Second-Order Linear Equations: General Theory, Vibrations and Forced Response
Pages 5 to 8 of the class notes, in their order: the second-order linear equation and its existence theorem, superposition and the general solution, the Wronskian, constant coefficients (Cases 01, 02, 03), inhomogeneous equations with the educated guess and the example $x'' + x' - 6x = f$ solved for three forcings, and the two strategies for variable coefficients: reduction of order and variation of constants. Four slips of the notes are corrected with the check that proves them; Robinson material beyond the notes is labelled as a complement.
For example, the equation of a spring: $p(t)$ is the friction of the system, $q(t)$ the spring constant and $f(t)$ the forcing. We have a basic well-posedness (existence and uniqueness) result in the style of Picard.
With a mass $m$, a damping coefficient $\mu$ and a spring stiffness $k$, Newton's law $m x'' = -\mu x' - kx + F(t)$ divided by $m$ is exactly this form with $p = \mu/m$, $q = k/m$ and $f = F/m$ (Section 5). Robinson starts from the general second-order equation:
Newton's second law is second order: the force fixes the acceleration $\ddot x$, not the velocity. A general second-order equation reads [Robinson, Eq. (11.1), p. 101]
$$\frac{d^2x}{dt^2} = f(\dot x, x, t). \tag{11.1}$$1.1 Why two initial conditions
Take the simplest case $\ddot x = f(t)$ (11.2), the motion of a particle under a prescribed force per unit mass. Integrate twice from $t_0$:
- $\dot x(t) - \dot x(t_0) = \displaystyle\int_{t_0}^{t}f(s)\,ds$ (Barrow's rule on $\ddot x = \frac{d}{dt}\dot x$). One constant, $\dot x(t_0)$, appears.
- $x(t) - x(t_0) = \displaystyle\int_{t_0}^{t}\dot x(s)\,ds = \dot x(t_0)(t - t_0) + \int_{t_0}^{t}\!\!\int_{t_0}^{s}f(r)\,dr\,ds$. A second constant, $x(t_0)$, appears.
Two integrations produce two arbitrary constants, so the position alone does not determine the motion: a piece of chalk can be dropped or thrown from the same point [p. 102].
1.2 Existence and uniqueness
Given $\ddot x = f(\dot x, x, t)$, if $f$, $\dfrac{\partial f}{\partial x}$ and $\dfrac{\partial f}{\partial \dot x}$ are continuous in a cube $\{(a_1, a_2) \times (b_1, b_2) \times (t_1, t_2)\}$, then the IVP with $x(t_0) = x_0$ and $\dot x(t_0) = y_0$ has a unique solution (if $t_0 \in (t_1, t_2)$, $x_0 \in (a_1, a_2)$ and $y_0 \in (b_1, b_2)$).
Robinson's statement of the same theorem, with $x_1 = x$ and $x_2 = \dot x$:
Suppose $f(x_2, x_1, t)$, $\partial f/\partial x_1$ and $\partial f/\partial x_2$ are continuous for $a_1 < x_1 < a_2$, $b_1 < x_2 < b_2$, $t_1 < t < t_2$. Then for all initial conditions
$$x(t_0) = x_0, \qquad \dot x(t_0) = y_0 \tag{11.4}$$with $a_1 < x_0 < a_2$, $b_1 < y_0 < b_2$, $t_1 < t_0 < t_2$, the equation $\ddot x = f(\dot x, x, t)$ has a unique solution, with two continuous derivatives, on some interval $I$ containing $t_0$.
Here $x_1$ stands for $x$ and $x_2$ for $\dot x$: the theorem is Theorem 6.2 applied to the equivalent first-order system $\dot x_1 = x_2$, $\dot x_2 = f(x_2, x_1, t)$ (Topic 4, Theorem 25.3). For linear equations with continuous coefficients the solution in fact exists on the whole interval where the coefficients are continuous (a standard result, not stated in the book); with constant coefficients, for all $t \in \mathbb R$ (Theorem 16.1, p. 150) — no blow-up is possible.
2. Superposition, independence and the general solution
If $x_1$ and $x_2$ are solutions to the linear homogeneous ODE $x'' + px' + qx = 0$, then any linear combination $\alpha x_1 + \beta x_2$ of $x_1$ and $x_2$ is also a solution.
Remark. If $f(t) = 0$ the equation is homogeneous; if $f(t) \neq 0$ it is inhomogeneous.
A second-order linear equation $a_2(t)\ddot x + a_1(t)\dot x + a_0(t)x = g(t)$ with $a_2(t) \neq 0$ can be divided by $a_2$ to give the standard form [Eqs. (11.6)–(11.7), pp. 102–103]
$$\ddot x + p(t)\,\dot x + q(t)\,x = f(t) \quad\text{(inhomogeneous)}, \qquad \ddot x + p(t)\,\dot x + q(t)\,x = 0 \quad\text{(homogeneous)}. \tag{11.6–11.7}$$Writing $L[x] = \ddot x + p\dot x + qx$ (11.10), the equations are $L[x] = f$ and $L[x] = 0$, and $L$ is linear (11.11):
$$L[\alpha x_1 + \beta x_2] = \alpha L[x_1] + \beta L[x_2].$$If $x_1$ and $x_2$ solve the homogeneous equation (11.7), so does $x = \alpha x_1 + \beta x_2$ for any constants $\alpha$, $\beta$.
Differentiation is linear: $\dot x = \alpha\dot x_1 + \beta\dot x_2$ and $\ddot x = \alpha\ddot x_1 + \beta\ddot x_2$. Therefore
$$\ddot x + p\dot x + qx = \alpha\big(\ddot x_1 + p\dot x_1 + qx_1\big) + \beta\big(\ddot x_2 + p\dot x_2 + qx_2\big) = \alpha\cdot 0 + \beta\cdot 0 = 0.$$The functions $x_1(t), x_2(t), \dots, x_n(t)$ are linearly independent on an interval $I$ if the only solution of $\alpha_1x_1 + \alpha_2x_2 + \dots + \alpha_nx_n = 0$ is $\alpha_1 = \alpha_2 = \dots = \alpha_n = 0$.
For two functions this means "not proportional": if $x_2 = cx_1$ then $cx_1 - x_2 = 0$ is a non-trivial combination; conversely, if $\alpha_1x_1 + \alpha_2x_2 = 0$ with $\alpha_1 \neq 0$, then $x_1 = -(\alpha_2/\alpha_1)x_2$ [pp. 104–105]. So $e^{t}$ and $e^{2t}$ are independent, while $\sin 2t$ and $\sin t\cos t$ are not ($\sin 2t = 2\sin t\cos t$).
Any solution to $x'' + px' + qx = 0$ can be written as a linear combination $x(t) = C_1x_1(t) + C_2x_2(t)$, where $x_1$ and $x_2$ are linearly independent solutions of the same problem. By the existence and uniqueness theorem it is easy to construct two linearly independent solutions.
The example of the notes. To prove $x(t) = C_1x_1(t) + C_2x_2(t)$, take the IVP $x'' + px' + qx = 0$, $x(t_0) = x_0$, $x'(t_0) = y_0$, where $x_0$ and $y_0$ can vary. Imposing both conditions on $C_1x_1 + C_2x_2$ gives the linear system
$$\begin{pmatrix} x_1(t_0) & x_2(t_0) \\ x_1'(t_0) & x_2'(t_0)\end{pmatrix}\begin{pmatrix} C_1 \\ C_2\end{pmatrix} = \begin{pmatrix} x_0 \\ y_0\end{pmatrix},$$which has a solution for every $x_0$, $y_0$ only when the determinant is non-zero [Notes, p. 5]. That determinant is the Wronskian of Section 3. Robinson's proof of the theorem:
Two independent solutions exist. By Theorem 11.1 there are solutions with $x_1(t_0) = 1$, $\dot x_1(t_0) = 0$ and $x_2(t_0) = 0$, $\dot x_2(t_0) = 1$. If $x_2 = cx_1$ then $x_2(t_0) = c$ forces $c = 0$, but then $\dot x_2(t_0) = 0 \neq 1$. So they are not proportional.
