S.V.D.

Singular Value Decomposition. The ultimate theorem of linear algebra that dismantles any matrix into pure rotations and scalings.

01 // The Dissection

Not every matrix is a clean, square, invertible box. But the SVD theorem guarantees that absolutely any matrix (even rectangular ones) can be factored into three fundamental actions. It is the geometric equivalent of separating an object into its raw DNA.

A = U \Sigma V^T

02 // The Pipeline

Let's watch exactly how a matrix $A$ transforms a perfect circle of vectors into a slanted ellipse. We apply the SVD right-to-left: first we rotate the grid ( $V^T$ ), then we stretch it strictly along the X and Y axes ( $\Sigma$ ), and finally, we rotate it into its new position ( $U$ ).

SVD Pipeline Visualizer

A = U \Sigma V^T

Full Transformation Matrix

The starting state: A perfect unit circle.

Step 1

V^T

Input Rotation

An Orthogonal Matrix. It does not stretch or squash space; it only spins it. It aligns the raw input vectors to prepare them for scaling.

Step 2

\Sigma

The Singular Values

A Diagonal Matrix. This contains the absolute "strength" or scaling factor of the transformation. It is the heart of the matrix where all physical stretching occurs.

Step 3

U

Output Rotation

Another Orthogonal Matrix. It takes the newly stretched shape and spins it into its final physical orientation in output space.

Real-World Application

Low-Rank Compression

Because the singular values in $\Sigma$ are sorted from largest to smallest, the SVD tells us exactly which pieces of data are important and which are just "noise".

If we delete the smallest singular values (Truncated SVD), we can compress megabytes of image data or machine learning matrices into kilobytes, while preserving almost all of the visual or mathematical meaning.

Rank-k Approximation

Linear Algebra Conquered

You have mastered the grid. You are ready for Calculus.

Next: Calculus