Sandhya Indurkar

Math foundations

Eigenvalues: Stretch Factors Along Special Directions

Eigenvalues as stretch factors along principal directions

The idea

Most linear transforms rotate and stretch at once. Eigenvalues are the stretch amounts along directions that do not rotate. For a covariance or correlation matrix, the largest eigenvalue tells you how much variance lives on the dominant axis. That is the same intuition behind PC1 variance explained.

Eigenvalues answer: How strongly does this transform amplify motion along each special direction?

Example: stretch factors along eigen-directions

Eigenvalues tell you how much a symmetric transform stretches along each special direction. Arrow length tracks |λ|.

Revenue and orders co-move. Eigenvalues rank how much each axis stretches variance.

Matrix Acol 1col 2
row 12.401.10
row 21.100.90

λ₁ (largest |λ|)

2.86

λ₂

0.64

PC1-style share

82%

PC2-style share

18%

Eigenvalue magnitudes |λ|

Revenue and orders co-move. Eigenvalues rank how much each axis stretches variance. λ₁ = 2.86 carries 82% of stretch. One composite axis (like PC1) explains most variance.

The math

Eigenpair

Av = λv

Vector v is an eigenvector. Scalar λ is its eigenvalue. Applying A to v only scales v; it does not turn v onto a new direction.

Characteristic equation

det(A − λI) = 0

For a 2×2 matrix, this quadratic gives two eigenvalues (possibly equal). Solve for λ, then find v from (A − λI)v = 0.

Stretch and PCA

|λ₁| ≥ |λ₂| ⇒ PC1-style dominance

On symmetric covariance matrices, eigenvalues equal variance along each principal component. The ratio λ₁ / (λ₁ + λ₂) matches variance explained by the first axis.

A simple application

In PCA readouts, rank eigenvalues before you trust a one-chart summary. If λ₁ explains most variance, a composite KPI score is reasonable. In dynamical systems, eigenvalues near 1 mean slow decay; values above 1 mean instability. Check both the stretch and the direction (eigenvectors post).