate case.<br />
We assume that F is a p + 1 times differentiable mapping and for some h ∈ R n consider<br />
the sequence<br />
x k+1 = x k − Λ −1<br />
h (f 1(x 0 + ωh + x k ), . . . , f p (x 0 + ωh + x k )) , k = 1, 2, . . . , (2)<br />
(<br />
where 0 < ω < 1/2. An operator Λ h = f 1 ′(x 0), f 2 ′′(x<br />
1<br />
0)[h], . . . ,<br />
(p−1)! f p (p)<br />
)<br />
(x 0 )[h, . . . , h] for<br />
x ∈ R n is called p-factor operator. To construct this operator we decompose R m into a direct<br />
sum Y 1 ⊕· · ·⊕Y p and define auxiliary mappings f i : R n → Y i , f i (x) = P Yi F (x), i = 1, . . . , p,<br />
such that f (k)<br />
i (x 0 ) = 0, k = 1, . . . , i − 1 and P Yi : R m → Y i – the projection operator, where<br />
Y i is closed subspace of Y (see [1]). For a linear operator Λ h we define its right inverse Λ −1<br />
h<br />
and Λ −1 y is an element x ∈ Rn such that ‖x‖ = min {‖z‖ : Λ h (z) = y}. By the “norm” of<br />
Λ −1<br />
h<br />
h<br />
we mean the number ‖Λ−1<br />
h ‖ = sup ‖y‖=1 inf{‖x‖ : Λ h x = y, x ∈ R n }. If Im Λ h = R m ,<br />
then the sequence (2) is convergent to the solution of (1).<br />
We illustrate the basic idea of the method with some numerical examples.<br />
Key words: p-regularity, singularity, contracting mapping, multimapping, p-factor operator.<br />
References<br />
[1] Tretýakov A. A., Marsden J. E.: Factor-analysis of nonlinear mappings: p-regularity theory,<br />
Commun. Pure Appl. Math. 2, 425–445 (2003)<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 102<br />
Conference in Numerical Analysis 2014 (NumAn 2014)<br />
September 2-5, 2014<br />
Chania, Greece<br />
Solving the Fredholm integral equation of the second kind by<br />
global spline quasi-interpolation of the kernel<br />
P. Sablonnière a and D. Barrera b<br />
a INSA & IRMAR, Rennes, France<br />
b Department of Applied Mathematics, University of Granada,<br />
Granada, Spain<br />
Paul.Sablonniere@insa-rennes.fr,dbarrera@ugr.es<br />
Abstract<br />
For solving the linear Fredholm integral equation of the second kind<br />
u(x) = f(x) +<br />
∫ b<br />
a<br />
k(x, t)u(t)dt,<br />
we propose to approximate the kernel by tensor products or blending sums of univariate spline<br />
quasi-interpolants (abbr. QI). These QIs have the general form Qf := ∑ j∈J c j(f)B j , where<br />
the B j are B-splines defined on some partition of [a, b] and the coefficients c j (f) are linear<br />
functionals based on values of the function f on some finite subset S := {s j , j ∈ J} of the<br />
interval I := [a, b]. Thus, the kernel will be approximated<br />
1. either by the tensor product of two univariate quasi-interpolants Q 1 and Q 2 in the variables<br />
x and t:<br />
k(x, t) ≈ (Q 1 ⊗ Q 2 )k(x, t) = ∑ K i,j B i (x)B j (t),<br />
i,j<br />
where the coefficients K i,j are linear combinations of values k(s i , s j ), for (i, j) ∈ J × J,<br />
2. or by the continuous blending sum of the two univariate quasi-interpolants Q 1 and Q 2 :<br />
k(x, t) ≈ (Q 1 ⊕ Q 2 )k(x, t) := (Q 1 ⊗ Id + Id ⊗ Q 2 − Q 1 ⊗ Q 2 )k(x, t)<br />
= ∑ i<br />
˜ki (t)B i (x) + ∑ j<br />
k j (x)B j (t) − ∑ i,j<br />
K i,j B i (x)B j (t),<br />
where the functions ˜k i (t) = c i (k(., t)) (resp. k j (x) = c j (k(x, ·))) are linear combinations<br />
of left sections k(s k , t) (resp. right sections k(x, s l )) of the kernel.<br />
When substituting these approximate (degenerate) kernels in the Fredhom equation, we get<br />
two types of approximate solutions:<br />
• u(x) = f(x) + ∑ X i B i (x) in the tensor-product case,<br />
• u(x) = f(x) + ∑ X i B i (x) + ∑ Y j k j (x) in the continuous blending case,<br />
The vectors of variables X i and Y j are then solutions of systems of linear equations. The two<br />
methods can be used with any type of spline QIs, although only C 1 quadratic and C 2 cubic<br />
splines will be considered.<br />
Key words: Fredholm equation, quasi-interpolation, tensor product, boolean sum.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book Conference of Abstracts in Numerical Analysis 2014 (NumAn 2014) 103<br />
September 2-5, 2014<br />
Chania,Greece<br />
On the comparison between fitted and unfitted finite element<br />
methods for the approximation of void electromigration<br />
Andrea Sacconi a ,<br />
a Department of Mathematics, Imperial College London,<br />
London, United Kingdom<br />
a.sacconi11@imperial.ac.uk<br />
Abstract<br />
Microelectronic circuits usually contain small voids or cracks, and if those defects are large<br />
enough to sever the line, they cause an open circuit. Two fully practical finite element methods<br />
for the temporal analysis of the migration of voids in the presence of surface diffusion and<br />
electric loading are presented. We simulate a bulk-interface coupled system, with a moving<br />
interface governed by a fourth-order geometric evolution equation and a bulk where the electric<br />
potential is computed. A fitted approach (where the interface grid is always extracted from the<br />
boundary of the bulk grid) and an unfitted approach (where there is no perfect matching between<br />
the two grids) are analysed. A comparison between the two methods, in terms of experimental<br />
order of convergence (when the exact solution to free boundary problem is known), CPU time,<br />
and coupling operations (e.g., smoothing/re-meshing of the grids, intersection between elements<br />
of the two grids), is presented in detail. Several numerical simulations are performed in order to<br />
test the accuracy of the methods.<br />
Key words: void electromigration, fitted, unfitted, finite element methods, re-meshing, smoothing.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 104<br />
Conference in Numerical Analysis 2014 (NumAn 2014)<br />
September 2-5, 2014<br />
Chania,Greece<br />
A Numerical Mesh-Less Method for Solving Unsteady<br />
Compressible Flows<br />
Samad Sattarzadeh a and Alireza Jahangirian a<br />
a Department of Aerospace Engineering, Amirkabir University of Technology,<br />
Tehran, Tehran, Iran<br />
sattarzadeh@aut.ac.ir,ajahan@aut.ac.ir<br />
Abstract<br />
A dual-time implicit mesh-less method for unsteady compressible flows is presented. Polynomial<br />
least-square (PLS) method is used to estimate the spatial derivatives at each node. The<br />
points distributed are moved based on the boundary movement. The unsteady flows over stationary<br />
and moving objects at different flow conditions are solved. Results indicate the computational<br />
efficiency of the method in comparison with the similar explicit and finite volume<br />
approaches.<br />
In the recent years researchers have tried to solve the numerical computation of flow with<br />
complex stationary and/or moving boundaries by using Euler and Navier-Stokes equations in<br />
different regimes. One of the main problems in the Computational Fluid Dynamics (CFD) for<br />
numerical flow simulation around complex geometries is the quality of the mesh. Mesh-less<br />
methods are more advantageous, especially in the moving and large deformations. The reason<br />
is that replacing and moving points are much simpler than changing or replacing the edges and<br />
