The Matrix Exponential

Can a single formula solve every linear system of ODEs at once? This holds regardless of the eigenstructure of the coefficient matrix. For scalar equations $x' = ax$ the answer is the familiar $x(t) = e^{at}x_0$ . It is natural to ask whether a matrix version of the exponential could play the same universal role for $\mathbf{x}' = A\mathbf{x}$ . The matrix exponential $e^{At}$ answers this question completely. It is defined for every square matrix $A$ , real or complex, diagonalizable or not. It packages the entire solution theory of linear systems into one object. This includes eigenvalues, generalized eigenvectors, resonance terms from repeated roots, and oscillation from complex pairs. All of this satisfies a matrix-valued differential equation. This matters because it replaces case-by-case solution recipes with a single unifying structure. Instead of memorizing separate procedures for distinct real eigenvalues, complex conjugate pairs, and defective (repeated) eigenvalues, one shows something once and for all. Namely, that $e^{At}$ is the fundamental matrix solution of $\mathbf{x}'=A\mathbf{x}$ , normalized to equal the identity at $t=0$ . One then computes it in each case using whatever eigenstructure is available. The exponential also clarifies why the classical recipes work. They are simply different algorithms for evaluating the same underlying series. The big picture is that $e^{At}$ is to the matrix equation $\mathbf{x}'=A\mathbf{x}$ exactly what $e^{at}$ is to the scalar equation $x'=ax$ . The caveat is that matrix multiplication does not commute in general. This is the source of nearly every extra subtlety. These include the failure of $e^{A+B}=e^Ae^B$ in general, the need for the Jordan form, and the appearance of polynomial-in- $t$ terms. Once this bridge is built, the long-run behavior of solutions becomes clear. Decay, growth, rotation, and resonance become a direct readout of the eigenvalues of $A$ . This article defines $e^{At}$ rigorously as a convergent matrix power series. It proves $e^{At}$ is the fundamental matrix of $\mathbf{x}'=A\mathbf{x}$ with $e^{A\cdot 0}=I$ . It establishes its algebraic properties: the group law when matrices commute, invertibility, and the effect of similarity transformations. It also gives complete algorithms for computing it via diagonalization and via the Jordan form. These cover real, complex, and repeated eigenvalue cases uniformly.

Key takeaways

$e^{At}$ is defined by the convergent series $\sum_{k=0}^\infty \frac{(At)^k}{k!}$ for any square matrix $A$ .
It solves the matrix IVP $\frac{d}{dt}e^{At} = Ae^{At}$ , $e^{A\cdot 0}=I$ , making it the fundamental matrix of $\mathbf{x}'=A\mathbf{x}$ .
$e^{At}$ is always invertible, with $(e^{At})^{-1}=e^{-At}$ .
If $A=PDP^{-1}$ is diagonalizable, then $e^{At}=Pe^{Dt}P^{-1}$ , reducing computation to scalar exponentials.
For non-diagonalizable $A$ , the Jordan form gives $e^{At} = P e^{Jt} P^{-1}$ where each Jordan block contributes polynomial-in- $t$ correction terms.
Key idea: $e^{At}$ turns the linear ODE $\mathbf{x}'=A\mathbf{x}$ into pure linear algebra — all solution behavior is read off from the eigenstructure of $A$ .

Where this lives today

The matrix exponential is the computational backbone of control theory, where the state-transition matrix $e^{At}$ propagates the state of a linear time-invariant system $\dot{\mathbf{x}}=A\mathbf{x}+B\mathbf{u}$ forward in time via the variation-of-parameters formula, and stability of a controller reduces to checking that all eigenvalues of the closed-loop matrix have negative real part (so $e^{At}\to 0$ ). In <a class="wiki-link" href="/wiki/math-diffeq-phase-portraits-for-linear-systems" data-track="wiki_article.body.wiki_link">Phase Portraits for Linear Systems</a>, the qualitative behavior — nodes, saddles, spirals, centers — is exactly the geometric picture drawn by the trajectories $t\mapsto e^{At}\mathbf{x}_0$ , and the eigenvalue classification used there is a direct consequence of the computations shown here. In Markov chain theory and continuous-time stochastic processes, the exponential of a generator matrix $Q$ gives the transition probability matrix $P(t) = e^{Qt}$ of a continuous-time Markov chain, with row sums preserved by properties of the exponential analogous to those proved above. In numerical analysis, computing $e^{At}$ accurately and efficiently (via scaling-and-squaring combined with Padé approximants, the "expm" algorithm used in software like MATLAB and Julia) is a well-studied problem precisely because naive truncation of the defining series can be numerically unstable for matrices with large norm or highly non-normal structure. In quantum mechanics, the time-evolution operator $U(t) = e^{-iHt/\hbar}$ for a time-independent Hamiltonian $H$ is a matrix (or operator) exponential, and its unitarity for Hermitian $H$ is the exact analogue of Part 4 of the theorem specialized to skew-Hermitian generators. In statistics, the matrix exponential of a covariance-generating drift matrix appears in continuous-time autoregressive processes and Ornstein–Uhlenbeck-type models used in finance and signal processing.

