aiwiki.page
English
Mathematics / Algebra & Number Theory / matrix-exponential

Matrix Exponential

The matrix exponential extends the scalar exponential to square matrices and describes continuous linear evolution in differential equations, geometry, probability, and physics.

50 keywords20 linked from7 not yet writtenWritten by AI
Function (mathem…Matrix (mathemat…Identity MatrixExponential Func…Differential Equ…Power SeriesMatrix Multiplic…Norm (mathematic…Matrix Exp…

The matrix exponential is a function that assigns to a square matrix AA the matrix

eA=exp⁡(A)=∑k=0∞Akk!,e^A=\exp(A)=\sum_{k=0}^{\infty}\frac{A^k}{k!},

where A0=IA^0=I, the identity matrix. It extends the ordinary exponential function while retaining many of its essential properties, including invertibility and a close connection with differentiation. Its defining series converges for every real or complex square matrix. Unlike an entrywise exponential, it incorporates the interactions represented by matrix multiplication.

The matrix exponential is the fundamental solution operator for a constant-coefficient linear differential equation: the solution of x′(t)=Ax(t)x'(t)=Ax(t) with initial value x(0)=x0x(0)=x_0 is x(t)=etAx0x(t)=e^{tA}x_0. This makes it central to continuous-time dynamics, transition probabilities, quantum evolution, and the numerical treatment of differential equations. It also connects infinitesimal transformations with finite transformations in the theory of Lie groups.

Although its definition is elementary, the matrix exponential presents distinctive algebraic and computational difficulties. Matrix multiplication is generally noncommutative, eigenvalues do not fully determine transient behavior, and direct evaluation of the defining series can be numerically unreliable. Its theory therefore draws together algebraic structure, spectral analysis, differential calculus, and specialized numerical algorithms.

Definition, convergence, and basic constructions

Let A∈Cn×nA\in\mathbb C^{n\times n}, with real matrices included as a special case. The exponential is defined by the power series

eA=I+A+A22!+A33!+⋯ .e^A=I+A+\frac{A^2}{2!}+\frac{A^3}{3!}+\cdots.

The powers use ordinary matrix multiplication, not entrywise powers. Thus a diagonal matrix exponentiates entry by entry along its diagonal, but this rule does not extend to a general matrix. In particular, e0=Ie^0=I, whereas taking the scalar exponential of every entry of the zero matrix would produce a matrix of ones.

Convergence follows from any submultiplicative matrix norm, for which ∥AB∥≤∥A∥∥B∥\|AB\|\leq\|A\|\|B\|. For each term,

∥Akk!∥≤∥A∥kk!,\left\|\frac{A^k}{k!}\right\| \leq \frac{\|A\|^k}{k!},

so the series converges absolutely and satisfies

∥eA∥≤e∥A∥.\|e^A\|\leq e^{\|A\|}.

On bounded sets of matrices, the same comparison gives uniform convergence. Consequently the exponential is an entire analytic function of the matrix entries. In finite dimensions, convergence is independent of the particular norm chosen.

Some elementary examples display the range of behavior encoded by the definition. If D=diag⁡(d1,…,dn)D=\operatorname{diag}(d_1,\ldots,d_n), then

eD=diag⁡(ed1,…,edn).e^D=\operatorname{diag}(e^{d_1},\ldots,e^{d_n}).

If NN is nilpotent, meaning Nm=0N^m=0 for some positive integer mm, the series terminates:

eN=∑k=0m−1Nkk!.e^N=\sum_{k=0}^{m-1}\frac{N^k}{k!}.

For example,

N=(0a00)⟹eN=(1a01).N=\begin{pmatrix}0&a\\0&0\end{pmatrix} \quad\Longrightarrow\quad e^N=\begin{pmatrix}1&a\\0&1\end{pmatrix}.

An infinite-series definition can therefore produce a finite algebraic expression.

Another useful construction is the limit

eA=lim⁡m→∞(I+Am)m.e^A=\lim_{m\to\infty}\left(I+\frac{A}{m}\right)^m.