volumes. Another attractive property of mesh-less methods is the ability of adding and subtracting<br />
nodes from the pre-existing nodes. Several mesh-less methods have been used with<br />
different privileges and drawbacks. So, it should be noted that choosing one method depends on<br />
the desired applications. In these methods, the approximation of the characteristics or derivatives<br />
is based on a group of nodes which can be nominated as neighbors. Mesh-less methods<br />
need more nodes in comparison with the finite difference method to achieve the same order of<br />
accuracy. As a result, the Navier-Stokes equations are solved by great bandwidth of the matrix<br />
in these methods. Therefore, the computational memory is unavoidably extensive. Nowadays<br />
by increasing the power of computational instruments, this problem is solved automatically.<br />
To show the ability of the method, computational results are compared using experimental<br />
and other reliable numerical data.<br />
Key words: Mesh-less method, unsteady flow, compressible flow.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 105<br />
Conference in Numerical Analysis 2014 (NumAn 2014)<br />
September 2-5, 2014<br />
Chania,Greece<br />
A Numerical Adaptive Mesh-Less Method for Solution of<br />
Compressible Flows<br />
Samad Sattarzadeh a , Alireza Jahangirian a and Mehdi Ebrahimi a<br />
a Department of Aerospace Engineering, Amirkabir University of Technology,<br />
Tehran, Tehran, Iran<br />
sattarzadeh@aut.ac.ir,ajahan@aut.ac.ir,Mebrahimi@aut.ac.ir<br />
Abstract<br />
Nowadays mesh-less methods for numerical simulation of fluid flows has attracted much interest.<br />
This is because of the advantages of the method in comparison with alternative ones. At<br />
first, this method is used to solve boundary problems such as heat transfer and solid mechanics.<br />
Its advantages such as being less sensitive to the location of points; easier to change the cloud of<br />
points; to generate the point easily specially in the complex geometries or in the critical zones.<br />
It is shown that the time of point generating is less than generating the grid in the field. In<br />
the other hand, as it is obvious, the results depend on the quality of the grid especially in the<br />
critical zones, such as shock waves. There are different ways to increase the quality of the point<br />
cloud. One way is to use more points which may decrease the efficiency and performance. The<br />
other way to increase the quality of the grid and the performance is adaptation methods. There<br />
are several works on adapting the mesh-based methods but a few works have been done on<br />
mesh-less methods.<br />
In this paper, an adaptation mesh-less method is developed to solve the compressible flow<br />
equations. The comparison shows that the results are in good agreement with experimental and<br />
other reliable numerical data. In addition it has better convergence behavior than un-adapted<br />
point distribution.<br />
Key words: Mesh-less, Adaptive method, Compressible flow.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book Conference of Abstracts in Numerical Analysis 2014 (NumAn 2014) 106<br />
September 2-5, 2014<br />
Chania,Greece<br />
Solving CT reconstruction with a particle physics tool (RooFit)<br />
Enrico Jr. Schioppa a , Wouter Verkerke a , Jan Visser a and Els Koffeman a<br />
a Nikhef, Dutch national institute for subatomic physics<br />
Amsterdam, The Netherlands<br />
e.schioppa@nikhef.nl<br />
Abstract<br />
Spectral X-ray CT makes use of novel detector technologies that provide energy information.<br />
This information can be naturally included in the reconstruction phase when the algorithm<br />
is based on a statistical formulation. In this framework, the full dataset can be described in<br />
terms of a likelihood function whose expectation values are the density vectors of the different<br />
materials present in the sample. The problem of image reconstruction is thus translated into a<br />
problem of multivariate maximization.<br />
From the formal point of view, the same type of problem is often encountered in the analysis of<br />
large amounts of data in particle physics. For this purpose, the RooFit tool was developed during<br />
the years by the high energy physics community. In this work, we present first studies and<br />
results on the possibility to employ RooFit to implement a spectral CT reconstruction algorithm.<br />
Key words: Spectral X-ray CT, maximum likelihood reconstruction.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 107<br />
Ziggurat Algorithm for Sampling<br />
from Bivariate Distributions<br />
Efraim Shmerling a<br />
a Department of Computer Science and Mathematics, Ariel University,<br />
Ariel 40700, Israel<br />
efraimsh@yahoo.com,efraimsh@ariel.ac.il<br />
Abstract<br />
It is shown that the Ziggurat algorithm designed for sampling from monotone decreasing univariate<br />
distributions can be extended to continuous bivariate distributions with necessary modifications.<br />
A bivariate version of the Ziggurat algorithm is presented. Results of experimental<br />
sampling from a bivariate Gamma distribution utilizing a RNG implementing the algorithm are<br />
given.<br />
Key words: Ziggurat algorithm, bivariate gamma distribution, random number generation.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 108<br />
Fokas transform method for classes of advection-diffusion<br />
IBVPs 1<br />
A.G. Sifalakis ∗ , M.G. Papadomanolaki, E.P. Papadopoulou and<br />
Y.G. Saridakis<br />
Applied Mathematics and Computers Laboratory (AMCL)<br />
Technical University of Crete<br />
Chania 73100, Greece<br />
∗ sifal@science.tuc.gr<br />
Abstract<br />
It is now well established that Fokas transform approach for the solution of linear PDE problems,<br />
yields novel integral representations of the solution in the complex plane that, for appropriately<br />
chosen integration contours, decay exponentially fast and converge uniformly at the<br />
boundaries. Motivated by these method-inherent advantages and the fact that their coupling with<br />
simple quadrature integration rules produce practical, powerful and efficient methods, recently<br />
we considered applying them for the solution of discontinuous advection-diffusion equations<br />
that model the evolution of aggressive forms of primary brain tumors in heterogeneous brain<br />
tissue. The purpose of the present work is two-folded:<br />
• To review our recent results on the Fokas method for multi-domain linear advectiondiffusion<br />
equations with discontinuous diffusivity for brain tumor models<br />
• To examine the behavior of the Fokas method for classes of advection-diffusion equations<br />
with linear in t diffusivity in the real half-line, as the first step of extending the above<br />
results in non-constant diffusitivy models.<br />
1 The present research work has been co-financed by the European Union (European Social Fund ESF) and Greek<br />
national funds through the Operational Program Education and Lifelong Learning of the National Strategic Reference<br />
Framework (NSRF) - Research Funding Program: THALIS. Investing in knowledge society through the European<br />
