Mahalanobis Data Whitening

• Mahalanobis data whitening is a canonical linear transformation that orthogonalizes multivariate data by removing correlations and standardizing variance based on empirical covariance.
• It employs spectral decomposition methods like eigendecomposition (or SVD) to construct the whitening matrix, ensuring data is aligned to the identity covariance for precise comparison.
• Practical implementations address numerical stability and computational efficiency, using regularization and FFT-based techniques in applications such as neuroimaging and signal processing.

Mahalanobis data whitening is a canonical linear transformation that removes correlations and standardizes variance among multivariate data dimensions by orthogonalizing with respect to the empirical covariance structure. This process yields whitened data suitable for analysis and dimensionality reduction, enabling rigorous comparison across samples, removal of individual-specific signatures, and alignment to chosen statistical templates. In applications such as neuroimaging, signal processing, and statistical inference, Mahalanobis whitening provides a mathematically optimal and interpretable preprocessing step, tightly connected to metrics on the manifold of covariance matrices such as the Bures distance.

1. Formal Definition and Mathematical Foundations

Let $X \in \mathbb{R}^{p \times n}$ denote data with $p$ variables and $n$ samples, assumed zero-mean. The empirical covariance is $$\Sigma = \frac{1}{n} X X^\top \in \mathbb{R}^{p \times p}$$. The Mahalanobis whitening transformation seeks a matrix $W^{-1/2}$ such that the transformed data $X_w = W^{-1/2} X$ have covariance $\mathrm{Cov}(X_w) = I_p$, the $p$-dimensional identity.

A canonical choice is $W = \Sigma$; the whitening matrix is constructed via spectral decomposition:

• $\Sigma = Q \Lambda Q^\top$, $Q$ orthogonal, $\Lambda = \mathrm{diag}(\lambda_1, ..., \lambda_p)$, $\lambda_i \geq 0$.
• $\Sigma^{-1/2} = Q \Lambda^{-1/2} Q^\top$, with $$\Lambda^{-1/2} = \mathrm{diag}(\lambda_1^{-1/2}, ..., \lambda_p^{-1/2})$$.

Hence,

$X_w = Q \Lambda^{-1/2} Q^\top X$

Alternative symmetric and ZCA transforms are also derived from this construction and maintain $\mathrm{Cov}(X_w) = I$ (Jacobson et al., 10 Nov 2025, Spurek et al., 2013).

2. Two-Stage De-individualization and Preprocessing

Mahalanobis whitening is often preceded by de-meaning and scaling to ensure zero mean and unit variance per variable. In neuroimaging (e.g., fMRI data), the workflow decomposes as follows (Jacobson et al., 10 Nov 2025):

1. De-meaning: For scan matrix $S \in \mathbb{R}^{p \times T}$ (regions $\times$ time),
   • Subtract per-row (region) mean:

   $$\bar{S}_{i, :} = S_{i, :} - \frac{1}{T} \sum_{t=1}^T S_{i, t}$$

• (Optional) Normalize by standard deviation per region.

1. Mahalanobis Whitening: Compute time-covariance,

$\Sigma_S = \frac{1}{T} \bar{S} \bar{S}^\top$

Form the whitening transform via eigendecomposition of $\Sigma_S$, apply $W^{-1/2}$ to yield $S_w$.

2. Segment Extraction and Comparison: Extract contiguous task segments $T_i$ from $S_w$ and measure separation via Frobenius norm,

$d_M(T_i, T_j) = \|T_i - T_j\|_F$

This "two-stage de-individualization" pipeline robustly removes both individual- and session-level covariance structure, resulting in data where only experimental variation is retained (Jacobson et al., 10 Nov 2025).

3. Consistency and Toeplitz Covariance Estimation in Stationary Processes

For data matrices from stationary processes with separable covariance structure $X = C_N^{1/2} Z R_M^{1/2}$, Mahalanobis whitening requires consistent estimation of the column covariance $R_M$ (Tian et al., 2020):

• Unbiased Toeplitz Estimator $\hat{R}_M$ achieves "ratio consistency"—for long-range dependent (LRD) processes, the spectral norm distance $$\|\hat R_M^{-1/2} R_M \hat R_M^{-1/2} - \xi_N I_M\| \to 0$$, where $\xi_N = \frac{1}{N}\operatorname{Tr} C_N$.
• The whitening map $Y_w = \hat R_M^{-1/2} X$ yields approximately white data in columns.
• Efficient construction leverages Toeplitz structure via FFT and matrix square root solvers.

