All Classes Namespaces Functions Variables Typedefs Enumerations Enumerator Groups Pages
Non linear optimization module

Detailed Description

* #include <unsupported/Eigen/NonLinearOptimization>

This module provides implementation of two important algorithms in non linear optimization. In both cases, we consider a system of non linear functions. Of course, this should work, and even work very well if those functions are actually linear. But if this is so, you should probably better use other methods more fitted to this special case.

One algorithm allows to find an extremum of such a system (Levenberg Marquardt algorithm) and the second one is used to find a zero for the system (Powell hybrid "dogleg" method).

This code is a port of minpack ( Minpack is a very famous, old, robust and well-reknown package, written in fortran. Those implementations have been carefully tuned, tested, and used for several decades.

The original fortran code was automatically translated using f2c ( in C, then c++, and then cleaned by several different authors. The last one of those cleanings being our starting point :

Finally, we ported this code to Eigen, creating classes and API coherent with Eigen. When possible, we switched to Eigen implementation, such as most linear algebra (vectors, matrices, stable norms).

Doing so, we were very careful to check the tests we setup at the very beginning, which ensure that the same results are found.


The tests are placed in the file unsupported/test/NonLinear.cpp.

There are two kinds of tests : those that come from examples bundled with cminpack. They guaranty we get the same results as the original algorithms (value for 'x', for the number of evaluations of the function, and for the number of evaluations of the jacobian if ever).

Other tests were added by myself at the very beginning of the process and check the results for levenberg-marquardt using the reference data on Since then i've carefully checked that the same results were obtained when modifiying the code. Please note that we do not always get the exact same decimals as they do, but this is ok : they use 128bits float, and we do the tests using the C type 'double', which is 64 bits on most platforms (x86 and amd64, at least). I've performed those tests on several other implementations of levenberg-marquardt, and (c)minpack performs VERY well compared to those, both in accuracy and speed.

The documentation for running the tests is on the wiki

API : overview of methods

Both algorithms can use either the jacobian (provided by the user) or compute an approximation by themselves (actually using Eigen Numerical differentiation module). The part of API referring to the latter use 'NumericalDiff' in the method names (exemple: LevenbergMarquardt.minimizeNumericalDiff() )

The methods LevenbergMarquardt.lmder1()/lmdif1()/lmstr1() and HybridNonLinearSolver.hybrj1()/hybrd1() are specific methods from the original minpack package that you probably should NOT use until you are porting a code that was previously using minpack. They just define a 'simple' API with default values for some parameters.

All algorithms are provided using Two APIs :

As an example, the method LevenbergMarquardt::minimize() is implemented as follow :

* Status LevenbergMarquardt<FunctorType,Scalar>::minimize(FVectorType &x, const int mode)
* {
* Status status = minimizeInit(x, mode);
* do {
* status = minimizeOneStep(x, mode);
* } while (status==Running);
* return status;
* }


The easiest way to understand how to use this module is by looking at the many examples in the file unsupported/test/NonLinearOptimization.cpp.


class  Eigen::HybridNonLinearSolver< FunctorType, Scalar >
 Finds a zero of a system of n nonlinear functions in n variables by a modification of the Powell hybrid method ("dogleg"). More...
class  Eigen::LevenbergMarquardt< _FunctorType >
 Performs non linear optimization over a non-linear function, using a variant of the Levenberg Marquardt algorithm. More...