classy-comparisons

CLASSy Comparisons

Use this skill when you need to compare CAMB against a local install of the latest released CLASS/classy version available from the upstream vX.Y.Z tags. CLASS/classy does not run on Windows, so these commands assume a Unix-like shell.

Installed Location

• Work directory: ${CLASS_WORKDIR:-${TMPDIR:-/tmp}/classy-comparisons}
• Tarball: ${CLASS_WORKDIR}/class_public-${CLASS_VERSION}.tar.gz
• Source tree: ${CLASS_WORKDIR}/class_public-${CLASS_VERSION}
• CLASS binary: ${CLASS_WORKDIR}/class_public-${CLASS_VERSION}/class
• Python wrapper target: ${CLASS_WORKDIR}/classy-local

This install is local to the current machine. Reinstall with the latest released CLASS tag and report the version:

CLASS_WORKDIR=${CLASS_WORKDIR:-${TMPDIR:-/tmp}/classy-comparisons}
export CLASS_WORKDIR
mkdir -p "${CLASS_WORKDIR}"
cd "${CLASS_WORKDIR}"
CLASS_TAG=$(
  python -c 'import re, subprocess
tags = subprocess.check_output(
    ["git", "ls-remote", "--tags", "--refs", "https://github.com/lesgourg/class_public.git", "v*"],
    text=True,
)
versions = []
for line in tags.splitlines():
    tag = line.rsplit("/", 1)[-1]
    if re.fullmatch(r"v\d+\.\d+\.\d+", tag):
        versions.append((tuple(map(int, tag[1:].split("."))), tag))
print(max(versions)[1])'
)
CLASS_VERSION=${CLASS_TAG#v}
CLASS_ROOT=${CLASS_WORKDIR}/class_public-${CLASS_VERSION}
CLASS_TARBALL=${CLASS_WORKDIR}/class_public-${CLASS_VERSION}.tar.gz
CLASS_URL=https://github.com/lesgourg/class_public/archive/refs/tags/${CLASS_TAG}.tar.gz
echo "Installing CLASS ${CLASS_VERSION} from ${CLASS_URL}"
rm -rf "${CLASS_ROOT}"
curl -L -o "${CLASS_TARBALL}" "${CLASS_URL}"
mkdir -p "${CLASS_ROOT}"
tar -xzf "${CLASS_TARBALL}" -C "${CLASS_ROOT}" --strip-components=1
cd "${CLASS_ROOT}"
make -j"$(python -c 'import os; print(os.cpu_count() or 1)')"
rm -rf "${CLASS_WORKDIR}/classy-local"
python -m pip install --target "${CLASS_WORKDIR}/classy-local" .
echo "Installed CLASS ${CLASS_VERSION} in ${CLASS_ROOT}"

Quick Checks

Run the CLASS binary with the reference ini file:

CLASS_WORKDIR=${CLASS_WORKDIR:-${TMPDIR:-/tmp}/classy-comparisons}
export CLASS_WORKDIR
CLASS_ROOT=$(
  python -c 'import os, pathlib
roots = sorted(pathlib.Path(os.environ["CLASS_WORKDIR"]).glob("class_public-[0-9]*"))
print(roots[-1])'
)
echo "Using CLASS ${CLASS_ROOT##*/class_public-} from ${CLASS_ROOT}"
cd "${CLASS_ROOT}"
./class explanatory.ini

Check that the Python wrapper imports:

CLASS_WORKDIR=${CLASS_WORKDIR:-${TMPDIR:-/tmp}/classy-comparisons}
export CLASS_WORKDIR
PYTHONPATH="${CLASS_WORKDIR}/classy-local${PYTHONPATH:+:$PYTHONPATH}" python -c "from classy import Class; import os, pathlib; root = sorted(pathlib.Path(os.environ['CLASS_WORKDIR']).glob('class_public-[0-9]*'))[-1]; print(f'Using CLASS {root.name.removeprefix(\"class_public-\")}'); print(Class)"

How to Run Comparisons

There is no dedicated CAMB-vs-CLASS driver in CAMB at the moment, so comparisons should be run from the root of the CAMB checkout with the local classy module on PYTHONPATH.

1. Start from the root of the CAMB checkout:

cd /path/to/camb

2. Make the local classy install visible to Python:

CLASS_WORKDIR=${CLASS_WORKDIR:-${TMPDIR:-/tmp}/classy-comparisons}
export CLASS_WORKDIR
export PYTHONPATH="${CLASS_WORKDIR}/classy-local${PYTHONPATH:+:$PYTHONPATH}"

3. Run a comparison driver from the repo root so import camb resolves to the CAMB checkout:

python your_comparison_driver.py

Matched Conventions

Always match recombination, helium, CMB temperature, lensing, units, and output normalization explicitly. For RECFAST comparisons:

pars.set_classes(recombination_model="Recfast")
class_params["recombination"] = "RECFAST"

Set the helium mass fraction explicitly on both sides; do not rely on a BBN default or a code-specific inferred value:

pars.set_cosmology(..., YHe=0.245, TCMB=2.7255)
class_params["YHe"] = 0.245
class_params["T_cmb"] = 2.7255

For Planck-era RECFAST matching in CAMB, use pars.Recomb.set_params(recfast_approx_model="planck"). CLASS/RecFastCLASS defaults are close to these Planck-era RECFAST knobs, but robust comparison drivers should still set them explicitly when the exact RECFAST approximation is part of the test.