Two are enough. Let $x_1$, $x_2$ be any two independent solutions and $x$ any solution, with $x(t_0) = x_0$, $\dot x(t_0) = v_0$. We want $\alpha$, $\beta$ with $\alpha x_1 + \beta x_2$ matching these data:
$$\begin{pmatrix} x_1(t_0) & x_2(t_0) \\ \dot x_1(t_0) & \dot x_2(t_0)\end{pmatrix}\begin{pmatrix}\alpha \\ \beta\end{pmatrix} = \begin{pmatrix} x_0 \\ v_0\end{pmatrix}. \tag{11.15}$$This is solvable for every right-hand side if and only if the determinant $x_1(t_0)\dot x_2(t_0) - x_2(t_0)\dot x_1(t_0)$ is non-zero. Suppose instead it were zero (11.16). Then the columns are proportional: there is a constant $c$ with $x_1(t_0) = cx_2(t_0)$ and $\dot x_1(t_0) = c\dot x_2(t_0)$ (11.17). (Robinson divides by $x_2(t_0)$ and $\dot x_2(t_0)$ to find $c$; they cannot both vanish, because then uniqueness would force $x_2 \equiv 0$, and if one vanishes the same conclusion follows directly — footnote, p. 106.) By superposition $y = cx_2$ is a solution, and it has the same data at $t_0$ as $x_1$. By uniqueness $x_1 = cx_2$ for all $t$, contradicting independence. Hence the determinant is non-zero, (11.15) has a solution, and by uniqueness $x = \alpha x_1 + \beta x_2$ everywhere.
$L$ maps the vector space $C^2(I)$ of twice continuously differentiable functions to $C^0(I)$. The solution set of $L[x] = 0$ is its kernel, a vector space, and the theorem says $\dim\ker L = 2$ with basis $\{x_1, x_2\}$. For an $n$th-order equation, $\dim\ker L = n$ (Sec. 16.2).
We find a solution for $C_1$ and $C_2$ if $W(x_1, x_2) \neq 0$.
3.1 What equation does $W$ satisfy?
Let $x_1$ and $x_2$ solve $x'' + p(t)x' + q(t)x = 0$ [Notes, p. 5].
- Product rule on both products: $$W' = \big(x_1'x_2' + x_1x_2''\big) - \big(x_2'x_1' + x_2x_1''\big) = x_1x_2'' - x_2x_1''.$$ The terms $x_1'x_2'$ cancel.
- Each $x_j$ is a solution, so $x_j'' = -px_j' - qx_j$. Substituting, $$W' = x_1\big(-px_2' - qx_2\big) - x_2\big(-px_1' - qx_1\big) = -p\,x_1x_2' + p\,x_2x_1' - q\,x_1x_2 + q\,x_2x_1.$$
- The $q$ terms cancel, and $-p\,x_1x_2' + p\,x_2x_1' = -p\,(x_1x_2' - x_2x_1')$:
The notes box the result as $W'(t) = p(t)\,W(t)$, without the minus sign, but their own last line before it reads $-p\,x_1x_2' + p\,x_2x_1'$, which is $-p\,W$. The solution the notes give next, $W(t) = Ce^{-\int p(t)\,dt}$, is the solution of $W' = -pW$, so only the boxed line is affected.
Check. For $x'' + x' - 6x = 0$ ($p = 1$) take $x_1 = e^{-3t}$, $x_2 = e^{2t}$: $W = e^{-3t}\cdot 2e^{2t} - e^{2t}\cdot(-3e^{-3t}) = 5e^{-t}$. Then $W' = -5e^{-t} = -1\cdot W$ ✓, while $p\,W = +5e^{-t} \neq W'$.
$W' = -pW$ is a first-order linear (and separable) equation, solved as in Topic 2: $W'/W = -p$, $\ln|W| = -\int p\,dt + c$, so
The exponential is never zero. So, as the notes conclude (also by Picard's theorem, since $W \equiv 0$ is the solution of $W' = -pW$ through $W(t_0) = 0$): either $W \equiv 0$ (everywhere) or $W \neq 0$ (never zero) — either $0$ for all $t$, or $\neq 0$ for all $t$. One evaluation at a single time decides independence on the whole interval. With constant $p$, $W(t) = W(t_0)\,e^{-p(t - t_0)}$.
3.2 Complement (Robinson): the Wronskian test
- For any two differentiable functions: if $W[x_1, x_2](t_0) \neq 0$ at some $t_0 \in I$, then $x_1$, $x_2$ are linearly independent on $I$.
- For two solutions of $\ddot x + p\dot x + qx = 0$: they are linearly independent if and only if $W \neq 0$; and then $W(t) \neq 0$ for every $t \in I$.
(1) Suppose $\alpha x_1(t) + \beta x_2(t) = 0$ for all $t \in I$. Differentiating, $\alpha\dot x_1(t) + \beta\dot x_2(t) = 0$ too. At $t = t_0$ these two equations read $\begin{pmatrix} x_1 & x_2 \\ \dot x_1 & \dot x_2\end{pmatrix}\begin{pmatrix}\alpha\\\beta\end{pmatrix} = \mathbf 0$ with a matrix of determinant $W(t_0) \neq 0$, which is invertible; multiplying by its inverse gives $\alpha = \beta = 0$.
(2) "If": part (1). "Only if": the proof of the previous theorem showed that if $W(t_0) = 0$ at any single point (equation (11.16)) the solutions are proportional. So for independent solutions $W$ never vanishes; for dependent ones it vanishes identically.
$f_1 = t^2|t|$ and $f_2 = t^3$ are linearly independent on $\mathbb R$ (they agree for $t > 0$ and are opposite for $t < 0$, so no single constant relates them), yet $\dot f_1 = 3t|t|$ and $W = t^2|t|\cdot 3t^2 - t^3\cdot 3t|t| = 0$ for every $t$. A zero Wronskian proves dependence only for solutions of a linear ODE.
4. Homogeneous equations with constant coefficients
Consider $x'' + px' + qx = 0$ for $p, q \in \mathbb R$. We use the previous theorem to find the general solution, since any solution can be written as a linear combination of two independent ones. We look for solutions of the type $x(t) = e^{kt}$, with $k$ to be determined.
By the chain rule $(e^{kt})' = ke^{kt}$ and $(e^{kt})'' = k^2e^{kt}$. Substituting:
$$(e^{kt})'' + p\,(e^{kt})' + q\,e^{kt} = k^2e^{kt} + pk\,e^{kt} + q\,e^{kt} = \big(k^2 + pk + q\big)e^{kt} = 0.$$Since $e^{kt} \neq 0$, the characteristic polynomial must vanish:
$$k^2 + pk + q = 0, \qquad k = \frac{-p \pm \sqrt{p^2 - 4q}}{2}.$$The sign of $p^2 - 4q$ gives the three cases of the notes [Notes, p. 5].