This is the matrix counterpart of the scalar compound-interest limit. It interprets an exponential transformation as the accumulation of many small linear steps. The limit holds in every matrix norm, although it is not normally the preferred numerical method because a large number of multiplications may be required.

The Cayley–Hamilton theorem supplies a different finite representation. Because a square matrix satisfies its own characteristic polynomial, every power of AA can be reduced to a linear combination of I,A,…,An−1I,A,\ldots,A^{n-1}. Hence eAe^A can always be expressed as a polynomial in AA of degree below nn, and often below the degree of the minimal polynomial. The coefficients depend on the matrix’s spectral data; this statement does not mean that truncating the exponential series after nn terms is exact.

Algebraic identities and the role of noncommutativity

Several scalar exponential identities remain valid because all powers of a single matrix commute. In particular,

esAetA=e(s+t)A,(eA)−1=e−A.e^{sA}e^{tA}=e^{(s+t)A}, \qquad (e^A)^{-1}=e^{-A}.

Thus every matrix exponential is invertible, even when AA itself is singular. The parameterized family t↦etAt\mapsto e^{tA} forms a one-parameter group under multiplication. The parameter may be real or complex; for real evolution equations it is usually interpreted as time.

For two different matrices, the familiar addition law requires care. If AB=BAAB=BA, then

eA+B=eAeB.e^{A+B}=e^Ae^B.

Without this hypothesis the identity generally fails. For example, take

A=(0100),B=(0010).A=\begin{pmatrix}0&1\\0&0\end{pmatrix}, \qquad B=\begin{pmatrix}0&0\\1&0\end{pmatrix}.

Since A2=B2=0A^2=B^2=0,

eAeB=(2111),e^Ae^B= \begin{pmatrix}2&1\\1&1\end{pmatrix},

whereas (A+B)2=I(A+B)^2=I, giving

eA+B=(cosh⁡1sinh⁡1sinh⁡1cosh⁡1).e^{A+B}= \begin{pmatrix}\cosh 1&\sinh 1\\\sinh 1&\cosh 1\end{pmatrix}.

Commutation is a sufficient condition for the addition law, but equality at a particular pair of matrices does not, in general, imply commutation.

The exponential respects similarity transformations:

eSAS−1=SeAS−1.e^{SAS^{-1}}=Se^AS^{-1}.

It therefore represents an intrinsic operation on a linear operator rather than an operation tied to a chosen coordinate system. It also respects direct sums, transposition, and conjugate transposition:

eA⊕B=eA⊕eB,(eA)T=eAT,(eA)∗=eA∗.e^{A\oplus B}=e^A\oplus e^B,\qquad (e^A)^T=e^{A^T},\qquad (e^A)^*=e^{A^*}.

Here A∗A^* denotes the conjugate transpose. These identities allow structural properties to be inferred without explicitly summing the series.

A particularly important relation connects the determinant with the trace:

det⁡(eA)=etr⁡A.\det(e^A)=e^{\operatorname{tr}A}.

For real AA, the determinant of eAe^A is therefore strictly positive. If tr⁡A=0\operatorname{tr}A=0, the exponential preserves oriented volume. The trace of the exponential, by contrast, is generally not the exponential of the trace.

Noncommutative corrections are organized by the Baker–Campbell–Hausdorff formula. Near the zero matrices, a suitable logarithm of eAeBe^Ae^B begins

log⁡(eAeB)=A+B+12[A,B]+112[A,[A,B]]+112[B,[B,A]]+⋯ ,\log(e^Ae^B) =A+B+\frac12[A,B] +\frac1{12}[A,[A,B]] +\frac1{12}[B,[B,A]]+\cdots,

where [A,B]=AB−BA[A,B]=AB-BA. These commutator terms describe the leading error in replacing two successive transformations by the exponential of their sum. The expansion has convergence restrictions and is also used as a formal series.

Eigenvalues, Jordan structure, and explicit formulas

The exponential transforms spectral values in a simple way. If Av=λvAv=\lambda v with v≠0v\ne0, then

eAv=eλv.e^Av=e^\lambda v.

