10.3 Cauchy Problems for the Wave and Heat Equations; Transform Methods
The wave and heat equations on the whole line and on a half-line: d'Alembert's formula, domains of dependence and influence, energy and reflection at an end, Duhamel's principle; the Fourier transform and its table, the heat kernel and the Cauchy problem for heat flow; and the Laplace transform in time for semi-infinite rods and strings.
On a bounded interval (§10.2) the boundary conditions pick out a discrete list of modes, and separation of variables writes every solution as a Fourier series. On the whole line there are no ends and no discrete modes, yet the two most important evolution equations can be solved exactly, by closed formulas. For the wave equation $u_{tt}=c^2u_{xx}$ it is d'Alembert's formula: every solution is a wave moving left plus a wave moving right, both at speed $c$. For the heat equation $u_t=k\,u_{xx}$ it is a convolution with the heat kernel, a Gaussian that spreads out as time passes. The two formulas show the deepest difference between hyperbolic and parabolic equations. A wave carries information at the finite speed $c$, it can be run backwards, and it keeps its corners. Heat is felt everywhere at once, it smooths every corner instantly, and it cannot be run backwards. The two transforms of the title are the general tools behind these formulas. The Fourier transform in $x$ diagonalises $\partial_x$ on the line, and the Laplace transform in $t$ turns a problem on a half-line, with the boundary value given as a function of time, into an ODE in $x$.
GATE MA asks about all of this. MCQs and NATs ask you to evaluate d'Alembert's formula at a point, to find the value of a forced wave at a point through Duhamel's principle or a particular solution, to solve the heat equation with Gaussian or trigonometric initial data, to read a Fourier transform off the table, or to solve a semi-infinite rod or string by the Laplace transform. MSQs test judgement: which points a given value depends on, whether a disturbance ever dies out, which quantities are conserved, when the heat problem has a unique solution, and what the maximum principle allows. Throughout, $c\gt0$ is the wave speed and $k\gt0$ the diffusivity. The characteristics of §10.1 and the classification of §10.2 are in the background, and the Laplace transform is the one of §6.2, with its notation and table.
The Wave Equation on the Line: d'Alembert's Formula
The equation $u_{tt}=c^2u_{xx}$ describes small transverse vibrations of a long taut string: $u(x,t)$ is the displacement, and $c=\sqrt{T/\rho}$ for tension $T$ and linear density $\rho$. It is hyperbolic, with characteristics $x\pm ct=\text{const}$ (§10.2). For a $C^2$ function the mixed partial derivatives commute, so the wave operator factors into two transport operators:
$$u_{tt}-c^2u_{xx}=\left(\partial_t-c\,\partial_x\right)\left(\partial_t+c\,\partial_x\right)u .$$Each factor is a first-order operator of the kind solved by characteristics in §10.1. That factorisation is the whole proof of d'Alembert's formula.
Let $c\gt0$, $f\in C^2(\mathbb{R})$ and $g\in C^1(\mathbb{R})$. The Cauchy problem (initial value problem) for the wave equation on the line is
$$u_{tt}=c^2u_{xx}\quad(x\in\mathbb{R},\ t\gt0),$$ $$u(x,0)=f(x),\qquad u_t(x,0)=g(x)\quad(x\in\mathbb{R}).$$$f$ is the initial displacement and $g$ the initial velocity. A classical solution is a function $u\in C^2(\mathbb{R}\times[0,\infty))$ (twice continuously differentiable for $t\gt0$, with all derivatives up to order $2$ extending continuously to $t=0$) that satisfies the equation and both initial conditions.
Under the hypotheses of Definition 10.11 the Cauchy problem has exactly one classical solution, namely
$$\begin{aligned}u(x,t)&=\frac{f(x+ct)+f(x-ct)}{2}\\&\quad+\frac{1}{2c}\int_{x-ct}^{x+ct}g(y)\,dy .\end{aligned}$$Equivalently $u(x,t)=F(x+ct)+G(x-ct)$ with $F=\tfrac12f+\tfrac1{2c}\Gamma$, $G=\tfrac12f-\tfrac1{2c}\Gamma$ and $\Gamma(y)=\int_0^yg$. $G(x-ct)$ is a profile moving right at speed $c$ and $F(x+ct)$ one moving left. Moreover, for data $(f_1,g_1)$ and $(f_2,g_2)$ the solutions satisfy $|u_1-u_2|\le\sup|f_1-f_2|+t\,\sup|g_1-g_2|$, so the problem is well posed.
Proof. Step 1: two transport equations. Let $u$ be a classical solution and put $w=u_t+cu_x$, a $C^1$ function. Since $u_{xt}=u_{tx}$,
$$w_t-c\,w_x=u_{tt}+cu_{xt}-cu_{tx}-c^2u_{xx}=0 .$$Step 2: solve for $w$ along characteristics. Fix $(x,t)$ with $t\gt0$ and put $\varphi(\tau)=w\left(x+c(t-\tau),\tau\right)$ for $0\le\tau\le t$. By the chain rule $\varphi'(\tau)=-c\,w_x+w_t=0$, so $\varphi$ is constant. Comparing $\tau=t$ with $\tau=0$, and using $u_t(y,0)=g(y)$ and $u_x(y,0)=f'(y)$,
$$w(x,t)=w(x+ct,0)=\psi(x+ct),\qquad \psi=g+c\,f' .$$Step 3: solve for $u$. Now $u_t+cu_x=\psi(x+ct)$. Put $\chi(\tau)=u\left(x-c(t-\tau),\tau\right)$. Then $\chi'(\tau)=cu_x+u_t=\psi\left(x-ct+2c\tau\right)$, and integrating from $0$ to $t$ (substituting $y=x-ct+2c\tau$):
$$u(x,t)=f(x-ct)+\frac1{2c}\int_{x-ct}^{x+ct}\psi(y)\,dy .$$Since $\frac1{2c}\int_{x-ct}^{x+ct}cf'=\frac12\left[f(x+ct)-f(x-ct)\right]$, this is exactly d'Alembert's formula. So every solution is given by the formula, which proves uniqueness.
Step 4: the formula is a solution. Write it as $F(x+ct)+G(x-ct)$ with $F,G$ as in the statement. They are $C^2$ because $f\in C^2$ and $\Gamma'=g\in C^1$. For any such function $u_{tt}=c^2\left(F''+G''\right)=c^2u_{xx}$. At $t=0$, $u=F+G=f$, and $u_t=cF'(x+ct)-cG'(x-ct)$ gives $u_t(x,0)=c\cdot\tfrac1c\Gamma'(x)=g(x)$. Finally, the estimate follows from the formula: the first term changes by at most $\sup|f_1-f_2|$, and the integral term by at most $\tfrac1{2c}\cdot2ct\cdot\sup|g_1-g_2|$. ∎
The formula still makes sense when $f$ is only continuous and $g$ only integrable. It then defines a generalised (weak) solution, the natural model for a plucked string, whose initial shape has corners. The proof shows what happens to such corners: $u$ is built from $f$ and $\Gamma$ evaluated along characteristics, so a kink of $f$ at $x_0$ is not smoothed out. It splits into two kinks that travel along the characteristics $x=x_0\pm ct$. Singularities of the data travel along characteristics, and only along them.
The formula says exactly which data a value depends on. That is the content of the next definition and theorem, and the main source of GATE MSQs on this topic.
For $u_{tt}=c^2u_{xx}$ the characteristics are the lines $x-ct=\text{const}$ and $x+ct=\text{const}$. Let $(x_0,t_0)$ have $t_0\gt0$.
- The domain of dependence of $(x_0,t_0)$ is the interval $[x_0-ct_0,\ x_0+ct_0]$ of the initial line. The characteristic triangle $\Delta(x_0,t_0)$ is the closed triangle with this base and apex $(x_0,t_0)$, bounded by the two characteristics through $(x_0,t_0)$.
- The domain of influence of an interval $[a,b]$ of the initial line is the set of points $(x,t)$, $t\ge0$, whose domain of dependence meets $[a,b]$:
The domain of influence is the region between the leftmost characteristic from $a$ and the rightmost characteristic from $b$.
Let $u$ be the solution of Theorem 10.21.
