PaperlayerView exact revision on arXiv

Source-linked AI summary

Joint Bayesian component separation and CMB power spectrum estimation

H. K. Eriksen, J. B. Jewell, C. Dickinson, A. J. Banday, K. M. Gorski, C. R. Lawrence

arXiv:0709.1058v2astro-ph

TL;DR

Foreground contamination and systematic uncertainties must be integrated into CMB power-spectrum inference without making the computation impractical. The paper develops a Gibbs sampler that jointly samples CMB and foreground amplitudes while conditionally sampling spectral parameters, yielding exact joint posteriors and demonstrating practical WMAP-scale analysis. Its current implementation assumes identical beam responses across frequency bands and has degeneracy and prior-related scope constraints.

  • Problem

    Foreground and systematic uncertainties must be propagated into CMB power spectra and cosmological parameters because foregrounds can rival temperature signals and dominate polarization.

  • Method

    The method combines Gibbs sampling with joint sampling of CMB and amplitude parameters and conditional sampling of foreground spectral parameters.

  • Results

    The method enables exact marginalization over general foreground models and demonstrated analysis of 3-yr WMAP-like data, with foreground parameters constrained to within a few percent.

  • Takeaways & Limitations

    Foreground uncertainties can be propagated into CMB power-spectrum and cosmological-parameter posteriors within the joint Bayesian framework.

  • Takeaways & Limitations

    The implementation assumes identical beam responses across frequency bands, restricting analysis to a common lowest resolution; offset degeneracies may also require external or internal priors.

Abstract

from arXiv · show

We describe and implement an exact, flexible, and computationally efficient algorithm for joint component separation and CMB power spectrum estimation, building on a Gibbs sampling framework. Two essential new features are 1) conditional sampling of foreground spectral parameters, and 2) joint sampling of all amplitude-type degrees of freedom (e.g., CMB, foreground pixel amplitudes, and global template amplitudes) given spectral parameters. Given a parametric model of the foreground signals, we estimate efficiently and accurately the exact joint foreground-CMB posterior distribution, and therefore all marginal distributions such as the CMB power spectrum or foreground spectral index posteriors. The main limitation of the current implementation is the requirement of identical beam responses at all frequencies, which restricts the analysis to the lowest resolution of a given experiment. We outline a future generalization to multi-resolution observations. To verify the method, we analyse simple models and compare the results to analytical predictions. We then analyze a realistic simulation with properties similar to the 3-yr WMAP data, downgraded to a common resolution of 3 degree FWHM. The results from the actual 3-yr WMAP temperature analysis are presented in a companion Letter.

1. INTRODUCTION

CMB experiments require accurate foreground separation and systematic-uncertainty propagation because foregrounds can rival temperature signals and dominate polarization. The paper extends Gibbs sampling to jointly analyze foregrounds and CMB quantities while retaining favorable computational scaling.

  • Systematic-effect uncertainties must be propagated through the CMB power spectrum and cosmological parameters to avoid underestimated final uncertainties.
  • Foreground emission must be separated accurately because it can rival the CMB temperature signal and dominate polarization over most of the sky.
  • Gibbs sampling reduces likelihood-evaluation scaling from O(Npix^3) dense covariance inversion to typically O(Npix^3/2) or O(Npix^2), depending on noise correlations.The lower scaling applies to white-noise data, while correlated-noise data typically incur the higher scaling.
  • The Gibbs framework can propagate systematic uncertainties end-to-end whenever a feasible sampling algorithm exists for the modeled effect.The paper gives beam uncertainty as an example of sampling a realization within the Markov chain.
  • The paper incorporates frequency-dependent foreground signals into a Gibbs sampler using fixed templates or pixel-level amplitudes with spectral response functions.The current code assumes identical angular resolution across frequency bands and outlines a future multi-resolution generalization.
  • At 3° FWHM, a realistic 3-yr WMAP-like simulation yields the exact likelihood up to approximately ℓ∼50–60.The same tool is also used in a companion analysis of real 3-yr WMAP temperature data.

