Linear Algebra › Section E · Eigenvalues and eigenvectors › Topic E.7
Applications to differential equations
When several quantities change continuously and each rate depends on all of them, the result is a system of differential equations. Eigenvectors split such a system into simple exponential pieces, and the eigenvalues predict whether solutions decay, grow or oscillate.
From one equation to a system
From calculus, the equation \(y' = ky\) has the solutions \(y = Ce^{kt}\): exponential growth when \(k > 0\) and decay when \(k < 0\). Many real situations couple several quantities. In two connected, well-mixed compartments, the amount in each changes at a rate that depends on both amounts. Writing \(\mathbf{x}(t) = (x_1(t), x_2(t))\) and differentiating entry by entry, such a model has the form
with a constant square matrix \(A\). Together with an initial condition \(\mathbf{x}(0) = \mathbf{x}_0\), it is an initial value problem. The solutions of the system form a vector space, because the equation is linear in \(\mathbf{x}\).
Eigenvectors give exponential solutions
Guess a solution that keeps a fixed direction and changes only in size: \(\mathbf{x}(t) = e^{\lambda t}\mathbf{v}\) for a fixed vector \(\mathbf{v}\). Its derivative is \(\lambda e^{\lambda t}\mathbf{v}\), while \(A\mathbf{x}(t) = e^{\lambda t}A\mathbf{v}\). These match for every \(t\) exactly when \(A\mathbf{v} = \lambda\mathbf{v}\). So every eigenpair of \(A\) produces a solution, sometimes called an eigenfunction of the system. Along the line through an eigenvector, the system behaves just like the one-variable equation \(y' = \lambda y\): the solution stays on the line and grows or decays at the rate set by \(\lambda\).
The general solution
Because the equation is linear, any combination of solutions is again a solution. If \(A\) is \(n \times n\) with \(n\) linearly independent eigenvectors \(\mathbf{v}_1, \dots, \mathbf{v}_n\) and eigenvalues \(\lambda_1, \dots, \lambda_n\), the general solution is
Setting \(t = 0\) makes every exponential equal to 1, so the initial condition becomes \(c_1\mathbf{v}_1 + \dots + c_n\mathbf{v}_n = \mathbf{x}_0\), the linear system \(P\mathbf{c} = \mathbf{x}_0\). Compare "Discrete dynamical systems": the powers \(\lambda^k\) of a step-by-step model are replaced by the exponentials \(e^{\lambda t}\) of a continuous one.
Worked example: two compartments
In a model of two well-mixed compartments, each loses material at 3 times its own amount per hour and receives material at 1 times the other's amount per hour. So \(\mathbf{x}' = A\mathbf{x}\) with
and the starting amounts are \(\mathbf{x}(0) = (5, 1)\). 1. Trace -6 and determinant 8 give \((\lambda + 2)(\lambda + 4) = 0\). 2. Eigenvectors: (1, 1) for -2 and (1, -1) for -4. 3. Solve \(c_1(1, 1) + c_2(1, -1) = (5, 1)\): \(c_1 = 3\) and \(c_2 = 2\). 4. So
5. Check: at \(t = 0\) this is (5, 1), and the derivative of the first entry, \(-6e^{-2t} - 8e^{-4t}\), equals \(-3x_1 + x_2\). The difference between the compartments dies out faster than the total.
Decoupling
The eigenvectors work because they decouple the system. With \(A = PDP^{-1}\), introduce new variables \(\mathbf{y} = P^{-1}\mathbf{x}\), so that \(\mathbf{x} = P\mathbf{y}\). Substituting, \(P\mathbf{y}' = AP\mathbf{y}\), and multiplying by \(P^{-1}\) gives
Since \(D\) is diagonal, this is just \(n\) separate equations \(y_i' = \lambda_i y_i\), each solved by \(y_i = c_ie^{\lambda_i t}\). Converting back with \(\mathbf{x} = P\mathbf{y}\) reproduces the general solution. In the original variables every rate involves every quantity; in eigenvector coordinates each new variable evolves on its own.
Attractors, repellers and saddles
For a 2 × 2 system with real eigenvalues, the signs decide the picture. If both eigenvalues are negative, every term \(e^{\lambda t}\) decays and all solutions approach the origin, an attractor or sink. If both are positive, solutions move away from the origin, a repeller or source. If one is positive and one negative, the origin is a saddle point: solutions starting on the line of the negative eigenvalue's eigenvector approach the origin, and all others eventually head off along the positive one's direction. Note the contrast with discrete systems, where the test is \(|\lambda| < 1\). Here the test is the sign of \(\lambda\), because \(e^{\lambda t} \to 0\) exactly when \(\lambda < 0\).