4.1 Case 01: two distinct real roots, $p^2 > 4q$
$x_1(t) = e^{k_1t}$ and $x_2(t) = e^{k_2t}$ are solutions. Their Wronskian is $W = e^{k_1t}\,k_2e^{k_2t} - e^{k_2t}\,k_1e^{k_1t} = (k_2 - k_1)\,e^{(k_1 + k_2)t} \neq 0$, so they are independent and
$$x(t) = C_1e^{k_1t} + C_2e^{k_2t}.$$$k^2 + k - 6 = (k - 2)(k + 3) = 0$, so $k = 2, -3$ and $x = Ae^{2t} + Be^{-3t}$, $\dot x = 2Ae^{2t} - 3Be^{-3t}$. At $t = 0$: $A + B = 1$, $2A - 3B = 2$. Subtracting twice the first, $-5B = 0$, so $B = 0$, $A = 1$: $\ x(t) = e^{2t}$.
If one root is $0$ (e.g. $\ddot x - 2\dot x = 0$, $k = 0, 2$), the corresponding solution is the constant $e^{0t} = 1$: $x = A + Be^{2t}$ (Example 12.2).
4.2 Case 02: one repeated real root, $p^2 = 4q$
The root is $k = -p/2$, so $2k + p = 0$, and the notes give [Notes, p. 5]
$$x(t) = C_1e^{kt} + C_2\,t\,e^{kt}.$$Why $te^{kt}$ is a solution. By the product rule, $(te^{kt})' = (1 + kt)e^{kt}$ and $(te^{kt})'' = (2k + k^2t)e^{kt}$. Then
$$(te^{kt})'' + p\,(te^{kt})' + q\,te^{kt} = \Big[\big(k^2 + pk + q\big)\,t + \big(2k + p\big)\Big]e^{kt} = \big[0\cdot t + 0\big]e^{kt} = 0,$$because $k$ is a root and $2k + p = 0$ for the double root. Wronskian: $e^{kt}(1 + kt)e^{kt} - te^{kt}\,ke^{kt} = e^{2kt} \neq 0$. Section 8.1 derives $te^{kt}$ by reduction of order.
$(k + 1)^2 = 0$, $k = -1$ twice. $x = (A + Bt)e^{-t}$, $\dot x = (B - A - Bt)e^{-t}$. Then $A = 0$ and $B - A = 1$: $\ x(t) = te^{-t}$.
4.3 Case 03: complex conjugate roots, $p^2 < 4q$
The roots are [Notes, p. 5]
$$k_{1,2} = -\frac{p}{2} \pm \frac{\sqrt{4q - p^2}}{2}\,i = \rho \pm \omega i, \qquad \rho = -\frac p2,\quad \omega = \frac{\sqrt{4q - p^2}}{2}.$$- The two exponentials are solutions (the computation above holds for complex $k$): $$x(t) = C_1e^{(\rho + \omega i)t} + C_2e^{(\rho - \omega i)t}.$$
- Euler's formula $e^{i\theta} = \cos\theta + i\sin\theta$, with $\cos(-\omega t) = \cos\omega t$ and $\sin(-\omega t) = -\sin\omega t$: $$x(t) = C_1e^{\rho t}\big(\cos\omega t + i\sin\omega t\big) + C_2e^{\rho t}\big(\cos(-\omega t) + i\sin(-\omega t)\big) = (C_1 + C_2)\,e^{\rho t}\cos\omega t + i\,(C_1 - C_2)\,e^{\rho t}\sin\omega t.$$
- If we choose $C_1$ and $C_2$ complex conjugates, $C_1 = \tfrac12(A - iB)$ and $C_2 = \overline{C_1} = \tfrac12(A + iB)$ with $A, B$ real, then $C_1 + C_2 = A$ and $i(C_1 - C_2) = i\,(-iB) = B$ are real, and
$\rho < 0$: decaying oscillation; $\rho = 0$: pure oscillation (simple harmonic motion); $\rho > 0$: growing oscillation. The Wronskian of $e^{\rho t}\cos\omega t$ and $e^{\rho t}\sin\omega t$ is $\omega e^{2\rho t} \neq 0$ (Robinson Exercise 11.1 (iii)).
$k = \frac{-2\pm\sqrt{4 - 20}}{2} = -1 \pm 2i$, so $x = e^{-t}(A\cos 2t + B\sin 2t)$ and, by the product rule, $\dot x = e^{-t}\big[(2B - A)\cos 2t - (2A + B)\sin 2t\big]$. Then $A = 1$, $2B - A = 0$, so $B = \tfrac12$: $\ x(t) = e^{-t}\big(\cos 2t + \tfrac12\sin 2t\big)$.
| Roots of $k^2 + pk + q = 0$ | Fundamental pair | General solution |
|---|---|---|
| Case 01: real, $k_1 \neq k_2$ | $e^{k_1t}$, $e^{k_2t}$ | $C_1e^{k_1t} + C_2e^{k_2t}$ |
| Case 02: repeated, $k = -p/2$ | $e^{kt}$, $te^{kt}$ | $C_1e^{kt} + C_2te^{kt}$ |
| Case 03: complex, $\rho \pm i\omega$ | $e^{\rho t}\cos\omega t$, $e^{\rho t}\sin\omega t$ | $e^{\rho t}(A\cos\omega t + B\sin\omega t)$ |
For $a_nx^{(n)} + \dots + a_1\dot x + a_0x = 0$ the trial $e^{kt}$ gives $a_nk^n + \dots + a_0 = 0$. Each simple real root contributes $e^{kt}$; a root repeated $m$ times contributes $e^{kt}, te^{kt}, \dots, t^{m-1}e^{kt}$; each complex pair $\rho \pm i\omega$ contributes $e^{\rho t}\cos\omega t$, $e^{\rho t}\sin\omega t$ (times $1, t, \dots$ if repeated). The homogeneous part of Example 16.2 ($x^{(4)} - 16x = 64\sin 2t$): $x^{(4)} - 16x = 0$ has $k^4 = 16$, $k = \pm 2, \pm 2i$, so $x = Ae^{2t} + Be^{-2t} + C\cos 2t + D\sin 2t$.
5. Complement: the mass–spring–damper
The notes name the spring as the model behind $x'' + px' + qx = f$. Robinson develops it in full; the three cases of Section 4 become the three damping regimes.
5.1 Model
Hypotheses. A rigid mass $m$ moves along one axis (one degree of freedom $x$, measured from equilibrium). The spring obeys Hooke's law, restoring force $-kx$; the dashpot gives a force proportional and opposite to the velocity, $-\mu\dot x$, with $\mu > 0$ [Robinson Sec. 13.1, p. 120; Sec. 13.3, p. 123]. Newton's second law:
$$m\ddot x = -kx - \mu\dot x + F(t) \quad\Longleftrightarrow\quad m\ddot x + \mu\dot x + kx = F(t).$$| Symbol | Meaning | SI unit |
|---|---|---|
| $m$ | Mass | kg |
| $k$ | Spring stiffness | N/m = kg/s² |
| $\mu$ | Damping coefficient | N s/m = kg/s |
| $\omega = \sqrt{k/m}$ | Natural angular frequency | rad/s |
| $\lambda = \mu/m$ | Damping rate | s⁻¹ |
Dividing by $m$, the free ($F = 0$) equation is [Eq. (13.6), p. 124]
$$\ddot x + \lambda\dot x + \omega^2x = 0, \qquad \lambda = \frac{\mu}{m} > 0, \quad \omega^2 = \frac{k}{m}. \tag{13.6}$$The text says "dividing by $m$ … and setting $\lambda = \mu/k$". Dividing $m\ddot x + \mu\dot x + kx = 0$ by $m$ gives the coefficient $\mu/m$ for $\dot x$, so the correct definition is $\lambda = \mu/m$. Dimensional check: $[\mu/m] = (\text{kg/s})/\text{kg} = \text{s}^{-1}$, as required for $\lambda\dot x$ to have the units of $\ddot x$; $[\mu/k] = (\text{kg/s})/(\text{kg/s}^2) = \text{s}$ would not.
