Spectral Internal Linear Combination (SpILC)

• SpILC is a component separation method that uses spectral templates and constraints to preserve targeted spectral energy distributions while minimizing variance.
• It extends standard ILC by incorporating multiple constraints, such as deprojecting foreground contaminants and handling multidimensional spectral subspaces.
• Applications include enhanced CMB analysis and polarization studies, with implementations directly optimizing auto- and cross-spectra for improved power spectrum estimations.

Searching arXiv for recent and foundational papers on Spectral Internal Linear Combination (SpILC), generalized ILC, and related constrained/moment-based variants. I’ll look up the relevant arXiv records to ground the article in the current literature. Spectral Internal Linear Combination (SpILC) denotes a class of component-separation methods in which linear combinations of multifrequency data are constrained by spectral or mixing information about the components to be preserved or suppressed. In the literature, “Spectral ILC” is usually understood as an ILC that explicitly exploits known or modeled frequency behavior through spectral templates, mixing vectors, or spectral subspaces; in a narrower later usage, it also denotes estimators constructed directly from auto- and cross-spectra rather than from component-separated maps (Remazeilles et al., 2011, Kwok et al., 12 Sep 2025). Across these usages, the defining structure is a constrained minimum-variance problem in frequency space.

1. Scope and terminology

In its broadest sense, SpILC is the spectral generalization of the Internal Linear Combination. Standard ILC preserves one component with known frequency scaling and suppresses all others only through variance minimization. Spectral ILC enriches this construction by introducing one or more explicit spectral constraints: preserve a target spectral energy distribution (SED), null one or more contaminant SEDs, preserve a multidimensional spectral subspace, or constrain responses to remain below prescribed bounds (Abylkairov et al., 2020).

This broad usage includes several closely related constructions. Generalized or multidimensional ILC treats foregrounds as a low-rank spectral subspace rather than as a single rigid SED (Remazeilles et al., 2011). Constrained ILC and partially constrained ILC use equality or inequality constraints on known contaminant SEDs, respectively (Abylkairov et al., 2020). Moment ILC augments the constraint set with spectral derivatives generated by moment expansions of foreground laws (Rotti et al., 2020). The Least Response Method replaces hard orthogonality to foreground templates by minimization of an upper bound on foreground-plus-noise variance built from prior ranges of foreground parameters (Maillard et al., 2024). In polarization, complex-coefficient and spin-wavelet variants extend the same idea to the spin-2 field $Q\pm iU$ (Fernández-Cobos et al., 2016, Rogers et al., 2016).

A narrower usage appears in the 2025 work “CMB component-separated power spectrum estimation by Spectral Internal Linear Combination,” where SpILC is defined as a minimum-variance, foreground-cleaned estimator built directly at the level of auto- and cross-power spectra. In that formulation, the optimization target is the two-point statistic itself rather than a map, and the constraint structure can be weaker than map-level cILC when some components are known to be spatially uncorrelated (Kwok et al., 12 Sep 2025).

2. Foundational constrained-variance formalism

A standard multifrequency data model writes, at pixel, harmonic, or localized-wavelet position $p$,

$$\mathbf{y}(p)=\mathbf{a}\,s(p)+\mathbf{A}_f\,\mathbf{s}_f(p)+\mathbf{n}(p),$$

where $\mathbf{a}$ is the SED vector of the desired component, $\mathbf{A}_f$ collects foreground SEDs, and $\mathbf{n}$ is instrumental noise. The ILC estimator is

$\hat{s}(p)=\mathbf{w}^T\mathbf{y}(p),$

with unit response

$\mathbf{w}^T\mathbf{a}=1,$

and minimum variance

$$\min_{\mathbf{w}}\ \mathbf{w}^T\mathbf{R}\mathbf{w},$$

where $$\mathbf{R}=\langle \mathbf{y}\mathbf{y}^T\rangle-\langle \mathbf{y}\rangle\langle \mathbf{y}^T\rangle$$ is the empirical covariance. The solution is

$$\mathbf{w}^T_{\rm ILC}=\big(\mathbf{a}^T\mathbf{R}^{-1}\mathbf{a}\big)^{-1}\mathbf{a}^T\mathbf{R}^{-1}.$$

This is already spectral in the minimal sense that it preserves a specified SED while using the covariance to downweight everything else (Abylkairov et al., 2020).