Social Fund.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 109<br />
Conference in Numerical Analysis 2014 (NumAn 2014)<br />
September 2-5, 2014<br />
Chania,Greece<br />
Approximate algorithm for single valued nonexpansive and<br />
multi-valued strictly pseudo contractive mappings in Hilbert<br />
spaces<br />
Wutiphol Sintunavarat a<br />
a Department of Mathematics and Statistics, Faculty of Science and Technology,<br />
Thammasat University Rangsit Center,<br />
Pathumthani 12121, Thailand.<br />
wutiphol@mathstat.sci.tu.ac.th, poom teun@hotmail.com<br />
Abstract<br />
In this talk, we introduce and study a new one-step iterative process to approximate common<br />
fixed points for single valued nonexpansive and multi-valued strictly pseudo contractive mappings<br />
in Hilbert spaces. Also, we give the strong convergence theorems of the purposed process<br />
under some appropriate additional conditions.<br />
Key words: Single valued nonexpansive mappings, Multi-valued strictly pseudo contractive mappings,<br />
Common fixed points, Strong convergence.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book Conference of Abstracts in Numerical Analysis 2014 (NumAn 2014) 110<br />
September 2-5, 2014<br />
Chania,Greece<br />
Application of an image registration method<br />
based on maximization of mutual information<br />
C. Spanakis a 1,a 2<br />
, K. Marias b and N. A. Kampanis c<br />
a 1<br />
Institute of Computer Science, Foundation<br />
of Research and Technology, Greece<br />
a 2<br />
Department of Sciences, Technical University of Crete,<br />
Chania, Crete, Greece<br />
b Institute of Computer Science, Foundation<br />
of Research and Technology, Greece<br />
c Institute of Applied and Computational Mathematics,<br />
Foundation of Research and Technology, Greece<br />
kspan@ics.forth.gr, kmarias@ics.forth.gr, kampanis@iacm.forth.gr<br />
Abstract<br />
Image Registration is the process of transforming sets of data acquired at different time-points,<br />
sensors and viewpoints into a single coordinate system. It is widely used in computer vision,<br />
medical imaging and satellite image analysis. Although it has been a central research topic in<br />
computer vision and medical image analysis for a long time, there are still unresolved issues and<br />
success rates seem to be data-dependent. There are many categories of methods that that are able<br />
to align images, but usually they are either specialized and accurate for specific types of data<br />
or more generic and error-prone frequently stumbling upon pitfalls. In this work, we describe<br />
our implementation and results on Maes method (Maes et al. 1997). By using three different<br />
variants of mutual information (used as the similarity measure), we present indicative results<br />
from different imaging domains and discuss the drawbacks/pitfalls of the method especially<br />
with regard to initial transformation selection and the initial direction vectors. The results are<br />
quite accurate with translation and rotation when dealing with images of good quality. However,<br />
the choice of a starting point and the initial direction vectors proved to be two critical factors<br />
for the success of the method, since different starting point or/and different initial direction<br />
vectors may lead to different optimal alignment registration results between the images. In<br />
order to solve this problem, we propose an extension of this method by enhancing its global<br />
optimization scheme by means of stochastic optimization.<br />
Key words: Image Registration, Mutual Information, Genetic Algorithms.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 111<br />
Conference in Numerical Analysis 2014 (NumAn 2014)<br />
September 2-5, 2014<br />
Chania,Greece<br />
Successive approximations for optimal control in some nonlinear<br />
systems with small parameter<br />
Alexander Spivak<br />
Department of Computer Science, HIT, Holon Institute of Technology,<br />
52 Golomb str., Holon, ISRAEL<br />
spivak@hit.ac.il<br />
Abstract<br />
An important class of nonlinear control systems is bilinear systems. Such systems are linear on<br />
phase coordinates when the control is fixed, and linear on the control when the coordinates are<br />
fixed. The first point for the study of bilinear systems is to investigate the dynamic processes of<br />
nuclear reactors, kinetics of neutrons, and heat transfer. Further investigations show that many<br />
processes in engineering, biology, ecology and other areas can be described by the bilinear<br />
systems . It is shown that bilinear systems may be applied to describe some chemical reactions<br />
and many physical processes in the growth of the human population.<br />
In this work we consider nonlinear stochastic systems that can be described in the form<br />
ẋ(t) = ɛf 1 (t, x) + B(t)x(t)u(t) + σẇ(t),<br />
x(0) = x 0 , 0 ≤ t ≤ T. (1)<br />
Here the vector x(t) is from the Euclidean space E n , the control u(t) ∈ E m , the matrices σ<br />
and B have continuous and bounded elements, ɛ ≥ 0 is a small parameter, and the initial vector<br />
x 0 ∈ E n and the constant T ≥ 0 are given. The function f 1 (t, x) ∈ E n is continuous in the<br />
totality of its arguments, and satisfies some constraints, w(t) is standard Wiener process. The<br />
matrix σ(t) in (1) is such that σ(t)σ ′ (t) is positive definite. We understand the equation (1) in<br />
the sense of Ito. Note that if ɛ = 0, initial system (1) is stochastic bilinear, that is, it contains a<br />
nonlinearity of the form x(t)u(t). The problem is to find a control u minimizing the functional<br />
J(0, u), where<br />
[<br />
J(t, u) = M<br />
x ′ (t)H 1 x(t) +<br />
∫ T<br />
t<br />
( x ′ (s)H 2 (s)x(s)+<br />
]<br />
+ u ′ (s)H 3 (s)u(s) + f(s, x(s))) ds . (2)<br />
Here H i , i = 1, 2, 3 are given matrices, so that H 1 , H 2 (t) are non-negative defined, H 3 (t) is<br />
positive defined in the interval [0, T ], and the matrices H 2 (t) and H 3 (t) are measurable and<br />
bounded. The vector f(t, x) is determined later.<br />
When ɛ = 0, the optimal control synthesis is found in an exact analytic form. Successive<br />
approximations to the optimal control are constructed with the help of the perturbation method.<br />
Error estimates of the suggested method are presented.<br />
Key words: Successive Approximations, Small Parameter, Perturbation Theory.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 112<br />
Inverse moment problems with applications in<br />
shape reconstruction<br />
Nikos Stylianopoulos<br />
Department of Mathematics and Statistics, University of Cyprus,<br />
Nicosia, Cyprus<br />
nikos@ucy.ac.cy<br />
Abstract<br />
Let µ be a finite positive Borel measure with compact support in the complex plane, and let<br />
{p n (µ, z)} ∞ n=0 denote the sequence of the orthonormal polynomials, with positive leading coefficients,<br />
defined by the inner product<br />
∫<br />
⟨f, g⟩ µ := f(z)g(z)dµ(z).<br />
The purpose of the talk is to report on some recent developments regarding the asymptotics of<br />
{p n (µ, z)} ∞ n=0 , in cases when µ belongs to a special class of measures that includes area-type<br />
