math.wiki
Calculus › Differential equations › Systems and transforms › Linear systems of differential equations
Sign in

Linear systems of differential equations

When several quantities each change at a rate set by all of them, how are they solved together, and why do the eigenvectors of one matrix solve every such system with constant coefficients?

Where you are 6 in Systems and transforms

Why this exists

Connected tanks, competing populations, coupled springs: in each, several quantities change at rates that depend on all of them at once. At a koi farm, the treatment in a pond falls as pumps carry water to a quarantine tank and rises as water comes back, while the tank’s amount does the opposite. Collecting the unknowns in one vector \mathbf{x}(t) turns such a model into \mathbf{x}' = A\mathbf{x}. For the discrete process \mathbf{x}_{k+1} = A\mathbf{x}_k, Matrix powers, dynamical systems and steady states split the motion along eigenvectors, each piece multiplied by \lambda per step. The same splitting works in continuous time with e^{\lambda t} in place of \lambda^k. An equation of order n is a system too, so Constant-coefficient equations is a special case.

Definition

In plain words: a linear system gives the rate of change of each unknown as a combination of all the unknowns, with coefficients that may depend on time, plus a forcing term.

Let I be an open interval, P(t) an n \times n matrix whose entries p_{ij} are continuous on I, and \mathbf{f}(t) a column of n functions continuous on I. The equation

\mathbf{x}' = P(t)\,\mathbf{x} + \mathbf{f}(t)

is a first-order linear system, homogeneous when \mathbf{f} = \mathbf{0}. Entry by entry it says x_i' = \sum_j p_{ij}(t)\,x_j + f_i(t). A solution is a differentiable vector function on an open interval K \subseteq I satisfying the equation at every point of K. An initial value problem adds \mathbf{x}(t_0) = \mathbf{x}_0 with t_0 \in I. When P(t) = A is a constant real matrix, \mathbf{x}' = A\mathbf{x} has constant coefficients.

Vector functions \mathbf{x}_1, \dots, \mathbf{x}_n on I are linearly dependent if c_1\mathbf{x}_1 + \dots + c_n\mathbf{x}_n = \mathbf{0} at every t \in I for constants not all 0, and linearly independent otherwise (Linear independence). Their Wronskian is

W(t) = \det\begin{bmatrix} \mathbf{x}_1(t) & \cdots & \mathbf{x}_n(t) \end{bmatrix}.

Then n linearly independent solutions of the homogeneous system on I form a fundamental set, and the matrix X(t) with those columns is a fundamental matrix.

Intuition

At each point \mathbf{x}, the system \mathbf{x}' = A\mathbf{x} prescribes the velocity A\mathbf{x}, which usually points off the line through \mathbf{0} and \mathbf{x}. Along an eigenvector it does not: if A\mathbf{v} = \lambda\mathbf{v}, the velocity at s\mathbf{v} is \lambda s\,\mathbf{v}. A solution starting on that line stays on it, with s' = \lambda s, so s = s_0e^{\lambda t}. These straight-line solutions e^{\lambda t}\mathbf{v} are the system’s version of e^{rx}.

In Figure 1, for A = \begin{bmatrix} 1 & 2 \\ 3 & 2 \end{bmatrix} of Example 2, points on the line through (2, 3) move out like e^{4t} and points on the line through (1, -1) move in like e^{-t}. Any other start is a combination of the two directions, each part moves on its own, and the curve bends from the second line toward the first.

x₁ x₂
Figure 1Solutions of \mathbf{x}' = A\mathbf{x} for A = \begin{bmatrix} 1 & 2 \\ 3 & 2 \end{bmatrix} (Example 2). The thick lines are the straight-line solutions e^{4t}\mathbf{v}_1 along \mathbf{v}_1 = (2, 3) and e^{-t}\mathbf{v}_2 along \mathbf{v}_2 = (1, -1), with their negatives. Every other solution c_1e^{4t}\mathbf{v}_1 + c_2e^{-t}\mathbf{v}_2 follows the second line in the past and the first in the future.

Results

1. Superposition. If \mathbf{x}_1, \mathbf{x}_2 solve \mathbf{x}' = P(t)\mathbf{x}, so does c_1\mathbf{x}_1 + c_2\mathbf{x}_2 for all constants. If \mathbf{x}_p solves \mathbf{x}' = P\mathbf{x} + \mathbf{f} on I, the solutions on I are exactly \mathbf{x}_p + \mathbf{x}_h with \mathbf{x}_h solving the homogeneous system.

2. Existence and uniqueness. If P and \mathbf{f} are continuous on I, t_0 \in I and \mathbf{x}_0 \in \mathbb{R}^n, the initial value problem has a solution on all of I, and every solution on an open interval K \subseteq I containing t_0 is its restriction to K.

