A topic in the Open Knowledge Graph — a free, open map of 15,290 topics and the order to learn them in.
The composition of two linear transformations T: Rⁿ → Rᵐ and S: Rᵐ → Rᵖ is the linear transformation S ∘ T: Rⁿ → Rᵖ given by (S ∘ T)(x) = S(T(x)). If A is the standard matrix of T and B is the standard matrix of S, then the standard matrix of S ∘ T is the matrix product BA. This reveals why matrix multiplication is defined the way it is: it encodes function composition. Because function composition is not commutative, matrix multiplication is generally not commutative either (BA ≠ AB).
Apply two successive geometric transformations (e.g., rotate then reflect) and observe that the combined effect equals multiplication by the product of the two matrices in the correct order. Experiment with reversing the order to see non-commutativity explicitly.
The key insight is that matrix multiplication is not an arbitrary algebraic recipe — it is function composition in disguise. You've already learned that a matrix represents a linear transformation: a rule that rotates, scales, shears, or reflects every vector in the input space. Now suppose you want to apply two transformations in sequence: first T, then S. This is exactly what function composition means — (S ∘ T)(x) = S(T(x)) — and multiplying the matrices in the correct order is how that composition is encoded.
To build concrete intuition: suppose T rotates vectors 90° counterclockwise (represented by matrix A) and S reflects across the x-axis (represented by matrix B). Applying T then S to a vector v means computing S(T(v)) = B(Av) = (BA)v. The combined transformation "rotate then reflect" is captured by the single matrix BA. This is why the transformation applied first appears *rightmost* in the product: the rightmost matrix acts first, matching the way function composition is written right-to-left.
The non-commutativity of matrix multiplication (BA ≠ AB in general) is not an algebraic quirk — it directly reflects that "rotate then reflect" is a different transformation than "reflect then rotate." Test this geometrically: take the point (1, 0), rotate it 90° counterclockwise to (0, 1), then reflect across the x-axis to (0, −1). Now reverse the order: reflect (1, 0) to (1, 0) first (already on the x-axis), then rotate to (0, 1). Different result. The matrices are doing the honest arithmetic of this asymmetry.
Associativity, on the other hand, holds because function composition is associative: it doesn't matter how you group a chain of transformations, as long as you preserve their order. A(BC) = (AB)C because applying C, then B, then A gives the same result regardless of whether you compute B∘C first or A∘B first. This means you can parenthesize a long chain of matrix multiplications however is computationally convenient — but you can never reorder the factors.