Abstract
We present a method for ultrafast confrontation of the Wilkinson Microwave Anisotropy Probe (WMAP) cosmic microwave background observations with theoretical models, implemented as a publicly available software package called CMBFIT. The method transforms cosmological parameters into a set of "normal" parameters where the WMAP likelihood surface is accurately fit by the exponential of a quartic or sextic polynomial. This approach combines the speed of previous physics-based approximations with precision cosmology grade accuracy ($\Delta \ln \mathcal{L} \sim 0.05$). We provide precalibrated fits for 6-parameter $\Lambda$CDM models and 7-parameter extensions including spatial curvature, gravitational waves, and a running spectral index.
Introduction: The Computational Bottleneck
In modern cosmology, determining fundamental constants—such as dark matter density or the universe's expansion rate—requires comparing theoretical models against high-precision data like that from the Wilkinson Microwave Anisotropy Probe (WMAP). The standard methodology involves using "Boltzmann solvers" (e.g., CMBFAST) to predict the CMB power spectrum for a given parameter set and calculating the likelihood of that prediction against observed data.
However, these solvers are computationally intensive, often requiring seconds to minutes per calculation. When combined with Markov Chain Monte Carlo (MCMC) methods that necessitate hundreds of thousands of samples, the process becomes a massive bottleneck. This is particularly problematic when researchers aim to perform joint analyses by combining CMB data with other large-scale structure datasets, such as the Sloan Digital Sky Survey (SDSS).
Methodology: Normal Parameters and Polynomial Fitting
To bypass the slow Boltzmann solvers, the authors developed CMBFIT, which utilizes a three-step approximation strategy:
- Sampling: A dense sample of the likelihood surface is generated using traditional, high-fidelity methods (CMBFAST).
- Parameter Transformation: Raw cosmological parameters often exhibit complex, non-linear degeneracies. CMBFIT transforms these into "normal parameters"—physically motivated variables like the acoustic scale ($\Theta_s$) and peak height ratios ($h_2, h_3$). In this transformed space, the likelihood surface becomes significantly more regular, resembling a multivariate Gaussian.
- Polynomial Fitting: The log-likelihood ($\ln \mathcal{L}$) is then modeled using 4th or 6th-order polynomials. While a standard Gaussian approximation (2nd-order) misses subtle non-Gaussianities in WMAP data, higher-order terms capture these nuances with high precision.
The choice of polynomial degree is critical. As shown in Figure 1, increasing the order $n$ improves the fit on the training set, but can lead to overfitting on the test set if the degree is too high for the available data density. The authors use a training/test set approach to find the optimal balance.
Results: Speed and Fidelity
The primary advantage of CMBFIT is its extreme speed. Evaluating a pre-calculated polynomial is mathematically trivial, reducing likelihood calculation times from minutes to less than a millisecond. This allows an MCMC chain of 200,000 steps to be completed in approximately one minute on a standard workstation.
Crucially, this speed does not come at the cost of accuracy. The marginalized distributions produced by CMBFIT are virtually indistinguishable from those generated by full Boltzmann solvers.
As detailed in Figure 3, the 6-parameter "vanilla" $\Lambda$CDM models achieve a test accuracy ($\Delta \ln \mathcal{L}_{test}$) of 0.05 to 0.08. Even extended 7-parameter models, including those with spatial curvature ($k$), running spectral index ($\alpha$), or gravitational waves ($r$), maintain sufficient accuracy for precision cosmology.
Discussion and Limitations
While CMBFIT offers a significant performance boost, it is model-dependent. The polynomial coefficients are specific to the parameter space used during training. Testing entirely new physical theories (e.g., modified gravity) would require generating a new set of training samples.
Furthermore, high-order polynomials can behave unpredictably outside their training range. To ensure stability during the initial "burn-in" phase of an MCMC, the authors implement a "patching" technique, using a simple Gaussian tail at the $3\sigma$ boundary. Despite these constraints, CMBFIT remains a powerful tool for rapid cosmological parameter estimation and joint-data analysis.
Summary updated