nls-newton
Explicit Newton-Raphson solver for scalar and vector NLEs
Getting Started
The solver is a couple of C/C++ compliant files, one source and one header. As a result, the source file (.cpp) can be renamed (.c) freely, and used in both languages. The solver is therefore distributed as raw source code that needs to be compiled.
The key point is in the following functions:
double nlsnewton()
void nlsnewton_vec()
As they are the "actual solver", with the closest syntax as MATLAB fsolve. Some extras have been added as the relaxation parameter omega, and the dimensionality of the problem. The solver's architecture is based on dynamic allocation, which allows it to be faster than fsolve in most cases (by tuning omega).
The solver yields a txt file with vector or matrix output, as it can be seen in the /tests/ folder.
Prerequisites
To implement the solver, a standard C/C++ IDE and compiler needs to be installed on the operative system. To name a few, Visual Studio 20xx (Visual C++) for Windows or Emacs (GCC) for Linux.
Installing
-
Follow the instructions for downloading the IDE (Visual Studio, Emacs et similia) of your preference, and then build a new C or C++ project.
-
In such project, simply copy and paste the source files from /source/. Remember to paste the nlssolver.c or nlssolver.cpp file in the "resources" folder and all the nlssolver.h files in the "header" folder of the current project.
-
Compile your project either in debug or release, your choice. The nlsmain.c file is an example implementation of the solver.
The solver's function set is called by:
#include"nlssolver.h"
and the rest is just declaration and definition of objective functions and solver parameters. The latter part is no different from any common MATLAB file when using fsolve.
Once the code is compiled, you are good to go!
Running the tests
No need to think of new functions for testing!
Two test functions are already provided within the main cpp file, in the form of standard C/C++ functions. Such functions are also nonlinear, aiming to be worst-case scenarios for systems.
Deployment
The solver can be deployed within any C/C++ IDE and compiler, and virtually to any machine that support the former tools. Data analysis is performed by any tool that can process a text file, for instance Gnuplot.
For this instance, I added a MATLAB file in /source/ to exemplify the data retreival and validation procedure.
Contributing and Improvements
Contributing and redistributing the code is free!
You may choose to modify and improve it to your liking. Possible improvements may include:
-
Increase the accuracy. Can the relaxation parameter omega be modified at runtime, or be adaptive?
-
Ease the computational load. The jacobian evaluation is explicit, meaning that also its inverse is computed to solve the linear system for each Newton step. This method is by no means fast. However there are algorithms that allow linear systems to be solved without inversion (LU decomposition, Cholesky Decomposition, Gauss-Jordan elimination and so on). What would be the fastest or most efficient method?
-
Expand to non-square systems. The solver can now handle DIM * DIM square systems (at least invertible jacobians). Using the pseudo-jacobian method, it is possible to solve M * DIM nonlinear systems with M > DIM. How can this be implemented, in accordance to improvement 2?
In fact, they are limited only by intuition and imagination.
Authors
- Andrea Giuseppe Landella - alandella
License
This project is licensed under The GNU General Public License v3.0 - see the LICENSE.md file for details.
Acknowledgments
- Inspiration
- Curiosity
- Boredom
The latter may imply the former two sometimes. Or vice versa?