For short-range dependent (SRD) processes, both unbiased and biased Toeplitz estimators are norm consistent, but only the unbiased estimator provides ratio consistency in the presence of LRD (Tian et al., 2020).

4. Connections to Bures Geometry and Quantum Metrics

On the manifold of $p \times p$ positive semidefinite matrices, geodesic distances are naturally measured by the Bures metric:

$$d_B(A, B) = \sqrt{\operatorname{Tr} A + \operatorname{Tr} B - 2\operatorname{Tr}\left( (A^{1/2} B A^{1/2})^{1/2} \right)}$$

Mahalanobis whitening aligns all sample covariances to the identity, which is the unique minimizer (up to congruence) of Bures distance to the standardized family. The choice $W = \Sigma$ represents the "Bures mean" of $\Sigma$ and $I$, and therefore Mahalanobis whitening is optimally aligned in the geometry of covariance matrices as justified by quantum fidelity and optimal transport perspectives (Jacobson et al., 10 Nov 2025).

5. Implementation Considerations and Regularization

Several practical issues arise in Mahalanobis whitening:

• Covariance Estimation: For moderate $T$ (samples), empirical $\Sigma$ may be ill-conditioned. Remedies include ridge regularization ($$\Sigma_{\mathrm{reg}} = (1-\alpha)\Sigma + \alpha I$$), Ledoit–Wolf shrinkage, and robust estimators (minimum covariance determinant, graphical lasso).
• Numerical Stability: Small eigenvalues cause large entries in $\Sigma^{-1/2}$; enforce $\lambda_i \geq \epsilon > 0$.
• Computational Complexity: Eigen-decomposition costs $O(p^3)$ for $p$ variables; for Toeplitz matrices, FFT and sine/cosine transforms reduce complexity to $O(p \log p)$.
• Alternative Construction: Singular value decomposition (SVD) can compute $\Sigma^{-1/2}$ efficiently: $$X = UDV^\top \implies \Sigma = UD^2U^\top \implies \Sigma^{-1/2} = UD^{-1}U^\top$$.

These techniques facilitate robust whitening even in high-dimensional, noisy, or temporally correlated scenarios (Jacobson et al., 10 Nov 2025, Tian et al., 2020).

6. Impact on Dimensionality Reduction and Statistical Inference

After Mahalanobis whitening, all data directions are standardized with unit variance:

• PCA: Applying standard PCA to whitened data ranks components by sampling noise rather than genuine signal; typically, PCA is performed prior to whitening.
• Manifold Learning (Isomap, UMAP): Whitening neutralizes subject-specific variance, ensuring that subsequent embeddings and clusterings reflect only stimulus or task-related structure.
• Signal Detection and Compression: Whitened data enables accurate signal detection (e.g., spike separation via Marčenko–Pastur law), estimation of component strengths, and nearly optimal principal component projection—even under long-range dependence (Tian et al., 2020).

7. Generalizations and Optimality Criteria

The classical Mahalanobis whitening can be extended and justified via cross-entropy minimization over the affine group. Setting $Y = \{y_1, ..., y_n\} \subset \mathbb{R}^N$, with mean $m_Y$ and covariance $\Sigma_Y$, the map $y \mapsto z = \Sigma_Y^{-1/2}(y - m_Y)$ produces data with zero mean and identity covariance (Spurek et al., 2013).

The cross-entropy between empirical and Gaussian distributions yields the optimal choice of affine parameters. For a fixed center $m$, the minimizer of the criterion is

$$\Sigma = \Sigma_Y \left(\Sigma_Y - \frac{(m - m_Y)(m - m_Y)^T}{1 + \|m - m_Y\|_{\Sigma_Y}^2}\right)^{-1} \Sigma_Y$$

and the corresponding whitening map is $y \mapsto \Sigma^{-1/2}(y - m)$. Classical whitening is recovered when $m = m_Y$.

Practical implementation involves computing $m_Y$, $\Sigma_Y$, choosing $m$, estimating or fixing $\Sigma$ accordingly, followed by eigen-decomposition and application of the whitening map (Spurek et al., 2013).

---

Mahalanobis data whitening constitutes a mathematically rigorous, computationally tractable, and robust approach to statistical preprocessing. Its role in aligning data geometrically via covariance structure, supporting reliable inference under complex dependencies, and integrating with optimal transport and quantum information metrics is well-established across statistical signal processing, neuroimaging, and machine learning.