measures and arc-length measures. This leads to algorithms for recovering the shape of the<br />
support of µ, from a finite set of the moments<br />
∫<br />
µ i,j := z i z j dµ(z) i, j = 0, 1, . . . , n,<br />
and thus, via the Radon transform, to applications in 2D geometric tomography.<br />
Key words: Orthogonal Polynomials, Inverse Moment Problems, Shape Reconstruction, Geometric Tomography.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 113<br />
Method for solving degenerate sub-definite nonlinear equations<br />
Ewa Szczepanik a , Alexey Tretyakov b<br />
a Department of Computer Science, Faculty of Sciences,<br />
Siedlce University of Natural Sciences and Humanities, Siedlce, Poland<br />
b Department of Mathematics and Physics ,<br />
Siedlce University of Natural Sciences and Humanities, Siedlce, Poland<br />
System Research Institute of the Polish Academy of Sciences, Warsaw, Poland<br />
Dorodnicyn Computing Center of the Russian Academy of Sciences,<br />
Moscow, Russia<br />
ewa.szczepanik@ii.uph.edu.pl,tret@uph.edu.pl<br />
We consider system of nonlinear equations<br />
⎡<br />
F (x) =<br />
Abstract<br />
⎢<br />
⎣<br />
f 1 (x)<br />
...<br />
f m (x)<br />
⎤<br />
⎥<br />
⎦ = 0,<br />
where x ∈ R n , m ≤ n, and F ′ (x ∗ ) is degenerate at the solution point x ∗ .<br />
Denote the feasible set as follows M(x ∗ ) = {x ∈ R n |F (x) = 0}.<br />
It is known that in nondegenerate case ( when F ′ (x ∗ ) is nondegenerate) the Newton-Gauss<br />
method converges to some point ˜x ∗ ∈ M(x ∗ )([1] pp. 228), and the rate of convergence is<br />
quadratic.<br />
In the complete system of nonlinear equations one of the main result of the p- regularity<br />
theory is p-factor method. Scheme of this method is as follows<br />
x k+1 = x k − {ψ ′ p(x k )} −1 ψ p (x k ),<br />
where p-factor operator ψ p (x k ) has following form<br />
ψ p (x k ) = P 1 F ′ (x k ) + P 2 F ′′ (x k )[h] + ... + P p F (p) (x k )[h] [p−1] .<br />
Here P 1 is ortoprojection onto Im(F ′ (x ∗ )) ⊥ and P i , i = 2, ..., p also ortoprojection on the same<br />
defined sets and the element h(‖h‖ = 1) we construct in such a way that p-factor matrix<br />
P 1 F ′ (x ∗ ) + P 2 F ′′ (x ∗ )h + ... + P p F (p) (x ∗ )[h] [p−1]<br />
is nonsingular at the solution point x ∗ = 0 (p-regular along h).<br />
However for degenerate sub-definite nonlinear equations has been considered only the case<br />
for p = 2. In this paper we present the generalization of the p-factor method for sub-definite<br />
case and p ≥ 2. We will also describe a numerical algorithm of the p-factor method and will<br />
give numerical results.<br />
References<br />
[1] A.F. IZMAILOV, A. A.TRETYAKOV, Factor-analysis of nonlinear mappings, Nauka,<br />
Moscow, 1994.[in Russian]<br />
Key words: nonlinear equation, p-factor operator, 1-regularity, singularity, necessary and<br />
sufficient condi-tions, nonregular constraints, 1-factor methods<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 114<br />
Stochastic optimization for a problem of saltwater intrusion in<br />
coastal aquifers with heterogeneous hydraulic conductivity 1<br />
P. N. Stratis a , G.P. Karatzas b , E.P. Papadopoulou a and Y.G. Saridakis a<br />
a Applied Mathematics and Computers Laboratory<br />
b Environmental Engineering Department<br />
Technical University of Crete<br />
73100 Chania, Greece<br />
pstratis@science.tuc.gr<br />
Abstract<br />
In the present study we implement the stochastic optimization technique ALOPEX, in order to<br />
control the problem of saltwater intrusion in coastal aquifers. The objective is to maximize the<br />
total volume of freshwater pumped by the wells of the aquifer, while protecting the aquifer from<br />
salt water intrusion. Extending previous results, we examine some cases of non-homogeneous<br />
aquifers, divided in rectangular areas with different values for the hydraulic conductivity parameter.<br />
At the same time, appropriate penalties strategies are used to produce different management<br />
policies. Numerical experimentation with several test cases of non-homogeneous aquifers and<br />
two real case coastal aquifers (Vathi and Hersonissos aquifers in Greece) are presented.<br />
Key words: ALOPEX stochastic optimization, non-homogeneous coastal aquifers, saltwater intrusion,<br />
pumping management, hydraulic conductivity<br />
1 The present research work has been co-financed by the European Union (European Social Fund ESF) and Greek<br />
national funds through the Operational Program Education and Lifelong Learning of the National Strategic Reference<br />
Framework (NSRF) - Research Funding Program: THALIS. Investing in knowledge society through the European<br />
Social Fund.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 115<br />
Numerical simulations for 1+2 dimensional coupled nonlinear<br />
Schrödinger type equations<br />
Thiab Taha<br />
Department of Computer Science, University of Georgia,<br />
Athens, GA, USA<br />
thiab@cs.uga.edu<br />
Abstract<br />
The coupled nonlinear Schrödinger equation is of tremendous importance in both theory and<br />
applications. Coupled nonlinear Schrödinger equation (CNLS) is the vectorial version of the<br />
nonlinear Schrödinger equation (NLS). The NLS equation is the main governing equation in<br />
the area of optical solitons. In this paper, we perform numerical simulations of two dimensional<br />
CNLS equation using the finite difference methods: a) the Explicit Finite Difference method and<br />
b) the Implicit Finite Difference method (Alternating Directions Implicit method). The methods<br />
are implemented. Our preliminary numerical results have shown that these methods give good<br />
results.<br />
Key words: Parallel Algorithms, MPI, Finite Difference.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 116<br />
Conference in Numerical Analysis 2014 (NumAn 2014)<br />
September 2-5, 2014<br />
Chania,Greece<br />
A Numerical Model for the prediction of flooding in Water<br />
Rivers<br />
Katerina Tsakiri a , Antonios Marsellos b .<br />
a School of Computing, Engineering and Mathematics, University of Brighton,<br />
BN2 4GJ, United Kingdom<br />
b School of Environment and Technology, University of Brighton, BN2 4GJ,<br />
United Kingdom<br />
k.tsakiri@brighton.ac.uk,a.marsellos@brighton.ac.uk<br />
Abstract<br />
A numerical model is presented for the explanation and prediction of the daily water discharge<br />
time series derived by three locations nearby Mohawk River, New York during the period<br />
2005-2013. For the analysis of the model, we use daily water discharge time series, daily data<br />
of ground water level and the climatic variables in Mohawk River, New York. A methodology is<br />
used for the decomposition of the time series of all the variables into different components (long,<br />
seasonal and short term component). The long term component describes the fluctuations of a<br />
time series defined as being longer than a given threshold; the seasonal component describes the<br />
year-to-year fluctuations, while the short term component describes the short term variations.<br />
The Kolmogorov-Zurbenko (KZ) filter is used for the decomposition of the time series. The KZ<br />