Ask the tutor: How does the sign pattern of the eigenvalues of $A$ determine whether trajectories of $\mathbf{x}'=A\mathbf{x}$ spiral inward, spiral outward, or trace closed loops in the phase plane?

A

Intuition

Think first about why the series converges at all. Fix any matrix norm $\|\cdot\|$ that is submultiplicative, meaning $\|AB\|\le \|A\|\|B\|$ (the operator norm and the Frobenius norm both work). Then $\|A^k\| \le \|A\|^k$ , so the series of norms $\sum_k \|A\|^k t^k/k!$ is dominated by the scalar series for $e^{\|A\|t}$ , which converges for every $t$ . This is the exact matrix analogue of the ratio test showing $\sum t^k/k!$ converges absolutely everywhere — the noncommutativity of matrix multiplication never enters the convergence argument, only the algebra of what the sum equals. Once convergence is settled, the key intuition is dynamical: $e^{At}\mathbf{x}_0$ traces out the trajectory of the system $\mathbf{x}'=A\mathbf{x}$ starting at $\mathbf{x}_0$ . If $A$ happens to be diagonal, say $A=\mathrm{diag}(\lambda_1,\dots,\lambda_n)$ , then $A^k$ is diagonal with entries $\lambda_i^k$ , and the series splits into $n$ independent scalar exponential series — so $e^{At}=\mathrm{diag}(e^{\lambda_1 t},\dots,e^{\lambda_n t})$ . Every other case (diagonalizable, complex eigenvalues, defective matrices) is obtained by conjugating this diagonal picture by a change of basis, or by adding polynomial correction terms when eigenvectors run out.

Figure

The exponential map takes an eigenvalue "spectrum" on the complex plane to a family of scalar exponentials

e^{\lambda t}

, one per eigenvalue

The theorem

e^{At}

e^{At}

A

Worked examples

Example 1: Distinct real eigenvalues, non-symmetric matrix. Let

A = \begin{pmatrix} 1 & 4 \\ 2 & 3\end{pmatrix}.

The characteristic polynomial is $\det(A-\lambda I) = (1-\lambda)(3-\lambda) - 8 = \lambda^2 - 4\lambda - 5 = (\lambda-5)(\lambda+1)$ , so the eigenvalues are $\lambda_1=5$ , $\lambda_2=-1$ . For $\lambda_1=5$ : solve $(A-5I)\mathbf{v}=0$ , i.e. $\begin{pmatrix}-4 & 4\\ 2 & -2\end{pmatrix}\mathbf{v}=0$ , giving $\mathbf{v}_1 = \binom{1}{1}$ . For $\lambda_2=-1$ : solve $(A+I)\mathbf{v}=0$ , i.e. $\begin{pmatrix}2 & 4 \\ 2 & 4\end{pmatrix}\mathbf{v}=0$ , giving $\mathbf{v}_2 = \binom{2}{-1}$ (choosing $v=2,-1$ to clear the fraction from $v_1=-2v_2$ ). So $P = \begin{pmatrix}1 & 2\\ 1 & -1\end{pmatrix}$ , $\det P = 1\cdot(-1) - 2\cdot 1 = -3$ , and

P^{-1} = \frac{1}{-3}\begin{pmatrix}-1 & -2 \\ -1 & 1\end{pmatrix} = \begin{pmatrix} 1/3 & 2/3 \\ 1/3 & -1/3\end{pmatrix}.

