FAAN: Factor Analysis for Anisotropic Noise

• FAAN is a factor analysis method that models covariance as a low-rank signal term plus anisotropic noise with nonuniform variances.
• It employs iterative coordinate descent and eigen-decomposition updates to jointly estimate latent factors and structured noise while ensuring diagonal matching.
• Applications span array processing, portfolio selection, and functional data analysis, providing enhanced subspace recovery and risk assessment.

Factor Analysis for Anisotropic Noise (FAAN) denotes factor-analytic estimation of a covariance matrix when the noise or idiosyncratic component is not isotropic. In its basic form, the observed covariance is written as a low-rank plus structured-noise decomposition, such as $R = SS' + \Sigma$ in the real-valued setting or $R = AA^H + \Psi$ in the complex setting, with the low-rank term representing latent factors or signal subspace and the noise covariance estimated jointly rather than fixed to $\sigma^2 I$. In the literature, the term is used both for the general problem of factor analysis under anisotropic, nonuniform, heteroscedastic, or colored noise, and for a specific maximum-likelihood coordinate descent algorithm introduced for diagonal but unequal noise variances (Stoica et al., 2023).

1. Covariance models and the meaning of anisotropy

The canonical FAAN model in the real-valued literature is

$R = SS' + \Sigma,$

where $S \in \mathbb{R}^{p \times r}$ is a rank-$r$ factor-loading matrix and $$\Sigma = \operatorname{diag}(\sigma_1^2,\dots,\sigma_p^2)$$ is diagonal. Anisotropic noise means that the diagonal entries $\sigma_i^2$ are not necessarily equal. In the signal-processing literature this is called nonuniform noise, and in the statistical literature it is called heteroscedastic noise (Stoica et al., 2023).

A broader complex-valued formulation appears in extended complex factor analysis, where

$R = A A^H + \Psi,$

with $A \in \mathbb{C}^{P \times Q}$ tall and $R = AA^H + \Psi$0, while $R = AA^H + \Psi$1 is a structured Hermitian noise covariance. The structure of $R = AA^H + \Psi$2 is represented by a mask $R = AA^H + \Psi$3 of zeros and ones through

$R = AA^H + \Psi$4

with a selection matrix $R = AA^H + \Psi$5 extracting the free entries. This formulation includes classical FA when $R = AA^H + \Psi$6, but also colored noise patterns such as Toeplitz, banded, and known support patterns (Sardarabadi et al., 2018).

A further extension treats anisotropy through a sparse positive semidefinite idiosyncratic covariance. In that notation, observations satisfy $R = AA^H + \Psi$7, or equivalently a low-rank component $R = AA^H + \Psi$8 and a sparse noise covariance $R = AA^H + \Psi$9. Off-diagonal entries in $\sigma^2 I$0 are allowed, so anisotropy is not restricted to unequal marginal variances but can include sparse correlated noise. This differs from diagonal-noise FAAN, and the notation also differs: in the 2023 FAAN paper $\sigma^2 I$1 denotes the loading matrix, whereas in the $\sigma^2 I$2 paper $\sigma^2 I$3 denotes the idiosyncratic covariance (Wang et al., 1 Apr 2025).

These formulations share the same structural principle: the covariance is decomposed into a low-rank signal term and a noise term whose anisotropy must be estimated rather than ignored. A common misconception is that anisotropic noise in FAAN always means diagonal noise. The literature does not support that restriction: some FAAN methods assume diagonal but unequal variances, while others explicitly admit colored, masked, banded, Toeplitz, or sparse correlated noise (Sardarabadi et al., 2018).

2. Objective functions, reparameterizations, and identifiability

For the Gaussian FA/PCA covariance model $\sigma^2 I$4, the negative normal log-likelihood, up to additive constants, is

$\sigma^2 I$5

Equivalently, one minimizes

$\sigma^2 I$6

which is also the quasi-likelihood beyond Gaussian settings. Using the determinant lemma and Woodbury identity, the model admits a whitened reparameterization

$\sigma^2 I$7

with $\sigma^2 I$8 orthonormal and $\sigma^2 I$9 diagonal. If $R = SS' + \Sigma,$0 has eigenvalues $R = SS' + \Sigma,$1, then the factor update reduces to $R = SS' + \Sigma,$2 (Stoica et al., 2023).

