tomo_builder population_stats mismatch with selected_pairs

[c-d-leonard]

Currently, if you call the following:

tomography = TomographyBuilder(
    lens_survey="desi",
    source_survey="lsst",
    lens_sample='lrg',
    source_sample=None,
    lens_year=None,
    source_year="1",
    lens_role="lens",
    source_role="source",
    overlap_threshold=0.1,
    source_behind_lens=True,
    shared_overrides={
        "bins": {
            "count": 5,
        },
    },
)

tomo_inputs = tomography.prepare_bins()

bin_pairs = tomo_inputs["bin_pairs"]

print('bin_pairs=', bin_pairs)

You get

bin_pairs= [(0, 4)]

i.e. you select only the lowest redshift desi LRG bin and the highest of 5 source LSST Year 1 bins.

However, tomo_builder._population_stats doesn't receive any information about this bin pair selection, so if you naively pass tomo_inputs above as follows:

    cosmo=cosmo,
    lens_result=tomo_inputs["lens_result"],
    source_result=tomo_inputs["source_result"],
    lens_population_stats=tomo_inputs["lens_population_stats"],
    source_population_stats=tomo_inputs["source_population_stats"],
    bin_pairs=bin_pairs,
    rp_bin_edges=rp_bin_edges,
    area_deg2=5000.0,
    sigma_e=0.26,
    galaxy_bias=params['b'],
    k=np.geomspace(10**(-4), 3.0, 5000),
    hankel_kwargs={
        "r_min": 0.6,
        "r_max": 110,
        "k_min": 10**(-4),
        "k_max": 30.0,
        "orders": (2,), 
        "n_zeros": 480000, 
        "n_zeros_step": 1000,
        "prune_r": 0,
        "verbose": True,
        "max_iterations": 1000, 
    },
    taper=False,
)


dsf_cov_dict = covariance_builder.covariance_for_pair(lens_bin_index=0, 
                                                      source_bin_index=0)

the shape noise you get out is using the n_eff_per_arcmin value for the original lowest-z of the five source bins, not the highest one as I would have expected.

If instead you replace the last line with
dsf_cov_dict = covariance_builder.covariance_for_pair(lens_bin_index=0, source_bin_index=4)

you get the shape noise as I expected.

I don't think this is necessarily a full blow bug but clearly there is potential for silent failure and lack of clarity on whether the bin indices required to be passed for covariance_for_pair are the original indices or the new ones after selection. There should potentially be a warning thrown to the user or something? I'm not sure the solution but I'm worried the current case will create silent analysis failures.