The Matrix Exponential and Systems of Differential Equations
We want to solve a system of differential equations where the rates of change of x and y depend on both variables at once:
dtdxdtdy=x+2y=3x+2y
Because y is in the equation for dtdx and x is in the equation for dtdy, we cannot solve for either variable on its own. Putting the two variables into a vector turns the coupled system into a single matrix equation that looks just like dtdx=3x.
Writing a coupled system as dtdv=Av
Writing the coordinates as a vector v and the coefficients as a matrix A puts the entire system into one equation:
dtdv[xy]=A[1322]v[xy]⟹dtdv=Av
Here dtdv just means taking the derivative of each entry in the vector, [dtdxdtdy]. Just as dtdx=3x has the solution x(t)=e3tx0, this matrix equation has a solution of the form v(t)=eAtv0.
1
Defining the matrix exponential eAt
The matrix exponential eAt is the 2×2 matrix that plays the role of e3t, giving the solution v(t)=eAtv0 to dtdv=Av from any starting vector v0.
To see how a matrix can go into an exponent, remember the Taylor series for e3t, whose derivative is 3 times itself:
e3t=1+3t+2!(3t)2+3!(3t)3+⋯
Because this infinite series only uses powers, division by numbers, and addition, we can plug a square matrix A into the exact same formula.
Defining the matrix exponential eAt
dtdxx(t)=3x=e3tx0dtdvv(t)=Av=eAtv0
eAt=?
Replacing the number 1 with the identity matrix I and 3 with A defines the matrix exponential:
Just like e3t,eAt has derivative dtd(eAt)=AeAt and starts at eA⋅0=I when t=0. That makes the solution to dtdv=Av starting at v0:
v(t)=eAtv0
For a general matrix A, adding up I+At+2!A2t2+⋯ by hand means summing four infinite series at once, one for each entry of the 2×2 matrix. When the matrix is diagonal, however, x and y never mix, and raising the matrix to powers only raises its diagonal entries to powers:
eDt=exp([4t00−t])=[e4t00e−t]
2
Solving a system with eigenvectors
For a coupled matrix, the off-diagonal numbers mix x and y together, so we cannot just put each entry into an exponent. We want to solve our original coupled system from the starting point (5,5):
dtd[xy]=[1322][xy],[x(0)y(0)]=[55]
Geometrically, your position v has velocity Av. At most points, Av points in a new direction and pushes the particle off the line through v, but along an eigenvector line (Av=λv) the velocity stays on that line and the motion simplifies to dtdv=λv.
Solving a system with eigenvectors
Text solution
Find the eigenvalues of A to get the growth ratesSetting det(A−λI)=0 gives the growth rates λ for the system.
Split the starting point into eigenvector piecesBecause v1 and v2 point in different directions, we can write [55] as a combination of them by adding the two row equations.
As t grows, the e−t term shrinks to zero, so the particle moves closer to the v1=[23] line as its v2=[1−1] piece shrinks, and moves outward along v1, growing like e4t.
Starting directly on either eigenvector line keeps the particle on that line forever, while any other starting point bends toward the v1 line as its v2 piece shrinks by e−t, and moves outward along v1 as its v1 piece grows by e4t. A starting point close to the v2 line first comes in toward the origin along v2 before it turns out along v1.
Every starting point splits into a shrinking piece along one eigenvector line and a growing piece along the other, so as the shrinking piece dies out, the path bends toward the growing line and heads out along it.
3
Computing eAt by diagonalization
Splitting a starting vector into eigenvectors solved the system for one initial point, but we can also put those same three steps into a single formula to compute the matrix eAt itself for A=[1322].
When A has enough independent eigenvectors to make an invertible matrix P, we can factor A=PDP−1 and pull P and P−1 outside the Taylor series for eAt.
Computing eAt by diagonalization
Reading from right to left when multiplying a starting vector v0, factoring P and P−1 out of the series turns eAt into our three-step eigenvector method:
Set up P,eDt, and P−1 from the eigenvectors and eigenvaluesPut the eigenvectors v1=[23] and v2=[1−1] into the columns of P and match that column order with λ1=4 and λ2=−1 on the diagonal of eDt.
Plugging t=0 into your final matrix is the quickest way to catch arithmetic mistakes: since e0=1,eA⋅0 must simplify to the identity matrix [1001]. Also keep in mind that eAt=PeDtP−1 requires A to have enough independent eigenvectors to make P invertible; if a matrix is missing an eigenvector, P−1 does not exist even though the infinite series eAt still does.