filter, which separates the long term variations from the short term variations in a time series,<br />
provides a simple design and the smallest level of interferences between the scales of a time series.<br />
The application of the KZ filter in an example of Schoharie Creek (nearby Mohawk River)<br />
has improved the prediction of the water discharge up to 81%. This methodology has been also<br />
applied for Sussex Rivers in United Kingdom.<br />
Key words: Flooding Prediction, decomposition of time series, KZ filter.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 117<br />
Interface Rexation Methods for the solution of Multi-Physics<br />
Problems 1<br />
Panagiota Tsompanopoulou<br />
Department of Computer and Communications Engineering,<br />
University of Thessaly, Volos, Greece<br />
yota@inf.uth.gr<br />
Abstract<br />
Multi-domain multi-physics problems simulate real world problems demanding efficient and<br />
high accuracy solutions. Domain Decomposition methods are well known methods that treat<br />
such kind of problems, but they first discretize the global problem (even if it is already partitioned<br />
by its physics) and then decompose it at the linear algebra level. Several techniques,<br />
mainly iteratively, are used to solve the set of the strongly coupled systems of linear equations<br />
that arise.<br />
Interface Relaxation (IR) methodology is a different and relatively new way to study such problems.<br />
The idea behind IR is to confront the global problem as closer as possible to its nature by<br />
realizing and utilizing its basic properties and behavior. Subproblems arise either by the physics<br />
of the original problem or by computational and parallelization issues. These “small” problems<br />
are studied independently of each other and appropriate methods (FEM, FD etc.) are used for<br />
their solution. However these subproblems are coupled on the common interfaces so as to satisfy<br />
the conditions resulting from the global problem’s properties (e.g., continuity and smoothness<br />
of the solution of the global problem). Initial guesses are considered on the interfaces, passed as<br />
boundary conditions to the “small” problems. These are solved concurrently and the resulting<br />
approximations are used by an IR method to relax the value and/or the derivative to get better<br />
estimates of the solution on the interfaces. These new estimates are passed again as boundary<br />
conditions to the small problems and the procedure iterates until convergence is achieved.<br />
When studying IR methods, one should consider issues from both mathematical analysis, computational<br />
complexity and software/hardware viewpoint. Mathematical analysis is often derived<br />
for model problems representatives of the original multiphysics applications since it is not possible<br />
and practical to get analysis for the realistic problems. A variety of software packages<br />
for the solution of simple non-multiphysics problems exist but they have to be combined under<br />
suitable software and hardware environments. Thus software reuse is of great importance when<br />
implementing IR methods.<br />
In this review we consider the Interface Relaxation methods proposed for the solution of multidomain<br />
multi-physics problems from both theoretical and implementation perspectives.<br />
Key words: Interface relaxation, multiphysics problems, Elliptic PDEs, Parabolic PDEs, software reuse.<br />
1 The present research work has been co-financed by the European Union (European Social Fund ESF) and Greek<br />
national funds through the Operational Program Education and Lifelong Learning of the National Strategic Reference<br />
Framework (NSRF) - Research Funding Program: THALIS. Investing in knowledge society through the European<br />
Social Fund.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 118<br />
Conference in Numerical Analysis 2014 (NumAn 2014)<br />
September 2-5, 2014<br />
Chania,Greece<br />
An order 19-rational integrator<br />
L. A. Ukpebor<br />
Department of Mathematics,<br />
Ambrose Alli University, Ekpoma, Nigeria<br />
lukeukpebor@gmail.com<br />
Abstract<br />
In this research paper, an order 19-rational integrator was developed for solving stiff initial-value<br />
problems of differential equations of the form:<br />
y ′ = f(x, y); y(x 0 ) = y 0<br />
The consistency and convergence of this method were established following the steps of Lambert<br />
J.D [?], where he stated that a one-step numerical integrator of the form<br />
y n+1 = y n + h n φ(x n , y n , h n )<br />
is convergent if and only if it is consistent.<br />
The application of this method to some selected stiff problems showed that it compared favourably<br />
with existing methods in terms of efficiency and accuracy.<br />
Key words: initial value problems, rational integrator, consistency, convergence, stiff problems.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 119<br />
Conference in Numerical Analysis 2014 (NumAn 2014)<br />
September 2-5, 2014<br />
Chania,Greece<br />
A Meshfree Method with Fundamental Solutions for<br />
Inhomogeneous Elastic Wave Problems<br />
Svilen S Valtchev a,b , Carlos J S Alves a,c and Nuno F M Martins a,d<br />
a CEMAT, ULisbon, Portugal,<br />
b ESTG, Polytechnic Institute of Leiria, Portugal<br />
c Department of Mathematics, ULisbon, Portugal<br />
d Department of Mathematics, FCT, Universidade Nova de Lisboa, Portugal<br />
ssv@math.ist.utl.pt,carlos.alves@math.ist.utl.pt,nfm@fct.unl.pt<br />
Abstract<br />
We consider the numerical solution of the inhomogeneous Cauchy-Navier equations of elastodynamics,<br />
assuming time-harmonic variation for the displacement field U(x, t) = u(x)e −iωt<br />
of an isotropic material with Lamé constants λ and µ and density ρ. The resulting elliptic PDE,<br />
posed in a bounded simply connected domain Ω is coupled with Dirichlet boundary conditions<br />
and solved trough a meshfree method, based on the Method of Fundamental Solutions (MFS).<br />
{ µ∆u + (λ + µ)∇(∇ · u) + ρω 2 u = f in Ω<br />
u = g<br />
In particular, an extension, from the scalar [1, 2] to the vector case, of the MFS is applied and the<br />
displacement field u is approximated in terms of a linear combination of fundamental solutions<br />
(Kupradze tensors) of the corresponding homogeneous PDE with different source points and test<br />
frequencies. The applicability of the numerical method is justified in terms of density results [3].<br />
The high accuracy and the convergence of the proposed method will be illustrated through 2D<br />
numerical simulations. Convex and non-convex domains and different sets of boundary data and<br />
body forces will be considered. Interior elastic wave scattering problems will also be addressed.<br />
Key words: Method of Fundamental Solutions, Inhomogeneous BVP, Elastic Wave Propagation.<br />
on Γ<br />
References<br />
[1] C.J.S. Alves and C.S. Chen, A new method of fundamental solutions applied to nonhomogeneous elliptic<br />
problems. Adv. Comput. Math., 23, 125(18pp), 2005.<br />
[2] C.J.S. Alves and S.S. Valtchev, A Kansa type method using fundamental solutions applied to elliptic<br />
