Choosing a Solver

Fitting in lightning is biased toward the Bayesian exploration of parameter space rather than maximum likelihood estimation. As such the majority of our plotting and post-processing functions assume that you’ve used emcee for fitting.

We do however include an option for fitting the SED model with the L-BFGS-B minimization algorithm, results from which are compatible with the lightning_postprocess function. There is additionally an option to solve the problem with L-BFGS-B and then, if it converges, perform a brief exploration of parameter around the best fitting solution (converting any preexisting bounds on the parameters to uniform priors). This method is potentially much faster than a brute force fit with emcee, though it risks getting stuck in a local minimum and is not possible when the solver fails to converge. The results from this MCMC followup are naturally compatible with the MCMC plotting functions.

In the future, we anticipate adding the option to use a nested sampler for exploration of the parameter space.