Abstract A new method for estimating the angular power spectrum $C_{\ell}$ from cosmic microwave background (CMB) maps is presented, which has the following desirable properties. (1) It is unbeatable in the sense that no other method can measure $C_{\ell}$ with smaller error bars. (2) It is quadratic, which makes the statistical properties of the measurements easy to compute and use for estimation of cosmological parameters. (3) It is computationally faster than rival high-precision methods such as the nonlinear maximum-likelihood technique, with the crucial steps scaling as $n^2$ rather than $n^3$, where $n$ is the number of map pixels. (4) It is applicable to any survey geometry whatsoever, with arbitrary regions masked out and arbitrary noise behavior. (5) It is not a "black-box" method, but quite simple to understand intuitively: it corresponds to a high-pass filtering and edge softening of the original map followed by a straight expansion in truncated spherical harmonics. It is argued that this method is computationally feasible even for future high-resolution CMB experiments with $n \sim 10^6\text{--}10^7$. It is shown that $C_{\ell}$ computed with this method is useful not merely for graphical presentation purposes, but also as an intermediate (and arguably necessary) step in the data analysis pipeline, reducing the data set to a more manageable size before the final step of constraining Gaussian cosmological models and parameters—while retaining all the cosmological information that was present in the original map.
Introduction
The angular power spectrum ($C_\ell$) of the Cosmic Microwave Background (CMB) is often described as a "gold mine" for modern cosmology. Encoded within its peaks and valleys are the fundamental parameters of our universe, including the Hubble constant, the density of baryonic and dark matter, and the cosmological constant. However, extracting this information from raw sky maps presents a formidable statistical and computational challenge. As satellite missions like MAP (WMAP) and Planck were being conceptualized, the community faced a "computational wall": standard estimation techniques were becoming too slow to handle the massive influx of pixel data.
This paper introduces a quadratic estimator that is both computationally efficient and statistically "lossless." By shifting the scaling of the analysis from $O(n^3)$ to $O(n^2)$, this method enables the processing of megapixel maps while ensuring that no cosmological information is discarded during the compression from maps to power spectra.
The Computational Challenge
The primary hurdle in CMB analysis is the size of the dataset. For a map with $n$ pixels, the standard Maximum Likelihood (ML) estimation methods typically scale as $n^3$. While manageable for the few thousand pixels of early experiments, this scaling becomes prohibitive for the million-pixel maps produced by modern satellites.
The goal is data compression: reducing millions of pixels into a few hundred or thousand power spectrum values ($C_\ell$). However, most compression methods are "lossy," meaning they discard some of the statistical power inherent in the original map. To be truly effective, a method must be "lossless," ensuring that the Fisher information matrix—the measure of how much information a dataset contains about a parameter—remains unchanged after the transformation.
The Quadratic Estimator
Tegmark proposes a quadratic estimator for the power spectrum of the form:
$$\tilde{C}_\ell = \mathbf{x}^t \mathbf{E}_\ell \mathbf{x} - b_\ell$$
Where $\mathbf{x}$ is the map data, $\mathbf{E}_\ell$ is a weight matrix, and $b_\ell$ is a bias subtraction term. Unlike non-linear maximum likelihood searches, which require iterative and expensive optimizations, this quadratic form allows for direct calculation.
Key Innovations:
- Optimality and the Cramer-Rao Bound: Using the Fisher information matrix formalism, the paper proves that this specific estimator is "unbeatable." It achieves the smallest possible error bars theoretically allowed by the laws of statistics.
- Map "Tilting" ($C^{-1}$): A critical step involves multiplying the map by the inverse of its covariance matrix ($\mathbf{C}^{-1}$). This operation acts as a sophisticated high-pass filter. Since the CMB power spectrum is "red" (dominated by large-scale fluctuations), the $C^{-1}$ operation suppresses these large-scale modes—which are often plagued by cosmic variance—allowing the estimator to focus on the more subtle, information-rich small-scale signals.
- Natural Apodization: The method automatically handles survey geometry. It performs "edge softening" or apodization, naturally downweighting pixels near the boundaries of a mask or survey area. This prevents "ringing" in the Fourier domain and leads to cleaner, narrower window functions.
Computational Efficiency
The breakthrough in speed comes from the realization that the most expensive part of the calculation—solving linear systems involving the covariance matrix—can be handled via iterative solvers like the Gauss-Seidel method. This reduces the complexity to $O(n^2)$, making the analysis of $10^6$ to $10^7$ pixels feasible on contemporary (late-1990s) hardware. This efficiency transforms power spectrum estimation from a final endpoint into a viable intermediate step in a larger data analysis pipeline.
Validation with COBE/DMR
To demonstrate the method's efficacy, it was applied to the 4-year COBE/DMR dataset. After applying a custom galactic cut to remove foreground interference, the resulting 3,881-pixel map was processed to produce a binned power spectrum. The results were consistent with previous, more computationally taxing methods, but provided a clearer, more rigorous path toward the high-resolution requirements of the MAP and Planck missions.
Conclusion and Limitations
The quadratic method presented here provides a robust, intuitive, and computationally feasible framework for CMB analysis. It frames power spectrum estimation not just as a way to make a plot, but as a necessary data compression step that retains 100% of the cosmological information needed for parameter estimation.
However, some caveats remain. The method assumes that both the CMB signal and the instrumental noise are Gaussian. While this is a standard assumption, non-Gaussian foregrounds (like dust or point sources) must be meticulously removed before this estimator is applied. Additionally, for very low multipoles (large angular scales), the Gaussian approximation of the $C_\ell$ estimates themselves may begin to weaken, potentially requiring supplementary linear measures to maintain absolute information retention.
Summary updated