The current SciPy version of the Levenberg-Marquardt algorithm only accepts an array of floats, which will be squared and summed in the algorithm itself. This means that the likelihood fitting, where negative numbers are the norm, is not able to use this algorithm and has to use a scalar minimization. The difference in number of iterations in order to reach convergence seems to be sufficient to warrent investigating the possibility of implementing our own version of the algorithm to make use of the increased speed. The point is that the algorithm should be able to handle complex values.
The current SciPy version of the Levenberg-Marquardt algorithm only accepts an array of floats, which will be squared and summed in the algorithm itself. This means that the likelihood fitting, where negative numbers are the norm, is not able to use this algorithm and has to use a scalar minimization. The difference in number of iterations in order to reach convergence seems to be sufficient to warrent investigating the possibility of implementing our own version of the algorithm to make use of the increased speed. The point is that the algorithm should be able to handle complex values.