PDEs. Advances in meshfree techniques, Computational Methods in Applied Sciences Series, vol. 5,<br />
241(16pp), Springer, 2007.<br />
[3] C.J.S. Alves, N.F.M. Martins and N. C. Roberty, Identification and reconstruction of elastic body forces.<br />
Inverse Problems, 30, 055015(18pp), 2014.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 120<br />
On the Numerical Solution of Power Flow Problems 1<br />
Manolis Vavalis and Dimitris Zimeris<br />
Department of Electrical and Computer Engineering, University of Thessaly,<br />
Volos, Greece<br />
{mav,dzimeris}@uth.gr<br />
Abstract<br />
The power generation, transmission and distribution system has been widely recognized as one<br />
of the most complex man-made systems and the related power flow analysis as the main ingredient<br />
of many related studies.<br />
The power flow problem is modeled through a system of non-linear equations that relate<br />
the bus voltages to the power generation and consumption. Its solution is used to access the<br />
stability of the power system and perform contingency analysis. It is also required by other<br />
related and relatively new problems, for example the optimal power flow problem, the financial<br />
transmission rights mechanisms and many others.<br />
Several recent advances (the liberalization of the energy markets, the emerging of smart grid<br />
technologies, the stochasticity in the power production due to utilization of renewable energy<br />
sources, the decentralization of the energy production) have recently increased the complexity<br />
of the power flow problems significantly.<br />
The envisioned interconnection of national power systems with global energy markets will<br />
be based on truly large scale, continent-wide power flow simulations where the efficiency of the<br />
numerical solution of the power flow equations is expected to be a vital component.<br />
This paper consists an up-to-day review of the various numerical methods that have been<br />
very recently proposed for the solution of power flow equations and several other related problems.<br />
These methods are examined from both the theoretical (convergence analysis) and the<br />
practical (efficiency, robustness, numerical stability, implementation) viewpoint.<br />
We also propose several new research directions which, we believe, have the potential to<br />
lead us to next generation power grid simulation engines. Engines that are capable to support<br />
operational large scale modern power grid systems associated with open energy markets, paying<br />
special attention to the information flow in addition to the power flow.<br />
Key words: Numerical linear algebra, Newton’s method, power flow equations, GMRES, preconditioning,<br />
basic iterative methods.<br />
1 The present research work has been co-financed by the European Union (European Social Fund ESF) and Greek<br />
national funds through the Operational Program ”Education and Lifelong Learning” of the National Strategic Reference<br />
Framework (NSRF) - Research Funding Program: THALIS. Investing in knowledge society through the European<br />
Social Fund.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 121<br />
Conference in Numerical Analysis 2014 (NumAn 2014)<br />
September 2-5, 2014<br />
Chania,Greece<br />
Towards robust parallel solvers for tridiagonal<br />
systems for multiGPUs<br />
Ioannis E. Venetis a , Alexandros Kouris a , Nikolaos Nikoloutsakos a ,<br />
Alexandros Sobczyk a and Efstratios Gallopoulos a<br />
a Computer Engineering & Informatics Department, University of Patras,<br />
Patras, Achaia, Greece<br />
venetis@ceid.upatras.gr,kouris@ceid.upatras.gr,<br />
nikoloutsa@ceid.upatras.gr,sobczyk@ceid.upatras.gr,<br />
stratis@ceid.upatras.gr<br />
Abstract<br />
We recently ([2]) presented an algorithm for nonsingular tridiagonal systems that is robust in<br />
that it can handle arbitrary partitionings of the matrix, even when the resulting diagonal blocks<br />
are exactly singular. We also showed an implementation on an NVIDIA GPU card that demonstrated<br />
performance that is very close to state of the art solvers, e.g. [1]. We extend here this<br />
work to clusters of GPUs using a combination of CUDA with MPI. The algorithm is based on<br />
ideas first presented by Sameh and Kuck in [3]; it is based on Givens rotations and requires no<br />
pivoting, which makes the algorithm simpler and more robust than existing ones for the GPU<br />
and for multiGPUs.<br />
References<br />
[1] L.-W. Chang, J.A. Stratton, H.S. Kim, and W.-M.W. Hwu. A scalable, numerically stable,<br />
high-performance tridiagonal solver using GPUs. In Proc. Int’l. Conf. High Performance<br />
Computing, Networking Storage and Analysis, SC ’12, pages 27:1–27:11, Los Alamitos,<br />
CA, USA, 2012. IEEE Computer Society Press.<br />
[2] I. E. Venetis, A. Kouris, A. Sobczyk, E. Gallopoulos and A.H. Sameh. Revisiting the<br />
Spike-based framework for GPU banded solvers: A Givens rotation approach for tridiagonal<br />
systems in CUDA. Parallel Matrix Algorithms and Applications Workshop, Lugano,<br />
July 2014.<br />
[3] A.H. Sameh and D.J. Kuck. On stable parallel linear system solvers. J. Assoc. Comput.<br />
Mach., 25(1):81–91, January 1978.<br />
Key words: Tridiagonal systems, Givens rotations, Spike, GPU, CUDA, MPI.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 122<br />
Some new perturbation bounds of generalized polar<br />
decomposition<br />
X.-L. Hong, L.-S. Meng and B. Zheng<br />
School of Mathematics and Statistics, Lanzhou University<br />
Lanzhou, Gansu Province, P.R.China<br />
hongxiaoli2007@163.com,menglsh07@lzu.edu.cn,bzheng@lzu.edu.cn<br />
Abstract<br />
Let A, Ã = A + E ∈ Cm×n have the (generalized) polar decompositions<br />
A = QH and à = ˜Q ˜H, (1)<br />
where Q is subunitary and H is Hermitian positive semi-definite. We present the following<br />
new bounds of the positive (semi-)definite polar factor and the (sub) unitary polar factor for<br />
the (generalized) polar decomposition under the general unitarily invariant norm ∥·∥and the<br />
spectral norm ∥·∥ 2 , which are stated as in the following theorem.<br />
Theorem. Let A, Ã = A + E ∈ Cm×n r have the (generalized) polar decompositions in (1).<br />
(1). When r and N ∈ C n×n<br />
> . Hence, all perturbation bounds in the above theorem<br />
can be naturally extended to the case of the weighted polar decomposition of A, which also<br />
improved the known perturbation bounds for the weighted polar decomposition.<br />
Key words: Perturbation bounds; Positive semi-definite polar factor; Subunitary polar factor; Generalized<br />
polar decomposition; Weighted polar decomposition; Unitarily invariant norm; Spectral norm.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 123<br />
Chebyshev accelerated preconditioned MHSS iteration methods<br />
for a class of block two-by-two linear systems<br />
Zeng-Qi Wang a ,<br />
a Department of Mathematics, Shanghai Jiao Tong University ,<br />
Shanghai, China<br />
wangzengqi@sjtu.edu.cn<br />
Abstract<br />
The preconditioned modified Hermitian and skew-Hermitian iteration method [1] is efficient<br />
for solving the the following block two-by-two systems of linear equations<br />
