Sunday, October 31, 2021

The Singular Value Decomposition without Algebra

 It is an unfortunate reality that many people try to learn linear algebra with algebra when what’s going on is essentially a geometric matter (that means you can understand it with pictures). I previously described this point of view in Understanding Regression with Geometry and here I’m going to attempt something similar with the Singular Value Decomposition.

A Vector is not a List of Numbers

A Vector is a Vector is a Vector

A Vector Lives in a Vector Space

A Matrix is not an Array of Numbers

A Matrix is a (Linear) Map between Vector Spaces

A Linear Map Satisfies Three Properties

What Linear Maps Do

Things a Linear Map Can Do.

The Singular Value Decomposition

Viewed Correctly, “Smushing” is a scaling of a single vector

Official Statement

A non-square diagonal matrix with 1, 2.3, -7, and 0 on the diagonal

Partial Proof Sketch

The Spectral Theorem

The Matrix Transpose

The Transpose Flips a Matrix and Puts it on its side

Getting the Singular Values

Following a map by its transpose results in a map from the source back to itself

Conclusion

References / Further Reading

No comments: