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; the wave equation — d'Alembert's formula, domains of dependence and influence, the vibrating string by separation of variables and conservation of energy; and the heat kernel on the whole line.
The heat equation $u_t=k\,u_{xx}$ and the wave equation $u_{tt}=c^2u_{xx}$ differ by a single time derivative, and §4.1 showed that this is the difference between a parabolic and a hyperbolic equation. This section makes that difference concrete. Separation of variables solves both on a finite interval, and the solutions look alike — Fourier series whose $n$-th term carries a time factor — but the time factors tell opposite stories: $e^{-kn^2\pi^2t/L^2}$ for heat, so high modes die almost instantly and every initial temperature is smoothed and forgotten, against $\cos(n\pi ct/L)$ for waves, so every mode rings for ever and corners in the initial shape travel without blurring. Heat obeys a maximum principle and loses "energy" $\int u^2$; waves have no maximum principle but conserve the energy $\frac12\int(u_t^2+c^2u_x^2)$. On the whole line, waves move at the finite speed $c$ (d'Alembert's formula), while heat is felt everywhere at once (the heat kernel).
These equations model cooling rods and cables, diffusion of pollutants and of heat in soil, option prices after a change of variables (the Black–Scholes equation of §4.1), vibrating strings of musical instruments, and sound and signals in one dimension. CSIR NET examines them directly: "find $u(x,t)$ at this point", "which mode decays slowest?", "what is the final temperature of an insulated rod?", "evaluate d'Alembert's formula", "what is the domain of dependence?", "which quantity is conserved?", and Part C statements on uniqueness, maximum principles and periodicity. Throughout, $k\gt0$ is the diffusivity, $c\gt0$ the wave speed, $L\gt0$ the length of the rod or string, and the Fourier series facts of §4.2 are used freely.
Where the Equations Come From; the Problems to Solve
Heat. Consider a thin rod of cross-section $A$, density $\rho$, specific heat $\sigma$ and thermal conductivity $\kappa$, insulated along its length, with temperature $u(x,t)$. The heat in a piece $[a,b]$ is $\int_a^b\rho\sigma Au\,dx$, and Fourier's law says heat flows from hot to cold at the rate $-\kappa Au_x$ across each section. Conservation of energy for $[a,b]$ reads $$\frac{d}{dt}\int_a^b\rho\sigma A\,u\,dx=\kappa A\,u_x(b,t)-\kappa A\,u_x(a,t)=\int_a^b\kappa A\,u_{xx}\,dx .$$ Since $[a,b]$ is arbitrary and the integrands are continuous, $\rho\sigma u_t=\kappa u_{xx}$, i.e. $u_t=ku_{xx}$ with $k=\kappa/(\rho\sigma)$. An end held at a fixed temperature gives a Dirichlet condition $u=T$; an insulated end gives a Neumann condition $u_x=0$ (no flux); an end losing heat to surroundings at temperature $T_e$ by Newton's law gives the Robin condition $-\kappa u_x=h(u-T_e)$ at $x=L$.
Waves. A string of linear density $\rho$ is stretched with tension $T$ and vibrates in a plane with small transverse displacement $u(x,t)$. The tension acts along the string; its vertical component at a section is $T\sin\alpha\approx T\tan\alpha=Tu_x$ when the slope is small, and the horizontal components balance. Newton's second law for the piece $[a,b]$ is $\int_a^b\rho u_{tt}\,dx=Tu_x(b,t)-Tu_x(a,t)$, so $u_{tt}=c^2u_{xx}$ with $c=\sqrt{T/\rho}$. A fixed end gives $u=0$; an end attached to a ring sliding without friction on a vertical rod gives $u_x=0$ (a "free" end).
The heat problem on a rod is: $u_t=ku_{xx}$ for $0\lt x\lt L$, $t\gt0$; $u(x,0)=f(x)$; and one boundary condition at each end, of Dirichlet ($u=T_i$), Neumann ($u_x=0$) or Robin type. A classical solution is a function $u$ continuous on $[0,L]\times[0,\infty)$ (for Dirichlet ends; this forces the compatibility $f(0)=T_1$, $f(L)=T_2$), with $u_t,u_{xx}$ continuous in $(0,L)\times(0,\infty)$, $u_x$ continuous up to every Neumann or Robin end for $t\gt0$, satisfying all the conditions.
The wave problem on a string is: $u_{tt}=c^2u_{xx}$ for $0\lt x\lt L$, $t\gt0$; $u(x,0)=f(x)$, $u_t(x,0)=g(x)$; and a condition at each end. A classical solution is $u\in C^2([0,L]\times[0,\infty))$ satisfying all of them. On the whole line ($-\infty\lt x\lt\infty$) there are no boundary conditions and the problem is the Cauchy problem.
A steady state (equilibrium) of an initial–boundary value problem is a time-independent solution $w(x)$ of the equation and the boundary conditions (the initial condition is ignored). For the heat equation $kw''=0$, so $w$ is linear: with ends at $T_1$ and $T_2$, $w(x)=T_1+(T_2-T_1)\dfrac xL$; with both ends insulated every constant is a steady state.
The Heat Equation on a Rod
Separate $u=X(x)T(t)$ in $u_t=ku_{xx}$: $XT'=kX''T$, so $\dfrac{T'}{kT}=\dfrac{X''}{X}=-\lambda$. With $u(0,t)=u(L,t)=0$ the spatial factor solves the Dirichlet Sturm–Liouville problem $X''+\lambda X=0$, $X(0)=X(L)=0$ of §2.3: $\lambda_n=(n\pi/L)^2$, $X_n=\sin\frac{n\pi x}L$, and then $T_n=e^{-k\lambda_nt}$. With insulated ends the eigenfunctions are $\cos\frac{n\pi x}L$, $n\ge0$, and the mode $n=0$ is the constant.
- (Ends at temperature $0$) Let $f$ be continuous and piecewise $C^1$ on $[0,L]$ with $f(0)=f(L)=0$, and $b_n=\frac2L\int_0^Lf(x)\sin\frac{n\pi x}{L}\,dx$. Then $$u(x,t)=\sum_{n=1}^\infty b_n\,e^{-k n^2\pi^2t/L^2}\sin\frac{n\pi x}{L}$$ is a classical solution of $u_t=ku_{xx}$, $u(0,t)=u(L,t)=0$, $u(x,0)=f(x)$. It is $C^\infty$ on $[0,L]\times(0,\infty)$, and $|u(x,t)|\le\big(\sum|b_n|\big)e^{-k\pi^2t/L^2}$.
