Weeks' Method for Numerical Laplace transform inversion with GPU acceleration 1.0

The numerical inversion of the Laplace transform is a long standing problem due its implicit ill-posedness. These functions implement one of the more well known numerical inversion algorithms, the Weeks method. Particularly new here is the use of graphics processing unit [GPU] computing to accelerate the method.To assist the user, a wrapper (WeeksMethod.m) to the core inversion functions is provided. It allows a user to input the Laplace space function F(s) to invert as a string and a sequence of inversion times to obtain approximate f(t) values. Optional inputs are the relative error tolerance (0.05 for 5%) for each time and a selection of the search algorithm for the method's parameters. The wrapped code will refine the f(t) approximations until an internal estimate of the relative error is below the user requested tolerance. For some classes of F(s) functions, it is not possible to reach the input tolerance. Therefore the code can also return a flag for each time to indicate if the tolerance was met and the internally computed estimates of the absolute and relative errors.The optional input search algorithm switch defines how the code will determine optimal values for the method's parameters. There are three choices: 0-local CPU fminbnd search, 1-global CPU search, 2-global GPU search. The tool has been written so that it will run in standard MATLAB on a central processing unit [CPU]. However, to rapidly perform a global search for optimal values of the method's parameters, GPU computing has also been used. To be more precise, in the Weeks method a function f(t) is represented as a sum of Laguerre polynomialsf(t) approx sum_{n=0}^{N-1}a_{n}e^{(sigma - b)t}L_{n}(2bt)a_{n} = a_{n}(F(s),sigma,b)Critical to the accuracy of the expansion is the choice of values for the (sigma,b) parameters. This can be accomplished from a minimization of an absolute error estimate for f_{N}(t). The calculation of the error estimate for each (sigma,b) pair requires that one compute the expansion coefficients a_{n} through a fast Fourier transform [FFT]. This tool uses the parallelization of a GPU to perform a global minimization by simultaneously computing an absolute error estimate E(sigma,b) for the approximate f(t) for each (sigma,b) pair. The code also has a CPU implementation of the E(sigma,b) global search. The GPU implementation is considerably faster. A local minimization of the error estimate based on MATLAB's fminbnd function is also implemented for CPU computations. From the estimates, the (sigma,b) pair which corresponds to the smallest absolute error is selected for the expansion.The default mode for the tool is to run on a CPU. The GPU code makes use of JACKET from AccelerEyes. For the GPU computations, one must have a NVIDIA brand graphics card installed and JACKET. These functions were developed using an NVIDIA Tesla c2070 GPU, MATLAB 2010b, and JACKET 1.6 in Debian Linux.Documentation is provided in the help for each file, in the references, and in an examples script (ScrExample.m).If you use the programs or adapt them to your own needs, we would appreciate a note letting us know who you are and about your work. The files are provided as is with no guarantees.

Requirements:	No special requirements
Platforms:	Matlab
Keyword:	Accomplished,  Calculation,  Choice,  Coefficients,  Compute,  Expansion,  Minimization,  Requires,  Sigmab
Users rating:	0/10