5.2 No damping: simple harmonic motion
With $\lambda = 0$: $\ddot x = -\omega^2x$ (13.1), roots $\pm i\omega$, so $x = A\cos\omega t + B\sin\omega t$ (13.2). Writing $M = \sqrt{A^2 + B^2}$ and choosing $\phi$ with $\cos\phi = A/M$, $\sin\phi = B/M$:
$$x = M\big(\cos\phi\cos\omega t + \sin\phi\sin\omega t\big) = M\cos(\omega t - \phi). \tag{13.3}$$The motion has amplitude $M$, period $2\pi/\omega = 2\pi\sqrt{m/k}$ independent of the amplitude, and lags $M\cos\omega t$ by the time $\phi/\omega$. The same equation governs the small oscillations of a pendulum, $\ddot\theta = -(g/L)\theta$, with period $2\pi\sqrt{L/g}$ (Sec. 13.2).
5.3 The three damping regimes [Sec. 13.3, pp. 124–126]
The auxiliary equation $k^2 + \lambda k + \omega^2 = 0$ has roots $k = \tfrac12\big(-\lambda \pm\sqrt{\lambda^2 - 4\omega^2}\big)$.
- Over-damped, $\lambda > 2\omega$: two real roots $-k_1$, $-k_2$, both negative because $0 < \sqrt{\lambda^2 - 4\omega^2} < \lambda$. $x = Ae^{-k_1t} + Be^{-k_2t}$: decay with no oscillation.
- Critically damped, $\lambda = 2\omega$: repeated root $-\lambda/2 = -\omega$. $x = (A + Bt)e^{-\omega t}$: no oscillation, and the fastest return to equilibrium; any less damping allows oscillation.
- Under-damped, $\lambda < 2\omega$: roots $-\lambda/2 \pm i\sigma$ with $\sigma = \tfrac12\sqrt{4\omega^2 - \lambda^2}$. $$x = e^{-\lambda t/2}\big(A\cos\sigma t + B\sin\sigma t\big) = Me^{-\lambda t/2}\cos(\sigma t - \phi). \tag{13.7}$$ The amplitude decays exponentially and successive maxima are separated by the constant time $2\pi/\sigma$: setting $\dot x = -Me^{-\lambda t/2}\big[\tfrac\lambda2\cos(\sigma t - \phi) + \sigma\sin(\sigma t - \phi)\big] = 0$ gives $\tan(\sigma t - \phi) = -\lambda/(2\sigma)$, whose solutions are spaced by $\pi/\sigma$ (alternating maxima and minima).
Engineers write $\lambda = 2\zeta\omega$, i.e. $\zeta = \lambda/(2\omega) = \mu/(2\sqrt{km})$, the damping ratio: $\zeta < 1$ under-damped, $\zeta = 1$ critical, $\zeta > 1$ over-damped. The critical damping coefficient is $\mu_c = 2\sqrt{km}$ and the damped frequency is $\sigma = \omega\sqrt{1 - \zeta^2}$.
Model $M\ddot x + k\dot x + \lambda x = 0$ with $M \approx 4\times10^5\ \text{kg}$, damping $k \approx 5\times10^4\ \text{kg/s}$, stiffness $\lambda \approx 10^7\ \text{kg/s}^2$ (the exercise uses $k$ and $\lambda$ the other way round from Sec. 13.3). Natural frequency $\sqrt{\lambda/M} = \sqrt{25} = 5\ \text{rad/s}$, i.e. $5/(2\pi) \approx 0.8\ \text{Hz}$. Critical damping would be $2\sqrt{M\lambda} = 2\sqrt{4\times10^{12}} = 4\times10^6\ \text{kg/s}$, so the actual damping is $5\times10^4/4\times10^6 \approx 1.3\%$ of critical. Pedestrians walking in step added a force $\approx 300\dot x$ each, i.e. negative damping: with enough people the net coefficient of $\dot x$ changes sign and oscillations grow.
A landing-gear leg carrying $m = 2000\ \text{kg}$ on an oleo of stiffness $k = 2\times10^5\ \text{N/m}$ has $\omega = \sqrt{k/m} = 10\ \text{rad/s}$ and $\mu_c = 2\sqrt{km} = 2\sqrt{4\times10^8} = 4\times10^4\ \text{N s/m}$. Designers choose $\zeta$ of the order of $0.5$–$0.7$ to avoid repeated bouncing (under-damping) without making the strut too stiff in compression (over-damping). With $\zeta = 0.6$: $\lambda = 12\ \text{s}^{-1}$, $\sigma = 10\sqrt{1 - 0.36} = 8\ \text{rad/s}$, and the amplitude falls by $e^{-\lambda T/2} = e^{-6\cdot 2\pi/8} \approx 0.009$ per damped period $T = 2\pi/\sigma$.
6. Inhomogeneous equations with constant coefficients
6.1 Structure of the general solution
The general solution $x(t)$ of $(\ast)$ is the sum of the general solution $x_H(t)$ of the associated homogeneous problem ($f \equiv 0$) and a particular solution $x_N(t)$ of $(\ast)$:
$$x(t) = x_H(t) + x_N(t).$$Proof [Notes, p. 6]. Let $x$ be any solution of $(\ast)$ and take $\Phi(t) = x(t) - x_N(t)$, so $\Phi' = x' - x_N'$ and $\Phi'' = x'' - x_N''$. Then
$$\Phi'' + p\Phi' + q\Phi = \big(x'' - x_N''\big) + p\big(x' - x_N'\big) + q\big(x - x_N\big) = \big(x'' + px' + qx\big) - \big(x_N'' + px_N' + qx_N\big) = f(t) - f(t) = 0.$$Since $\Phi$ satisfies the homogeneous equation, $\Phi = x_H$ for some choice of the constants (Section 2), and $x = \Phi + x_N = x_H + x_N$.
The last line of the proof reads "$\Phi(t) = x_H(t) = x(t) - x_N(t) \Rightarrow x(t) = x_H(t) - x_N(t)$". Solving $x_H = x - x_N$ for $x$ gives $x = x_H + x_N$, as stated at the top of the same page. (The middle line also writes $q(x - x_N')$ for $q(x - x_N)$.)
Check. With $f = 12$ in the example of Section 6.3, $x_N = -2$. The function $x_H - x_N$ with $x_H = 0$ is $x = 2$, and $x'' + x' - 6x = -12 \neq 12$; the function $x_H + x_N = -2$ gives $-6(-2) = 12$ ✓.
Robinson states the same result as a lemma, calling $x_H$ the complementary function and $x_N$ a particular integral:
Let $y(t) = Ay_1(t) + By_2(t)$ be the general solution of $L[y] = 0$ (the complementary function) and $x_p$ any one solution of $L[x_p] = f$ (a particular integral). Then the general solution of $L[x] = f$ is
$$x(t) = Ay_1(t) + By_2(t) + x_p(t).$$It is a solution: by linearity $L[y + x_p] = L[y] + L[x_p] = 0 + f = f$. Every solution has this form: if $L[x] = f$, then $L[x - x_p] = L[x] - L[x_p] = f - f = 0$, so $x - x_p$ solves the homogeneous equation and equals $Ay_1 + By_2$ for some $A$, $B$ (Section 2). The two constants fit any initial conditions.
6.2 The educated guess
To find $x_N(t)$ we make an educated guess, with trial and error, based on the form of $f(t)$; then we plug $x_N$ into the full equation and solve for its coefficients [Notes, p. 6].
- Polynomial of order $n$: $\ x_N(t) = C_nt^n + C_{n-1}t^{n-1} + \dots + C_2t^2 + C_1t + C_0$ (all powers).