Constrained ILC extends the same quadratic program to several known SEDs. If $$\mathbf{A}=[\,\mathbf{a}\ \mathbf{b}_1\ \cdots\ \mathbf{b}_m\,]$$ and $\mathbf{e}^T=[1\ 0\ \cdots\ 0]$, the optimization becomes

$$\min_{\mathbf{w}}\ \mathbf{w}^T\mathbf{R}\mathbf{w} \quad\text{subject to}\quad \mathbf{w}^T\mathbf{A}=\mathbf{e},$$

with solution

$$\mathbf{w}^T_{\rm cILC} = \mathbf{e}^T\big(\mathbf{A}^T\mathbf{R}^{-1}\mathbf{A}\big)^{-1}\mathbf{A}^T\mathbf{R}^{-1}.$$

At this level, SpILC is simply the principle that spectral templates define the admissible response space of the estimator: target SEDs are preserved, undesired SEDs are nulled, and the covariance supplies the minimum-variance weights (Abylkairov et al., 2020).

3. Multidimensional spectral subspaces and localized implementations

The 2011 generalized-ILC formulation extends ILC beyond a single mixing vector by treating complicated foregrounds as a multidimensional spectral subspace. In a given needlet domain $\mathcal{D}^{(j)}_k$, the data are modeled as

$$\mathbf{y}=\mathbf{A}\mathbf{x}+\mathbf{n},\qquad \mathbf{A}= \begin{bmatrix} \mathbf{a} & \mathbf{F} \end{bmatrix},\qquad \mathbf{x}= \begin{bmatrix} s\\ \mathbf{g} \end{bmatrix},$$

where $\mathbf{a}$ is the CMB mixing vector and $\mathbf{F}$ is an unknown foreground mixing matrix whose column space, rather than individual columns, is inferred from the data. The foreground emission is $\mathbf{f}=\mathbf{F}\mathbf{g}$, with covariance

$$\mathbf{R}_{\rm FG}=\langle \mathbf{f}\mathbf{f}^T\rangle = \mathbf{F}\mathbf{G}\mathbf{F}^T.$$

Instead of a scalar estimator, the method reconstructs a vector-valued foreground map $\hat{\mathbf{f}}=\mathbf{W}\mathbf{y}$ with the constraint

$\mathbf{W}\mathbf{F}=\mathbf{F},$

and minimizes

$$\min_{\mathbf{W}}\ \mathrm{Tr}\big(\mathbf{W}\mathbf{R}\mathbf{W}^T\big) \quad\text{subject to}\quad \mathbf{W}\mathbf{F}=\mathbf{F}.$$

The solution is

$$\mathbf{W}_{\rm FG} = \mathbf{F}\big(\mathbf{F}^T\mathbf{R}^{-1}\mathbf{F}\big)^{-1}\mathbf{F}^T\mathbf{R}^{-1}.$$

This is the subspace analogue of the one-dimensional ILC solution: the preserved object is now a spectral subspace rather than a single SED vector (Remazeilles et al., 2011).

The same work implements this construction in needlet space. Harmonic bandpass filters $h_\ell^{(j)}$ satisfy

$\sum_j [h_\ell^{(j)}]^2=1,$

and define localized coefficients

$$\gamma^{(j)}(p)=\sum_{\ell m} y_{\ell m}\, h_\ell^{(j)}\,Y_{\ell m}(p).$$

Local covariance is estimated as

$$\widehat{\mathbf{R}}_{ab}^{(j)}(p) = \frac{1}{N_p}\sum_{p'\in\mathcal{D}_p}\gamma_a^{(j)}(p')\gamma_b^{(j)}(p').$$

After noise whitening,

$$\mathbf{R}_N^{-1/2}\widehat{\mathbf{R}}\mathbf{R}_N^{-1/2} = \widehat{\mathbf{U}}\,\widehat{\mathbf{\Delta}}\,\widehat{\mathbf{U}}^T,$$

significant eigenmodes with eigenvalues $>1+\varepsilon$ define the signal subspace, from which the CMB direction and the orthogonal foreground complement are extracted. In an extended model

$$\mathbf{A}= \begin{bmatrix} \mathbf{a} & \mathbf{b} & \mathbf{F} \end{bmatrix},$$

the same construction yields a foreground subspace orthogonal to both CMB and thermal SZ, so that the foreground estimate has vanishing CMB and SZ contamination in expectation. This localized, low-rank, subspace-preserving formulation is one of the clearest antecedents of later SpILC language (Remazeilles et al., 2011).

