The Matrix Exponential

A system of quantities that all influence each other's rate of change — say, capital flowing between three linked trading strategies, each bleeding into or draining from the others at fixed rates — is described by a system of coupled linear differential equations. Solving each equation separately is impossible because they're tangled together. The matrix exponential solves the whole system at once, producing a single matrix that tells you exactly where every quantity will be at any future time, given only where it started.

An analogy: compound growth, but for several linked accounts

Ordinary compound growth says a single quantity growing continuously at rate $r$ becomes $x_0 e^{rt}$ after time $t$ — the familiar exponential. Now imagine several accounts that don't just grow on their own, but also leak into and receive from each other continuously. There's no single growth rate anymore; there's a whole matrix of rates. The matrix exponential is exactly the multi-account version of $e^{rt}$: instead of one number growing, an entire vector of linked quantities evolves together, and the "growth rate" $r$ is replaced by a matrix $A$ describing every account's self-growth and cross-leakage simultaneously.

The math, one symbol at a time

For a system $\frac{d}{dt}x(t) = Ax(t)$ with initial condition $x(0)$, the solution is

$$x(t) = e^{At} x(0), \qquad e^{At} = \sum_{k=0}^{\infty} \frac{(At)^k}{k!}.$$

In words: $e^{At}$ is defined the same way as the ordinary exponential's power series, just with the matrix $A$ (times $t$) plugged in for the scalar exponent, and matrix powers used instead of number powers. Computing that infinite sum directly is impractical, but if $A$ is diagonalizable as $A = PDP^{-1}$, then $e^{At} = Pe^{Dt}P^{-1}$, where $e^{Dt}$ is just a diagonal matrix of ordinary scalar exponentials $e^{\lambda_i t}$ — the infinite matrix sum collapses to exponentiating each eigenvalue separately.

Worked example 1: a two-strategy capital flow system

Two strategies pass capital back and forth: strategy 1 grows at rate 1 but leaks to strategy 2 at rate 1; strategy 2 shrinks at rate $-1$ but nothing flows into it otherwise. $A = \begin{pmatrix}1 & 0\\1 & -1\end{pmatrix}$ has eigenvalues $\lambda_1 = 1$, $\lambda_2 = -1$ (it's already lower triangular, so eigenvalues sit on the diagonal), with eigenvectors $(2,1)$ and $(0,1)$. Diagonalizing and exponentiating each eigenvalue: capital along the $\lambda_1=1$ direction grows like $e^t$, while capital along the $\lambda_2=-1$ direction decays like $e^{-t}$. Starting from $x(0) = (1,0)$, expanding in the eigenvector basis and recombining gives $x(t) = (e^t, \, \frac{1}{2}(e^t - e^{-t}))$ — strategy 1 grows exponentially, and strategy 2 first absorbs inflow, then its own decay term eventually dominates.

Worked example 2: a continuous-time Markov chain transition matrix

A trading state (idle, active) switches with generator matrix $Q = \begin{pmatrix}-2 & 2\\1 & -1\end{pmatrix}$ (rows sum to zero, off-diagonals are transition rates). The probability of being in each state after time $t$, starting idle, is the first row of $e^{Qt}$. $Q$'s eigenvalues are $0$ and $-3$; the $\lambda=0$ eigenvalue never decays and corresponds to the long-run steady state, while the $\lambda=-3$ term $e^{-3t}$ decays to zero — after about $t=1$, $e^{-3} \approx 0.05$, so the chain is already 95% of the way to its steady-state split of time between idle and active, read directly off the surviving $\lambda=0$ eigenvector.

Watch how a mean-reverting path settles toward its long-run level at a rate set by its decay parameter — the same exponential-decay mechanism that a negative eigenvalue produces inside $e^{At}$.

What this means in practice

The matrix exponential is the standard tool for propagating any continuous-time linear system forward: continuous-time Markov chain transition probabilities, term-structure models built on linear ODEs, and coupled mean-reverting processes. It's also the bridge between discrete-time matrix powers (see Diagonalization and Similarity) and continuous time — $e^{At}$ is what a transition matrix $A^n$ becomes when the number of steps $n$ is replaced by a continuous clock.