In the constrained least-squares formulation for extended complex FA, the criterion is instead

$R = SS' + \Sigma,$3

with $R = SS' + \Sigma,$4. The parameter vector is

$R = SS' + \Sigma,$5

and a Gauss–Newton step solves

$R = SS' + \Sigma,$6

The paper’s main structural result is that the $R = SS' + \Sigma,$7-block of the Gauss–Newton system can be diagonalized in closed form once $R = SS' + \Sigma,$8 is enforced, so the rank deficiency induced by the unitary ambiguity $R = SS' + \Sigma,$9 is isolated explicitly (Sardarabadi et al., 2018).

Identifiability is central throughout FAAN. In the real-valued model, only the subspace $S \in \mathbb{R}^{p \times r}$0 is identifiable; $S \in \mathbb{R}^{p \times r}$1 itself is identifiable only up to right-multiplication by an orthogonal matrix. The 2023 FAAN paper states Ledermann’s bound

$S \in \mathbb{R}^{p \times r}$2

with generic global identifiability for $S \in \mathbb{R}^{p \times r}$3, generic local identifiability for $S \in \mathbb{R}^{p \times r}$4, and generic local unidentifiability for $S \in \mathbb{R}^{p \times r}$5. It also gives the data-dependent Guttman lower bound

$S \in \mathbb{R}^{p \times r}$6

which provides a lower bound on the rank required to represent $S \in \mathbb{R}^{p \times r}$7 exactly as $S \in \mathbb{R}^{p \times r}$8 (Stoica et al., 2023).

In the complex EFA setting, the nonuniqueness is $S \in \mathbb{R}^{p \times r}$9-dimensional in the complex case, so $r$0 constraints are required. The paper proposes either $r$1 with $r$2 diagonal and positive, or a lower-triangular $r$3 with real positive diagonal. This is not only an identifiability device but also the mechanism that makes the reduced Gauss–Newton system nonsingular (Sardarabadi et al., 2018).

3. The FAAN coordinate-descent algorithm

The specific method named FAAN by Stoica and Babu is a coordinate descent algorithm that iteratively maximizes the normal likelihood function for

$r$4

Its first block update fixes $r$5, forms the whitened sample covariance

$r$6

computes its top $r$7 eigenpairs $r$8, sets $r$9, and updates the factor strengths through

$$\Sigma = \operatorname{diag}(\sigma_1^2,\dots,\sigma_p^2)$$0

This yields the closed-form factor step (Stoica et al., 2023).

The second block update fixes $$\Sigma = \operatorname{diag}(\sigma_1^2,\dots,\sigma_p^2)$$1 and $$\Sigma = \operatorname{diag}(\sigma_1^2,\dots,\sigma_p^2)$$2, forms

$$\Sigma = \operatorname{diag}(\sigma_1^2,\dots,\sigma_p^2)$$3

and updates each $$\Sigma = \operatorname{diag}(\sigma_1^2,\dots,\sigma_p^2)$$4 by solving a scalar quadratic

$$\Sigma = \operatorname{diag}(\sigma_1^2,\dots,\sigma_p^2)$$5

where

$$\Sigma = \operatorname{diag}(\sigma_1^2,\dots,\sigma_p^2)$$6

The feasible minimizer is

$$\Sigma = \operatorname{diag}(\sigma_1^2,\dots,\sigma_p^2)$$7

This noise update is carried out as a Gauss–Seidel sweep and may be repeated a few times per outer iteration (Stoica et al., 2023).

The reconstructed low-rank covariance is

$$\Sigma = \operatorname{diag}(\sigma_1^2,\dots,\sigma_p^2)$$8

and one may equivalently take

$$\Sigma = \operatorname{diag}(\sigma_1^2,\dots,\sigma_p^2)$$9

A notable property is diagonal matching: at a stationary point,

$\sigma_i^2$0

so FAAN preserves the sample variances on the diagonal. The same paper emphasizes that the iterates remain feasible, no projection steps are needed, and the method exhibits monotonic decrease of $\sigma_i^2$1 together with excellent reliability in the reported experiments (Stoica et al., 2023).