- Dependence. $u(x_0,t_0)$ depends only on the values of $f$ at the two points $x_0\pm ct_0$ and on $g$ on $[x_0-ct_0,x_0+ct_0]$. If two data pairs agree on this interval, the two solutions agree on the whole triangle $\Delta(x_0,t_0)$.
- Influence. If $f$ and $g$ vanish outside $[a,b]$, then $u(x,t)=0$ outside the domain of influence, that is, whenever $x\lt a-ct$ or $x\gt b+ct$.
- After the wave has passed. With $f,g$ as in (2), if $x-ct\le a$ and $x+ct\ge b$, then
So the string returns to rest behind the wave front if and only if $\int_a^bg=0$. In particular, when $g\equiv0$ the string is at rest there.
Proof. (1) The formula uses $f(x_0\pm ct_0)$ and $g$ on $[x_0-ct_0,x_0+ct_0]$, nothing else. A point $(x,t)$ of $\Delta(x_0,t_0)$ has $x-ct\ge x_0-ct_0$ and $x+ct\le x_0+ct_0$, so its own domain of dependence lies inside that of $(x_0,t_0)$. (2) If $x\gt b+ct$, then $x-ct\gt b$, so the interval $[x-ct,x+ct]$ lies in $(b,\infty)$. There $f$ vanishes at both endpoints and $g$ vanishes on the whole interval, so $u=0$. The case $x\lt a-ct$ is the same. (3) A continuous $f$ vanishing outside $[a,b]$ has $f(a)=f(b)=0$, so both values $f(x\pm ct)$ are $0$. Also $[x-ct,x+ct]\supseteq[a,b]$, and $g=0$ off $[a,b]$, so the integral equals $\int_a^bg$. ∎
Part (3) says that the sharp form of Huygens' principle fails in one space dimension. A localised disturbance in $f$ passes a fixed observer and leaves no trace. A localised velocity disturbance leaves the string permanently displaced by $\frac1{2c}\int g$. In three space dimensions the sharp principle holds, which is why a sound is heard as a pulse and not as a pulse followed by a permanent hum.
Let $u$ be a classical solution with data $f\in C^2$ and $g\in C^1$ vanishing outside $[a,b]$. Then the energy
$$E(t)=\frac12\int_{-\infty}^{\infty}\left(u_t^2+c^2u_x^2\right)dx$$is finite and constant in $t$: $E(t)=E(0)=\frac12\int\left(g^2+c^2f'^2\right)dx$. Here $K(t)=\frac12\int u_t^2$ is the kinetic energy and $P(t)=\frac{c^2}2\int u_x^2$ the potential energy. Only their sum is conserved.
Proof. By Theorem 10.21 the solution is given by d'Alembert's formula, so by Theorem 10.22(2) $u$ vanishes on the open sets $x\lt a-ct$ and $x\gt b+ct$. Hence $u_t$ and $u_x$ vanish there too, and by continuity on their boundaries. For $t$ in a bounded interval $[0,T]$ the integrand is $C^1$ and vanishes outside the fixed compact interval $[a-cT,b+cT]$, so we may differentiate under the integral sign:
$$E'(t)=\int\left(u_tu_{tt}+c^2u_xu_{xt}\right)dx=c^2\int\left(u_tu_{xx}+u_xu_{tx}\right)dx .$$We used $u_{tt}=c^2u_{xx}$ and $u_{xt}=u_{tx}$. The last integrand is $\partial_x\left(u_tu_x\right)$, so the integral equals $\left[u_tu_x\right]$ evaluated at the ends of $[a-cT,b+cT]$, which is $0$. ∎
The same computation on a bounded interval with fixed or free ends gives $E'=0$ as well. It also proves uniqueness there: the difference of two solutions has zero data, hence zero energy, hence $u_t=u_x=0$ and $u\equiv0$. The half-line problem of the next theorem is solved by a reflection trick instead.
Let $f\in C^2[0,\infty)$ and $g\in C^1[0,\infty)$, and consider $u_{tt}=c^2u_{xx}$ for $x\gt0$, $t\gt0$, with $u(x,0)=f(x)$ and $u_t(x,0)=g(x)$.
- Fixed end, $u(0,t)=0$. Suppose the compatibility conditions $f(0)=g(0)=f''(0)=0$ hold. Then the unique $C^2$ solution is d'Alembert's formula applied to the odd extensions of $f$ and $g$. For $x\ge ct$ it is the usual formula, and for $0\le x\lt ct$ $$\begin{aligned}u(x,t)&=\tfrac12\left[f(x+ct)-f(ct-x)\right]\\&\quad+\tfrac1{2c}\int_{ct-x}^{x+ct}g(y)\,dy .\end{aligned}$$
- Free end, $u_x(0,t)=0$. Suppose $f'(0)=g'(0)=0$. Then the solution uses the even extensions. For $0\le x\lt ct$ $$\begin{aligned}u(x,t)&=\tfrac12\left[f(x+ct)+f(ct-x)\right]\\&\quad+\tfrac1{2c}\left[\int_0^{x+ct}\!g+\int_0^{ct-x}\!g\right].\end{aligned}$$
Proof. (1) Let $f_o,g_o$ be the odd extensions, $f_o(-x)=-f(x)$. Then $f_o$ is continuous because $f(0)=0$. Its derivative is even and continuous. Its second derivative is odd, so it is continuous at $0$ exactly when $f''(0)=0$. Thus $f_o\in C^2(\mathbb{R})$, and likewise $g_o\in C^1(\mathbb{R})$ because $g(0)=0$. Let $U$ be the whole-line solution with data $f_o,g_o$. The function $-U(-x,t)$ solves the same equation with data $-f_o(-x)=f_o(x)$ and $-g_o(-x)=g_o(x)$, so by uniqueness (Theorem 10.21) $-U(-x,t)=U(x,t)$. $U$ is odd in $x$, hence $U(0,t)=0$, and its restriction to $x\ge0$ solves the half-line problem.
For uniqueness, let $v$ be any $C^2$ solution on the quarter plane. From $v(0,t)=0$ we get $v_{tt}(0,t)=0$, hence $v_{xx}(0,t)=0$. So the odd extension of $v$ in $x$ is $C^2$ and solves the whole-line problem with data $f_o,g_o$, and it must equal $U$. For the explicit form when $x\lt ct$: $f_o(x-ct)=-f(ct-x)$, and in $\int_{x-ct}^{x+ct}g_o$ the part over $[x-ct,ct-x]$ cancels because $g_o$ is odd. (2) is identical with even extensions. Then $U(-x,t)=U(x,t)$ gives $U_x(0,t)=0$, and an even $C^2$ extension needs $f'(0)=0$ (and $g'(0)=0$ for $g$). ∎
The physics is in the sign. A pulse travelling towards a fixed end comes back upside down: its image, an inverted copy behind the wall, emerges as it disappears. At a free end it comes back the right way up, and at the moment of reflection the end moves with twice the incoming amplitude.
The Non-Homogeneous Wave Equation: Duhamel's Principle
A string driven by an external force per unit mass $F(x,t)$ obeys $u_{tt}-c^2u_{xx}=F$. By linearity, the solution with data $f,g$ is d'Alembert's solution plus the solution of the forced equation with zero data. To find the latter, think of the force as a stream of impulses. During a short interval $[\tau,\tau+d\tau]$ it changes the velocity by $F(x,\tau)\,d\tau$, which launches a free wave from time $\tau$ on. Adding up these free waves is Duhamel's principle, the PDE version of variation of parameters (§6.1).
Let $F$ and $F_x$ be continuous on $\mathbb{R}\times[0,\infty)$. Then
$$u_p(x,t)=\frac1{2c}\int_0^t\!\!\int_{x-c(t-\tau)}^{x+c(t-\tau)}F(y,\tau)\,dy\,d\tau$$is $C^2$ on $\mathbb{R}\times[0,\infty)$ and solves $u_{tt}-c^2u_{xx}=F$ with $u(x,0)=0$ and $u_t(x,0)=0$. The double integral is $\frac1{2c}\iint_{\Delta(x,t)}F$, the integral over the characteristic triangle. For $f\in C^2$ and $g\in C^1$, the unique classical solution of
$$u_{tt}-c^2u_{xx}=F,$$ $$u(x,0)=f,\qquad u_t(x,0)=g$$is d'Alembert's solution (Theorem 10.21) plus $u_p$.