Check: $AP \stackrel{?}{=} PD$ . Column 1: $A\binom{1}{1} = \binom{1+4}{2+3}=\binom{5}{5} = 5\binom{1}{1}$ . ✓. Column 2: $A\binom{2}{-1} = \binom{2-4}{4-3} = \binom{-2}{1} = -1\cdot\binom{2}{-1}$ . ✓. By the corollary,

e^{At} = \begin{pmatrix}1 & 2\\ 1 & -1\end{pmatrix}\begin{pmatrix}e^{5t} & 0\\ 0 & e^{-t}\end{pmatrix}\begin{pmatrix}1/3 & 2/3\\ 1/3 & -1/3\end{pmatrix}.

Multiplying the first two factors:

\begin{pmatrix}e^{5t} & 2e^{-t}\\ e^{5t} & -e^{-t}\end{pmatrix},

and then multiplying by $P^{-1}$ entrywise:

e^{At} = \begin{pmatrix} \tfrac{1}{3}e^{5t} + \tfrac{2}{3}e^{-t} & \tfrac{2}{3}e^{5t} - \tfrac{2}{3}e^{-t} \\[4pt] \tfrac{1}{3}e^{5t} - \tfrac{1}{3}e^{-t} & \tfrac{2}{3}e^{5t} + \tfrac{1}{3}e^{-t} \end{pmatrix}.

As a check, at $t=0$ this gives $\begin{pmatrix}1/3+2/3 & 2/3-2/3\\ 1/3-1/3 & 2/3+1/3\end{pmatrix} = \begin{pmatrix}1&0\\0&1\end{pmatrix}=I$ , as required. Also differentiating the $(1,1)$ entry gives $\tfrac{5}{3}e^{5t} - \tfrac{2}{3}e^{-t}$ , and computing $(Ae^{At})_{11} = 1\cdot(\tfrac13e^{5t}+\tfrac23e^{-t}) + 4\cdot(\tfrac13e^{5t}-\tfrac13 e^{-t}) = \tfrac53 e^{5t} - \tfrac23 e^{-t}$ , matching. ✓

Example 2: A defective (non-diagonalizable) $3\times 3$ matrix with a repeated eigenvalue. Let

A = \begin{pmatrix} 2 & 1 & 0 \\ 0 & 2 & 1 \\ 0 & 0 & 2\end{pmatrix} + \begin{pmatrix}0&0&0\\0&0&0\\0&0&0\end{pmatrix},

i.e. $A$ is already a single $3\times 3$ Jordan block with eigenvalue $\lambda = 2$ :

A = \begin{pmatrix} 2 & 1 & 0\\ 0 & 2 & 1\\ 0 & 0 & 2\end{pmatrix}.

Write $A = 2I + N$ where

N = \begin{pmatrix}0&1&0\\0&0&1\\0&0&0\end{pmatrix}

is nilpotent: $N^2 = \begin{pmatrix}0&0&1\\0&0&0\\0&0&0\end{pmatrix}$ , and $N^3 = 0$ (direct computation: $N^2$ has a single nonzero entry in position $(1,3)$ , and $N^3 = N\cdot N^2$ shifts that entry off the matrix entirely, giving the zero matrix). Since $2I$ and $N$ commute (as $2I$ commutes with every matrix), Part 3 of the theorem gives

e^{At} = e^{2It}e^{Nt} = e^{2t} I \cdot e^{Nt}.

Because $N^3=0$ , the series for $e^{Nt}$ terminates after two nontrivial terms:

e^{Nt} = I + Nt + \frac{N^2t^2}{2} = \begin{pmatrix}1 & t & t^2/2\\ 0 & 1 & t \\ 0 & 0 & 1\end{pmatrix}.

Therefore

e^{At} = e^{2t}\begin{pmatrix}1 & t & t^2/2\\ 0 & 1 & t\\ 0 & 0 & 1\end{pmatrix} = \begin{pmatrix} e^{2t} & t e^{2t} & \tfrac{t^2}{2}e^{2t}\\ 0 & e^{2t} & te^{2t}\\ 0 & 0 & e^{2t}\end{pmatrix}.

