Skip to main content
TheoremPath
More Demos

Spectral Geometry

Random Matrix / Spectral Geometry Lab

Learn how pure noise creates a spectral bulk, why nearly square data matrices become ill-conditioned, and when a PCA spike separates from the noise.

Marchenko-Pastur

Noise Has A Shape

See what pure random covariance produces before adding any signal.

A random data matrix does not give a tight pile of eigenvalues. The spectrum spreads into a predictable bulk whose edges depend on the aspect ratio.

aspect ratio gamma
0.65
noise lower edge
0.038
noise upper edge
3.262
ridge condition
28.4
Random matrix spectrumPure noise still has structure: a bulk, edges, and a conditioning penalty.0123456lower edgeupper edgeReading ruleBefore trusting a component, ask whether noise alone would make an eigenvalue that large.
Chapter
Diagnosis

Noise already creates large eigenvalues

The top sample eigenvalue is not evidence by itself. Compare it against the upper edge of the noise bulk before calling it signal.

gives the noise-only spectral edges.

Research translation

Spectrum is a first-pass sanity check. If your claimed factor, feature, or principal component is not beyond the noise bulk, the model may be fitting the geometry of finite samples rather than a real direction.

Before: Matrix Mechanics LabNext: High-Dimensional Probability LabTheory: Random Matrix Overview
PrerequisiteMatrix MechanicsEigenvalues, SVD, rank, and row reduction.TheoryRandom Matrix TheoryMarchenko-Pastur, covariance, and PCA.ApplicationCovariance EstimationWhy sample covariance breaks in high dimension.Next LabHigh-Dimensional ProbabilityThin shells, concentration, and spike thresholds.

TheoremPath · 584 topics · High-dimensional probability lab