Proof. For $\tau\ge0$ and $t\ge\tau$ put
$$v(x,t;\tau)=\frac1{2c}\int_{x-c(t-\tau)}^{x+c(t-\tau)}F(y,\tau)\,dy .$$By Theorem 10.21, with time shifted by $\tau$, $v(\cdot,\cdot\,;\tau)$ solves the homogeneous equation for $t\gt\tau$ with $v=0$ and $v_t=F(\cdot,\tau)$ at $t=\tau$. Explicitly, $v_t=\tfrac12\left[F(x+c(t-\tau),\tau)+F(x-c(t-\tau),\tau)\right]$ and $v_x=\tfrac1{2c}\left[F(x+c(t-\tau),\tau)-F(x-c(t-\tau),\tau)\right]$. These are continuous, and so are the second derivatives, which involve $F_x$ only. Now $u_p(x,t)=\int_0^tv(x,t;\tau)\,d\tau$. Leibniz's rule for an integral with variable upper limit gives
$$\partial_tu_p=v(x,t;t)+\int_0^tv_t\,d\tau=\int_0^tv_t(x,t;\tau)\,d\tau ,$$ $$\partial_t^2u_p=v_t(x,t;t)+\int_0^tv_{tt}\,d\tau=F(x,t)+\int_0^tv_{tt}\,d\tau ,$$because $v(x,t;t)=0$ and $v_t(x,t;t)=F(x,t)$. Also $\partial_x^2u_p=\int_0^tv_{xx}\,d\tau$. Since $v_{tt}=c^2v_{xx}$ for each $\tau$, we get $\partial_t^2u_p-c^2\partial_x^2u_p=F$. At $t=0$ both $u_p$ and $\partial_tu_p$ are integrals over $[0,0]$, hence $0$. Uniqueness: the difference of two solutions of the full problem solves the homogeneous equation with zero data, so it vanishes by Theorem 10.21. ∎
Two consequences are used constantly. First, $u(x_0,t_0)$ depends on $F$ in the whole solid triangle $\Delta(x_0,t_0)$, not just on its base: the forced wave still has finite speed. Second, for a constant force $F\equiv F_0$ the triangle has area $ct^2$, so $u_p=\frac{1}{2c}F_0\cdot ct^2=\frac12F_0t^2$, which is free fall.
To solve $u_{tt}-c^2u_{xx}=F$ with $u(x,0)=f$ and $u_t(x,0)=g$:
- Look for a particular solution $u^*$ of the PDE alone when $F$ is simple. For $F=F(t)$ try $u^*(t)$ with $u^{*\prime\prime}=F$. For $F=F(x)$ try $u^*(x)$ with $-c^2u^{*\prime\prime}=F$. For $F=\sin(ax)h(t)$ try $u^*=A(t)\sin(ax)$, so that $A''+c^2a^2A=h$. Polynomial forcing usually has a polynomial $u^*$.
- Subtract. $w=u-u^*$ solves the homogeneous equation with data $f-u^*(\cdot,0)$ and $g-u^*_t(\cdot,0)$. Apply d'Alembert's formula.
- Otherwise use Duhamel: integrate $F$ over the characteristic triangle and divide by $2c$.
- Check the equation and both initial conditions, or evaluate both routes at one point.
The Fourier Transform
On the line, separation of variables produces the functions $e^{i\omega x}$ for every real $\omega$, not just for a discrete list of frequencies. A Fourier series $\sum c_ne^{inx}$ becomes an integral $\frac1{2\pi}\int\hat f(\omega)e^{i\omega x}\,d\omega$. Its use for PDEs comes from one rule: the transform turns $\partial_x$ into multiplication by $i\omega$. Below, $L^1(\mathbb{R})$ is the space of (Lebesgue) integrable functions, $\int|f|\lt\infty$ (§4.3). For the functions in this section, absolutely convergent improper Riemann integrals are enough.
For $f\in L^1(\mathbb{R})$ the Fourier transform of $f$ is the function $\hat f=\mathcal{F}\{f\}$ defined by
$$\hat f(\omega)=\int_{-\infty}^{\infty}f(x)\,e^{-i\omega x}\,dx\qquad(\omega\in\mathbb{R}),$$and the inverse transform of an integrable $\phi$ is
$$\mathcal{F}^{-1}\{\phi\}(x)=\frac1{2\pi}\int_{-\infty}^{\infty}\phi(\omega)\,e^{i\omega x}\,d\omega .$$The convolution of $f,g\in L^1$ is $(f*g)(x)=\int_{-\infty}^{\infty}f(x-y)\,g(y)\,dy$. It is defined for almost every $x$, and for every $x$ if one of the functions is also bounded. Conventions differ. Some texts use $e^{+i\omega x}$ in the forward transform, or put $\frac1{\sqrt{2\pi}}$ in front of both integrals. The results below then change by constant factors or by $\omega\mapsto-\omega$, so check which convention a question uses.
Let $f,g\in L^1(\mathbb{R})$, $a\in\mathbb{R}$ and $\lambda\ne0$.
- Bounded and continuous. $\hat f$ is continuous and $|\hat f(\omega)|\le\int|f|$. Moreover $\hat f(\omega)\to0$ as $|\omega|\to\infty$ (the Riemann–Lebesgue lemma).
- Shift, modulation, scaling. $\mathcal{F}\{f(x-a)\}=e^{-ia\omega}\hat f(\omega)$, $\ \mathcal{F}\{e^{iax}f(x)\}=\hat f(\omega-a)$, and $\mathcal{F}\{f(\lambda x)\}=\tfrac1{|\lambda|}\hat f\left(\tfrac\omega\lambda\right)$.
- Derivatives. If $f\in C^1$ and $f'\in L^1$, then $\mathcal{F}\{f'\}=i\omega\hat f(\omega)$. If also $f\in C^2$ and $f''\in L^1$, then $\mathcal{F}\{f''\}=-\omega^2\hat f(\omega)$.
- Multiplication by $x$. If $xf(x)\in L^1$, then $\hat f\in C^1$ and $\mathcal{F}\{xf(x)\}=i\,\hat f\,'(\omega)$.
- Convolution. $f*g\in L^1$ and $\mathcal{F}\{f*g\}=\hat f\,\hat g$.
Proof. (1) $|e^{-i\omega x}|=1$ gives the bound. If $\omega_n\to\omega$, the integrands $f(x)e^{-i\omega_nx}$ converge pointwise and are dominated by $|f|$, so $\hat f(\omega_n)\to\hat f(\omega)$ by the dominated convergence theorem (§4.3). For the indicator of $[\alpha,\beta]$ one computes $\hat f(\omega)=\frac{e^{-i\omega\alpha}-e^{-i\omega\beta}}{i\omega}\to0$. Step functions are dense in $L^1$, and the bound in (1) carries the limit over to every $f$ (we omit the density argument). (2) Substitute $z=x-a$ in the first integral. In the second, $e^{iax}e^{-i\omega x}=e^{-i(\omega-a)x}$. In the third, substitute $z=\lambda x$, with $dx=dz/|\lambda|$ once the orientation is fixed for $\lambda\lt0$.