This displays the classic signature of a repeated eigenvalue with a deficient eigenspace: the diagonal entries are $e^{2t}$ , but the superdiagonal structure of $N$ forces polynomial factors $t$ and $t^2/2$ to appear alongside $e^{2t}$ , exactly the terms familiar from the "repeated roots" method for scalar-reducible systems, now derived directly from the exponential series rather than assumed as an ansatz. Check at $t=0$ : the matrix is $\begin{pmatrix}1&0&0\\0&1&0\\0&0&1\end{pmatrix}=I$ . ✓ Differentiating the $(1,2)$ entry $te^{2t}$ gives $e^{2t}+2te^{2t}$ ; and $(Ae^{At})_{12} = 2\cdot(te^{2t}) + 1\cdot(e^{2t}) + 0 = e^{2t}+2te^{2t}$ , matching. ✓

Why we need the commutativity hypothesis

AB=BA

Scope, edge cases & common traps

Situation	What to use	Common pitfall
$A$ diagonalizable (distinct or repeated eigenvalues, full eigenspace)	$e^{At}=Pe^{Dt}P^{-1}$	Forgetting $P^{-1}$ must be the inverse, not the transpose, unless $P$ is orthogonal
$A$ has a complex conjugate eigenvalue pair $a\pm bi$	Diagonalize over $\mathbb C$ , or use real block $\begin{pmatrix}a&b\\-b&a\end{pmatrix}\to e^{at}\begin{pmatrix}\cos bt & \sin bt\\ -\sin bt & \cos bt\end{pmatrix}$	Mixing a complex eigenvector basis into a "real" answer without taking real/imaginary parts correctly
$A$ defective (fewer eigenvectors than eigenvalue multiplicity)	Jordan form: $A=2I+N$ decomposition, exploit nilpotency of $N$	Assuming $e^{At}$ is diagonal-looking; missing the polynomial-in- $t$ terms
Need $e^{A}e^{B}$ for two different matrices	Only equals $e^{A+B}$ if $AB=BA$	Silently assuming the group law for any two matrices
Need $e^{At}$ for large $t$ , only asymptotics wanted	Read off dominant eigenvalue's real part directly, skip full computation	Forgetting a repeated eigenvalue contributes $t^k e^{\lambda t}$ growth that can beat a nearby larger but non-repeated eigenvalue for small-to-moderate $t$

The mathematician behind it

Origins · 1 minGiuseppe Peano and how this was discovered

Giuseppe Peano

1858–1932 · Turin, Italy

Giuseppe Peano is best remembered for the Peano axioms of arithmetic and for Peano's existence theorem in ODEs, but he was also among the first, in an 1888 paper, to study systems of linear differential equations with variable and constant coefficients using explicit matrix-series methods, effectively constructing what we now call the fundamental matrix and, in the constant-coefficient case, the matrix exponential. His work sat at the crossroads of the emerging theory of linear algebra (itself still coalescing from the work of Cayley and Sylvester) and the classical theory of linear ODEs developed earlier by Cauchy and Lagrange.

History

- 1858: Giuseppe Peano born in Spinetta, near Cuneo, Italy.

- 1858: Arthur Cayley publishes A Memoir on the Theory of Matrices, laying groundwork for matrix algebra including the Cayley–Hamilton theorem.

- 1888: Peano publishes Intégration par séries des équations différentielles linéaires, giving series (Peano–Baker series) solutions to linear ODE systems, the direct ancestor of the constant-coefficient matrix exponential.

- Early 1900s: The Jordan canonical form, developed by Camille Jordan in the 1870s, is adopted as the standard tool for computing

e^{At}

when

A

is not diagonalizable.

- Mid-20th century: The matrix exponential becomes a standard chapter in linear systems theory, control theory (state transition matrices), and numerical linear algebra (Padé approximation methods for computing

e^{At}

efficiently).

Why does the matrix exponential series always converge, even though matrix multiplication doesn't commute?

Convergence is a question about the size of the terms $A^k t^k/k!$ , not about their algebraic order. Using any submultiplicative matrix norm, $\|A^k\| \le \|A\|^k$ , so the whole matrix series is term-by-term dominated in norm by the ordinary scalar series $\sum_k \|A\|^k|t|^k/k!$ , which converges to $e^{\|A\|\,|t|}$ for every real number $\|A\|\,|t|$ . This scalar bound forces the matrix series to converge absolutely, by the standard comparison test extended to normed vector spaces (here, the finite-dimensional and hence complete space of $n\times n$ matrices). Noncommutativity would only threaten identities that rearrange or factor the series, such as $e^{A+B}=e^Ae^B$ , never the bare fact of convergence.

e^{At}

the same as taking the exponential of each entry of

At

No, and this is one of the most common beginner mistakes. Entrywise exponentiation ignores the matrix products $A^2, A^3,\dots$ entirely and has no connection to differential equations. The matrix exponential $e^{At}$ genuinely mixes the entries of $A$ through repeated matrix multiplication, which is exactly what is needed so that $\frac{d}{dt}e^{At}=Ae^{At}$ holds. For a nondiagonal matrix the two notions differ dramatically — for instance, entrywise exponentiation of an off-diagonal $1$ leaves it $e^1=e\approx 2.718$ , whereas the true matrix exponential of a Jordan block produces a $t$ -dependent term $te^{\lambda t}$ in that slot, as shown in the worked example.