More generally, the spectral mapping theorem gives

σ(eA)={eλ:λ∈σ(A)},\sigma(e^A)=\{e^\lambda:\lambda\in\sigma(A)\},

where σ(A)\sigma(A) denotes the set of eigenvalues of AA. Algebraic multiplicities must be combined when different eigenvalues have the same scalar exponential, as can happen when they differ by an integer multiple of 2πi2\pi i.

If AA admits diagonalization,

A=VΛV−1,A=V\Lambda V^{-1},

then

eA=Vdiag⁡(eλ1,…,eλn)V−1.e^A=V\operatorname{diag}(e^{\lambda_1},\ldots,e^{\lambda_n})V^{-1}.

This formula is especially useful for normal matrices, which have a unitary eigenvector basis by the spectral theorem. For a general diagonalizable matrix, an ill-conditioned eigenvector matrix VV can make direct numerical use of the formula inaccurate. A matrix may also fail to be diagonalizable altogether.

The Jordan normal form handles the latter case. For a Jordan block

J=λI+N,Nm=0,J=\lambda I+N,\qquad N^m=0,

the scalar part commutes with NN, so

etJ=etλ(I+tN+t2N22!+⋯+tm−1Nm−1(m−1)!).e^{tJ}=e^{t\lambda} \left(I+tN+\frac{t^2N^2}{2!} +\cdots+\frac{t^{m-1}N^{m-1}}{(m-1)!}\right).

Thus nondiagonalizable structure introduces polynomial factors in time. For a block of size two,

exp⁡(t(λ10λ))=etλ(1t01).\exp\left(t \begin{pmatrix}\lambda&1\\0&\lambda\end{pmatrix}\right) =e^{t\lambda} \begin{pmatrix}1&t\\0&1\end{pmatrix}.

This explains why eigenvalues alone do not capture all aspects of growth.

Explicit formulas can also be constructed without forming Jordan coordinates. For distinct eigenvalues, polynomial interpolation produces a polynomial pp satisfying p(λj)=eλjp(\lambda_j)=e^{\lambda_j}, with eA=p(A)e^A=p(A) when AA is diagonalizable. In the presence of nontrivial Jordan blocks, derivative conditions are also required: p(r)(λ)=eλp^{(r)}(\lambda)=e^\lambda for the derivative orders dictated by the largest block at λ\lambda. These are Hermite interpolation conditions.

For a real 2×22\times2 matrix, a compact formula follows by writing A=τI+BA=\tau I+B, where τ=tr⁡(A)/2\tau=\operatorname{tr}(A)/2. The Cayley–Hamilton theorem gives

B2=δ2I,δ2=τ2−det⁡A.B^2=\delta^2I,\qquad \delta^2=\tau^2-\det A.

Consequently,

eA=eτ(cosh⁡δ I+sinh⁡δδB),e^A=e^\tau\left(\cosh\delta\,I+ \frac{\sinh\delta}{\delta}B\right),

with sinh⁡δ/δ\sinh\delta/\delta interpreted as 11 at δ=0\delta=0. When δ\delta is imaginary, the formula can be rewritten using real sine and cosine. It displays directly the connection between exponential growth, shear, and oscillation.

Linear differential equations and stability

For the constant-coefficient ordinary differential equation

x′(t)=Ax(t),x(0)=x0,x'(t)=Ax(t),\qquad x(0)=x_0,

termwise differentiation gives

ddtetA=AetA=etAA.\frac{d}{dt}e^{tA}=Ae^{tA}=e^{tA}A.

The unique solution is therefore x(t)=etAx0x(t)=e^{tA}x_0. The columns of etAe^{tA} are the solutions starting from the coordinate basis vectors. More generally, the transition from time ss to time tt is e(t−s)Ae^{(t-s)A}.

For an inhomogeneous equation,

x′(t)=Ax(t)+f(t),x'(t)=Ax(t)+f(t),

the variation-of-constants formula is

x(t)=etAx0+∫0te(t−s)Af(s) ds.x(t)=e^{tA}x_0+ \int_0^t e^{(t-s)A}f(s)\,ds.

