Chapter 37 — Key Takeaways

The one idea

Eigenvalues are the destiny of a linear dynamical system. The system $\mathbf{x}' = A\mathbf{x}$ is solved exactly by $\mathbf{x}(t) = e^{At}\mathbf{x}(0)$, and the long-run behavior — decay, growth, oscillation — is dictated entirely by the eigenvalues of $A$: their real parts decide stability, their imaginary parts decide oscillation, and together they classify the phase portrait. This is recurring theme #6 of the book (eigenvalues reveal what a matrix really does) reaching its most dramatic form: in a world that evolves in time, the eigenvalues are the future. The matrix exponential is the machine that turns the eigenvalues into the flow of time — the same Taylor series that defines $e^{at}$, now fed a matrix.

The big ideas, in order

  1. A linear system is a velocity field. $\mathbf{x}'=A\mathbf{x}$ plants the velocity vector $A\mathbf{x}$ at each point; trajectories are streamlines and the whole flow pattern is the phase portrait. The origin is always an equilibrium (since $A\mathbf{0}=\mathbf{0}$), and the central question is whether it is stable.

  2. The matrix exponential is the same series as the scalar one, and always converges. $e^{At} = \sum_{k=0}^\infty (At)^k/k!$ converges for every square matrix and every $t$ (dominated by the scalar series for $e^{\lVert A\rVert|t|}$), so $e^{At}$ always exists — even when $A$ cannot be diagonalized. At $t=0$ it is $I$.

  3. Compute it by diagonalization when you can. If $A=PDP^{-1}$, then $e^{At}=Pe^{Dt}P^{-1}$, where $e^{Dt}$ is the diagonal matrix of $e^{\lambda_i t}$. Geometrically: change to eigen-coordinates, let each mode exponentiate at its own rate $e^{\lambda_i t}$, change back. The eigenvalues are the rates of the independent modes.

  4. $e^{At}$ solves the system because $\frac{d}{dt}e^{At}=Ae^{At}$. This one derivative fact (proved by differentiating the series term by term) makes $\mathbf{x}(t)=e^{At}\mathbf{x}(0)$ satisfy both the equation and the initial condition. The matrix $e^{At}$ is the flow / state-transition matrix — the transformation that is the passage of time.

  5. The solution is a superposition of eigenmodes. Writing $\mathbf{x}(0)=\sum c_i\mathbf{v}_i$ in the eigenbasis gives $\mathbf{x}(t) = \sum_i c_i e^{\lambda_i t}\mathbf{v}_i$. Each eigenvector is a mode growing/decaying like $e^{\lambda_i t}$; the eigenvalue closest to zero sets the slow relaxation timescale, and the ratio of fastest to slowest is the stiffness.

  6. Complex eigenvalues mean oscillation. For $\lambda = \alpha\pm\beta i$, Euler's formula splits $e^{\lambda t}=e^{\alpha t}(\cos\beta t + i\sin\beta t)$ into an envelope $e^{\alpha t}$ (real part = decay/growth) and an oscillation at frequency $\beta$ (imaginary part = rhythm). Real eigenvalues cannot oscillate — oscillation is the fingerprint of complex eigenvalues.

  7. The defective case needs Jordan form and produces $t\,e^{\lambda t}$. A repeated eigenvalue short of eigenvectors is split as $\lambda I + N$ with $N$ nilpotent; $e^{Nt}=I+Nt$ terminates, giving $e^{Jt}=e^{\lambda t}\begin{bmatrix}1&t\\0&1\end{bmatrix}$. The missing eigenvector is paid for with a polynomial factor $t^{m}$ — but the exponential still dominates, so stability is still set by $\operatorname{Re}\lambda$.

  8. The stability dictionary (state the conditions!). Asymptotically stable $\iff$ all $\operatorname{Re}\lambda<0$; unstable $\iff$ some $\operatorname{Re}\lambda>0$; borderline when the largest real part is exactly $0$ (bounded oscillation if non-defective, unstable if defective). Real part negative ⇒ decay; positive ⇒ growth; zero ⇒ knife-edge.