How is

e^{At}

actually computed by hand for exam-style problems?

There are two standard routes. If $A$ has $n$ linearly independent eigenvectors (i.e. $A$ is diagonalizable, guaranteed for example when all eigenvalues are distinct), diagonalize $A=PDP^{-1}$ , exponentiate the diagonal entries of $D$ scaled by $t$ , and conjugate back: $e^{At}=Pe^{Dt}P^{-1}$ . If $A$ is defective (a repeated eigenvalue with too few eigenvectors), pass to the Jordan form $A=PJP^{-1}$ , split each Jordan block as $\lambda I + N$ with $N$ nilpotent, and use $e^{(\lambda I+N)t}=e^{\lambda t}e^{Nt}$ where $e^{Nt}$ is a finite sum because $N^m=0$ for some $m\le$ block size. Complex conjugate eigenvalue pairs are handled by either working in $\mathbb C$ and taking real parts at the end, or using the real canonical block directly.

Why is

e^{At}

called "the" fundamental matrix rather than "a" fundamental matrix?

Any invertible matrix solution $\Psi(t)$ of $\Psi'=A\Psi$ is called a fundamental matrix, and there are infinitely many (any $\Psi(t)C$ for constant invertible $C$ is another one). What singles out $e^{At}$ is the normalization $e^{A\cdot 0}=I$ : it is the unique fundamental matrix equal to the identity at $t=0$ . This normalization is exactly what makes the formula $\mathbf{x}(t)=e^{At}\mathbf{x}_0$ solve the initial value problem directly, without any extra algebra to match initial conditions, which is why it earns the definite article in most treatments.

Does

e^{At}\to 0

t\to\infty

tell us the system is stable?

Yes, and this is one of the main payoffs of the theory. If every eigenvalue of $A$ has strictly negative real part, then in the Jordan form each block contributes terms of the shape $t^k e^{\lambda t}$ with $\mathrm{Re}(\lambda)<0$ , and since exponential decay always beats polynomial growth, $t^k e^{\lambda t}\to 0$ as $t\to\infty$ for every fixed $k$ . Since $e^{At}=Pe^{Jt}P^{-1}$ for fixed invertible $P$ , this forces $e^{At}\to 0$ entrywise, and hence every solution $\mathbf{x}(t)=e^{At}\mathbf{x}_0\to \mathbf{0}$ . This is precisely the asymptotic stability criterion used when reading off long-term behavior in Phase Portraits for Linear Systems.

What goes wrong if I try to define

e^{At}

only using eigenvalues, without eigenvectors or Jordan structure?

Eigenvalues alone determine the exponents $e^{\lambda t}$ that appear, but not how they combine to reconstruct the full matrix, nor whether extra polynomial factors of $t$ are needed. Two matrices can share exactly the same eigenvalues (even the same characteristic polynomial) yet have completely different exponentials if one is diagonalizable and the other is defective — compare $\begin{pmatrix}2&0\\0&2\end{pmatrix}$ (exponential $e^{2t}I$ ) against $\begin{pmatrix}2&1\\0&2\end{pmatrix}$ (exponential with a $te^{2t}$ term), both having the sole eigenvalue $2$ with algebraic multiplicity $2$ . The eigenvectors (or generalized eigenvectors via the Jordan form) are indispensable extra data.

How does this connect to complex or repeated eigenvalues covered elsewhere?

The matrix exponential is the unifying object behind the case-by-case recipes taught under Complex and Repeated Eigenvalues. A complex conjugate pair $a\pm bi$ produces, via diagonalization over $\mathbb{C}$ followed by taking real and imaginary parts, the familiar rotation-and-scaling block $e^{at}\begin{pmatrix}\cos bt & \sin bt \\ -\sin bt & \cos bt\end{pmatrix}$ seen in Systems of First-Order Linear ODEs. A repeated eigenvalue with a deficient eigenspace produces the polynomial-times-exponential terms $t^k e^{\lambda t}$ via the nilpotent part of the Jordan decomposition, exactly the "extra $t$ factor" ansatz used in the repeated-root method. The exponential explains why those ansätze work, rather than requiring them to be guessed.