3. Liouville’s formula. If \mathbf{x}_1, \dots, \mathbf{x}_n solve \mathbf{x}' = P(t)\mathbf{x} on I and t_0 \in I, then

W(t) = W(t_0)\,e^{\int_{t_0}^{t} \operatorname{tr} P(s)\,ds}.

So W is identically 0 on I or never 0 there.

4. Fundamental sets. For n solutions of \mathbf{x}' = P(t)\mathbf{x} on I these are equivalent: (i) W(t_0) \ne 0 at some t_0 \in I; (ii) W has no zero on I; (iii) the solutions are linearly independent. Then every solution on I is X(t)\,\mathbf{c} for exactly one constant vector \mathbf{c} (the general solution), and \mathbf{x}(t_0) = \mathbf{x}_0 gives \mathbf{c} = X(t_0)^{-1}\mathbf{x}_0.

5. Equations of order n. Let a_0, \dots, a_{n-1} be continuous on I, and let P(t) be the companion matrix: 1 in each entry (i, i + 1), last row (-a_0, \dots, -a_{n-1}), 0 elsewhere. Then y solves

y^{(n)} + a_{n-1}y^{(n-1)} + \dots + a_1y' + a_0y = 0

exactly when \mathbf{x} = (y, y', \dots, y^{(n-1)}) solves \mathbf{x}' = P(t)\mathbf{x}, and every solution of the system has this form. If the a_k are constants, the eigenvalues of P are the roots of \lambda^n + a_{n-1}\lambda^{n-1} + \dots + a_0, each with eigenspace spanned by (1, \lambda, \dots, \lambda^{n-1}).

6. The eigenvalue method. Let A be a real n \times n matrix.

  • (a) For a constant vector \mathbf{v} \ne \mathbf{0}, e^{\lambda t}\mathbf{v} solves \mathbf{x}' = A\mathbf{x} if and only if A\mathbf{v} = \lambda\mathbf{v}.
  • (b) Let \mathbf{v}_1, \dots, \mathbf{v}_n be linearly independent eigenvectors with real eigenvalues \lambda_1, \dots, \lambda_n (they exist when A has n distinct real eigenvalues). Every solution on an open interval is the restriction of
\mathbf{x} = c_1e^{\lambda_1t}\mathbf{v}_1 + \dots + c_ne^{\lambda_nt}\mathbf{v}_n

for exactly one choice of constants. With P = \begin{bmatrix} \mathbf{v}_1 & \cdots & \mathbf{v}_n \end{bmatrix} and e^{tD} = \operatorname{diag}(e^{\lambda_1t}, \dots, e^{\lambda_nt}), the solution with \mathbf{x}(0) = \mathbf{x}_0 is Pe^{tD}P^{-1}\mathbf{x}_0, defined on \mathbb{R}.

7. Complex eigenvalues. Let A be real with eigenvalue \lambda = \alpha + \beta i, \beta \ne 0, eigenvector \mathbf{v}, \mathbf{a} = \operatorname{Re}\mathbf{v} and \mathbf{b} = \operatorname{Im}\mathbf{v}. Then

\begin{aligned} \mathbf{x}_1 &= e^{\alpha t}(\mathbf{a}\cos\beta t - \mathbf{b}\sin\beta t), \\ \mathbf{x}_2 &= e^{\alpha t}(\mathbf{a}\sin\beta t + \mathbf{b}\cos\beta t) \end{aligned}

are linearly independent real solutions of \mathbf{x}' = A\mathbf{x}, the real and imaginary parts of e^{\lambda t}\mathbf{v}. For n = 2, every real solution is c_1\mathbf{x}_1 + c_2\mathbf{x}_2 for exactly one pair of real constants.

Why it is true

Result 1. Differentiation and multiplication by P(t) are linear, so \mathbf{x} \mapsto \mathbf{x}' - P\mathbf{x} is linear: it sends c_1\mathbf{x}_1 + c_2\mathbf{x}_2 to c_1(\mathbf{x}_1' - P\mathbf{x}_1) + c_2(\mathbf{x}_2' - P\mathbf{x}_2), and it sends \mathbf{x} to \mathbf{f} exactly when it sends \mathbf{x} - \mathbf{x}_p to \mathbf{0}. \square

Result 2 is stated without proof in general. Existence and uniqueness proves Picard’s theorem for one equation; the same iteration, with |y| replaced by the length of a vector and the Lipschitz constant by a bound for the entries of P on a closed subinterval, proves it for systems, on every closed subinterval of I at once. For constant A with a basis of eigenvectors, Result 6(b) proves it outright, as Result 7 does for 2 \times 2 matrices with complex eigenvalues.