(3) First, $f(x)\to0$ as $|x|\to\infty$: $f(x)=f(0)+\int_0^xf'$ has a limit as $x\to\infty$ because $f'\in L^1$, and a non-zero limit would contradict $\int|f|\lt\infty$. The same holds at $-\infty$. Integrating by parts on $[-R,R]$,
$$\int_{-R}^{R}f'e^{-i\omega x}dx=\Big[fe^{-i\omega x}\Big]_{-R}^{R}+i\omega\int_{-R}^{R}fe^{-i\omega x}dx ,$$and letting $R\to\infty$ gives $i\omega\hat f$. Apply this twice for $f''$. (4) Since $|e^{-ihx}-1|\le|hx|$, the difference quotients
$$\frac{\hat f(\omega+h)-\hat f(\omega)}{h}=\int f(x)e^{-i\omega x}\,\frac{e^{-ihx}-1}{h}\,dx$$have integrands dominated by $|xf(x)|$. By dominated convergence they tend to $\int(-ix)f(x)e^{-i\omega x}\,dx$. So $\hat f\,'=\mathcal{F}\{-ixf\}$, which is the claim. (5) By Tonelli's theorem $\iint|f(x-y)||g(y)|\,dy\,dx=\int|f|\int|g|\lt\infty$, so Fubini's theorem (§1.3, §4.3) allows us to swap the integrals:
$$\int(f*g)(x)e^{-i\omega x}dx=\int g(y)e^{-i\omega y}\left(\int f(x-y)e^{-i\omega(x-y)}dx\right)dy=\hat f(\omega)\hat g(\omega).\ ∎$$Proof. Let $\phi(\omega)=\mathcal{F}\{e^{-ax^2}\}$. Since $xe^{-ax^2}\in L^1$, Theorem 10.26(4) gives $\phi'(\omega)=\mathcal{F}\{-ixe^{-ax^2}\}$. Now $-ixe^{-ax^2}=\tfrac{i}{2a}\left(e^{-ax^2}\right)'$, so by Theorem 10.26(3)
$$\phi'(\omega)=\frac{i}{2a}\cdot i\omega\,\phi(\omega)=-\frac{\omega}{2a}\,\phi(\omega).$$This linear ODE gives $\phi(\omega)=\phi(0)\,e^{-\omega^2/(4a)}$, and $\phi(0)=\int e^{-ax^2}dx=\sqrt{\pi/a}$ is the Gaussian integral (§1.3). ∎
- Inversion. If $f\in L^1$ is continuous and $\hat f\in L^1$, then $f(x)=\dfrac1{2\pi}\displaystyle\int\hat f(\omega)e^{i\omega x}\,d\omega$ for every $x$. If instead $f\in L^1$ is piecewise $C^1$, then $\lim_{R\to\infty}\frac1{2\pi}\int_{-R}^{R}\hat f(\omega)e^{i\omega x}d\omega=\tfrac12\left[f(x^+)+f(x^-)\right]$ at every $x$.
- Uniqueness. If $\hat f=\hat g$, then $f=g$ almost everywhere, and at every point where both are continuous.
- Duality. Under the hypotheses of (1), $\mathcal{F}\{\hat f\}(\omega)=2\pi f(-\omega)$.
- Plancherel. If $f\in L^1\cap L^2$, then $\hat f\in L^2$ and $\displaystyle\int|f|^2dx=\frac1{2\pi}\int|\hat f|^2d\omega$. More generally $\int f\bar g=\frac1{2\pi}\int\hat f\,\overline{\hat g}$.
Proof (sketch of (1)–(3)). Insert a Gaussian damping factor, $\varepsilon\gt0$. By Fubini's theorem and Theorem 10.27 (with $a=\varepsilon$, evaluated at the frequency $y-x$),
$$\frac1{2\pi}\int\hat f(\omega)e^{-\varepsilon\omega^2}e^{i\omega x}d\omega=\int f(y)\,\frac{e^{-(x-y)^2/(4\varepsilon)}}{\sqrt{4\pi\varepsilon}}\,dy .$$As $\varepsilon\to0^+$, the left side tends to the inversion integral by dominated convergence (dominated by $|\hat f|$). The right side is the heat kernel of the next subsection, with $k=1$ and $t=\varepsilon$, convolved with $f$, and it tends to $f(x)$ at every point of continuity by the approximate-identity property in Theorem 10.29(4). The piecewise $C^1$ case runs like the Dirichlet-kernel proof for Fourier series. (2) follows from (1) applied to $f-g$, with the general case by approximation. (3) is (1) with $x$ replaced by $-\omega$. (4) is the integral analogue of Parseval's identity for Fourier series (§8.3). Apply (1) to $h=f*\tilde f$ with $\tilde f(x)=\overline{f(-x)}$: then $\hat h=|\hat f|^2$ and $h(0)=\int|f|^2$, and the general case is an approximation argument we omit. ∎
The table below collects the transforms GATE uses. Lines 1, 3 and 4 are direct integrations, line 2 follows from line 1 by duality, and line 5 is Theorem 10.27.
| $f(x)$ | $\hat f(\omega)=\int f(x)e^{-i\omega x}dx$ | conditions |
|---|---|---|
| $e^{-a|x|}$ | $\dfrac{2a}{a^2+\omega^2}$ | $a\gt0$ |
| $\dfrac{1}{x^2+a^2}$ | $\dfrac{\pi}{a}\,e^{-a|\omega|}$ | $a\gt0$ |
| $\chi_{[-a,a]}(x)$ (box) | $\dfrac{2\sin a\omega}{\omega}$ (value $2a$ at $\omega=0$) | $a\gt0$ |
| $(a-|x|)_+$ (triangle) | $\dfrac{4\sin^2(a\omega/2)}{\omega^2}$ | $a\gt0$ |
| $e^{-ax^2}$ | $\sqrt{\pi/a}\;e^{-\omega^2/(4a)}$ | $a\gt0$ |
| $f(x-b)$; $\ e^{ibx}f(x)$ | $e^{-ib\omega}\hat f(\omega)$; $\ \hat f(\omega-b)$ | $b\in\mathbb{R}$ |
| $f'(x)$; $\ f''(x)$; $\ xf(x)$ | $i\omega\hat f$; $\ -\omega^2\hat f$; $\ i\hat f\,'$ | Theorem 10.26 |
| $(f*g)(x)$; $\ \hat f(x)$ | $\hat f(\omega)\hat g(\omega)$; $\ 2\pi f(-\omega)$ | Theorems 10.26, 10.28 |
Half-line transforms. For $f$ defined on $[0,\infty)$, the Fourier sine and cosine transforms are $\mathcal{F}_s\{f\}(\omega)=\int_0^\infty f(x)\sin\omega x\,dx$ and $\mathcal{F}_c\{f\}(\omega)=\int_0^\infty f(x)\cos\omega x\,dx$. They are, up to the factors $-2i$ and $2$, the Fourier transforms of the odd and even extensions of $f$. Integrating by parts twice, when $f$ and $f'$ tend to $0$ at infinity,
$$\mathcal{F}_s\{f''\}=\omega f(0)-\omega^2\mathcal{F}_s\{f\},$$ $$\mathcal{F}_c\{f''\}=-f'(0)-\omega^2\mathcal{F}_c\{f\}.$$The sine transform therefore suits a prescribed value $u(0,t)$, and the cosine transform a prescribed flux $u_x(0,t)$. For the problems GATE sets, the odd and even extensions (Theorem 10.24 and Exercise III.9) or the Laplace transform in $t$ (the last subsection) are usually quicker.
The Heat Equation on the Line
Take the Fourier transform in $x$ of $u_t=ku_{xx}$, $u(x,0)=f(x)$, assuming for the moment that everything converges. Theorem 10.26(3) turns $u_{xx}$ into $-\omega^2\hat u$, so for each $\omega$ we get an ODE in $t$:
$$\hat u_t(\omega,t)=-k\omega^2\,\hat u(\omega,t),\qquad\hat u(\omega,0)=\hat f(\omega),$$so $\hat u(\omega,t)=\hat f(\omega)\,e^{-k\omega^2t}$. Each frequency is damped at the rate $k\omega^2$, so high frequencies (sharp features) die first. By Theorem 10.27 with $a=\frac1{4kt}$, the factor $e^{-k\omega^2t}$ is the transform of $\frac{1}{\sqrt{4\pi kt}}e^{-x^2/(4kt)}$. The convolution theorem then turns the product back into a convolution. That formula is now defined and proved directly, so that it holds even for data, such as a step or $\cos x$, that have no Fourier transform at all.
- $\Phi\gt0$, $\Phi\in C^\infty\left(\mathbb{R}\times(0,\infty)\right)$ and $\Phi_t=k\Phi_{xx}$.
- $\int_{-\infty}^{\infty}\Phi(x,t)\,dx=1$ for every $t\gt0$.
- $\hat\Phi(\omega,t)=e^{-k\omega^2t}$, and $\Phi(\cdot,t)*\Phi(\cdot,s)=\Phi(\cdot,t+s)$ (the semigroup property).
