Firecrown

Firecrown — DESC likelihood and parameter inference framework

You're helping a user write code with Firecrown, DESC's framework for implementing and running cosmological likelihoods. Firecrown sits downstream of TXPipe (which produces sacc files) and upstream of samplers (CosmoSIS, NumCosmo, Cobaya); it uses pyccl for all theory predictions.

Overview

Package layout:

• firecrown.likelihood — Likelihood, Statistic, Source, GaussFamily, ConstGaussian, TwoPoint, WeakLensing, NumberCounts, and all systematics
• firecrown.modeling_tools — ModelingTools (wraps pyccl.Cosmology) and CCLFactory
• firecrown.parameters — ParamsMap, RequiredParameters, UpdatableCollection
• firecrown.connector — cosmosis, numcosmo, cobaya sub-packages

Lifecycle every call: update(params) → tools.prepare(cosmo) → compute_theory_vector(tools) → compute_loglike(tools)

Systematics are applied sequentially: each systematic in a source's list receives the current source state and returns a modified version — order matters.

Entry point — every Firecrown likelihood file must define:

def build_likelihood(build_parameters: dict) -> tuple[Likelihood, ModelingTools]:
    ...
    return lk, tools

The samplers and TXPipe both call load_likelihood_from_script(path, build_parameters).

Minimal 3×2pt example (real-space):

import firecrown.likelihood.weak_lensing as wl
import firecrown.likelihood.number_counts as nc
from firecrown.likelihood.two_point import TwoPoint
from firecrown.likelihood.gaussian import ConstGaussian
from firecrown.modeling_tools import ModelingTools
import sacc

def build_likelihood(build_parameters):
    sacc_data = sacc.Sacc.load_fits(build_parameters["sacc_file"])

    sources = {}
    for name in sacc_data.tracers:
        if name.startswith("source"):
            sources[name] = wl.WeakLensing(sacc_tracer=name, systematics=[
                wl.PhotoZShift(sacc_tracer=name),
                wl.MultiplicativeShearBias(sacc_tracer=name),
            ])
        elif name.startswith("lens"):
            sources[name] = nc.NumberCounts(sacc_tracer=name, systematics=[
                nc.LinearBiasSystematic(sacc_tracer=name),
            ])

    stats = {}
    for dt in sacc_data.get_data_types():
        for t1, t2 in sacc_data.get_tracer_combinations(dt):
            stat = TwoPoint(source0=sources[t1], source1=sources[t2], sacc_data_type=dt)
            stats[f"{dt}_{t1}_{t2}"] = stat

    lk = ConstGaussian(statistics=list(stats.values()))
    lk.read(sacc_data)
    return lk, ModelingTools()

See references/api.md for systematic parameters, sacc data-type strings, and sampler connectors.

TXPipe usage

TXPipe connects to Firecrown via sacc files; some internal stages also call load_likelihood_from_script directly for theory predictions. See references/txpipe.md for the full data flow, stage list, and tracer naming conventions.

Reference