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

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.