- Approximate identity. For every $\delta\gt0$, $\int_{|x|\ge\delta}\Phi(x,t)\,dx\to0$ as $t\to0^+$. Hence, if $f$ is bounded and continuous at $x_0$, then $\int\Phi(x-y,t)f(y)\,dy\to f(x_0)$ as $(x,t)\to(x_0,0^+)$.
Proof. (1) Logarithmic differentiation gives $\dfrac{\Phi_t}{\Phi}=-\dfrac1{2t}+\dfrac{x^2}{4kt^2}$ and $\dfrac{\Phi_x}{\Phi}=-\dfrac{x}{2kt}$. Hence
$$k\Phi_{xx}=k\Phi\left(\frac{x^2}{4k^2t^2}-\frac1{2kt}\right)=\Phi\left(\frac{x^2}{4kt^2}-\frac1{2t}\right)=\Phi_t .$$(2) Substitute $x=\sqrt{4kt}\,z$: the integral becomes $\frac1{\sqrt\pi}\int e^{-z^2}dz=1$. (3) The first claim is Theorem 10.27 with $a=\frac1{4kt}$. The transform of the convolution is $e^{-k\omega^2t}e^{-k\omega^2s}=e^{-k\omega^2(t+s)}$, so the semigroup property follows from uniqueness (Theorem 10.28(2)). (4) The same substitution gives $\int_{|x|\ge\delta}\Phi=\frac2{\sqrt\pi}\int_{\delta/\sqrt{4kt}}^{\infty}e^{-z^2}dz\to0$. Now let $|f|\le M$ and $\varepsilon\gt0$, and choose $\delta$ with $|f(y)-f(x_0)|\lt\varepsilon$ whenever $|y-x_0|\lt2\delta$. By (2),
$$\int\Phi(x-y,t)f(y)\,dy-f(x_0)=\int\Phi(x-y,t)\left[f(y)-f(x_0)\right]dy .$$Let $|x-x_0|\lt\delta$. Where $|y-x_0|\lt2\delta$ the integrand is at most $\varepsilon\,\Phi$, contributing at most $\varepsilon$. Where $|y-x_0|\ge2\delta$ we have $|x-y|\ge\delta$, so that part is at most $2M\int_{|z|\ge\delta}\Phi(z,t)\,dz$, which is less than $\varepsilon$ for small $t$. ∎
Uniqueness for the heat equation rests on a maximum principle, which we prove in the form needed.
Proof. First suppose $w_t\lt kw_{xx}$ strictly. Suppose the maximum of $w$ over the compact set $R$ were attained at a point $(x_0,t_0)$ with $|x_0|\lt L$ and $0\lt t_0\le T$. Then $w_{xx}(x_0,t_0)\le0$, since $x\mapsto w(x,t_0)$ has an interior maximum. Also $w_t(x_0,t_0)\ge0$: it equals $0$ if $t_0\lt T$, and is a one-sided derivative $\ge0$ if $t_0=T$. This contradicts $w_t\lt kw_{xx}$. In general, put $w_\varepsilon=w-\varepsilon t$ with $\varepsilon\gt0$. Then $(w_\varepsilon)_t=w_t-\varepsilon\lt k(w_\varepsilon)_{xx}$, so
$$\max_Rw\le\max_Rw_\varepsilon+\varepsilon T=\max_{\partial_pR}w_\varepsilon+\varepsilon T\le\max_{\partial_pR}w+\varepsilon T .$$Let $\varepsilon\to0$. ∎
Let $f$ be bounded and continuous on $\mathbb{R}$. Define $u(x,0)=f(x)$ and, for $t\gt0$,
$$u(x,t)=\int_{-\infty}^{\infty}\Phi(x-y,t)\,f(y)\,dy=\int_{-\infty}^{\infty}\frac{e^{-(x-y)^2/(4kt)}}{\sqrt{4\pi kt}}\,f(y)\,dy .$$- $u\in C^\infty\left(\mathbb{R}\times(0,\infty)\right)$ and $u_t=ku_{xx}$ for $t\gt0$. This uses only that $f$ is bounded and piecewise continuous.
- $u$ is continuous on $\mathbb{R}\times[0,\infty)$: the initial values are attained.
- $\inf f\le u(x,t)\le\sup f$.
- Uniqueness among bounded solutions. Let $v$ be continuous on $\mathbb{R}\times[0,\infty)$, with $v_t,v_{xx}$ continuous for $t\gt0$, $v_t=kv_{xx}$ and $v(x,0)=f(x)$. If $v$ is bounded on $\mathbb{R}\times[0,T]$ for every $T$, then $v=u$.
- If moreover $f\in L^1$, then $\int u(x,t)\,dx=\int f$ for all $t$ (conservation of heat).
Proof. (1) Each derivative $\partial_t^j\partial_x^m\Phi(x-y,t)$ is a polynomial in $x-y$ and $t^{-1/2}$ times $e^{-(x-y)^2/(4kt)}$. For $t$ in $[t_1,t_2]\subset(0,\infty)$ and $x$ in a bounded set it is dominated by $C(1+|y|^N)e^{-y^2/(8kt_2)}$ for some $C$ and $N$, and this is integrable against the bounded $f$. So we may differentiate under the integral sign, and $\Phi_t=k\Phi_{xx}$ passes to $u$. (2) is Theorem 10.29(4). (3) $u(x,t)$ is an average of the values of $f$ with the positive weight $\Phi(x-\cdot,t)$ of total mass $1$.
(4) Let $|v|,|u|\le B/2$ on $\mathbb{R}\times[0,T]$, so that $w=v-u$ satisfies $|w|\le B$, solves the heat equation, and has $w(x,0)=0$. Fix $L\gt0$ and compare $w$ with
$$h(x,t)=\frac{B}{L^2}\left(x^2+2kt\right),$$which satisfies $h_t=\frac{2kB}{L^2}=kh_{xx}$. On $R=[-L,L]\times[0,T]$ the function $z=w-h$ satisfies $z_t=kz_{xx}$. On the bottom, $z=-Bx^2/L^2\le0$. On the sides, $h(\pm L,t)\ge B\ge w$, so $z\le0$. By Theorem 10.30, $z\le0$ on $R$, that is, $w(x,t)\le\frac{B}{L^2}(x^2+2kt)$. The same argument applied to $-w$ gives $|w(x,t)|\le\frac{B}{L^2}(x^2+2kt)$. Fix $(x,t)$ and let $L\to\infty$: then $w(x,t)=0$. Since $T$ was arbitrary, $v=u$. (5) By Fubini's theorem and Theorem 10.29(2), $\iint\Phi(x-y,t)f(y)\,dy\,dx=\int f(y)\,dy$. ∎
Three consequences describe heat flow. Infinite speed: if $f\ge0$ is continuous and not identically zero, then $u(x,t)\gt0$ for every $x$ and every $t\gt0$, because $\Phi\gt0$ everywhere. Heat released in a small region is felt at once arbitrarily far away, though only in a Gaussian-small amount. Smoothing: a jump in $f$ disappears at once, since $u(\cdot,t)$ is $C^\infty$ for every $t\gt0$. Irreversibility: the formula cannot be run backwards. The function $\frac1ne^{-kn^2t}\sin nx$ solves the equation and has tiny initial data when $n$ is large, but at $t=-1$ it equals $\frac1ne^{kn^2}\sin nx$, which is huge. Arbitrarily small changes in the data produce enormous changes in the backward solution, so the backward problem is ill-posed. Finally, a source term is handled exactly as for waves. Under mild conditions on $F$ (bounded, continuous and $C^1$ in $x$, say), Duhamel's principle gives
$$u_F(x,t)=\int_0^t\!\!\int_{-\infty}^{\infty}\Phi(x-y,t-s)\,F(y,s)\,dy\,ds ,$$which solves $u_t=ku_{xx}+F$ with $u(x,0)=0$. Each layer of heat added at time $s$ spreads by the kernel for the remaining time $t-s$. We state this without proof.