- $f(t) = Ce^{\alpha t}$. Case 1: if $\alpha$ is not a root of $k^2 + pk + q = 0$, then $x_N(t) = Ae^{\alpha t}$. Case 2: if $\alpha$ is a root, then $x_N(t) = Ate^{\alpha t}$.
- $f(t) = \alpha\cos\omega t + \beta\sin\omega t$: $x_N(t) = A\cos\omega t + B\sin\omega t$; if it is already in $x_H(t)$, then $x_N(t) = At\cos\omega t + Bt\sin\omega t$.
- Sums: if $f$ is the sum of two or more terms, solve for each term and add ("decomposition").
- Higher order: the procedure also works for higher-order equations: get the $n$ roots, build $x_H$ from terms $C_ne^{k_nt}$, and find $x_N$ with the same guesses.
Why the rules work. Derivatives of a polynomial of degree $n$ are polynomials of degree $\le n$, derivatives of $e^{\alpha t}$ are multiples of $e^{\alpha t}$, and derivatives of $\cos\omega t$, $\sin\omega t$ are combinations of the same two functions. So $x_N'' + px_N' + qx_N$ has the same form as the guess, and matching coefficients gives linear equations. If a term of the guess solves the homogeneous equation it contributes $0$ and cannot match $f$: that is why the guess is multiplied by $t$ ("Case 2", resonance).
If $\alpha$ is a double root of $k^2 + pk + q = 0$, both $e^{\alpha t}$ and $te^{\alpha t}$ solve the homogeneous equation and the guess must be $At^2e^{\alpha t}$. Robinson's rule of thumb covers every case: multiply by $t$ until no term of the guess solves the homogeneous equation. The table below collects his guesses.
| Forcing $f(t)$ | Standard guess $x_p$ | If a term of the guess solves $L[y] = 0$ |
|---|---|---|
| Polynomial of degree $n$ | $c_nt^n + \dots + c_1t + c_0$ (all powers) | Multiply by $t$ (e.g. root $k = 0$) |
| $ce^{\kappa t}$ | $Ce^{\kappa t}$ | $Cte^{\kappa t}$; $Ct^2e^{\kappa t}$ if $\kappa$ is a repeated root |
| $\alpha\sin\sigma t + \beta\cos\sigma t$ | $C\sin\sigma t + D\cos\sigma t$ (both, always) | $t(C\sin\sigma t + D\cos\sigma t)$ — resonance |
| Sum $\alpha f_1 + \beta f_2$ | $\alpha x_1 + \beta x_2$ with $L[x_j] = f_j$ | Treat each part separately (14.7) |
| Products, e.g. $te^{-t}\cos 2t$ | Product of the guesses: $(At + B)e^{-t}\cos 2t + (Ct + D)e^{-t}\sin 2t$ | As above |
Take the standard guess. If it contains terms that satisfy the homogeneous equation, multiply the guess by $t$; repeat until no term of the guess solves the homogeneous equation.
6.3 The example of the notes: $x'' + x' - 6x = f(t)$
Step 1: find $x_H(t)$ [Notes, p. 6]. Here $p = 1$, $q = -6$. The characteristic polynomial is $k^2 + pk + q = k^2 + k - 6 = (k + 3)(k - 2) = 0$, so $k = -3, 2$ (Case 01) and
$$x_H(t) = C_1e^{-3t} + C_2e^{2t}.$$Step 2: find $x_N(t)$ for three forcings.
- A constant is a polynomial of degree $0$: define $x_N(t) = C$. Then $x_N' = 0$ and $x_N'' = 0$.
- Plug in: $0 + 0 - 6C = 12$, so $C = -2$ and $x_N(t) = -2$.
- General solution: $\ x(t) = C_1e^{-3t} + C_2e^{2t} - 2$.
Check: $(-2)'' + (-2)' - 6(-2) = 12$ ✓. As the constant is the equilibrium of the forced equation, $x \to -2$ only along $C_2 = 0$; any $C_2 \neq 0$ makes $e^{2t}$ dominate.
- Degree-1 polynomial: $x_N(t) = a_1t + a_0$ (the notes call the coefficients $C_1$, $C_0$; other letters avoid a clash with the constants of $x_H$). Then $x_N' = a_1$ and $x_N'' = 0$.
- Plug in: $0 + a_1 - 6(a_1t + a_0) = 36t$, i.e. $-6a_1\,t + (a_1 - 6a_0) = 36t + 0$.
- Match the coefficient of $t$: $-6a_1 = 36$, so $a_1 = -6$. Match the constants: $a_1 - 6a_0 = 0$, so $a_0 = a_1/6 = -1$.
- $x_N(t) = -6t - 1$ and $\ x(t) = C_1e^{-3t} + C_2e^{2t} - 6t - 1$.
Check: $0 + (-6) - 6(-6t - 1) = -6 + 36t + 6 = 36t$ ✓. (The notes write the first matching as $6C_1t = 36t$ before concluding $C_1 = -6$; the coefficient of $t$ is $-6C_1$, and the value $-6$ is right.)
- $\alpha = -2$ is not a root ($k = -3, 2$): Case 1, $x_N = Ae^{-2t}$, with $x_N' = -2Ae^{-2t}$ and $x_N'' = 4Ae^{-2t}$ (chain rule).
- Plug in: $4Ae^{-2t} - 2Ae^{-2t} - 6Ae^{-2t} = -4Ae^{-2t} = 4e^{-2t}$, so $A = -1$.
- $x_N(t) = -e^{-2t}$ and $\ x(t) = C_1e^{-3t} + C_2e^{2t} - e^{-2t}$.
Check: $(-4 + 2 + 6)\,e^{-2t} = 4e^{-2t}$ ✓.
By linearity the particular solutions add: $x_N = -2 + (-6t - 1) + (-e^{-2t}) = -6t - 3 - e^{-2t}$. Check: $x_N' = -6 + 2e^{-2t}$, $x_N'' = -4e^{-2t}$, and $x_N'' + x_N' - 6x_N = (-4 + 2 + 6)e^{-2t} - 6 + 36t + 18 = 4e^{-2t} + 12 + 36t$ ✓.
6.4 Complement (Robinson): more examples
The complementary function is $Ae^{2t} + Be^{-3t}$. Try the full cubic $x_p = Ct^3 + Dt^2 + Et + F$, so $\dot x_p = 3Ct^2 + 2Dt + E$ and $\ddot x_p = 6Ct + 2D$. Substituting and collecting powers:
$$-6Ct^3 + (3C - 6D)t^2 + (6C + 2D - 6E)t + (2D + E - 6F) = 216t^3.$$Matching coefficients from the top: $-6C = 216 \Rightarrow C = -36$; $3C - 6D = 0 \Rightarrow D = -18$; $6C + 2D - 6E = 0 \Rightarrow E = (-216 - 36)/6 = -42$; $2D + E - 6F = 0 \Rightarrow F = (-36 - 42)/6 = -13$. Hence $x = Ae^{2t} + Be^{-3t} - 36t^3 - 18t^2 - 42t - 13$. A bare $Ct^3$ would fail: its derivatives generate $t^2$, $t$ and constant terms that must be cancelled.
$e^{-3t}$ already solves the homogeneous equation, so $Ce^{-3t}$ gives $0 = 5e^{-3t}$. Multiply by $t$: $x_p = Cte^{-3t}$, with $\dot x_p = C(1 - 3t)e^{-3t}$ and $\ddot x_p = C(-6 + 9t)e^{-3t}$ (product rule twice). Then
$$\ddot x_p + \dot x_p - 6x_p = C\big[(-6 + 9t) + (1 - 3t) - 6t\big]e^{-3t} = -5Ce^{-3t} = 5e^{-3t} \;\Rightarrow\; C = -1.$$The $te^{-3t}$ terms cancel, as they must. $x = Ae^{2t} + Be^{-3t} - te^{-3t}$.