CAMB Neutrino Mapping

Map CAMB massive neutrinos after pars.set_cosmology(...) as one CLASS ncdm species per CAMB mass eigenstate:

T_std = (4 / 11) ** (1 / 3)
n = pars.nu_mass_eigenstates
class_params.update({
    "N_ur": pars.num_nu_massless,
    "N_ncdm": n,
    "deg_ncdm": ",".join(str(pars.nu_mass_numbers[i]) for i in range(n)),
    "T_ncdm": ",".join(
        str(T_std * (pars.nu_mass_degeneracies[i] / pars.nu_mass_numbers[i]) ** 0.25)
        for i in range(n)
    ),
    "omega_ncdm": ",".join(str(pars.omnuh2 * pars.nu_mass_fractions[i]) for i in range(n)),
})

Use omega_ncdm, not m_ncdm, when trying to reproduce CAMB exactly: CAMB's mnu is first converted to omnuh2, then CAMB solves the thermal mass that gives that density, including a small velocity correction.

Example: CAMB mnu=0.21, nnu=3.044, num_massive_neutrinos=3, neutrino_hierarchy="degenerate" becomes one CLASS species with N_ur=0, N_ncdm=1, deg_ncdm=3, T_ncdm=(4/11)**(1/3)*(3.044/3)**0.25, and omega_ncdm=pars.omnuh2.

For a hand-specified CAMB model with general eigenstate degeneracies, use the same rule. For example, if CAMB has two eigenstates with nu_mass_numbers=[2, 1], nu_mass_degeneracies=[2.1, 0.944], and nu_mass_fractions=[0.75, 0.25], then CLASS should use deg_ncdm="2,1", T_ncdm=f"{T_std*(2.1/2)**0.25},{T_std*(0.944/1)**0.25}", and omega_ncdm=f"{pars.omnuh2*0.75},{pars.omnuh2*0.25}".

CLASS neutrino-related parameters to check:

• Physics mapping: N_ur, N_ncdm, omega_ncdm/Omega_ncdm/m_ncdm, T_ncdm, deg_ncdm, and ksi_ncdm. For CAMB standard thermal neutrinos use ksi_ncdm=0 or leave it unset.
• Nonthermal distributions: use_ncdm_psd_files, ncdm_psd_filenames, and ncdm_psd_parameters. Leave unset for CAMB thermal Fermi-Dirac neutrinos.
• Numerical neutrino controls: tol_ncdm_bg, tol_ncdm, tol_M_ncdm, ncdm_N_momentum_bins, ncdm_quadrature_strategy, ncdm_maximum_q, l_max_ncdm, ncdm_fluid_approximation, and ncdm_fluid_trigger_tau_over_tau_k. Record these when changed.
• Massless-neutrino perturbation parameters: ceff2_ur, cvis2_ur, ur_fluid_approximation, and ur_fluid_trigger_tau_over_tau_k. Leave at defaults for standard CAMB comparisons unless deliberately testing altered massless-neutrino physics.

Comparison Rules

• Match cosmological parameters explicitly. Do not assume CAMB and CLASS defaults use the same neutrino, helium, or recombination conventions.
• Compare like with like: unlensed with unlensed, lensed with lensed, raw C_l with raw C_l, and the same unit normalization.
• Put both codes on a common ell, k, or redshift grid before taking differences.
• Record any non-default CLASS precision file or precision overrides used in the comparison.
• When debugging a discrepancy, first confirm that CLASS runs standalone before changing CAMB or CLASS precision settings.

Lensed CMB Accuracy Checks

• Lensed spectra have high-ell convolution edge effects. For a plotted range ending at lmax_plot, compute CLASS lensed spectra beyond the plotted range and truncate only for the comparison. As a starting point, use l_max_scalars = lmax_plot + 2000
• For purely numerical comparisons, match the effective lensing/source cutoff, not just the named accuracy setting. If CAMB's reference has max_eta_k, first convert it to a physical source cutoff using CAMB's conformal age:

camb_kmax_1Mpc = camb_max_eta_k / camb_eta0

• For CLASS 3.2 and later, the default CMB lensing-potential calculation uses the full-Limber scheme (want_lcmb_full_limber = yes) for lCl. Match the CAMB source reach by setting

class_params["k_max_limber_over_l_max_scalars"] = camb_kmax_1Mpc / class_l_max_scalars

Here class_l_max_scalars is the CLASS compute value, including any high-ell margin used before truncating to the plotted range. If you change the CLASS compute l_max_scalars, recompute this ratio so the physical k_max [1/Mpc] remains fixed.

• Do not use k_max_tau0_over_l_max to match the high-k CMB lensing reach for CLASS 3.2+ full-Limber lCl comparisons. That parameter controls the primary CMB transfer/source cutoff and can make flat lensed-CMB runs much slower without matching the full-Limber lensing cutoff being tested. Use it only when deliberately matching primary-CMB transfer cutoffs, or explicitly setting want_lcmb_full_limber = no.

• P_k_max_h/Mpc and P_k_max_1/Mpc control output matter-power calculations. They are not the CMB-lensing source cutoff and should not be used to match CAMB lensed-CMB source reach unless the comparison is explicitly about exported matter power.

Timing Hygiene

• Set thread-control environment variable OMP_NUM_THREADS before importing CLASS if specific thread number is needed. Disgard first run which may be doing one off cached calculations (esp. for non-flat).

CLASS-native Commands

Useful CLASS commands from the extracted source tree:

CLASS_WORKDIR=${CLASS_WORKDIR:-${TMPDIR:-/tmp}/classy-comparisons}
export CLASS_WORKDIR
CLASS_ROOT=$(
  python -c 'import os, pathlib
roots = sorted(pathlib.Path(os.environ["CLASS_WORKDIR"]).glob("class_public-[0-9]*"))
print(roots[-1])'
)
echo "Using CLASS ${CLASS_ROOT##*/class_public-} from ${CLASS_ROOT}"
cd "${CLASS_ROOT}"
./class default.ini
./class explanatory.ini cl_permille.pre
python CPU.py -h