The Laplace Transform in Time: Half-Line Problems
Now take a rod or string occupying $x\gt0$, with its end condition given as a function of time: an end held at a temperature $p(t)$, or an end moved up and down. The space variable no longer ranges over $\mathbb{R}$, but time still ranges over $[0,\infty)$, and the initial conditions are exactly what the Laplace transform needs. We use the notation of §6.2. For a function $\varphi$ of $t$, $\mathcal{L}\{\varphi\}(s)=\int_0^\infty e^{-st}\varphi(t)\,dt$. For the solution we write $U(x,s)=\int_0^\infty e^{-st}u(x,t)\,dt$, and $P=\mathcal{L}\{p\}$. Because $u$ denotes the solution, we write the unit step function of §6.2 as $H$: $H(t-a)=0$ for $t\lt a$ and $1$ for $t\ge a$. The second shifting theorem then reads $\mathcal{L}\{\varphi(t-a)H(t-a)\}=e^{-as}\mathcal{L}\{\varphi\}(s)$ for $a\ge0$.
Let $J$ be an interval, and let $u,u_t,u_x,u_{xx}$ be continuous on $J\times[0,\infty)$ with $|u|,|u_t|,|u_x|,|u_{xx}|\le Me^{\alpha t}$ there. Then for $s\gt\alpha$:
$$\mathcal{L}\{u_t\}=sU(x,s)-u(x,0),$$ $$\mathcal{L}\{u_x\}=U_x(x,s),\qquad\mathcal{L}\{u_{xx}\}=U_{xx}(x,s).$$If also $u_{tt}$ is continuous with $|u_{tt}|\le Me^{\alpha t}$, then $\mathcal{L}\{u_{tt}\}=s^2U-s\,u(x,0)-u_t(x,0)$. The heat equation $u_t=ku_{xx}$ becomes $kU_{xx}-sU=-u(x,0)$, and the wave equation becomes $c^2U_{xx}-s^2U=-s\,u(x,0)-u_t(x,0)$. Both are ODEs in $x$ with parameter $s$. A boundary condition $u(0,t)=p(t)$ becomes $U(0,s)=P(s)$.
Proof. The $t$-rules are the derivative theorem of §6.2 applied for each fixed $x$. Integration by parts gives $\int_0^Te^{-st}u_t\,dt=e^{-sT}u(x,T)-u(x,0)$ $+s\int_0^Te^{-st}u\,dt$, and $e^{-sT}u(x,T)\to0$ by the growth bound. Apply this twice for $u_{tt}$. For the $x$-rules, differentiate under the integral sign. This is justified because $|e^{-st}u_x|\le Me^{-(s-\alpha)t}$, a bound that is integrable on $[0,\infty)$ and independent of $x$ (Leibniz's rule with dominated convergence). The same holds for $u_{xx}$. ∎
The transformed heat equation has solutions $e^{\pm x\sqrt{s/k}}$. Inverting $e^{-x\sqrt{s/k}}$ needs three new entries for the Laplace table of §6.2.
Proof. A lemma. For $q\gt0$ let $J=\int_0^\infty e^{-(v-q/v)^2}dv$. The substitution $v=q/r$ gives $J=\int_0^\infty e^{-(r-q/r)^2}\,\frac{q}{r^2}\,dr$. Adding this to the definition,
$$2J=\int_0^\infty e^{-(v-q/v)^2}\left(1+\frac{q}{v^2}\right)dv=\int_{-\infty}^{\infty}e^{-w^2}dw=\sqrt\pi ,$$with $w=v-q/v$, which increases from $-\infty$ to $\infty$ and has $dw=(1+q/v^2)\,dv$. Since $(v-q/v)^2=v^2-2q+q^2/v^2$, this says
$$\int_0^\infty e^{-v^2-q^2/v^2}dv=\frac{\sqrt\pi}{2}\,e^{-2q}.$$First transform. In $\int_0^\infty e^{-st}\frac{a}{2\sqrt\pi}t^{-3/2}e^{-a^2/(4t)}dt$ substitute $t=\frac{a^2}{4v^2}$. Then $t^{-3/2}=\frac{8v^3}{a^3}$, $|dt|=\frac{a^2}{2v^3}dv$, and $\frac{a^2}{4t}=v^2$, so the integral becomes
$$\frac{2}{\sqrt\pi}\int_0^\infty e^{-v^2-(a^2s/4)/v^2}\,dv=e^{-a\sqrt s},$$by the lemma with $q=\frac{a\sqrt s}{2}$. Second. Call the first function $g_a$. With $r=\frac{a}{2\sqrt\tau}$ we get $g_a(\tau)\,d\tau=-\frac{2}{\sqrt\pi}e^{-r^2}dr$, so $\int_0^tg_a(\tau)\,d\tau=\frac2{\sqrt\pi}\int_{a/(2\sqrt t)}^{\infty}e^{-r^2}dr=\operatorname{erfc}\frac{a}{2\sqrt t}$. By the integration theorem of §6.2, $\mathcal{L}\{\int_0^t\varphi\}=\mathcal{L}\{\varphi\}/s$. Third. Differentiate the second identity with respect to $a$. On the left, $-\partial_a\operatorname{erfc}\frac{a}{2\sqrt t}=\frac{1}{\sqrt{\pi t}}e^{-a^2/(4t)}$. On the right, $-\partial_a\frac{e^{-a\sqrt s}}{s}=\frac{e^{-a\sqrt s}}{\sqrt s}$. Moving $\partial_a$ inside $\mathcal{L}$ is allowed by dominated convergence, since the derivative is bounded by $(\pi t)^{-1/2}$, which is integrable against $e^{-st}$. ∎
- $x\in\mathbb{R}$, initial data given: d'Alembert (waves) or the heat kernel (heat) directly. The Fourier transform in $x$ is the systematic route and explains both. Use it for other constant-coefficient equations too, e.g. $u_t=ku_{xx}-bu_x-au$.
- $x\gt0$, the boundary value at $x=0$ a given function of $t$, simple initial data: the Laplace transform in $t$. Solve the ODE in $x$, discard the growing exponential $e^{+x\sqrt{s/k}}$ or $e^{+sx/c}$ because $u$ is bounded, match $U(0,s)=P(s)$, and invert. Waves invert by the second shifting theorem, heat by Theorem 10.33, and a general $p(t)$ by the convolution theorem.
- $x\gt0$, zero value or zero flux at $x=0$, initial data given: odd or even extension (Theorem 10.24, Exercise III.9), or the Fourier sine or cosine transform.
- Bounded interval: separation of variables and Fourier series (§10.2). The Laplace transform also works, but inverting it needs residues.
- Corners do not smooth out for waves. A plucked string $f=(1-|x|)_+$ has no classical solution. d'Alembert's formula gives a generalised one whose kinks travel along $x=\pm1\pm ct$ and $x=\pm ct$ forever. Heat, in contrast, smooths a corner instantly.
- Huygens' principle fails in one dimension. Compactly supported velocity data leave the permanent displacement $\frac1{2c}\int g$ behind the front (Example 1(b), Example 2(b)). Only when $g\equiv0$, or $\int g=0$, does the string return to rest.
- Domain of dependence ≠ two points. $u(x_0,t_0)$ depends on $f$ at two points only, but on $g$ on the whole interval $[x_0-ct_0,x_0+ct_0]$, and on a forcing $F$ in the whole solid triangle.
- Kinetic energy alone is not conserved, and neither is $\int u\,dx$ in general. Only $E=K+P$ is (Example 2(d)).
- Odd versus even. A fixed end needs the odd extension (inverted reflection) and a free end the even one (upright reflection). Mixing them up flips the sign of the reflected pulse.
- Uniqueness for heat needs a growth condition. Tychonoff constructed a non-zero solution of $u_t=u_{xx}$ on $\mathbb{R}\times(0,\infty)$, continuous up to $t=0$ with $u(x,0)=0$. It grows faster than any $e^{ax^2}$. Theorem 10.31(4) assumes boundedness, and growth up to $e^{ax^2}$ would also do.
- Infinite speed of propagation. For $f\ge0$ with compact support and $f\not\equiv0$, $u(x,t)\gt0$ at every $x$ for every $t\gt0$. "Heat has not yet reached $x=10$" is false, although $u$ there is tiny.
- The backward heat equation is ill-posed. $\frac1ne^{-kn^2t}\sin nx$ has small data at $t=0$ and is huge at $t=-1$.
