Computing e^{tA}: diagonalization, Jordan blocks and linear systems
Compute the matrix exponential through similarity, diagonalization and Jordan form, and use it to solve linear systems of differential equations, including the variation-of-constants formula.
§6.2Estimated reading time: 13 min
The previous unit on matrix exponentials defined etA=n≥0∑tnAn/n!, showed that it solves x′=Ax and proved eA+B=eAeB when AB=BA. This unit turns those facts into a method. The idea is that the exponential is easy for diagonal matrices and for matrices of the form scalar plus nilpotent, and the similarity transformations of Jordan form and diagonalization carry every matrix to such a form. The result is an explicit solution of every constant-coefficient linear system, and a clear description of its long-time behavior in terms of eigenvalues. All matrices are square and complex or real.
Similarity and the exponential
Proposition 6.2.1
(Similarity rule for the exponential)
If P is invertible then etPAP−1=PetAP−1.
Proof
(Canceling P−1P in every power)
(PAP−1)n=PAnP−1 by cancellation of P−1P. Hence each partial sum of the series for etPAP−1 equals P times the corresponding partial sum for etA times P−1, and multiplication by fixed matrices is continuous, so the limits agree.
Proposition 6.2.2
(Exponential of a diagonal matrix)
If D=diag(λ1,…,λn) then etD=diag(eλ1t,…,eλnt).
Proof
(The series acts entrywise on the diagonal)
Dk=diag(λ1k,…,λnk), so the series acts entrywise on the diagonal and gives the scalar exponential series in each entry.
Consequently, if A=PDP−1 is diagonalizable, then
etA=Pdiag(eλ1t,…,eλnt)P−1,
where the columns of P are eigenvectors for λ1,…,λn.
Proposition 6.2.3
(Scalar plus nilpotent)
Let A=λI+N with Nm=0. Then
etA=eλtk=0∑m−1k!tkNk.
Proof
(Splitting off the commuting scalar part)
λI commutes with N, so etA=etλIetN=eλtI⋅etN by the commuting product rule. The series for etN stops at k=m−1 because Nk=0 for k≥m.
In particular, for a Jordan block Jm(λ)=λI+N of size m the exponential has (i,j) entry eλttj−i/(j−i)! for j≥i and 0 below the diagonal. For a general matrix with Jordan decomposition A=PJP−1, J=diag(Jm1(λ1),…), we get etA=PetJP−1 with etJ block diagonal. Every entry of etA is therefore a linear combination of functions tkeλt, with λ an eigenvalue and k less than the size of the largest Jordan block for λ.
Theorem 6.2.4
(Determinant of the exponential)
detetA=ettrA; in particular etA is always invertible.
Proof
(Triangularize and read off the diagonal)
Write A=PJP−1 with J upper triangular (Jordan form). Powers of an upper triangular matrix are upper triangular with diagonal entries λik, so etJ is upper triangular with diagonal eλit. By the similarity rule, detetA=detetJ=i∏eλit=et∑λi=ettrA.
For the matrix of the first worked example below the trace is 2, and indeed the determinant of the matrix found there is 41((e3t+e−t)2−(e3t−e−t)2)=e3te−t=e2t.
Linear systems
Theorem 6.2.5
(Solutions of the homogeneous system)
For A∈Mn, x0∈Cn, the initial value problem x′=Ax, x(0)=x0 has the unique solution x(t)=etAx0. The solution space of x′=Ax is an n-dimensional vector space, with basis etAe1,…,etAen.
Proof
(Existence by differentiation, uniqueness by freezing e−tAy)
Existence follows from dtdetA=AetA. For uniqueness, if y is another solution, then dtd(e−tAy(t))=−Ae−tAy+e−tAAy=0 because A commutes with e−tA; hence e−tAy(t)=y(0)=x0 and y(t)=etAx0, using etAe−tA=I. The map x0↦etAx0 is a linear bijection from Cn to the solution space.
Proof X-ray
Why $e^{-tA}y$ is the right quantity to differentiate
The proof does not try to compare y with etAx0 directly. It multiplies by e−tA, which is the one factor that cancels the growth of every solution at once. The derivative has two terms: one from differentiating e−tA and one from differentiating y. They cancel only because A and e−tA commute, which holds since e−tA is a limit of polynomials in A; for a non-constant coefficient matrix A(t) this step fails. A function with zero derivative on an interval is constant (apply the one-variable fact to each coordinate), and the value at t=0 fixes the constant. Invertibility of etA, from the determinant formula, is what allows us to multiply back. Existence needs only the termwise differentiation of the series; uniqueness is where the cancellation is spent.
Theorem 6.2.6
(Variation of constants)
For continuous b:R→Cn the problem x′=Ax+b(t), x(0)=x0 has the unique solution
x(t)=etAx0+∫0te(t−s)Ab(s)ds.
Proof
(Differentiate the formula and use uniqueness for the difference)
Write x(t)=etAx0+etA∫0te−sAb(s)ds. By the product rule and the fundamental theorem of calculus, x′=AetAx0+AetA∫0te−sAbds+etAe−tAb(t)=Ax(t)+b(t), and x(0)=x0. If x1,x2 both solve the problem, x1−x2 solves the homogeneous problem with zero initial value, so is 0 by the previous theorem.