Result 3. Let X have the solutions as columns and rows \mathbf{r}_1, \dots, \mathbf{r}_n. Since X' = PX, row i of X' is \sum_j p_{ij}\mathbf{r}_j. The determinant is a sum of products with one factor from each row, so by the product rule W' is the sum over i of \det X with row i differentiated. Determinants are linear in each row (Properties of determinants: row operations, products, transposes), so that term is \sum_j p_{ij} times \det X with \mathbf{r}_j in row i, which is 0 for j \ne i (two equal rows). Hence W' = (\operatorname{tr} P)\,W, and Result 2(b) of Linear equations and integrating factors gives the formula. \square

Result 4. (i) implies (ii) by Result 3. (ii) implies (iii): if X(t)\mathbf{c} = \mathbf{0} on I, then \mathbf{c} = \mathbf{0} because X(t_0) is invertible. (iii) implies (i): if W(t_0) = 0, some \mathbf{c} \ne \mathbf{0} has X(t_0)\mathbf{c} = \mathbf{0}; then X\mathbf{c} and \mathbf{0} solve the same initial value problem, so X\mathbf{c} = \mathbf{0} on I by Result 2. Finally, for a solution \mathbf{x}, the equation X(t_0)\mathbf{c} = \mathbf{x}(t_0) has exactly one solution, and then X\mathbf{c} = \mathbf{x} by Result 2. \square

Result 5. For \mathbf{x} = (y, \dots, y^{(n-1)}), rows 1 to n - 1 of the system say (y^{(k)})' = y^{(k+1)}, always true, and the last row is the equation. Conversely, rows 1 to n - 1 give x_{k+1} = x_1^{(k)} by induction, so \mathbf{x} has this form with y = x_1. If P\mathbf{v} = \lambda\mathbf{v}, rows 1 to n - 1 say v_{k+1} = \lambda v_k, so \mathbf{v} = v_1(1, \lambda, \dots, \lambda^{n-1}) with v_1 \ne 0, and for such \mathbf{v} the last row holds exactly when -(a_0 + a_1\lambda + \dots + a_{n-1}\lambda^{n-1})\,v_1 = \lambda^nv_1, that is, when \lambda is a root. \square

Result 6. (a) (e^{\lambda t}\mathbf{v})' = \lambda e^{\lambda t}\mathbf{v} and A(e^{\lambda t}\mathbf{v}) = e^{\lambda t}A\mathbf{v} agree exactly when A\mathbf{v} = \lambda\mathbf{v}, since e^{\lambda t} \ne 0.

(b) P is invertible and AP = PD with D = \operatorname{diag}(\lambda_1, \dots, \lambda_n) (Diagonalization). Let \mathbf{x} solve on an open interval K and put \mathbf{y} = P^{-1}\mathbf{x}, its coordinates in the eigenvector basis. Then

\mathbf{y}' = P^{-1}A\mathbf{x} = P^{-1}AP\,\mathbf{y} = D\mathbf{y},

that is, y_i' = \lambda_iy_i for each i: the system comes apart into n scalar equations. By Result 2(b) of Linear equations and integrating factors, y_i = c_ie^{\lambda_it}, so \mathbf{x} = P\mathbf{y} = \sum_i c_ie^{\lambda_it}\mathbf{v}_i. Conversely every such sum solves, by (a) and Result 1. The constants are unique because \mathbf{x} = Pe^{tD}\mathbf{c} with Pe^{tD} invertible, and \mathbf{c} = P^{-1}\mathbf{x}_0. Eigenvectors for distinct eigenvalues are independent (Finding an eigenvector). \square

Result 7. By Euler’s formula (Complex and repeated roots), e^{\lambda t} = e^{\alpha t}(\cos\beta t + i\sin\beta t) has derivative \lambda e^{\lambda t}, so as in 6(a), \mathbf{w} = e^{\lambda t}\mathbf{v} satisfies \mathbf{w}' = A\mathbf{w}. Expanding, \mathbf{w} = \mathbf{x}_1 + i\,\mathbf{x}_2, and since A is real the real and imaginary parts give \mathbf{x}_1' = A\mathbf{x}_1 and \mathbf{x}_2' = A\mathbf{x}_2.

At t = 0 the solutions are \mathbf{a} and \mathbf{b}, so independence of \mathbf{a}, \mathbf{b} is enough. If they were dependent, \mathbf{v} = \mu\mathbf{u} with \mathbf{u} real and nonzero and \mu complex, and A\mathbf{u} = \lambda\mathbf{u} with A\mathbf{u} real would make \lambda real.