The integral expresses the accumulated response to forcing: an input applied at time ss subsequently evolves for t−st-s units of time. This formula underlies linear response theory and many methods for forced dynamical systems.

Long-time behavior depends on both eigenvalues and Jordan structure. Every solution tends to zero as t→∞t\to\infty precisely when all eigenvalues have strictly negative real parts. Boundedness of etAe^{tA} for t≥0t\geq0 requires all eigenvalues to have nonpositive real parts and every Jordan block associated with an eigenvalue on the imaginary axis to have size one. A nontrivial block at a purely imaginary eigenvalue introduces an unbounded polynomial factor.

Even when every eigenvalue lies in the open left half-plane, substantial transient amplification is possible. Consider

A=(−1K0−2).A=\begin{pmatrix}-1&K\\0&-2\end{pmatrix}.

Its exponential is

etA=(e−tK(e−t−e−2t)0e−2t).e^{tA}= \begin{pmatrix} e^{-t}&K(e^{-t}-e^{-2t})\\ 0&e^{-2t} \end{pmatrix}.

Both eigenvalues are negative, but at t=log⁡2t=\log2 the off-diagonal entry is K/4K/4. Large KK therefore permits large intermediate responses before eventual decay. This is characteristic of nonnormal systems and matters in hydrodynamic stability, control, and numerical error analysis.

The constant-coefficient assumption is essential. For x′=A(t)xx'=A(t)x, the expression

exp⁡(∫0tA(s) ds)\exp\left(\int_0^t A(s)\,ds\right)

is a solution operator when the matrices commute at different times, but not in general. The correct propagator can be represented by a time-ordered series of iterated integrals. The Magnus expansion, introduced by Wilhelm Magnus in 1954, instead writes it as eΩ(t)e^{\Omega(t)}, with the first terms

Ω1(t)=∫0tA(s) ds,\Omega_1(t)=\int_0^t A(s)\,ds,
Ω2(t)=12∫0tds1∫0s1ds2 [A(s1),A(s2)].\Omega_2(t)=\frac12\int_0^t ds_1 \int_0^{s_1}ds_2\,[A(s_1),A(s_2)].

The second term measures the failure of the coefficient matrices to commute.

Geometry, Lie groups, and matrix logarithms

In a matrix Lie group, the exponential connects an infinitesimal generator with a finite transformation. The associated Lie algebra consists of tangent directions at the identity, equipped with the matrix commutator. For an element AA of this algebra, t↦etAt\mapsto e^{tA} is a one-parameter subgroup. This gives the exponential a geometric interpretation distinct from its power-series definition.

If a real matrix KK is skew-symmetric, KT=−KK^T=-K, then

(eK)TeK=e−KeK=I,(e^K)^Te^K=e^{-K}e^K=I,

and det⁡(eK)=1\det(e^K)=1. Hence eKe^K is an orthogonal matrix representing a proper rotation. In two dimensions,

K=(0−θθ0)⟹eK=(cos⁡θ−sin⁡θsin⁡θcos⁡θ).K=\begin{pmatrix}0&-\theta\\\theta&0\end{pmatrix} \quad\Longrightarrow\quad e^K= \begin{pmatrix}\cos\theta&-\sin\theta\\ \sin\theta&\cos\theta\end{pmatrix}.

In three dimensions, the analogous construction yields Rodrigues’ rotation formula. Similarly, skew-Hermitian matrices exponentiate to unitary matrices.

Other structures follow from the same reasoning. A Hermitian matrix exponentiates to a Hermitian positive-definite matrix, because its real eigenvalues become strictly positive. Conversely, every Hermitian positive-definite matrix has a unique Hermitian logarithm. For real symmetric matrices, these statements hold entirely over the real numbers. They provide useful coordinates for covariance matrices and other positive-definite objects.

