Condition numbers of indefinite rank 2 ghost Wishart matrices

Ramis Movassagh, Alan EdelmanLinear 2015 · DOI · arXiv

In the realm of multivariate statistics, Wishart matrices are fundamental tools for covariance estimation. While standard Wishart matrices are typically positive definite, certain high-sensitivity engineering applications—most notably hypersensitive ground-based radar systems—require the analysis of indefinite matrices, where eigenvalues can be both positive and negative.

This paper, authored by Ramis Movassagh and Alan Edelman from MIT, provides a rigorous mathematical treatment of this specific class of random matrices. The authors derive the exact probability distribution for the condition numbers of indefinite, rank-2 "ghost" Wishart matrices, filling a significant gap in the statistical literature.

The Problem: Indefinite Matrices and Numerical Stability

The "condition number" (the ratio of the largest to smallest eigenvalue) is a critical metric for numerical stability and signal resolution. While the distribution of condition numbers for standard Wishart matrices is well-documented, the indefinite case remained largely unexplored. The authors sought to find the exact probability density function (PDF) for the condition number of these indefinite matrices when they have a rank of 2.

Methodology: The "Ghost" Framework

The authors analyzed matrices of the form $A = W^T W \Sigma$, where $W$ is an $L \times 2$ matrix of standard normals and $\Sigma$ is an indefinite diagonal matrix. To achieve a unified result, they employed several key technical strategies:

  • Ghost $\beta$-Normals: Instead of treating real ($\beta=1$), complex ($\beta=2$), and quaternionic ($\beta=4$) cases as separate problems, the authors used the "ghost" matrix framework. This allows $\beta$ to be treated as a continuous parameter, covering all standard types and their interpolations within a single formula.
  • Reduced QR Factorization: By decomposing $W = QR$, they simplified the problem to the condition number of a smaller $2 \times 2$ matrix, $R \Sigma R^T$.
  • Special Functions: The derivation involved transforming joint distributions and integrating out angular components, resulting in closed-form solutions expressed through the Gauss hypergeometric function ($_{2}F_{1}$).
Comparison of theoretical PDF (red) and Monte Carlo simulation (black dots) for real matrices with $L=4$.

Key Results and Validation

The primary achievement of the paper is the explicit PDF for the condition number for any $\beta$ and any dimension $L$. For the specific case of real matrices ($\beta=1$) with $L=2$ and symmetric $\Sigma$ ($x_1 = -x_2$), the density simplifies to an elegant form: $$f(\sigma) = \frac{2}{\pi \sqrt{\sigma}(\sigma + 1)}$$

To validate these theoretical derivations, the authors performed extensive Monte Carlo simulations, generating $10^5$ random trials for various parameters. As shown in the figures, the theoretical curves (red) perfectly track the empirical data (black dots) across different regimes.

Theoretical vs. empirical density for the $L=2$ case, showing a monotonically decreasing distribution.

Insights and Future Directions

The paper demonstrates that the "ghost" $\beta$ approach is a powerful tool for deriving unified formulas across physics and algebra. One notable observation is the deterministic limit: as $\beta \to \infty$, the randomness "freezes out," and the distribution collapses into a delta function, making the condition number predictable based solely on matrix dimensions.

While this work is strictly limited to rank-2 matrices ($L \times 2$), it provides the necessary foundation for future generalizations to $L \times N$ indefinite matrices where $N > 2$. These results are immediately applicable to improving the precision and reliability of hypersensitive signal processing systems.

Summary updated

BibTeX
@inproceedings{movassagh2015condition,
  author = {Ramis Movassagh and Alan Edelman},
  title = {{Condition numbers of indefinite rank 2 ghost Wishart matrices}},
  booktitle = {Linear},
  year = {2015},
  doi = {10.1016/j.laa.2015.05.027},
}
Figure