Matrix Cholesky not positive definite in CosmoMC

[Caterina Umilta]

Hi,

I am running into problems with CosmoMC with the Planck-BK likelihood. When I try to run it, the run stops immediately and gives out this output error:

Code: Select all

 Number of MPI processes:           1
 file_root:first_simulation
 Random seeds: 24343, 15003 rand_inst:   1
 WARNING: bin_window 1 outside cl_lmin-cl_max range: /n/home/windows/bpwf_bin1.txt
 WARNING: bin_window 2 outside cl_lmin-cl_max range: /n/home/windows/bpwf_bin2.txt
 WARNING: bin_window 3 outside cl_lmin-cl_max range: /n/home/windows/bpwf_bin3.txt
 WARNING: bin_window 4 outside cl_lmin-cl_max range: /n/home/windows/bpwf_bin4.txt
 WARNING: bin_window 5 outside cl_lmin-cl_max range: /n/home/windows/bpwf_bin5.txt
 Matrix_Cholesky: not positive definite 26
 MpiStop:            0

I do not use a propose matrix but I tried to change the propose width of parameters, with no success. The code does not even run for

Code: Select all

action = 4

I also tried without success

Code: Select all

action=2,0

I am able to successfully run the same dataset file with the same (or very similar) parameters with Cobaya. I am also able to run different dataset files with CosmoMC and the same likelihood.

Do you have any suggestion on how to debug this? I am wondering what in particular can trigger the "not positive definite" matrix this early in the run.

[Antony Lewis]

What data file is it? What exactly is the difference with the cases that do run?

[Caterina Umilta]

Hi Antony,

It is not one of the provided dataset file, it is made from a custom simulation set. The difference with respect to the cases that run is in the simulation characteristics, in particular the foregrounds.

I have done some checks on the datasets and it looks like it is the bandpower covariance matrix (the file named "covmat" in the dataset file) that is not positive-definite and causes the issue.

What is strange is that this same dataset runs smoothly with Cobaya. I am wondering if Cobaya does some extra step for non positive-definite matrices or if the algorithm is set up in a completely different way so that this problem never comes up.