Matrix Theory and Applications I

The matrix is the data structure of modern AI. This course is about understanding it, computing with it, and — the part that is usually skipped — computing with it reliably.

Sixteen weeks built on Golub & Van Loan, Matrix Computations, 4th edition, with every week tied to specific sections of the book and to a problem you will meet in machine learning, vision, or scientific computing.

What you will be able to do

  1. Analyse — state and use the core theorems of matrix analysis, and judge when a matrix problem is well- or ill-conditioned.
  2. Compute — select and derive an appropriate decomposition or iterative method, and reason about its cost and stability.
  3. Implement — turn an algorithm from the board into working code, and verify it against a reference implementation.
  4. Apply — recognise the matrix problem hiding inside a machine learning, vision, or scientific computing task, and solve it.

How the course is built

The five parts run in the book’s own order, which is not the order most syllabi use:

Why the SVD appears twice. Golub and Van Loan introduce the SVD on page 76, in the analysis chapter, and defer its computation 410 pages to §8.6. We follow them. What the SVD is needs only norms, so it lands in Week 3 and is then available as an instrument for the rest of the course; how to compute it stably is an algorithm that cannot be motivated before QR iteration, so it lands in Week 14.

Prerequisites

Linear algebra, calculus, and a first course in numerical analysis. You should be comfortable writing and debugging code in MATLAB, Python, or Julia — the assignments are language-agnostic, but every one of them requires you to implement something and check it.