The algorithm occupies an intermediate position between PCA and full covariance estimation. It is more structured than unconstrained covariance fitting, because the signal part is low rank, but less restrictive than white-noise PCA, because the diagonal noise variances are estimated jointly. This is why FAAN is used when ordinary EVD/PCA is biased by unknown nonuniform noise (Stoica et al., 2023).

4. Structured-noise and second-order extensions

The constrained least-squares Gauss–Newton approach for extended complex factor analysis generalizes FAAN beyond diagonal noise. With the Jacobian block

$\sigma_i^2$2

the paper proves a closed-form eigendecomposition

$\sigma_i^2$3

after enforcing $\sigma_i^2$4. The null space corresponding to the unitary ambiguity is spanned by $\sigma_i^2$5, while the nonzero spectrum is diagonalized in terms of $\sigma_i^2$6, $\sigma_i^2$7, and $\sigma_i^2$8. The resulting solver first computes $\sigma_i^2$9 from

$R = A A^H + \Psi,$0

and then computes $R = A A^H + \Psi,$1 in reduced coordinates without explicitly constructing $R = A A^H + \Psi,$2 or $R = A A^H + \Psi,$3. The paper states that $R = A A^H + \Psi,$4 is positive definite, $R = A A^H + \Psi,$5 is positive semidefinite, and $R = A A^H + \Psi,$6, so iterative solvers converge efficiently (Sardarabadi et al., 2018).

The same paper is explicit about a limitation: the proposed algorithm does not guarantee the positive (semi-)definiteness of $R = A A^H + \Psi,$7. The authors state that this will be addressed in future works. This limitation distinguishes the method from formulations that impose positive semidefiniteness directly on the noise covariance (Sardarabadi et al., 2018).

For maximum-likelihood factor analysis in array processing, the 2025 ECME paper formulates the covariance as

$R = A A^H + \Psi,$8

and describes FAAN through the reparameterization

$R = A A^H + \Psi,$9

The FAAN iteration alternates an eigendecomposition-based update of $A \in \mathbb{C}^{P \times Q}$0 from $A \in \mathbb{C}^{P \times Q}$1 and a sequential sensor-wise update of $A \in \mathbb{C}^{P \times Q}$2. The ECME variant keeps the same signal-subspace step in spirit but replaces the sequential noise update by the closed-form conditional maximization

$A \in \mathbb{C}^{P \times Q}$3

The paper states that ECME has almost the same per-iteration complexity as FAAN, while numerical results show faster stable convergence and higher computational efficiency (Gong, 4 Aug 2025).

A different extension is the $A \in \mathbb{C}^{P \times Q}$4-regularized low-rank-plus-sparse covariance model

$A \in \mathbb{C}^{P \times Q}$5

with $A \in \mathbb{C}^{P \times Q}$6, $A \in \mathbb{C}^{P \times Q}$7, and $A \in \mathbb{C}^{P \times Q}$8. Here anisotropic noise is represented by a sparse PSD $A \in \mathbb{C}^{P \times Q}$9 that may have off-diagonal entries. The method uses a barrier-augmented objective, a hard-thresholding $R = AA^H + \Psi$00 proximal operator, the stationary-point equation $R = AA^H + \Psi$01, and Newton iterations inside an interior-point scheme. This replaces diagonal-noise FAAN by a sparse-correlated-noise variant with explicit positive semidefiniteness handling (Wang et al., 1 Apr 2025).

5. Application domains

In array processing and direction-of-arrival estimation, FAAN is used when the sensor noise is nonuniform and therefore ordinary subspace estimation from the sample covariance is biased. The 2023 FAAN paper studies narrowband-source covariance estimation for DOA estimation under diagonal anisotropic sensor noise, and the 2025 ECME paper considers the stochastic model

$R = AA^H + \Psi$02

After estimating the factor loading $R = AA^H + \Psi$03, the signal subspace is passed to MUSIC or Root-MUSIC. The ECME paper reports that RMSEs of DOA estimates from both FAAN and ECME approach the stochastic CRLBs as the number of snapshots increases, with ECME showing better RMSE at fixed iteration budgets because of its faster convergence (Gong, 4 Aug 2025).