For n = 2: \bar{\mathbf{v}} is an eigenvector for \bar\lambda \ne \lambda, so the proof of 6(b) runs with the complex matrix \begin{bmatrix} \mathbf{v} & \bar{\mathbf{v}} \end{bmatrix} (for complex y, y' = \lambda y makes the real and imaginary parts of e^{-\lambda t}y constant, by the mean value theorem). So a real solution is c\,e^{\lambda t}\mathbf{v} + d\,e^{\bar\lambda t}\bar{\mathbf{v}} for unique complex c, d; conjugating shows d = \bar c, so \mathbf{x} = 2\operatorname{Re}(c\,\mathbf{w}) = 2(\operatorname{Re}c)\,\mathbf{x}_1 - 2(\operatorname{Im}c)\,\mathbf{x}_2. \square

Worked examples

Example 1

From one equation to a system

Write y''' - 2y'' - y' + 2y = 0 as a first-order system, find the eigenvalues and eigenvectors of its matrix, and find the general solution of the equation.

Solution

Here a_0 = 2, a_1 = -1, a_2 = -2, so by Result 5 the vector \mathbf{x} = (y, y', y'') solves

\mathbf{x}' = \begin{bmatrix} 0 & 1 & 0 \\ 0 & 0 & 1 \\ -2 & 1 & 2 \end{bmatrix}\mathbf{x}.

The eigenvalues are the roots of

\begin{aligned} \lambda^3 - 2\lambda^2 - \lambda + 2 &= \lambda^2(\lambda - 2) - (\lambda - 2) \\ &= (\lambda - 2)(\lambda - 1)(\lambda + 1), \end{aligned}

with eigenvectors (1, \lambda, \lambda^2): (1, 2, 4), (1, 1, 1) and (1, -1, 1). The matrix sends (1, 2, 4) to (2, 4, 8). The eigenvalues are distinct, so by Result 6(b) every solution of the system is

\begin{aligned} \mathbf{x} &= c_1e^{2t}\begin{bmatrix} 1 \\ 2 \\ 4 \end{bmatrix} + c_2e^{t}\begin{bmatrix} 1 \\ 1 \\ 1 \end{bmatrix} \\ &\quad + c_3e^{-t}\begin{bmatrix} 1 \\ -1 \\ 1 \end{bmatrix}, \end{aligned}

and by Result 5 the solutions of the equation are its first entries, y = c_1e^{2t} + c_2e^{t} + c_3e^{-t}.

Example 2

Two exponentials in the plane

Solve \mathbf{x}' = \begin{bmatrix} 1 & 2 \\ 3 & 2 \end{bmatrix}\mathbf{x}, \mathbf{x}(0) = (3, 7), and check Liouville’s formula for the two straight-line solutions.

Solution

The trace is 3 and the determinant -4, so the The characteristic equation and finding eigenvalues is \lambda^2 - 3\lambda - 4 = (\lambda - 4)(\lambda + 1) = 0. For \lambda = 4, A - 4I = \begin{bmatrix} -3 & 2 \\ 3 & -2 \end{bmatrix} gives \mathbf{v}_1 = (2, 3); for \lambda = -1, A + I = \begin{bmatrix} 2 & 2 \\ 3 & 3 \end{bmatrix} gives \mathbf{v}_2 = (1, -1). By Result 6(b), \mathbf{x} = c_1e^{4t}\mathbf{v}_1 + c_2e^{-t}\mathbf{v}_2, and \mathbf{x}(0) = (3, 7) reads 2c_1 + c_2 = 3, 3c_1 - c_2 = 7, so c_1 = 2 and c_2 = -1:

\begin{aligned} \mathbf{x} &= 2e^{4t}\begin{bmatrix} 2 \\ 3 \end{bmatrix} - e^{-t}\begin{bmatrix} 1 \\ -1 \end{bmatrix} \\ &= \begin{bmatrix} 4e^{4t} - e^{-t} \\ 6e^{4t} + e^{-t} \end{bmatrix}. \end{aligned}

Check the first row: x_1' = 16e^{4t} + e^{-t} = x_1 + 2x_2. The Wronskian of e^{4t}\mathbf{v}_1 and e^{-t}\mathbf{v}_2 is

W = \det\begin{bmatrix} 2e^{4t} & e^{-t} \\ 3e^{4t} & -e^{-t} \end{bmatrix} = -5e^{3t},

and Liouville’s formula with \operatorname{tr} A = 3 gives W(0)\,e^{3t} = -5e^{3t} too.

Example 3

Complex eigenvalues

Solve \mathbf{x}' = \begin{bmatrix} 1 & -5 \\ 1 & -3 \end{bmatrix}\mathbf{x}, \mathbf{x}(0) = (1, 2).

Solution

The trace is -2 and the determinant 2, so \lambda^2 + 2\lambda + 2 = 0 and \lambda = -1 \pm i. For \lambda = -1 + i,

A - \lambda I = \begin{bmatrix} 2 - i & -5 \\ 1 & -2 - i \end{bmatrix}.