- (Insulated ends) Let $f$ be continuous and piecewise $C^1$ on $[0,L]$, and $a_n=\frac2L\int_0^Lf(x)\cos\frac{n\pi x}{L}\,dx$. Then $$u(x,t)=\frac{a_0}{2}+\sum_{n=1}^\infty a_n\,e^{-kn^2\pi^2t/L^2}\cos\frac{n\pi x}{L}$$ is a classical solution with $u_x(0,t)=u_x(L,t)=0$. The mean temperature $\frac1L\int_0^Lu(x,t)\,dx=\frac{a_0}2$ is constant in time, and $|u(x,t)-\frac{a_0}2|\le\big(\sum_{n\ge1}|a_n|\big)e^{-k\pi^2t/L^2}$.
Proof. (1) For $t\gt0$. Fix $t_0\gt0$. The $n$-th term differentiated $i$ times in $x$ and $j$ times in $t$ is bounded, for $t\ge t_0$, by $$|b_n|\Big(\frac{n\pi}L\Big)^{i}\Big(\frac{kn^2\pi^2}{L^2}\Big)^{j}e^{-kn^2\pi^2t_0/L^2}\le B\,C_{ij}\,n^{i+2j}e^{-kn^2\pi^2t_0/L^2},\qquad B=\max_n|b_n|,$$ and $\sum_nn^me^{-\gamma n^2}\lt\infty$ for every $m$ and $\gamma\gt0$. By the Weierstrass M-test all differentiated series converge uniformly on $[0,L]\times[t_0,\infty)$, so $u$ is $C^\infty$ there and may be differentiated term by term. Each term satisfies the equation and the boundary conditions, hence so does $u$.
Up to $t=0$. The odd $2L$-periodic extension of $f$ is continuous (because $f(0)=f(L)=0$) and piecewise $C^1$, so $\sum|b_n|\lt\infty$ by the uniform convergence theorem for Fourier series (§4.2). Each term is bounded by $|b_n|$ on $[0,L]\times[0,\infty)$; the series converges uniformly there, $u$ is continuous, and $u(x,0)=\sum b_n\sin\frac{n\pi x}L=f(x)$. The bound follows from $e^{-kn^2\pi^2t/L^2}\le e^{-k\pi^2t/L^2}$ for $n\ge1$.
(2) The same argument with cosines; the even extension of $f$ is continuous without any end-point condition, so $\sum|a_n|\lt\infty$. The derivative series $u_x=-\sum a_n\frac{n\pi}Le^{-\cdots}\sin\frac{n\pi x}L$ vanishes at $x=0,L$ for $t\gt0$. Integrating term by term (uniform convergence), $\int_0^L\cos\frac{n\pi x}L\,dx=0$ for $n\ge1$, so $\frac1L\int_0^Lu\,dx=\frac{a_0}2$. (Directly: $\frac{d}{dt}\int_0^Lu\,dx=k\int_0^Lu_{xx}\,dx=k\big[u_x\big]_0^L=0$ — no heat crosses an insulated end.) ∎
Two remarks complete the picture. Non-zero end temperatures: if $u(0,t)=T_1$ and $u(L,t)=T_2$, write $u=w+v$ with the steady state $w$ of Definition 4.10; then $v$ solves the problem with zero ends and initial value $f-w$, and $u\to w$ uniformly, exponentially fast at the rate $k\pi^2/L^2$. Rough data: if $f$ is only piecewise continuous (for example $f\equiv T_0$ with ice-cold ends, which violates compatibility at the corners), the series still converges for $t\gt0$ and gives a $C^\infty$ solution there, which tends to $f$ at every point of continuity as $t\to0^+$ — but not uniformly. The heat equation smooths instantly: whatever the data, the solution is $C^\infty$ for $t\gt0$.
Uniqueness of these solutions, and their continuous dependence on the data, come from two principles that do not use series at all.
Let $Q_T=(0,L)\times(0,T]$ and let $\Gamma_T=\big([0,L]\times\{0\}\big)\cup\big(\{0,L\}\times[0,T]\big)$ be its parabolic boundary (bottom and sides, not the top). Let $u$ be continuous on $\overline{Q_T}$ with $u_t,u_{xx}$ continuous in $Q_T$, and suppose $u_t-ku_{xx}\le0$ in $Q_T$. Then $$\max_{\overline{Q_T}}u=\max_{\Gamma_T}u .$$ If $u_t-ku_{xx}\ge0$, then $\min_{\overline{Q_T}}u=\min_{\Gamma_T}u$; a solution of the heat equation satisfies both.
Proof. For $\varepsilon\gt0$ let $v=u+\varepsilon x^2$; then $v_t-kv_{xx}=u_t-ku_{xx}-2k\varepsilon\lt0$ in $Q_T$. The continuous function $v$ attains its maximum over the compact set $\overline{Q_T}$ at some $(x_0,t_0)$. Suppose $(x_0,t_0)\notin\Gamma_T$, i.e. $0\lt x_0\lt L$ and $0\lt t_0\le T$. As a function of $x$, $v(\cdot,t_0)$ has an interior maximum at $x_0$, so $v_{xx}(x_0,t_0)\le0$. As a function of $t$ on $(0,t_0]$, $v(x_0,\cdot)$ is largest at $t_0$, so $v_t(x_0,t_0)\ge0$ (a one-sided derivative if $t_0=T$; it is zero if $t_0\lt T$). Then $v_t-kv_{xx}\ge0$ at $(x_0,t_0)$, a contradiction. Hence $\max_{\overline{Q_T}}v=\max_{\Gamma_T}v$, and $$\max_{\overline{Q_T}}u\le\max_{\overline{Q_T}}v=\max_{\Gamma_T}v\le\max_{\Gamma_T}u+\varepsilon L^2 .$$ Let $\varepsilon\to0$. The statement for the minimum follows by applying this to $-u$. ∎
Physically: without sources, a rod cannot become hotter than the hottest temperature it started with or was held at on its ends. The top of $Q_T$ is excluded for a reason — the maximum can occur there if, for example, the ends are being heated.