4
Complex eigenvalues and spiraling solutions
When a system rotates vectors in the plane, there are no real straight-line eigenvectors, and det(A−λI)=0 gives complex eigenvalues λ=a±bi. We want to find the real path (x(t),y(t)) for a system with rotation:
dtd[xy]=[01−2−2][xy],[x(0)y(0)]=[31]
To turn a complex eigenvalue λ=a+bi into real functions of time, we use Euler's formula to split the exponent into a real growth factor and a rotation:
Find the complex eigenvalues and one eigenvectorWe only need the eigenvector for one of the two complex conjugate eigenvalues, so we plug λ=−1+i into (A−λI)v=0 and use the simpler bottom row.
det[−λ1−2−2−λ]=λ2+2λ+2=(λ+1)2+1=0⟹λ=−1±i
For λ=−1+i:[1−i1−2−1−i][xy]=[00]⟹x=(1+i)y⟹v=[1+i1]
Expand eλtv and split into real and imaginary partsMultiplying (cost+isint) into the vector using i2=−1 gives two independent real solutions, u(t) and w(t).
Combine the two real solutions to match the starting pointSetting [x(t)y(t)]=c1u(t)+c2w(t) and plugging in t=0 (where e0=1,cos0=1, and sin0=0) gives c1 and c2.
Starting at (3,1), the cosine and sine terms rotate the particle around the origin while e−t shrinks its distance to (0,0).
5
Long-term behavior from the eigenvalues
Often we don't need the full formula for x(t) and y(t). We just want to look at the eigenvalues of A and know what happens to the system as t→∞.
In discrete systems vk+1=Avk, each step multiplies by λ, so λk shrinks or grows without bound depending on whether the magnitude ∣λ∣ is smaller or larger than 1. In continuous differential equations dtdv=Av,λ is in the exponent of eλt, which changes the condition we check.
Long-term behavior from the eigenvalues
Because e(a+bi)t=eat(cosbt+isinbt), the imaginary part bi only brings in cosbt and sinbt, which keep repeating, so it makes the path go around the origin without making it grow or shrink over time. All growth or decay comes from eat, which shrinks when the real part a=Re(λ) is negative and grows when Re(λ) is positive:
In the complex plane, crossing the vertical line Re(λ)=0 changes inward spirals (Re(λ)<0) into closed orbits (Re(λ)=0) and then outward spirals (Re(λ)>0). The imaginary part b sets the rotation speed, and when b=0 on the real axis, the paths do not rotate at all.
Which side of the vertical line a complex eigenvalue sits on sets whether solution paths shrink, orbit, or grow, while its height only sets their spin speed.
Key concepts
Every coupled linear system dtdv=Av starting at v0 has the solution v(t)=eAtv0, where eAt=I+At+2!A2t2+3!A3t3+⋯. How you compute that solution depends on the matrix A:
When not to find eigenvectors
If the matrix is already diagonal, D=[400−1], the variables never mix, so just put the diagonal entries into exponents: eDt=[e4t00e−t]. If A is missing an eigenvector (such as A=[0040]), P−1 does not exist, so compute A2,A3,… and plug them straight into the Taylor series I+At+2!A2t2+⋯.
Distinct real eigenvalues (λ1,λ2)
To solve for one starting point v0, split v0=c1v1+c2v2 along the eigenvectors and multiply each piece by eλt, giving v(t)=c1eλ1tv1+c2eλ2tv2. To find the full matrix eAt, use eAt=PeDtP−1 and check that t=0 gives I.
Complex eigenvalues (λ=a±bi)
Find one eigenvector v for λ=a+bi, expand eat(cosbt+isinbt)v=u(t)+iw(t), and use the real and imaginary parts as your two real solutions: [x(t)y(t)]=c1u(t)+c2w(t).
Long-term behavior (t→∞)
Look at the sign of Re(λ) compared to 0 (not ∣λ∣ compared to 1). When both Re(λ)<0, paths decay to [00]; when both Re(λ)>0, paths grow without bound; opposite signs (λ1>0,λ2<0) form a saddle that pulls paths in along v2 and sends them out along v1; and pure imaginary eigenvalues λ=±bi (Re(λ)=0) give closed loops.
6
Practice
1
Find the matrix exponential eDt for D=[−2005], and use it to solve dtd[xy]=[−2005][xy] with starting point [x(0)y(0)]=[4−3].
Look at the off-diagonal zeros: do x and y mix, and what happens to the zero entries in every term of the Taylor series for eDt?
