Galaxy bias as cobaya theory

[itrharrison]

This will implement a cobaya theory class to calculate power spectra with a bias term (e.g. linear galaxy bias).

To begin with this will just be a constant linear bias
Pk_gg = b_lin**2. Pk_mm
but other subclasses can contain more sophisticated bias models (including ones calculated by external codes such as velcileptors or fastpt).

Merging will close #33.

[itrharrison]

I'm attempting to run tests evaluating the model using the one likelihood as a dummy likelihood. You can see the tests here, but the info is:

    info = {"params": {
                       "b_lin": 1.,
                       "H0": 70.,
                       "ombh2": 0.0245,
                       "omch2": 0.1225,
                       "ns": 0.96,
                       "As": 2.2e-9,
                       "tau": 0.05
                       },
            "likelihood": {"one": None},
            "theory": {"camb": None,
                       "linear_bias": {"external": Linear_bias}
                       },
            "sampler": {"evaluate": None},
            "debug": True
           }

This currently fails both with and without the linear bias theory, when trying to run the sampler:

>       updated_info, sampler = run(info)

soliket/tests/test_bias.py:64:
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
../../../../opt/anaconda3/envs/cobaya/lib/python3.9/site-packages/cobaya/run.py:157: in run
    sampler.run()
../../../../opt/anaconda3/envs/cobaya/lib/python3.9/site-packages/cobaya/samplers/evaluate/evaluate.py:66: in run
    self.logposterior = self.model.logposterior(reference_point)
../../../../opt/anaconda3/envs/cobaya/lib/python3.9/site-packages/cobaya/model.py:419: in logposterior
    like = self._loglikes_input_params(input_params,
../../../../opt/anaconda3/envs/cobaya/lib/python3.9/site-packages/cobaya/model.py:310: in _loglikes_input_params
    result = self.logps(input_params, return_derived=return_derived, cached=cached,
../../../../opt/anaconda3/envs/cobaya/lib/python3.9/site-packages/cobaya/model.py:245: in logps
    compute_success = component.check_cache_and_compute(
../../../../opt/anaconda3/envs/cobaya/lib/python3.9/site-packages/cobaya/theory.py:260: in check_cache_and_compute
    if self.calculate(state, want_derived, **params_values_dict) is False:
../../../../opt/anaconda3/envs/cobaya/lib/python3.9/site-packages/cobaya/theories/camb/camb.py:485: in calculate
    params, results = self.provider.get_CAMB_transfers()
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

self = <cobaya.theory.Provider object at 0x161e98f70>
name = 'get_CAMB_transfers'

    def __getattr__(self, name):
        if name.startswith('get_'):
            requirement = name[4:]
            try:
                return getattr(self.requirement_providers[requirement], name)
            except KeyError:  # requirement not listed (parameter or result)
>               raise AttributeError
E               AttributeError

../../../../opt/anaconda3/envs/cobaya/lib/python3.9/site-packages/cobaya/theory.py:462: AttributeError

[cmbant]

I guess your calculate function always gets the Pk even if the class isn't actually asked to calculate anything.

Rather than sampling a mock, you could call get_model() and then add_requirements (https://cobaya.readthedocs.io/en/latest/models.html?highlight=get_model#model.Model.add_requirements) to tell it something that you need to compute. There's an example at https://cobaya.readthedocs.io/en/latest/cosmo_model.html

Or maybe you can add a "requires" argument for Pkgg to the likelihood in the yaml (not tried it).

[itrharrison]

That works, thanks! I had tried add_requirements but hadn't forced the model computation.

[damonge]

I am not sure I would pass these downstream as interpolators, as opposed to arrays. Other downstream codes (e.g. CCL, but there must be others) will use their own interpolating functions based on input arrays, so you'll need to resample these upstream interpolators and reinterpolate them, which will be inefficient and give rise to errors.

[itrharrison]

(sorry for the delay, was moving house)

Okay, I agree this seems sensible. There are a couple of SOLikeT likelihoods (xcorr, clusters) which request Pk_interpolator but it would be trivial to have them request grids and do any interpolation inside the likelihood.