The problem $u_t=ku_{xx}$ in $(0,L)\times(0,T]$, $u(x,0)=f(x)$, $u(0,t)=\alpha(t)$, $u(L,t)=\beta(t)$ has at most one solution continuous on $\overline{Q_T}$ with $u_t,u_{xx}$ continuous in $Q_T$. If $u_1,u_2$ are solutions with data $(f_i,\alpha_i,\beta_i)$, then $$\max_{\overline{Q_T}}|u_1-u_2|\le\max\Big\{\max_{[0,L]}|f_1-f_2|,\ \max_{[0,T]}|\alpha_1-\alpha_2|,\ \max_{[0,T]}|\beta_1-\beta_2|\Big\},$$ and if $f_1\le f_2$, $\alpha_1\le\alpha_2$, $\beta_1\le\beta_2$, then $u_1\le u_2$.
Proof. $w=u_1-u_2$ solves the heat equation, so by Theorem 4.20 both $w$ and $-w$ are bounded by their maxima on $\Gamma_T$, where $w$ equals the differences of the data. Equal data give $w\equiv0$; ordered data give $w\le\max_{\Gamma_T}w\le0$. ∎
Let $u$ be a classical solution of $u_t=ku_{xx}$ on $(0,L)\times(0,\infty)$ with $u$, $u_x$, $u_t$ continuous on $[0,L]\times(0,\infty)$, $u$ continuous on $[0,L]\times[0,\infty)$, and at each end either $u=0$ or $u_x=0$. Then $E(t)=\int_0^Lu(x,t)^2\,dx$ satisfies $$E'(t)=-2k\int_0^Lu_x^2\,dx\le0 .$$ Consequently each of these problems has at most one classical solution. If both ends are Dirichlet ends, $E(t)\le e^{-2k\pi^2t/L^2}E(0)$.
Proof. On $[0,L]\times[t_0,t_1]$ with $t_0\gt0$ the integrand $2uu_t$ is continuous, so we may differentiate under the integral sign: $E'=\int_0^L2uu_t\,dx=2k\int_0^Luu_{xx}\,dx$ $=2k\big[uu_x\big]_0^L-2k\int_0^Lu_x^2\,dx$, and the boundary term vanishes at each end because $u=0$ or $u_x=0$ there. If $u_1,u_2$ solve the same problem, their difference $w$ has $E_w(0)=0$, $E_w\ge0$, $E_w$ non-increasing on $(0,\infty)$ and continuous at $0$; so $E_w\equiv0$ and $w\equiv0$. For two Dirichlet ends, the Rayleigh quotient of $X''+\lambda X=0$, $X(0)=X(L)=0$ (§2.3) gives $\int_0^Lu_x^2\ge\frac{\pi^2}{L^2}\int_0^Lu^2$, so $E'\le-\frac{2k\pi^2}{L^2}E$, and Gronwall's inequality (§1.1) gives the decay. ∎
The energy method also covers Neumann and mixed ends, where Theorem 4.20 does not immediately give uniqueness, and Robin ends $u_x(L,t)=-\sigma u(L,t)$ with $\sigma\ge0$ (the boundary term is then $-2k\sigma u(L,t)^2\le0$).
The Wave Equation on the Line: d'Alembert's Formula
§4.1 showed that the characteristics of $u_{tt}=c^2u_{xx}$ are the lines $x\pm ct=\text{const}$, and that in the characteristic coordinates $\xi=x+ct$, $\eta=x-ct$ the equation becomes $u_{\xi\eta}=0$. This gives the whole solution of the Cauchy problem.
Let $f\in C^2(\mathbb{R})$ and $g\in C^1(\mathbb{R})$. The Cauchy problem $u_{tt}=c^2u_{xx}$ ($x\in\mathbb{R}$, $t\gt0$), $u(x,0)=f(x)$, $u_t(x,0)=g(x)$ has exactly one solution $u\in C^2(\mathbb{R}\times[0,\infty))$, namely $$u(x,t)=\frac12\big[f(x+ct)+f(x-ct)\big]+\frac1{2c}\int_{x-ct}^{x+ct}g(s)\,ds .$$
Proof. Every solution has this form. With $\xi=x+ct$, $\eta=x-ct$ and $u(x,t)=v(\xi,\eta)$, the chain rule gives $u_{tt}-c^2u_{xx}=-4c^2v_{\xi\eta}$ (the transformation formula for the principal part in §4.1, or directly). The half-plane $t\ge0$ becomes the half-plane $\xi\ge\eta$, and each line $\xi=\text{const}$ or $\eta=\text{const}$ meets it in an interval. So $v_{\xi\eta}=0$ means $v_\xi$ does not depend on $\eta$, $v_\xi=\phi(\xi)$, and integrating, $v=F(\xi)+G(\eta)$ with $F'=\phi$: $u=F(x+ct)+G(x-ct)$ for some $F,G\in C^2$.
The data fix $F$ and $G$. At $t=0$: $F+G=f$ and $cF'-cG'=g$. Integrating the second, $F-G=\frac1c\int_0^xg(s)\,ds+K$. Hence $F(x)=\frac12f(x)+\frac1{2c}\int_0^xg+\frac K2$ and $G(x)=\frac12f(x)-\frac1{2c}\int_0^xg-\frac K2$, and $$u=F(x+ct)+G(x-ct)=\tfrac12\big[f(x+ct)+f(x-ct)\big]+\tfrac1{2c}\Big(\int_0^{x+ct}g-\int_0^{x-ct}g\Big),$$ which is the formula; in particular the solution is unique. Conversely, the formula is $C^2$ when $f\in C^2$ and $g\in C^1$ (the integral term is $C^2$ because its derivatives involve $g$ and $g'$ only), it has the form $F(x+ct)+G(x-ct)$, hence solves the equation, and $u(x,0)=f(x)$, $u_t(x,0)=\frac c2(f'(x)-f'(x))$ $+\frac1{2c}(cg(x)+cg(x))=g(x)$. ∎
So every solution is a superposition of two waves of fixed shape, $F$ moving left and $G$ moving right at speed $c$. An initial displacement with zero velocity splits into two half-size copies that separate; an initial velocity alone produces a plateau that spreads in both directions.