Worked example: a saddle
Solve \(\mathbf{x}' = A\mathbf{x}\) with \(\mathbf{x}(0) = (1, 2)\) and classify the origin.
1. The trace is 0 and the determinant is -9, so \(\lambda^2 - 9 = 0\) and \(\lambda = \pm 3\). 2. \(A - 3I\) has rows (-2, 4) and (2, -4), giving (2, 1); \(A + 3I\) has rows (4, 4) and (2, 2), giving (1, -1). 3. One positive and one negative eigenvalue: a saddle point. 4. Solve \(c_1(2, 1) + c_2(1, -1) = (1, 2)\): \(c_1 = 1\) and \(c_2 = -1\). 5. So \(\mathbf{x}(t) = e^{3t}(2, 1) - e^{-3t}(1, -1)\), which equals (1, 2) at \(t = 0\). As \(t\) grows, the solution runs out along the direction (2, 1).
Complex eigenvalues: oscillation
When \(A\) is real with eigenvalues \(a \pm bi\), the solution \(e^{\lambda t}\mathbf{v}\) is complex, but its real and imaginary parts are both real solutions, and together they give the general real solution. The key identity is
The sines and cosines make solutions rotate around the origin, and the factor \(e^{at}\) controls their size. If \(a < 0\) the trajectories spiral in, a spiral attractor; if \(a > 0\) they spiral out; if \(a = 0\) they travel round closed curves forever. For example, eigenvalues \(-1 \pm 2i\) give solutions that oscillate with period \(\pi\) while shrinking like \(e^{-t}\).
Worked example: an undamped spring
A mass on a spring, with no friction and suitable units, obeys \(y'' + y = 0\). 1. Let \(x_1 = y\) and \(x_2 = y'\). Then \(x_1' = x_2\) and \(x_2' = -x_1\), so \(A\) has rows (0, 1) and (-1, 0). 2. Trace 0 and determinant 1 give \(\lambda^2 + 1 = 0\), so \(\lambda = \pm i\). 3. The first row of \(A - iI\) is \((-i, 1)\), so the eigenvector's second entry is \(i\) times its first, and \(\mathbf{v} = (1, i)\). 4. Then
5. Its real part \((\cos t, -\sin t)\) and imaginary part \((\sin t, \cos t)\) are solutions, so \(y = c_1\cos t + c_2\sin t\). With \(a = 0\), the motion is a steady oscillation that neither grows nor fades.
Common mistakes to avoid
First, carrying over the discrete test: for \(\mathbf{x}' = A\mathbf{x}\), eigenvalues -0.5 and -3 both mean decay, and an eigenvalue of 0.5 means growth, even though it is less than 1. Second, writing \(\lambda^t\) instead of \(e^{\lambda t}\). Third, reading the constants \(c_i\) straight from \(\mathbf{x}(0)\); they are the weights in \(P\mathbf{c} = \mathbf{x}_0\), not the entries of \(\mathbf{x}_0\). Fourth, judging complex eigenvalues by the modulus: in continuous time only the real part \(a\) decides spiral in or out.
Key takeaways
Each eigenpair of \(A\) gives a solution \(e^{\lambda t}\mathbf{v}\) of \(\mathbf{x}' = A\mathbf{x}\). With \(n\) independent eigenvectors, the general solution is \(\sum c_ie^{\lambda_i t}\mathbf{v}_i\), with weights from \(P\mathbf{c} = \mathbf{x}_0\). The change of variables \(\mathbf{y} = P^{-1}\mathbf{x}\) decouples the system into \(y_i' = \lambda_i y_i\). Negative eigenvalues give an attractor, positive ones a repeller, mixed signs a saddle. Complex eigenvalues \(a \pm bi\) give rotation, spiralling in if \(a < 0\) and out if \(a > 0\).
Flashcards
Flashcard
A linear system of differential equations with constant matrix A
Test your understanding
Question 1 of 5
A has eigenvector v with Av = −2v. Which function solves x′ = Ax?
Question 2 of 5
For x′ = Ax, the 2 × 2 matrix A has eigenvalues −1 and −4. What happens to solutions as t grows?
Question 3 of 5
A has eigenvector (1, 0) for λ = 1 and (1, 1) for λ = −2. Which function solves x′ = Ax with x(0) = (3, 2)?
Question 4 of 5
A real 2 × 2 matrix A has eigenvalues −0.5 ± 3i. How do the solutions of x′ = Ax behave?
Question 5 of 5
A has eigenvector (2, 1) for λ = 3 and (1, −1) for λ = −3. Which initial condition gives a solution of x′ = Ax that approaches the origin?