The second row gives x_1 = (2 + i)\,x_2, so \mathbf{v} = (2 + i, 1); the first row agrees, since (2 - i)(2 + i) = 5. With \mathbf{a} = (2, 1), \mathbf{b} = (1, 0), \alpha = -1, \beta = 1, Result 7 gives

\begin{aligned} \mathbf{x}_1 &= e^{-t}\begin{bmatrix} 2\cos t - \sin t \\ \cos t \end{bmatrix}, \\ \mathbf{x}_2 &= e^{-t}\begin{bmatrix} 2\sin t + \cos t \\ \sin t \end{bmatrix}. \end{aligned}

At t = 0 these are \mathbf{a} and \mathbf{b}, so 2c_1 + c_2 = 1 and c_1 = 2, giving c_2 = -3 and

\mathbf{x} = e^{-t}\begin{bmatrix} \cos t - 8\sin t \\ 2\cos t - 3\sin t \end{bmatrix}.

Check the second row: x_2' = e^{-t}(-5\cos t + \sin t) = x_1 - 3x_2.

Example 4

A repeated eigenvalue with enough eigenvectors

Solve \mathbf{x}' = \begin{bmatrix} 2 & 1 & 1 \\ 1 & 2 & 1 \\ 1 & 1 & 2 \end{bmatrix}\mathbf{x}, \mathbf{x}(0) = (3, 0, 0).

Solution

Write A = I + J, where every entry of J is 1. Since J(1, 1, 1) = (3, 3, 3), A(1, 1, 1) = 4(1, 1, 1). A vector with v_1 + v_2 + v_3 = 0 has J\mathbf{v} = \mathbf{0}, so A\mathbf{v} = \mathbf{v}: the eigenvalue 1 has a whole plane of eigenvectors, with basis (1, -1, 0) and (1, 0, -1). The three eigenvectors are independent, since

\det\begin{bmatrix} 1 & 1 & 1 \\ 1 & -1 & 0 \\ 1 & 0 & -1 \end{bmatrix} = 3.

Result 6(b) asks for independent eigenvectors, not distinct eigenvalues, so

\begin{aligned} \mathbf{x} &= c_1e^{4t}(1, 1, 1) \\ &\quad + e^{t}\big(c_2(1, -1, 0) + c_3(1, 0, -1)\big). \end{aligned}

At t = 0: c_1 + c_2 + c_3 = 3, c_1 = c_2 and c_1 = c_3, so all three are 1 and

\mathbf{x} = \big(e^{4t} + 2e^{t},\ e^{4t} - e^{t},\ e^{4t} - e^{t}\big).

Check the first row: x_1' = 4e^{4t} + 2e^{t} = 2x_1 + x_2 + x_3. Every column of A adds to 4, so the sum x_1 + x_2 + x_3 = 3e^{4t} solves s' = 4s on its own.

Example 5

Kenji’s quarantine tank

At a koi farm, Kenji has a pond holding 4{,}000 liters and a quarantine tank holding 1{,}000 liters, joined by two pumps: one moves 200 liters per minute from pond to tank, the other 200 liters per minute back. She dissolves 100 grams of a treatment in the tank, none in the pond, and starts the pumps. Each body of water stays well mixed. Find the amount in each after t minutes, and when the tank’s amount is within 10\% of its final value.

Solution

Let m_1 and m_2 be the grams in the pond and the tank. The pond loses \displaystyle 200 \cdot \frac{m_1}{4000} = 0.05\,m_1 grams per minute and gains \displaystyle 200 \cdot \frac{m_2}{1000} = 0.2\,m_2; the tank does the reverse, so

\mathbf{m}' = \begin{bmatrix} -0.05 & 0.2 \\ 0.05 & -0.2 \end{bmatrix}\mathbf{m}, \qquad \mathbf{m}(0) = \begin{bmatrix} 0 \\ 100 \end{bmatrix}.

The trace is -0.25 and the determinant 0, so \lambda(\lambda + 0.25) = 0. For \lambda = 0, 0.05\,m_1 = 0.2\,m_2 gives (4, 1); for \displaystyle \lambda = -\frac{1}{4}, \displaystyle A + \frac14 I = \begin{bmatrix} 0.2 & 0.2 \\ 0.05 & 0.05 \end{bmatrix} gives (1, -1). Then 4c_1 + c_2 = 0 and c_1 - c_2 = 100, so c_1 = 20, c_2 = -80, and

\mathbf{m} = \begin{bmatrix} 80 - 80e^{-t/4} \\ 20 + 80e^{-t/4} \end{bmatrix}.