- No Fourier transform does not mean no solution. $\cos x$, $1$, $x^2$ and the step $H$ are not in $L^1$, but the kernel formula still solves the heat problem. For instance $f=x^2$ gives $u=x^2+2kt$, a valid but unbounded solution.
- Fourier conventions. With $\frac1{\sqrt{2\pi}}$ in front, $e^{-x^2/2}$ is its own transform and $\mathcal{F}\{e^{-|x|}\}=\sqrt{2/\pi}\,\frac{1}{1+\omega^2}$. Check the convention before comparing options.
- Keep the decaying exponential. In a Laplace solution on $x\gt0$, boundedness kills $e^{+x\sqrt{s/k}}$. Forgetting this leaves an unknown constant, and keeping the wrong exponential gives a solution that blows up as $x\to\infty$.
- erf versus erfc. A rod initially at $0$ with its end held at $T_0$ has $u=T_0\operatorname{erfc}\frac{x}{2\sqrt{kt}}$. A rod initially at $T_0$ with its end held at $0$ has $u=T_0\operatorname{erf}\frac{x}{2\sqrt{kt}}$. The argument has $2\sqrt{kt}$, not $\sqrt{kt}$ or $2kt$.
- CSIR Unit 3 · §4.3 The Heat and Wave Equations — the heat equation on a rod with Dirichlet and Neumann ends, steady states, the maximum principle and the energy method on an interval, the vibrating string by separation of variables, conservation of energy and periodicity, and a second proof of d'Alembert's formula through characteristic coordinates.
Two Labs
Exercises
- 1 Solve $u_{tt}=9u_{xx}$ with $u(x,0)=\sin x$ and $u_t(x,0)=3\cos x$, and describe the solution in words.
- 2 Solve $u_{tt}=u_{xx}$ with $u(x,0)=x^2$ and $u_t(x,0)=x$.
- 3 For $u_{tt}=4u_{xx}$ with $f=0$ and $g=e^{-x^2}$, express $u(0,t)$ through $\operatorname{erf}$ and find $\lim_{t\to\infty}u(0,t)$.
- 4 For $u_{tt}=4u_{xx}$, find the domain of dependence of the point $(3,2)$ and the domain of influence of the interval $[0,1]$.
- 5 Prove that the solution of the Cauchy problem is a single wave $G(x-ct)$ moving right if and only if $g=-cf'$.
- 6 Prove that if $f$ and $g$ are odd, then $u(0,t)=0$ for all $t$, and that if they are even, then $u_x(0,t)=0$.
- 7 Parallelogram rule. Let $A,B,C,D$ be the vertices, in order, of a parallelogram whose sides are characteristics of $u_{tt}=c^2u_{xx}$. Prove that every $C^2$ solution satisfies $u(A)+u(C)=u(B)+u(D)$.
- 8 Let $f$ and $g$ have period $2L$. Show that $u(x,t+\tfrac{2L}c)-u(x,t)=\tfrac1c\int_0^{2L}g$. Deduce that $u$ is periodic in $t$ if and only if $\int_0^{2L}g=0$.
- 9 Equipartition. Let $f=0$ and let $g\in C^1$ vanish outside $[-a,a]$. Prove that the kinetic and potential energies are equal, $K(t)=P(t)=\tfrac12E$, for all $t\ge a/c$.
- 1 Solve $u_{tt}-u_{xx}=2$ with zero initial data, using Duhamel's formula.
- 2 Solve $u_{tt}-c^2u_{xx}=e^{-t}$ with zero initial data.
- 3 Solve $u_{tt}-u_{xx}=x$ with $u(x,0)=0$ and $u_t(x,0)=1$.
- 4 Solve $u_{tt}-4u_{xx}=\cos x$ with zero initial data, and find $u(0,\tfrac\pi2)$.
- 5 Show from Duhamel's formula that if $F=0$ in the triangle $\Delta(x_0,t_0)$, then the forced part $u_p(x_0,t_0)$ vanishes. Give an $F$, vanishing on the initial line, for which $u_p(0,1)\ne0$.
- 6 A semi-infinite string $x\gt0$ with $c=1$ is fixed at $x=0$, with $f=0$ and $g(x)=xe^{-x^2}$. Find $u(x,t)$ and $u(1,1)$.
- 7 A semi-infinite string $x\gt0$ with $c=1$ has a free end at $x=0$, and $f(x)=e^{-(x-3)^2}$, $g=0$. Find $u(0,3)$ and compare it with the height of each travelling half-pulse.
- 8 Prove that a $C^2$ solution of the fixed-end problem of Theorem 10.24 can exist only if $f(0)=g(0)=f''(0)=0$.
- 1 Find $\mathcal{F}\{xe^{-x^2}\}$ in two ways: by Theorem 10.26(4), and by writing $xe^{-x^2}=-\tfrac12\left(e^{-x^2}\right)'$.
- 2 Find $\mathcal{F}\{e^{-|x-2|}\}$.
- 3 Find $\mathcal{F}\{\chi_{[0,1]}\}$.
- 4 Show that $\chi_{[-1,1]}*\chi_{[-1,1]}=(2-|x|)_+$, and deduce the transform of the triangle $(2-|x|)_+$.
- 5 Prove that if $f\in L^1$ is real and even, then $\hat f$ is real and even, with $\hat f(\omega)=2\int_0^\infty f(x)\cos\omega x\,dx$.
- 6 Solve $u_t=2u_{xx}$ with $u(x,0)=e^{-x^2/2}$, and find $u(0,2)$.
- 7 Solve $u_t=u_{xx}$ with $u(x,0)=1$ for $|x|\lt1$ and $0$ otherwise. Express $u$ through $\operatorname{erf}$ and evaluate $u(0,1)$.
- 8 Show that the kernel formula with $f(x)=x^3$ gives $u=x^3+6ktx$, and check that it solves $u_t=ku_{xx}$.
- 9 Half-line by images. Show that $u(x,t)=\int_0^\infty K(x,y,t)f(y)\,dy$, with $K=\Phi(x-y,t)-\Phi(x+y,t)$, solves $u_t=ku_{xx}$ on $x\gt0$ with $u(0,t)=0$ and $u(x,0)=f(x)$ (for bounded continuous $f$ with $f(0)=0$). Compute it for $f\equiv1$ and compare with Example 7(a).
- 10 Solve $u_t=ku_{xx}-au$ ($a$ constant) with $u(x,0)=f(x)$ by substituting $u=e^{-at}v$.
- 1 Find $\mathcal{L}^{-1}\left\{\dfrac{e^{-3\sqrt s}}{s}\right\}$.
- 2 Find $\mathcal{L}^{-1}\left\{\dfrac{e^{-2\sqrt s}}{\sqrt s}\right\}$.
- 3 Modelling. A long rod ($k=1\ \text{cm}^2/\text{s}$) initially at $0\,^\circ\text{C}$ has its end held at $100\,^\circ\text{C}$ from $t=0$. Find the temperature at $x=1$ cm after $1$ s, and the time at which the point $x=2$ cm reaches $50\,^\circ\text{C}$.
- 4 Solve $u_t=ku_{xx}$ on $x\gt0$ with $u(x,0)=20$, $u(0,t)=0$ and $u$ bounded, by the Laplace transform.
- 5 Solve $u_{tt}=4u_{xx}$ on $x\gt0$ with zero initial data, $u(0,t)=t^2$ and $u$ bounded, and find $u(1,1)$.
- 6 A string falling under gravity. Solve $u_{tt}=c^2u_{xx}-g_0$ on $x\gt0$ with $u(x,0)=u_t(x,0)=0$, $u(0,t)=0$ and $u$ bounded for each $t$. Describe the shape.
- 7 Use the convolution theorem to show that the bounded solution of $u_t=ku_{xx}$ on $x\gt0$ with $u(x,0)=0$ and $u(0,t)=p(t)$ ($p$ bounded and continuous) is $$u(x,t)=\int_0^tp(\tau)\,\frac{x\,e^{-x^2/(4k(t-\tau))}}{2\sqrt{\pi k}\,(t-\tau)^{3/2}}\,d\tau .$$
- 8 Explain why the Laplace transform in $t$ is not suited to the backward heat problem, and why, for a bounded solution, the transform in $t$ of the heat equation on $x\gt0$ cannot contain $e^{+x\sqrt{s/k}}$.