Complementary function $(A + Bt)e^{-t}$. Try $x_p = C\sin 2t + D\cos 2t$ (both terms: the $\dot x$ term turns cosines into sines). $\dot x_p = 2C\cos 2t - 2D\sin 2t$, $\ddot x_p = -4C\sin 2t - 4D\cos 2t$, so
$$\ddot x_p + 2\dot x_p + x_p = (4C - 3D)\cos 2t - (3C + 4D)\sin 2t = 100\cos 2t.$$$3C + 4D = 0$ and $4C - 3D = 100$ give $C = 16$, $D = -12$ (check: $64 + 36 = 100$, $48 - 48 = 0$). So $x = (A + Bt)e^{-t} + 16\sin 2t - 12\cos 2t$. As $t \to \infty$ the transient dies and the steady oscillation has amplitude $\sqrt{16^2 + 12^2} = 20$.
$\cos t$ and $\sin t$ solve the homogeneous equation, so try $x_p = t(C\sin t + D\cos t)$. Differentiating twice with the product rule, $\ddot x_p = 2C\cos t - 2D\sin t - t(C\sin t + D\cos t)$, so $\ddot x_p + x_p = 2C\cos t - 2D\sin t = 8\cos t$: $C = 4$, $D = 0$. $x = A\sin t + B\cos t + 4t\sin t$: the forced part grows linearly in time.
7. Complement: forced vibrations and resonance
The resonance rule of the notes ("if it is already in $x_H$, multiply by $t$") has a physical meaning, developed by Robinson.
Force the oscillator at angular frequency $\alpha$: $\ \ddot x + \omega^2x = a\cos\alpha t$ (15.3).
- $\alpha \neq \omega$ [Sec. 15.1.1]: try $x_p = C\cos\alpha t$ (no sine needed — there is no $\dot x$ term): $(-\alpha^2 + \omega^2)C = a$, so $x = M\cos(\omega t - \phi) + \dfrac{a}{\omega^2 - \alpha^2}\cos\alpha t$ (15.4). Bounded, but large when $\alpha \approx \omega$. From rest, $x = \dfrac{\cos\alpha t - \cos\omega t}{\omega^2 - \alpha^2}\cdot a = \dfrac{2a}{\omega^2 - \alpha^2}\sin\dfrac{(\omega + \alpha)t}{2}\sin\dfrac{(\omega - \alpha)t}{2}$: a fast oscillation modulated by a slow envelope — beats (Exercises 15.1–15.2).
- $\alpha = \omega$ [Sec. 15.1.2]: the rule of thumb gives $x_p = \dfrac{a}{2\omega}\,t\sin\omega t$: the amplitude grows without bound. This is resonance (Robinson recalls the 1831 Broughton bridge collapse under marching troops). The same result follows from the beats formula by letting $\alpha \to \omega$ with L'Hôpital's rule (Ex. 15.3).
7.1 Damped forcing: the frequency response [Sec. 15.2, pp. 145–148]
For $\ddot x + \lambda\dot x + \omega^2x = a\cos\alpha t$ with $\lambda > 0$, try $x_p = C\sin\alpha t + D\cos\alpha t$. Matching sines and cosines:
$$C(\omega^2 - \alpha^2) = D\lambda\alpha, \qquad D(\omega^2 - \alpha^2) + C\lambda\alpha = a.$$Solving (substitute $C$ from the first into the second):
$$C = \frac{a\lambda\alpha}{(\omega^2 - \alpha^2)^2 + \lambda^2\alpha^2}, \qquad D = \frac{a(\omega^2 - \alpha^2)}{(\omega^2 - \alpha^2)^2 + \lambda^2\alpha^2}, \qquad x_p = C\sin\alpha t + D\cos\alpha t. \tag{15.5}$$The complementary function decays (all roots have negative real part), so after a transient only $x_p$ remains, with amplitude
$$R(\alpha) = \sqrt{C^2 + D^2} = \frac{a}{\sqrt{(\omega^2 - \alpha^2)^2 + \lambda^2\alpha^2}} = a\,F(\alpha), \qquad F(\alpha) = \frac{1}{\sqrt{(\omega^2 - \alpha^2)^2 + \lambda^2\alpha^2}}. \tag{15.6}$$The denominator is smallest where $\frac{d}{d\alpha}\big[(\omega^2 - \alpha^2)^2 + \lambda^2\alpha^2\big] = -4\alpha(\omega^2 - \alpha^2) + 2\lambda^2\alpha = 0$ (chain rule on the square), i.e. $2\alpha\,(2\alpha^2 - 2\omega^2 + \lambda^2) = 0$. For $\lambda^2 < 2\omega^2$ this gives an interior peak at $\alpha^2 = \omega^2 - \tfrac12\lambda^2$, where $R_{\max} = \dfrac{2a}{\lambda\sqrt{4\omega^2 - \lambda^2}} \approx \dfrac{a}{\lambda\omega}$ for small $\lambda$. The response is finite for any $\lambda > 0$ but inversely proportional to the damping: pseudo-resonance. For $\lambda^2 \ge 2\omega^2$ there is no peak: $R(\alpha)$ decreases monotonically from $R(0) = a/\omega^2$.
- Limits: $\alpha \to 0$: $R \to a/\omega^2$ (static deflection under a steady force); $\alpha \to \infty$: $R \approx a/\alpha^2 \to 0$ (inertia dominates).
- Dimensions: $a$ is force per unit mass [m/s²]; $[\omega^2 - \alpha^2] = [\lambda\alpha] = \text{s}^{-2}$, so $[R] = \text{m}$ ✓.
Rotor blades, engine mounts and wing structures are tuned so that their natural frequencies stay away from forcing frequencies such as rotor or shaft speeds. Helicopter ground resonance and aeroelastic flutter are more dangerous still: as in the Millennium Bridge example, the coupling with the excitation acts as negative damping, and the amplitude grows exponentially rather than linearly.
8. Variable coefficients: reduction of order
In general we will not be able to find all solutions, but we will study two different strategies.
8.1 Strategy 1: reduction of order
This method finds a second solution of a homogeneous problem when one is already known [Notes, p. 7]. Suppose $u(t)$ solves $x'' + p(t)x' + q(t)x = 0$, and look for a second solution of the form $x(t) = y(t)\,u(t)$.
- Product rule: $x' = y'u + yu'$, and differentiating again, $x'' = y''u + y'u' + y'u' + yu'' = y''u + 2y'u' + yu''$.
- Plug $x$ into the equation: $$\big[y''u + 2y'u' + yu''\big] + p(t)\big[y'u + yu'\big] + q(t)\big[yu\big] = 0.$$
- Group by $y$, $y'$ and $y''$: $$\big[u'' + p(t)u' + q(t)u\big]\,y + \big[2u' + p(t)u\big]\,y' + u\,y'' = 0.$$ The first bracket is $0$ because $u$ is a solution, so $y$ itself disappears.