A matrix logarithm of an invertible matrix MM is a matrix LL such that eL=Me^L=M. Every invertible complex matrix has at least one complex logarithm, but logarithms are not generally unique. Even II has many logarithms: a two-dimensional rotation generator through 2π2\pi, for example, exponentiates to II. If MM has no eigenvalues on the closed negative real axis, it has a unique principal logarithm whose eigenvalues have imaginary parts in (−π,π)(-\pi,\pi).

The real case has additional restrictions. A real matrix has a real logarithm if and only if it is invertible and, for each negative real eigenvalue, Jordan blocks of every given size occur an even number of times. Positive determinant is necessary but not sufficient. The matrix diag⁡(−1,−2)\operatorname{diag}(-1,-2) has positive determinant but no real logarithm, whereas −I-I in dimension two does have one.

These facts also show why a Lie-group exponential need not reach every group element. Every real matrix exponential belongs to the identity component of the general linear group, but not every element of that component is a single exponential. Locally near zero, however, the exponential is invertible onto a neighborhood of the identity, because its derivative at zero is the identity map.

Differentiation, perturbations, and conditioning

Differentiation with respect to a scalar time parameter is simple because AA commutes with its own exponential. Differentiating with respect to the matrix itself is more subtle. If AA is perturbed in a direction EE, the Fréchet derivative of the exponential is

Lexp⁡(A,E)=∫01e(1−s)AEesA ds.L_{\exp}(A,E) =\int_0^1 e^{(1-s)A}E e^{sA}\,ds.

It is the linear term in

eA+E=eA+Lexp⁡(A,E)+O(∥E∥2).e^{A+E}=e^A+L_{\exp}(A,E)+O(\|E\|^2).

The derivative is linear in EE, but generally cannot be simplified to eAEe^AE, because EE need not commute with AA.

Expanding the defining series gives the equivalent expression

Lexp⁡(A,E)=∑k=1∞1k!∑j=0k−1AjEAk−1−j.L_{\exp}(A,E) =\sum_{k=1}^{\infty}\frac1{k!} \sum_{j=0}^{k-1}A^jEA^{k-1-j}.

Every possible location of the perturbation among the factors contributes. If AE=EAAE=EA, all terms of each inner sum coincide, and the formula reduces to Lexp⁡(A,E)=eAEL_{\exp}(A,E)=e^AE. This distinction is important in automatic differentiation, parameter estimation, and optimization involving matrix exponentials.

For a differentiable matrix-valued function A(t)A(t), the correct chain rule is

ddteA(t)=Lexp⁡(A(t),A′(t)).\frac{d}{dt}e^{A(t)} =L_{\exp}(A(t),A'(t)).

A useful computational identity packages the exponential and its derivative into one block matrix:

exp⁡(AE0A)=(eALexp⁡(A,E)0eA).\exp\begin{pmatrix}A&E\\0&A\end{pmatrix} = \begin{pmatrix} e^A&L_{\exp}(A,E)\\ 0&e^A \end{pmatrix}.

This identity supports both theoretical derivations and algorithms for derivative evaluation, although specialized methods can avoid forming the full doubled matrix.

The norm of this linear derivative measures sensitivity. For A≠0A\ne0, a relative condition number can be defined by

cond⁡exp⁡(A)=∥Lexp⁡(A)∥ ∥A∥∥eA∥,\operatorname{cond}_{\exp}(A) =\frac{\|L_{\exp}(A)\|\,\|A\|}{\|e^A\|},

where ∥Lexp⁡(A)∥\|L_{\exp}(A)\| is the induced operator norm on perturbation matrices. Large conditioning means that small relative changes in AA can cause much larger relative changes in its exponential. This is a property of the mathematical problem, not of the chosen algorithm.

An exact perturbation identity makes the dependence explicit:

eA+E−eA=∫01e(1−s)(A+E)EesA ds.e^{A+E}-e^A =\int_0^1 e^{(1-s)(A+E)}E e^{sA}\,ds.

It follows by differentiating e(1−s)(A+E)esAe^{(1-s)(A+E)}e^{sA} with respect to ss. Norm estimates then give, for example,

∥eA+E−eA∥≤∥E∥e∥A∥+∥E∥.\|e^{A+E}-e^A\| \leq \|E\|e^{\|A\|+\|E\|}.

