Foundations

Singular Value Decomposition, visualized

The Singular Value Decomposition writes any matrix as a rotation, a set of stretches, and another rotation: A = UΣVᵀ. The cleanest way to see it is what a matrix does to the unit circle — it always turns it into an ellipse. The lengths of that ellipse's two semi-axes are exactly the singular values.

SVD is the most useful decomposition in data science: it powers PCA, image and data compression, recommender systems, and the pseudoinverse that solves least squares. Drag the matrix and watch the singular values grow, shrink, and — when one hits zero — collapse the ellipse to a line.

Drag the matrix — the circle becomes an ellipse

Singular values σ₁, σ₂

2.29, 0.87

the longest and shortest stretch — the ellipse's semi-axes

Area scale σ₁ · σ₂

2

equals |determinant| — how much area grows

Free · runs entirely in your browser · nothing to install

How to use it

  1. Drag the four matrix entries and watch the faint unit circle map to a solid ellipse.
  2. Read the two singular values — they are the lengths of the ellipse's longest and shortest semi-axes.
  3. Shrink the matrix until the smaller singular value hits zero and the ellipse flattens to a line (a rank-1 matrix).

What you'll take away

  • Why every matrix maps the unit circle to an ellipse.
  • What the singular values mean geometrically — the stretch factors.
  • How a zero singular value signals a rank-deficient (singular) matrix.

Want to actually build this?

This demo is one moment inside a full Math to Machine lesson — predict, build, and explain the concept, with an AI tutor that gives hints, not answers. The first five lessons are free.

FAQ

What is singular value decomposition?
SVD factors any matrix into three simple steps — a rotation, a scaling along perpendicular axes, and another rotation. The scaling amounts are the singular values, and they reveal the matrix's rank and most important directions.
Why is SVD important in machine learning?
Keeping only the largest singular values gives the best low-rank approximation of a matrix, which underlies PCA, data and image compression, recommender systems, and numerically stable least-squares solutions.

More Foundations tools