Chapter 35 — Key Takeaways
The one idea
The matrix was never the transformation. A linear map $T:V\to W$ between abstract vector spaces is the real object; choosing a basis for the domain and one for the codomain merely photographs it as a matrix, and a different choice of bases gives a different photograph of the very same map. Once you hold the map fixed and treat its matrix as a basis-dependent view, the whole book reorganizes around the transformation as the thing that is real — recurring theme #1, finally stated in full. Kernel and image are the null and column spaces freed from coordinates; rank–nullity conserves dimension; and every $n$-dimensional space is $\mathbb{R}^n$ in disguise.
The big ideas, in order
-
A linear transformation is defined by two axioms, not by a matrix. $T$ is linear iff it is additive ($T(\mathbf{u}+\mathbf{v})=T(\mathbf{u})+T(\mathbf{v})$) and homogeneous ($T(c\mathbf{v})=cT(\mathbf{v})$). These are exactly the two facts that made $\mathbf{x}\mapsto A\mathbf{x}$ linear in Chapter 7, with the coordinates removed — so "linear map" is meaningful between any two vector spaces. Every linear map fixes $\mathbf{0}$ and preserves linear combinations, which is why a map is completely determined by its action on a basis.
-
Differentiation is the anchor. $\frac{d}{dx}$ on polynomials is a linear operator: the sum rule and constant-multiple rule of calculus are precisely the two axioms. The product rule and chain rule are nonlinear, which is why the operator view captures the linear heart of calculus.
-
The matrix of a map is built from images of basis vectors. With bases $B$ of $V$ and $C$ of $W$, the matrix $[T]_{C\leftarrow B}$ has $j$-th column $[T(\mathbf{b}_j)]_C$, and satisfies $[T(\mathbf{v})]_C = [T]_{C\leftarrow B}\,[\mathbf{v}]_B$. Applying the abstract map becomes ordinary matrix–vector multiplication on coordinate vectors.
-
The matrix is basis-dependent; the operator is not. Change basis and the matrix transforms by the similarity $[T]_{\tilde B}=P^{-1}[T]_B P$ of Chapter 16 (proved this chapter). So one operator has many similar matrices, all sharing determinant, trace, rank, characteristic polynomial, and eigenvalues — the operator's invariants. The anchor shows it vividly: $D$ is the superdiagonal $1,2,3$ in the monomial basis but a clean shift in the factorial basis.
-
Kernel and image generalize null space and column space. $\ker T=\{\mathbf{v}:T(\mathbf{v})=\mathbf{0}\}$ (a subspace of the domain) and $\operatorname{im}T=\{T(\mathbf{v})\}$ (a subspace of the codomain). In coordinates they are exactly $N([T])$ and $C([T])$. $T$ is injective iff $\ker T=\{\mathbf{0}\}$; surjective iff $\operatorname{im}T=W$. For $D$: kernel = constants (the "$+C$"), image = lower-degree polynomials.
-
Abstract Rank–Nullity conserves dimension. $\dim\ker T+\dim\operatorname{im}T=\dim V$ for finite-dimensional $V$ — crushed directions plus surviving directions equal the whole domain. Proved by extending a kernel basis to a basis of $V$. On $D:\mathbb{P}_3\to\mathbb{P}_3$, $1+3=4$. A corollary: for an operator $T:V\to V$ on a finite-dimensional space, injective $\iff$ surjective (fails in infinite dimensions).
-
Every $n$-dimensional space is isomorphic to $\mathbb{R}^n$. An isomorphism is a bijective linear map; two finite-dimensional spaces are isomorphic iff they have the same dimension. The coordinate map $\mathbf{v}\mapsto[\mathbf{v}]_B$ is the isomorphism $V\cong\mathbb{R}^n$ — which is exactly why coordinates work: we think abstractly and compute in $\mathbb{R}^n$, faithfully, because the two are the same space relabeled. Dimension is the complete invariant.
-
Nilpotence is differentiation's signature. $D^{n+1}=0$ on $\mathbb{P}_n$ — differentiate a degree-$n$ polynomial $n+1$ times and nothing remains. A nilpotent operator has only the eigenvalue $0$, and it is the purest seed of the defective matrices that Jordan form (Chapter 36) studies.
Skills you gained
- Test whether a map is linear by checking additivity and homogeneity (and the fast pre-screen $T(\mathbf{0})=\mathbf{0}$), and produce a counterexample when it is not.
- Build the matrix of an abstract linear map in chosen bases by computing images of basis vectors — for differentiation, evaluation, the shift, integration, and the transpose operator.
- Find the kernel and image of an abstract map and read off injectivity, surjectivity, and rank/nullity.
- Apply rank–nullity to count solution-space dimensions, reconcile finite- and infinite-dimensional behavior, and reason about redundancy in codes.
- Change the basis of an operator via the similarity $P^{-1}[T]_BP$, and recognize which quantities are basis-independent invariants.
- Recognize the coordinate map as an isomorphism $V\cong\mathbb{R}^n$, and explain why finite-dimensional linear algebra is the same in every $n$-dimensional space.
Terms to know
linear transformation · abstract vector space · additivity · homogeneity · linear operator · matrix of a linear map ($[T]_{C\leftarrow B}$) · coordinate vector · similarity · kernel · image (range) · injective · surjective · rank · nullity · rank–nullity theorem · isomorphism · coordinate map · differentiation operator · shift operator · evaluation map · integration operator · nilpotent operator
Connections — backward and forward
- Back to Chapter 7: "a matrix is a function that transforms space" was stated for $\mathbb{R}^n$; this chapter removes the coordinates and reveals the linear map as the real object.
- Back to Chapter 16: the change-of-basis similarity $P^{-1}AP$ is exactly how an operator's matrix transforms; we proved it here for abstract operators.
- Back to Chapter 13–14: the null space and column space, and matrix rank–nullity, are the $V=\mathbb{R}^n$ special cases of kernel, image, and abstract rank–nullity.
- Back to Chapter 5: the abstract vector-space axioms made polynomials, functions, and matrices into vector spaces; this chapter studies the maps between them.
- Back to Chapter 34: the companion liberation — that chapter freed the dot product from $\mathbb{R}^n$; this one frees the matrix. Together they are the two halves of coordinate-free linear algebra.
- Forward to Chapter 36: the operators whose matrices cannot be diagonalized in any basis lead to generalized eigenvectors and the Jordan normal form; the nilpotent $D$ of this chapter is the prototype.
- Forward to Chapter 37: the matrix exponential $e^{At}$ solves the operator equation $\mathbf{x}'=A\mathbf{x}$, fusing the differentiation operator with the matrix viewpoint into the engine of differential equations.
- Cross-book: the operator view of $\frac{d}{dx}$ is the derivative as an operator, and physical observables are operators in quantum mechanics — both are linear transformations on abstract (often infinite-dimensional) spaces, represented by matrices only after a basis is chosen.
Remember this: $T:V\to W$ is the noun; $[T]_{C\leftarrow B}$ is one of its many photographs. Hold the transformation fixed, treat its matrix as a coordinate-dependent view, and the kernel/image, rank–nullity, and isomorphism theorems of this chapter put every linear map — finite matrix, differential operator, or quantum observable — under one roof.