4. Moment, partial, and least-response formulations

One branch of SpILC development sharpens the spectral model itself. In Moment ILC, the frequency dependence of foregrounds is expanded around pivot parameters, so that a single physical foreground generates a hierarchy of moment SEDs. The constrained ILC matrix $V$ then includes not only target signals such as the CMB, $y$, and $\mu$, but also base foreground SEDs and their derivatives. The recovered component vector is

$$\hat{\vec{\tau}}_i = (V^T\mathcal{C}^{-1}V)^{-1}V^T\mathcal{C}^{-1}\mathbf{d}_i,$$

so each moment becomes an explicit constrained spectral mode. This construction was introduced to recover tiny monopolar spectral distortion signals in the presence of realistic foregrounds and instrumental noise, and it was also shown to improve CMB anisotropy measurements by reducing foreground biases and signal uncertainties when using the $\mu$ILC (Rotti et al., 2020).

A second branch relaxes hard nulling. Partially constrained ILC interpolates between standard ILC and cILC by replacing exact deprojection with inequality constraints such as

$|\mathbf{w}^T\mathbf{b}_1|\le\epsilon.$

For one partially constrained foreground, the method either reduces to standard ILC when the inequality is inactive or lands on one of the boundary solutions

$$\mathbf{w}_{\rm pcILC,\pm}^T\mathbf{b}_1=\pm\epsilon.$$

The stated motivation is to optimize the tradeoff between foreground bias and variance. For cleaning tSZ contamination at $\ell\in[3000,4800]$, if a small tSZ residual of $20\%$ of the standard ILC residual can be tolerated, the variance of the CMB temperature map is reduced by at least $50\%$ over the cILC value; for simultaneous tSZ and CIB constraints, cILC variance can be $>40\times$ ILC on high-$\ell$ scales, whereas pcILC significantly reduces the variance while bounding both residuals (Abylkairov et al., 2020).

A third branch replaces explicit hard orthogonality by covariance penalization built from prior foreground ranges. The Least Response Method assumes bounded parameter domains and amplitude bounds for foreground SEDs, constructs

$$\boldsymbol{\Phi}=\sum_{m=1}^M A_m^2\, q_{ij}^m,\qquad q_{ij}^m=\frac{1}{V_\Omega}\int_\Omega f_m(\nu_i,\mathbf{P})f_m(\nu_j,\mathbf{P})\,d\mathbf{P},$$

and minimizes

$$\boldsymbol{\omega}(\boldsymbol{\Phi}+\mathbf{C})\boldsymbol{\omega}^T$$

subject to

$\boldsymbol{\omega}\mathbf{I}_d^T=1.$

The resulting weights are

$$\boldsymbol{\omega}_{\rm LRM} = \mathbf{I}_d(\boldsymbol{\Phi}+\mathbf{C})^{-1} \left[\mathbf{I}_d(\boldsymbol{\Phi}+\mathbf{C})^{-1}\mathbf{I}_d^T\right]^{-1}.$$

For CMB spectral distortions, this is a prior-informed spectral ILC in which poorly defined foreground shapes are handled through an effective covariance rather than a finite set of exact nulling constraints; in the quoted comparisons, LRM yields a smaller total response to foreground plus noise than MILC in the considered dust+CIB and all-foreground cases (Maillard et al., 2024).

5. Power-spectrum SpILC

The 2025 formulation gives the term SpILC a more specific meaning: minimum-variance, foreground-cleaned estimators of CMB power spectra constructed directly from channel auto- and cross-spectra. For frequency channels $i,j$,

$$\hat{C}_\ell^{ij}=\frac{1}{N_p}\sum_p x_p^i x_p^{j*},$$

and the SpILC estimator is

$$\hat{K}_\ell = W_\ell^{ij}\,\hat{C}_\ell^{ij}, \qquad W_\ell^{ij}=W_\ell^{ji}.$$

The target is not a map but an unbiased estimator of $C_\ell^{ss}$, the auto-spectrum of the component of interest (Kwok et al., 12 Sep 2025).