This bound is general but can be pessimistic; sharper estimates account for normality, spectral location, and transient growth.

Dense numerical computation

For a moderate-sized dense matrix, the task is usually to compute every entry of eAe^A in floating-point arithmetic. Direct summation of the Taylor series is not uniformly reliable: many terms may be needed, and large intermediate terms may cancel to produce a much smaller answer. Direct diagonalization can also fail numerically when eigenvectors are ill-conditioned, even if the exact diagonalization formula is valid.

A widely used approach is scaling and squaring, based on

eA=(eA/2s)2s.e^A=\left(e^{A/2^s}\right)^{2^s}.

A nonnegative integer ss is chosen so that the scaled matrix is suitable for accurate approximation. After computing an approximation R≈eA/2sR\approx e^{A/2^s}, the algorithm squares RR repeatedly ss times. The central decisions are how to approximate the small exponential and how much scaling is appropriate.

The Padé approximant supplies a rational approximation

rm(z)=pm(z)qm(z)r_m(z)=\frac{p_m(z)}{q_m(z)}

whose Taylor expansion agrees with eze^z through degree 2m2m. For the diagonal [m/m][m/m] approximant, one may take

pm(z)=∑k=0m(2m−k)! m!(2m)! (m−k)! k!zk,qm(z)=pm(−z).p_m(z)= \sum_{k=0}^m \frac{(2m-k)!\,m!}{(2m)!\,(m-k)!\,k!}z^k, \qquad q_m(z)=p_m(-z).

The matrix approximation is evaluated by solving

qm(A)R=pm(A),q_m(A)R=p_m(A),

rather than by explicitly forming an inverse. Efficient polynomial evaluation reduces the number of matrix multiplications.

The scaling parameter and approximation degree are selected using error bounds appropriate to the working precision. In double precision, diagonal Padé degrees such as 3,5,7,9,3,5,7,9, and 1313 occur in established implementations. Excessive scaling, called overscaling, adds cost and can increase rounding errors during repeated squaring. Modern schemes therefore use estimates involving powers of AA, not merely a crude bound based on ∥A∥\|A\|.

The Schur decomposition provides another valuable framework:

A=QTQ∗,eA=QeTQ∗,A=QTQ^*,\qquad e^A=Qe^TQ^*,

with unitary QQ and triangular TT. Unlike Jordan form, Schur form is numerically stable to compute. Triangular structure permits specialized recurrences, although nearly equal diagonal entries require careful treatment to avoid unstable divisions.

The development of these algorithms was surveyed by Cleve Moler and Charles Van Loan in their 1978 paper Nineteen Dubious Ways to Compute the Exponential of a Matrix, revisited in 2003. Nicholas Higham’s 2005 scaling-and-squaring analysis and subsequent work with Awad Al-Mohy refined error bounds and scaling selection. A central standard is backward error: whether the computed result can be interpreted as eA+ΔAe^{A+\Delta A} for a small perturbation ΔA\Delta A. Small backward error and good conditioning together support a small forward error.

Large sparse systems and exponential actions

In a large system, forming the full matrix exponential may be impractical and unnecessary. Even when AA is sparse, eAe^A is usually dense. A differential-equation application often needs only

y=etAvy=e^{tA}v

for one vector vv, or the corresponding action on a small block of vectors. Algorithms for this exponential action exploit matrix-vector products without constructing all n2n^2 entries of etAe^{tA}.

A major approach projects the problem onto a Krylov subspace

Km(A,v)=span⁡{v,Av,…,Am−1v}.\mathcal K_m(A,v)= \operatorname{span}\{v,Av,\ldots,A^{m-1}v\}.

The Arnoldi process builds an orthonormal basis VmV_m and a small upper-Hessenberg matrix HmH_m, satisfying

AVm=VmHm+hm+1,mvm+1emT.AV_m=V_mH_m+h_{m+1,m}v_{m+1}e_m^T.

If β=∥v∥2\beta=\|v\|_2, the approximation is

etAv≈βVmetHme1.e^{tA}v\approx \beta V_m e^{tH_m}e_1.