E1. GATE MA · MCQ Let $u$ solve $u_{tt}=4u_{xx}$ for $x\in\mathbb{R}$, $t\gt0$, with $u(x,0)=\sin x$ and $u_t(x,0)=4x$. Then $u\left(\tfrac\pi2,\tfrac\pi4\right)$ equals (A) $\dfrac{\pi^2}{2}$; (B) $1+\dfrac{\pi^2}{2}$; (C) $\pi^2$; (D) $2\pi^2$.
E2. GATE MA · MCQ Let $u$ be the bounded solution of $u_t=u_{xx}$ on $\mathbb{R}\times(0,\infty)$ with $u(x,0)=\cos^2x$. Then $u\left(0,\tfrac{\log2}{4}\right)$ equals (A) $\tfrac34$; (B) $\tfrac12$; (C) $\tfrac12\left(1+2^{-1/4}\right)$; (D) $1$.
E3. GATE MA · MSQ CSIR Part C Let $f\in C^2(\mathbb{R})$ and $g\in C^1(\mathbb{R})$ vanish outside $[-1,1]$, not both identically zero, and let $u$ solve $u_{tt}=u_{xx}$, $u(x,0)=f$, $u_t(x,0)=g$. Which of the following is/are true for every such $f,g$? (One or more options may be correct.) (A) $u(x,t)=0$ whenever $|x|\gt t+1$. (B) $u(5,t)=0$ for all $t\ge6$. (C) If $g\equiv0$, then $u(x,t)=0$ whenever $|x|\le t-1$. (D) $\int_{-\infty}^{\infty}\left(u_t^2+u_x^2\right)dx$ does not depend on $t$.
E4. GATE MA · MSQ Consider $u_t=u_{xx}$ on $\mathbb{R}\times(0,\infty)$. Which of the following statements is/are true? (One or more options may be correct.) (A) If $f\ge0$ is continuous, not identically zero and vanishes outside $[0,1]$, then the solution $u=\Phi*f$ satisfies $u(100,1)\gt0$. (B) If $u(x,0)=e^{-x^2}$ and $u=\Phi*f$, then $u(0,t)=\dfrac1{1+4t}$. (C) The only solution of $u_t=u_{xx}$ that is continuous on $\mathbb{R}\times[0,\infty)$ with $u(x,0)=0$ is $u\equiv0$. (D) If $f\in L^1$ is bounded and continuous and $u=\Phi*f$, then $\int u(x,t)\,dx=\int f(x)\,dx$ for all $t\gt0$.
E5. GATE MA · NAT Let $u$ be given by Duhamel's formula for $u_{tt}-u_{xx}=F(x,t)$ with zero initial data, where $F(x,t)=\max(0,1-|x|)$ for all $t\ge0$. The value of $u(0,2)$ is ______ (correct to two decimal places).
E6. GATE MA · NAT Let $u$ be the bounded solution of $u_{tt}=4u_{xx}$ for $x\gt0$, $t\gt0$, with $u(x,0)=u_t(x,0)=0$ and $u(0,t)=te^{-t}$. The value of $u(3,4)$ is ______ (correct to two decimal places).
Suggested plan — 3 lectures × 60 min.
- Waves on the line. The factorisation of the wave operator, Theorem 10.21 by the transport-equation proof (link it to §10.1), Example 1. Definition 10.12 and Theorem 10.22, drawn on the board as cones. Project Interactive 1 in "plucked" and "struck" modes and ask the class to predict the value at the probe before releasing Play. Energy (Theorem 10.23), then Theorem 10.24 in the two half-line modes of the lab, and Examples 2 and 3.
- Forcing and the Fourier transform. Duhamel's principle (Theorem 10.25) with the impulse picture first, then the proof. Example 4 by both routes. Definition 10.13, Theorem 10.26 (prove (3) and (5)), the Gaussian (Theorem 10.27) by its ODE, inversion and Plancherel as statements with the Gaussian-regularisation idea, the table, Example 5.
- Heat and the Laplace transform. Derive the heat kernel by the Fourier transform. Theorem 10.29, the maximum principle (Theorem 10.30) and uniqueness (Theorem 10.31(4)) carefully. Project Interactive 2: "box with ripples" shows the high frequencies dying first, and the $u(5,t)$ readout shows infinite speed. Example 6. Then Theorems 10.32–10.33 (state 10.33 and prove only the lemma), the strategy box, Example 7, and GATE corner E2 and E6 live.
Discussion questions. (1) Both equations are linear with constant coefficients, yet one has finite and the other infinite speed of propagation. Which step of each solution formula is responsible? (2) Why does a struck string not return to rest in one dimension, while a sound in a room does die away? (3) In Example 7(a) the gradient at the end is infinite at $t=0^+$. Is that physically reasonable, and what does the model ignore?
Misconceptions to watch for. Using $\tfrac1c$ instead of $\tfrac1{2c}$; thinking the domain of dependence is the two points $x_0\pm ct_0$; assuming a disturbance always passes and leaves the string at rest; using the even extension for a fixed end; mixing Fourier conventions; claiming uniqueness for the heat equation without a growth condition; writing $\operatorname{erfc}\frac{x}{\sqrt{kt}}$; keeping the growing exponential in a Laplace solution.
Suggested homework. I(1, 3, 4, 7, 9), II(2, 4, 6, 7, 8), III(1, 4, 6, 7, 9), IV(3, 4, 6, 7); GATE corner E3, E5.
Research question. In 1862 William Thomson (later Lord Kelvin) estimated the age of the Earth from one measurement: how fast temperature increases with depth in mines. How did a half-line heat problem give him a number, why was the number wrong by a factor of about fifty, and what does the episode teach about the assumptions behind a model?
Why it matters. Kelvin modelled the Earth's crust as a half-space $x\gt0$ ($x$ the depth). He assumed it started at a uniform molten temperature $T_0$ and had its surface held at $0$ ever since. That is Exercise IV.4. The surface gradient $T_0/\sqrt{\pi kt}$ then determines $t$. The same half-line model is used today to date the cooling of ocean floor away from mid-ocean ridges, to design the depth of water pipes against frost, and to analyse laser heating of materials.
What to do.
- From $u=T_0\operatorname{erf}\frac{x}{2\sqrt{kt}}$ derive $u_x(0,t)=\frac{T_0}{\sqrt{\pi kt}}$, and solve for the age $t$ in terms of $T_0$, $k$ and the measured gradient $G$.
- Use values close to Kelvin's: $T_0\approx3900\,^\circ\text{C}$, $k\approx1.2\times10^{-6}\ \text{m}^2/\text{s}$, $G\approx0.036\,^\circ\text{C}/\text{m}$. Compute $t$ in years.
- Find how sensitive $t$ is to each input: what happens if $G$ is halved, or $k$ doubled?
- Read about the two physical effects the model leaves out: heat produced by radioactive decay in the crust, and convection in the mantle. Explain qualitatively how each changes the predicted age.
Tools: pencil and paper for 1 and 3; a calculator or Python for 2; an encyclopaedia or a history-of-science article for 4.
Hand in: the formula for $t$, the computed age, a short sensitivity table, and a paragraph on what the model got right (the $\sqrt t$ law) and what it got wrong.
Going further. Ocean-floor depth below a ridge grows like $\sqrt{\text{age}}$ for exactly the reason of Example 7(a). Look up the observed depth–age curve of the Pacific floor and test the $\sqrt t$ law against it.
🧪 Self-check quiz
Q1. For $u_{tt}=4u_{xx}$ on $\mathbb{R}$, the domain of dependence of the point $(x,t)=(1,3)$ is
Q2. With $\hat f(\omega)=\int f(x)e^{-i\omega x}dx$, the Fourier transform of $e^{-2|x|}$ is
Q3. The bounded solution of $u_t=u_{xx}$ with $u(x,0)=e^{-x^2}$ has $u(0,t)$ equal to
Q4. Taking the Laplace transform in $t$ of $u_t=ku_{xx}$ on $x\gt0$ with $u(x,0)=0$ and $u$ bounded gives $U(x,s)=$
Q5. A pulse travels along a semi-infinite string towards the end $x=0$, which is fixed. After reflection the pulse