Removing Real‐World Foregrounds from Cosmic Microwave Background Maps

Max TegmarkThe Astrophysical Journal 1998 · DOI · arXiv

Abstract: Most work on foreground removal has treated the case where the frequency dependence of all components is perfectly known and independent of position. In contrast, real-world foregrounds are generally not perfectly correlated between frequencies, with the spectral index varying slightly with position and (in the case of some radio sources) with time. A method incorporating this complication is presented, and illustrated with an application to the upcoming satellite missions MAP and Planck. We find that even spectral index variations as small as 0.1 can have a substantial impact on how channels should be combined and on attainable accuracy.

Introduction

This 1998 paper by Max Tegmark addresses a critical hurdle in observational cosmology: how to clean the "dirt" off our pictures of the early universe. As Cosmic Microwave Background (CMB) experiments moved toward the high-precision era of the MAP (later WMAP) and Planck missions, the simplifying assumptions of the 1990s—specifically that foregrounds have a uniform "color" across the sky—were no longer sufficient.

To study the CMB, astronomers must look through our own galaxy, which glows with synchrotron radiation, free-free emission, and thermal dust. Traditional cleaning methods assumed these foregrounds were perfectly correlated across different frequencies. In the real world, physical properties vary by location, meaning the spectral index is not a constant but a fluctuating field. If a cleaning algorithm ignores these variations, it leaves behind residuals that can be mistaken for cosmological signals, potentially biasing measurements of fundamental parameters like the Hubble constant or the density of the universe.

Methodology: Modeling Frequency Decoherence

Tegmark generalizes CMB foreground removal by relaxing the assumption of perfect frequency correlation. He frames the problem as a signal-to-noise optimization, treating foregrounds as a form of highly correlated noise.

A key innovation is the introduction of a "frequency coherence" parameter, $\xi$, which is inversely proportional to the dispersion of the spectral index ($\Delta\alpha$). By working in "mode space" (Fourier or wavelets), the algorithm exploits the fact that different foregrounds dominate at different angular scales. This allows for a weight vector ($\mathbf{w}$) that minimizes total error by balancing residual foreground contamination against instrumental noise. Tegmark compares various linear inversion techniques, demonstrating that his generalized approach is more robust than simpler marginalization methods when channel counts are limited.

Performance and Mission Comparison

The paper uses technical specifications—frequency channels, beam widths, and sensitivity levels—from the COBE, MAP, and Planck missions to simulate performance.

Tradeoff curves for COBE, MAP, and Planck missions showing the minimum achievable foreground residual for a given noise level.

As shown in the tradeoff curves above, the inclusion of many channels in the Planck mission allows for much better cleaning than earlier missions like COBE. However, the "floor" created by spectral index variation remains a limiting factor for all missions.

The Impact of Spectral Index Variation

A major insight is the "greediness trap": if you assume foregrounds are perfectly correlated ($\xi = \infty$) when they actually vary slightly, trying to "perfectly" subtract them leads to a massive spike in instrumental noise without actually removing the foregrounds effectively. Instead of deterministic subtraction, Tegmark treats foregrounds statistically as noise components with specific covariance matrices.

Performance curves for the Planck mission showing how cleaning accuracy degrades as frequency coherence ($\xi$) decreases.

The figure above demonstrates how the achievable accuracy for the Planck mission is sensitive to the coherence parameter. The paper shows that even a spectral index variation as small as $\Delta\alpha \sim 0.1$ is enough to fundamentally change how we should weight different frequency channels. Tegmark establishes that frequency coherence is roughly the inverse of the spectral index dispersion, providing a simple heuristic for researchers to estimate the reliability of cross-frequency extrapolations.

Limitations and Future Work

While this statistical approach is powerful, it has limitations. The toy model for frequency coherence assumes a Gaussian-like decay of correlation, which may not capture the full complexity of the interstellar medium. Furthermore, the effectiveness of the method depends on knowing the spectral index dispersion ($\Delta\alpha$), highlighting the need for better empirical maps of these variations. Finally, because real-world foregrounds are highly non-Gaussian, "throwing out" the most contaminated regions (like the galactic plane) remains a necessary first step before applying these statistical methods.

Summary updated

Abstract

Most work on foreground removal has treated the case where the frequency dependence of all components is perfectly known and independent of position. In contrast, real-world foregrounds are generally not perfectly correlated between frequencies, with the spectral index varying slightly with position and (in the case of some radio sources) with time. A method incorporating this complication in presented, and illustrated with an application to the upcoming satellite missions MAP and Planck. We find that even spectral index variations as small as 0.1 can have a substantial impact on how channels should be combined and on attainable accuracy.
BibTeX
@article{tegmark1998removing,
  author = {Max Tegmark},
  title = {{Removing Real‐World Foregrounds from Cosmic Microwave Background Maps}},
  journal = {The Astrophysical Journal},
  year = {1998},
  doi = {10.1086/305905},
}
Figure