  9. Four phase-portrait types, classified by eigenvalues. Real same-sign ⇒ node (stable if negative); real opposite-sign ⇒ saddle (always unstable); complex with nonzero real part ⇒ spiral (stable if real part negative); purely imaginary ⇒ center (closed orbits). The trace-determinant plane is the shortcut: trace controls stability, determinant and discriminant control type.

  10. $e^{A+B}\ne e^Ae^B$ unless $AB=BA$. Matrix non-commutativity reaches into the exponential. The law holds if and only if $A$ and $B$ commute. Always-true consequences: $e^{A}e^{-A}=I$ (so $e^{At}$ is always invertible, inverse $e^{-At}$) and $e^{At}e^{As}=e^{A(t+s)}$ (the flow composes in time).

Skills you gained

  • Compute the matrix exponential $e^{At}$ three ways: the raw power series, diagonalization $Pe^{Dt}P^{-1}$ (for diagonalizable $A$), and the Jordan splitting (for defective $A$) — and know to call scipy.linalg.expm in practice.
  • Solve $\mathbf{x}'=A\mathbf{x}$ with $\mathbf{x}(t)=e^{At}\mathbf{x}(0)$ and as a sum of eigenmodes, and verify against scipy.integrate.solve_ivp.
  • Read stability off the signs of the real parts of the eigenvalues, stating the conditions precisely.
  • Classify any $2\times2$ phase portrait as node, saddle, spiral, or center — from the eigenvalues or, faster, from the trace, determinant, and discriminant.
  • Recognize when the exponential law fails, checking $AB=BA$ before splitting $e^{A+B}$.

Terms to know

matrix exponential ($e^{At}$) · system of differential equations · linear dynamical system · trajectory · phase portrait · equilibrium · asymptotic stability · fundamental / state-transition matrix · eigenmode · stiffness · node · saddle · spiral (focus) · center · trace-determinant plane · underdamped / critically damped / overdamped · nilpotent · generalized eigenvector.

How this ties to the book

  • Recurring theme #6 (eigenvalues reveal what a matrix really does): its most consequential appearance — eigenvalues as the fate of a time-evolving system. We ranked the web with a dominant eigenvalue in Chapter 29; here we predict stability with the whole spectrum.
  • Recurring theme #1 (a matrix is a transformation): $e^{At}$ is the transformation that is the passage of time — the flow that carries every state forward.
  • Recurring theme #3 (computation validates theory): every numeric result was confirmed three ways — closed form, expm, and solve_ivp — agreeing to machine precision.
  • Builds directly on: diagonalization $A=PDP^{-1}$ and powers of a matrix (Chapter 25); complex eigenvalues and Euler's formula (Chapter 26); the Jordan normal form and generalized eigenvectors (Chapter 36); the characteristic polynomial and trace/determinant relations (Chapter 24); the Spectral Theorem's orthogonal eigenvectors for symmetric matrices (Chapter 27).

Forward references

  • Chapter 38 (Numerical Linear Algebra) is the reality check this chapter twice anticipated: why the diagonalization route fails for defective matrices (ill-conditioned $P$), and why scipy.linalg.expm succeeds with a different, numerically stable algorithm (scaling-and-squaring). The condition number will quantify the sensitivity we kept gesturing at.
  • Cross-book: the link between differential equations and eigenvalues founded here is the gateway to dynamical-systems theory; and the operator $e^{-iHt}$ — the matrix exponential of a Hermitian Hamiltonian — is exactly time evolution in quantum mechanics, where the eigenvalues of $H$ are the energy levels and the qubit of Chapters 5/21/27/34 evolves by precisely this $e^{At}$.

The one sentence to keep: Find the eigenvalues of $A$ and look at their real parts — negative is decay, positive is growth, imaginary is oscillation — and you know the fate of $\mathbf{x}'=A\mathbf{x}$ before computing a single trajectory.