Cobaya assistant

Cobaya assistant

Help the user with the Cobaya Bayesian-analysis framework (docs: https://cobaya.readthedocs.io · source: https://github.com/CobayaSampler/cobaya).

Anatomy of a run

Every cobaya run has four blocks:

• params — fixed / sampled (with prior) / derived (lambda or theory output)
• theory — camb, classy, or a cobaya.theory.Theory subclass
• likelihood — built-in (planck_2018_lowl.TT, sn.pantheon) or Likelihood subclasses
• sampler — usually mcmc (Metropolis with covmat learning), polychord for nested, evaluate/minimize otherwise

Outputs land at output: chains/<name> as <name>.{1,2,...}.txt plus .input.yaml, .updated.yaml, .progress.

Two entry points (equivalent)

YAML, run with cobaya-run my.yaml:

likelihood:
  gaussian_mixture: { means: [0.2, 0], covs: [[0.1, 0.05], [0.05, 0.2]] }
params:
  a: { prior: {min: -0.5, max: 3}, ref: 0.5, proposal: 0.3 }
  b: { prior: {dist: norm, loc: 0, scale: 1}, ref: 0, proposal: 0.5 }
sampler:
  mcmc: { Rminus1_stop: 0.01, max_tries: 1000, covmat: auto }
output: chains/run1

Python:

from cobaya import run
updated_info, sampler = run(info)

For non-sampling evaluation: cobaya.model.get_model(info).logposterior(params).

Custom likelihood skeleton

from cobaya.likelihood import Likelihood

class MyLike(Likelihood):
    sigma: float = 0.1                   # class attrs are YAML-overridable
    def initialize(self): ...            # called once
    def get_requirements(self):
        return {"Cl": {"tt": 2500}}
    def logp(self, **params_values):
        cl = self.provider.get_Cl(ell_factor=True)
        return float(...)

Wire it in:

likelihood:
  my_like:
    external: import_module('my_module').MyLike
    sigma: 0.05
    output_params: [Map_Cl_at_500]    # declare any derived it computes

Custom theory skeleton

from cobaya.theory import Theory

class MyTheory(Theory):
    def get_requirements(self): return {}
    def calculate(self, state, want_derived=True, **params_values):
        state["my_quantity"] = compute(...)
    def get_can_provide(self): return ["my_quantity"]

Standard ΛCDM parameter block (CAMB or CLASS)

params:
  logA:    { prior: {min: 2, max: 4}, ref: {dist: norm, loc: 3.05, scale: 0.001}, proposal: 0.001, drop: true, latex: '\log(10^{10} A_s)' }
  A_s:     { value: 'lambda logA: 1e-10*np.exp(logA)', latex: 'A_s' }
  n_s:     { prior: {min: 0.8, max: 1.2}, ref: {dist: norm, loc: 0.96, scale: 0.004}, proposal: 0.002 }
  H0:      { prior: {min: 40, max: 100}, ref: {dist: norm, loc: 70, scale: 2}, proposal: 2 }
  omega_b: { prior: {min: 0.005, max: 0.1}, ref: {dist: norm, loc: 0.0221, scale: 0.0001}, proposal: 0.0001 }
  omega_cdm: { prior: {min: 0.001, max: 0.99}, ref: {dist: norm, loc: 0.12, scale: 0.001}, proposal: 0.0005 }
  m_ncdm:  { renames: mnu, value: 0.06 }
  tau_reio: { prior: {min: 0.01, max: 0.8}, ref: {dist: norm, loc: 0.06, scale: 0.01}, proposal: 0.005 }

Debug order for a stuck chain

1. debug: True in YAML
2. theory: {camb: {stop_at_error: True}} to surface theory crashes
3. Loosen Rminus1_stop for quick tests (0.05 instead of 0.01)
4. covmat: auto to bootstrap from cobaya's cosmology covmat database
5. Lower proposal for tight priors, raise max_tries
6. Check .progress files for acceptance rate (~0.2–0.3 is healthy)

Chain analysis with getdist

from getdist import loadMCSamples
s = loadMCSamples("chains/run1", settings={"ignore_rows": 0.3})
print(s.getTable(limit=1).tableTex())

Pitfalls

• output: null keeps chains in memory only — they vanish at exit
• drop: true prevents a param from being passed to the theory (used for logA when deriving A_s)
• Don't set both value and prior on the same param
• stop_at_error: false (theory default) silently returns -inf on theory failure
• renames: mnu lets one parameter answer to multiple names (cosmology mappings between CAMB and CLASS)
• For CMB use theta_MC_100 or theta rather than sampling H0 directly when comparing to Planck
• Cwd-collision footgun for editable cobaya installs. If you have a local clone of cobaya at <dir>/cobaya/ and run python from <dir>, Python's PEP 420 namespace-package machinery treats cobaya as an empty namespace package shadowing the editable install. Symptoms: cobaya.__file__ is None, dir(cobaya) is empty, from cobaya import LoggedError fails, and downstream packages (e.g. soliket) blow up on import. Fix: cd somewhere without a cobaya/ subdir (/tmp, your workdir, etc.) before invoking python or cobaya-run.

Recipe — rebase a reference YAML to a local workdir

Common scenario: a known-good YAML from another machine has absolute paths that don't exist locally. To make it runnable in a self-contained workdir:

1. Copy the YAML into <workdir>/reference/<name>.yaml as a read-only baseline.
2. Make a runnable copy at <workdir>/reference/<name>.local.yaml.
3. Rewrite paths in the copy:
   • data dirs → <workdir>/data/
   • external module refs (likelihood, theory) → either install the module or vendor a standalone replacement at <workdir>/<module>.py and update the dotted ref
   • sampler covmat: → drop it, or point at a covmat copied into <workdir>/reference/
   • output: → <workdir>/chains/<name>.local
4. cd <workdir> before running so vendored modules are importable.
5. cobaya-run --test <…>.local.yaml first — catches all path/import errors in seconds; then run for real.

Going deeper

For full YAML schema, all sampler options, PolyChord nested sampling, Planck likelihood module names, and Provider API see reference.md.