For a point $(x_0,t_0)$ with $t_0\gt0$, the two characteristics through it, $x-x_0=\pm c(t-t_0)$, cut the line $t=0$ at $x_0\mp ct_0$. The interval $[x_0-ct_0,\,x_0+ct_0]$ is the domain of dependence of $(x_0,t_0)$, and the triangle it forms with $(x_0,t_0)$ is the characteristic triangle. For a point $x_1$ of the initial line, the wedge $\{(x,t):|x-x_1|\le ct\}$ is its domain of influence.
- $u(x_0,t_0)$ depends only on $f(x_0\pm ct_0)$ and on $g$ restricted to the domain of dependence $[x_0-ct_0,x_0+ct_0]$.
- If $f$ and $g$ vanish outside $[a,b]$, then $u(x,t)=0$ whenever $x\lt a-ct$ or $x\gt b+ct$: the data influence only the union of the domains of influence of their points.
- If $u_1,u_2$ are the solutions with data $(f_1,g_1)$, $(f_2,g_2)$, then for every $t\ge0$ $$\sup_{x}|u_1(x,t)-u_2(x,t)|\le\sup|f_1-f_2|+t\,\sup|g_1-g_2| .$$
Proof. (1) Read it off d'Alembert's formula. (2) If $x\gt b+ct$ then $x\pm ct\gt b$ and the interval $[x-ct,x+ct]$ lies to the right of $b$, so every term vanishes; similarly for $x\lt a-ct$. (3) Apply the formula to $u_1-u_2$, whose data are $f_1-f_2$ and $g_1-g_2$: the first bracket is at most $\sup|f_1-f_2|$ in absolute value and the integral at most $\frac1{2c}\cdot2ct\,\sup|g_1-g_2|$. ∎
In one space dimension a disturbance does not leave the string at rest behind it when it comes from an initial velocity: after the plateau has passed a point, $u$ stays at $\frac1{2c}\int g$ (the sharp form of Huygens' principle fails in one dimension; it holds for the wave equation in three dimensions, which is why sound in air arrives and then stops). The formula also makes sense when $f$ is merely continuous and $g$ merely integrable; the resulting generalised solution has kinks that travel along characteristics — exactly the curves along which §4.1 allowed second derivatives to jump.
Boundaries and reflections. For a semi-infinite string $x\ge0$ with a fixed end $u(0,t)=0$, extend $f$ and $g$ to odd functions on $\mathbb{R}$ and apply d'Alembert's formula: the solution is odd in $x$, so it vanishes at $x=0$ (the method of images). For $x\lt ct$ this gives $$u(x,t)=\frac12\big[f(x+ct)-f(ct-x)\big]+\frac1{2c}\int_{ct-x}^{x+ct}g(s)\,ds,$$ while for $x\ge ct$ d'Alembert's formula holds unchanged. A pulse reflects from a fixed end upside down. A free end $u_x(0,t)=0$ corresponds to the even extension, and a pulse reflects from it the right way up (Exercise II.6). For a classical solution the odd extension must be $C^2$, which requires $f(0)=f''(0)=0$ and $g(0)=0$.
The Vibrating String: Separation of Variables and Energy
For a string with fixed ends, $u=X(x)T(t)$ gives $\frac{T''}{c^2T}=\frac{X''}{X}=-\lambda$, the same Sturm–Liouville problem as for the rod, and $T''+c^2\lambda_nT=0$ with $\omega_n=\frac{n\pi c}L$: the normal modes $\sin\frac{n\pi x}{L}\big(a_n\cos\omega_nt+b_n\sin\omega_nt\big)$. Mode $n$ vibrates with frequency $\nu_n=\frac{\omega_n}{2\pi}=\frac{n}{2L}\sqrt{T/\rho}$ and has $n-1$ fixed interior points (nodes); $\nu_1$ is the fundamental and $\nu_n=n\nu_1$ are its harmonics — the reason a string sounds musical, and the content of Mersenne's laws (pitch rises with tension, falls with length and density).
Let $f\in C^2[0,L]$ with $f(0)=f(L)=f''(0)=f''(L)=0$ and $g\in C^1[0,L]$ with $g(0)=g(L)=0$, and let $F$, $G$ be the odd $2L$-periodic extensions of $f$, $g$. Then the problem $u_{tt}=c^2u_{xx}$, $u(0,t)=u(L,t)=0$, $u(x,0)=f(x)$, $u_t(x,0)=g(x)$ has the unique classical solution $$u(x,t)=\frac12\big[F(x+ct)+F(x-ct)\big]+\frac1{2c}\int_{x-ct}^{x+ct}G(s)\,ds=\sum_{n=1}^\infty\Big(a_n\cos\frac{n\pi ct}{L}+b_n\sin\frac{n\pi ct}{L}\Big)\sin\frac{n\pi x}{L},$$ with $a_n=\frac2L\int_0^Lf(x)\sin\frac{n\pi x}L\,dx$ and $b_n=\frac{2}{n\pi c}\int_0^Lg(x)\sin\frac{n\pi x}{L}\,dx$. The solution is periodic in $t$ with period $2L/c$.
Proof. Smoothness of the extensions. $F$ is odd about $0$ and, being odd and $2L$-periodic, also odd about $L$: $F(L+s)=F(s-L)=-F(L-s)$. For an odd function built from a $C^2$ piece, $F'$ is automatically continuous at the reflection points ($F'$ is even about them), while $F$ and $F''$ (odd about them) are continuous there exactly when $f$ and $f''$ vanish at $0$ and $L$. So $F\in C^2(\mathbb{R})$; likewise $G\in C^1(\mathbb{R})$ because $g(0)=g(L)=0$.
It solves the problem. By Theorem 4.23 the d'Alembert function is a $C^2$ solution on $\mathbb{R}\times[0,\infty)$ with $u(x,0)=F(x)$, $u_t(x,0)=G(x)$, which equal $f,g$ on $[0,L]$. At $x=0$: $F(ct)+F(-ct)=0$ and $\int_{-ct}^{ct}G=0$ by oddness. At $x=L$: $F(L+ct)+F(L-ct)=0$ and $\int_{L-ct}^{L+ct}G=0$ because $F$, $G$ are odd about $L$. Uniqueness is Theorem 4.26 below.
