Conclusion
This chapter turned matrices into objects we can inspect.
At the start, Ax = b was a question about reachability. Can the columns of
A, with some weights from x, make the target b?
From there, the chapter built a vocabulary for structure:
- span describes everything reachable by combining available directions
- independence asks whether a direction adds something new
- a basis gives a non-redundant coordinate frame
- rank counts independent output directions
- column space describes reachable outputs
- null space describes input directions that disappear
- eigenvectors show directions that a matrix does not turn
- diagonalization looks for coordinates where a matrix becomes simple scaling
- SVD rewrites a matrix as input directions, strengths, and output directions
- low-rank approximation keeps strong directions and discards weaker ones
- PCA applies this direction-and-strength thinking to centered data
The Main Skill
The main skill is not memorizing every formula. It is learning to ask better questions when a matrix appears.
What can this matrix reach? What does it collapse? Which directions are independent? Which directions are strong? Which details are being discarded?
Those questions make later ML ideas easier to read. Feature matrices, embeddings, projections, attention maps, covariance matrices, and layer weights all become less opaque when you can ask what structure they preserve or lose.
What Comes Next
Calculus comes next.
Linear algebra describes spaces and transformations. Calculus describes how quantities change. Together, they explain gradients, backpropagation, optimization, and training.
Keep This Question Nearby
When you meet a matrix in ML, ask:
What structure does this matrix reveal, preserve, compress, or discard?