The 2023 FAAN paper also treats portfolio selection. Asset returns are modeled by

$R = AA^H + \Psi$04

with $R = AA^H + \Psi$05 representing common risk and $R = AA^H + \Psi$06 representing idiosyncratic heteroscedastic risk. For mean–variance portfolios one uses $R = AA^H + \Psi$07, while for minimum-variance portfolios with unit-sum constraint one uses

$R = AA^H + \Psi$08

The paper states that, on CRSP returns with $R = AA^H + \Psi$09, FAAN combined with BIC to choose $R = AA^H + \Psi$10 yields lower out-of-sample portfolio standard deviation than SCM and MTP2, especially for short lookback windows (Stoica et al., 2023).

In discretely observed functional data, the multivariate reformulation

$R = AA^H + \Psi$11

casts each sampled curve as a high-dimensional observation whose covariance decomposes into common components and idiosyncratic noise. The paper allows $R = AA^H + \Psi$12 to be non-spherical and to exhibit short-range dependence across grid locations. PCA-based factor extraction estimates the common components through

$R = AA^H + \Psi$13

while ML/EM approaches are cited as the route when $R = AA^H + \Psi$14 is modeled explicitly as non-diagonal. The paper further states that eigenfunctions can be estimated from raw discretizations at rate $R = AA^H + \Psi$15 without smoothing, and develops tests of the iid-noise null using the empirical variance of averaged periodogram ordinates (Hörmann et al., 2020).

These applications show that FAAN is not tied to a single inferential task. In array processing it is a subspace-estimation device under unknown nonuniform noise; in finance it is a covariance estimator under heteroscedastic idiosyncratic risk; in functional data it is a preprocessing and diagnostic framework that distinguishes common signal variation from anisotropic measurement noise.

6. Limitations, misconceptions, and comparative position

A recurrent misconception is that FAAN is merely PCA with a diagonal rescaling. The literature does not support that view. EVD/PCA assumes white noise, or at least known noise that can be prewhitened, whereas FAAN estimates the low-rank signal covariance and the anisotropic noise jointly. Under unknown colored noise, EVD is biased; this is one of the explicit motivations for replacing EVD by FA or extended FA (Sardarabadi et al., 2018).

Another misconception is that FAAN denotes a single algorithm. The term is used more broadly. In the 2023 paper it names a coordinate descent ML method for diagonal anisotropic noise; in the 2018 complex EFA paper it is a natural description of FA under structured colored noise; in the 2025 ECME paper it denotes the baseline MLFA iteration against which ECME is compared; and in the $R = AA^H + \Psi$16 paper it is realized through a sparse PSD noise covariance. This suggests that FAAN is best understood as a family of factor-analysis methods adapted to anisotropic noise models rather than as one invariant procedure.

The limitations are method-specific. The 2023 FAAN algorithm assumes diagonal noise and therefore does not model cross-sensor or cross-coordinate noise correlations. The 2025 ECME paper states explicitly that both FAAN and ECME assume diagonal, sensor-uncorrelated noise and that model mismatch from colored or coherent noise may degrade performance (Gong, 4 Aug 2025). The 2018 Gauss–Newton EFA method handles structured colored noise, but the algorithm does not guarantee positive semidefiniteness of $R = AA^H + \Psi$17 (Sardarabadi et al., 2018). The $R = AA^H + \Psi$18 interior-point formulation enforces PSD constraints, but it is nonconvex, parameter-sensitive through $R = AA^H + \Psi$19 and $R = AA^H + \Psi$20, and computationally expensive because second-order steps require matrix inverses and reduced linear solves (Wang et al., 1 Apr 2025).

Convergence behavior also varies. The 2023 FAAN paper emphasizes monotonic decrease and excellent reliability. The 2018 constrained least-squares method reports competitive convergence and relative robustness across noise realizations, but also notes that the problem is non-convex and some failures can occur. The 2025 ECME paper reports faster stable convergence than FAAN at almost the same per-iteration complexity. These are not contradictory claims: they refer to different formulations, objectives, and update rules, even when the underlying covariance model is similar (Stoica et al., 2023).

The comparative position of FAAN within covariance estimation is therefore precise. It occupies the regime where a low-rank latent structure is retained, isotropic-noise assumptions are relaxed, and the noise model is estimated jointly with the factor structure. What changes across the literature is the admissible noise class—diagonal unequal variances, masked Hermitian structure, or sparse PSD covariance—and the optimization machinery used to fit it.