classy-szlite 0.2.2

Fast, differentiable cosmology in pure JAX. CMB Cls, matter Pk (linear and nonlinear), distances, derived parameters, and halo-model tSZ Cl^yy via high-accuracy ede-v2 CosmoPower emulators.

classy_szlite

Fast, differentiable cosmology in pure JAX.

classy_szlite provides JIT-compiled, jax.grad-friendly access to:

• CMB angular power spectra — TT, TE, EE
• Linear and nonlinear matter Pk — P(k, z), Pnl(k, z)
• Distances — H(z), comoving χ(z), angular-diameter D_A(z)
• Derived parameters — σ8, Ω_m, S8
• Halo-model tSZ Cl^yy — Arnaud 2010 GNFW pressure profile

Backed by the high-accuracy ede-v2 CosmoPower emulators. LCDM-equivalent cosmology is the default — fEDE = 0.001 (the package's default) reproduces standard ΛCDM to emulator precision, and the same emulator covers the full early-dark-energy parameter space if you want to explore it.

Runtime dependencies: jax, numpy, mcfit.

Install

pip install classy_szlite

Or from source:

git clone https://github.com/CLASS-SZ/classy_szlite
cd classy_szlite
pip install -e .

You also need the CosmoPower emulator .npz files at ~/class_sz_data/ (or the path in $CLASSY_SZLITE_DATA_DIR). See Installation.

Quick start

import jax.numpy as jnp
import classy_szlite as csl

cosmo = csl.CosmoParams()                      # Planck-18 + ede-v2 LCDM-equivalent defaults

# Derived parameters
csl.derived(cosmo)
# → {'sigma_8': 0.812, 'Omega_m': 0.311, 'S8': 0.827, 'der_full': ...}

# CMB Cls (dimensionless D_ℓ; × Tcmb² for μK²)
csl.cl_TTTEEE(cosmo)
# → {'ell', 'tt', 'te', 'ee'}

# Matter Pk at multiple z
k, pk  = csl.Pk(cosmo,  [0., 0.5, 1., 2.])
k, pnl = csl.Pnl(cosmo, [0., 0.5, 1., 2.])

# Distances
Hz, chi, Da = csl.distances(cosmo, [0.1, 0.5, 1.0])

# Halo-model tSZ Cl^yy
profile = csl.ProfileParamsA10(P0=8.13, beta=5.48, B=1.25)
ell = jnp.geomspace(2, 9000, 80)
cl_1h, cl_2h = csl.cl_yy(cosmo, profile, ell)

# MCMC fast path: precompute cosmology + halo grids once → ~5 ms/call
ev = csl.cl_yy_factory(cosmo, ell)
cl_1h, cl_2h = ev(profile)

Throughput

Warm-call timing, n = 100 calls per benchmark, freshly randomised inputs:

Function	mean ± std (ms)	calls/s
derived	0.54 ± 0.04	1850 ± 150
cl_TTTEEE	2.52 ± 0.14	400 ± 25
Pk	1.49 ± 0.12	670 ± 55
distances	1.29 ± 0.09	770 ± 60
cl_yy (full pipeline)	17.84 ± 0.58	56 ± 2
cl_yy_factory (fixed-cosmo)	5.38 ± 0.42	185 ± 15
cl_yy_factory + jax.grad	17.12 ± 1.01	58 ± 3

Reference platform: macOS arm64 (M-series CPU), single-thread JAX. See Throughput for a more detailed table + reproduction script.

JAX gradients

All public functions are JAX-traceable. The factory closure is the recommended path for gradient-based inference at fixed cosmology:

import jax
ev = csl.cl_yy_factory(cosmo, ell)

def loss(P0, beta):
    cl_1h, cl_2h = ev(csl.ProfileParamsA10(P0=P0, beta=beta, B=1.25))
    return jnp.sum(cl_1h + cl_2h)

d_loss = jax.grad(loss, argnums=(0, 1))(8.13, 5.48)

Gradients also work through the full pipeline (cosmology + profile) and w.r.t. the whole CosmoParams pytree — see Gradients.

Documentation

Full docs at https://classy-szlite.readthedocs.io.

License

MIT.