In the one-component case, SpILC and ordinary map-level ILC are equivalent at the spectral level. With $t_\mu=a^ia^j$ and spectral covariance matrix

$$D_\ell^{\mu\nu} = C_\ell^{(i|k}C_\ell^{|j)l} = \frac{1}{2}\big(C_\ell^{ik}C_\ell^{jl}+C_\ell^{il}C_\ell^{jk}\big),$$

the weights are

$$w_\ell^\mu = \frac{D^{-1}_{\ell,\mu\nu}t_\nu}{t_\rho D^{-1}_{\ell,\rho\tau}t_\tau}.$$

In this limit, the paper proves that

$$W^{\rm SpILC}_{\ell,ij}=w^{\rm ILC}_{\ell,i}w^{\rm ILC}_{\ell,j},$$

so the SpILC spectrum is exactly the auto-spectrum of the standard ILC map (Kwok et al., 12 Sep 2025).

The genuinely new regime arises with deprojection. For two modeled components,

$x_p^i=a^is_p+b^iy_p+n_p^i,$

strong-cSpILC imposes the full spectral set

$$W_\ell^{ij}a^ia^j=1,\qquad W_\ell^{ij}b^ib^j=0,\qquad W_\ell^{ij}a^{(i}b^{j)}=0,$$

and is mathematically equivalent to map-level cILC. Weak-cSpILC drops the cross-term constraint whenever $C_\ell^{sy}=0$ is physically known, retaining only

$W_\ell^{ij}a^ia^j=1,\qquad W_\ell^{ij}b^ib^j=0.$

This yields fewer constraints than cILC while preserving unbiasedness of the target power spectrum under the stated uncorrelatedness assumption. The corresponding weights are

$$w^\mu_\ell = \frac{ (\mathbf{u}^TD_\ell^{-1}\mathbf{u})D^{-1}_{\ell,\mu\rho}t^\rho - (\mathbf{t}^TD_\ell^{-1}\mathbf{u})D^{-1}_{\ell,\mu\rho}u^\rho }{ (\mathbf{t}^TD_\ell^{-1}\mathbf{t})(\mathbf{u}^TD_\ell^{-1}\mathbf{u}) - (\mathbf{t}^TD_\ell^{-1}\mathbf{u})^2 },$$

where $\mathbf{t}$ encodes $a^ia^j$ and $\mathbf{u}$ encodes $b^ib^j$ (Kwok et al., 12 Sep 2025).

The variance is minimized using

$$\mathrm{Cov}(\hat{C}_\ell^{jk},\hat{C}_\ell^{mn}) = \frac{2}{N_p}C_\ell^{jm}C_\ell^{kn}, \qquad \mathrm{Var}(\hat{K}_\ell) = W_\ell^{jk}W_\ell^{mn}\frac{2}{N_p}C_\ell^{jm}C_\ell^{kn}.$$

Because $W_\ell^{ij}$ is not required to factorize as $w_iw_j$, spectral-level optimization has extra freedom: auto-spectrum weights can even be negative, which cannot occur at map level. In Simons Observatory-like simulations with both tSZ and CIB fully deprojected, SpILC produces CMB power spectrum estimators with more than 7 times smaller errorbars than constrained ILC at $\ell\gtrsim 4000$; at those scales, cILC reaches $\sigma_{\rm cILC}/\sigma_{\rm ILC}\gtrsim 40$, whereas $\sigma_{\rm SpILC}/\sigma_{\rm ILC}\approx 5$ (Kwok et al., 12 Sep 2025).

6. Polarization, spectral distortions, and methodological limits

In polarization, the natural object is the spin-2 field

$P=Q+iU.$

A two-spin complex-coefficient ILC estimates

$$\hat{P}_{\rm CMB}(p)=\sum_{j=1}^{N_\nu}\omega_j P_j(p), \qquad \omega_j=\omega_j^{(R)}+i\omega_j^{(I)},$$

with constraints

$$\sum_j \omega_j^{(R)}=1,\qquad \sum_j \omega_j^{(I)}=0.$$

The minimized quantity is the covariant polarization variance

$$\langle |\hat{P}_{\rm CMB}|^2\rangle = \langle \hat{Q}_{\rm CMB}^2+\hat{U}_{\rm CMB}^2\rangle.$$

