Bispectrum-Based SMICA Analysis

• Bispectrum-based SMICA is an extension of the SMICA framework that integrates three-point correlations to capture non-Gaussian signals in CMB data.
• It employs a hybrid two-step methodology: first fitting the standard power spectrum, then conditioning on bispectrum estimates via a binned approach to manage computational complexity.
• Simulations demonstrate that the method reliably recovers foreground bispectra while propagating uncertainty, enhancing constraints on primordial non-Gaussianity.

Bispectrum-based SMICA is an extension of the Spectral Matching Independent Component Analysis (SMICA) framework for component separation in cosmic microwave background (CMB) data analysis, designed to exploit information beyond second-order statistics. In contrast to standard SMICA, which operates under the assumption of Gaussianity and focuses on matching empirical and model power spectra, the bispectrum-based formalism targets non-Gaussian features—especially those arising from foreground components such as polarized dust and synchrotron emission. The key innovation lies in incorporating the three-point correlation function, or bispectrum, directly into the component separation process, rather than treating non-Gaussianity as a post-cleaning residual property. The approach is motivated by the fact that CMB foregrounds often exhibit appreciable non-Gaussianity, making their accurate characterization essential for robust measurements of primordial non-Gaussianity and CMB signals (Citran et al., 27 Nov 2025).

1. Standard SMICA: Second-Order Statistical Matching

SMICA models the sky’s observed spherical-harmonic coefficients $a^d_{\ell m}$ at each frequency channel $d$ as a sum of $N_{\rm comp}$ linearly mixed, statistically independent source components $s^c_{\ell m}$ and instrumental noise $n^d_{\ell m}$: $$a^d_{\ell m} = \sum_{c=1}^{N_{\rm comp}} A^{dc} s^c_{\ell m} + n^d_{\ell m}.$$ Assuming statistical isotropy, Gaussianity, independent noise, and mutual independence of components, the Gaussian likelihood for the data vector $\{\widehat a^d_{\ell m}\}$ at each $(\ell, m)$ is

$$\mathcal{L}(A, C^c_\ell, \sigma^d_\ell | \widehat a^d_{\ell m}) = \prod_{\ell, m} \frac{1}{\sqrt{(2\pi)^{N_{\rm freq}} \det C_\ell}} \exp\left[-\frac{1}{2} (\widehat a_{\ell m})^\dagger C_\ell^{-1} \widehat a_{\ell m}\right],$$

where

$$C_\ell^{dd'} = \sum_{c} A^{dc} C^c_\ell A^{d'c} + \delta^{dd'} (\sigma^d_\ell)^2.$$

This is equivalent, up to an additive constant, to minimizing the spectral-matching Kullback–Leibler divergence: $d$0 with

$d$1

and empirical cross-spectra $d$2. The blind separation is achieved by minimizing this divergence over the mixing matrix $d$3, component power spectra $d$4, and instrumental noise spectra $d$5 (Citran et al., 27 Nov 2025).

2. Incorporation of Bispectrum and Multivariate Edgeworth Expansion

To access non-Gaussian information, Bispectrum-based SMICA extends the standard framework by incorporating three-point (bispectrum) correlations. The initial strategy proceeds by expanding the likelihood with a Multivariate Edgeworth Expansion (MEE) about the Gaussian: $d$6 where $d$7 are Hermite-tensor corrections. To leading order, the modified joint probability reads

$d$8

where $d$9 denotes the observed bispectrum and $N_{\rm comp}$0 is a variance-weighted inner product with

$N_{\rm comp}$1

The full log-likelihood (equation 38 in the reference) is

$N_{\rm comp}$2

However, direct joint optimization is impractical because: (a) the MEE is only reliable for small bispectrum amplitudes; and (b) the parameter space for bispectra $N_{\rm comp}$3 is vast ($N_{\rm comp}$4 degrees of freedom), far exceeding typical SMICA parameter counts (Citran et al., 27 Nov 2025).

3. Hybrid Bispectrum-based Methodology and Binned Estimation

Given the prohibitive dimensionality of joint likelihood optimization with the MEE, the approach transitions to a hybrid two-step scheme:

1. Standard SMICA Power Spectrum Fit: First, a conventional SMICA run fits for $N_{\rm comp}$5, $N_{\rm comp}$6, and $N_{\rm comp}$7 ignoring bispectrum information.
2. Conditioned Bispectrum Estimation: With these parameters fixed, bispectra are then estimated by maximizing a Gaussian likelihood for $N_{\rm comp}$8: $N_{\rm comp}$9 A simple multi-component model is adopted for the bispectrum: $s^c_{\ell m}$0 Templates (local, equilateral, orthogonal) are used for dust and synchrotron, and a primordial CMB bispectrum proportional to $s^c_{\ell m}$1 can be included.

