JOVANA
Library Glossary Getting Started Three Levels Fields How it works Mission
Join the mission
All guides

Linear dynamical systems and the matrix exponential

Let the state evolve over time. Diagonalization turns a coupled system into independent scalar exponentials, and the eigenvalues alone decide whether everything settles, oscillates, or blows up.

Two clocks: discrete steps and continuous flow

A linear dynamical system comes in two flavors. Discrete time: x_{t+1} = A x_t, so x_t = A^t x_0 — exactly the Markov-chain iteration from guide 2, now without the probability constraint. Continuous time: dx/dt = A x, whose solution is x(t) = e^{At} x_0 using the matrix exponential. Both are governed by the eigenvalues of A.

Diagonalization decouples the system

If A = Q Lambda Q^{-1} is diagonalizable, change coordinates to the eigenbasis with y = Q^{-1} x. The coupled system becomes n independent scalar equations: dy_i/dt = lambda_i y_i, each solved by y_i(t) = e^{lambda_i t} y_i(0). The matrix exponential is then trivial: e^{At} = Q diag(e^{lambda_1 t}, ..., e^{lambda_n t}) Q^{-1}.

# Coupled springs / two tanks: dx/dt = A x with
#   A = [ -2 ,  1 ;
#          1 , -2 ]
# eigenvalues:  lambda_1 = -1 ,  lambda_2 = -3   (both negative)
# eigenvectors: q1 = (1, 1)/sqrt2 ,  q2 = (1,-1)/sqrt2
# general solution:
#   x(t) = c1 e^{-1 t} q1  +  c2 e^{-3 t} q2
# both modes decay -> x(t) -> 0 .  The system is STABLE.
Each eigenvalue is one independent decaying (or growing) mode.

Eigenvalues alone decide stability

Since every mode behaves like e^{lambda_i t} (continuous) or lambda_i^t (discrete), stability is read straight off the spectrum. Complex eigenvalues alpha +/- i*beta add oscillation at frequency beta with envelope e^{alpha t}; it is the real part that decides growth or decay.

  1. Continuous dx/dt = A x: stable iff every eigenvalue has real part < 0 (left half-plane).
  2. Discrete x_{t+1} = A x_t: stable iff every eigenvalue has |lambda| < 1 (inside the unit circle).
  3. Marginal cases sit exactly on the boundary; one eigenvalue past it and the system blows up.