2. REVIEW OF BASIC ALGORITHMS

The paper reviews Gibbs sampling for CMB sky and power-spectrum inference, then extends the setup to multi-frequency, beam-convolved data and parametric foreground models. Sampling avoids infeasible grid or dense-matrix approaches while retaining exact posterior draws after burn-in.

  • CMB Gibbs sampling: Direct grid evaluation is infeasible because the required number of grid points grows exponentially with the number of free parameters.
  • CMB Gibbs sampling: Direct joint sampling is difficult and would generally require dense S+N covariance inversion with O(Npix^3) scaling.
  • CMB Gibbs sampling: The CMB Gibbs sampler alternates draws from P(s|Cℓ,d) and P(Cℓ|s,d) to obtain joint samples from P(s,Cℓ|d) after burn-in.
  • CMB Gibbs sampling: Given the CMB sky signal, the conditional power-spectrum distribution is an inverse Gamma distribution that has a simple textbook sampling algorithm.
  • CMB Gibbs sampling: The conditional sky-signal distribution is Gaussian with mean (S^-1+N^-1)^-1N^-1d and covariance (S^-1+N^-1)^-1.
  • Numerical implementation: Conjugate Gradients avoids brute-force inversion for current approximately 10^6 × 10^6 systems, while practical scaling is O(Npix^3/2) for white noise and O(Npix^2) for correlated noise.
  • Extensions: The reviewed framework is generalized to multi-frequency observations with instrumental beams and to parametric frequency models for total sky signals.

3. JOINT CMB AND FOREGROUND SAMPLING

The joint sampler targets the full CMB–foreground posterior by combining flexible foreground models with joint amplitude draws and conditional spectral-parameter sampling. Joint amplitude sampling improves mixing, while direct inversion sampling avoids costly Metropolis–Hastings burn-in for univariate spectral parameters.

  • Joint model: The target posterior is P(s,Cℓ,aν,i,bj,ck,θk|d) for matched-beam experiments with Aν=A.This setting supports low-resolution analysis of high-resolution experiments such as WMAP and Planck.
  • Joint model: The model combines the CMB, free-amplitude spatial templates, fixed-spectrum global templates, pixel-level foreground amplitudes, and instrumental noise.Fixed templates incorporate constraints from other measurements, while pixel-level spectra can vary spatially.
  • Sampling scheme: All amplitude-type degrees of freedom are sampled jointly, while nonlinear spectral parameters are sampled conditionally and the CMB power-spectrum sampler remains unchanged.
  • Amplitude sampling: Conditional amplitude sampling mixes poorly because strong correlations couple CMB, monopole, dipole, and template amplitudes through CMB cosmic variance.The resulting long Markov-chain correlation lengths make the approach prohibitively inefficient for general applications.
  • Amplitude sampling: Joint amplitude sampling draws from a four-component Gaussian distribution P(s,aν,i,bj,ck|Cℓ,θk,d), directly addressing the amplitude correlations.
  • Validation: The joint sampler reaches the correct regime rapidly and explores it efficiently, whereas conditional sampling converges slowly with long correlation lengths in the tested simulation.The comparison uses CMB, monopole, and synchrotron-template amplitudes.
  • Spectral-parameter sampling: The direct inversion sampler replaces Metropolis–Hastings for spectral parameters and draws from the exact univariate conditional distribution without acceptance-probability calculations.It is limited to univariate problems and may require alternatives for more complicated parameter dependencies.

4. MARGINALIZATION, PRIORS AND DEGENERACIES

The section examines marginalization, prior choice, and degeneracies in joint foreground–CMB inference. It shows that strong parameter degeneracies can distort marginal estimates, while suitable priors preserve meaningful inference, especially for the CMB component.