Series form. By the uniform convergence theorem for Fourier series (§4.2), $F=\sum a_n\sin\frac{n\pi x}L$ and $G=\sum g_n\sin\frac{n\pi x}L$ uniformly, with $g_n$ the sine coefficients of $g$. Substituting and using $\tfrac12[\sin\alpha(x+ct)+\sin\alpha(x-ct)]=\sin\alpha x\cos\alpha ct$ and $\frac1{2c}\int_{x-ct}^{x+ct}\sin\alpha s\,ds=\frac{1}{c\alpha}\sin\alpha x\sin\alpha ct$ (term-by-term integration of a uniformly convergent series) gives the series with $b_n=\frac{L}{n\pi c}g_n$, as stated. Periodicity: $F(x\pm c(t+2L/c))=F(x\pm ct\pm2L)=F(x\pm ct)$, and the integral of $G$ over the shifted interval changes by integrals of $G$ over whole periods, which vanish because $G$ is odd. ∎
Let $u\in C^2([0,L]\times[0,\infty))$ solve $u_{tt}=c^2u_{xx}$, and at each end let either $u=0$ or $u_x=0$. Then the energy $$E(t)=\frac12\int_0^L\big(u_t^2+c^2u_x^2\big)\,dx$$ (kinetic plus potential, per unit density) is constant. Consequently the initial–boundary value problem has at most one classical solution. For the series solution of Theorem 4.25, $E=\dfrac L4\displaystyle\sum_{n\ge1}\omega_n^2\big(a_n^2+b_n^2\big)$ with $\omega_n=\frac{n\pi c}L$.
Proof. Differentiating under the integral sign (the integrand is $C^1$ on compact sets) and using the equation, $$E'(t)=\int_0^L\big(u_tu_{tt}+c^2u_xu_{xt}\big)\,dx=c^2\int_0^L\big(u_tu_{xx}+u_xu_{xt}\big)\,dx=c^2\int_0^L\big(u_tu_x\big)_x\,dx=c^2\big[u_tu_x\big]_0^L .$$ At a fixed end $u(0,t)=0$ for all $t$, so $u_t(0,t)=0$; at a free end $u_x=0$. Either way the boundary term vanishes. If two solutions have the same data, their difference $w$ has $E_w(0)=\frac12\int(0+c^2\cdot0)=0$, so $E_w\equiv0$, $w_t=w_x=0$, $w$ is constant, and $w=w(x,0)=0$. For the series, $u_t$ and $u_x$ are sine and cosine series in $x$ with coefficients $\omega_n(-a_n\sin\omega_nt+b_n\cos\omega_nt)$ and $\frac{n\pi}{L}(a_n\cos\omega_nt+b_n\sin\omega_nt)$; Parseval's identity on $[0,L]$ ($\int_0^L\sin^2=\int_0^L\cos^2=\frac L2$) gives $E=\frac L4\sum\omega_n^2\big[(-a_n\sin+b_n\cos)^2$ $+(a_n\cos+b_n\sin)^2\big]$ $=\frac L4\sum\omega_n^2(a_n^2+b_n^2)$. ∎
The Heat Kernel, and Heat versus Waves
The heat kernel (fundamental solution of the heat equation on $\mathbb{R}$) is $$\Phi(x,t)=\frac{1}{\sqrt{4\pi kt}}\,e^{-x^2/(4kt)},\qquad x\in\mathbb{R},\ t\gt0 .$$ It is the temperature produced by a unit amount of heat released at $x=0$ at time $0$: a Gaussian of total mass $1$ and standard deviation $\sqrt{2kt}$.
Let $f$ be bounded and continuous on $\mathbb{R}$ and put $u(x,t)=\displaystyle\int_{-\infty}^{\infty}\Phi(x-y,t)\,f(y)\,dy$ for $t\gt0$. Then:
- $u\in C^\infty(\mathbb{R}\times(0,\infty))$ and $u_t=ku_{xx}$;
- $u(x,t)\to f(x_0)$ as $(x,t)\to(x_0,0^+)$, so $u$ extended by $f$ is continuous on $\mathbb{R}\times[0,\infty)$;
- $\inf f\le u\le\sup f$; and if $f\ge0$ with $f\not\equiv0$, then $u(x,t)\gt0$ for every $x$ and every $t\gt0$ (infinite speed of propagation);
- (Tychonoff) $u$ is the only solution continuous on $\mathbb{R}\times[0,\infty)$ with $u(x,0)=f(x)$ that satisfies a growth bound $|u(x,t)|\le Ae^{ax^2}$; without such a bound uniqueness fails.
Proof sketch. (1) Direct differentiation: $\Phi_t=\Phi\big(-\frac1{2t}+\frac{x^2}{4kt^2}\big)$ and $\Phi_{xx}=\Phi\big(-\frac1{2kt}+\frac{x^2}{4k^2t^2}\big)$, so $\Phi_t=k\Phi_{xx}$. The Gaussian decay of $\Phi$ and of all its derivatives, uniformly for $t$ in compact subsets of $(0,\infty)$, justifies differentiating under the integral sign. (2) $\int\Phi(x,t)\,dx=1$ (substitute $x=\sqrt{4kt}\,s$ and use $\int e^{-s^2}ds=\sqrt\pi$), and for each $\delta\gt0$, $\int_{|x|\gt\delta}\Phi(x,t)\,dx\to0$ as $t\to0^+$. So $\Phi(\cdot,t)$ is an approximate identity, and the argument used for the Poisson kernel in §4.2 (split near and far from $x_0$, use uniform continuity on a neighbourhood) gives (2). (3) $\Phi\gt0$ with total mass $1$ makes $u$ a weighted average of $f$; if $f\ge0$ is positive somewhere, the integrand is positive on an interval. (4) The uniqueness proof applies the maximum principle to a large rectangle together with a comparison function that grows like $e^{ax^2}$; Tychonoff constructed a non-zero solution with $u(x,0)=0$ that grows faster than any such bound. ∎
For example, $f(x)=e^{-x^2}$ gives $u=\dfrac{1}{\sqrt{1+4kt}}\,e^{-x^2/(1+4kt)}$ (Exercise IV.2), and the step $f=1$ for $x\gt0$, $0$ for $x\lt0$ gives $u=\frac12\big[1+\operatorname{erf}\big(x/\sqrt{4kt}\big)\big]$ (Exercise IV.3) — positive at every $x$ as soon as $t\gt0$.