Long-time behavior. Because the entries of etA are combinations of tkeλt, if every eigenvalue satisfies Reλ<0 then etA→0 and every solution tends to 0 as t→∞, since tkeλt→0. If some eigenvalue has Reλ>0 then etAx0 is unbounded for suitable x0: take x0 an eigenvector for λ, giving etAx0=eλtx0 of modulus e(Reλ)t∣x0∣.
Worked examples
Worked example 6.2.7
(A diagonalizable matrix)
Let A=(1221). The eigenvalues are 3 with eigenvector (1,1) and −1 with eigenvector (1,−1). Take P=(111−1), P−1=21(111−1). Then
Check: at t=0 this is I, and its derivative at 0 is 21(2442)=A. For x0=(1,0), x(t)=(21(e3t+e−t),21(e3t−e−t)).
Worked example 6.2.8
(A repeated eigenvalue)
Solve y′′−4y′+4y=0, y(0)=1, y′(0)=0. Put x=(y,y′), so x′=Ax with A=(0−414). The characteristic polynomial is λ2−4λ+4=(λ−2)2. With N=A−2I=(−2−412), N2=(4−48−8−2+2−4+4)=0. By the scalar-plus-nilpotent formula,
etA=e2t(I+tN)=e2t(1−2t−4tt1+2t).
Hence x(t)=etA(1,0)T=e2t(1−2t,−4t), so y=(1−2t)e2t. Direct check: y′=−4te2t, giving y′(0)=0, and y′′=(−4−8t)e2t, so y′′−4y′+4y=e2t(−4−8t+16t+4−8t)=0.
Worked example 6.2.9
(Variation of constants with a nilpotent matrix)
Solve x′=Ax+b with A=(0010), b=(0,1)T constant, x(0)=0. Here A2=0, so etA=I+tA=(10t1) and
Check: x1′=t=x2 and x2′=1, as required for x′=Ax+b.
Worked example 6.2.10
(Spiral decay)
Let A=(−12−2−1)=−I+2J with J=(01−10). Since −I commutes with 2J, etA=e−te2tJ=e−t(cos2tsin2t−sin2tcos2t). The solution with x0=(1,0) is x(t)=e−t(cos2t,sin2t), a spiral with ∣x(t)∣=e−t→0, consistent with eigenvalues −1±2i having negative real part.
Common mistakes and subtle points
Counterexample mode
The exponential is not taken entrywise
Claim: etA is the matrix with entries etaij. Take A=(0010). The claim gives E(t)=(11et1). But A2=0 gives etA=(10t1). At t=1 the entries e and 1 of E differ from 1 and 0. Also E′=(00et0) while AE=(1010), so E does not solve X′=AX. Repair: the exponential is the power series in the matrix; only for diagonal matrices does it act entrywise.
Common mistake
(Forgetting to change back to the original basis)
For A=(1221) the diagonal factor is etD=diag(e3t,e−t), but etA=PetDP−1 has off-diagonal entries 21(e3t−e−t)=0. Reporting etD as etA also gives the wrong solution for x0=(1,0): it yields (e3t,0), whereas the true second component is 21(e3t−e−t)=0.
Common mistake
(A double eigenvalue always brings a factor t)
The factors tk appear exactly when A is not diagonalizable at that eigenvalue. For A=2I2, the double eigenvalue 2 gives etA=e2tI with no t; for the Jordan block (2012) one gets e2t(10t1).
A higher-order scalar equation is turned into a first-order system by taking x=(y,y′,…); the first component of the solution is the answer.
Exercises
Quick check
(Compute etA for A=(3012), and solve x′=Ax, x(0)=(1,1)T.)
Solution · Answer
The eigenvalues are 3 (eigenvector (1,0)) and 2 (since A−2I=(1010), eigenvector (1,−1)). Take P=(101−1); then P2=I, so P−1=P. Thus
Derivative at 0: (303−22)=A. Then x(t)=etA(1,1)T=(2e3t−e2t,e2t).
Quick check
(Let A=(1−11−1). Show A2=0 and compute etA.)
Solution · Answer
A2=(1−1−1+11−1−1+1)=0. With N=A, the scalar-plus-nilpotent formula (taking λ=0, m=2) gives etA=I+tA=(1+t−tt1−t).
Quick check
(Let A=(102−3). Show that etA is upper triangular with diagonal (et,e−3t) and find detetA.)
Solution · Answer
Powers of an upper triangular matrix are upper triangular, with diagonal entries equal to the powers of the diagonal entries of A. So every partial sum of the series, and hence its limit, is upper triangular with diagonal entries n∑tnan/n!=eta for a=1,−3. Then detetA=ete−3t=e−2t=ettrA.
Quick check
(Find x(t) for x′=Ax, x(0)=(1,0)T, where A=(−12−2−1), and determine t→∞limx(t).)
Solution · Answer
A=−I+2J and −I commutes with 2J, so etA=e−t(cos2tsin2t−sin2tcos2t) and x(t)=e−t(cos2t,sin2t). Then ∣x(t)∣=e−t→0, so the limit is (0,0).
Quick check
(Solve x′=(200−1)x+(20), x(0)=0, by variation of constants.)
Solution · Answer
etA=diag(e2t,e−t) and b=(2,0)T. Then x(t)=∫0tdiag(e2(t−s),e−(t−s))(2,0)Tds=(2∫0te2(t−s)ds,0)=(e2t−1,0). Check: x1′=2e2t=2(e2t−1)+2 and x2=0 satisfies x2′=−x2.