Hi Thejs,

In KiDS-1000 we had some problems with the evidence estimate reported by MultiNest, which was significantly different from the fiducial CosmoSIS result. This is caused by the implementation of Gaussian priors in MontePython by adding an additional Gaussian likelihood for parameters with Gaussian priors. CosmoSIS instead directly samples from a Gaussian distribution, so that in doesn't require mixing the likelihood and the prior. While the posteriors are in perfect agreement, it leads to inconsistent evidence estimates, which for example becomes relevant when quantifying tensions between datasets via the Suspiciousness index.

I added the option to directly sample from a Gaussian distribution (by mapping the interval [0,1] onto a Gaussian in map_from_unit_interval). Mean and sigma can be defined in the .param file by simply adding ‘gaussian’, mean, and sigma to the list. The option to define priors as either flat or Gaussian was already set up, but I believe it wasn't really used.

At the moment it only works with MultiNest, but at least adding this feature to the PolyChord sampler would be straightforward.
Feel free to add it to the repo if you find it useful!

Best,
Benjamin