Key Takeaways — The Characteristic Polynomial and How to Find Eigenvalues

DataField.Dev

Key Takeaways — The Characteristic Polynomial and How to Find Eigenvalues

The big ideas

Eigenvalues are where $A - \lambda I$ goes singular. A number $\lambda$ is an eigenvalue of $A$ precisely when $(A - \lambda I)\mathbf{v} = \mathbf{0}$ has a nonzero solution — that is, when $A - \lambda I$ has a nontrivial null space, is rank-deficient, and therefore has determinant zero. This is not a recipe pulled from a hat; it is the eigen-equation of Chapter 23 rewritten as the singularity condition of Chapter 11. The whole chapter is contained in the equivalence $$ \lambda \text{ is an eigenvalue} \;\Longleftrightarrow\; \det(A - \lambda I) = 0. $$
The characteristic polynomial is the singularity detector. $p_A(\lambda) = \det(A - \lambda I)$ is a degree-$n$ polynomial whose roots are exactly the eigenvalues. Finding eigenvalues is finding its roots. Its constant term is $\det(A)$; its degree guarantees $n$ eigenvalues counted with multiplicity over $\mathbb{C}$.
The method is three steps, every time. (1) Form $A - \lambda I$. (2) Compute $p_A(\lambda) = \det(A - \lambda I)$ and find its roots — the eigenvalues. (3) For each eigenvalue, solve $(A - \lambda I)\mathbf{v} = \mathbf{0}$ — a null-space computation — for the eigenvectors. Nothing here is new machinery; it is the determinant, the null space, and row reduction you already owned, organized into a procedure.
Two multiplicities, and they can disagree. Algebraic multiplicity counts how many times $\lambda$ is a root; geometric multiplicity counts the dimension of the eigenspace $E_\lambda = N(A - \lambda I)$. They obey $1 \le m_g \le m_a$. When $m_g < m_a$ for some eigenvalue, the matrix is defective — short of eigenvectors — and (as Chapter 25 shows) cannot be diagonalized. A repeated eigenvalue does not automatically mean defective; you must compute the eigenspace dimension.
Trace and determinant know the spectrum. For every square matrix, $\operatorname{tr}(A) = \sum_i \lambda_i$ and $\det(A) = \prod_i \lambda_i$, both provable by comparing coefficients of $p_A$. These give instant checks on any eigenvalue computation, and they explain why a matrix is singular exactly when $0$ is one of its eigenvalues.
By-hand factoring does not scale — real solvers iterate. There is no radical formula for polynomial roots beyond degree $4$ (Abel–Ruffini), and forming the characteristic polynomial is numerically unstable anyway. Production software finds eigenvalues by iteration — power iteration (Chapter 23/29) and the QR algorithm (Chapter 38) — never by rooting the polynomial. The polynomial is for understanding; iteration is for computing.

Skills you gained

Forming and expanding the characteristic polynomial of a $2 \times 2$ and a $3 \times 3$ matrix (expanding along the zero-richest row or column).
Finding all eigenvalues as roots, then each eigenvector as a null-space basis, with a direct eigen-equation check.
Computing algebraic and geometric multiplicity and diagnosing a defective matrix.
Using $\operatorname{tr}(A) = \sum\lambda_i$ and $\det(A) = \prod\lambda_i$ as fast sanity checks.
Recognizing complex eigenvalues as the signature of a rotation, and $\lambda = 1$ as the guaranteed eigenvalue of a stochastic matrix.
Reading the eigenvalues of a system off its physics (natural frequencies, population growth rates) and trusting numpy for the parts the hand method cannot reach.

Terms to know

characteristic polynomial · characteristic equation · eigenvalue · eigenvector · eigenspace · spectrum · algebraic multiplicity · geometric multiplicity · defective matrix · trace · stochastic (Markov) matrix · stable age distribution · secular equation

Connections to the recurring themes

Geometry and algebra are two views of one object (Theme 2). This chapter is the theme's sharpest demonstration: a single eigenvalue is simultaneously a root of a polynomial, a stretch factor you can watch in the visualizer, a value that makes a matrix singular, and a contribution to trace and determinant. The characteristic polynomial is the hinge between these views.
The four fundamental subspaces (Theme 5). Eigenvectors are null-space vectors — $E_\lambda = N(A - \lambda I)$ — so the entire eigenvalue computation is the Part III null-space machinery applied to a shifted matrix. Eigenvalue $\Leftrightarrow$ rank drop $\Leftrightarrow$ nonzero null space.
Eigenvalues reveal what a matrix really does (Theme 6). Strip away the coordinate system and a matrix is a set of independent stretches along its eigen-directions, with the eigenvalues as the stretch factors — the skeleton on which the transformation hangs.
Computation validates theory and theory guides computation (Theme 3). We proved why the method works, computed by hand to build intuition, and verified with numpy — and the closing section drew the honest line between what is provable on paper and what is computable in practice.

Forward references — where this leads

Chapter 25 (Diagonalization) is the immediate payoff. When a matrix has a full set of independent eigenvectors (no defect), they become the columns of $P$ and the eigenvalues the diagonal of $D$, so $A = PDP^{-1}$. This turns the hard problem of $A^{100}$ into the trivial $D^{100}$ — the engine behind every Markov chain, population model, and linear dynamical system. The eigenvalues you learned to find here are the diagonal of $D$; the eigenvectors are the columns of $P$. The whole reason multiplicity mattered is that diagonalization requires $m_g = m_a$ for every eigenvalue.
Chapter 26 (Complex Eigenvalues) confronts the matrices whose characteristic polynomial has no real roots — the rotations of §24.9 — and reads rotation angle and scaling straight off the complex eigenvalues.
Chapter 27 (Spectral Theorem) explains why the symmetric matrices of the vibration case study have orthogonal eigenvectors and never defect.
Chapter 29 (PageRank) and Chapter 38 (Numerical Linear Algebra) deliver the iterative algorithms — power iteration and the QR algorithm — that find eigenvalues of the giant matrices this chapter's hand method cannot touch, completing the story §24.10 began.