The eigenvalue 0 gives the equilibrium solution (80, 20), where both concentrations are 0.02 grams per liter; the total stays 100 because each column of the matrix adds to 0. The tank’s amount is within 10\% of 20 when 80e^{-t/4} \le 2, that is

t \ge 4\ln 40 \approx 14.76 \text{ minutes}.
5 10 15 20 20 40 60 80 100 pond tank t m
Figure 2Kenji’s pond and tank (Example 5): m_1 = 80 - 80e^{-t/4} and m_2 = 20 + 80e^{-t/4} grams after t minutes. Both approach the equilibrium solution (80, 20), an eigenvector for the eigenvalue 0. The dot marks t = 4\ln 40 \approx 14.76.
Example 6

A short proof: Abel’s identity from Liouville’s formula

Let p, q be continuous on I and y_1, y_2 solve y'' + p(t)\,y' + q(t)\,y = 0 there. Prove that W(y_1, y_2)(t) = W(t_0)\,e^{-\int_{t_0}^{t} p(s)\,ds}, the identity of Constant-coefficient equations.

Solution

By Result 5 with a_0 = q and a_1 = p, the vectors \mathbf{x}_i = (y_i, y_i') solve \mathbf{x}' = P(t)\mathbf{x} with

P(t) = \begin{bmatrix} 0 & 1 \\ -q(t) & -p(t) \end{bmatrix}, \qquad \operatorname{tr} P = -p.

Their Wronskian is \det\begin{bmatrix} y_1 & y_2 \\ y_1' & y_2' \end{bmatrix} = y_1y_2' - y_1'y_2, the Wronskian of y_1, y_2, so Result 3 gives the identity. \square

Boundaries

  • Not enough eigenvectors. For A = \begin{bmatrix} 2 & 1 \\ 0 & 2 \end{bmatrix} the eigenvectors are the multiples of (1, 0). Solving from the bottom row, x_2 = c_2e^{2t}, then x_1' - 2x_1 = c_2e^{2t} gives x_1 = (c_1 + c_2t)\,e^{2t}: a factor t, as for the repeated root in Complex and repeated roots. The general case needs generalized eigenvectors, not covered here.
  • Constant coefficients only. Results 1 to 4 hold for P(t); the eigenvalue method does not. The matrix of Exercise 2 has the single eigenvalue \displaystyle \frac{1}{t} with eigenvector (1, 0), yet e^{t/t}(1, 0) = (e, 0) is not a solution.

Where this goes next

Phase portraits draws every solution curve of a 2 \times 2 system and sorts them by eigenvalues into nodes, saddles, spirals and centers; Figure 1 and Example 3 are two. Nonlinear systems and linearization studies a nonlinear system near an equilibrium through a linear one. The Laplace transform solves forced systems by algebra.

Exercises

Systems and fundamental sets

  1. Write u'' = -2u + v, v'' = u - 2v as a first-order system \mathbf{x}' = A\mathbf{x} with \mathbf{x} = (u, u', v, v'), and check that u = v = \cos t gives a solution.
  2. Verify that (t, 0) and (t^2, t) solve \mathbf{x}' = P(t)\mathbf{x} on (0, \infty) with P(t) = \begin{bmatrix} 1/t & 1 \\ 0 & 1/t \end{bmatrix}, compute their Wronskian, check it against Liouville’s formula, and solve with \mathbf{x}(1) = (1, 3).
  3. Find the 2 \times 2 matrix A for which e^{3t}(1, 2) and e^{-t}(1, -2) both solve \mathbf{x}' = A\mathbf{x}.

Real eigenvalues

  1. Find the general solution of \mathbf{x}' = \begin{bmatrix} 5 & -1 \\ 3 & 1 \end{bmatrix}\mathbf{x}.
  2. Solve \mathbf{x}' = \begin{bmatrix} -4 & 2 \\ 1 & -3 \end{bmatrix}\mathbf{x}, \mathbf{x}(0) = (1, 4).
  3. Solve \mathbf{x}' = \begin{bmatrix} 1 & 2 \\ 2 & 4 \end{bmatrix}\mathbf{x}, \mathbf{x}(0) = (0, 5).
  4. Solve \mathbf{x}' = \begin{bmatrix} 1 & 0 & 0 \\ 2 & 3 & 0 \\ 1 & 1 & -1 \end{bmatrix}\mathbf{x}, \mathbf{x}(0) = (1, 3, 2).
  5. Solve y''' - y'' - 4y' + 4y = 0, y(0) = 1, y'(0) = 0, y''(0) = 0 through its companion system.
  6. Find the general solution of \mathbf{x}' = \begin{bmatrix} 3 & 0 & 0 \\ 0 & 1 & 2 \\ 0 & 2 & 1 \end{bmatrix}\mathbf{x}.

Complex eigenvalues

  1. Find the general solution of \mathbf{x}' = \begin{bmatrix} 3 & -2 \\ 4 & -1 \end{bmatrix}\mathbf{x} in real form.
  2. Solve \mathbf{x}' = \begin{bmatrix} -2 & -5 \\ 1 & 2 \end{bmatrix}\mathbf{x}, \mathbf{x}(0) = (1, 0), and show that every solution of this system has period 2\pi.
  3. Find the general solution of \mathbf{x}' = \begin{bmatrix} 0 & -1 & 0 \\ 1 & 0 & 0 \\ 0 & 0 & -2 \end{bmatrix}\mathbf{x} in real form, and the solution with \mathbf{x}(0) = (1, 1, 1).