To address computational intractability, the bispectrum is binned. Multipole binning, as in equation 81, segments $s^c_{\ell m}$2 into 12 predefined bins, mapping the order $s^c_{\ell m}$3 initial bispectrum configurations to 364 bins per even/odd component. The even part is defined as: $s^c_{\ell m}$4 with $s^c_{\ell m}$5 as the Wigner-3j weighted normalization (Citran et al., 27 Nov 2025).

4. Simulation-Based Validation and Performance

Simulations were performed on 400 Monte Carlo realizations of polarization-only (E and B modes) LiteBIRD-like skies, for $s^c_{\ell m}$6, and 15 frequency channels between 40 and 402 GHz. Components included Gaussian CMB ($s^c_{\ell m}$7), PySM d0 dust and s0 synchrotron templates (no spatial SED variation), and instrumental noise. Power-spectrum fits used partial sky masking ($s^c_{\ell m}$8), while the bispectrum estimator operated on the unmasked sky to reduce mask-induced variance.

Empirically, applying the full joint likelihood (PS+BS) yielded parameter shifts in $s^c_{\ell m}$9 that were negligible compared to Gaussian-only SMICA ($n^d_{\ell m}$0 relative). This indicates that at current sensitivities and sky coverage, bispectrum corrections are too weak to improve power spectrum or mixing parameter constraints.

In the conditioned bispectrum estimation, the method robustly recovers dust and synchrotron bispectra (particularly in squeezed and equilateral bins), with mean residuals within the $n^d_{\ell m}$1 scatter of simulations. The squeezed configurations show the highest detectability. For primordial non-Gaussianity, the fitted value of $n^d_{\ell m}$2 using the multi-frequency multi-component bispectrum fit is $n^d_{\ell m}$3, consistent with the traditional pipeline result ($n^d_{\ell m}$4), but with the advantage that foreground uncertainty is naturally included in the bispectrum fit (Citran et al., 27 Nov 2025).

5. Advantages, Limitations, and Open Issues

The bispectrum-based SMICA methodology offers significant conceptual and practical advances:

• Foreground Uncertainty Propagation: By performing bispectrum estimation within the component separation step, one accounts for both foreground uncertainty and channel covariance in non-Gaussianity constraints.
• Simultaneous Separation: Multiple non-Gaussian components can be handled jointly, yielding a single estimator for both foreground and primordial bispectra, instead of the standard two-step process (clean map $n^d_{\ell m}$5 bispectrum).
• Dimensionality Reduction: Binning and simple component modeling render the estimation tractable, compressing the vast bispectrum space.

However, several limitations remain:

• Power Spectrum Constraints: The bispectrum-augmented likelihood does not tighten spectral or mixing matrix parameter constraints, due to the smallness of bispectrum corrections at low $n^d_{\ell m}$6 and, potentially, the need to include higher-order moments like the trispectrum.
• Model Dependence: Foreground bispectrum models are currently empirical (e.g., local/equilateral/orthogonal shapes scaled by power spectra), with residual mask dependence; physics-driven templates could reduce modeling bias.
• Computational Cost: Joint PS+BS fits with the full bispectrum parameterization are infeasible; binning remains essential.
• Extension to Larger Datasets: Application to higher $n^d_{\ell m}$7 (e.g., Planck), full T/E/B coverage, and masked skies will require new algorithmic optimizations and control of mask-induced variance via linear-correction terms (Citran et al., 27 Nov 2025).

6. Impact and Prospects for CMB Component Separation

The demonstration by Citran et al. shows that bispectrum-based SMICA can reliably recover both foreground and primordial bispectra while properly propagating multi-frequency uncertainty. The approach constitutes a substantive advance for CMB data analysis, particularly in the precise separation of non-Gaussian foregrounds and in constraining the primordial bispectrum. Future development is focused on scaling to higher resolution, incorporating physically motivated foreground models, and optimizing algorithmic efficiency for next-generation CMB polarization experiments (Citran et al., 27 Nov 2025).

A plausible implication is that, as instrument sensitivity and resolution increase, the synergetic use of bispectrum-based and standard second-order techniques will be crucial—especially for robust measurements of $n^d_{\ell m}$8 and foreground characterization in the presence of complex, non-Gaussian contaminants.