core Estimated learning time: 6 h

1.5 SVD and low-rank approximation

You can compress a matrix and know what you threw away.

Before:00. Orientation & SetupUnlocks:03. Data Handling & Analysis04. Classical AI — Agents, Search & Knowledge Representation

The singular value decomposition factorises any matrix — square or not — into rotate, stretch, rotate, and truncating it gives the best possible low-rank approximation, with a guarantee about what was lost. It is the workhorse behind PCA, recommender factorisation and LoRA fine-tuning. It sits at the summit of the linear algebra sequence because it composes everything before it. The common confusion is thinking SVD needs a square matrix; its whole value is that it does not.

Work through these

  • The SVD as rotate–stretch–rotate

    Every matrix, of any shape, can be read as a rotation, then a stretch along axes, then another rotation. This is the most generally useful decomposition in applied linear algebra.

  • Truncated SVD and the Eckart–Young theorem

    Keeping only the largest stretch factors gives the best possible approximation of a matrix at that size, and there is a theorem saying so. Compression, denoising and recommendation all rest on this result.

  • Applications: PCA, latent semantic analysis, LoRA

    The same decomposition underlies dimensionality reduction, one classical text-analysis method, and the low-rank adapters used to fine-tune large models. Seeing one technique behind all three is the point of this item.

  • Condition number and numerical stability

    The ratio between the largest and smallest stretch factors predicts how much a small change in input can distort the output. It is the number that tells you whether to trust a numerical result.

Sign in to keep your progress.

Free resources

Links last checked 29 Aug 2026.

Stuck here?

Ask a mentor. A real person answers, and they can see exactly which topic you're on. Usually within a couple of working days.

Checking your session…

Topics shown in module order.