Skip to content

CHANGELOG.md

Latest commit

 

History

History
164 lines (149 loc) · 9.79 KB

CHANGELOG.md

File metadata and controls

164 lines (149 loc) · 9.79 KB
Raw

CAMB v2 Change Summary

This is a high-level summary of important user-facing and result-facing changes in CAMB v2. It is not a complete commit log; see the GitHub history and the detailed notes under docs/changelog/ for tuning, validation, and implementation details.

Accuracy Defaults And Diagnostics

  • CAMB now enables targeted accuracy improvements by default with the internal AccuracyTarget = 1. These changes are aimed at modern high-precision CMB, lensing, and matter-power use cases. Set AccuracyTarget = 0 in an .ini file, or camb.config.AccuracyTarget = 0 from Python, for behavior closer to the CAMB 1.x numerical-error profile.
  • The new camb.check_accuracy module and camb check_accuracy command compare a requested calculation to a higher-accuracy reference, report CMB, lensing potential, matter-power, and derived-parameter differences, optionally make plots, and can search for minimal accuracy boosts.
  • CAMBparams.set_for_lmax(..., lens_potential_accuracy=None) is now the public default. None selects an automatic high-accuracy lensing-potential/kmax setting, max(4, (lmax - 1500) / 500).
  • Explicit lens_potential_accuracy values keep their old meaning. In particular, use lens_potential_accuracy=0 to reproduce the old low-k default behavior. set_params_cosmomc keeps its historical default lens_potential_accuracy=1; pass None there to opt into the new automatic rule.
  • The automatic lens-potential rule is calibrated for lensed CMB spectra and lensing-potential stability at the relevant accuracy target. At high multipoles, remaining numerical errors in lensed spectra can be much smaller than the uncertainty from non-linear matter modelling.
  • lens_output_margin is now a first-class Python and .ini parameter. It consistently controls how far above the requested lensed output range CAMB calculates internally, including the Fortran lensing convolution support.

Non-Flat Models And Hyperspherical Bessel Functions

  • Non-flat scalar line-of-sight integration has been substantially refactored and sped up. The main changes are Numerov/source-grid integration, improved high-oscillation cutoffs, near-flat shifted-ν approximations, and direct Olver evaluation in high-substep ranges.
  • Near-flat open and closed models can reuse flat Bessel table machinery where controlled local error estimates allow it. This improves speed near the flat limit while preserving continuity checks.
  • The branch includes new hyperspherical Bessel implementations and validation paths: Olver-style approximations, small-chi/open-small-nu fallbacks, Airy utilities, and Python-accessible math utilities for testing.
  • In the documented branch comparison, default non-flat/matter-power cases were about three times faster than the CAMB 1.x baseline in that test set. Focused non-flat scalar integration timings show larger speedups in some curvature regimes, with low-level numerical changes documented in docs/changelog/.

Lensing Calculations

  • CAMB has a new optimized lensing method selector. It keeps the long-standing fast curved-sky method for ordinary runs and uses a full Gauss-Legendre curved-sky correlation method when AccurateBB=True.
  • The direct curved-sky lensing implementation can also be selected explicitly, and Python calls such as get_lensed_cls_with_spectrum can temporarily override the lensing method for comparisons.
  • The full-sky correlation code was optimized substantially, including cached Gauss-Legendre nodes/weights, inlined accumulation, recurrence-based factors, and faster Legendre tables exposed through camb.mathutils.
  • Low-l EE tapering and high-L template extension behavior have been clarified and made more consistent between the Python and Fortran lensing paths.

Matter Power And Non-Linear Modelling

  • Matter-power accuracy tuning was updated for massive neutrinos, photon and massless-neutrino hierarchy depths, and transfer-high-precision cases. The goal is better default agreement with boosted references without requiring broad global accuracy boosts.
  • HMCode/Halofit evaluation was cleaned up and optimized. Cached HMCode redshift-local quantities give speedups of order 10-20% in the documented matter-power benchmarks, with only tiny changes from removing unintended single-precision round trips.
  • CAMB now includes an SPkNonLinear model for the SP(k) baryon-suppression prescription, wrapping a base Halofit/HMCode model. It includes documented validity ranges, MCMC-friendly boundary behavior, and protections against double-counting baryonic feedback.
  • New non-linear model hooks include ExternalNonLinearRatio for externally supplied non-linear ratios and SecondOrderPK for second-order perturbative matter-power ratios.
  • HMcode-2020's BAO wiggle extraction is now only performed once (at z=0) and scaled by the growth factor for cosmologies with low neutrino mass fraction (below the new Halofit.HMcode_wiggle_max_fnu, default 0.01) and a dark energy model with scale-independent low-redshift growth (a new TDarkEnergyModel.assume_scale_indep_lowz_growth() method, true by default for w/wa and PPF dark energy with unit sound speed, general quintessence, and the early dark energy/axion-effective-fluid models). get_matter_power_spectrum also now computes non-linear ratios for all requested redshifts in a single call instead of one call per redshift. Together these give up to about 15% faster non-linear matter-power evaluation, at accuracy costs many orders of magnitude below HMcode's own uncertainty. See docs/changelog/2026-08-04-hmcode-wiggle-reuse-batched-nonlinear.md.