- Define $z = y'$ [Notes, p. 7]: $z$ satisfies the first-order equation $$u\,z' + \big[2u' + p(t)\,u\big]\,z = 0.$$
The notes stop here. To finish (Robinson Exercise 17.7), the equation is separable: where $u \neq 0$ and $z \neq 0$, $\dfrac{z'}{z} = -\dfrac{2u'}{u} - p$. Integrate each term: by the chain rule $\dfrac{d}{dt}\ln|z(t)| = \dfrac{z'}{z}$ and $\dfrac{d}{dt}\ln|u(t)| = \dfrac{u'}{u}$ (equivalently, substitute $s = u(t)$, $ds = u'(t)\,dt$, so $\int\frac{u'}{u}\,dt = \int\frac{ds}{s} = \ln|s|$). Hence $\ln|z| = -2\ln|u| - \int p\,dt + c$ and
$$z = \frac{A\,e^{-\int p\,dt}}{u^2}, \qquad y = \int z\,dt, \qquad x_2 = u\int\frac{e^{-\int p\,dt}}{u^2}\,dt.$$For $x'' + px' + qx = 0$ with a double root $k = -p/2$, take $u = e^{kt}$. Then $e^{-\int p\,dt} = e^{-pt} = e^{2kt}$ and $u^2 = e^{2kt}$, so $z = A$, $y = At + B$ and $x = (At + B)\,e^{kt}$: the second solution $te^{kt}$ of Case 02 comes out of the method.
Robinson writes the same computation for $a(t)\ddot x + b(t)\dot x + c(t)x = 0$:
- Substitute $x = u(t)\,y(t)$ (the constant $A$ in $Au$ is promoted to a function). Product rule: $\dot x = \dot uy + u\dot y$, $\ \ddot x = \ddot uy + 2\dot u\dot y + u\ddot y$.
- Substitute and group by $y$, $\dot y$, $\ddot y$: $$y\,\big[a\ddot u + b\dot u + cu\big] + a\big(2\dot u\dot y + u\ddot y\big) + bu\dot y = 0.$$ The bracket is zero because $u$ is a solution — this is why the method works.
- $y$ itself has disappeared. With $z = \dot y$ we get the first-order linear equation $\ \big[a u\big]\dot z + \big[2a\dot u + bu\big]z = 0$ (17.3).
- Solve for $z$ (separable or integrating factor), integrate once to get $y$, and set $x = uy$.
In the normalised form $\ddot x + p\dot x + qx = 0$ the result is $z = \dfrac{A\,e^{-\int p\,dt}}{u^2}$ (Robinson Exercise 17.7), so $x_2 = u\displaystyle\int\frac{e^{-\int p\,dt}}{u^2}\,dt$. Robinson advises applying the method rather than memorising this formula.
For $\ddot x - 2\lambda\dot x + \lambda^2x = 0$ with known $u = e^{\lambda t}$, set $x = ye^{\lambda t}$: $\dot x = (\lambda y + \dot y)e^{\lambda t}$, $\ddot x = (\lambda^2y + 2\lambda\dot y + \ddot y)e^{\lambda t}$. Substituting and cancelling $e^{\lambda t} \neq 0$:
$$\lambda^2y + 2\lambda\dot y + \ddot y - 2\lambda^2y - 2\lambda\dot y + \lambda^2y = \ddot y = 0 \;\Longrightarrow\; y = A + Bt \;\Longrightarrow\; x = (A + Bt)e^{\lambda t}. \tag{17.4}$$The second solution $te^{\lambda t}$ of Section 4.2 is now derived, not guessed.
- Check $u$: $\dot u = -t^{-2}$, $\ddot u = 2t^{-3}$, so $t^2(2t^{-3}) - 2t(-t^{-2}) - 4t^{-1} = (2 + 2 - 4)t^{-1} = 0$ ✓.
- $x = y/t$: $\dot x = \dfrac{\dot y}{t} - \dfrac{y}{t^2}$, $\ddot x = \dfrac{\ddot y}{t} - \dfrac{2\dot y}{t^2} + \dfrac{2y}{t^3}$.
- Substitute: $t\ddot y - 2\dot y + \dfrac{2y}{t} - 2\dot y + \dfrac{2y}{t} - \dfrac{4y}{t} = t\ddot y - 4\dot y = 0$.
- $z = \dot y$: $\dot z - \dfrac4t z = 0$. Integrating factor $e^{-4\ln t} = t^{-4}$: $\frac{d}{dt}(t^{-4}z) = 0$, so $z = ct^4$.
- $y = \int ct^4\,dt = \dfrac{c}{5}t^5 + b$ and $x = y/t = at^4 + b/t$. The new independent solution is $x_2 = t^4$.
Check: $t^2(12t^2) - 2t(4t^3) - 4t^4 = (12 - 8 - 4)t^4 = 0$ ✓. Wronskian: $t^{-1}\cdot4t^3 - t^4\cdot(-t^{-2}) = 5t^2 \neq 0$ for $t > 0$ ✓.
9. Strategy 2: variation of constants
To find a particular solution of the complete equation $x'' + p(t)x' + q(t)x = f(t)$ when two linearly independent solutions $x_1$, $x_2$ of the homogeneous equation are known. Idea: look for a solution $x(t) = u_1(t)\,x_1(t) + u_2(t)\,x_2(t)$, where we need to solve for $u_1$ and $u_2$.
9.1 Derivation, step by step
- First derivative (product rule on each term): $$x' = u_1'x_1 + u_1x_1' + u_2'x_2 + u_2x_2'.$$
- Impose $u_1'x_1 + u_2'x_2 = 0$. Two unknown functions and one equation leave one free condition; this choice keeps $u_j''$ out of $x''$. Where $x_1 \neq 0$ it gives $$u_1' = -u_2'\,\frac{x_2}{x_1}. \tag{$\ast$}$$ Then $x' = u_1x_1' + u_2x_2'$.
- Second derivative: $\ x'' = u_1'x_1' + u_1x_1'' + u_2'x_2' + u_2x_2''$.
- Plug in and regroup by $u_1$ and $u_2$: $$\big(u_1'x_1' + u_1x_1'' + u_2'x_2' + u_2x_2''\big) + p\big(u_1x_1' + u_2x_2'\big) + q\big(u_1x_1 + u_2x_2\big) = f,$$ $$u_1\big(x_1'' + px_1' + qx_1\big) + u_2\big(x_2'' + px_2' + qx_2\big) + u_1'x_1' + u_2'x_2' = f.$$ Both brackets are $0$ because $x_1$ and $x_2$ satisfy the homogeneous equation, so $$u_1'x_1' + u_2'x_2' = f(t).$$
- Substitute $(\ast)$: $$\left(-u_2'\frac{x_2}{x_1}\right)x_1' + u_2'x_2' = u_2'\left(x_2' - \frac{x_2x_1'}{x_1}\right) = u_2'\,\frac{x_1x_2' - x_2x_1'}{x_1} = u_2'\,\frac{W}{x_1} = f.$$
- Solve: $u_2' = \dfrac{f\,x_1}{W}$, and from $(\ast)$, $u_1' = -\dfrac{f\,x_1}{W}\cdot\dfrac{x_2}{x_1} = -\dfrac{f\,x_2}{W}$, with $W = x_1x_2' - x_2x_1'$.
The notes write $u_1' = \dfrac{f\,x_2}{W}$ and $u_1 = \displaystyle\int\frac{x_2f}{W}$, without the minus sign. Their own relation $(\ast)$, $u_1' = -u_2'x_2/x_1$, with $u_2' = fx_1/W$, gives $u_1' = -fx_2/W$. Robinson's formula (18.8) has the minus sign too.
Check with the example of Section 6.3, $f = 12$. $x_1 = e^{-3t}$, $x_2 = e^{2t}$, $W = 5e^{-t}$ (Section 3).
- $u_2' = \dfrac{12\,e^{-3t}}{5e^{-t}} = \dfrac{12}{5}e^{-2t}$. With $s = -2t$, $ds = -2\,dt$: $\int\tfrac{12}{5}e^{-2t}dt = -\tfrac{6}{5}\int e^{s}ds = -\tfrac65e^{-2t}$ (check: $\frac{d}{dt}\big(-\tfrac65e^{-2t}\big) = \tfrac{12}{5}e^{-2t}$), so $u_2 = -\dfrac65e^{-2t}$ and $u_2x_2 = -\dfrac65$.
