Chapter 8 — Key Takeaways

The one idea

Matrix multiplication is the composition of transformations: $AB$ means "do $B$, then do $A$." It is not an arbitrary arithmetic rule to be memorized — the row-times-column formula is the consequence of computing the columns of the composite transformation. Once you read $AB$ as a single combined motion of space whose columns are $A$ applied to the columns of $B$, the product rule becomes inevitable and the most famous surprise of the subject — that $AB \ne BA$ — becomes obvious: the order in which you transform space matters.

The big ideas, in order

1. Addition and scalar multiplication are entrywise and easy. $A + B$ adds matching entries (same-shape only); $cA$ scales every entry. Geometrically, addition is the parallel combination (send the same input through both, add the outputs); it is commutative and makes the $m\times n$ matrices a vector space (Chapter 5).
2. Multiplication is composition — the rich operation. $(AB)\mathbf{x} = A(B\mathbf{x})$ for every $\mathbf{x}$: this is the definition. The matrix on the right acts first, exactly like $f(g(x))$.
3. The product rule is derived, not memorized. Columns-as-images (Chapter 7) forces it: the $j$-th column of $AB$ is $A$ applied to the $j$-th column of $B$. Spelling that out gives $(AB)_{ij} = \sum_k a_{ik}b_{kj}$ = row $i$ of $A$ dotted with column $j$ of $B$ — the row-times-column rule as the arithmetic shadow of composition.
4. Inner dimensions must match. $B$'s outputs must be valid inputs to $A$: an $(m\times n)(n\times p)$ product is $m\times p$. The inner $n$ cancels; the outer dimensions survive. Conformable = inner dimensions agree.
5. Matrix multiplication is NOT commutative. $AB \ne BA$ in general, because composition is order-dependent. The visualizer shows it: shear-then-rotate and rotate-then-shear produce different parallelograms. Commuting is the special case (shared structure / shared eigenvectors), not the default.
6. Powers iterate a transformation. For square $A$, $A^n$ = "apply $A$ $n$ times." Shears add ($S^n$ shears by $n$); rotations add angles ($R(\theta)^n = R(n\theta)$); adjacency-matrix powers count paths. High powers reveal long-run behavior — the doorway to eigenvalues (Chapter 23) and PageRank (Chapter 29).
7. The transpose $A^{\mathsf{T}}$ is the adjoint. Flip rows and columns; its meaning is $(A\mathbf{x})\cdot\mathbf{y} = \mathbf{x}\cdot(A^{\mathsf{T}}\mathbf{y})$. It reverses order in products: $(AB)^{\mathsf{T}} = B^{\mathsf{T}}A^{\mathsf{T}}$. It is not the inverse; the two coincide only for orthogonal matrices.
8. Structural laws hold — except commutativity. Multiplication is associative ($(AB)C = A(BC)$, both "do $C$, then $B$, then $A$") and distributive (left and right, separately). Square matrices form a non-commutative ring.
9. The identity $I$ is the "1" of multiplication. $AI = IA = A$, because composing with "do nothing" changes nothing. It anchors the inverse: $A^{-1}A = I$ means "do $A$, then undo it, equals do nothing" (Chapter 9).

The five readings of a product (fluency = switching freely)

Skills you gained

• Add and scalar-multiply matrices, and read both as combining transformations.
• Multiply matrices by hand three ways (composition/columns/entries) and explain why they agree.
• Derive the product rule from "apply $B$, then $A$" instead of memorizing it.
• Predict and verify that $AB \ne BA$ — both algebraically and with the visualizer's two pictures.
• Compute the transpose and apply the order-reversal rule $(AB)^{\mathsf{T}} = B^{\mathsf{T}}A^{\mathsf{T}}$.
• Use associativity, distributivity, the identity, and matrix powers $A^n$.
• Implement matmul from scratch as composition and verify it against numpy.

Terms to know

matrix addition, scalar multiplication, Hadamard (entrywise) product, matrix multiplication, composition of transformations, product rule, inner/outer dimensions, conformable, non-commutativity, commute, matrix power, transpose, adjoint, symmetric matrix, associativity, distributivity, identity matrix, zero matrix, non-commutative ring.

How this connects to the recurring themes

• Theme 1 (transformations are the point). Multiplication is literally the composition of transformations; the entire chapter refuses to present it as mere arithmetic.
• Theme 2 (geometry = algebra). The product matrix and the two-step motion are one object; the row-times-column rule and the side-by-side parallelograms describe the same fact.
• Theme 3 (computation validates theory). Every printed number matched numpy, and your from-scratch matmul now backs composition with code.
• Theme 6 (eigenvalues reveal what a matrix does). Powers $A^n$ and the question of commuting both point straight at eigenvectors — the long-run behavior and the shared-structure condition are eigen-facts in disguise.

Toolkit contribution

toolkit/matrices.py gains matmul(A, B) — the matrix product built from scratch as composition (column $j$ of $AB$ is apply(A, col_j_of_B)), plus add, scale, and identity. All verified against numpy's @ on square and non-square conformable pairs; the build raises a clear error on inner-dimension mismatch and confirms matmul(R, S) != matmul(S, R).

Forward references

• Chapter 9 — The inverse $A^{-1}$: the transformation that undoes $A$ ($A^{-1}A = I$); exists iff $A$ loses no information. The order-reversal $(AB)^{-1} = B^{-1}A^{-1}$ mirrors the transpose rule met here.
• Chapter 11 — The determinant: $\det(AB) = \det(A)\det(B)$ (the area-scalings multiply, as our step-by-step composition previewed).
• Chapter 12 — Computer graphics: composing model/view/projection matrices, with homogeneous coordinates folding translation into the product (Case Study 1).
• Chapter 18 — The dot product and the transpose-as-adjoint identity made rigorous.
• Chapter 23 & 29 — Eigenvectors and PageRank: matrix powers $A^n$ and long-run behavior; the steady state of Case Study 2's Markov chain is an eigenvector with eigenvalue 1.
• Chapters 30–33 — The outer-product reading powers the SVD, low-rank approximation, and the layer-by-layer composition inside neural networks.