Recombination, Reionization, And Backgrounds

  • The default BBN consistency relation now uses the September 2024 PRIMAT helium and deuterium table, replacing the 2021 PRIMAT table. For typical Planck-like models this lowers the default helium mass fraction by about 2e-4, with sub-per-mille effects on fixed-parameter CMB spectra.
  • RECFAST now uses a Rosenbrock integrator while stiff. The new path is tuned against high-accuracy internal references and scales with CAMB accuracy boosts. It is intended to improve the speed/accuracy tradeoff of the recombination background calculation.
  • The default RECFAST approximation is now the recfast_cosmorec fit, including the helium-rate correction calibrated against direct CosmoRec histories. Planck-era RECFAST parameters remain available as recfast_planck and are explicitly used by Planck-specific compatibility inputs.
  • The CosmoRec wrapper was updated for the newer CosmoRec vX interface and exposes the relevant CosmoRec controls through CAMB's recombination model.
  • Reionization models now have an optional approximate heating switch that raises the baryon temperature and sound speed during reionization. It is off by default and is intended for order-of-magnitude low-redshift matter-power effects rather than precision thermal-history modelling.
  • Thermal massive-neutrino background density and pressure now use direct smooth fits over the intermediate mass range, reducing setup/global state and modestly speeding repeated background evaluations.
  • set_cosmology(mnu=...) now treats mnu as the physical sum of neutrino masses and calculates Omega_nu h^2 from the actual thermal density, rather than from the non-relativistic proxy mnu/94.07. At fixed mnu this raises omnuh2 fractionally by about 5e-5 for one 0.06eV neutrino, and by more for lighter eigenstates (about 5e-4 for three degenerate 0.02eV masses). Very light neutrinos, which previously inverted to zero mass and zero Omega_nu, now get real masses and the correct relativistic density.
  • set_cosmology(omnuh2_active=...) (with mnu=None) sets Omega_nu h^2 for the active neutrinos directly. meffsterile keeps its Planck-paper definition.
  • standard_neutrino_neff is now applied consistently to the eigenstate degeneracies and mass splitting as well as the density conversion; previously the Fortran side always used 3.044 for the former.
  • The neutrino mass inversion used to set the eigenstate masses from Omega_nu h^2 is more accurate, and no longer fails just above the minimum mass sum for the inverted hierarchy. See docs/changelog/2026-07-29-neutrino-mass-density-mapping.md.

Python Interface And New Capabilities

  • A Python bispectrum wrapper is now available as camb.bispectrum. It runs the existing Fortran CMB-lensing or local-primordial bispectrum calculation using normal CAMBparams objects, writes large tables directly to files, and returns small Fisher summaries when the library is built with Fisher support.
  • Documentation now includes pages for the bispectrum wrapper, SP(k), nonlinear models, check-accuracy workflow, and math utilities.
  • CAMB now targets Python 3.10+ and uses the ruff/pre-commit toolchain for Python formatting and linting.
  • The development tree includes updated devcontainer and CI configuration, but those changes are primarily for contributors rather than result-facing users.

Compatibility Notes

  • Numerical outputs can change relative to CAMB 1.x because the v2 branch has a higher default accuracy target, different non-flat algorithms, updated lensing support, tuned matter-power accuracy settings, and RECFAST changes.
  • For closer 1.x-style numerical behavior, start with AccuracyTarget = 0, explicit lens_potential_accuracy=0 in set_for_lmax, and fixed legacy matter-power settings where comparing against older runs.
  • Users comparing old and new results should use camb check_accuracy and compare at fixed physical output ranges and k ranges. Avoid judging changes only from sparse grid-index differences, especially for matter power and high-l lensing.
  • 32 bit builds no longer supported
  • Added experimental support for LLVM flang compiler (all platforms)