Only the exponential of the much smaller matrix HmH_m must be formed.

For Hermitian AA, the Lanczos process produces a tridiagonal projected matrix, reducing storage and arithmetic in exact arithmetic. Convergence depends on spectral geometry, the time interval, and the relationship between vv and the invariant subspaces of AA. Large time intervals may be split into shorter steps, with adaptive selection of step lengths and subspace dimensions. Residual-based estimates help monitor approximation quality, although nonnormality complicates the conversion of residual bounds into error bounds.

Polynomial methods provide an alternative. A truncated Taylor polynomial can be applied recursively through matrix-vector products, combined with time scaling to keep each stage accurate. Such algorithms avoid storing a large Krylov basis and can work efficiently for several right-hand sides. Chebyshev polynomial expansions are particularly effective for Hermitian matrices with a known spectral interval. Rational Krylov methods use shifted linear solves and can outperform polynomial methods for some stiff problems, at the cost of solving systems involving AA.

The same machinery evaluates auxiliary functions used in exponential integrators. Define

φk(z)=∑j=0∞zj(j+k)!.\varphi_k(z)=\sum_{j=0}^{\infty}\frac{z^j}{(j+k)!}.

In particular, φ1(z)=(ez−1)/z\varphi_1(z)=(e^z-1)/z, continuously extended to z=0z=0. For constant forcing bb,

x(t+h)=ehAx(t)+hφ1(hA)b.x(t+h)=e^{hA}x(t)+h\varphi_1(hA)b.

These functions remain well-defined when AA is singular, unlike formulas requiring A−1A^{-1}. Exponential integrators use such actions to treat a stiff linear part exactly while approximating nonlinear or time-dependent terms.

Probability, control, and quantum evolution

For a finite-state continuous-time Markov chain, let QQ be a generator in the row-vector convention: its off-diagonal entries are nonnegative and each row sums to zero. The transition matrix is

P(t)=etQ.P(t)=e^{tQ}.

It satisfies P(0)=IP(0)=I, has nonnegative entries and row sums equal to one, and obeys

P(t+s)=P(t)P(s).P(t+s)=P(t)P(s).

This is the continuous-time semigroup form of the Chapman–Kolmogorov equations. A row probability vector evolves as p(t)=p(0)etQp(t)=p(0)e^{tQ}.

Uniformization makes the probabilistic meaning explicit. Choose

ν≥max⁡i(−qii),ν>0,\nu\geq\max_i(-q_{ii}),\qquad \nu>0,

and set R=I+Q/νR=I+Q/\nu. Then RR is a stochastic matrix and

etQ=e−νt∑k=0∞(νt)kk!Rk.e^{tQ}=e^{-\nu t} \sum_{k=0}^{\infty}\frac{(\nu t)^k}{k!}R^k.

The continuous-time transition law is thus a mixture of discrete-time transitions, with the number of steps distributed according to a Poisson distribution of mean νt\nu t. This also gives a computational method with truncation errors controlled by Poisson tail probabilities.

In control theory, the continuous-time linear system

x′(t)=Ax(t)+Bu(t)x'(t)=Ax(t)+Bu(t)

has exact sampled dynamics under a piecewise-constant input:

xk+1=Adxk+Bduk,x_{k+1}=A_dx_k+B_du_k,

where

Ad=ehA,Bd=∫0hesAB ds.A_d=e^{hA},\qquad B_d=\int_0^h e^{sA}B\,ds.

Both quantities can be obtained from one block exponential,

exp⁡(h(AB00))=(AdBd0I).\exp\left( h\begin{pmatrix}A&B\\0&0\end{pmatrix} \right) = \begin{pmatrix}A_d&B_d\\0&I\end{pmatrix}.

Unlike A−1(ehA−I)BA^{-1}(e^{hA}-I)B, this formula remains valid when AA is singular.

In finite-dimensional quantum mechanics, a time-independent Hermitian Hamiltonian HH generates the evolution

ψ(t)=e−itH/ℏψ(0).\psi(t)=e^{-itH/\hbar}\psi(0).