( ) ( ) ( )<br />
W −T y p<br />
Ax ≡<br />
= ≡ g.<br />
T W z q<br />
It could be written as the following procedure:<br />
⎧ ( ) ( )<br />
αV + W 0 y (k+ 1 2 )<br />
⎪⎨ 0 αV + W z (k+ 1 2 )<br />
( ) ( )<br />
αV + T 0 y<br />
(k+1)<br />
⎪⎩ 0 αV + T z (k+1)<br />
=<br />
=<br />
( ) ( ) ( αV T y<br />
(k) p<br />
−T αV z (k) +<br />
q<br />
( ) ( αV −W y (k+ 1 2 )<br />
W<br />
αV<br />
z (k+ 1 2 ) )<br />
)<br />
,<br />
( q<br />
+<br />
−p<br />
)<br />
,<br />
where α is a given positive constant and V ∈ R n×n is a prescribed symmetric positive definite<br />
matrix. The Chebyshev semi-iteration method is fulfilled for accelerating the above iteration<br />
method. It could be verified that the Chebyshev accelerated PMHSS iteration method is a parameter<br />
free method. It converges unconditionally. The new method is utilized on solving the<br />
distributed control problems. Numerical experiments shows that the performance of the Chebyshev<br />
accelerated PMHSS iteration method is independent on not only the mesh size and the<br />
regularization parameter of the cost functional.<br />
Key words: Chebyshev semi-iteration, PMHSS iteration, PDE-constrained optimization, block two-bytwo<br />
matrices.<br />
References<br />
[1] Bai Z-Z, Benzi M, Chen F, Wang Z-Q (2013) Preconditioned MHSS iteration methods for<br />
a class of block two-by-two linear systems with applications to distributed control problems.<br />
IMA Journal of Numerical Analysis 33:343-369<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book Conference of Abstracts in Numerical Analysis 2014 (NumAn 2014) 124<br />
September 2-5, 2014<br />
Chania,Greece<br />
The WR-HSS Methods for Non-Self-Adjoint Positive Definite<br />
Linear Differential Equations and Applications to the Unsteady<br />
Discrete Elliptic Problem<br />
Xi Yang a ,<br />
a Dept. Math., Nanjing University of Aeronautics and Astronautics,<br />
Nanjing 210016, Jiangsu, P.R. China<br />
yangxi@lsec.cc.ac.cn<br />
Abstract<br />
We consider the numerical methods for non-self-adjoint positive definite linear differential equations,<br />
L(x) = B ẋ + A x = q, x(0) = x 0 , (1)<br />
with B being Hermitian and A being non-Hermitian positive definite, and their corresponding<br />
applications to the unsteady discrete elliptic problem, which is derived from spatial discretization<br />
of the unsteady elliptic problem with Dirichlet boundary condition, i.e.,<br />
{ ∂u<br />
∂t − ∇ · [a(x)∇u(x)] + ∑ d ∂<br />
j=1 ∂x j<br />
(p(x)u(x)) = f(x), u(x, 0) = u 0 (x), x ∈ Ω<br />
Dirichlet Boundary Condition.<br />
(2)<br />
Taking into account the idea of Hermitian/skew-Hermitian splitting (HSS) in [1], we establish<br />
a class of waveform relaxation iteration methods based on the HSS splitting of the non-selfadjoint<br />
positive definite linear operator L, i.e., WR-HSS methods. We analyze these WR-HSS<br />
methods with the help of Fourier Transform. Similarly to the HSS methods for solving linear<br />
algebraic equations, we find that the WR-HSS methods are unconditionally convergent to the<br />
solution of (1). In addition, we derive the upper bound of the contraction factor of the WR-HSS<br />
methods which is only dependent on the Hermitian part of L. Finally, the applications of these<br />
WR-HSS methods to the unsteady discrete elliptic problem demonstrate their effectiveness and<br />
the corresponding theoretical results.<br />
Key words: elliptic problem, Hermitian/skew-Hermitian splitting, waveform relaxation.<br />
References<br />
[1] Z.-Z. Bai, G.H. Golub and M.K. Ng, Hermitian and skew-Hermitian splitting methods for<br />
non-Hermitian positive definite linear systems, SIAM J. Matrix Anal. Appl., 24(2003), 603-<br />
626.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book Conference of Abstracts in Numerical Analysis 2014 (NumAn 2014) 125<br />
September 2-5, 2014<br />
Chania,Greece<br />
Sensitivity of the Domain Decomposition Method to<br />
Perturbation of the Transmission Conditions<br />
Anastasiya Zaitseva a and Vadim Lisitsa a,b<br />
a Novosibirsk State University,<br />
Novosibirsk, Russia<br />
b Institute of Petroleum Geology and Geophysics of SB RAS,<br />
Novosibirsk, Russia<br />
zaf1990@mail.ru, lisitsavv@ipgg.sbras.ru<br />
Abstract<br />
Nowadays Domain Decomposition (DD) method is one of the common tools to construct preconditioner<br />
to solve 3D Helmholtz equation, especially in geophysical applications. There are<br />
numerous papers devoted to construction of optimal transmission conditions to improve convergence<br />
of the DD. However, these researches are focused on the differential statements and no<br />
perturbation is typically assumed. Whereas solution of the 3D Helmholtz equation requires the<br />
use of the numerical methods such as finite differences of finite elements, thus a numerical error<br />
is introduced in the transmission conditions as a result of numerical approximation. Moreover,<br />
in some cases it is worth using different numerical methods in adjoint subdomains, which makes<br />
the considered perturbations nonsymmetric.<br />
In this paper, a simplest 1D Helmholtz equation was considered and the perturbation of<br />
the Dirichlet-to-Neumann map based transmission conditions are considered. It was proved,<br />
that if the perturbation is symmetric (the same numerical methods and discretizations are used<br />
in the adjoint subdomains) the numerical solution converges to the true solution for almost<br />
all practically meaningful cases. However, if the perturbation is nonsymmetric, i.e. different<br />
numerical methods are used, different discretizations are applied, different approximations of<br />
the boundary operators are utilized, the numerical solution converges, but not to the solution of<br />
the original problem. In this case, an irreducible error presents, which linearly depends on the<br />
perturbation of the transmission conditions.<br />
The research was done under financial support of the Russian Foundation for Basic Research<br />
grants no. 13-05-00076, 13-05-12051, 14-05-00049, 14-05-93090, 14-01-31340, fellowship<br />
SP-150.2012.5 of the President of the Russian Federation, and integration projects of SB RAS<br />
127 and 130.<br />
Key words: Domain decomposition, Helmholtz equation.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 126<br />
An improved model of heart rate kinetics<br />
Maria Zakynthinaki<br />
Applied Mathematics and Computers Laboratory<br />
Technical University of Crete<br />
Chania, Greece<br />
marzak@science.tuc.gr<br />
Abstract<br />
The heart rate in response to movement (exercise) is modeled as a dynamical system and its<br />
temporal evolution is given as the solution of a system of two coupled differential equations.<br />
The model assumes the heart rate kinetics to be a function of exercise intensity (which can also<br />
be time-dependent), blood lactate and the current cardiovascular condition of the individual.<br />
By means of numerical optimization the model can be fit to experimental heart rate time series<br />
data and provide important information regarding an individual’s cardiovascular condition.<br />