| heat $u_t=ku_{xx}$ | wave $u_{tt}=c^2u_{xx}$ | |
|---|---|---|
| type, data | parabolic; $u(x,0)$ only | hyperbolic; $u(x,0)$ and $u_t(x,0)$ |
| mode $n$ on $[0,L]$ | $e^{-kn^2\pi^2t/L^2}$: decays | $\cos,\sin\frac{n\pi ct}{L}$: never decays |
| regularity | $C^\infty$ for $t\gt0$, any data | as smooth as the data, no better |
| speed of propagation | infinite | exactly $c$ |
| time reversal | backward problem ill-posed | reversible ($t\mapsto-t$) |
| principle | maximum principle; $\int u^2$ decreases | no maximum principle; energy conserved |
| long-time behaviour | tends to the steady state | periodic (string) or passes by (line) |
- The backward heat equation is ill-posed. $u_n=\frac1ne^{-kn^2(t-T)}\sin nx$ has data at $t=T$ tending to $0$ uniformly and values at $t=0$ that blow up (Example 3(b)).
- The wave equation has no maximum principle. $u=\sin x\sin t$ on $[0,\pi]\times[0,\frac\pi2]$ vanishes on the bottom and sides but equals $1$ at $(\frac\pi2,\frac\pi2)$.
- Insulated rods do not cool to zero. With $u_x=0$ at both ends the mode $\lambda_0=0$ survives and $u\to\frac1L\int_0^Lf$. Forgetting the constant term is the commonest error in Neumann problems.
- Heat content is conserved only for insulated ends (Example 3(d)); with Dirichlet ends heat leaks out.
- Mixed ends cool four times more slowly. With $u(0,t)=0$, $u_x(L,t)=0$ the eigenvalues are $\big((n-\frac12)\frac\pi L\big)^2$; the slowest mode decays like $e^{-k\pi^2t/(4L^2)}$, not $e^{-k\pi^2t/L^2}$.
- Incompatible data. $f\equiv T_0$ with ends at $0$: the series solves the equation for $t\gt0$ but $u$ is discontinuous at the corners $(0,0)$, $(L,0)$, and the convergence near $t=0$ is not uniform (Gibbs).
- Uniqueness on the line needs a growth condition. Tychonoff's solution vanishes at $t=0$ but is not identically zero; it grows faster than $e^{ax^2}$.
- The factor in d'Alembert's formula is $\frac1{2c}$, the integration limits are $x\pm ct$, and the displacement term has $\frac12$: $\frac12[f(x+ct)+f(x-ct)]$, not $f(x+ct)+f(x-ct)$.
- Velocity data leave a wake. In one dimension $u(x,t)\to\frac1{2c}\int g$ at every fixed $x$; only displacement pulses pass by completely (Example 7(b)).
- Odd versus even extension. A fixed end needs the odd extension (reflection upside down), a free end the even one. Using the wrong one gives a function that violates the boundary condition.
- Classical solutions need compatibility. For the fixed string, $f''(0)=f''(L)=0$ is needed for $u\in C^2$; the plucked string is only a generalised solution, with kinks moving along characteristics.
Exercises
- 1 Solve $u_t=2u_{xx}$, $0\lt x\lt\pi$, $u(0,t)=u(\pi,t)=0$, $u(x,0)=3\sin x-\sin4x$.
- 2 Solve $u_t=u_{xx}$, $0\lt x\lt1$, $u(0,t)=u(1,t)=0$, $u(x,0)=\sin^3\pi x$.
- 3 A rod $0\le x\le L$ at uniform temperature $T_0$ has its ends plunged into ice ($u=0$) at $t=0$. Find $u(x,t)$, and the centre temperature at the time when $k\pi^2t/L^2=1$.
- 4 Solve the insulated-rod problem on $[0,\pi]$ with $u(x,0)=\cos^2x$.
- 5 For $u(0,t)=0$, $u_x(L,t)=0$, find the eigenfunctions and the decay rate of the slowest mode, and compare with two Dirichlet ends.
- 6 Find the steady state of $u_t=ku_{xx}+q$ ($q$ a constant heat source) with $u(0,t)=u(L,t)=0$, and its maximum.
- 7 Find the steady state of $u_t=u_{xx}$ on $(0,1)$ with $u(0,t)=0$ and the Robin condition $u_x(1,t)+u(1,t)=1$.
- 8 Show that $x(\pi-x)\le\pi\sin x$ on $[0,\pi]$, and deduce from the comparison principle that the solution with zero ends and $u(x,0)=x(\pi-x)$ satisfies $0\le u(x,t)\le\pi e^{-kt}\sin x$.
- 9 For $u=\sum b_ne^{-kn^2t}\sin nx$ on $[0,\pi]$, show $E(t)=\int_0^\pi u^2\,dx=\frac\pi2\sum b_n^2e^{-2kn^2t}$ and deduce $E(t)\le e^{-2kt}E(0)$, in agreement with Theorem 4.22.
- 10 Modelling. A copper wire of length $L=0.5\ \text{m}$ ($k=1.1\times10^{-4}\ \text{m}^2/\text{s}$) with ends at $0\,^\circ$C has an initial temperature whose first sine coefficient is $b_1=40\,^\circ$C. Estimate how long it takes the first mode to fall below $1\,^\circ$C, and how much faster the second mode decays.
- 1 Solve $u_{tt}=u_{xx}$ on $\mathbb{R}$ with $u(x,0)=\sin x$, $u_t(x,0)=\cos x$.
- 2 Solve $u_{tt}=9u_{xx}$ with $u(x,0)=x^2$, $u_t(x,0)=x$, and check your answer.
- 3 For $u_{tt}=4u_{xx}$ with $f=0$ and $g=1$ on $(-1,1)$, $g=0$ elsewhere, find $u(0,t)$ for all $t\ge0$ and sketch $u(\cdot,t)$ for $t=\frac14$ and $t=2$.
- 4 For $c=5$ and data vanishing outside $[0,2]$, when can the point $x=22$ first be disturbed? Describe the region where $u$ may be non-zero.