- Correct sign: $u_1' = -\dfrac{12\,e^{2t}}{5e^{-t}} = -\dfrac{12}{5}e^{3t}$; with $s = 3t$, $ds = 3\,dt$, $\int -\tfrac{12}{5}e^{3t}dt = -\tfrac45\int e^{s}ds = -\tfrac45e^{3t}$, so $u_1 = -\dfrac45e^{3t}$, $u_1x_1 = -\dfrac45$ and $x_N = -\dfrac45 - \dfrac65 = -2$ ✓ (the result of Section 6.3).
- Sign of the notes: $u_1 = +\dfrac45e^{3t}$ gives $x_N = \dfrac45 - \dfrac65 = -\dfrac25$, and $-6\cdot\left(-\dfrac25\right) = \dfrac{12}{5} \neq 12$ ✗.
9.2 Complement (Robinson): the same result without dividing by $x_1$
Step 2 divides by $x_1$, which may vanish (for instance $x_1 = \sin t$ at $t = 0$). Robinson solves the two conditions as a linear system with Cramer's rule, which needs only $W \neq 0$ and gives the same formulas [Robinson, Ch. 18, pp. 164–166].
The two conditions of steps 2 and 4 form a linear system for $u_1'$, $u_2'$:
$$\begin{pmatrix} x_1 & x_2 \\ x_1' & x_2'\end{pmatrix}\begin{pmatrix} u_1' \\ u_2'\end{pmatrix} = \begin{pmatrix} 0 \\ f\end{pmatrix}.$$Its determinant is the Wronskian $W = x_1x_2' - x_2x_1'$, never zero for independent solutions (Section 3). Cramer's rule:
$$u_1' = \frac{\begin{vmatrix} 0 & x_2 \\ f & x_2'\end{vmatrix}}{W} = -\frac{x_2\,f}{W}, \qquad u_2' = \frac{\begin{vmatrix} x_1 & 0 \\ x_1' & f\end{vmatrix}}{W} = \frac{x_1\,f}{W}.$$Robinson writes the forcing as $g(t)$:
Constants of integration only add multiples of $x_1$, $x_2$, which belong to the complementary function.
$x_1 = e^{-3t}$, $x_2 = e^{2t}$, $W = e^{-3t}\cdot 2e^{2t} - e^{2t}\cdot(-3e^{-3t}) = 5e^{-t}$. Then $\dfrac{x_2g}{W} = \dfrac{e^{2t}\cdot5e^{-3t}}{5e^{-t}} = 1$ and $\dfrac{x_1g}{W} = \dfrac{e^{-3t}\cdot5e^{-3t}}{5e^{-t}} = e^{-5t}$, so
$$x_p = -e^{-3t}\int 1\,dt + e^{2t}\int e^{-5t}\,dt = -te^{-3t} + e^{2t}\left(-\frac{e^{-5t}}{5}\right) = -te^{-3t} - \frac15e^{-3t}.$$The second term is a multiple of $x_1$ and is absorbed into $Ae^{-3t}$: the particular integral $-te^{-3t}$ of Example 14.8 is recovered.
The book writes the second integral as $5e^{2t}\int e^{-5t}dt$ and the result as $-te^{-3t} - e^{-3t}$. The factor $5$ was already cancelled by $W = 5e^{-t}$, so the term is $e^{2t}\cdot(-\tfrac15e^{-5t}) = -\tfrac15e^{-3t}$. Check: $\frac{d}{dt}\big(-\tfrac15e^{-5t}\big) = e^{-5t}$, so $\int e^{-5t}dt = -\tfrac15e^{-5t}$. The slip is harmless for the general solution, since any multiple of $e^{-3t}$ is absorbed into the arbitrary constant, as Robinson notes.
- $x_1 = \sin t$, $x_2 = \cos t$; ansatz $x = u\sin t + v\cos t$ with $\dot u\sin t + \dot v\cos t = 0$ (18.10).
- Then $\dot x = u\cos t - v\sin t$, $\ddot x = \dot u\cos t - u\sin t - \dot v\sin t - v\cos t$, and the equation reduces to $\dot u\cos t - \dot v\sin t = \tan t$ (18.11).
- Multiply (18.10) by $\sin t$, (18.11) by $\cos t$ and add: $\dot u = \tan t\cos t = \sin t$. Multiply (18.10) by $\cos t$, (18.11) by $\sin t$ and subtract: $\dot v = -\dfrac{\sin^2t}{\cos t} = \cos t - \sec t$.
- Integrate: $u = -\cos t$; $v = \sin t - \ln|\sec t + \tan t|$, using $\int\sec t\,dt = \ln|\sec t + \tan t|$ (differentiate to check: $\frac{\sec t\tan t + \sec^2t}{\sec t + \tan t} = \sec t$).
- $x_p = -\cos t\sin t + \big(\sin t - \ln|\sec t + \tan t|\big)\cos t = -\cos t\,\ln|\sec t + \tan t|$.
Check (write $\ell = \ln|\sec t + \tan t|$, $\dot\ell = \sec t$): $\dot x_p = \sin t\,\ell - 1$, $\ddot x_p = \cos t\,\ell + \sin t\sec t = \cos t\,\ell + \tan t$, so $\ddot x_p + x_p = \tan t$ ✓.
10. Summary and lecture map
| Result | Statement | Notes | Robinson | Lecture |
|---|---|---|---|---|
| Existence and uniqueness | $f$, $\partial f/\partial x$, $\partial f/\partial\dot x$ continuous in a cube; data $x(t_0)$, $\dot x(t_0)$ | p. 5 | Thm 11.1, p. 102 | Sep 22 |
| Superposition | $\alpha x_1 + \beta x_2$ solves $x'' + px' + qx = 0$ | p. 5 | Sec. 11.2, p. 103 | Sep 22 |
| General solution | $x = C_1x_1 + C_2x_2$ with $x_1$, $x_2$ independent | p. 5 | Sec. 11.3, pp. 104–106 | Sep 22 |
| Wronskian | $W = x_1x_2' - x_2x_1'$; $W' = -pW$, $W = Ce^{-\int p}$: $W \equiv 0$ or never $0$ | p. 5 | Sec. 11.4, Ex. 11.2 | Sep 22 |
| Constant coefficients | $k^2 + pk + q = 0$; Cases 01, 02, 03 | p. 5 | Ch. 12, Sec. 16.1 | Sep 22–24 |
| Mass–spring–damper | Over/critical/under-damped | p. 5 (spring) | Ch. 13 | Sep 22 |
| Inhomogeneous | $x = x_H + x_N$; educated guess; $x'' + x' - 6x = 12,\ 36t,\ 4e^{-2t}$ | p. 6 | Ch. 14 | Sep 24–29 |
| Resonance | Guess times $t$; damped peak at $\alpha^2 = \omega^2 - \lambda^2/2$ | p. 6 | Ch. 15 | — |
| Reduction of order | $x = yu$ gives a first-order equation for $z = y'$ | p. 7 | Ch. 17 | Sep 24–29 |
| Variation of constants | $u_1' = -fx_2/W$, $u_2' = fx_1/W$ | p. 8 | Ch. 18 | Oct 1 |
Topic 4 rewrites $\ddot x + p\dot x + qx = f$ as a first-order system for $(x, \dot x)$ and replaces "two independent solutions" by a fundamental matrix, the Wronskian by a determinant of vector solutions, and the auxiliary equation by an eigenvalue problem.