Forcing and models

  1. At the koi farm, Hiro joins a 2{,}000-liter pond and a 3{,}000-liter pond with pumps moving 240 liters per minute each way, and dissolves 150 grams of salt in the first pond. Find both amounts after t minutes and when the second pond holds 90\% of its final amount.
  2. Layla runs clean water at 100 liters per minute into a 1{,}000-liter tank holding 200 grams of dye; it overflows at the same rate into a 500-liter tank of clean water, which drains at 100 liters per minute. Find both amounts, and the largest amount of dye in the second tank and when it occurs.
  3. Find the constant solution of \mathbf{x}' = \begin{bmatrix} -4 & 2 \\ 1 & -3 \end{bmatrix}\mathbf{x} + \begin{bmatrix} 6 \\ 1 \end{bmatrix}, and solve with \mathbf{x}(0) = (3, 5).

Proofs

  1. Prove that if \mathbf{x} solves \mathbf{x}' = A\mathbf{x} on \mathbb{R}, then so do \mathbf{x}(t - s) for each constant s and \mathbf{x}'.
  2. Let A have n linearly independent eigenvectors whose eigenvalues are all negative. Prove that every solution of \mathbf{x}' = A\mathbf{x} tends to \mathbf{0} as t \to \infty.
  3. Let A\mathbf{v} = \lambda\mathbf{v} with \mathbf{v} \ne \mathbf{0}, and let \mathbf{x} solve \mathbf{x}' = A\mathbf{x} with \mathbf{x}(t_0) = \mathbf{v}. Prove that \mathbf{x}(t) = e^{\lambda(t - t_0)}\mathbf{v} for every t, so the solution never leaves the line of \mathbf{v} and never reaches \mathbf{0}.
  4. Let X(t) be a fundamental matrix of \mathbf{x}' = P(t)\mathbf{x} on I. Prove that X(t)\,C is one for every invertible constant matrix C, and that every fundamental matrix has this form.
  5. Show that (1, t) and (t, t^2) are linearly independent on \mathbb{R} although their Wronskian is 0 at every t, and conclude that they do not both solve any \mathbf{x}' = P(t)\mathbf{x} with P continuous on \mathbb{R}.
  6. Let every column of the real matrix A add to 0. Prove that x_1 + \dots + x_n is constant along every solution of \mathbf{x}' = A\mathbf{x}, and that 0 is an eigenvalue of A.
Answers
  1. A = \begin{bmatrix} 0 & 1 & 0 & 0 \\ -2 & 0 & 1 & 0 \\ 0 & 0 & 0 & 1 \\ 1 & 0 & -2 & 0 \end{bmatrix}. With u = v = \cos t, u'' = -\cos t = -2u + v, and likewise for v.
  2. \mathbf{x} = (3t^2 - 2t,\ 3t). Both rows check, W = \det\begin{bmatrix} t & t^2 \\ 0 & t \end{bmatrix} = t^2, and with \displaystyle \operatorname{tr} P = \frac{2}{t} Liouville gives W(1)\,e^{2\ln t} = t^2. The condition gives c_2 = 3, c_1 = -2.
  3. A = \begin{bmatrix} 1 & 1 \\ 4 & 1 \end{bmatrix}, which is PDP^{-1} with P = \begin{bmatrix} 1 & 1 \\ 2 & -2 \end{bmatrix}, D = \operatorname{diag}(3, -1), by Result 6(a).
  4. \mathbf{x} = c_1e^{2t}(1, 3) + c_2e^{4t}(1, 1), from (\lambda - 2)(\lambda - 4) = 0.
  5. \mathbf{x} = (3e^{-2t} - 2e^{-5t},\ 3e^{-2t} + e^{-5t}): eigenvalues -2, -5, eigenvectors (1, 1), (2, -1), constants 3, -1.
  6. \mathbf{x} = (-2 + 2e^{5t},\ 1 + 4e^{5t}): eigenvalues 0, 5, eigenvectors (2, -1), (1, 2), constants -1, 2.
  7. \mathbf{x} = (e^{t},\ -e^{t} + 4e^{3t},\ e^{3t} + e^{-t}). The matrix is triangular, with eigenvalues 1, 3, -1 and eigenvectors (1, -1, 0), (0, 4, 1), (0, 0, 1); all constants are 1.