Numerical simulations can also provide predictions for any given exercise intensity, even those<br />
that no data exist for. This is of great importance, not only for efficiently designing training<br />
sessions for healthy subjects, but also for providing a complete means of heart rate analysis in<br />
population groups for which direct heart rate recordings at intense exercises are not possible or<br />
not allowed, such as elderly or pregnant women. Examples of successful fit of the proposed<br />
model to recorded heart rate time series data, as well as heart rate kinetics simulations will be<br />
presented.<br />
Key words: Cardiac dynamics, numerical models, numerical optimization, numerical simulation.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 127<br />
Conference in Numerical Analysis 2014 (NumAn 2014)<br />
September 2-5, 2014<br />
Chania,Greece<br />
Normalizations of the Proposal Density<br />
in Markov Chain Monte Carlo Algorithms<br />
Antoine Zambelli a<br />
a Anderson School of Management, University of California Los Angeles,<br />
Los Angeles, CA, USA<br />
antoine.zambelli.2014@anderson.ucla.edu<br />
Abstract<br />
We explore the effects of normalizing the proposal density in Markov Chain Monte Carlo algorithms,<br />
in the context of a nonlinear inverse problem. Our problem is that of reconstructing the<br />
conductivity term K in the 2-dimensional heat equation<br />
u xx + u yy = 2H<br />
Kδ u (1)<br />
given temperatures at the boundary points, given by d. A Metropolis-Hastings MCMC algorithm<br />
is implemented to do so. Markov Chains produce a probability distribution of possible<br />
solutions conditional on the observed data. We generate a candidate solution K ′ and solve the<br />
forward problem, obtaining d ′ . In this way, at step n the probability of setting K n+1 = K ′ is<br />
given by<br />
{<br />
α(K ′ |K n ) ≡ min<br />
1, P (K′ |d)g(K n |K ′ )<br />
P (K n |d)g(K ′ |K n )<br />
For our given proposal density g, this is initially computed as<br />
α = min<br />
{<br />
1, e<br />
∑<br />
−1 n,m<br />
2σ 2 i,j=1<br />
[(d ij −d ′ ij) 2 −(d ij −d nij ) 2] }<br />
}<br />
(2)<br />
{<br />
= min 1, e −D} (3)<br />
We identify certain issues with this construction, stemming from large and fluctuating values of<br />
D. Using this framework, we develop normalization terms z 0 , z and parameters λ that preserve<br />
the inherently sparse information at our disposal, rewriting (3) as<br />
α = z 0 e −λzD (4)<br />
We examine the results of this variant of the MCMC algorithm on the reconstructions of several<br />
2-dimensional conductivity functions.<br />
Key words: Ill-posed, Inverse Problems, MCMC, Normalization, Numerical Analysis.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book Conference of Abstracts in Numerical Analysis 2014 (NumAn 2014) 128<br />
September 2-5, 2014<br />
Chania,Greece<br />
A local preconditioned alternating direction iteration method for<br />
generalized saddle point problems<br />
Guo-Feng Zhang a and Zhong Zheng a<br />
a School of Mathematics and Statistics, Lanzhou University,<br />
Lanzhou, 730000, China<br />
gf zhang@lzu.edu.cn, zhengzh13@lzu.edu.cn<br />
Abstract<br />
In this paper, a local preconditioned alternating direction iteration method is presented for solving<br />
the generalized saddle point problems. By using the new method, we only need solver two<br />
linear sub-system of linear equations with symmetric and definite positive coefficient matrices<br />
per iteration step for solving the generalized saddle point problems. The convergence of the<br />
new iteration method is analyzed and some spectral properties of the preconditioned matrix are<br />
discussed. Numerical examples are reported to confirm the efficiency of the proposed method.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 129<br />
Conference in Numerical Analysis 2014 (NumAn 2014)<br />
September 2-5, 2014<br />
Chania,Greece<br />
Katservich Algorithm Based on Spherical Detector for<br />
Cone-Beam CT and the Implementation on GPU<br />
Yan Zhang, Qian Li<br />
Harbin Institute of Technology Shenzhen Graduate School,<br />
Shenzhen, Guangdong, China, 518055<br />
ianzh@foxmail.com,lqiankm@foxmail.com<br />
Abstract<br />
Katsevich algorithm is an exact cone-beam reconstruction algorithm of filtered backprojection<br />
(FBP) type. In this paper, an implementation for a spherical detector is proposed. It reduces the<br />
error generated by geometric shapes such as curved or plan detector. CT image reconstruction<br />
using spherical detector makes the speed of reconstruction faster and the quality of image better.<br />
Since CT image reconstruction has a huge amount of computation, it is difficult to meet the requirements<br />
of both fast and accurate reconstruction using CPU. This paper takes the advantages<br />
of GPU which is programmable and parallelizable to accelerate the image reconstruction.<br />
Key words: cone-beam CT, Katservich algorithm, spherical detector, GPU, acceleration<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book of Abstracts 130<br />
A Riemannian Newton Algorithm for Nonlinear Eigenvalue<br />
Problems<br />
Zhi Zhao a , Zheng-Jian Bai b and Xiao-Qing Jin a<br />
a Department of Mathematics, University of Macau,<br />
Macao, People’s Republic of China<br />
b School of Mathematical Sciences, Xiamen University,<br />
Xiamen 361005, People’s Republic of China<br />
zhaozhi231@163.com, zjbai@xmu.edu.cn, xqjin@umac.mo<br />
Abstract<br />
We give the formulation of a Riemannian Newton algorithm for solving a class of nonlinear<br />
eigenvalue problems by minimizing a total energy function subject to the orthogonality constraint.<br />
Under some mild assumptions, we establish the global and quadratic convergence of the<br />
proposed method. Moreover, the positive definiteness condition of the Riemannian Hessian of<br />
the total energy function at a solution is derived. Some numerical tests are reported to illustrate<br />
the efficiency of the proposed method for solving large-scale problems.<br />
Key words: nonlinear eigenvalue problem, Riemannian Newton algorithm, Stiefel manifold, Grassmann<br />
manifold.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014
NumAn2014 Book Conference of Abstracts in Numerical Analysis 2014 (NumAn 2014) 131<br />
September 2-5, 2014<br />
Chania,Greece<br />
Finite element approximations for a linear stochastic<br />
Cahn-Hilliard-Cook equation<br />
Georgios E. Zouraris a<br />
a Department of Mathematics and Applied Mathematics,<br />
University of Crete,<br />
Heraklion, Crete<br />
Greece<br />
zouraris@math.uoc.gr<br />
Abstract<br />
We consider an initial- and Dirichlet boundary- value problem for a linear stochastic Cahn-<br />
Hilliard-Cook equation driven by an additive noise. We approximate its solution using, for<br />
the discretization in space, a finite element method, and for the discretization in time, a timestepping<br />
method. For the proposed numerical method, we derive strong a priori error estimates.<br />
Key words: stochastic Cahn-Hilliard-Cook equation, finite element method, error estimates.<br />
numan2014.amcl.tuc.gr — Conference on Numerical Analysis, Chania, Greece, Sept 2-5, 2014