  • Degeneracies: The offset–amplitude degeneracy has almost no effect on the CMB component, which depends on the sum of the two components rather than their internal division.The degeneracy remains important when the foreground components themselves are of interest.
  • Priors: Jeffreys’ rule should generally replace a uniform prior when no informative prior is available, because naive uniform priors can produce biased estimates.For the spectral index example, Jeffreys’ prior compensates for the larger parameter volume associated with steep spectra.
  • Marginalization: Relaxing the prior on the synchrotron spectral index broadens and shifts the marginal free-free template amplitude posterior away from its true value.A strong prior with ∆βs = 0.01 keeps the amplitude posterior near the true zero value.
  • Degeneracies: Unknown frequency-band offsets are strongly degenerate with the overall foreground zero-level, making their individual absolute values unidentifiable without external information.A constant can be added to the foreground amplitude while subtracting a scaled value from each offset without changing the fit.
  • Degeneracies: Spatial variation in the spectral index partially resolves the offset–foreground degeneracy, but weaker variations make the covariance matrix more ill-conditioned.The Giardino model has condition number 2 × 10^7; reducing fluctuations tenfold increases it by two orders of magnitude.
  • Degeneracies: Strong degeneracies make marginal distributions difficult to summarize and require prohibitive numbers of samples to explore fully.The marginal mean is reliable only for mildly degenerate, non-Gaussian joint distributions.

5. CODE VERIFICATION

Commander’s three conditional distributions are tested against analytical calculations at low resolution. Across CMB power spectrum, Gaussian amplitudes, and spectral parameters, the implementation reproduces the analytical posteriors exactly.

  • Verification design: The verification suite tests P(Cℓ|s), the joint amplitude distribution, and P(θk|s, aν,i, bj, ck, d) against analytical expressions.These tests assess both the general sampling algorithms and their specific implementation in Commander.
  • CMB power spectrum verification: Commander reproduces CMB power spectrum distributions perfectly for four tested multipoles.The comparison uses low-resolution simulations, brute-force pixel-space likelihood slices, and Rao-Blackwellized Commander outputs.
  • Gaussian amplitude sampler: Commander reproduces the exact analytical marginal distributions for the joint Gaussian amplitude sampler.The test fixes the CMB power spectrum and synchrotron spectral-index map before comparing analytical and sampled amplitude distributions.
  • Spectral parameter verification: Commander’s spectral-index sampler shows perfect agreement with direct grid evaluation of the three-dimensional joint posterior.The single-pixel model includes an unknown synchrotron amplitude and spectral index plus an unknown lowest-frequency offset.
  • Overall result: All conditional distributions currently implemented in Commander have been verified.The verification covers the CMB power spectrum, Gaussian amplitudes, and spectral parameters.

6. APPLICATION TO SIMULATED 3-YR WMAP DATA

The simulated 3-year WMAP analysis reconstructs CMB and foreground components accurately, with strong convergence and negligible noise relative to cosmic variance on the largest scales. Model mismatch and unmodelled smoothed noise produce localized biases and underestimated uncertainties, especially beyond the primary low-ℓ regime.

  • Simulation and noise: Both instrumental and regularization signal-to-noise ratios are ≳500 at ℓ≤50, making them negligible there compared with cosmic variance.The CMB-to-regularization noise ratio reaches unity near ℓ∼120, while the regularization contribution is below 2% at ℓmax=150.
  • Convergence: About 200 iterations are required for the Gibbs sampler to reach equilibrium, after which 4000 samples achieve excellent convergence across marginal statistics.The Gelman-Rubin statistic is below 1.01 for the CMB spectrum up to ℓ∼100 and template amplitudes, and below 1.05 for CMB-pixel and foreground amplitudes.
  • Component separation: The reconstructed CMB and foreground maps are visually compelling, with no obvious CMB foreground residuals and recognizable structures in the foreground maps.The spectral-index map clearly distinguishes known synchrotron and free-free regions.
  • Model mismatch: A single power-law foreground model and dust-spectrum mismatch generate correlated foreground residuals and a spectral-index bias of about −0.1.The dust template amplitude is overestimated by 1–2%, while the high-latitude prior contributes to the spectral-index bias in noise-dominated regions.
  • Limitations: Errors in the CMB and foreground amplitudes are underestimated by a factor of ∼1.5 to 2 because smoothed instrumental noise and simplified foreground models are omitted.The unmodelled noise has negligible impact on cosmological scales ℓ≤30, but causes a slight bias at ℓ≳100 in the 3-year WMAP data.