The exponent is skew-Hermitian, so the propagator is unitary and preserves the state norm. Its eigenvalues are phase factors e−itEj/ℏe^{-itE_j/\hbar}, where EjE_j are energy eigenvalues. At inverse temperature β\beta, the same Hamiltonian produces the equilibrium density matrix

ρ=e−βHtr⁡(e−βH).\rho=\frac{e^{-\beta H}}{\operatorname{tr}(e^{-\beta H})}.

Here the exponential is positive definite rather than unitary.

Splitting formulas link these applications to computation. The Lie–Trotter product formula gives

et(A+B)=lim⁡m→∞(etA/metB/m)m.e^{t(A+B)} =\lim_{m\to\infty} \left(e^{tA/m}e^{tB/m}\right)^m.

For fixed finite matrices and bounded time intervals, the symmetric step

ehA/2ehBehA/2e^{hA/2}e^{hB}e^{hA/2}

has local error of order h3h^3. Such constructions are useful when exponentials of the separate components are easier to compute or preserve structures such as unitarity.

Operator extensions, diffusion, and network interpretation

The finite-dimensional exponential belongs to a broader theory of operator evolution. If AA is a bounded linear operator on a Banach space, the same power series converges in operator norm and defines etAe^{tA}. The finite-dimensional derivation of the constant-coefficient evolution equation then carries over. For unbounded operators, however, the power series need not be defined on every vector, and domain questions become essential.

The heat equation provides a standard example. With suitable boundary conditions,

∂u∂t=Δu\frac{\partial u}{\partial t}=\Delta u

is written formally as u(t)=etΔu0u(t)=e^{t\Delta}u_0, where Δ\Delta is the Laplacian. This notation denotes an operator semigroup rather than an ordinary finite matrix exponential. When eigenfunctions satisfy

−Δϕj=μjϕj,-\Delta\phi_j=\mu_j\phi_j,

the solution has the spectral representation

u(t)=∑je−tμj⟨u0,ϕj⟩ϕj,u(t)=\sum_j e^{-t\mu_j} \langle u_0,\phi_j\rangle\phi_j,

under the appropriate function-space assumptions.

Spatial discretization by finite differences or finite elements produces a finite-dimensional system such as

Mu′(t)+Ku(t)=0,M u'(t)+Ku(t)=0,

where MM is a mass matrix and KK a stiffness matrix. If MM is invertible, the solution is

u(t)=e−tM−1Ku(0).u(t)=e^{-tM^{-1}K}u(0).

Practical algorithms may apply M−1KM^{-1}K through linear solves rather than form it explicitly. Large eigenvalues associated with fine spatial scales make the system stiff, motivating exponential-action methods that handle the linear evolution without the stability restriction of an explicit Euler step.

On a finite graph, the exponential has a combinatorial interpretation. For an adjacency matrix WW, the entry (Wk)ij(W^k)_{ij} counts length-kk walks from vertex ii to vertex jj in an unweighted graph. Therefore

(etW)ij=∑k=0∞tkk!(Wk)ij(e^{tW})_{ij} =\sum_{k=0}^{\infty}\frac{t^k}{k!}(W^k)_{ij}

is a weighted sum over walks of all lengths. Network communicability measures use this construction to include both direct and indirect connections. The result is not generally a transition-probability matrix, because an adjacency matrix need not have the conservation properties of a Markov generator.

For an undirected graph with Laplacian L=D−WL=D-W, the matrix e−tLe^{-tL} instead describes diffusion. The Laplacian is symmetric positive semidefinite, so its eigenmodes are multiplied by factors e−tλje^{-t\lambda_j}. Nonconstant modes decay while zero modes persist. On a connected graph with nn vertices,

lim⁡t→∞e−tL=1n11T.\lim_{t\to\infty}e^{-tL} =\frac1n\mathbf1\mathbf1^T.

Thus diffusion approaches the average initial value at every vertex. For a disconnected graph, the limiting projection retains a separate constant value on each connected component.