In harmonic $E/B$ language, the complex weight acts as a rotation-plus-scaling matrix,

$$\begin{pmatrix} \omega_j^{(R)} & -\omega_j^{(I)}\\ \omega_j^{(I)} & \omega_j^{(R)} \end{pmatrix} = |\omega_j| \begin{pmatrix} \cos(2\phi_j) & -\sin(2\phi_j)\\ \sin(2\phi_j) & \cos(2\phi_j) \end{pmatrix},$$

so the residual remains a proper spin-2 field. This contrasts with independent $Q/U$ weighting, which the paper shows can introduce unphysical spin-4 residual structure, and with independent $E/B$ weighting, which can induce parity-mixed residuals (Fernández-Cobos et al., 2016).

Spin-SILC further combines spectral ILC with spin, directional, scale-discretized wavelets. The spin-2 wavelet coefficients of $P$ are linearly combined at each scale, position, and orientation,

$$W_{jnk}^{\,{}_{2}P,\mathrm{ILC}} = \sum_{c=1}^{N_c}\omega_{jnk}^c\,W_{jnk}^{\,{}_{2}P,c},$$

subject to

$\sum_{c=1}^{N_c} a^c\omega_{jnk}^c=1+0i.$

The resulting coefficients satisfy

$$W_{\tilde E}^{\,{}_{0}\Psi^j}(\rho)=-\mathrm{Re}\!\left[W_P^{\,{}_{2}\Psi^j}(\rho)\right],\qquad W_{\tilde B}^{\,{}_{0}\Psi^j}(\rho)=-\mathrm{Im}\!\left[W_P^{\,{}_{2}\Psi^j}(\rho)\right],$$

so component separation and $E/B$ decomposition are performed concurrently (Rogers et al., 2016).

A related polarization development combines moment-expanded foreground SEDs with polarization ILC in pixel space. The constraint matrix includes the zeroth-order synchrotron and dust SEDs and their first- and second-order spectral derivatives; the optimal real weights are

$$\mathbf{w}^T = \mathbf{e}^T(\mathbf{F}^T\mathbf{C}^{-1}\mathbf{F})^{-1}\mathbf{F}^T\mathbf{C}^{-1}.$$

Applied to simulated WMAP and Planck frequencies, this heavily constrained moment-PILC reconstructs thermal-dust and synchrotron $Q,U$ templates with correlations around $r\sim0.99$ in the quoted optimal configuration and recovers dust and synchrotron $EE/BB$ spectra consistent with the inputs within the stated uncertainties (Adak, 2021).

For spectral distortions, the SpILC viewpoint becomes purely frequency-space. Classical spectral ILC preserves the distortion spectrum $\mathbf{I}_d$ and minimizes $\boldsymbol{\omega}\mathbf{D}\boldsymbol{\omega}^T$. MILC instead nulls Taylor-expanded foreground moments, while LRM minimizes $$\boldsymbol{\omega}(\boldsymbol{\Phi}+\mathbf{C})\boldsymbol{\omega}^T$$ under unit response to the target distortion. In the $\mu$-distortion instrument study, the optics spectrum $I_{\rm opt}=\tau_{\rm opt}B(\nu,T_{\rm opt})$ becomes another spectral contaminant, and the quoted optimum for $\mu$ measurements occurs near $T_{\rm opt}\sim 9$ K, where the trade-off between spectral degeneracy and photon noise is minimized (Maillard et al., 2024).

Across these formulations, the principal limitations are consistent. They include sensitivity to SED mis-specification and decorrelation, empirical covariance errors, restricted channel count, and, in weak-cSpILC, reliance on assumptions such as $C_\ell^{sy}=0$. For CIB-like contaminants, significant residual bias can remain even when nominal deprojection is enforced if the true SED is not rigid or if frequency decorrelation is large (Abylkairov et al., 2020). For spectral-level SpILC, if components assumed uncorrelated in the weak-constraint construction are in fact correlated, dropping the corresponding spectral constraints can bias the estimator (Kwok et al., 12 Sep 2025). This suggests that SpILC is best understood not as a single algorithm but as a spectral optimization principle whose performance depends on how accurately the preserved and nulled spectral subspaces represent the actual sky.