- 5 A semi-infinite string ($c=1$) fixed at $x=0$ starts from rest with $f(x)=x^3$. Find $u$ and $u(1,2)$, and check the boundary condition.
- 6 Show that for a free end ($u_x(0,t)=0$) the even extension of the data gives the solution, and that a pulse is reflected without inversion.
- 7 For which $c$ is $u=e^{-(x-4t)^2}$ a solution of $u_{tt}=c^2u_{xx}$? Describe it.
- 8 (Parallelogram rule) Let $ABCD$ be a parallelogram whose sides are characteristic lines of $u_{tt}=c^2u_{xx}$, with $A,C$ opposite. Prove $u(A)+u(C)=u(B)+u(D)$ for every $C^2$ solution.
- 1 Solve $u_{tt}=u_{xx}$ on $(0,\pi)$ with fixed ends, $u(x,0)=0$, $u_t(x,0)=\sin x+\sin2x$.
- 2 Solve $u_{tt}=4u_{xx}$ on $(0,1)$ with fixed ends, $u(x,0)=0.1\sin\pi x+0.05\sin3\pi x$, $u_t(x,0)=0$. What is the period?
- 3 A string of length $L$ is plucked at $x=\frac L3$ to height $h$ and released from rest. Show $a_n=\dfrac{9h}{n^2\pi^2}\sin\dfrac{n\pi}3$, and identify the missing harmonics.
- 4 (Struck string) A string at rest in its equilibrium position is struck so that $g=v_0$ on $|x-\frac L2|\lt\delta$ and $g=0$ elsewhere. Find $b_n$.
- 5 Modelling. A guitar string has vibrating length $0.65\ \text{m}$, tension $70\ \text{N}$ and linear density $0.40\ \text{g/m}$. Find its fundamental frequency, and the tension needed to tune it to $329.63\ \text{Hz}$.
- 6 Show that the single mode $u=A\sin\frac{n\pi x}L\cos\omega_nt$ has energy $\frac L4A^2\omega_n^2$, and find how it splits between kinetic and potential energy at time $t$.
- 7 Solve the free–free string ($u_x(0,t)=u_x(L,t)=0$) by separation, and interpret the mode $n=0$.
- 8 For the damped string $u_{tt}+2\beta u_t=c^2u_{xx}$ ($\beta\gt0$) with fixed ends, show that $E(t)=\frac12\int_0^L(u_t^2+c^2u_x^2)\,dx$ decreases, and find the separated solutions when $\beta\lt\omega_1$.
- 1 Verify that $\Phi(x,t)=(4\pi kt)^{-1/2}e^{-x^2/(4kt)}$ satisfies $\Phi_t=k\Phi_{xx}$ and $\int_{-\infty}^\infty\Phi\,dx=1$.
- 2 Solve the heat equation on $\mathbb{R}$ with $u(x,0)=e^{-x^2}$.
- 3 Solve it with $u(x,0)=1$ for $x\gt0$, $0$ for $x\lt0$, and show $u(x,t)\gt0$ for every $x$ when $t\gt0$.
- 4 Prove with the energy method that the heat problem with both ends insulated has at most one classical solution.
- 5 Show that the wave equation obeys no maximum principle: find a solution on $[0,\pi]\times[0,T]$ whose maximum is not attained on the bottom or the sides.
- 6 For $u_{tt}=c^2u_{xx}$ on $(0,L)$ with $u(0,t)=0$ and $u_x(L,t)=-\sigma u(L,t)$ ($\sigma\gt0$, an elastic support), show that $E(t)=\frac12\int_0^L(u_t^2+c^2u_x^2)\,dx+\frac{c^2\sigma}2u(L,t)^2$ is conserved.
- 7 (Equipartition) For the series solution of Theorem 4.25, show that the time averages of the kinetic and potential energies over one period are equal.
- 8 (Ångström's method) Show that $u=Ae^{-\kappa x}\cos(\omega t-\kappa x)$ with $\kappa=\sqrt{\omega/(2k)}$ solves $u_t=ku_{xx}$ for $x\gt0$ and satisfies $u(0,t)=A\cos\omega t$. How can $k$ be found from measurements at two points?
E1. CSIR Part B GATE MA Let $u$ solve $u_t=u_{xx}$ on $0\lt x\lt\pi$, $t\gt0$, with $u(0,t)=u(\pi,t)=0$ and $u(x,0)=\sin x+2\sin3x$. Then $u(\frac\pi2,1)$ equals (A) $e^{-1}+2e^{-9}$; (B) $e^{-1}-2e^{-9}$; (C) $e^{-1}-2e^{-3}$; (D) $3e^{-1}$.
E2. CSIR Part B Let $u$ solve $u_{tt}=4u_{xx}$ on $\mathbb{R}$ with $u(x,0)=x^2$ and $u_t(x,0)=\cos x$. Then $u(0,\frac\pi4)$ equals (A) $\frac{\pi^2}4+\frac12$; (B) $\frac{\pi^2}{16}+\frac12$; (C) $\frac{\pi^2}4+1$; (D) $\frac{\pi^2}4$.
E3. CSIR Part B A rod $0\le x\le2$ with both ends insulated has initial temperature $u(x,0)=x$ and satisfies $u_t=ku_{xx}$. Then $\lim_{t\to\infty}u(1.5,t)$ is (A) $0$; (B) $1$; (C) $1.5$; (D) a number depending on $k$.
E4. CSIR Part C Let $u$ solve $u_t=u_{xx}$ on $0\lt x\lt1$, $t\gt0$, $u(0,t)=u(1,t)=0$, $u(x,0)=f(x)$, where $f\not\equiv0$ is continuous and piecewise $C^1$ with $f(0)=f(1)=0$, and let $b_1=2\int_0^1f(x)\sin\pi x\,dx$. Which of the following are true? (One or more options may be correct.) (A) $\max_x|u(x,t)|\le\max|f|$ for all $t\gt0$. (B) $\int_0^1u(x,t)\,dx$ does not depend on $t$. (C) $e^{\pi^2t}u(x,t)\to b_1\sin\pi x$ uniformly in $x$ as $t\to\infty$. (D) If $f$ changes sign in $(0,1)$, then $u(\cdot,t)$ changes sign in $(0,1)$ for every $t\gt0$.
