Matrix Exponential Method

Core Idea

The solution to dx/dt = Ax, x(0) = x₀ is x(t) = e^(At)x₀, where e^(At) is the matrix exponential. If A = PDP⁻¹ (diagonalizable), then e^(At) = Pe^(Dt)P⁻¹ is easy to compute. The matrix exponential encodes all solution behavior and naturally handles initial conditions. It is the matrix analog of the scalar solution x(t) = e^(at)x₀.

Explainer

You already know how to solve the scalar ODE dx/dt = ax: the answer is x(t) = e^(at)x₀, where x₀ is the initial condition. The system dx/dt = Ax, where x is a vector and A is a matrix, is the exact same equation with scalars replaced by matrices. The solution has the same form: x(t) = e^(At)x₀. The only new ingredient is understanding what e^(At) means when A is a matrix.

The matrix exponential is defined by the same power series as the scalar exponential: e^(At) = I + At + (At)²/2! + (At)³/3! + ⋯, where I is the identity matrix. This definition is always valid but computing it directly requires summing infinitely many matrices — impractical for explicit solutions. This is where diagonalization from linear algebra rescues you. If A = PDP⁻¹, where D is diagonal, then:

• (At)^n = (PDP⁻¹)^n · t^n = P D^n P⁻¹ · t^n (the middle matrices collapse)
• So e^(At) = P e^(Dt) P⁻¹

The matrix e^(Dt) is trivial because D is diagonal: e^(Dt) is just the diagonal matrix with entries e^(λ₁t), e^(λ₂t), ..., where λ₁, λ₂, ... are the eigenvalues of A. Diagonalization transforms the hard problem of exponentiating a general matrix into the easy problem of exponentiating a diagonal one.

The geometric picture is illuminating. The columns of P are eigenvectors of A. In the eigenvector basis, the coupled system dx/dt = Ax decouples into independent scalar equations, each of the form du/dt = λu. The matrix exponential reassembles these decoupled solutions back into the original coordinates. Each eigenvalue governs the behavior of one "mode" of the system: negative real eigenvalues correspond to decaying modes, positive real eigenvalues to growing modes, and purely imaginary eigenvalues to oscillatory modes. The long-term behavior of the system is dominated by the eigenvalue with the largest real part — a fact that makes the eigenvalue structure of A the central object in stability analysis.