Because D is diagonal, each diagonal entry goes into its own exponential while the off-diagonal zeros stay 0 in every power of D:
eDt=[e−2t00e5t]
(Watch out not to turn the off-diagonal zeros into e0=1, which would turn eDt into [e−2t11e5t] and give [4e−2t−34−3e5t].) Multiplying eDt by the starting point [4−3] gives:
[x(t)y(t)]=[e−2t00e5t][4−3]=[4e−2t−3e5t]
So x(t)=4e−2t and y(t)=−3e5t.
2
Solve the coupled system dtdx=2x+3y,dtdy=4x+y with starting point [x(0)y(0)]=[5−2] by finding the eigenvalues and eigenvectors of the matrix.
Write the system with A=[2431], find the two eigenvector lines where the variables do not mix, and solve c1v1+c2v2=[5−2].
Now split the starting vector [5−2] into pieces along v1 and v2 (be careful not to use the starting coordinates 5 and −2 directly as c1 and c2, which would give x(t)=5e5t−6e−2t and y(t)=5e5t+8e−2t):
(Make sure to swap the diagonal entries when inverting P; forgetting to swap 2 and 1 gives [2−1−11], which leads to [4e3t−e2t2e3t−e2t−2e3t+e2t−e3t+e2t] and fails the t=0 check by giving [31−10].) Multiplying PeDtP−1 gives:
At t=0, this equals [2−11−1−2+2−1+2]=[1001]=I.
4
The matrix A=[0040] has only one eigenvector direction, so it cannot be diagonalized as PDP−1. Compute A2, then use the Taylor series eAt=I+At+2!A2t2+3!A3t3+⋯ directly to find eAt and check that dtd(eAt)=AeAt.
Multiply A by itself first. How many terms of the infinite Taylor series are non-zero?
Multiplying A by itself gives:
A2=[0040][0040]=[0000]
Because A2 is the zero matrix, every higher power Ak for k≥2 is also [0000]. All terms from 2!A2t2 onward are zero, leaving only the first two terms (don't forget to include the identity matrix I=[1001] at the start, or you would get [004t0]):
eAt=I+At=[1001]+[004t0]=[104t1]
Taking the derivative of each entry gives dtd(eAt)=[0040], which matches AeAt=[0040][104t1]=[0040].
5
Solve dtd[xy]=[21−50][xy] with starting point [x(0)y(0)]=[51], and describe the path as t increases.
Find one eigenvector v for λ=1+2i, and expand et(cos2t+isin2t)v using i2=−1 to separate the real part u(t) and imaginary part w(t).
First find the eigenvalues of A=[21−50]:
det[2−λ1−5−λ]=λ2−2λ+5=(λ−1)2+4=0⟹λ=1±2i
For λ=1+2i, the bottom row of (A−λI)v=0 is 1x+(−1−2i)y=0, so x=(1+2i)y, giving v=[1+2i1]. Expand e(1+2i)tv using Euler's formula (watch the sign on 2i2sin2t=−2sin2t in the top entry; writing +2sin2t instead leads to x(t)=et(5cos2t+4sin2t)):
The path starts at (5,1) and spirals outward away from the origin because Re(λ)=1>0.
6
For each matrix A below, determine the long-term behavior of solutions to the continuous system dtdv=Av as t→∞ and compare it with the discrete system vk+1=Avk as k→∞: (a) A has eigenvalues λ1=−0.5 (with eigenvector v1=[12]) and λ2=0.5 (with eigenvector v2=[1−1]). (b) A has pure imaginary eigenvalues λ=±2i.
Check how eλt behaves (sign of Re(λ) compared to 0) versus how λk behaves (magnitude ∣λ∣ compared to 1).
(a) In continuous time, λ1=−0.5<0 and λ2=0.5>0 have opposite signs (don't use the discrete test ∣0.5∣<1 to conclude that the λ2=0.5 piece decays; in e0.5t, a positive exponent grows). The system forms a saddle: starting points on the v1=[12] line decay to [00] like e−0.5t, while every other starting point has a piece along v2=[1−1] that grows like e0.5t without bound. In discrete time, both eigenvalues have magnitude ∣λ∣=0.5<1, so (−0.5)k and (0.5)k both go to 0 and every starting point decays to [00].
(b) In continuous time, Re(λ)=0, so the size factor is e0t=1 and non-zero solutions orbit the origin in closed loops forever without growing or shrinking. In discrete time, ∣±2i∣=2>1, so each eigenvector piece is multiplied by a number of magnitude 2 at every step, the pieces grow like 2k, and every non-zero starting point grows without bound as k→∞.