E5. CSIR Part C Let $u$ solve $u_{tt}=u_{xx}$ on $\mathbb{R}\times(0,\infty)$ with $u(x,0)=f(x)$, $u_t(x,0)=g(x)$, where $f,g$ are smooth and vanish outside $[-1,1]$. Which of the following are true? (A) $u(x,t)=0$ whenever $|x|\gt1+t$. (B) For every $x$, $u(x,t)\to0$ as $t\to\infty$. (C) If $g\equiv0$, then $u(0,t)=0$ for all $t\ge1$. (D) $\frac12\int_{-\infty}^{\infty}(u_t^2+u_x^2)\,dx$ is independent of $t$.
E6. CSIR Part C GATE MA Let $u$ solve $u_{tt}=u_{xx}$ on $0\lt x\lt\pi$ with $u(0,t)=u(\pi,t)=0$, $u(x,0)=\sin2x$, $u_t(x,0)=\sin x$. Which of the following are true? (A) $u$ is periodic in $t$ with period $2\pi$. (B) $u(\frac\pi2,t)=\sin t$ for all $t$. (C) $\frac12\int_0^\pi(u_t^2+u_x^2)\,dx=\frac{5\pi}4$ for all $t$. (D) $u(x,\frac\pi4)=0$ for all $x$.
Suggested plan — 3 lectures × 60 min.
- Heat. Derive the heat equation from conservation and Fourier's law (ten minutes, with the three kinds of end condition). Separation and Theorem 4.19 — insist on the role of $\sum|b_n|\lt\infty$ from §4.2. Examples 1 and 2; project Interactive 1, first Dirichlet then insulated ends with the same data, and ask for the final temperature before pressing Play. Then the mixed rod.
- Principles. Theorem 4.20 (the $\varepsilon x^2$ trick is the same as in §4.2 — point this out), Theorems 4.21–4.22 and Example 3; the backward-heat example is worth a dramatic numerical illustration ($n=10$, $T=1$: a factor $e^{100}$). Then the wave equation: derivation from the string, Theorem 4.23 via §4.1's canonical form, Definition 4.11 and Theorem 4.24, Example 4.
- Strings. Reflection by images and Example 5; project Interactive 2 (cases 1–3), watching the characteristic triangle and the energy readouts. Theorems 4.25–4.26 and Example 6 (compute the energy both ways on the board). Example 7, the heat kernel (Theorem 4.27) and the heat-versus-wave table. End with E4 and E5 live.
Discussion questions. (1) Why does a smooth temperature profile appear instantly even from a discontinuous one, while a kink in a string never disappears? Relate this to the factors $e^{-kn^2t}$ and $\cos\omega_nt$. (2) The heat kernel is positive everywhere for $t\gt0$: heat "travels at infinite speed". Is this physical? What model would fix it, and of which type would it be? (3) Plucking a guitar string near the bridge sounds brighter than plucking it in the middle. Explain with the coefficients of Exercise III.3.
Misconceptions to watch for. Using $e^{-knt}$ instead of $e^{-kn^2\pi^2t/L^2}$; final temperature $0$ for insulated rods; forgetting to subtract the steady state before expanding; applying the maximum principle to the wave equation; $\frac1c$ for $\frac1{2c}$ in d'Alembert's formula; even extension for a fixed end; believing kinetic energy alone is conserved; thinking the heat equation can be run backwards.
Suggested homework. I(2, 3, 4, 5, 8), II(2, 3, 5, 8), III(2, 3, 5, 6), IV(2, 3, 6, 8); Exam corner E2, E4, E6.
Research question. How deep does winter reach into the ground? The surface temperature of soil rises and falls with the day and with the year. Model the soil as a semi-infinite rod, $u_t=ku_{xx}$ for depth $x\gt0$, with a periodic surface temperature, and find how fast the daily and the annual temperature waves are damped with depth, and how much they lag behind the surface.
Why it matters. The answer decides how deep water pipes must be buried in cold countries, why traditional wine cellars and root cellars a few metres underground stay at nearly constant temperature, how permafrost responds to climate change, and how ground-source heat pumps are designed. The same mathematics is the basis of Anders Ångström's 1861 method for measuring thermal diffusivity, still used in materials laboratories (as the laser-flash and "Ångström" methods).
What to do.
- Verify (Exercise IV.8) that $u=Ae^{-\kappa x}\cos(\omega t-\kappa x)$, $\kappa=\sqrt{\omega/(2k)}$, solves the heat equation with surface temperature $A\cos\omega t$. Interpret the factor $e^{-\kappa x}$ and the phase $\kappa x$.
- For a typical soil, $k\approx5\times10^{-7}\ \text{m}^2/\text{s}$. Compute $\kappa$ for the daily cycle and for the annual cycle, the depth at which the amplitude drops to $1/e$ of its surface value, and the depth at which the annual wave is exactly half a year out of phase.
- Find temperature records for one site at two depths (many national meteorological services and soil-monitoring networks publish them). Estimate $k$ from the amplitude ratio and the phase lag using the formula of Exercise IV.8. Do the two estimates (from amplitude and from phase) agree?
- Why is the solution not unique without a condition at large depth? Which condition is physically right?
Tools: A calculator or spreadsheet; Python with NumPy for fitting sinusoids to real data (least squares on $a\cos\omega t+b\sin\omega t$).
Hand in: A two-page note with the verification, a table of $\kappa$, damping depths and lags for the daily and annual cycles, and your estimate of $k$ from real data with a comment on its reliability.
Going further. Real soil has layers with different diffusivities. Solve the two-layer problem with continuity of temperature and heat flux at the interface; which layer controls the damping?
🧪 Self-check quiz
Q1. For $u_t=ku_{xx}$ on $(0,L)$ with $u=0$ at both ends, the $n$-th Fourier mode decays like
Q2. The solution of $u_{tt}=4u_{xx}$ on $\mathbb{R}$ with $u(x,0)=x$, $u_t(x,0)=1$ is
Q3. For $u_{tt}=c^2u_{xx}$ on the line, the value $u(x_0,t_0)$ depends on the initial velocity only through its values on
Q4. A rod with both ends insulated starts with temperature $f(x)$ on $[0,L]$. As $t\to\infty$ its temperature tends to
Q5. For a vibrating string with fixed ends, which quantity is constant in time?