  8. \displaystyle y = \frac{4}{3}e^{t} - \frac{1}{2}e^{2t} + \frac{1}{6}e^{-2t}. The companion matrix has last row (-4, 4, 1) and eigenvalues 1, 2, -2, with eigenvectors (1, \lambda, \lambda^2).
  9. \mathbf{x} = c_1e^{3t}\mathbf{v}_1 + c_2e^{3t}\mathbf{v}_2 + c_3e^{-t}\mathbf{v}_3 with \mathbf{v}_1 = (1, 0, 0), \mathbf{v}_2 = (0, 1, 1) and \mathbf{v}_3 = (0, 1, -1): the eigenvalue 3 has two independent eigenvectors.
  10. \mathbf{x} = c_1\mathbf{x}_1 + c_2\mathbf{x}_2 with \mathbf{x}_1 = e^{t}(\cos 2t,\ \cos 2t + \sin 2t) and \mathbf{x}_2 = e^{t}(\sin 2t,\ \sin 2t - \cos 2t), from \lambda = 1 + 2i, \mathbf{v} = (1, 1 - i).
  11. \mathbf{x} = (\cos t - 2\sin t,\ \sin t). For \lambda = i, \mathbf{v} = (-2 + i, 1) gives \mathbf{x}_1 = (-2\cos t - \sin t, \cos t) and \mathbf{x}_2 = (\cos t - 2\sin t, \sin t), with c_1 = 0, c_2 = 1. By Result 7 every solution is a combination of these, so has period 2\pi.
  12. \mathbf{x} = c_1\mathbf{x}_1 + c_2\mathbf{x}_2 + c_3\mathbf{x}_3 with \mathbf{x}_1 = (\cos t, \sin t, 0) and \mathbf{x}_2 = (\sin t, -\cos t, 0) from \mathbf{v} = (1, -i, 0) for \lambda = i, and \mathbf{x}_3 = e^{-2t}(0, 0, 1); they are independent at t = 0 (Result 4). The solution is (\cos t - \sin t,\ \sin t + \cos t,\ e^{-2t}).
  13. m_1 = 60 + 90e^{-t/5}, m_2 = 90 - 90e^{-t/5}; m_2 = 81 at t = 5\ln 10 \approx 11.51 minutes. The matrix \begin{bmatrix} -0.12 & 0.08 \\ 0.12 & -0.08 \end{bmatrix} has eigenvalues 0, -0.2 and eigenvectors (2, 3), (1, -1).
  14. m_1 = 200e^{-t/10}, m_2 = 200e^{-t/10} - 200e^{-t/5}, from m_1' = -0.1\,m_1, m_2' = 0.1\,m_1 - 0.2\,m_2. The maximum is 50 grams at t = 10\ln 2 \approx 6.93 minutes, where m_2' = 0 gives \displaystyle e^{-t/10} = \frac12.
  15. The constant solution is (2, 1), since A(2, 1) = (-6, -1). By Result 1 and Exercise 5, x_1 = 2 + 3e^{-2t} - 2e^{-5t} and x_2 = 1 + 3e^{-2t} + e^{-5t}.
  16. By the chain rule, \mathbf{z}(t) = \mathbf{x}(t - s) has \mathbf{z}'(t) = \mathbf{x}'(t - s) = A\mathbf{z}(t). Since \mathbf{x}' = A\mathbf{x} is differentiable, (\mathbf{x}')' = A\mathbf{x}'. \square
  17. By Result 6(b), \mathbf{x} = \sum_i c_ie^{\lambda_it}\mathbf{v}_i with each \lambda_i < 0, and each e^{\lambda_it} \to 0. \square
  18. By Result 6(a) and Result 1, e^{\lambda(t - t_0)}\mathbf{v} solves, with value \mathbf{v} at t_0, so it equals \mathbf{x} by Result 2. Its values are nonzero multiples of \mathbf{v}. \square
  19. (XC)' = X'C = P(XC), and \det(XC) = \det X\det C \ne 0, so by Result 4 the columns form a fundamental set. If Y is a fundamental matrix, Result 4 writes each column as X\mathbf{c}_j, so Y = XC with C = X(t_0)^{-1}Y(t_0), invertible since \det Y(t_0) \ne 0. \square
  20. W = t^2 - t^2 = 0. If c_1(1, t) + c_2(t, t^2) = \mathbf{0} for all t, the first entry c_1 + c_2t vanishes at t = 0 and t = 1, so c_1 = c_2 = 0. Were both solutions, Result 4 would make W never 0. \square
  21. With \mathbf{1} = (1, \dots, 1), the hypothesis is \mathbf{1}^{T}A = \mathbf{0}^{T}, so s = \mathbf{1}^{T}\mathbf{x} has s' = \mathbf{1}^{T}A\mathbf{x} = 0 and is constant. Also A^{T}\mathbf{1} = \mathbf{0}, so \det A = \det A^{T} = 0. \square
math.wiki · CC-BY-SA-4.0