Key Takeaways — Chapter 23: Eigenvalues and Eigenvectors

The big ideas

  • An eigenvector is an invariant direction. Apply a matrix to a typical vector and it rotates to point somewhere new; apply it to an eigenvector and the vector stays on its own line, only stretched. The eigenvectors are the few directions a transformation does not turn. This geometric picture — visible as the dashed eigen-lines that stay on themselves in the visualizer — comes first; the algebra describes it.

  • The eigen-equation $A\mathbf{v} = \lambda\mathbf{v}$ is the whole subject in one line. It says that on the special vector $\mathbf{v}$, the entire matrix $A$ acts exactly like the single scalar $\lambda$ — transforming equals scaling. The pair $(\lambda, \mathbf{v})$ is an eigenpair; the list of all eigenvalues is the spectrum.

  • The eigenvalue is a stretch factor, and its sign and size say everything. $\lambda > 1$ grows the direction, $0 < \lambda < 1$ shrinks it, $\lambda = 1$ fixes it, $\lambda < 0$ flips it, and $\lambda = 0$ collapses it (so $0$ is an eigenvalue exactly when the matrix is singular). For long-run repeated application, the magnitude $|\lambda|$ is the growth rate, and the largest-$|\lambda|$ direction — the dominant eigenvector — wins.

  • An eigenvector is really a direction, not one arrow. Any nonzero scalar multiple of an eigenvector is again an eigenvector with the same eigenvalue, so eigenvectors are determined only up to scale (numpy returns the unit-length representative). The zero vector is never an eigenvector — it satisfies the equation for every $\lambda$, which is exactly why we forbid it.

  • Eigenvalues reveal what a matrix really does (recurring theme #6). The eigenvectors are the transformation's own natural axes; in those coordinates the matrix becomes diagonal, just a list of stretch factors. The entries of a matrix depend on your coordinate system; its eigenvalues do not. Similar matrices share eigenvalues — the eigenvalues are the coordinate-free truth about a transformation.

Skills you gained

  • Recognize an eigenvector geometrically and predict eigenvalues of familiar transformations (identity → all $\lambda=1$; scaling $cI$ → all $\lambda=c$; reflection → $\pm1$; projection → $1$ and $0$; rotation → none real).
  • Find the eigenvalues of a $2\times2$ matrix by setting $\det(A - \lambda I) = 0$ — a condition you can now derive from "$(A-\lambda I)\mathbf{v}=\mathbf{0}$ needs a nonzero solution," not just memorize — and find each eigenvector as the null space of $A - \lambda I$.
  • Use the trace–determinant shortcut: for a $2\times2$, the eigenvalues sum to $\operatorname{tr}(A)$ and multiply to $\det(A)$.
  • Describe the eigenspace $E_\lambda = N(A - \lambda I)$ as a subspace, and use the eigen-decomposition trick to apply $A$ (and $A^k$) to any vector by splitting it along the eigen-axes.
  • Verify eigenpairs with np.linalg.eig, understanding that its eigenvectors are unit-length and may differ from yours by a scalar or sign.
  • Sketch the dominant eigenvector via power iteration (your toolkit contribution) and explain why it converges.

Terms to know

eigenvector, eigenvalue, eigen-equation ($A\mathbf{v}=\lambda\mathbf{v}$), invariant direction, eigenpair, eigenspace ($E_\lambda = N(A-\lambda I)$), spectrum, stretch factor, dominant eigenvector, characteristic equation ($\det(A-\lambda I)=0$), trace, defective matrix, power iteration, Rayleigh quotient, steady state, normal mode.

How this connects

  • Backward. Eigenvectors live on the four fundamental subspaces of Part III (theme #5): the eigenspace is the null space $N(A - \lambda I)$, and $\lambda = 0$ ties to the ordinary null space $N(A)$ and singularity (Chapters 9, 11, 13). Finding eigenvectors is the null-space computation of Chapter 13 applied to shifted matrices. The transformation viewpoint (Chapter 7) and similarity/change of basis (Chapter 16) are what make "coordinate-free essence" precise.
  • Forward in Part V. Chapter 24 turns $\det(A - \lambda I) = 0$ into a reliable method (the characteristic polynomial) and distinguishes algebraic from geometric multiplicity — the subtlety we previewed with the deficient shear. Chapter 25 (diagonalization) cashes in a full set of eigenvectors as $A = PDP^{-1}$, making matrix powers trivial. Chapter 26 rescues the rotations with no real eigenvectors via complex eigenvalues. Chapter 27 (the Spectral Theorem) shows symmetric matrices have orthogonal eigenvectors. Chapter 29 (PageRank) computes the dominant eigenvector of the web by power iteration.
  • Across the book and beyond. The dominant eigenvector is the engine of PageRank (Chapter 29) and of PCA (Chapter 32), where the eigenvectors of a covariance matrix are the data's principal directions. In physics, eigenvalues are the measurable values of observables and eigenstates in quantum mechanics, and the natural frequencies of any vibrating structure. The matrix exponential (Chapter 37) solves $\mathbf{x}' = A\mathbf{x}$ with the eigenvalues setting stability.

The one sentence to remember

Almost every vector a matrix touches gets rotated, but a precious few special directions are only stretched — those invariant directions are the eigenvectors, their stretch factors are the eigenvalues, and together they are the skeleton on which the entire transformation hangs.