7. CONCLUSIONS

The paper presents joint posterior sampling for CMB power spectra and flexible foreground models, enabling exact marginalization and uncertainty propagation. The current implementation is effective for large-scale analyses but assumes identical beams across frequencies.

  • The algorithm draws joint samples from P(C_ℓ, s, θ|d), enabling exact marginalization over general foreground models.Foreground uncertainties propagate through the CMB power spectrum and cosmological parameters.
  • Degeneracies in parametric models, including synchrotron–free-free spectra and unknown frequency offsets, require careful treatment.
  • The method can constrain a very general foreground model to within a few percent in all parameters for 3-yr WMAP temperature data.
  • The present implementation assumes identical beam responses across frequency bands, restricting analyses to the lowest available angular resolution.A harmonic-space generalization could accommodate individual beams and foreground power-spectrum modeling.
  • The authors project that the method could analyze Planck data to ℓ∼100–200, while recommending it as a baseline strategy for large-scale Planck analysis.This projection is based on extrapolation with angular resolution and instrumental noise.

APPENDIX THE TEMPLATE AMPLITUDE COUPLING MATRIX

The appendix expands the conditional Gaussian sampling machinery by writing the amplitude-coupling matrix in explicit block form. Its blocks encode noise-weighted couplings among the CMB, template, and foreground amplitudes.

  • The conditional distribution P(s, a_ν,i, b_j, c_k|C_ℓ, θ_k, d) is sampled as a Gaussian using its joint mean and inverse covariance.
  • The inverse covariance is organized into blocks containing noise-weighted products among amplitude operators A, T, F, and G.
  • The appendix writes each matrix block explicitly to support implementation of the sampling algorithm.
  • Noise independence between frequency channels makes cross-frequency terms in the template-amplitude block vanish.

JEFFREYS’ IGNORANCE PRIOR FOR A SPECTRAL INDEX

The appendix derives the full Jeffreys ignorance prior for a power-law foreground spectral index. The derivation incorporates noise and the antenna-to-thermodynamic conversion factor.

  • Unbiased marginal parameter estimates for a power-law foreground require a proper ignorance prior for the spectral index.
  • The derivation starts from the frequency-dependent data model and likelihood under independent noise between frequency channels.
  • Jeffreys’ prior is obtained from the ensemble-averaged second derivative of the log-likelihood.
  • The resulting expression includes the noise properties and antenna-to-thermodynamic conversion factor a(ν).

TEMPLATE ORTHOGONALITY CONSTRAINTS

The appendix addresses degeneracies between frequency-dependent template offsets and free-amplitude foreground components. It derives orthogonality constraints that make the amplitude solution unique, while acknowledging the strength and limitations of the imposed prior.

  • Unknown frequency offsets and foreground zero-levels can trade an arbitrary constant between amplitudes without materially changing χ².
  • The degeneracy can be addressed using external calibration or by removing foreground-like spectral components from the offsets.
  • The framework generalizes to any collection of fixed spatial templates with free amplitudes at every frequency.
  • For CMB and free-amplitude foreground components, the amplitude equations are solved after imposing the prior c=0.
  • Equation C4 supplies one orthogonality constraint per template, implemented in the conjugate-gradient solver through projection operators.
  • The constraint is a strong prior that can interpret random offset fluctuations with foreground-like spectra as foreground emission.It also prevents the sampler from exploring the joint distribution of offsets and foregrounds.
  • For experiments requiring this constraint, such as differential observatories, constructing a more self-consistent alternative is difficult.
Loading 0709.1058v2…